xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 12545b85834f32e0fdb831489503c3c4f6db715a) 
1 /*-
2  * Copyright (c) 2016-2020 Netflix, Inc.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  */
26 
27 #include <sys/cdefs.h>
28 #include "opt_inet.h"
29 #include "opt_inet6.h"
30 #include "opt_ipsec.h"
31 #include "opt_ratelimit.h"
32 #include "opt_kern_tls.h"
33 #if defined(INET) || defined(INET6)
34 #include <sys/param.h>
35 #include <sys/arb.h>
36 #include <sys/module.h>
37 #include <sys/kernel.h>
38 #ifdef TCP_HHOOK
39 #include <sys/hhook.h>
40 #endif
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/mbuf.h>
46 #include <sys/proc.h>		/* for proc0 declaration */
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/sysctl.h>
50 #include <sys/systm.h>
51 #ifdef STATS
52 #include <sys/qmath.h>
53 #include <sys/tree.h>
54 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
55 #else
56 #include <sys/tree.h>
57 #endif
58 #include <sys/refcount.h>
59 #include <sys/queue.h>
60 #include <sys/tim_filter.h>
61 #include <sys/smp.h>
62 #include <sys/kthread.h>
63 #include <sys/kern_prefetch.h>
64 #include <sys/protosw.h>
65 #ifdef TCP_ACCOUNTING
66 #include <sys/sched.h>
67 #include <machine/cpu.h>
68 #endif
69 #include <vm/uma.h>
70 
71 #include <net/route.h>
72 #include <net/route/nhop.h>
73 #include <net/vnet.h>
74 
75 #define TCPSTATES		/* for logging */
76 
77 #include <netinet/in.h>
78 #include <netinet/in_kdtrace.h>
79 #include <netinet/in_pcb.h>
80 #include <netinet/ip.h>
81 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
82 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
83 #include <netinet/ip_var.h>
84 #include <netinet/ip6.h>
85 #include <netinet6/in6_pcb.h>
86 #include <netinet6/ip6_var.h>
87 #include <netinet/tcp.h>
88 #define	TCPOUTFLAGS
89 #include <netinet/tcp_fsm.h>
90 #include <netinet/tcp_seq.h>
91 #include <netinet/tcp_timer.h>
92 #include <netinet/tcp_var.h>
93 #include <netinet/tcp_log_buf.h>
94 #include <netinet/tcp_syncache.h>
95 #include <netinet/tcp_hpts.h>
96 #include <netinet/tcp_ratelimit.h>
97 #include <netinet/tcp_accounting.h>
98 #include <netinet/tcpip.h>
99 #include <netinet/cc/cc.h>
100 #include <netinet/cc/cc_newreno.h>
101 #include <netinet/tcp_fastopen.h>
102 #include <netinet/tcp_lro.h>
103 #ifdef NETFLIX_SHARED_CWND
104 #include <netinet/tcp_shared_cwnd.h>
105 #endif
106 #ifdef TCP_OFFLOAD
107 #include <netinet/tcp_offload.h>
108 #endif
109 #ifdef INET6
110 #include <netinet6/tcp6_var.h>
111 #endif
112 #include <netinet/tcp_ecn.h>
113 
114 #include <netipsec/ipsec_support.h>
115 
116 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
117 #include <netipsec/ipsec.h>
118 #include <netipsec/ipsec6.h>
119 #endif				/* IPSEC */
120 
121 #include <netinet/udp.h>
122 #include <netinet/udp_var.h>
123 #include <machine/in_cksum.h>
124 
125 #ifdef MAC
126 #include <security/mac/mac_framework.h>
127 #endif
128 #include "sack_filter.h"
129 #include "tcp_rack.h"
130 #include "tailq_hash.h"
131 #include "rack_bbr_common.h"
132 
133 uma_zone_t rack_zone;
134 uma_zone_t rack_pcb_zone;
135 
136 #ifndef TICKS2SBT
137 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
138 #endif
139 
140 VNET_DECLARE(uint32_t, newreno_beta);
141 VNET_DECLARE(uint32_t, newreno_beta_ecn);
142 #define V_newreno_beta VNET(newreno_beta)
143 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
144 
145 #define	M_TCPFSB	__CONCAT(M_TCPFSB, STACKNAME)
146 #define	M_TCPDO		__CONCAT(M_TCPDO, STACKNAME)
147 
148 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb_" __XSTRING(STACKNAME), "TCP fast send block");
149 MALLOC_DEFINE(M_TCPDO, "tcp_do_" __XSTRING(STACKNAME), "TCP deferred options");
150 MALLOC_DEFINE(M_TCPPCM, "tcp_pcm_" __XSTRING(STACKNAME), "TCP PCM measurement information");
151 
152 struct sysctl_ctx_list rack_sysctl_ctx;
153 struct sysctl_oid *rack_sysctl_root;
154 
155 #define CUM_ACKED 1
156 #define SACKED 2
157 
158 /*
159  * The RACK module incorporates a number of
160  * TCP ideas that have been put out into the IETF
161  * over the last few years:
162  * - Matt Mathis's Rate Halving which slowly drops
163  *    the congestion window so that the ack clock can
164  *    be maintained during a recovery.
165  * - Yuchung Cheng's RACK TCP (for which its named) that
166  *    will stop us using the number of dup acks and instead
167  *    use time as the gage of when we retransmit.
168  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
169  *    of Dukkipati et.al.
170  * RACK depends on SACK, so if an endpoint arrives that
171  * cannot do SACK the state machine below will shuttle the
172  * connection back to using the "default" TCP stack that is
173  * in FreeBSD.
174  *
175  * To implement RACK the original TCP stack was first decomposed
176  * into a functional state machine with individual states
177  * for each of the possible TCP connection states. The do_segment
178  * functions role in life is to mandate the connection supports SACK
179  * initially and then assure that the RACK state matches the conenction
180  * state before calling the states do_segment function. Each
181  * state is simplified due to the fact that the original do_segment
182  * has been decomposed and we *know* what state we are in (no
183  * switches on the state) and all tests for SACK are gone. This
184  * greatly simplifies what each state does.
185  *
186  * TCP output is also over-written with a new version since it
187  * must maintain the new rack scoreboard.
188  *
189  */
190 static int32_t rack_tlp_thresh = 1;
191 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
192 static int32_t rack_tlp_use_greater = 1;
193 static int32_t rack_reorder_thresh = 2;
194 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
195 						 * - 60 seconds */
196 static uint16_t rack_policer_rxt_thresh= 0;	/* 499 = 49.9%, 0 is off  */
197 static uint8_t rack_policer_avg_thresh = 0; /* 3.2 */
198 static uint8_t rack_policer_med_thresh = 0; /* 1 - 16 */
199 static uint16_t rack_policer_bucket_reserve = 20; /* How much % is reserved in the bucket */
200 static uint64_t rack_pol_min_bw = 125000;	/* 1mbps in Bytes per sec */
201 static uint32_t rack_policer_data_thresh = 64000;	/* 64,000 bytes must be sent before we engage */
202 static uint32_t rack_policing_do_bw_comp = 1;
203 static uint32_t rack_pcm_every_n_rounds = 100;
204 static uint32_t rack_pcm_blast = 0;
205 static uint32_t rack_pcm_is_enabled = 1;
206 static uint8_t rack_req_del_mss = 18;	/* How many segments need to be sent in a recovery episode to do policer_detection */
207 static uint8_t rack_ssthresh_rest_rto_rec = 0; /* Do we restore ssthresh when we have rec -> rto -> rec */
208 
209 static uint32_t rack_gp_gain_req = 1200;		/* Amount percent wise required to gain to record a round has "gaining" */
210 static uint32_t rack_rnd_cnt_req = 0x10005;		/* Default number of rounds if we are below rack_gp_gain_req where we exit ss */
211 
212 
213 static int32_t rack_rxt_scoreboard_clear_thresh = 2;
214 static int32_t rack_dnd_default = 0;		/* For rr_conf = 3, what is the default for dnd */
215 static int32_t rack_rxt_controls = 0;
216 static int32_t rack_fill_cw_state = 0;
217 static uint8_t rack_req_measurements = 1;
218 /* Attack threshold detections */
219 static uint32_t rack_highest_sack_thresh_seen = 0;
220 static uint32_t rack_highest_move_thresh_seen = 0;
221 static uint32_t rack_merge_out_sacks_on_attack = 0;
222 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
223 static int32_t rack_hw_pace_extra_slots = 0;	/* 2 extra MSS time betweens */
224 static int32_t rack_hw_rate_caps = 0; /* 1; */
225 static int32_t rack_hw_rate_cap_per = 0;	/* 0 -- off  */
226 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
227 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
228 static int32_t rack_hw_up_only = 0;
229 static int32_t rack_stats_gets_ms_rtt = 1;
230 static int32_t rack_prr_addbackmax = 2;
231 static int32_t rack_do_hystart = 0;
232 static int32_t rack_apply_rtt_with_reduced_conf = 0;
233 static int32_t rack_hibeta_setting = 0;
234 static int32_t rack_default_pacing_divisor = 250;
235 static uint16_t rack_pacing_min_seg = 0;
236 static int32_t rack_timely_off = 0;
237 
238 static uint32_t sad_seg_size_per = 800;	/* 80.0 % */
239 static int32_t rack_pkt_delay = 1000;
240 static int32_t rack_send_a_lot_in_prr = 1;
241 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
242 static int32_t rack_verbose_logging = 0;
243 static int32_t rack_ignore_data_after_close = 1;
244 static int32_t rack_enable_shared_cwnd = 1;
245 static int32_t rack_use_cmp_acks = 1;
246 static int32_t rack_use_fsb = 1;
247 static int32_t rack_use_rfo = 1;
248 static int32_t rack_use_rsm_rfo = 1;
249 static int32_t rack_max_abc_post_recovery = 2;
250 static int32_t rack_client_low_buf = 0;
251 static int32_t rack_dsack_std_based = 0x3;	/* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */
252 static int32_t rack_bw_multipler = 0;		/* Limit on fill cw's jump up to be this x gp_est */
253 #ifdef TCP_ACCOUNTING
254 static int32_t rack_tcp_accounting = 0;
255 #endif
256 static int32_t rack_limits_scwnd = 1;
257 static int32_t rack_enable_mqueue_for_nonpaced = 0;
258 static int32_t rack_hybrid_allow_set_maxseg = 0;
259 static int32_t rack_disable_prr = 0;
260 static int32_t use_rack_rr = 1;
261 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
262 static int32_t rack_persist_min = 250000;	/* 250usec */
263 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
264 static int32_t rack_honors_hpts_min_to =  1;	/* Do we honor the hpts minimum time out for pacing timers */
265 static uint32_t rack_max_reduce = 10;		/* Percent we can reduce slot by */
266 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
267 static int32_t rack_limit_time_with_srtt = 0;
268 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
269 static int32_t rack_enobuf_hw_boost_mult = 0;	/* How many times the hw rate we boost slot using time_between */
270 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
271 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
272 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
273 static int32_t rack_hw_check_queue = 0;		/* Do we always pre-check queue depth of a hw queue */
274 static int32_t rack_full_buffer_discount = 10;
275 /*
276  * Currently regular tcp has a rto_min of 30ms
277  * the backoff goes 12 times so that ends up
278  * being a total of 122.850 seconds before a
279  * connection is killed.
280  */
281 static uint32_t rack_def_data_window = 20;
282 static uint32_t rack_goal_bdp = 2;
283 static uint32_t rack_min_srtts = 1;
284 static uint32_t rack_min_measure_usec = 0;
285 static int32_t rack_tlp_min = 10000;	/* 10ms */
286 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
287 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
288 static const int32_t rack_free_cache = 2;
289 static int32_t rack_hptsi_segments = 40;
290 static int32_t rack_rate_sample_method = USE_RTT_LOW;
291 static int32_t rack_pace_every_seg = 0;
292 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
293 static int32_t rack_slot_reduction = 4;
294 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
295 static int32_t rack_cwnd_block_ends_measure = 0;
296 static int32_t rack_rwnd_block_ends_measure = 0;
297 static int32_t rack_def_profile = 0;
298 
299 static int32_t rack_lower_cwnd_at_tlp = 0;
300 static int32_t rack_always_send_oldest = 0;
301 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
302 
303 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
304 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
305 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
306 
307 /* Probertt */
308 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
309 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
310 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
311 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
312 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
313 
314 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
315 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
316 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
317 static uint32_t rack_probertt_use_min_rtt_exit = 0;
318 static uint32_t rack_probe_rtt_sets_cwnd = 0;
319 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
320 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
321 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
322 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
323 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
324 static uint32_t rack_probertt_filter_life = 10000000;
325 static uint32_t rack_probertt_lower_within = 10;
326 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
327 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
328 static int32_t rack_probertt_clear_is = 1;
329 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
330 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
331 
332 /* Part of pacing */
333 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
334 
335 /* Timely information:
336  *
337  * Here we have various control parameters on how
338  * timely may change the multiplier. rack_gain_p5_ub
339  * is associated with timely but not directly influencing
340  * the rate decision like the other variables. It controls
341  * the way fill-cw interacts with timely and caps how much
342  * timely can boost the fill-cw b/w.
343  *
344  * The other values are various boost/shrink numbers as well
345  * as potential caps when adjustments are made to the timely
346  * gain (returned by rack_get_output_gain(). Remember too that
347  * the gain returned can be overriden by other factors such as
348  * probeRTT as well as fixed-rate-pacing.
349  */
350 static int32_t rack_gain_p5_ub = 250;
351 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
352 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
353 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
354 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
355 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
356 static int32_t rack_gp_decrease_per = 80;	/* Beta value of timely decrease (.8) = 80 */
357 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
358 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
359 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
360 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
361 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
362 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
363 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
364 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
365 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
366 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
367 static int32_t rack_use_max_for_nobackoff = 0;
368 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
369 static int32_t rack_timely_no_stopping = 0;
370 static int32_t rack_down_raise_thresh = 100;
371 static int32_t rack_req_segs = 1;
372 static uint64_t rack_bw_rate_cap = 0;
373 static uint64_t rack_fillcw_bw_cap = 3750000;	/* Cap fillcw at 30Mbps */
374 
375 
376 /* Rack specific counters */
377 counter_u64_t rack_saw_enobuf;
378 counter_u64_t rack_saw_enobuf_hw;
379 counter_u64_t rack_saw_enetunreach;
380 counter_u64_t rack_persists_sends;
381 counter_u64_t rack_persists_acks;
382 counter_u64_t rack_persists_loss;
383 counter_u64_t rack_persists_lost_ends;
384 counter_u64_t rack_total_bytes;
385 #ifdef INVARIANTS
386 counter_u64_t rack_adjust_map_bw;
387 #endif
388 /* Tail loss probe counters */
389 counter_u64_t rack_tlp_tot;
390 counter_u64_t rack_tlp_newdata;
391 counter_u64_t rack_tlp_retran;
392 counter_u64_t rack_tlp_retran_bytes;
393 counter_u64_t rack_to_tot;
394 counter_u64_t rack_hot_alloc;
395 counter_u64_t tcp_policer_detected;
396 counter_u64_t rack_to_alloc;
397 counter_u64_t rack_to_alloc_hard;
398 counter_u64_t rack_to_alloc_emerg;
399 counter_u64_t rack_to_alloc_limited;
400 counter_u64_t rack_alloc_limited_conns;
401 counter_u64_t rack_split_limited;
402 counter_u64_t rack_rxt_clamps_cwnd;
403 counter_u64_t rack_rxt_clamps_cwnd_uniq;
404 
405 counter_u64_t rack_multi_single_eq;
406 counter_u64_t rack_proc_non_comp_ack;
407 
408 counter_u64_t rack_fto_send;
409 counter_u64_t rack_fto_rsm_send;
410 counter_u64_t rack_nfto_resend;
411 counter_u64_t rack_non_fto_send;
412 counter_u64_t rack_extended_rfo;
413 
414 counter_u64_t rack_sack_proc_all;
415 counter_u64_t rack_sack_proc_short;
416 counter_u64_t rack_sack_proc_restart;
417 counter_u64_t rack_sack_attacks_detected;
418 counter_u64_t rack_sack_attacks_reversed;
419 counter_u64_t rack_sack_attacks_suspect;
420 counter_u64_t rack_sack_used_next_merge;
421 counter_u64_t rack_sack_splits;
422 counter_u64_t rack_sack_used_prev_merge;
423 counter_u64_t rack_sack_skipped_acked;
424 counter_u64_t rack_ack_total;
425 counter_u64_t rack_express_sack;
426 counter_u64_t rack_sack_total;
427 counter_u64_t rack_move_none;
428 counter_u64_t rack_move_some;
429 
430 counter_u64_t rack_input_idle_reduces;
431 counter_u64_t rack_collapsed_win;
432 counter_u64_t rack_collapsed_win_seen;
433 counter_u64_t rack_collapsed_win_rxt;
434 counter_u64_t rack_collapsed_win_rxt_bytes;
435 counter_u64_t rack_try_scwnd;
436 counter_u64_t rack_hw_pace_init_fail;
437 counter_u64_t rack_hw_pace_lost;
438 
439 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
440 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
441 
442 
443 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
444 
445 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
446 	(tv) = (value) + slop;	 \
447 	if ((u_long)(tv) < (u_long)(tvmin)) \
448 		(tv) = (tvmin); \
449 	if ((u_long)(tv) > (u_long)(tvmax)) \
450 		(tv) = (tvmax); \
451 } while (0)
452 
453 static void
454 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
455 
456 static int
457 rack_process_ack(struct mbuf *m, struct tcphdr *th,
458     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
459     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val, int32_t orig_tlen);
460 static int
461 rack_process_data(struct mbuf *m, struct tcphdr *th,
462     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
463     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
464 static void
465 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
466    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
467 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
468 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
469     uint8_t limit_type);
470 static struct rack_sendmap *
471 rack_check_recovery_mode(struct tcpcb *tp,
472     uint32_t tsused);
473 static uint32_t
474 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack);
475 static void
476 rack_cong_signal(struct tcpcb *tp,
477 		 uint32_t type, uint32_t ack, int );
478 static void rack_counter_destroy(void);
479 static int
480 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
481 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
482 static void
483 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
484 static void
485 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
486     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
487 static void rack_dtor(void *mem, int32_t size, void *arg);
488 static void
489 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
490     uint32_t flex1, uint32_t flex2,
491     uint32_t flex3, uint32_t flex4,
492     uint32_t flex5, uint32_t flex6,
493     uint16_t flex7, uint8_t mod);
494 
495 static void
496 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
497    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
498    struct rack_sendmap *rsm, uint8_t quality);
499 static struct rack_sendmap *
500 rack_find_high_nonack(struct tcp_rack *rack,
501     struct rack_sendmap *rsm);
502 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
503 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
504 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
505 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
506 static void
507 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
508 			    tcp_seq th_ack, int line, uint8_t quality);
509 static void
510 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
511 
512 static uint32_t
513 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
514 static int32_t rack_handoff_ok(struct tcpcb *tp);
515 static int32_t rack_init(struct tcpcb *tp, void **ptr);
516 static void rack_init_sysctls(void);
517 
518 static void
519 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
520     struct tcphdr *th, int entered_rec, int dup_ack_struck,
521     int *dsack_seen, int *sacks_seen);
522 static void
523 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
524     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
525     struct rack_sendmap *hintrsm, uint32_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
526 
527 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
528 
529 
530 static void
531 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
532     struct rack_sendmap *rsm, uint32_t cts);
533 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
534 static int32_t rack_output(struct tcpcb *tp);
535 
536 static uint32_t
537 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
538     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
539     uint32_t cts, uint32_t segsiz);
540 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
541 static void rack_remxt_tmr(struct tcpcb *tp);
542 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
543 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
544 static int32_t rack_stopall(struct tcpcb *tp);
545 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
546 static uint32_t
547 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
548     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint32_t add_flag, int segsiz);
549 static void
550 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
551     struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz);
552 static int
553 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
554     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
555 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
556 static int
557 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
558     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
559     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
560 
561 static void
562 rack_peg_rxt(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t segsiz);
563 
564 static int
565 rack_do_closing(struct mbuf *m, struct tcphdr *th,
566     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
567     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
568 static int
569 rack_do_established(struct mbuf *m, struct tcphdr *th,
570     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
571     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
572 static int
573 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
574     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
575     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
576 static int
577 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
578     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
579     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
580 static int
581 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
582     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
583     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
584 static int
585 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
586     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
587     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
588 static int
589 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
590     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
591     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
592 static int
593 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
594     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
595     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
596 static void rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
597 struct rack_sendmap *
598 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
599     uint32_t tsused);
600 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
601     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
602 static void
603      tcp_rack_partialack(struct tcpcb *tp);
604 static int
605 rack_set_profile(struct tcp_rack *rack, int prof);
606 static void
607 rack_apply_deferred_options(struct tcp_rack *rack);
608 
609 int32_t rack_clear_counter=0;
610 
611 static uint64_t
612 rack_get_lt_bw(struct tcp_rack *rack)
613 {
614 	struct timeval tv;
615 	uint64_t tim, bytes;
616 
617 	tim = rack->r_ctl.lt_bw_time;
618 	bytes = rack->r_ctl.lt_bw_bytes;
619 	if (rack->lt_bw_up) {
620 		/* Include all the current bytes too */
621 		microuptime(&tv);
622 		bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
623 		tim += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
624 	}
625 	if ((bytes != 0) && (tim != 0))
626 		return ((bytes * (uint64_t)1000000) / tim);
627 	else
628 		return (0);
629 }
630 
631 static void
632 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
633 {
634 	struct sockopt sopt;
635 	struct cc_newreno_opts opt;
636 	struct newreno old;
637 	struct tcpcb *tp;
638 	int error, failed = 0;
639 
640 	tp = rack->rc_tp;
641 	if (tp->t_cc == NULL) {
642 		/* Tcb is leaving */
643 		return;
644 	}
645 	rack->rc_pacing_cc_set = 1;
646 	if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
647 		/* Not new-reno we can't play games with beta! */
648 		failed = 1;
649 		goto out;
650 
651 	}
652 	if (CC_ALGO(tp)->ctl_output == NULL)  {
653 		/* Huh, not using new-reno so no swaps.? */
654 		failed = 2;
655 		goto out;
656 	}
657 	/* Get the current values out */
658 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
659 	sopt.sopt_dir = SOPT_GET;
660 	opt.name = CC_NEWRENO_BETA;
661 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
662 	if (error)  {
663 		failed = 3;
664 		goto out;
665 	}
666 	old.beta = opt.val;
667 	opt.name = CC_NEWRENO_BETA_ECN;
668 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
669 	if (error)  {
670 		failed = 4;
671 		goto out;
672 	}
673 	old.beta_ecn = opt.val;
674 
675 	/* Now lets set in the values we have stored */
676 	sopt.sopt_dir = SOPT_SET;
677 	opt.name = CC_NEWRENO_BETA;
678 	opt.val = rack->r_ctl.rc_saved_beta.beta;
679 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
680 	if (error)  {
681 		failed = 5;
682 		goto out;
683 	}
684 	opt.name = CC_NEWRENO_BETA_ECN;
685 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
686 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
687 	if (error) {
688 		failed = 6;
689 		goto out;
690 	}
691 	/* Save off the values for restoral */
692 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
693 out:
694 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
695 		union tcp_log_stackspecific log;
696 		struct timeval tv;
697 		struct newreno *ptr;
698 
699 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
700 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
701 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
702 		log.u_bbr.flex1 = ptr->beta;
703 		log.u_bbr.flex2 = ptr->beta_ecn;
704 		log.u_bbr.flex3 = ptr->newreno_flags;
705 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
706 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
707 		log.u_bbr.flex6 = failed;
708 		log.u_bbr.flex7 = rack->gp_ready;
709 		log.u_bbr.flex7 <<= 1;
710 		log.u_bbr.flex7 |= rack->use_fixed_rate;
711 		log.u_bbr.flex7 <<= 1;
712 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
713 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
714 		log.u_bbr.flex8 = flex8;
715 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
716 			       0, &log, false, NULL, NULL, 0, &tv);
717 	}
718 }
719 
720 static void
721 rack_set_cc_pacing(struct tcp_rack *rack)
722 {
723 	if (rack->rc_pacing_cc_set)
724 		return;
725 	/*
726 	 * Use the swap utility placing in 3 for flex8 to id a
727 	 * set of a new set of values.
728 	 */
729 	rack->rc_pacing_cc_set = 1;
730 	rack_swap_beta_values(rack, 3);
731 }
732 
733 static void
734 rack_undo_cc_pacing(struct tcp_rack *rack)
735 {
736 	if (rack->rc_pacing_cc_set == 0)
737 		return;
738 	/*
739 	 * Use the swap utility placing in 4 for flex8 to id a
740 	 * restoral of the old values.
741 	 */
742 	rack->rc_pacing_cc_set = 0;
743 	rack_swap_beta_values(rack, 4);
744 }
745 
746 static void
747 rack_remove_pacing(struct tcp_rack *rack)
748 {
749 	if (rack->rc_pacing_cc_set)
750 		rack_undo_cc_pacing(rack);
751 	if (rack->r_ctl.pacing_method & RACK_REG_PACING)
752 		tcp_decrement_paced_conn();
753 	if (rack->r_ctl.pacing_method & RACK_DGP_PACING)
754 		tcp_dec_dgp_pacing_cnt();
755 	rack->rc_always_pace = 0;
756 	rack->r_ctl.pacing_method = RACK_PACING_NONE;
757 	rack->dgp_on = 0;
758 	rack->rc_hybrid_mode = 0;
759 	rack->use_fixed_rate = 0;
760 }
761 
762 static void
763 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
764 	       uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
765 {
766 	if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
767 		union tcp_log_stackspecific log;
768 		struct timeval tv;
769 
770 		memset(&log, 0, sizeof(log));
771 		log.u_bbr.flex1 = seq_end;
772 		log.u_bbr.flex2 = rack->rc_tp->gput_seq;
773 		log.u_bbr.flex3 = ack_end_t;
774 		log.u_bbr.flex4 = rack->rc_tp->gput_ts;
775 		log.u_bbr.flex5 = send_end_t;
776 		log.u_bbr.flex6 = rack->rc_tp->gput_ack;
777 		log.u_bbr.flex7 = mode;
778 		log.u_bbr.flex8 = 69;
779 		log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
780 		log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
781 		log.u_bbr.pkts_out = line;
782 		log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
783 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
784 		log.u_bbr.epoch = rack->r_ctl.current_round;
785 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
786 		if (rsm != NULL) {
787 			log.u_bbr.applimited = rsm->r_start;
788 			log.u_bbr.delivered = rsm->r_end;
789 			log.u_bbr.epoch = rsm->r_flags;
790 		}
791 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
792 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
793 		    &rack->rc_inp->inp_socket->so_rcv,
794 		    &rack->rc_inp->inp_socket->so_snd,
795 		    BBR_LOG_HPTSI_CALC, 0,
796 		    0, &log, false, &tv);
797 	}
798 }
799 
800 static int
801 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
802 {
803 	uint32_t stat;
804 	int32_t error;
805 
806 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
807 	if (error || req->newptr == NULL)
808 		return error;
809 
810 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
811 	if (error)
812 		return (error);
813 	if (stat == 1) {
814 #ifdef INVARIANTS
815 		printf("Clearing RACK counters\n");
816 #endif
817 		counter_u64_zero(rack_tlp_tot);
818 		counter_u64_zero(rack_tlp_newdata);
819 		counter_u64_zero(rack_tlp_retran);
820 		counter_u64_zero(rack_tlp_retran_bytes);
821 		counter_u64_zero(rack_to_tot);
822 		counter_u64_zero(rack_saw_enobuf);
823 		counter_u64_zero(rack_saw_enobuf_hw);
824 		counter_u64_zero(rack_saw_enetunreach);
825 		counter_u64_zero(rack_persists_sends);
826 		counter_u64_zero(rack_total_bytes);
827 		counter_u64_zero(rack_persists_acks);
828 		counter_u64_zero(rack_persists_loss);
829 		counter_u64_zero(rack_persists_lost_ends);
830 #ifdef INVARIANTS
831 		counter_u64_zero(rack_adjust_map_bw);
832 #endif
833 		counter_u64_zero(rack_to_alloc_hard);
834 		counter_u64_zero(rack_to_alloc_emerg);
835 		counter_u64_zero(rack_sack_proc_all);
836 		counter_u64_zero(rack_fto_send);
837 		counter_u64_zero(rack_fto_rsm_send);
838 		counter_u64_zero(rack_extended_rfo);
839 		counter_u64_zero(rack_hw_pace_init_fail);
840 		counter_u64_zero(rack_hw_pace_lost);
841 		counter_u64_zero(rack_non_fto_send);
842 		counter_u64_zero(rack_nfto_resend);
843 		counter_u64_zero(rack_sack_proc_short);
844 		counter_u64_zero(rack_sack_proc_restart);
845 		counter_u64_zero(rack_to_alloc);
846 		counter_u64_zero(rack_to_alloc_limited);
847 		counter_u64_zero(rack_alloc_limited_conns);
848 		counter_u64_zero(rack_split_limited);
849 		counter_u64_zero(rack_rxt_clamps_cwnd);
850 		counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
851 		counter_u64_zero(rack_multi_single_eq);
852 		counter_u64_zero(rack_proc_non_comp_ack);
853 		counter_u64_zero(rack_sack_attacks_detected);
854 		counter_u64_zero(rack_sack_attacks_reversed);
855 		counter_u64_zero(rack_sack_attacks_suspect);
856 		counter_u64_zero(rack_sack_used_next_merge);
857 		counter_u64_zero(rack_sack_used_prev_merge);
858 		counter_u64_zero(rack_sack_splits);
859 		counter_u64_zero(rack_sack_skipped_acked);
860 		counter_u64_zero(rack_ack_total);
861 		counter_u64_zero(rack_express_sack);
862 		counter_u64_zero(rack_sack_total);
863 		counter_u64_zero(rack_move_none);
864 		counter_u64_zero(rack_move_some);
865 		counter_u64_zero(rack_try_scwnd);
866 		counter_u64_zero(rack_collapsed_win);
867 		counter_u64_zero(rack_collapsed_win_rxt);
868 		counter_u64_zero(rack_collapsed_win_seen);
869 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
870 	} else if (stat == 2) {
871 #ifdef INVARIANTS
872 		printf("Clearing RACK option array\n");
873 #endif
874 		COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
875 	} else if (stat == 3) {
876 		printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
877 	} else if (stat == 4) {
878 #ifdef INVARIANTS
879 		printf("Clearing RACK out size array\n");
880 #endif
881 		COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
882 	}
883 	rack_clear_counter = 0;
884 	return (0);
885 }
886 
887 static void
888 rack_init_sysctls(void)
889 {
890 	struct sysctl_oid *rack_counters;
891 	struct sysctl_oid *rack_attack;
892 	struct sysctl_oid *rack_pacing;
893 	struct sysctl_oid *rack_timely;
894 	struct sysctl_oid *rack_timers;
895 	struct sysctl_oid *rack_tlp;
896 	struct sysctl_oid *rack_misc;
897 	struct sysctl_oid *rack_features;
898 	struct sysctl_oid *rack_measure;
899 	struct sysctl_oid *rack_probertt;
900 	struct sysctl_oid *rack_hw_pacing;
901 	struct sysctl_oid *rack_policing;
902 
903 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
904 	    SYSCTL_CHILDREN(rack_sysctl_root),
905 	    OID_AUTO,
906 	    "sack_attack",
907 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
908 	    "Rack Sack Attack Counters and Controls");
909 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
910 	    SYSCTL_CHILDREN(rack_sysctl_root),
911 	    OID_AUTO,
912 	    "stats",
913 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
914 	    "Rack Counters");
915 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
916 	    SYSCTL_CHILDREN(rack_sysctl_root),
917 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
918 	    &rack_rate_sample_method , USE_RTT_LOW,
919 	    "What method should we use for rate sampling 0=high, 1=low ");
920 	/* Probe rtt related controls */
921 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
922 	    SYSCTL_CHILDREN(rack_sysctl_root),
923 	    OID_AUTO,
924 	    "probertt",
925 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
926 	    "ProbeRTT related Controls");
927 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
928 	    SYSCTL_CHILDREN(rack_probertt),
929 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
930 	    &rack_atexit_prtt_hbp, 130,
931 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
932 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
933 	    SYSCTL_CHILDREN(rack_probertt),
934 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
935 	    &rack_atexit_prtt, 130,
936 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
937 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
938 	    SYSCTL_CHILDREN(rack_probertt),
939 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
940 	    &rack_per_of_gp_probertt, 60,
941 	    "What percentage of goodput do we pace at in probertt");
942 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
943 	    SYSCTL_CHILDREN(rack_probertt),
944 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
945 	    &rack_per_of_gp_probertt_reduce, 10,
946 	    "What percentage of goodput do we reduce every gp_srtt");
947 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
948 	    SYSCTL_CHILDREN(rack_probertt),
949 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
950 	    &rack_per_of_gp_lowthresh, 40,
951 	    "What percentage of goodput do we allow the multiplier to fall to");
952 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
953 	    SYSCTL_CHILDREN(rack_probertt),
954 	    OID_AUTO, "time_between", CTLFLAG_RW,
955 	    & rack_time_between_probertt, 96000000,
956 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
957 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
958 	    SYSCTL_CHILDREN(rack_probertt),
959 	    OID_AUTO, "safety", CTLFLAG_RW,
960 	    &rack_probe_rtt_safety_val, 2000000,
961 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
962 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
963 	    SYSCTL_CHILDREN(rack_probertt),
964 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
965 	    &rack_probe_rtt_sets_cwnd, 0,
966 	    "Do we set the cwnd too (if always_lower is on)");
967 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
968 	    SYSCTL_CHILDREN(rack_probertt),
969 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
970 	    &rack_max_drain_wait, 2,
971 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
972 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
973 	    SYSCTL_CHILDREN(rack_probertt),
974 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
975 	    &rack_must_drain, 1,
976 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
977 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
978 	    SYSCTL_CHILDREN(rack_probertt),
979 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
980 	    &rack_probertt_use_min_rtt_entry, 1,
981 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
982 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
983 	    SYSCTL_CHILDREN(rack_probertt),
984 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
985 	    &rack_probertt_use_min_rtt_exit, 0,
986 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
987 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
988 	    SYSCTL_CHILDREN(rack_probertt),
989 	    OID_AUTO, "length_div", CTLFLAG_RW,
990 	    &rack_probertt_gpsrtt_cnt_div, 0,
991 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
992 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
993 	    SYSCTL_CHILDREN(rack_probertt),
994 	    OID_AUTO, "length_mul", CTLFLAG_RW,
995 	    &rack_probertt_gpsrtt_cnt_mul, 0,
996 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
997 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
998 	    SYSCTL_CHILDREN(rack_probertt),
999 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
1000 	    &rack_min_probertt_hold, 200000,
1001 	    "What is the minimum time we hold probertt at target");
1002 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1003 	    SYSCTL_CHILDREN(rack_probertt),
1004 	    OID_AUTO, "filter_life", CTLFLAG_RW,
1005 	    &rack_probertt_filter_life, 10000000,
1006 	    "What is the time for the filters life in useconds");
1007 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1008 	    SYSCTL_CHILDREN(rack_probertt),
1009 	    OID_AUTO, "lower_within", CTLFLAG_RW,
1010 	    &rack_probertt_lower_within, 10,
1011 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
1012 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1013 	    SYSCTL_CHILDREN(rack_probertt),
1014 	    OID_AUTO, "must_move", CTLFLAG_RW,
1015 	    &rack_min_rtt_movement, 250,
1016 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
1017 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1018 	    SYSCTL_CHILDREN(rack_probertt),
1019 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
1020 	    &rack_probertt_clear_is, 1,
1021 	    "Do we clear I/S counts on exiting probe-rtt");
1022 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1023 	    SYSCTL_CHILDREN(rack_probertt),
1024 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
1025 	    &rack_max_drain_hbp, 1,
1026 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
1027 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1028 	    SYSCTL_CHILDREN(rack_probertt),
1029 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
1030 	    &rack_hbp_thresh, 3,
1031 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
1032 	/* Pacing related sysctls */
1033 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1034 	    SYSCTL_CHILDREN(rack_sysctl_root),
1035 	    OID_AUTO,
1036 	    "pacing",
1037 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1038 	    "Pacing related Controls");
1039 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1040 	    SYSCTL_CHILDREN(rack_pacing),
1041 	    OID_AUTO, "pcm_enabled", CTLFLAG_RW,
1042 	    &rack_pcm_is_enabled, 1,
1043 	    "Do we by default do PCM measurements?");
1044 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1045 	    SYSCTL_CHILDREN(rack_pacing),
1046 	    OID_AUTO, "pcm_rnds", CTLFLAG_RW,
1047 	    &rack_pcm_every_n_rounds, 100,
1048 	    "How many rounds before we need to do a PCM measurement");
1049 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1050 	    SYSCTL_CHILDREN(rack_pacing),
1051 	    OID_AUTO, "pcm_blast", CTLFLAG_RW,
1052 	    &rack_pcm_blast, 0,
1053 	    "Blast out the full cwnd/rwnd when doing a PCM measurement");
1054 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1055 	    SYSCTL_CHILDREN(rack_pacing),
1056 	    OID_AUTO, "rnd_gp_gain", CTLFLAG_RW,
1057 	    &rack_gp_gain_req, 1200,
1058 	    "How much do we have to increase the GP to record the round 1200 = 120.0");
1059 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1060 	    SYSCTL_CHILDREN(rack_pacing),
1061 	    OID_AUTO, "dgp_out_of_ss_at", CTLFLAG_RW,
1062 	    &rack_rnd_cnt_req, 0x10005,
1063 	    "How many rounds less than rnd_gp_gain will drop us out of SS");
1064 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1065 	    SYSCTL_CHILDREN(rack_pacing),
1066 	    OID_AUTO, "no_timely", CTLFLAG_RW,
1067 	    &rack_timely_off, 0,
1068 	    "Do we not use timely in DGP?");
1069 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1070 	    SYSCTL_CHILDREN(rack_pacing),
1071 	    OID_AUTO, "fullbufdisc", CTLFLAG_RW,
1072 	    &rack_full_buffer_discount, 10,
1073 	    "What percentage b/w reduction over the GP estimate for a full buffer (default=0 off)?");
1074 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1075 	    SYSCTL_CHILDREN(rack_pacing),
1076 	    OID_AUTO, "fillcw", CTLFLAG_RW,
1077 	    &rack_fill_cw_state, 0,
1078 	    "Enable fillcw on new connections (default=0 off)?");
1079 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1080 	    SYSCTL_CHILDREN(rack_pacing),
1081 	    OID_AUTO, "min_burst", CTLFLAG_RW,
1082 	    &rack_pacing_min_seg, 0,
1083 	    "What is the min burst size for pacing (0 disables)?");
1084 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1085 	    SYSCTL_CHILDREN(rack_pacing),
1086 	    OID_AUTO, "divisor", CTLFLAG_RW,
1087 	    &rack_default_pacing_divisor, 250,
1088 	    "What is the default divisor given to the rl code?");
1089 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1090 	    SYSCTL_CHILDREN(rack_pacing),
1091 	    OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1092 	    &rack_bw_multipler, 0,
1093 	    "What is the limit multiplier of the current gp_est that fillcw can increase the b/w too, 200 == 200% (0 = off)?");
1094 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1095 	    SYSCTL_CHILDREN(rack_pacing),
1096 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
1097 	    &rack_max_per_above, 30,
1098 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1099 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1100 	    SYSCTL_CHILDREN(rack_pacing),
1101 	    OID_AUTO, "allow1mss", CTLFLAG_RW,
1102 	    &rack_pace_one_seg, 0,
1103 	    "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1104 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1105 	    SYSCTL_CHILDREN(rack_pacing),
1106 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1107 	    &rack_limit_time_with_srtt, 0,
1108 	    "Do we limit pacing time based on srtt");
1109 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1110 	    SYSCTL_CHILDREN(rack_pacing),
1111 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1112 	    &rack_per_of_gp_ss, 250,
1113 	    "If non zero, what percentage of goodput to pace at in slow start");
1114 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1115 	    SYSCTL_CHILDREN(rack_pacing),
1116 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1117 	    &rack_per_of_gp_ca, 150,
1118 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1119 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1120 	    SYSCTL_CHILDREN(rack_pacing),
1121 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1122 	    &rack_per_of_gp_rec, 200,
1123 	    "If non zero, what percentage of goodput to pace at in recovery");
1124 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1125 	    SYSCTL_CHILDREN(rack_pacing),
1126 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1127 	    &rack_hptsi_segments, 40,
1128 	    "What size is the max for TSO segments in pacing and burst mitigation");
1129 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1130 	    SYSCTL_CHILDREN(rack_pacing),
1131 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1132 	    &rack_slot_reduction, 4,
1133 	    "When doing only burst mitigation what is the reduce divisor");
1134 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1135 	    SYSCTL_CHILDREN(rack_sysctl_root),
1136 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1137 	    &rack_pace_every_seg, 0,
1138 	    "If set we use pacing, if clear we use only the original burst mitigation");
1139 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1140 	    SYSCTL_CHILDREN(rack_pacing),
1141 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1142 	    &rack_bw_rate_cap, 0,
1143 	    "If set we apply this value to the absolute rate cap used by pacing");
1144 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1145 	    SYSCTL_CHILDREN(rack_pacing),
1146 	    OID_AUTO, "fillcw_cap", CTLFLAG_RW,
1147 	    &rack_fillcw_bw_cap, 3750000,
1148 	    "Do we have an absolute cap on the amount of b/w fillcw can specify (0 = no)?");
1149 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1150 	    SYSCTL_CHILDREN(rack_sysctl_root),
1151 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1152 	    &rack_req_measurements, 1,
1153 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1154 	/* Hardware pacing */
1155 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1156 	    SYSCTL_CHILDREN(rack_sysctl_root),
1157 	    OID_AUTO,
1158 	    "hdwr_pacing",
1159 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1160 	    "Pacing related Controls");
1161 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1162 	    SYSCTL_CHILDREN(rack_hw_pacing),
1163 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1164 	    &rack_hw_rwnd_factor, 2,
1165 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1166 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1167 	    SYSCTL_CHILDREN(rack_hw_pacing),
1168 	    OID_AUTO, "precheck", CTLFLAG_RW,
1169 	    &rack_hw_check_queue, 0,
1170 	    "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1171 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1172 	    SYSCTL_CHILDREN(rack_hw_pacing),
1173 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1174 	    &rack_enobuf_hw_boost_mult, 0,
1175 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1176 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1177 	    SYSCTL_CHILDREN(rack_hw_pacing),
1178 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1179 	    &rack_enobuf_hw_max, 2,
1180 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1181 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1182 	    SYSCTL_CHILDREN(rack_hw_pacing),
1183 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1184 	    &rack_enobuf_hw_min, 2,
1185 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1186 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1187 	    SYSCTL_CHILDREN(rack_hw_pacing),
1188 	    OID_AUTO, "enable", CTLFLAG_RW,
1189 	    &rack_enable_hw_pacing, 0,
1190 	    "Should RACK attempt to use hw pacing?");
1191 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1192 	    SYSCTL_CHILDREN(rack_hw_pacing),
1193 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1194 	    &rack_hw_rate_caps, 0,
1195 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1196 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1197 	    SYSCTL_CHILDREN(rack_hw_pacing),
1198 	    OID_AUTO, "uncap_per", CTLFLAG_RW,
1199 	    &rack_hw_rate_cap_per, 0,
1200 	    "If you go over b/w by this amount you will be uncapped (0 = never)");
1201 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1202 	    SYSCTL_CHILDREN(rack_hw_pacing),
1203 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1204 	    &rack_hw_rate_min, 0,
1205 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1206 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1207 	    SYSCTL_CHILDREN(rack_hw_pacing),
1208 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1209 	    &rack_hw_rate_to_low, 0,
1210 	    "If we fall below this rate, dis-engage hw pacing?");
1211 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1212 	    SYSCTL_CHILDREN(rack_hw_pacing),
1213 	    OID_AUTO, "up_only", CTLFLAG_RW,
1214 	    &rack_hw_up_only, 0,
1215 	    "Do we allow hw pacing to lower the rate selected?");
1216 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1217 	    SYSCTL_CHILDREN(rack_hw_pacing),
1218 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1219 	    &rack_hw_pace_extra_slots, 0,
1220 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1221 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1222 	    SYSCTL_CHILDREN(rack_sysctl_root),
1223 	    OID_AUTO,
1224 	    "timely",
1225 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1226 	    "Rack Timely RTT Controls");
1227 	/* Timely based GP dynmics */
1228 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229 	    SYSCTL_CHILDREN(rack_timely),
1230 	    OID_AUTO, "upper", CTLFLAG_RW,
1231 	    &rack_gp_per_bw_mul_up, 2,
1232 	    "Rack timely upper range for equal b/w (in percentage)");
1233 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1234 	    SYSCTL_CHILDREN(rack_timely),
1235 	    OID_AUTO, "lower", CTLFLAG_RW,
1236 	    &rack_gp_per_bw_mul_down, 4,
1237 	    "Rack timely lower range for equal b/w (in percentage)");
1238 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1239 	    SYSCTL_CHILDREN(rack_timely),
1240 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1241 	    &rack_gp_rtt_maxmul, 3,
1242 	    "Rack timely multiplier of lowest rtt for rtt_max");
1243 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1244 	    SYSCTL_CHILDREN(rack_timely),
1245 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1246 	    &rack_gp_rtt_mindiv, 4,
1247 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1248 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1249 	    SYSCTL_CHILDREN(rack_timely),
1250 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1251 	    &rack_gp_rtt_minmul, 1,
1252 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1253 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1254 	    SYSCTL_CHILDREN(rack_timely),
1255 	    OID_AUTO, "decrease", CTLFLAG_RW,
1256 	    &rack_gp_decrease_per, 80,
1257 	    "Rack timely Beta value 80 = .8 (scaled by 100)");
1258 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1259 	    SYSCTL_CHILDREN(rack_timely),
1260 	    OID_AUTO, "increase", CTLFLAG_RW,
1261 	    &rack_gp_increase_per, 2,
1262 	    "Rack timely increase perentage of our GP multiplication factor");
1263 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1264 	    SYSCTL_CHILDREN(rack_timely),
1265 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1266 	    &rack_per_lower_bound, 50,
1267 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1268 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1269 	    SYSCTL_CHILDREN(rack_timely),
1270 	    OID_AUTO, "p5_upper", CTLFLAG_RW,
1271 	    &rack_gain_p5_ub, 250,
1272 	    "Profile 5 upper bound to timely gain");
1273 
1274 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1275 	    SYSCTL_CHILDREN(rack_timely),
1276 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1277 	    &rack_per_upper_bound_ss, 0,
1278 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1279 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1280 	    SYSCTL_CHILDREN(rack_timely),
1281 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1282 	    &rack_per_upper_bound_ca, 0,
1283 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1284 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1285 	    SYSCTL_CHILDREN(rack_timely),
1286 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1287 	    &rack_do_dyn_mul, 0,
1288 	    "Rack timely do we enable dynmaic timely goodput by default");
1289 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1290 	    SYSCTL_CHILDREN(rack_timely),
1291 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1292 	    &rack_gp_no_rec_chg, 1,
1293 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1294 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1295 	    SYSCTL_CHILDREN(rack_timely),
1296 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1297 	    &rack_timely_dec_clear, 6,
1298 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1299 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1300 	    SYSCTL_CHILDREN(rack_timely),
1301 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1302 	    &rack_timely_max_push_rise, 3,
1303 	    "Rack timely how many times do we push up with b/w increase");
1304 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1305 	    SYSCTL_CHILDREN(rack_timely),
1306 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1307 	    &rack_timely_max_push_drop, 3,
1308 	    "Rack timely how many times do we push back on b/w decent");
1309 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1310 	    SYSCTL_CHILDREN(rack_timely),
1311 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1312 	    &rack_timely_min_segs, 4,
1313 	    "Rack timely when setting the cwnd what is the min num segments");
1314 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1315 	    SYSCTL_CHILDREN(rack_timely),
1316 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1317 	    &rack_use_max_for_nobackoff, 0,
1318 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1319 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1320 	    SYSCTL_CHILDREN(rack_timely),
1321 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1322 	    &rack_timely_int_timely_only, 0,
1323 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1324 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1325 	    SYSCTL_CHILDREN(rack_timely),
1326 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1327 	    &rack_timely_no_stopping, 0,
1328 	    "Rack timely don't stop increase");
1329 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1330 	    SYSCTL_CHILDREN(rack_timely),
1331 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1332 	    &rack_down_raise_thresh, 100,
1333 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1334 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1335 	    SYSCTL_CHILDREN(rack_timely),
1336 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1337 	    &rack_req_segs, 1,
1338 	    "Bottom dragging if not these many segments outstanding and room");
1339 
1340 	/* TLP and Rack related parameters */
1341 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1342 	    SYSCTL_CHILDREN(rack_sysctl_root),
1343 	    OID_AUTO,
1344 	    "tlp",
1345 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1346 	    "TLP and Rack related Controls");
1347 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1348 	    SYSCTL_CHILDREN(rack_tlp),
1349 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1350 	    &use_rack_rr, 1,
1351 	    "Do we use Rack Rapid Recovery");
1352 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1353 	    SYSCTL_CHILDREN(rack_tlp),
1354 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1355 	    &rack_max_abc_post_recovery, 2,
1356 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1357 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1358 	    SYSCTL_CHILDREN(rack_tlp),
1359 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1360 	    &rack_non_rxt_use_cr, 0,
1361 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1362 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1363 	    SYSCTL_CHILDREN(rack_tlp),
1364 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1365 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1366 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1367 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1368 	    SYSCTL_CHILDREN(rack_tlp),
1369 	    OID_AUTO, "limit", CTLFLAG_RW,
1370 	    &rack_tlp_limit, 2,
1371 	    "How many TLP's can be sent without sending new data");
1372 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1373 	    SYSCTL_CHILDREN(rack_tlp),
1374 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1375 	    &rack_tlp_use_greater, 1,
1376 	    "Should we use the rack_rtt time if its greater than srtt");
1377 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1378 	    SYSCTL_CHILDREN(rack_tlp),
1379 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1380 	    &rack_tlp_min, 10000,
1381 	    "TLP minimum timeout per the specification (in microseconds)");
1382 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1383 	    SYSCTL_CHILDREN(rack_tlp),
1384 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1385 	    &rack_always_send_oldest, 0,
1386 	    "Should we always send the oldest TLP and RACK-TLP");
1387 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1388 	    SYSCTL_CHILDREN(rack_tlp),
1389 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1390 	    &rack_lower_cwnd_at_tlp, 0,
1391 	    "When a TLP completes a retran should we enter recovery");
1392 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1393 	    SYSCTL_CHILDREN(rack_tlp),
1394 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1395 	    &rack_reorder_thresh, 2,
1396 	    "What factor for rack will be added when seeing reordering (shift right)");
1397 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1398 	    SYSCTL_CHILDREN(rack_tlp),
1399 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1400 	    &rack_tlp_thresh, 1,
1401 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1402 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1403 	    SYSCTL_CHILDREN(rack_tlp),
1404 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1405 	    &rack_reorder_fade, 60000000,
1406 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1407 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1408 	    SYSCTL_CHILDREN(rack_tlp),
1409 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1410 	    &rack_pkt_delay, 1000,
1411 	    "Extra RACK time (in microseconds) besides reordering thresh");
1412 
1413 	/* Timer related controls */
1414 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1415 	    SYSCTL_CHILDREN(rack_sysctl_root),
1416 	    OID_AUTO,
1417 	    "timers",
1418 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1419 	    "Timer related controls");
1420 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1421 	    SYSCTL_CHILDREN(rack_timers),
1422 	    OID_AUTO, "reset_ssth_rec_rto", CTLFLAG_RW,
1423 	    &rack_ssthresh_rest_rto_rec, 0,
1424 	    "When doing recovery -> rto -> recovery do we reset SSthresh?");
1425 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1426 	    SYSCTL_CHILDREN(rack_timers),
1427 	    OID_AUTO, "scoreboard_thresh", CTLFLAG_RW,
1428 	    &rack_rxt_scoreboard_clear_thresh, 2,
1429 	    "How many RTO's are allowed before we clear the scoreboard");
1430 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1431 	    SYSCTL_CHILDREN(rack_timers),
1432 	    OID_AUTO, "honor_hpts_min", CTLFLAG_RW,
1433 	    &rack_honors_hpts_min_to, 1,
1434 	    "Do rack pacing timers honor hpts min timeout");
1435 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1436 	    SYSCTL_CHILDREN(rack_timers),
1437 	    OID_AUTO, "hpts_max_reduce", CTLFLAG_RW,
1438 	    &rack_max_reduce, 10,
1439 	    "Max percentage we will reduce slot by for pacing when we are behind");
1440 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1441 	    SYSCTL_CHILDREN(rack_timers),
1442 	    OID_AUTO, "persmin", CTLFLAG_RW,
1443 	    &rack_persist_min, 250000,
1444 	    "What is the minimum time in microseconds between persists");
1445 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1446 	    SYSCTL_CHILDREN(rack_timers),
1447 	    OID_AUTO, "persmax", CTLFLAG_RW,
1448 	    &rack_persist_max, 2000000,
1449 	    "What is the largest delay in microseconds between persists");
1450 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1451 	    SYSCTL_CHILDREN(rack_timers),
1452 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1453 	    &rack_delayed_ack_time, 40000,
1454 	    "Delayed ack time (40ms in microseconds)");
1455 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1456 	    SYSCTL_CHILDREN(rack_timers),
1457 	    OID_AUTO, "minrto", CTLFLAG_RW,
1458 	    &rack_rto_min, 30000,
1459 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1460 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1461 	    SYSCTL_CHILDREN(rack_timers),
1462 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1463 	    &rack_rto_max, 4000000,
1464 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1465 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1466 	    SYSCTL_CHILDREN(rack_timers),
1467 	    OID_AUTO, "minto", CTLFLAG_RW,
1468 	    &rack_min_to, 1000,
1469 	    "Minimum rack timeout in microseconds");
1470 	/* Measure controls */
1471 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1472 	    SYSCTL_CHILDREN(rack_sysctl_root),
1473 	    OID_AUTO,
1474 	    "measure",
1475 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1476 	    "Measure related controls");
1477 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1478 	    SYSCTL_CHILDREN(rack_measure),
1479 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1480 	    &rack_wma_divisor, 8,
1481 	    "When doing b/w calculation what is the  divisor for the WMA");
1482 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1483 	    SYSCTL_CHILDREN(rack_measure),
1484 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1485 	    &rack_cwnd_block_ends_measure, 0,
1486 	    "Does a cwnd just-return end the measurement window (app limited)");
1487 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1488 	    SYSCTL_CHILDREN(rack_measure),
1489 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1490 	    &rack_rwnd_block_ends_measure, 0,
1491 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1492 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1493 	    SYSCTL_CHILDREN(rack_measure),
1494 	    OID_AUTO, "min_target", CTLFLAG_RW,
1495 	    &rack_def_data_window, 20,
1496 	    "What is the minimum target window (in mss) for a GP measurements");
1497 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1498 	    SYSCTL_CHILDREN(rack_measure),
1499 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1500 	    &rack_goal_bdp, 2,
1501 	    "What is the goal BDP to measure");
1502 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1503 	    SYSCTL_CHILDREN(rack_measure),
1504 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1505 	    &rack_min_srtts, 1,
1506 	    "What is the goal BDP to measure");
1507 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1508 	    SYSCTL_CHILDREN(rack_measure),
1509 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1510 	    &rack_min_measure_usec, 0,
1511 	    "What is the Minimum time time for a measurement if 0, this is off");
1512 	/* Features */
1513 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1514 	    SYSCTL_CHILDREN(rack_sysctl_root),
1515 	    OID_AUTO,
1516 	    "features",
1517 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1518 	    "Feature controls");
1519 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1520 	    SYSCTL_CHILDREN(rack_features),
1521 	    OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1522 	    &rack_hybrid_allow_set_maxseg, 0,
1523 	    "Should hybrid pacing allow the setmss command");
1524 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1525 	    SYSCTL_CHILDREN(rack_features),
1526 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1527 	    &rack_use_cmp_acks, 1,
1528 	    "Should RACK have LRO send compressed acks");
1529 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1530 	    SYSCTL_CHILDREN(rack_features),
1531 	    OID_AUTO, "fsb", CTLFLAG_RW,
1532 	    &rack_use_fsb, 1,
1533 	    "Should RACK use the fast send block?");
1534 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1535 	    SYSCTL_CHILDREN(rack_features),
1536 	    OID_AUTO, "rfo", CTLFLAG_RW,
1537 	    &rack_use_rfo, 1,
1538 	    "Should RACK use rack_fast_output()?");
1539 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1540 	    SYSCTL_CHILDREN(rack_features),
1541 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1542 	    &rack_use_rsm_rfo, 1,
1543 	    "Should RACK use rack_fast_rsm_output()?");
1544 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1545 	    SYSCTL_CHILDREN(rack_features),
1546 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1547 	    &rack_enable_mqueue_for_nonpaced, 0,
1548 	    "Should RACK use mbuf queuing for non-paced connections");
1549 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1550 	    SYSCTL_CHILDREN(rack_features),
1551 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1552 	    &rack_do_hystart, 0,
1553 	    "Should RACK enable HyStart++ on connections?");
1554 	/* Policer detection */
1555 	rack_policing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1556 	    SYSCTL_CHILDREN(rack_sysctl_root),
1557 	    OID_AUTO,
1558 	    "policing",
1559 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1560 	    "policer detection");
1561 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1562 	    SYSCTL_CHILDREN(rack_policing),
1563 	    OID_AUTO, "rxt_thresh", CTLFLAG_RW,
1564 	    &rack_policer_rxt_thresh, 0,
1565 	   "Percentage of retransmits we need to be a possible policer (499 = 49.9 percent)");
1566 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1567 	    SYSCTL_CHILDREN(rack_policing),
1568 	    OID_AUTO, "avg_thresh", CTLFLAG_RW,
1569 	    &rack_policer_avg_thresh, 0,
1570 	    "What threshold of average retransmits needed to recover a lost packet (1 - 169 aka 21 = 2.1)?");
1571 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1572 	    SYSCTL_CHILDREN(rack_policing),
1573 	    OID_AUTO, "med_thresh", CTLFLAG_RW,
1574 	    &rack_policer_med_thresh, 0,
1575 	    "What threshold of Median retransmits needed to recover a lost packet (1 - 16)?");
1576 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1577 	    SYSCTL_CHILDREN(rack_policing),
1578 	    OID_AUTO, "data_thresh", CTLFLAG_RW,
1579 	    &rack_policer_data_thresh, 64000,
1580 	    "How many bytes must have gotten through before we can start doing policer detection?");
1581 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1582 	    SYSCTL_CHILDREN(rack_policing),
1583 	    OID_AUTO, "bwcomp", CTLFLAG_RW,
1584 	    &rack_policing_do_bw_comp, 1,
1585 	    "Do we raise up low b/w so that at least pace_max_seg can be sent in the srtt?");
1586 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1587 	    SYSCTL_CHILDREN(rack_policing),
1588 	    OID_AUTO, "recmss", CTLFLAG_RW,
1589 	    &rack_req_del_mss, 18,
1590 	    "How many MSS must be delivered during recovery to engage policer detection?");
1591 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1592 	    SYSCTL_CHILDREN(rack_policing),
1593 	    OID_AUTO, "res_div", CTLFLAG_RW,
1594 	    &rack_policer_bucket_reserve, 20,
1595 	    "What percentage is reserved in the policer bucket?");
1596 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1597 	    SYSCTL_CHILDREN(rack_policing),
1598 	    OID_AUTO, "min_comp_bw", CTLFLAG_RW,
1599 	    &rack_pol_min_bw, 125000,
1600 	    "Do we have a min b/w for b/w compensation (0 = no)?");
1601 	/* Misc rack controls */
1602 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1603 	    SYSCTL_CHILDREN(rack_sysctl_root),
1604 	    OID_AUTO,
1605 	    "misc",
1606 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1607 	    "Misc related controls");
1608 #ifdef TCP_ACCOUNTING
1609 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1610 	    SYSCTL_CHILDREN(rack_misc),
1611 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1612 	    &rack_tcp_accounting, 0,
1613 	    "Should we turn on TCP accounting for all rack sessions?");
1614 #endif
1615 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1616 	    SYSCTL_CHILDREN(rack_misc),
1617 	    OID_AUTO, "dnd", CTLFLAG_RW,
1618 	    &rack_dnd_default, 0,
1619 	    "Do not disturb default for rack_rrr = 3");
1620 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1621 	    SYSCTL_CHILDREN(rack_misc),
1622 	    OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1623 	    &sad_seg_size_per, 800,
1624 	    "Percentage of segment size needed in a sack 800 = 80.0?");
1625 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1626 	    SYSCTL_CHILDREN(rack_misc),
1627 	    OID_AUTO, "rxt_controls", CTLFLAG_RW,
1628 	    &rack_rxt_controls, 0,
1629 	    "Retransmit sending size controls (valid  values 0, 1, 2 default=1)?");
1630 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1631 	    SYSCTL_CHILDREN(rack_misc),
1632 	    OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1633 	    &rack_hibeta_setting, 0,
1634 	    "Do we ue a high beta (80 instead of 50)?");
1635 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1636 	    SYSCTL_CHILDREN(rack_misc),
1637 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1638 	    &rack_apply_rtt_with_reduced_conf, 0,
1639 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1640 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1641 	    SYSCTL_CHILDREN(rack_misc),
1642 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1643 	    &rack_dsack_std_based, 3,
1644 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1645 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1646 	    SYSCTL_CHILDREN(rack_misc),
1647 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1648 	    &rack_prr_addbackmax, 2,
1649 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1650 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1651 	    SYSCTL_CHILDREN(rack_misc),
1652 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1653 	    &rack_stats_gets_ms_rtt, 1,
1654 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1655 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1656 	    SYSCTL_CHILDREN(rack_misc),
1657 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1658 	    &rack_client_low_buf, 0,
1659 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1660 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1661 	    SYSCTL_CHILDREN(rack_misc),
1662 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1663 	    &rack_def_profile, 0,
1664 	    "Should RACK use a default profile (0=no, num == profile num)?");
1665 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1666 	    SYSCTL_CHILDREN(rack_misc),
1667 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1668 	    &rack_enable_shared_cwnd, 1,
1669 	    "Should RACK try to use the shared cwnd on connections where allowed");
1670 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1671 	    SYSCTL_CHILDREN(rack_misc),
1672 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1673 	    &rack_limits_scwnd, 1,
1674 	    "Should RACK place low end time limits on the shared cwnd feature");
1675 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1676 	    SYSCTL_CHILDREN(rack_misc),
1677 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1678 	    &rack_disable_prr, 0,
1679 	    "Should RACK not use prr and only pace (must have pacing on)");
1680 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1681 	    SYSCTL_CHILDREN(rack_misc),
1682 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1683 	    &rack_verbose_logging, 0,
1684 	    "Should RACK black box logging be verbose");
1685 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1686 	    SYSCTL_CHILDREN(rack_misc),
1687 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1688 	    &rack_ignore_data_after_close, 1,
1689 	    "Do we hold off sending a RST until all pending data is ack'd");
1690 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1691 	    SYSCTL_CHILDREN(rack_misc),
1692 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1693 	    &rack_sack_not_required, 1,
1694 	    "Do we allow rack to run on connections not supporting SACK");
1695 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1696 	    SYSCTL_CHILDREN(rack_misc),
1697 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1698 	    &rack_send_a_lot_in_prr, 1,
1699 	    "Send a lot in prr");
1700 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1701 	    SYSCTL_CHILDREN(rack_misc),
1702 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1703 	    &rack_autosndbuf_inc, 20,
1704 	    "What percentage should rack scale up its snd buffer by?");
1705 
1706 
1707 	/* Sack Attacker detection stuff */
1708 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1709 	    SYSCTL_CHILDREN(rack_attack),
1710 	    OID_AUTO, "merge_out", CTLFLAG_RW,
1711 	    &rack_merge_out_sacks_on_attack, 0,
1712 	    "Do we merge the sendmap when we decide we are being attacked?");
1713 
1714 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1715 	    SYSCTL_CHILDREN(rack_attack),
1716 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1717 	    &rack_highest_sack_thresh_seen, 0,
1718 	    "Highest sack to ack ratio seen");
1719 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1720 	    SYSCTL_CHILDREN(rack_attack),
1721 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1722 	    &rack_highest_move_thresh_seen, 0,
1723 	    "Highest move to non-move ratio seen");
1724 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1725 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1726 	    SYSCTL_CHILDREN(rack_attack),
1727 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1728 	    &rack_ack_total,
1729 	    "Total number of Ack's");
1730 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1731 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1732 	    SYSCTL_CHILDREN(rack_attack),
1733 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1734 	    &rack_express_sack,
1735 	    "Total expresss number of Sack's");
1736 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1737 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1738 	    SYSCTL_CHILDREN(rack_attack),
1739 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1740 	    &rack_sack_total,
1741 	    "Total number of SACKs");
1742 	rack_move_none = counter_u64_alloc(M_WAITOK);
1743 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744 	    SYSCTL_CHILDREN(rack_attack),
1745 	    OID_AUTO, "move_none", CTLFLAG_RD,
1746 	    &rack_move_none,
1747 	    "Total number of SACK index reuse of positions under threshold");
1748 	rack_move_some = counter_u64_alloc(M_WAITOK);
1749 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750 	    SYSCTL_CHILDREN(rack_attack),
1751 	    OID_AUTO, "move_some", CTLFLAG_RD,
1752 	    &rack_move_some,
1753 	    "Total number of SACK index reuse of positions over threshold");
1754 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1755 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756 	    SYSCTL_CHILDREN(rack_attack),
1757 	    OID_AUTO, "attacks", CTLFLAG_RD,
1758 	    &rack_sack_attacks_detected,
1759 	    "Total number of SACK attackers that had sack disabled");
1760 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1761 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1762 	    SYSCTL_CHILDREN(rack_attack),
1763 	    OID_AUTO, "reversed", CTLFLAG_RD,
1764 	    &rack_sack_attacks_reversed,
1765 	    "Total number of SACK attackers that were later determined false positive");
1766 	rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1767 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1768 	    SYSCTL_CHILDREN(rack_attack),
1769 	    OID_AUTO, "suspect", CTLFLAG_RD,
1770 	    &rack_sack_attacks_suspect,
1771 	    "Total number of SACKs that triggered early detection");
1772 
1773 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1774 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1775 	    SYSCTL_CHILDREN(rack_attack),
1776 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1777 	    &rack_sack_used_next_merge,
1778 	    "Total number of times we used the next merge");
1779 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1780 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1781 	    SYSCTL_CHILDREN(rack_attack),
1782 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1783 	    &rack_sack_used_prev_merge,
1784 	    "Total number of times we used the prev merge");
1785 	/* Counters */
1786 	rack_total_bytes = counter_u64_alloc(M_WAITOK);
1787 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1788 	    SYSCTL_CHILDREN(rack_counters),
1789 	    OID_AUTO, "totalbytes", CTLFLAG_RD,
1790 	    &rack_total_bytes,
1791 	    "Total number of bytes sent");
1792 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1793 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1794 	    SYSCTL_CHILDREN(rack_counters),
1795 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1796 	    &rack_fto_send, "Total number of rack_fast_output sends");
1797 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1798 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1799 	    SYSCTL_CHILDREN(rack_counters),
1800 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1801 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1802 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1803 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1804 	    SYSCTL_CHILDREN(rack_counters),
1805 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1806 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1807 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1808 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1809 	    SYSCTL_CHILDREN(rack_counters),
1810 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1811 	    &rack_non_fto_send, "Total number of rack_output first sends");
1812 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1813 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1814 	    SYSCTL_CHILDREN(rack_counters),
1815 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1816 	    &rack_extended_rfo, "Total number of times we extended rfo");
1817 
1818 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1819 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1820 	    SYSCTL_CHILDREN(rack_counters),
1821 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1822 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1823 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1824 
1825 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1826 	    SYSCTL_CHILDREN(rack_counters),
1827 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1828 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1829 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1830 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1831 	    SYSCTL_CHILDREN(rack_counters),
1832 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1833 	    &rack_tlp_tot,
1834 	    "Total number of tail loss probe expirations");
1835 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1836 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1837 	    SYSCTL_CHILDREN(rack_counters),
1838 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1839 	    &rack_tlp_newdata,
1840 	    "Total number of tail loss probe sending new data");
1841 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1842 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1843 	    SYSCTL_CHILDREN(rack_counters),
1844 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1845 	    &rack_tlp_retran,
1846 	    "Total number of tail loss probe sending retransmitted data");
1847 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1848 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1849 	    SYSCTL_CHILDREN(rack_counters),
1850 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1851 	    &rack_tlp_retran_bytes,
1852 	    "Total bytes of tail loss probe sending retransmitted data");
1853 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1854 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1855 	    SYSCTL_CHILDREN(rack_counters),
1856 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1857 	    &rack_to_tot,
1858 	    "Total number of times the rack to expired");
1859 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1860 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1861 	    SYSCTL_CHILDREN(rack_counters),
1862 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1863 	    &rack_saw_enobuf,
1864 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1865 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1866 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1867 	    SYSCTL_CHILDREN(rack_counters),
1868 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1869 	    &rack_saw_enobuf_hw,
1870 	    "Total number of times a send returned enobuf for hdwr paced connections");
1871 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1872 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1873 	    SYSCTL_CHILDREN(rack_counters),
1874 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1875 	    &rack_saw_enetunreach,
1876 	    "Total number of times a send received a enetunreachable");
1877 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1878 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1879 	    SYSCTL_CHILDREN(rack_counters),
1880 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1881 	    &rack_hot_alloc,
1882 	    "Total allocations from the top of our list");
1883 	tcp_policer_detected = counter_u64_alloc(M_WAITOK);
1884 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1885 	    SYSCTL_CHILDREN(rack_counters),
1886 	    OID_AUTO, "policer_detected", CTLFLAG_RD,
1887 	    &tcp_policer_detected,
1888 	    "Total policer_detections");
1889 
1890 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1891 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1892 	    SYSCTL_CHILDREN(rack_counters),
1893 	    OID_AUTO, "allocs", CTLFLAG_RD,
1894 	    &rack_to_alloc,
1895 	    "Total allocations of tracking structures");
1896 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1897 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1898 	    SYSCTL_CHILDREN(rack_counters),
1899 	    OID_AUTO, "allochard", CTLFLAG_RD,
1900 	    &rack_to_alloc_hard,
1901 	    "Total allocations done with sleeping the hard way");
1902 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1903 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1904 	    SYSCTL_CHILDREN(rack_counters),
1905 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1906 	    &rack_to_alloc_emerg,
1907 	    "Total allocations done from emergency cache");
1908 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1909 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1910 	    SYSCTL_CHILDREN(rack_counters),
1911 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1912 	    &rack_to_alloc_limited,
1913 	    "Total allocations dropped due to limit");
1914 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1915 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1916 	    SYSCTL_CHILDREN(rack_counters),
1917 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1918 	    &rack_alloc_limited_conns,
1919 	    "Connections with allocations dropped due to limit");
1920 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1921 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1922 	    SYSCTL_CHILDREN(rack_counters),
1923 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1924 	    &rack_split_limited,
1925 	    "Split allocations dropped due to limit");
1926 	rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1927 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1928 	    SYSCTL_CHILDREN(rack_counters),
1929 	    OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1930 	    &rack_rxt_clamps_cwnd,
1931 	    "Number of times that excessive rxt clamped the cwnd down");
1932 	rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1933 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1934 	    SYSCTL_CHILDREN(rack_counters),
1935 	    OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1936 	    &rack_rxt_clamps_cwnd_uniq,
1937 	    "Number of connections that have had excessive rxt clamped the cwnd down");
1938 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1939 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1940 	    SYSCTL_CHILDREN(rack_counters),
1941 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1942 	    &rack_persists_sends,
1943 	    "Number of times we sent a persist probe");
1944 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1945 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1946 	    SYSCTL_CHILDREN(rack_counters),
1947 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1948 	    &rack_persists_acks,
1949 	    "Number of times a persist probe was acked");
1950 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1951 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1952 	    SYSCTL_CHILDREN(rack_counters),
1953 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1954 	    &rack_persists_loss,
1955 	    "Number of times we detected a lost persist probe (no ack)");
1956 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1957 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1958 	    SYSCTL_CHILDREN(rack_counters),
1959 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1960 	    &rack_persists_lost_ends,
1961 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1962 #ifdef INVARIANTS
1963 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1964 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1965 	    SYSCTL_CHILDREN(rack_counters),
1966 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1967 	    &rack_adjust_map_bw,
1968 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1969 #endif
1970 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1971 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1972 	    SYSCTL_CHILDREN(rack_counters),
1973 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1974 	    &rack_multi_single_eq,
1975 	    "Number of compressed acks total represented");
1976 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1977 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1978 	    SYSCTL_CHILDREN(rack_counters),
1979 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1980 	    &rack_proc_non_comp_ack,
1981 	    "Number of non compresseds acks that we processed");
1982 
1983 
1984 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1985 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1986 	    SYSCTL_CHILDREN(rack_counters),
1987 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1988 	    &rack_sack_proc_all,
1989 	    "Total times we had to walk whole list for sack processing");
1990 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1991 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1992 	    SYSCTL_CHILDREN(rack_counters),
1993 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1994 	    &rack_sack_proc_restart,
1995 	    "Total times we had to walk whole list due to a restart");
1996 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1997 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1998 	    SYSCTL_CHILDREN(rack_counters),
1999 	    OID_AUTO, "sack_short", CTLFLAG_RD,
2000 	    &rack_sack_proc_short,
2001 	    "Total times we took shortcut for sack processing");
2002 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
2003 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2004 	    SYSCTL_CHILDREN(rack_attack),
2005 	    OID_AUTO, "skipacked", CTLFLAG_RD,
2006 	    &rack_sack_skipped_acked,
2007 	    "Total number of times we skipped previously sacked");
2008 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
2009 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2010 	    SYSCTL_CHILDREN(rack_attack),
2011 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
2012 	    &rack_sack_splits,
2013 	    "Total number of times we did the old fashion tree split");
2014 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
2015 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2016 	    SYSCTL_CHILDREN(rack_counters),
2017 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
2018 	    &rack_input_idle_reduces,
2019 	    "Total number of idle reductions on input");
2020 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
2021 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2022 	    SYSCTL_CHILDREN(rack_counters),
2023 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
2024 	    &rack_collapsed_win_seen,
2025 	    "Total number of collapsed window events seen (where our window shrinks)");
2026 
2027 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
2028 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2029 	    SYSCTL_CHILDREN(rack_counters),
2030 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
2031 	    &rack_collapsed_win,
2032 	    "Total number of collapsed window events where we mark packets");
2033 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
2034 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2035 	    SYSCTL_CHILDREN(rack_counters),
2036 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
2037 	    &rack_collapsed_win_rxt,
2038 	    "Total number of packets that were retransmitted");
2039 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
2040 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2041 	    SYSCTL_CHILDREN(rack_counters),
2042 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
2043 	    &rack_collapsed_win_rxt_bytes,
2044 	    "Total number of bytes that were retransmitted");
2045 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
2046 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2047 	    SYSCTL_CHILDREN(rack_counters),
2048 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
2049 	    &rack_try_scwnd,
2050 	    "Total number of scwnd attempts");
2051 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
2052 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
2053 	    OID_AUTO, "outsize", CTLFLAG_RD,
2054 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
2055 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
2056 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
2057 	    OID_AUTO, "opts", CTLFLAG_RD,
2058 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
2059 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
2060 	    SYSCTL_CHILDREN(rack_sysctl_root),
2061 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2062 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
2063 }
2064 
2065 static uint32_t
2066 rc_init_window(struct tcp_rack *rack)
2067 {
2068 	return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
2069 
2070 }
2071 
2072 static uint64_t
2073 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
2074 {
2075 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
2076 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
2077 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2078 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
2079 	else
2080 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
2081 }
2082 
2083 static void
2084 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
2085 	uint64_t data, uint8_t mod, uint16_t aux,
2086 	struct tcp_sendfile_track *cur, int line)
2087 {
2088 #ifdef TCP_REQUEST_TRK
2089 	int do_log = 0;
2090 
2091 	/*
2092 	 * The rate cap one is noisy and only should come out when normal BB logging
2093 	 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
2094 	 * once per chunk and make up the BBpoint that can be turned on by the client.
2095 	 */
2096 	if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2097 		/*
2098 		 * The very noisy two need to only come out when
2099 		 * we have verbose logging on.
2100 		 */
2101 		if (rack_verbose_logging != 0)
2102 			do_log = tcp_bblogging_on(rack->rc_tp);
2103 		else
2104 			do_log = 0;
2105 	} else if (mod != HYBRID_LOG_BW_MEASURE) {
2106 		/*
2107 		 * All other less noisy logs here except the measure which
2108 		 * also needs to come out on the point and the log.
2109 		 */
2110 		do_log = tcp_bblogging_on(rack->rc_tp);
2111 	} else {
2112 		do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2113 	}
2114 
2115 	if (do_log) {
2116 		union tcp_log_stackspecific log;
2117 		struct timeval tv;
2118 		uint64_t lt_bw;
2119 
2120 		/* Convert our ms to a microsecond */
2121 		memset(&log, 0, sizeof(log));
2122 
2123 		log.u_bbr.cwnd_gain = line;
2124 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2125 		log.u_bbr.rttProp = tim;
2126 		log.u_bbr.bw_inuse = cbw;
2127 		log.u_bbr.delRate = rack_get_gp_est(rack);
2128 		lt_bw = rack_get_lt_bw(rack);
2129 		log.u_bbr.flex1 = seq;
2130 		log.u_bbr.pacing_gain = aux;
2131 		/* lt_bw = < flex3 | flex2 > */
2132 		log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2133 		log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2134 		/* Record the last obtained us rtt in inflight */
2135 		if (cur == NULL) {
2136 			/* Make sure we are looking at the right log if an overide comes in */
2137 			cur = rack->r_ctl.rc_last_sft;
2138 		}
2139 		if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2140 			log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2141 		else {
2142 			/* Use the last known rtt i.e. the rack-rtt */
2143 			log.u_bbr.inflight = rack->rc_rack_rtt;
2144 		}
2145 		if (cur != NULL) {
2146 			uint64_t off;
2147 
2148 			log.u_bbr.cur_del_rate = cur->deadline;
2149 			if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2150 				/* start = < lost | pkt_epoch > */
2151 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2152 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2153 				log.u_bbr.flex6 = cur->start_seq;
2154 				log.u_bbr.pkts_out = cur->end_seq;
2155 			} else {
2156 				/* start = < lost | pkt_epoch > */
2157 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2158 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2159 				/* end = < pkts_out | flex6 > */
2160 				log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2161 				log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2162 			}
2163 			/* first_send = <lt_epoch | epoch> */
2164 			log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2165 			log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2166 			/* localtime = <delivered | applimited>*/
2167 			log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2168 			log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2169 #ifdef TCP_REQUEST_TRK
2170 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2171 			log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2172 #endif
2173 			log.u_bbr.inhpts = 1;
2174 			log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2175 			log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2176 			log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2177 		} else {
2178 			log.u_bbr.flex7 = 0xffff;
2179 			log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2180 		}
2181 		/*
2182 		 * Compose bbr_state to be a bit wise 0000ADHF
2183 		 * where A is the always_pace flag
2184 		 * where D is the dgp_on flag
2185 		 * where H is the hybrid_mode on flag
2186 		 * where F is the use_fixed_rate flag.
2187 		 */
2188 		log.u_bbr.bbr_state = rack->rc_always_pace;
2189 		log.u_bbr.bbr_state <<= 1;
2190 		log.u_bbr.bbr_state |= rack->dgp_on;
2191 		log.u_bbr.bbr_state <<= 1;
2192 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2193 		log.u_bbr.bbr_state <<= 1;
2194 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2195 		log.u_bbr.flex8 = mod;
2196 		tcp_log_event(rack->rc_tp, NULL,
2197 		    &rack->rc_inp->inp_socket->so_rcv,
2198 		    &rack->rc_inp->inp_socket->so_snd,
2199 		    TCP_HYBRID_PACING_LOG, 0,
2200 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2201 
2202 	}
2203 #endif
2204 }
2205 
2206 #ifdef TCP_REQUEST_TRK
2207 static void
2208 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2209 {
2210 	if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2211 		union tcp_log_stackspecific log;
2212 		struct timeval tv;
2213 		uint64_t off;
2214 
2215 		/* Convert our ms to a microsecond */
2216 		memset(&log, 0, sizeof(log));
2217 
2218 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2219 		log.u_bbr.delRate = cur->sent_at_fs;
2220 
2221 		if ((cur->flags & TCP_TRK_TRACK_FLG_LSND) == 0) {
2222 			/*
2223 			 * We did not get a new Rules Applied to set so
2224 			 * no overlapping send occured, this means the
2225 			 * current byte counts are correct.
2226 			 */
2227 			log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2228 			log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2229 		} else {
2230 			/*
2231 			 * Overlapping send case, we switched to a new
2232 			 * send and did a rules applied.
2233 			 */
2234 			log.u_bbr.cur_del_rate = cur->sent_at_ls;
2235 			log.u_bbr.rttProp = cur->rxt_at_ls;
2236 		}
2237 		log.u_bbr.bw_inuse = cur->rxt_at_fs;
2238 		log.u_bbr.cwnd_gain = line;
2239 		off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2240 		log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2241 		/* start = < flex1 | flex2 > */
2242 		log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2243 		log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2244 		/* end = < flex3 | flex4 > */
2245 		log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2246 		log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2247 
2248 		/* localtime = <delivered | applimited>*/
2249 		log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2250 		log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2251 		/* client timestamp = <lt_epoch | epoch>*/
2252 		log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2253 		log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2254 		/* now set all the flags in */
2255 		log.u_bbr.pkts_out = cur->hybrid_flags;
2256 		log.u_bbr.lost = cur->playout_ms;
2257 		log.u_bbr.flex6 = cur->flags;
2258 		/*
2259 		 * Last send time  = <flex5 | pkt_epoch>  note we do not distinguish cases
2260 		 * where a false retransmit occurred so first_send  <-> lastsend may
2261 		 * include longer time then it actually took if we have a false rxt.
2262 		 */
2263 		log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2264 		log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2265 		/*
2266 		 * Compose bbr_state to be a bit wise 0000ADHF
2267 		 * where A is the always_pace flag
2268 		 * where D is the dgp_on flag
2269 		 * where H is the hybrid_mode on flag
2270 		 * where F is the use_fixed_rate flag.
2271 		 */
2272 		log.u_bbr.bbr_state = rack->rc_always_pace;
2273 		log.u_bbr.bbr_state <<= 1;
2274 		log.u_bbr.bbr_state |= rack->dgp_on;
2275 		log.u_bbr.bbr_state <<= 1;
2276 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2277 		log.u_bbr.bbr_state <<= 1;
2278 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2279 
2280 		log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2281 		tcp_log_event(rack->rc_tp, NULL,
2282 		    &rack->rc_inp->inp_socket->so_rcv,
2283 		    &rack->rc_inp->inp_socket->so_snd,
2284 		    TCP_HYBRID_PACING_LOG, 0,
2285 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2286 	}
2287 }
2288 #endif
2289 
2290 static inline uint64_t
2291 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2292 {
2293 	uint64_t ret_bw, ether;
2294 	uint64_t u_segsiz;
2295 
2296 	ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2297 	if (rack->r_is_v6){
2298 #ifdef INET6
2299 		ether += sizeof(struct ip6_hdr);
2300 #endif
2301 		ether += 14;	/* eheader size 6+6+2 */
2302 	} else {
2303 #ifdef INET
2304 		ether += sizeof(struct ip);
2305 #endif
2306 		ether += 14;	/* eheader size 6+6+2 */
2307 	}
2308 	u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2309 	ret_bw = bw;
2310 	ret_bw *= ether;
2311 	ret_bw /= u_segsiz;
2312 	return (ret_bw);
2313 }
2314 
2315 static void
2316 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2317 {
2318 #ifdef TCP_REQUEST_TRK
2319 	struct timeval tv;
2320 	uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2321 #endif
2322 
2323 	if (rack->r_ctl.bw_rate_cap == 0)
2324 		return;
2325 #ifdef TCP_REQUEST_TRK
2326 	if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2327 	    (rack->r_ctl.rc_last_sft != NULL)) {
2328 		/*
2329 		 * We have a dynamic cap. The original target
2330 		 * is in bw_rate_cap, but we need to look at
2331 		 * how long it is until we hit the deadline.
2332 		 */
2333 		struct tcp_sendfile_track *ent;
2334 
2335       		ent = rack->r_ctl.rc_last_sft;
2336 		microuptime(&tv);
2337 		timenow = tcp_tv_to_lusectick(&tv);
2338 		if (timenow >= ent->deadline) {
2339 			/* No time left we do DGP only */
2340 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2341 					   0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2342 			rack->r_ctl.bw_rate_cap = 0;
2343 			return;
2344 		}
2345 		/* We have the time */
2346 		timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2347 		if (timeleft < HPTS_MSEC_IN_SEC) {
2348 			/* If there is less than a ms left just use DGPs rate */
2349 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2350 					   0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2351 			rack->r_ctl.bw_rate_cap = 0;
2352 			return;
2353 		}
2354 		/*
2355 		 * Now lets find the amount of data left to send.
2356 		 *
2357 		 * Now ideally we want to use the end_seq to figure out how much more
2358 		 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2359 		 */
2360 		if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2361 			if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2362 				lenleft = ent->end_seq - rack->rc_tp->snd_una;
2363 			else {
2364 				/* TSNH, we should catch it at the send */
2365 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2366 						   0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2367 				rack->r_ctl.bw_rate_cap = 0;
2368 				return;
2369 			}
2370 		} else {
2371 			/*
2372 			 * The hard way, figure out how much is gone and then
2373 			 * take that away from the total the client asked for
2374 			 * (thats off by tls overhead if this is tls).
2375 			 */
2376 			if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2377 				lengone = rack->rc_tp->snd_una - ent->start_seq;
2378 			else
2379 				lengone = 0;
2380 			if (lengone < (ent->end - ent->start))
2381 				lenleft = (ent->end - ent->start) - lengone;
2382 			else {
2383 				/* TSNH, we should catch it at the send */
2384 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2385 						   0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2386 				rack->r_ctl.bw_rate_cap = 0;
2387 				return;
2388 			}
2389 		}
2390 		if (lenleft == 0) {
2391 			/* We have it all sent */
2392 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2393 					   0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2394 			if (rack->r_ctl.bw_rate_cap)
2395 				goto normal_ratecap;
2396 			else
2397 				return;
2398 		}
2399 		calcbw = lenleft * HPTS_USEC_IN_SEC;
2400 		calcbw /= timeleft;
2401 		/* Now we must compensate for IP/TCP overhead */
2402 		calcbw = rack_compensate_for_linerate(rack, calcbw);
2403 		/* Update the bit rate cap */
2404 		rack->r_ctl.bw_rate_cap = calcbw;
2405 		if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2406 		    (rack_hybrid_allow_set_maxseg == 1) &&
2407 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2408 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2409 			uint32_t orig_max;
2410 
2411 			orig_max = rack->r_ctl.rc_pace_max_segs;
2412 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2413 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2414 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2415 		}
2416 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2417 				   calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2418 		if ((calcbw > 0) && (*bw > calcbw)) {
2419 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2420 					   *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2421 			*capped = 1;
2422 			*bw = calcbw;
2423 		}
2424 		return;
2425 	}
2426 normal_ratecap:
2427 #endif
2428 	if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2429 #ifdef TCP_REQUEST_TRK
2430 		if (rack->rc_hybrid_mode &&
2431 		    rack->rc_catch_up &&
2432 		    (rack->r_ctl.rc_last_sft != NULL) &&
2433 		    (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2434 		    (rack_hybrid_allow_set_maxseg == 1) &&
2435 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2436 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2437 			uint32_t orig_max;
2438 
2439 			orig_max = rack->r_ctl.rc_pace_max_segs;
2440 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2441 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2442 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2443 		}
2444 #endif
2445 		*capped = 1;
2446 		*bw = rack->r_ctl.bw_rate_cap;
2447 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2448 				   *bw, 0, 0,
2449 				   HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2450 	}
2451 }
2452 
2453 static uint64_t
2454 rack_get_gp_est(struct tcp_rack *rack)
2455 {
2456 	uint64_t bw, lt_bw, ret_bw;
2457 
2458 	if (rack->rc_gp_filled == 0) {
2459 		/*
2460 		 * We have yet no b/w measurement,
2461 		 * if we have a user set initial bw
2462 		 * return it. If we don't have that and
2463 		 * we have an srtt, use the tcp IW (10) to
2464 		 * calculate a fictional b/w over the SRTT
2465 		 * which is more or less a guess. Note
2466 		 * we don't use our IW from rack on purpose
2467 		 * so if we have like IW=30, we are not
2468 		 * calculating a "huge" b/w.
2469 		 */
2470 		uint64_t srtt;
2471 
2472 		if (rack->dis_lt_bw == 1)
2473 			lt_bw = 0;
2474 		else
2475 			lt_bw = rack_get_lt_bw(rack);
2476 		if (lt_bw) {
2477 			/*
2478 			 * No goodput bw but a long-term b/w does exist
2479 			 * lets use that.
2480 			 */
2481 			ret_bw = lt_bw;
2482 			goto compensate;
2483 		}
2484 		if (rack->r_ctl.init_rate)
2485 			return (rack->r_ctl.init_rate);
2486 
2487 		/* Ok lets come up with the IW guess, if we have a srtt */
2488 		if (rack->rc_tp->t_srtt == 0) {
2489 			/*
2490 			 * Go with old pacing method
2491 			 * i.e. burst mitigation only.
2492 			 */
2493 			return (0);
2494 		}
2495 		/* Ok lets get the initial TCP win (not racks) */
2496 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2497 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2498 		bw *= (uint64_t)USECS_IN_SECOND;
2499 		bw /= srtt;
2500 		ret_bw = bw;
2501 		goto compensate;
2502 
2503 	}
2504 	if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2505 		/* Averaging is done, we can return the value */
2506 		bw = rack->r_ctl.gp_bw;
2507 	} else {
2508 		/* Still doing initial average must calculate */
2509 		bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2510 	}
2511 	if (rack->dis_lt_bw) {
2512 		/* We are not using lt-bw */
2513 		ret_bw = bw;
2514 		goto compensate;
2515 	}
2516 	lt_bw = rack_get_lt_bw(rack);
2517 	if (lt_bw == 0) {
2518 		/* If we don't have one then equate it to the gp_bw */
2519 		lt_bw = rack->r_ctl.gp_bw;
2520 	}
2521 	if (rack->use_lesser_lt_bw) {
2522 		if (lt_bw < bw)
2523 			ret_bw = lt_bw;
2524 		else
2525 			ret_bw = bw;
2526 	} else {
2527 		if (lt_bw > bw)
2528 			ret_bw = lt_bw;
2529 		else
2530 			ret_bw = bw;
2531 	}
2532 	/*
2533 	 * Now lets compensate based on the TCP/IP overhead. Our
2534 	 * Goodput estimate does not include this so we must pace out
2535 	 * a bit faster since our pacing calculations do. The pacing
2536 	 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2537 	 * we are using to do this, so we do that here in the opposite
2538 	 * direction as well. This means that if we are tunneled and the
2539 	 * segsiz is say 1200 bytes we will get quite a boost, but its
2540 	 * compensated for in the pacing time the opposite way.
2541 	 */
2542 compensate:
2543 	ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2544 	return(ret_bw);
2545 }
2546 
2547 
2548 static uint64_t
2549 rack_get_bw(struct tcp_rack *rack)
2550 {
2551 	uint64_t bw;
2552 
2553 	if (rack->use_fixed_rate) {
2554 		/* Return the fixed pacing rate */
2555 		return (rack_get_fixed_pacing_bw(rack));
2556 	}
2557 	bw = rack_get_gp_est(rack);
2558 	return (bw);
2559 }
2560 
2561 static uint16_t
2562 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2563 {
2564 	if (rack->use_fixed_rate) {
2565 		return (100);
2566 	} else if (rack->in_probe_rtt && (rsm == NULL))
2567 		return (rack->r_ctl.rack_per_of_gp_probertt);
2568 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2569 		  rack->r_ctl.rack_per_of_gp_rec)) {
2570 		if (rsm) {
2571 			/* a retransmission always use the recovery rate */
2572 			return (rack->r_ctl.rack_per_of_gp_rec);
2573 		} else if (rack->rack_rec_nonrxt_use_cr) {
2574 			/* Directed to use the configured rate */
2575 			goto configured_rate;
2576 		} else if (rack->rack_no_prr &&
2577 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2578 			/* No PRR, lets just use the b/w estimate only */
2579 			return (100);
2580 		} else {
2581 			/*
2582 			 * Here we may have a non-retransmit but we
2583 			 * have no overrides, so just use the recovery
2584 			 * rate (prr is in effect).
2585 			 */
2586 			return (rack->r_ctl.rack_per_of_gp_rec);
2587 		}
2588 	}
2589 configured_rate:
2590 	/* For the configured rate we look at our cwnd vs the ssthresh */
2591 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2592 		return (rack->r_ctl.rack_per_of_gp_ss);
2593 	else
2594 		return (rack->r_ctl.rack_per_of_gp_ca);
2595 }
2596 
2597 static void
2598 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2599 {
2600 	/*
2601 	 * Types of logs (mod value)
2602 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2603 	 * 2 = a dsack round begins, persist is reset to 16.
2604 	 * 3 = a dsack round ends
2605 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2606 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2607 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2608 	 */
2609 	if (tcp_bblogging_on(rack->rc_tp)) {
2610 		union tcp_log_stackspecific log;
2611 		struct timeval tv;
2612 
2613 		memset(&log, 0, sizeof(log));
2614 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2615 		log.u_bbr.flex1 <<= 1;
2616 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2617 		log.u_bbr.flex1 <<= 1;
2618 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2619 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2620 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2621 		log.u_bbr.flex4 = flex4;
2622 		log.u_bbr.flex5 = flex5;
2623 		log.u_bbr.flex6 = flex6;
2624 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2625 		log.u_bbr.flex8 = mod;
2626 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2627 		log.u_bbr.epoch = rack->r_ctl.current_round;
2628 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2629 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2630 		    &rack->rc_inp->inp_socket->so_rcv,
2631 		    &rack->rc_inp->inp_socket->so_snd,
2632 		    RACK_DSACK_HANDLING, 0,
2633 		    0, &log, false, &tv);
2634 	}
2635 }
2636 
2637 static void
2638 rack_log_hdwr_pacing(struct tcp_rack *rack,
2639 		     uint64_t rate, uint64_t hw_rate, int line,
2640 		     int error, uint16_t mod)
2641 {
2642 	if (tcp_bblogging_on(rack->rc_tp)) {
2643 		union tcp_log_stackspecific log;
2644 		struct timeval tv;
2645 		const struct ifnet *ifp;
2646 		uint64_t ifp64;
2647 
2648 		memset(&log, 0, sizeof(log));
2649 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2650 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2651 		if (rack->r_ctl.crte) {
2652 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2653 		} else if (rack->rc_inp->inp_route.ro_nh &&
2654 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2655 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2656 		} else
2657 			ifp = NULL;
2658 		if (ifp) {
2659 			ifp64 = (uintptr_t)ifp;
2660 			log.u_bbr.flex3 = ((ifp64  >> 32) & 0x00000000ffffffff);
2661 			log.u_bbr.flex4 = (ifp64 & 0x00000000ffffffff);
2662 		}
2663 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2664 		log.u_bbr.bw_inuse = rate;
2665 		log.u_bbr.flex5 = line;
2666 		log.u_bbr.flex6 = error;
2667 		log.u_bbr.flex7 = mod;
2668 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2669 		log.u_bbr.flex8 = rack->use_fixed_rate;
2670 		log.u_bbr.flex8 <<= 1;
2671 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2672 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2673 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2674 		if (rack->r_ctl.crte)
2675 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2676 		else
2677 			log.u_bbr.cur_del_rate = 0;
2678 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2679 		log.u_bbr.epoch = rack->r_ctl.current_round;
2680 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2681 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2682 		    &rack->rc_inp->inp_socket->so_rcv,
2683 		    &rack->rc_inp->inp_socket->so_snd,
2684 		    BBR_LOG_HDWR_PACE, 0,
2685 		    0, &log, false, &tv);
2686 	}
2687 }
2688 
2689 static uint64_t
2690 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2691 {
2692 	/*
2693 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2694 	 */
2695 	uint64_t bw_est, high_rate;
2696 	uint64_t gain;
2697 
2698 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2699 	bw_est = bw * gain;
2700 	bw_est /= (uint64_t)100;
2701 	/* Never fall below the minimum (def 64kbps) */
2702 	if (bw_est < RACK_MIN_BW)
2703 		bw_est = RACK_MIN_BW;
2704 	if (rack->r_rack_hw_rate_caps) {
2705 		/* Rate caps are in place */
2706 		if (rack->r_ctl.crte != NULL) {
2707 			/* We have a hdwr rate already */
2708 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2709 			if (bw_est >= high_rate) {
2710 				/* We are capping bw at the highest rate table entry */
2711 				if (rack_hw_rate_cap_per &&
2712 				    (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2713 					rack->r_rack_hw_rate_caps = 0;
2714 					goto done;
2715 				}
2716 				rack_log_hdwr_pacing(rack,
2717 						     bw_est, high_rate, __LINE__,
2718 						     0, 3);
2719 				bw_est = high_rate;
2720 				if (capped)
2721 					*capped = 1;
2722 			}
2723 		} else if ((rack->rack_hdrw_pacing == 0) &&
2724 			   (rack->rack_hdw_pace_ena) &&
2725 			   (rack->rack_attempt_hdwr_pace == 0) &&
2726 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2727 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2728 			/*
2729 			 * Special case, we have not yet attempted hardware
2730 			 * pacing, and yet we may, when we do, find out if we are
2731 			 * above the highest rate. We need to know the maxbw for the interface
2732 			 * in question (if it supports ratelimiting). We get back
2733 			 * a 0, if the interface is not found in the RL lists.
2734 			 */
2735 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2736 			if (high_rate) {
2737 				/* Yep, we have a rate is it above this rate? */
2738 				if (bw_est > high_rate) {
2739 					bw_est = high_rate;
2740 					if (capped)
2741 						*capped = 1;
2742 				}
2743 			}
2744 		}
2745 	}
2746 done:
2747 	return (bw_est);
2748 }
2749 
2750 static void
2751 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2752 {
2753 	if (tcp_bblogging_on(rack->rc_tp)) {
2754 		union tcp_log_stackspecific log;
2755 		struct timeval tv;
2756 
2757 		if ((mod != 1) && (rack_verbose_logging == 0))  {
2758 			/*
2759 			 * We get 3 values currently for mod
2760 			 * 1 - We are retransmitting and this tells the reason.
2761 			 * 2 - We are clearing a dup-ack count.
2762 			 * 3 - We are incrementing a dup-ack count.
2763 			 *
2764 			 * The clear/increment are only logged
2765 			 * if you have BBverbose on.
2766 			 */
2767 			return;
2768 		}
2769 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2770 		log.u_bbr.flex1 = tsused;
2771 		log.u_bbr.flex2 = thresh;
2772 		log.u_bbr.flex3 = rsm->r_flags;
2773 		log.u_bbr.flex4 = rsm->r_dupack;
2774 		log.u_bbr.flex5 = rsm->r_start;
2775 		log.u_bbr.flex6 = rsm->r_end;
2776 		log.u_bbr.flex8 = mod;
2777 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2778 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2779 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2780 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2781 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2782 		log.u_bbr.pacing_gain = rack->r_must_retran;
2783 		log.u_bbr.epoch = rack->r_ctl.current_round;
2784 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2785 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2786 		    &rack->rc_inp->inp_socket->so_rcv,
2787 		    &rack->rc_inp->inp_socket->so_snd,
2788 		    BBR_LOG_SETTINGS_CHG, 0,
2789 		    0, &log, false, &tv);
2790 	}
2791 }
2792 
2793 static void
2794 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2795 {
2796 	if (tcp_bblogging_on(rack->rc_tp)) {
2797 		union tcp_log_stackspecific log;
2798 		struct timeval tv;
2799 
2800 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2801 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2802 		log.u_bbr.flex2 = to;
2803 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2804 		log.u_bbr.flex4 = slot;
2805 		log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2806 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2807 		log.u_bbr.flex7 = rack->rc_in_persist;
2808 		log.u_bbr.flex8 = which;
2809 		if (rack->rack_no_prr)
2810 			log.u_bbr.pkts_out = 0;
2811 		else
2812 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2813 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2814 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2815 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2816 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2817 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2818 		log.u_bbr.pacing_gain = rack->r_must_retran;
2819 		log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2820 		log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2821 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2822 		log.u_bbr.lost = rack_rto_min;
2823 		log.u_bbr.epoch = rack->r_ctl.roundends;
2824 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2825 		log.u_bbr.bw_inuse <<= 32;
2826 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2827 		log.u_bbr.applimited = rack->rc_tp->t_flags2;
2828 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2829 		    &rack->rc_inp->inp_socket->so_rcv,
2830 		    &rack->rc_inp->inp_socket->so_snd,
2831 		    BBR_LOG_TIMERSTAR, 0,
2832 		    0, &log, false, &tv);
2833 	}
2834 }
2835 
2836 static void
2837 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2838 {
2839 	if (tcp_bblogging_on(rack->rc_tp)) {
2840 		union tcp_log_stackspecific log;
2841 		struct timeval tv;
2842 
2843 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2844 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2845 		log.u_bbr.flex8 = to_num;
2846 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2847 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2848 		if (rsm == NULL)
2849 			log.u_bbr.flex3 = 0;
2850 		else
2851 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2852 		if (rack->rack_no_prr)
2853 			log.u_bbr.flex5 = 0;
2854 		else
2855 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2856 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2857 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2858 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2859 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2860 		log.u_bbr.pacing_gain = rack->r_must_retran;
2861 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2862 		log.u_bbr.bw_inuse <<= 32;
2863 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2864 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2865 		    &rack->rc_inp->inp_socket->so_rcv,
2866 		    &rack->rc_inp->inp_socket->so_snd,
2867 		    BBR_LOG_RTO, 0,
2868 		    0, &log, false, &tv);
2869 	}
2870 }
2871 
2872 static void
2873 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2874 		 struct rack_sendmap *prev,
2875 		 struct rack_sendmap *rsm,
2876 		 struct rack_sendmap *next,
2877 		 int flag, uint32_t th_ack, int line)
2878 {
2879 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2880 		union tcp_log_stackspecific log;
2881 		struct timeval tv;
2882 
2883 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2884 		log.u_bbr.flex8 = flag;
2885 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2886 		log.u_bbr.cur_del_rate = (uintptr_t)prev;
2887 		log.u_bbr.delRate = (uintptr_t)rsm;
2888 		log.u_bbr.rttProp = (uintptr_t)next;
2889 		log.u_bbr.flex7 = 0;
2890 		if (prev) {
2891 			log.u_bbr.flex1 = prev->r_start;
2892 			log.u_bbr.flex2 = prev->r_end;
2893 			log.u_bbr.flex7 |= 0x4;
2894 		}
2895 		if (rsm) {
2896 			log.u_bbr.flex3 = rsm->r_start;
2897 			log.u_bbr.flex4 = rsm->r_end;
2898 			log.u_bbr.flex7 |= 0x2;
2899 		}
2900 		if (next) {
2901 			log.u_bbr.flex5 = next->r_start;
2902 			log.u_bbr.flex6 = next->r_end;
2903 			log.u_bbr.flex7 |= 0x1;
2904 		}
2905 		log.u_bbr.applimited = line;
2906 		log.u_bbr.pkts_out = th_ack;
2907 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2908 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2909 		if (rack->rack_no_prr)
2910 			log.u_bbr.lost = 0;
2911 		else
2912 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2913 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2914 		log.u_bbr.bw_inuse <<= 32;
2915 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2916 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2917 		    &rack->rc_inp->inp_socket->so_rcv,
2918 		    &rack->rc_inp->inp_socket->so_snd,
2919 		    TCP_LOG_MAPCHG, 0,
2920 		    0, &log, false, &tv);
2921 	}
2922 }
2923 
2924 static void
2925 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2926 		 struct rack_sendmap *rsm, int conf)
2927 {
2928 	if (tcp_bblogging_on(tp)) {
2929 		union tcp_log_stackspecific log;
2930 		struct timeval tv;
2931 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2932 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2933 		log.u_bbr.flex1 = t;
2934 		log.u_bbr.flex2 = len;
2935 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2936 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2937 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2938 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2939 		log.u_bbr.flex7 = conf;
2940 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2941 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2942 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2943 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2944 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2945 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2946 		if (rsm) {
2947 			log.u_bbr.pkt_epoch = rsm->r_start;
2948 			log.u_bbr.lost = rsm->r_end;
2949 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2950 			/* We loose any upper of the 24 bits */
2951 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2952 		} else {
2953 			/* Its a SYN */
2954 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2955 			log.u_bbr.lost = 0;
2956 			log.u_bbr.cwnd_gain = 0;
2957 			log.u_bbr.pacing_gain = 0;
2958 		}
2959 		/* Write out general bits of interest rrs here */
2960 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2961 		log.u_bbr.use_lt_bw <<= 1;
2962 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2963 		log.u_bbr.use_lt_bw <<= 1;
2964 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2965 		log.u_bbr.use_lt_bw <<= 1;
2966 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2967 		log.u_bbr.use_lt_bw <<= 1;
2968 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2969 		log.u_bbr.use_lt_bw <<= 1;
2970 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2971 		log.u_bbr.use_lt_bw <<= 1;
2972 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2973 		log.u_bbr.use_lt_bw <<= 1;
2974 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2975 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2976 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2977 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2978 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2979 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2980 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2981 		log.u_bbr.bw_inuse <<= 32;
2982 		if (rsm)
2983 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2984 		TCP_LOG_EVENTP(tp, NULL,
2985 		    &rack->rc_inp->inp_socket->so_rcv,
2986 		    &rack->rc_inp->inp_socket->so_snd,
2987 		    BBR_LOG_BBRRTT, 0,
2988 		    0, &log, false, &tv);
2989 
2990 
2991 	}
2992 }
2993 
2994 static void
2995 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2996 {
2997 	/*
2998 	 * Log the rtt sample we are
2999 	 * applying to the srtt algorithm in
3000 	 * useconds.
3001 	 */
3002 	if (tcp_bblogging_on(rack->rc_tp)) {
3003 		union tcp_log_stackspecific log;
3004 		struct timeval tv;
3005 
3006 		/* Convert our ms to a microsecond */
3007 		memset(&log, 0, sizeof(log));
3008 		log.u_bbr.flex1 = rtt;
3009 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3010 		log.u_bbr.flex7 = 1;
3011 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3012 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3013 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3014 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3015 		log.u_bbr.pacing_gain = rack->r_must_retran;
3016 		/*
3017 		 * We capture in delRate the upper 32 bits as
3018 		 * the confidence level we had declared, and the
3019 		 * lower 32 bits as the actual RTT using the arrival
3020 		 * timestamp.
3021 		 */
3022 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
3023 		log.u_bbr.delRate <<= 32;
3024 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
3025 		/* Lets capture all the things that make up t_rtxcur */
3026 		log.u_bbr.applimited = rack_rto_min;
3027 		log.u_bbr.epoch = rack_rto_max;
3028 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
3029 		log.u_bbr.lost = rack_rto_min;
3030 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
3031 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
3032 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
3033 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
3034 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
3035 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3036 		    &rack->rc_inp->inp_socket->so_rcv,
3037 		    &rack->rc_inp->inp_socket->so_snd,
3038 		    TCP_LOG_RTT, 0,
3039 		    0, &log, false, &tv);
3040 	}
3041 }
3042 
3043 static void
3044 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
3045 {
3046 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3047 		union tcp_log_stackspecific log;
3048 		struct timeval tv;
3049 
3050 		/* Convert our ms to a microsecond */
3051 		memset(&log, 0, sizeof(log));
3052 		log.u_bbr.flex1 = rtt;
3053 		log.u_bbr.flex2 = send_time;
3054 		log.u_bbr.flex3 = ack_time;
3055 		log.u_bbr.flex4 = where;
3056 		log.u_bbr.flex7 = 2;
3057 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3058 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3059 		log.u_bbr.bw_inuse <<= 32;
3060 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3061 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3062 		    &rack->rc_inp->inp_socket->so_rcv,
3063 		    &rack->rc_inp->inp_socket->so_snd,
3064 		    TCP_LOG_RTT, 0,
3065 		    0, &log, false, &tv);
3066 	}
3067 }
3068 
3069 
3070 static void
3071 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
3072 {
3073 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3074 		union tcp_log_stackspecific log;
3075 		struct timeval tv;
3076 
3077 		/* Convert our ms to a microsecond */
3078 		memset(&log, 0, sizeof(log));
3079 		log.u_bbr.flex1 = idx;
3080 		log.u_bbr.flex2 = rack_ts_to_msec(tsv);
3081 		log.u_bbr.flex3 = tsecho;
3082 		log.u_bbr.flex7 = 3;
3083 		log.u_bbr.rttProp = tsv;
3084 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3085 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3086 		log.u_bbr.bw_inuse <<= 32;
3087 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3088 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3089 		    &rack->rc_inp->inp_socket->so_rcv,
3090 		    &rack->rc_inp->inp_socket->so_snd,
3091 		    TCP_LOG_RTT, 0,
3092 		    0, &log, false, &tv);
3093 	}
3094 }
3095 
3096 
3097 static inline void
3098 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
3099 {
3100 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3101 		union tcp_log_stackspecific log;
3102 		struct timeval tv;
3103 
3104 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3105 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3106 		log.u_bbr.flex1 = line;
3107 		log.u_bbr.flex2 = tick;
3108 		log.u_bbr.flex3 = tp->t_maxunacktime;
3109 		log.u_bbr.flex4 = tp->t_acktime;
3110 		log.u_bbr.flex8 = event;
3111 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3112 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3113 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3114 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3115 		log.u_bbr.pacing_gain = rack->r_must_retran;
3116 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3117 		log.u_bbr.bw_inuse <<= 32;
3118 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3119 		TCP_LOG_EVENTP(tp, NULL,
3120 		    &rack->rc_inp->inp_socket->so_rcv,
3121 		    &rack->rc_inp->inp_socket->so_snd,
3122 		    BBR_LOG_PROGRESS, 0,
3123 		    0, &log, false, &tv);
3124 	}
3125 }
3126 
3127 static void
3128 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
3129 {
3130 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3131 		union tcp_log_stackspecific log;
3132 
3133 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3134 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3135 		log.u_bbr.flex1 = slot;
3136 		if (rack->rack_no_prr)
3137 			log.u_bbr.flex2 = 0;
3138 		else
3139 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3140 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3141 		log.u_bbr.flex6 = line;
3142 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3143 		log.u_bbr.flex8 = rack->rc_in_persist;
3144 		log.u_bbr.timeStamp = cts;
3145 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3146 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3147 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3148 		log.u_bbr.pacing_gain = rack->r_must_retran;
3149 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3150 		    &rack->rc_inp->inp_socket->so_rcv,
3151 		    &rack->rc_inp->inp_socket->so_snd,
3152 		    BBR_LOG_BBRSND, 0,
3153 		    0, &log, false, tv);
3154 	}
3155 }
3156 
3157 static void
3158 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
3159 {
3160 	if (tcp_bblogging_on(rack->rc_tp)) {
3161 		union tcp_log_stackspecific log;
3162 		struct timeval tv;
3163 
3164 		memset(&log, 0, sizeof(log));
3165 		log.u_bbr.flex1 = did_out;
3166 		log.u_bbr.flex2 = nxt_pkt;
3167 		log.u_bbr.flex3 = way_out;
3168 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3169 		if (rack->rack_no_prr)
3170 			log.u_bbr.flex5 = 0;
3171 		else
3172 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3173 		log.u_bbr.flex6 = nsegs;
3174 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3175 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
3176 		log.u_bbr.flex7 <<= 1;
3177 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
3178 		log.u_bbr.flex7 <<= 1;
3179 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
3180 		log.u_bbr.flex8 = rack->rc_in_persist;
3181 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3182 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3183 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3184 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3185 		log.u_bbr.use_lt_bw <<= 1;
3186 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3187 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3188 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3189 		log.u_bbr.pacing_gain = rack->r_must_retran;
3190 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3191 		log.u_bbr.bw_inuse <<= 32;
3192 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3193 		log.u_bbr.epoch = rack->rc_inp->inp_socket->so_snd.sb_hiwat;
3194 		log.u_bbr.lt_epoch = rack->rc_inp->inp_socket->so_rcv.sb_hiwat;
3195 		log.u_bbr.lost = rack->rc_tp->t_srtt;
3196 		log.u_bbr.pkt_epoch = rack->rc_tp->rfbuf_cnt;
3197 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3198 		    &rack->rc_inp->inp_socket->so_rcv,
3199 		    &rack->rc_inp->inp_socket->so_snd,
3200 		    BBR_LOG_DOSEG_DONE, 0,
3201 		    0, &log, false, &tv);
3202 	}
3203 }
3204 
3205 static void
3206 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
3207 {
3208 	if (tcp_bblogging_on(rack->rc_tp)) {
3209 		union tcp_log_stackspecific log;
3210 		struct timeval tv;
3211 
3212 		memset(&log, 0, sizeof(log));
3213 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3214 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3215 		log.u_bbr.flex4 = arg1;
3216 		log.u_bbr.flex5 = arg2;
3217 		log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3218 		log.u_bbr.flex6 = arg3;
3219 		log.u_bbr.flex8 = frm;
3220 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3221 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3222 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3223 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3224 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3225 		log.u_bbr.pacing_gain = rack->r_must_retran;
3226 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3227 		    &tptosocket(tp)->so_snd,
3228 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3229 	}
3230 }
3231 
3232 static void
3233 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3234 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3235 {
3236 	if (tcp_bblogging_on(rack->rc_tp)) {
3237 		union tcp_log_stackspecific log;
3238 		struct timeval tv;
3239 
3240 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3241 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3242 		log.u_bbr.flex1 = slot;
3243 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3244 		log.u_bbr.flex4 = reason;
3245 		if (rack->rack_no_prr)
3246 			log.u_bbr.flex5 = 0;
3247 		else
3248 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3249 		log.u_bbr.flex7 = hpts_calling;
3250 		log.u_bbr.flex8 = rack->rc_in_persist;
3251 		log.u_bbr.lt_epoch = cwnd_to_use;
3252 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3253 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3254 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3255 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3256 		log.u_bbr.pacing_gain = rack->r_must_retran;
3257 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3258 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3259 		log.u_bbr.bw_inuse <<= 32;
3260 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3261 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3262 		    &rack->rc_inp->inp_socket->so_rcv,
3263 		    &rack->rc_inp->inp_socket->so_snd,
3264 		    BBR_LOG_JUSTRET, 0,
3265 		    tlen, &log, false, &tv);
3266 	}
3267 }
3268 
3269 static void
3270 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3271 		   struct timeval *tv, uint32_t flags_on_entry)
3272 {
3273 	if (tcp_bblogging_on(rack->rc_tp)) {
3274 		union tcp_log_stackspecific log;
3275 
3276 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3277 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3278 		log.u_bbr.flex1 = line;
3279 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3280 		log.u_bbr.flex3 = flags_on_entry;
3281 		log.u_bbr.flex4 = us_cts;
3282 		if (rack->rack_no_prr)
3283 			log.u_bbr.flex5 = 0;
3284 		else
3285 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3286 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3287 		log.u_bbr.flex7 = hpts_removed;
3288 		log.u_bbr.flex8 = 1;
3289 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3290 		log.u_bbr.timeStamp = us_cts;
3291 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3292 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3293 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3294 		log.u_bbr.pacing_gain = rack->r_must_retran;
3295 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3296 		log.u_bbr.bw_inuse <<= 32;
3297 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3298 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3299 		    &rack->rc_inp->inp_socket->so_rcv,
3300 		    &rack->rc_inp->inp_socket->so_snd,
3301 		    BBR_LOG_TIMERCANC, 0,
3302 		    0, &log, false, tv);
3303 	}
3304 }
3305 
3306 static void
3307 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3308 			  uint32_t flex1, uint32_t flex2,
3309 			  uint32_t flex3, uint32_t flex4,
3310 			  uint32_t flex5, uint32_t flex6,
3311 			  uint16_t flex7, uint8_t mod)
3312 {
3313 	if (tcp_bblogging_on(rack->rc_tp)) {
3314 		union tcp_log_stackspecific log;
3315 		struct timeval tv;
3316 
3317 		if (mod == 1) {
3318 			/* No you can't use 1, its for the real to cancel */
3319 			return;
3320 		}
3321 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3322 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3323 		log.u_bbr.flex1 = flex1;
3324 		log.u_bbr.flex2 = flex2;
3325 		log.u_bbr.flex3 = flex3;
3326 		log.u_bbr.flex4 = flex4;
3327 		log.u_bbr.flex5 = flex5;
3328 		log.u_bbr.flex6 = flex6;
3329 		log.u_bbr.flex7 = flex7;
3330 		log.u_bbr.flex8 = mod;
3331 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3332 		    &rack->rc_inp->inp_socket->so_rcv,
3333 		    &rack->rc_inp->inp_socket->so_snd,
3334 		    BBR_LOG_TIMERCANC, 0,
3335 		    0, &log, false, &tv);
3336 	}
3337 }
3338 
3339 static void
3340 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3341 {
3342 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3343 		union tcp_log_stackspecific log;
3344 		struct timeval tv;
3345 
3346 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3347 		log.u_bbr.flex1 = timers;
3348 		log.u_bbr.flex2 = ret;
3349 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3350 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3351 		log.u_bbr.flex5 = cts;
3352 		if (rack->rack_no_prr)
3353 			log.u_bbr.flex6 = 0;
3354 		else
3355 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3356 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3357 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3358 		log.u_bbr.pacing_gain = rack->r_must_retran;
3359 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3360 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3361 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3362 		    &rack->rc_inp->inp_socket->so_rcv,
3363 		    &rack->rc_inp->inp_socket->so_snd,
3364 		    BBR_LOG_TO_PROCESS, 0,
3365 		    0, &log, false, &tv);
3366 	}
3367 }
3368 
3369 static void
3370 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3371 {
3372 	if (tcp_bblogging_on(rack->rc_tp)) {
3373 		union tcp_log_stackspecific log;
3374 		struct timeval tv;
3375 
3376 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3377 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3378 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3379 		if (rack->rack_no_prr)
3380 			log.u_bbr.flex3 = 0;
3381 		else
3382 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3383 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3384 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3385 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3386 		log.u_bbr.flex7 = line;
3387 		log.u_bbr.flex8 = frm;
3388 		log.u_bbr.pkts_out = orig_cwnd;
3389 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3390 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3391 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3392 		log.u_bbr.use_lt_bw <<= 1;
3393 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3394 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3395 		    &rack->rc_inp->inp_socket->so_rcv,
3396 		    &rack->rc_inp->inp_socket->so_snd,
3397 		    BBR_LOG_BBRUPD, 0,
3398 		    0, &log, false, &tv);
3399 	}
3400 }
3401 
3402 static void
3403 rack_counter_destroy(void)
3404 {
3405 	counter_u64_free(rack_total_bytes);
3406 	counter_u64_free(rack_fto_send);
3407 	counter_u64_free(rack_fto_rsm_send);
3408 	counter_u64_free(rack_nfto_resend);
3409 	counter_u64_free(rack_hw_pace_init_fail);
3410 	counter_u64_free(rack_hw_pace_lost);
3411 	counter_u64_free(rack_non_fto_send);
3412 	counter_u64_free(rack_extended_rfo);
3413 	counter_u64_free(rack_ack_total);
3414 	counter_u64_free(rack_express_sack);
3415 	counter_u64_free(rack_sack_total);
3416 	counter_u64_free(rack_move_none);
3417 	counter_u64_free(rack_move_some);
3418 	counter_u64_free(rack_sack_attacks_detected);
3419 	counter_u64_free(rack_sack_attacks_reversed);
3420 	counter_u64_free(rack_sack_attacks_suspect);
3421 	counter_u64_free(rack_sack_used_next_merge);
3422 	counter_u64_free(rack_sack_used_prev_merge);
3423 	counter_u64_free(rack_tlp_tot);
3424 	counter_u64_free(rack_tlp_newdata);
3425 	counter_u64_free(rack_tlp_retran);
3426 	counter_u64_free(rack_tlp_retran_bytes);
3427 	counter_u64_free(rack_to_tot);
3428 	counter_u64_free(rack_saw_enobuf);
3429 	counter_u64_free(rack_saw_enobuf_hw);
3430 	counter_u64_free(rack_saw_enetunreach);
3431 	counter_u64_free(rack_hot_alloc);
3432 	counter_u64_free(tcp_policer_detected);
3433 	counter_u64_free(rack_to_alloc);
3434 	counter_u64_free(rack_to_alloc_hard);
3435 	counter_u64_free(rack_to_alloc_emerg);
3436 	counter_u64_free(rack_to_alloc_limited);
3437 	counter_u64_free(rack_alloc_limited_conns);
3438 	counter_u64_free(rack_split_limited);
3439 	counter_u64_free(rack_multi_single_eq);
3440 	counter_u64_free(rack_rxt_clamps_cwnd);
3441 	counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3442 	counter_u64_free(rack_proc_non_comp_ack);
3443 	counter_u64_free(rack_sack_proc_all);
3444 	counter_u64_free(rack_sack_proc_restart);
3445 	counter_u64_free(rack_sack_proc_short);
3446 	counter_u64_free(rack_sack_skipped_acked);
3447 	counter_u64_free(rack_sack_splits);
3448 	counter_u64_free(rack_input_idle_reduces);
3449 	counter_u64_free(rack_collapsed_win);
3450 	counter_u64_free(rack_collapsed_win_rxt);
3451 	counter_u64_free(rack_collapsed_win_rxt_bytes);
3452 	counter_u64_free(rack_collapsed_win_seen);
3453 	counter_u64_free(rack_try_scwnd);
3454 	counter_u64_free(rack_persists_sends);
3455 	counter_u64_free(rack_persists_acks);
3456 	counter_u64_free(rack_persists_loss);
3457 	counter_u64_free(rack_persists_lost_ends);
3458 #ifdef INVARIANTS
3459 	counter_u64_free(rack_adjust_map_bw);
3460 #endif
3461 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3462 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3463 }
3464 
3465 static struct rack_sendmap *
3466 rack_alloc(struct tcp_rack *rack)
3467 {
3468 	struct rack_sendmap *rsm;
3469 
3470 	/*
3471 	 * First get the top of the list it in
3472 	 * theory is the "hottest" rsm we have,
3473 	 * possibly just freed by ack processing.
3474 	 */
3475 	if (rack->rc_free_cnt > rack_free_cache) {
3476 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3477 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3478 		counter_u64_add(rack_hot_alloc, 1);
3479 		rack->rc_free_cnt--;
3480 		return (rsm);
3481 	}
3482 	/*
3483 	 * Once we get under our free cache we probably
3484 	 * no longer have a "hot" one available. Lets
3485 	 * get one from UMA.
3486 	 */
3487 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
3488 	if (rsm) {
3489 		rack->r_ctl.rc_num_maps_alloced++;
3490 		counter_u64_add(rack_to_alloc, 1);
3491 		return (rsm);
3492 	}
3493 	/*
3494 	 * Dig in to our aux rsm's (the last two) since
3495 	 * UMA failed to get us one.
3496 	 */
3497 	if (rack->rc_free_cnt) {
3498 		counter_u64_add(rack_to_alloc_emerg, 1);
3499 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3500 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3501 		rack->rc_free_cnt--;
3502 		return (rsm);
3503 	}
3504 	return (NULL);
3505 }
3506 
3507 static struct rack_sendmap *
3508 rack_alloc_full_limit(struct tcp_rack *rack)
3509 {
3510 	if ((V_tcp_map_entries_limit > 0) &&
3511 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3512 		counter_u64_add(rack_to_alloc_limited, 1);
3513 		if (!rack->alloc_limit_reported) {
3514 			rack->alloc_limit_reported = 1;
3515 			counter_u64_add(rack_alloc_limited_conns, 1);
3516 		}
3517 		return (NULL);
3518 	}
3519 	return (rack_alloc(rack));
3520 }
3521 
3522 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3523 static struct rack_sendmap *
3524 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3525 {
3526 	struct rack_sendmap *rsm;
3527 
3528 	if (limit_type) {
3529 		/* currently there is only one limit type */
3530 		if (rack->r_ctl.rc_split_limit > 0 &&
3531 		    rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3532 			counter_u64_add(rack_split_limited, 1);
3533 			if (!rack->alloc_limit_reported) {
3534 				rack->alloc_limit_reported = 1;
3535 				counter_u64_add(rack_alloc_limited_conns, 1);
3536 			}
3537 			return (NULL);
3538 		}
3539 	}
3540 
3541 	/* allocate and mark in the limit type, if set */
3542 	rsm = rack_alloc(rack);
3543 	if (rsm != NULL && limit_type) {
3544 		rsm->r_limit_type = limit_type;
3545 		rack->r_ctl.rc_num_split_allocs++;
3546 	}
3547 	return (rsm);
3548 }
3549 
3550 static void
3551 rack_free_trim(struct tcp_rack *rack)
3552 {
3553 	struct rack_sendmap *rsm;
3554 
3555 	/*
3556 	 * Free up all the tail entries until
3557 	 * we get our list down to the limit.
3558 	 */
3559 	while (rack->rc_free_cnt > rack_free_cache) {
3560 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3561 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3562 		rack->rc_free_cnt--;
3563 		rack->r_ctl.rc_num_maps_alloced--;
3564 		uma_zfree(rack_zone, rsm);
3565 	}
3566 }
3567 
3568 static void
3569 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3570 {
3571 	if (rsm->r_flags & RACK_APP_LIMITED) {
3572 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
3573 			rack->r_ctl.rc_app_limited_cnt--;
3574 		}
3575 	}
3576 	if (rsm->r_limit_type) {
3577 		/* currently there is only one limit type */
3578 		rack->r_ctl.rc_num_split_allocs--;
3579 	}
3580 	if (rsm == rack->r_ctl.rc_first_appl) {
3581 		rack->r_ctl.cleared_app_ack_seq = rsm->r_start + (rsm->r_end - rsm->r_start);
3582 		rack->r_ctl.cleared_app_ack = 1;
3583 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3584 			rack->r_ctl.rc_first_appl = NULL;
3585 		else
3586 			rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3587 	}
3588 	if (rsm == rack->r_ctl.rc_resend)
3589 		rack->r_ctl.rc_resend = NULL;
3590 	if (rsm == rack->r_ctl.rc_end_appl)
3591 		rack->r_ctl.rc_end_appl = NULL;
3592 	if (rack->r_ctl.rc_tlpsend == rsm)
3593 		rack->r_ctl.rc_tlpsend = NULL;
3594 	if (rack->r_ctl.rc_sacklast == rsm)
3595 		rack->r_ctl.rc_sacklast = NULL;
3596 	memset(rsm, 0, sizeof(struct rack_sendmap));
3597 	/* Make sure we are not going to overrun our count limit of 0xff */
3598 	if ((rack->rc_free_cnt + 1) > RACK_FREE_CNT_MAX) {
3599 		rack_free_trim(rack);
3600 	}
3601 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3602 	rack->rc_free_cnt++;
3603 }
3604 
3605 static uint32_t
3606 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3607 {
3608 	uint64_t srtt, bw, len, tim;
3609 	uint32_t segsiz, def_len, minl;
3610 
3611 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3612 	def_len = rack_def_data_window * segsiz;
3613 	if (rack->rc_gp_filled == 0) {
3614 		/*
3615 		 * We have no measurement (IW is in flight?) so
3616 		 * we can only guess using our data_window sysctl
3617 		 * value (usually 20MSS).
3618 		 */
3619 		return (def_len);
3620 	}
3621 	/*
3622 	 * Now we have a number of factors to consider.
3623 	 *
3624 	 * 1) We have a desired BDP which is usually
3625 	 *    at least 2.
3626 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3627 	 *    but we allow it too to be more.
3628 	 * 3) We want to make sure a measurement last N useconds (if
3629 	 *    we have set rack_min_measure_usec.
3630 	 *
3631 	 * We handle the first concern here by trying to create a data
3632 	 * window of max(rack_def_data_window, DesiredBDP). The
3633 	 * second concern we handle in not letting the measurement
3634 	 * window end normally until at least the required SRTT's
3635 	 * have gone by which is done further below in
3636 	 * rack_enough_for_measurement(). Finally the third concern
3637 	 * we also handle here by calculating how long that time
3638 	 * would take at the current BW and then return the
3639 	 * max of our first calculation and that length. Note
3640 	 * that if rack_min_measure_usec is 0, we don't deal
3641 	 * with concern 3. Also for both Concern 1 and 3 an
3642 	 * application limited period could end the measurement
3643 	 * earlier.
3644 	 *
3645 	 * So lets calculate the BDP with the "known" b/w using
3646 	 * the SRTT has our rtt and then multiply it by the
3647 	 * goal.
3648 	 */
3649 	bw = rack_get_bw(rack);
3650 	srtt = (uint64_t)tp->t_srtt;
3651 	len = bw * srtt;
3652 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3653 	len *= max(1, rack_goal_bdp);
3654 	/* Now we need to round up to the nearest MSS */
3655 	len = roundup(len, segsiz);
3656 	if (rack_min_measure_usec) {
3657 		/* Now calculate our min length for this b/w */
3658 		tim = rack_min_measure_usec;
3659 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3660 		if (minl == 0)
3661 			minl = 1;
3662 		minl = roundup(minl, segsiz);
3663 		if (len < minl)
3664 			len = minl;
3665 	}
3666 	/*
3667 	 * Now if we have a very small window we want
3668 	 * to attempt to get the window that is
3669 	 * as small as possible. This happens on
3670 	 * low b/w connections and we don't want to
3671 	 * span huge numbers of rtt's between measurements.
3672 	 *
3673 	 * We basically include 2 over our "MIN window" so
3674 	 * that the measurement can be shortened (possibly) by
3675 	 * an ack'ed packet.
3676 	 */
3677 	if (len < def_len)
3678 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3679 	else
3680 		return (max((uint32_t)len, def_len));
3681 
3682 }
3683 
3684 static int
3685 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3686 {
3687 	uint32_t tim, srtts, segsiz;
3688 
3689 	/*
3690 	 * Has enough time passed for the GP measurement to be valid?
3691 	 */
3692 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3693 		/* Not enough bytes yet */
3694 		return (0);
3695 	}
3696 	if ((tp->snd_max == tp->snd_una) ||
3697 	    (th_ack == tp->snd_max)){
3698 		/*
3699 		 * All is acked quality of all acked is
3700 		 * usually low or medium, but we in theory could split
3701 		 * all acked into two cases, where you got
3702 		 * a signifigant amount of your window and
3703 		 * where you did not. For now we leave it
3704 		 * but it is something to contemplate in the
3705 		 * future. The danger here is that delayed ack
3706 		 * is effecting the last byte (which is a 50:50 chance).
3707 		 */
3708 		*quality = RACK_QUALITY_ALLACKED;
3709 		return (1);
3710 	}
3711 	if (SEQ_GEQ(th_ack,  tp->gput_ack)) {
3712 		/*
3713 		 * We obtained our entire window of data we wanted
3714 		 * no matter if we are in recovery or not then
3715 		 * its ok since expanding the window does not
3716 		 * make things fuzzy (or at least not as much).
3717 		 */
3718 		*quality = RACK_QUALITY_HIGH;
3719 		return (1);
3720 	}
3721 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3722 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3723 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3724 		/* Not enough bytes yet */
3725 		return (0);
3726 	}
3727 	if (rack->r_ctl.rc_first_appl &&
3728 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3729 		/*
3730 		 * We are up to the app limited send point
3731 		 * we have to measure irrespective of the time..
3732 		 */
3733 		*quality = RACK_QUALITY_APPLIMITED;
3734 		return (1);
3735 	}
3736 	/* Now what about time? */
3737 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3738 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3739 	if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3740 		/*
3741 		 * We do not allow a measurement if we are in recovery
3742 		 * that would shrink the goodput window we wanted.
3743 		 * This is to prevent cloudyness of when the last send
3744 		 * was actually made.
3745 		 */
3746 		*quality = RACK_QUALITY_HIGH;
3747 		return (1);
3748 	}
3749 	/* Nope not even a full SRTT has passed */
3750 	return (0);
3751 }
3752 
3753 static void
3754 rack_log_timely(struct tcp_rack *rack,
3755 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3756 		uint64_t up_bnd, int line, uint8_t method)
3757 {
3758 	if (tcp_bblogging_on(rack->rc_tp)) {
3759 		union tcp_log_stackspecific log;
3760 		struct timeval tv;
3761 
3762 		memset(&log, 0, sizeof(log));
3763 		log.u_bbr.flex1 = logged;
3764 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3765 		log.u_bbr.flex2 <<= 4;
3766 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3767 		log.u_bbr.flex2 <<= 4;
3768 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3769 		log.u_bbr.flex2 <<= 4;
3770 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3771 		log.u_bbr.flex3 = rack->rc_gp_incr;
3772 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3773 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3774 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3775 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3776 		log.u_bbr.flex8 = method;
3777 		log.u_bbr.cur_del_rate = cur_bw;
3778 		log.u_bbr.delRate = low_bnd;
3779 		log.u_bbr.bw_inuse = up_bnd;
3780 		log.u_bbr.rttProp = rack_get_bw(rack);
3781 		log.u_bbr.pkt_epoch = line;
3782 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3783 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3784 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3785 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3786 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3787 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3788 		log.u_bbr.cwnd_gain <<= 1;
3789 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3790 		log.u_bbr.cwnd_gain <<= 1;
3791 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3792 		log.u_bbr.cwnd_gain <<= 1;
3793 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3794 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3795 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3796 		    &rack->rc_inp->inp_socket->so_rcv,
3797 		    &rack->rc_inp->inp_socket->so_snd,
3798 		    TCP_TIMELY_WORK, 0,
3799 		    0, &log, false, &tv);
3800 	}
3801 }
3802 
3803 static int
3804 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3805 {
3806 	/*
3807 	 * Before we increase we need to know if
3808 	 * the estimate just made was less than
3809 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3810 	 *
3811 	 * If we already are pacing at a fast enough
3812 	 * rate to push us faster there is no sense of
3813 	 * increasing.
3814 	 *
3815 	 * We first caculate our actual pacing rate (ss or ca multiplier
3816 	 * times our cur_bw).
3817 	 *
3818 	 * Then we take the last measured rate and multipy by our
3819 	 * maximum pacing overage to give us a max allowable rate.
3820 	 *
3821 	 * If our act_rate is smaller than our max_allowable rate
3822 	 * then we should increase. Else we should hold steady.
3823 	 *
3824 	 */
3825 	uint64_t act_rate, max_allow_rate;
3826 
3827 	if (rack_timely_no_stopping)
3828 		return (1);
3829 
3830 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3831 		/*
3832 		 * Initial startup case or
3833 		 * everything is acked case.
3834 		 */
3835 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3836 				__LINE__, 9);
3837 		return (1);
3838 	}
3839 	if (mult <= 100) {
3840 		/*
3841 		 * We can always pace at or slightly above our rate.
3842 		 */
3843 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3844 				__LINE__, 9);
3845 		return (1);
3846 	}
3847 	act_rate = cur_bw * (uint64_t)mult;
3848 	act_rate /= 100;
3849 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3850 	max_allow_rate /= 100;
3851 	if (act_rate < max_allow_rate) {
3852 		/*
3853 		 * Here the rate we are actually pacing at
3854 		 * is smaller than 10% above our last measurement.
3855 		 * This means we are pacing below what we would
3856 		 * like to try to achieve (plus some wiggle room).
3857 		 */
3858 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3859 				__LINE__, 9);
3860 		return (1);
3861 	} else {
3862 		/*
3863 		 * Here we are already pacing at least rack_max_per_above(10%)
3864 		 * what we are getting back. This indicates most likely
3865 		 * that we are being limited (cwnd/rwnd/app) and can't
3866 		 * get any more b/w. There is no sense of trying to
3867 		 * raise up the pacing rate its not speeding us up
3868 		 * and we already are pacing faster than we are getting.
3869 		 */
3870 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3871 				__LINE__, 8);
3872 		return (0);
3873 	}
3874 }
3875 
3876 static void
3877 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3878 {
3879 	/*
3880 	 * When we drag bottom, we want to assure
3881 	 * that no multiplier is below 1.0, if so
3882 	 * we want to restore it to at least that.
3883 	 */
3884 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3885 		/* This is unlikely we usually do not touch recovery */
3886 		rack->r_ctl.rack_per_of_gp_rec = 100;
3887 	}
3888 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3889 		rack->r_ctl.rack_per_of_gp_ca = 100;
3890 	}
3891 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3892 		rack->r_ctl.rack_per_of_gp_ss = 100;
3893 	}
3894 }
3895 
3896 static void
3897 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3898 {
3899 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3900 		rack->r_ctl.rack_per_of_gp_ca = 100;
3901 	}
3902 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3903 		rack->r_ctl.rack_per_of_gp_ss = 100;
3904 	}
3905 }
3906 
3907 static void
3908 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3909 {
3910 	int32_t  calc, logged, plus;
3911 
3912 	logged = 0;
3913 
3914 	if (rack->rc_skip_timely)
3915 		return;
3916 	if (override) {
3917 		/*
3918 		 * override is passed when we are
3919 		 * loosing b/w and making one last
3920 		 * gasp at trying to not loose out
3921 		 * to a new-reno flow.
3922 		 */
3923 		goto extra_boost;
3924 	}
3925 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3926 	if (rack->rc_gp_incr &&
3927 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3928 		/*
3929 		 * Reset and get 5 strokes more before the boost. Note
3930 		 * that the count is 0 based so we have to add one.
3931 		 */
3932 extra_boost:
3933 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3934 		rack->rc_gp_timely_inc_cnt = 0;
3935 	} else
3936 		plus = (uint32_t)rack_gp_increase_per;
3937 	/* Must be at least 1% increase for true timely increases */
3938 	if ((plus < 1) &&
3939 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3940 		plus = 1;
3941 	if (rack->rc_gp_saw_rec &&
3942 	    (rack->rc_gp_no_rec_chg == 0) &&
3943 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3944 				  rack->r_ctl.rack_per_of_gp_rec)) {
3945 		/* We have been in recovery ding it too */
3946 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3947 		if (calc > 0xffff)
3948 			calc = 0xffff;
3949 		logged |= 1;
3950 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3951 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3952 		    (rack->rc_dragged_bottom == 0) &&
3953 		    (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3954 			rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3955 	}
3956 	if (rack->rc_gp_saw_ca &&
3957 	    (rack->rc_gp_saw_ss == 0) &&
3958 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3959 				  rack->r_ctl.rack_per_of_gp_ca)) {
3960 		/* In CA */
3961 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3962 		if (calc > 0xffff)
3963 			calc = 0xffff;
3964 		logged |= 2;
3965 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3966 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3967 		    (rack->rc_dragged_bottom == 0) &&
3968 		    (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3969 			rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3970 	}
3971 	if (rack->rc_gp_saw_ss &&
3972 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3973 				  rack->r_ctl.rack_per_of_gp_ss)) {
3974 		/* In SS */
3975 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3976 		if (calc > 0xffff)
3977 			calc = 0xffff;
3978 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3979 		if (rack->r_ctl.rack_per_upper_bound_ss &&
3980 		    (rack->rc_dragged_bottom == 0) &&
3981 		    (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3982 			rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3983 		logged |= 4;
3984 	}
3985 	if (logged &&
3986 	    (rack->rc_gp_incr == 0)){
3987 		/* Go into increment mode */
3988 		rack->rc_gp_incr = 1;
3989 		rack->rc_gp_timely_inc_cnt = 0;
3990 	}
3991 	if (rack->rc_gp_incr &&
3992 	    logged &&
3993 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3994 		rack->rc_gp_timely_inc_cnt++;
3995 	}
3996 	rack_log_timely(rack,  logged, plus, 0, 0,
3997 			__LINE__, 1);
3998 }
3999 
4000 static uint32_t
4001 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
4002 {
4003 	/*-
4004 	 * norm_grad = rtt_diff / minrtt;
4005 	 * new_per = curper * (1 - B * norm_grad)
4006 	 *
4007 	 * B = rack_gp_decrease_per (default 80%)
4008 	 * rtt_dif = input var current rtt-diff
4009 	 * curper = input var current percentage
4010 	 * minrtt = from rack filter
4011 	 *
4012 	 * In order to do the floating point calculations above we
4013 	 * do an integer conversion. The code looks confusing so let me
4014 	 * translate it into something that use more variables and
4015 	 * is clearer for us humans :)
4016 	 *
4017 	 * uint64_t norm_grad, inverse, reduce_by, final_result;
4018 	 * uint32_t perf;
4019 	 *
4020 	 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
4021 	 *             (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
4022 	 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
4023 	 * inverse /= 1000000;
4024 	 * reduce_by = (1000000 - inverse);
4025 	 * final_result = (cur_per * reduce_by) / 1000000;
4026 	 * perf = (uint32_t)final_result;
4027 	 */
4028 	uint64_t perf;
4029 
4030 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
4031 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
4032 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
4033 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
4034 		     (uint64_t)1000000)) /
4035 		(uint64_t)1000000);
4036 	if (perf > curper) {
4037 		/* TSNH */
4038 		perf = curper - 1;
4039 	}
4040 	return ((uint32_t)perf);
4041 }
4042 
4043 static uint32_t
4044 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
4045 {
4046 	/*
4047 	 *                                   highrttthresh
4048 	 * result = curper * (1 - (B * ( 1 -  ------          ))
4049 	 *                                     gp_srtt
4050 	 *
4051 	 * B = rack_gp_decrease_per (default .8 i.e. 80)
4052 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
4053 	 */
4054 	uint64_t perf;
4055 	uint32_t highrttthresh;
4056 
4057 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4058 
4059 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
4060 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
4061 					((uint64_t)highrttthresh * (uint64_t)1000000) /
4062 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
4063 	if (tcp_bblogging_on(rack->rc_tp)) {
4064 		uint64_t log1;
4065 
4066 		log1 = rtt;
4067 		log1 <<= 32;
4068 		log1 |= highrttthresh;
4069 		rack_log_timely(rack,
4070 				rack_gp_decrease_per,
4071 				(uint64_t)curper,
4072 				log1,
4073 				perf,
4074 				__LINE__,
4075 				15);
4076 	}
4077 	return (perf);
4078 }
4079 
4080 static void
4081 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
4082 {
4083 	uint64_t logvar, logvar2, logvar3;
4084 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
4085 
4086 	if (rack->rc_skip_timely)
4087 		return;
4088 	if (rack->rc_gp_incr) {
4089 		/* Turn off increment counting */
4090 		rack->rc_gp_incr = 0;
4091 		rack->rc_gp_timely_inc_cnt = 0;
4092 	}
4093 	ss_red = ca_red = rec_red = 0;
4094 	logged = 0;
4095 	/* Calculate the reduction value */
4096 	if (rtt_diff < 0) {
4097 		rtt_diff *= -1;
4098 	}
4099 	/* Must be at least 1% reduction */
4100 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
4101 		/* We have been in recovery ding it too */
4102 		if (timely_says == 2) {
4103 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
4104 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4105 			if (alt < new_per)
4106 				val = alt;
4107 			else
4108 				val = new_per;
4109 		} else
4110 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4111 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
4112 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4113 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4114 		} else {
4115 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4116 			rec_red = 0;
4117 		}
4118 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4119 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4120 		logged |= 1;
4121 	}
4122 	if (rack->rc_gp_saw_ss) {
4123 		/* Sent in SS */
4124 		if (timely_says == 2) {
4125 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4126 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4127 			if (alt < new_per)
4128 				val = alt;
4129 			else
4130 				val = new_per;
4131 		} else
4132 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4133 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4134 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4135 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4136 		} else {
4137 			ss_red = new_per;
4138 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4139 			logvar = new_per;
4140 			logvar <<= 32;
4141 			logvar |= alt;
4142 			logvar2 = (uint32_t)rtt;
4143 			logvar2 <<= 32;
4144 			logvar2 |= (uint32_t)rtt_diff;
4145 			logvar3 = rack_gp_rtt_maxmul;
4146 			logvar3 <<= 32;
4147 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4148 			rack_log_timely(rack, timely_says,
4149 					logvar2, logvar3,
4150 					logvar, __LINE__, 10);
4151 		}
4152 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4153 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4154 		logged |= 4;
4155 	} else if (rack->rc_gp_saw_ca) {
4156 		/* Sent in CA */
4157 		if (timely_says == 2) {
4158 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4159 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4160 			if (alt < new_per)
4161 				val = alt;
4162 			else
4163 				val = new_per;
4164 		} else
4165 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4166 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
4167 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4168 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4169 		} else {
4170 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4171 			ca_red = 0;
4172 			logvar = new_per;
4173 			logvar <<= 32;
4174 			logvar |= alt;
4175 			logvar2 = (uint32_t)rtt;
4176 			logvar2 <<= 32;
4177 			logvar2 |= (uint32_t)rtt_diff;
4178 			logvar3 = rack_gp_rtt_maxmul;
4179 			logvar3 <<= 32;
4180 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4181 			rack_log_timely(rack, timely_says,
4182 					logvar2, logvar3,
4183 					logvar, __LINE__, 10);
4184 		}
4185 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4186 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4187 		logged |= 2;
4188 	}
4189 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
4190 		rack->rc_gp_timely_dec_cnt++;
4191 		if (rack_timely_dec_clear &&
4192 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4193 			rack->rc_gp_timely_dec_cnt = 0;
4194 	}
4195 	logvar = ss_red;
4196 	logvar <<= 32;
4197 	logvar |= ca_red;
4198 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
4199 			__LINE__, 2);
4200 }
4201 
4202 static void
4203 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4204 		     uint32_t rtt, uint32_t line, uint8_t reas)
4205 {
4206 	if (tcp_bblogging_on(rack->rc_tp)) {
4207 		union tcp_log_stackspecific log;
4208 		struct timeval tv;
4209 
4210 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4211 		log.u_bbr.flex1 = line;
4212 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4213 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4214 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4215 		log.u_bbr.flex5 = rtt;
4216 		log.u_bbr.flex6 = rack->rc_highly_buffered;
4217 		log.u_bbr.flex6 <<= 1;
4218 		log.u_bbr.flex6 |= rack->forced_ack;
4219 		log.u_bbr.flex6 <<= 1;
4220 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4221 		log.u_bbr.flex6 <<= 1;
4222 		log.u_bbr.flex6 |= rack->in_probe_rtt;
4223 		log.u_bbr.flex6 <<= 1;
4224 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4225 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4226 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4227 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4228 		log.u_bbr.flex8 = reas;
4229 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4230 		log.u_bbr.delRate = rack_get_bw(rack);
4231 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4232 		log.u_bbr.cur_del_rate <<= 32;
4233 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4234 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4235 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4236 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4237 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4238 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4239 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4240 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4241 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4242 		log.u_bbr.rttProp = us_cts;
4243 		log.u_bbr.rttProp <<= 32;
4244 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4245 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4246 		    &rack->rc_inp->inp_socket->so_rcv,
4247 		    &rack->rc_inp->inp_socket->so_snd,
4248 		    BBR_LOG_RTT_SHRINKS, 0,
4249 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4250 	}
4251 }
4252 
4253 static void
4254 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4255 {
4256 	uint64_t bwdp;
4257 
4258 	bwdp = rack_get_bw(rack);
4259 	bwdp *= (uint64_t)rtt;
4260 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4261 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4262 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4263 		/*
4264 		 * A window protocol must be able to have 4 packets
4265 		 * outstanding as the floor in order to function
4266 		 * (especially considering delayed ack :D).
4267 		 */
4268 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4269 	}
4270 }
4271 
4272 static void
4273 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4274 {
4275 	/**
4276 	 * ProbeRTT is a bit different in rack_pacing than in
4277 	 * BBR. It is like BBR in that it uses the lowering of
4278 	 * the RTT as a signal that we saw something new and
4279 	 * counts from there for how long between. But it is
4280 	 * different in that its quite simple. It does not
4281 	 * play with the cwnd and wait until we get down
4282 	 * to N segments outstanding and hold that for
4283 	 * 200ms. Instead it just sets the pacing reduction
4284 	 * rate to a set percentage (70 by default) and hold
4285 	 * that for a number of recent GP Srtt's.
4286 	 */
4287 	uint32_t segsiz;
4288 
4289 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4290 	if (rack->rc_gp_dyn_mul == 0)
4291 		return;
4292 
4293 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4294 		/* We are idle */
4295 		return;
4296 	}
4297 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4298 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4299 		/*
4300 		 * Stop the goodput now, the idea here is
4301 		 * that future measurements with in_probe_rtt
4302 		 * won't register if they are not greater so
4303 		 * we want to get what info (if any) is available
4304 		 * now.
4305 		 */
4306 		rack_do_goodput_measurement(rack->rc_tp, rack,
4307 					    rack->rc_tp->snd_una, __LINE__,
4308 					    RACK_QUALITY_PROBERTT);
4309 	}
4310 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4311 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4312 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4313 		     rack->r_ctl.rc_pace_min_segs);
4314 	rack->in_probe_rtt = 1;
4315 	rack->measure_saw_probe_rtt = 1;
4316 	rack->r_ctl.rc_time_probertt_starts = 0;
4317 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4318 	if (rack_probertt_use_min_rtt_entry)
4319 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4320 	else
4321 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4322 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4323 			     __LINE__, RACK_RTTS_ENTERPROBE);
4324 }
4325 
4326 static void
4327 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4328 {
4329 	struct rack_sendmap *rsm;
4330 	uint32_t segsiz;
4331 
4332 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4333 		     rack->r_ctl.rc_pace_min_segs);
4334 	rack->in_probe_rtt = 0;
4335 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4336 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4337 		/*
4338 		 * Stop the goodput now, the idea here is
4339 		 * that future measurements with in_probe_rtt
4340 		 * won't register if they are not greater so
4341 		 * we want to get what info (if any) is available
4342 		 * now.
4343 		 */
4344 		rack_do_goodput_measurement(rack->rc_tp, rack,
4345 					    rack->rc_tp->snd_una, __LINE__,
4346 					    RACK_QUALITY_PROBERTT);
4347 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4348 		/*
4349 		 * We don't have enough data to make a measurement.
4350 		 * So lets just stop and start here after exiting
4351 		 * probe-rtt. We probably are not interested in
4352 		 * the results anyway.
4353 		 */
4354 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4355 	}
4356 	/*
4357 	 * Measurements through the current snd_max are going
4358 	 * to be limited by the slower pacing rate.
4359 	 *
4360 	 * We need to mark these as app-limited so we
4361 	 * don't collapse the b/w.
4362 	 */
4363 	rsm = tqhash_max(rack->r_ctl.tqh);
4364 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4365 		if (rack->r_ctl.rc_app_limited_cnt == 0)
4366 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4367 		else {
4368 			/*
4369 			 * Go out to the end app limited and mark
4370 			 * this new one as next and move the end_appl up
4371 			 * to this guy.
4372 			 */
4373 			if (rack->r_ctl.rc_end_appl)
4374 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4375 			rack->r_ctl.rc_end_appl = rsm;
4376 		}
4377 		rsm->r_flags |= RACK_APP_LIMITED;
4378 		rack->r_ctl.rc_app_limited_cnt++;
4379 	}
4380 	/*
4381 	 * Now, we need to examine our pacing rate multipliers.
4382 	 * If its under 100%, we need to kick it back up to
4383 	 * 100%. We also don't let it be over our "max" above
4384 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4385 	 * Note setting clamp_atexit_prtt to 0 has the effect
4386 	 * of setting CA/SS to 100% always at exit (which is
4387 	 * the default behavior).
4388 	 */
4389 	if (rack_probertt_clear_is) {
4390 		rack->rc_gp_incr = 0;
4391 		rack->rc_gp_bwred = 0;
4392 		rack->rc_gp_timely_inc_cnt = 0;
4393 		rack->rc_gp_timely_dec_cnt = 0;
4394 	}
4395 	/* Do we do any clamping at exit? */
4396 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4397 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4398 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4399 	}
4400 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4401 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4402 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4403 	}
4404 	/*
4405 	 * Lets set rtt_diff to 0, so that we will get a "boost"
4406 	 * after exiting.
4407 	 */
4408 	rack->r_ctl.rc_rtt_diff = 0;
4409 
4410 	/* Clear all flags so we start fresh */
4411 	rack->rc_tp->t_bytes_acked = 0;
4412 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4413 	/*
4414 	 * If configured to, set the cwnd and ssthresh to
4415 	 * our targets.
4416 	 */
4417 	if (rack_probe_rtt_sets_cwnd) {
4418 		uint64_t ebdp;
4419 		uint32_t setto;
4420 
4421 		/* Set ssthresh so we get into CA once we hit our target */
4422 		if (rack_probertt_use_min_rtt_exit == 1) {
4423 			/* Set to min rtt */
4424 			rack_set_prtt_target(rack, segsiz,
4425 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4426 		} else if (rack_probertt_use_min_rtt_exit == 2) {
4427 			/* Set to current gp rtt */
4428 			rack_set_prtt_target(rack, segsiz,
4429 					     rack->r_ctl.rc_gp_srtt);
4430 		} else if (rack_probertt_use_min_rtt_exit == 3) {
4431 			/* Set to entry gp rtt */
4432 			rack_set_prtt_target(rack, segsiz,
4433 					     rack->r_ctl.rc_entry_gp_rtt);
4434 		} else {
4435 			uint64_t sum;
4436 			uint32_t setval;
4437 
4438 			sum = rack->r_ctl.rc_entry_gp_rtt;
4439 			sum *= 10;
4440 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4441 			if (sum >= 20) {
4442 				/*
4443 				 * A highly buffered path needs
4444 				 * cwnd space for timely to work.
4445 				 * Lets set things up as if
4446 				 * we are heading back here again.
4447 				 */
4448 				setval = rack->r_ctl.rc_entry_gp_rtt;
4449 			} else if (sum >= 15) {
4450 				/*
4451 				 * Lets take the smaller of the
4452 				 * two since we are just somewhat
4453 				 * buffered.
4454 				 */
4455 				setval = rack->r_ctl.rc_gp_srtt;
4456 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
4457 					setval = rack->r_ctl.rc_entry_gp_rtt;
4458 			} else {
4459 				/*
4460 				 * Here we are not highly buffered
4461 				 * and should pick the min we can to
4462 				 * keep from causing loss.
4463 				 */
4464 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4465 			}
4466 			rack_set_prtt_target(rack, segsiz,
4467 					     setval);
4468 		}
4469 		if (rack_probe_rtt_sets_cwnd > 1) {
4470 			/* There is a percentage here to boost */
4471 			ebdp = rack->r_ctl.rc_target_probertt_flight;
4472 			ebdp *= rack_probe_rtt_sets_cwnd;
4473 			ebdp /= 100;
4474 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4475 		} else
4476 			setto = rack->r_ctl.rc_target_probertt_flight;
4477 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4478 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4479 			/* Enforce a min */
4480 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4481 		}
4482 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
4483 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4484 	}
4485 	rack_log_rtt_shrinks(rack,  us_cts,
4486 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4487 			     __LINE__, RACK_RTTS_EXITPROBE);
4488 	/* Clear times last so log has all the info */
4489 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4490 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4491 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4492 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
4493 }
4494 
4495 static void
4496 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4497 {
4498 	/* Check in on probe-rtt */
4499 
4500 	if (rack->rc_gp_filled == 0) {
4501 		/* We do not do p-rtt unless we have gp measurements */
4502 		return;
4503 	}
4504 	if (rack->in_probe_rtt) {
4505 		uint64_t no_overflow;
4506 		uint32_t endtime, must_stay;
4507 
4508 		if (rack->r_ctl.rc_went_idle_time &&
4509 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4510 			/*
4511 			 * We went idle during prtt, just exit now.
4512 			 */
4513 			rack_exit_probertt(rack, us_cts);
4514 		} else if (rack_probe_rtt_safety_val &&
4515 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4516 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4517 			/*
4518 			 * Probe RTT safety value triggered!
4519 			 */
4520 			rack_log_rtt_shrinks(rack,  us_cts,
4521 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4522 					     __LINE__, RACK_RTTS_SAFETY);
4523 			rack_exit_probertt(rack, us_cts);
4524 		}
4525 		/* Calculate the max we will wait */
4526 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4527 		if (rack->rc_highly_buffered)
4528 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4529 		/* Calculate the min we must wait */
4530 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4531 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4532 		    TSTMP_LT(us_cts, endtime)) {
4533 			uint32_t calc;
4534 			/* Do we lower more? */
4535 no_exit:
4536 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4537 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4538 			else
4539 				calc = 0;
4540 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4541 			if (calc) {
4542 				/* Maybe */
4543 				calc *= rack_per_of_gp_probertt_reduce;
4544 				if (calc > rack_per_of_gp_probertt)
4545 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4546 				else
4547 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4548 				/* Limit it too */
4549 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4550 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4551 			}
4552 			/* We must reach target or the time set */
4553 			return;
4554 		}
4555 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
4556 			if ((TSTMP_LT(us_cts, must_stay) &&
4557 			     rack->rc_highly_buffered) ||
4558 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4559 			      rack->r_ctl.rc_target_probertt_flight)) {
4560 				/* We are not past the must_stay time */
4561 				goto no_exit;
4562 			}
4563 			rack_log_rtt_shrinks(rack,  us_cts,
4564 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4565 					     __LINE__, RACK_RTTS_REACHTARGET);
4566 			rack->r_ctl.rc_time_probertt_starts = us_cts;
4567 			if (rack->r_ctl.rc_time_probertt_starts == 0)
4568 				rack->r_ctl.rc_time_probertt_starts = 1;
4569 			/* Restore back to our rate we want to pace at in prtt */
4570 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4571 		}
4572 		/*
4573 		 * Setup our end time, some number of gp_srtts plus 200ms.
4574 		 */
4575 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4576 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4577 		if (rack_probertt_gpsrtt_cnt_div)
4578 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4579 		else
4580 			endtime = 0;
4581 		endtime += rack_min_probertt_hold;
4582 		endtime += rack->r_ctl.rc_time_probertt_starts;
4583 		if (TSTMP_GEQ(us_cts,  endtime)) {
4584 			/* yes, exit probertt */
4585 			rack_exit_probertt(rack, us_cts);
4586 		}
4587 
4588 	} else if ((rack->rc_skip_timely == 0) &&
4589 		   (TSTMP_GT(us_cts, rack->r_ctl.rc_lower_rtt_us_cts)) &&
4590 		   ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt)) {
4591 		/* Go into probertt, its been too long since we went lower */
4592 		rack_enter_probertt(rack, us_cts);
4593 	}
4594 }
4595 
4596 static void
4597 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4598 		       uint32_t rtt, int32_t rtt_diff)
4599 {
4600 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4601 	uint32_t losses;
4602 
4603 	if ((rack->rc_gp_dyn_mul == 0) ||
4604 	    (rack->use_fixed_rate) ||
4605 	    (rack->in_probe_rtt) ||
4606 	    (rack->rc_always_pace == 0)) {
4607 		/* No dynamic GP multiplier in play */
4608 		return;
4609 	}
4610 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4611 	cur_bw = rack_get_bw(rack);
4612 	/* Calculate our up and down range */
4613 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4614 	up_bnd /= 100;
4615 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4616 
4617 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4618 	subfr /= 100;
4619 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4620 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4621 		/*
4622 		 * This is the case where our RTT is above
4623 		 * the max target and we have been configured
4624 		 * to just do timely no bonus up stuff in that case.
4625 		 *
4626 		 * There are two configurations, set to 1, and we
4627 		 * just do timely if we are over our max. If its
4628 		 * set above 1 then we slam the multipliers down
4629 		 * to 100 and then decrement per timely.
4630 		 */
4631 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4632 				__LINE__, 3);
4633 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4634 			rack_validate_multipliers_at_or_below_100(rack);
4635 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4636 	} else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4637 		/*
4638 		 * We are decreasing this is a bit complicated this
4639 		 * means we are loosing ground. This could be
4640 		 * because another flow entered and we are competing
4641 		 * for b/w with it. This will push the RTT up which
4642 		 * makes timely unusable unless we want to get shoved
4643 		 * into a corner and just be backed off (the age
4644 		 * old problem with delay based CC).
4645 		 *
4646 		 * On the other hand if it was a route change we
4647 		 * would like to stay somewhat contained and not
4648 		 * blow out the buffers.
4649 		 */
4650 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4651 				__LINE__, 3);
4652 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4653 		if (rack->rc_gp_bwred == 0) {
4654 			/* Go into reduction counting */
4655 			rack->rc_gp_bwred = 1;
4656 			rack->rc_gp_timely_dec_cnt = 0;
4657 		}
4658 		if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4659 			/*
4660 			 * Push another time with a faster pacing
4661 			 * to try to gain back (we include override to
4662 			 * get a full raise factor).
4663 			 */
4664 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4665 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4666 			    (timely_says == 0) ||
4667 			    (rack_down_raise_thresh == 0)) {
4668 				/*
4669 				 * Do an override up in b/w if we were
4670 				 * below the threshold or if the threshold
4671 				 * is zero we always do the raise.
4672 				 */
4673 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4674 			} else {
4675 				/* Log it stays the same */
4676 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4677 						__LINE__, 11);
4678 			}
4679 			rack->rc_gp_timely_dec_cnt++;
4680 			/* We are not incrementing really no-count */
4681 			rack->rc_gp_incr = 0;
4682 			rack->rc_gp_timely_inc_cnt = 0;
4683 		} else {
4684 			/*
4685 			 * Lets just use the RTT
4686 			 * information and give up
4687 			 * pushing.
4688 			 */
4689 			goto use_timely;
4690 		}
4691 	} else if ((timely_says != 2) &&
4692 		    !losses &&
4693 		    (last_bw_est > up_bnd)) {
4694 		/*
4695 		 * We are increasing b/w lets keep going, updating
4696 		 * our b/w and ignoring any timely input, unless
4697 		 * of course we are at our max raise (if there is one).
4698 		 */
4699 
4700 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4701 				__LINE__, 3);
4702 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4703 		if (rack->rc_gp_saw_ss &&
4704 		    rack->r_ctl.rack_per_upper_bound_ss &&
4705 		     (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4706 			    /*
4707 			     * In cases where we can't go higher
4708 			     * we should just use timely.
4709 			     */
4710 			    goto use_timely;
4711 		}
4712 		if (rack->rc_gp_saw_ca &&
4713 		    rack->r_ctl.rack_per_upper_bound_ca &&
4714 		    (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4715 			    /*
4716 			     * In cases where we can't go higher
4717 			     * we should just use timely.
4718 			     */
4719 			    goto use_timely;
4720 		}
4721 		rack->rc_gp_bwred = 0;
4722 		rack->rc_gp_timely_dec_cnt = 0;
4723 		/* You get a set number of pushes if timely is trying to reduce */
4724 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4725 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4726 		} else {
4727 			/* Log it stays the same */
4728 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4729 			    __LINE__, 12);
4730 		}
4731 		return;
4732 	} else {
4733 		/*
4734 		 * We are staying between the lower and upper range bounds
4735 		 * so use timely to decide.
4736 		 */
4737 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4738 				__LINE__, 3);
4739 use_timely:
4740 		if (timely_says) {
4741 			rack->rc_gp_incr = 0;
4742 			rack->rc_gp_timely_inc_cnt = 0;
4743 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4744 			    !losses &&
4745 			    (last_bw_est < low_bnd)) {
4746 				/* We are loosing ground */
4747 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4748 				rack->rc_gp_timely_dec_cnt++;
4749 				/* We are not incrementing really no-count */
4750 				rack->rc_gp_incr = 0;
4751 				rack->rc_gp_timely_inc_cnt = 0;
4752 			} else
4753 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4754 		} else {
4755 			rack->rc_gp_bwred = 0;
4756 			rack->rc_gp_timely_dec_cnt = 0;
4757 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4758 		}
4759 	}
4760 }
4761 
4762 static int32_t
4763 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4764 {
4765 	int32_t timely_says;
4766 	uint64_t log_mult, log_rtt_a_diff;
4767 
4768 	log_rtt_a_diff = rtt;
4769 	log_rtt_a_diff <<= 32;
4770 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4771 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4772 		    rack_gp_rtt_maxmul)) {
4773 		/* Reduce the b/w multiplier */
4774 		timely_says = 2;
4775 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4776 		log_mult <<= 32;
4777 		log_mult |= prev_rtt;
4778 		rack_log_timely(rack,  timely_says, log_mult,
4779 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4780 				log_rtt_a_diff, __LINE__, 4);
4781 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4782 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4783 			    max(rack_gp_rtt_mindiv , 1)))) {
4784 		/* Increase the b/w multiplier */
4785 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4786 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4787 			 max(rack_gp_rtt_mindiv , 1));
4788 		log_mult <<= 32;
4789 		log_mult |= prev_rtt;
4790 		timely_says = 0;
4791 		rack_log_timely(rack,  timely_says, log_mult ,
4792 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4793 				log_rtt_a_diff, __LINE__, 5);
4794 	} else {
4795 		/*
4796 		 * Use a gradient to find it the timely gradient
4797 		 * is:
4798 		 * grad = rc_rtt_diff / min_rtt;
4799 		 *
4800 		 * anything below or equal to 0 will be
4801 		 * a increase indication. Anything above
4802 		 * zero is a decrease. Note we take care
4803 		 * of the actual gradient calculation
4804 		 * in the reduction (its not needed for
4805 		 * increase).
4806 		 */
4807 		log_mult = prev_rtt;
4808 		if (rtt_diff <= 0) {
4809 			/*
4810 			 * Rttdiff is less than zero, increase the
4811 			 * b/w multiplier (its 0 or negative)
4812 			 */
4813 			timely_says = 0;
4814 			rack_log_timely(rack,  timely_says, log_mult,
4815 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4816 		} else {
4817 			/* Reduce the b/w multiplier */
4818 			timely_says = 1;
4819 			rack_log_timely(rack,  timely_says, log_mult,
4820 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4821 		}
4822 	}
4823 	return (timely_says);
4824 }
4825 
4826 static __inline int
4827 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4828 {
4829 	if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4830 	    SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4831 		/**
4832 		 * This covers the case that the
4833 		 * resent is completely inside
4834 		 * the gp range or up to it.
4835 		 *      |----------------|
4836 		 *      |-----| <or>
4837 		 *            |----|
4838 		 *            <or>   |---|
4839 		 */
4840 		return (1);
4841 	} else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4842 		   SEQ_GT(rsm->r_end, tp->gput_seq)){
4843 		/**
4844 		 * This covers the case of
4845 		 *      |--------------|
4846 		 *  |-------->|
4847 		 */
4848 		return (1);
4849 	} else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4850 		   SEQ_LT(rsm->r_start, tp->gput_ack) &&
4851 		   SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4852 
4853 		/**
4854 		 * This covers the case of
4855 		 *      |--------------|
4856 		 *              |-------->|
4857 		 */
4858 		return (1);
4859 	}
4860 	return (0);
4861 }
4862 
4863 static __inline void
4864 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4865 {
4866 
4867 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
4868 		return;
4869 	/*
4870 	 * We have a Goodput measurement in progress. Mark
4871 	 * the send if its within the window. If its not
4872 	 * in the window make sure it does not have the mark.
4873 	 */
4874 	if (rack_in_gp_window(tp, rsm))
4875 		rsm->r_flags |= RACK_IN_GP_WIN;
4876 	else
4877 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4878 }
4879 
4880 static __inline void
4881 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4882 {
4883 	/* A GP measurement is ending, clear all marks on the send map*/
4884 	struct rack_sendmap *rsm = NULL;
4885 
4886 	rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4887 	if (rsm == NULL) {
4888 		rsm = tqhash_min(rack->r_ctl.tqh);
4889 	}
4890 	/* Nothing left? */
4891 	while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4892 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4893 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4894 	}
4895 }
4896 
4897 
4898 static __inline void
4899 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4900 {
4901 	struct rack_sendmap *rsm = NULL;
4902 
4903 	if (tp->snd_una == tp->snd_max) {
4904 		/* Nothing outstanding yet, nothing to do here */
4905 		return;
4906 	}
4907 	if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4908 		/*
4909 		 * We are measuring ahead of some outstanding
4910 		 * data. We need to walk through up until we get
4911 		 * to gp_seq marking so that no rsm is set incorrectly
4912 		 * with RACK_IN_GP_WIN.
4913 		 */
4914 		rsm = tqhash_min(rack->r_ctl.tqh);
4915 		while (rsm != NULL) {
4916 			rack_mark_in_gp_win(tp, rsm);
4917 			if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4918 				break;
4919 			rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4920 		}
4921 	}
4922 	if (rsm == NULL) {
4923 		/*
4924 		 * Need to find the GP seq, if rsm is
4925 		 * set we stopped as we hit it.
4926 		 */
4927 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4928 		if (rsm == NULL)
4929 			return;
4930 		rack_mark_in_gp_win(tp, rsm);
4931 	}
4932 	/*
4933 	 * Now we may need to mark already sent rsm, ahead of
4934 	 * gput_seq in the window since they may have been sent
4935 	 * *before* we started our measurment. The rsm, if non-null
4936 	 * has been marked (note if rsm would have been NULL we would have
4937 	 * returned in the previous block). So we go to the next, and continue
4938 	 * until we run out of entries or we exceed the gp_ack value.
4939 	 */
4940 	rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4941 	while (rsm) {
4942 		rack_mark_in_gp_win(tp, rsm);
4943 		if (SEQ_GT(rsm->r_end, tp->gput_ack))
4944 			break;
4945 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4946 	}
4947 }
4948 
4949 static void
4950 rack_log_gp_calc(struct tcp_rack *rack, uint32_t add_part, uint32_t sub_part, uint32_t srtt, uint64_t meas_bw, uint64_t utim, uint8_t meth, uint32_t line)
4951 {
4952 	if (tcp_bblogging_on(rack->rc_tp)) {
4953 		union tcp_log_stackspecific log;
4954 		struct timeval tv;
4955 
4956 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4957 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4958 		log.u_bbr.flex1 = add_part;
4959 		log.u_bbr.flex2 = sub_part;
4960 		log.u_bbr.flex3 = rack_wma_divisor;
4961 		log.u_bbr.flex4 = srtt;
4962 		log.u_bbr.flex7 = (uint16_t)line;
4963 		log.u_bbr.flex8 = meth;
4964 		log.u_bbr.delRate = rack->r_ctl.gp_bw;
4965 		log.u_bbr.cur_del_rate = meas_bw;
4966 		log.u_bbr.rttProp = utim;
4967 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4968 		    &rack->rc_inp->inp_socket->so_rcv,
4969 		    &rack->rc_inp->inp_socket->so_snd,
4970 		    BBR_LOG_THRESH_CALC, 0,
4971 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4972 	}
4973 }
4974 
4975 static void
4976 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4977 			    tcp_seq th_ack, int line, uint8_t quality)
4978 {
4979 	uint64_t tim, bytes_ps, stim, utim;
4980 	uint32_t segsiz, bytes, reqbytes, us_cts;
4981 	int32_t gput, new_rtt_diff, timely_says;
4982 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4983 	int did_add = 0;
4984 
4985 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4986 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4987 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4988 		tim = us_cts - tp->gput_ts;
4989 	else
4990 		tim = 0;
4991 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4992 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4993 	else
4994 		stim = 0;
4995 	/*
4996 	 * Use the larger of the send time or ack time. This prevents us
4997 	 * from being influenced by ack artifacts to come up with too
4998 	 * high of measurement. Note that since we are spanning over many more
4999 	 * bytes in most of our measurements hopefully that is less likely to
5000 	 * occur.
5001 	 */
5002 	if (tim > stim)
5003 		utim = max(tim, 1);
5004 	else
5005 		utim = max(stim, 1);
5006 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
5007 	rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
5008 	if ((tim == 0) && (stim == 0)) {
5009 		/*
5010 		 * Invalid measurement time, maybe
5011 		 * all on one ack/one send?
5012 		 */
5013 		bytes = 0;
5014 		bytes_ps = 0;
5015 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5016 					   0, 0, 0, 10, __LINE__, NULL, quality);
5017 		goto skip_measurement;
5018 	}
5019 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
5020 		/* We never made a us_rtt measurement? */
5021 		bytes = 0;
5022 		bytes_ps = 0;
5023 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5024 					   0, 0, 0, 10, __LINE__, NULL, quality);
5025 		goto skip_measurement;
5026 	}
5027 	/*
5028 	 * Calculate the maximum possible b/w this connection
5029 	 * could have. We base our calculation on the lowest
5030 	 * rtt we have seen during the measurement and the
5031 	 * largest rwnd the client has given us in that time. This
5032 	 * forms a BDP that is the maximum that we could ever
5033 	 * get to the client. Anything larger is not valid.
5034 	 *
5035 	 * I originally had code here that rejected measurements
5036 	 * where the time was less than 1/2 the latest us_rtt.
5037 	 * But after thinking on that I realized its wrong since
5038 	 * say you had a 150Mbps or even 1Gbps link, and you
5039 	 * were a long way away.. example I am in Europe (100ms rtt)
5040 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
5041 	 * bytes my time would be 1.2ms, and yet my rtt would say
5042 	 * the measurement was invalid the time was < 50ms. The
5043 	 * same thing is true for 150Mb (8ms of time).
5044 	 *
5045 	 * A better way I realized is to look at what the maximum
5046 	 * the connection could possibly do. This is gated on
5047 	 * the lowest RTT we have seen and the highest rwnd.
5048 	 * We should in theory never exceed that, if we are
5049 	 * then something on the path is storing up packets
5050 	 * and then feeding them all at once to our endpoint
5051 	 * messing up our measurement.
5052 	 */
5053 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
5054 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
5055 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
5056 	if (SEQ_LT(th_ack, tp->gput_seq)) {
5057 		/* No measurement can be made */
5058 		bytes = 0;
5059 		bytes_ps = 0;
5060 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5061 					   0, 0, 0, 10, __LINE__, NULL, quality);
5062 		goto skip_measurement;
5063 	} else
5064 		bytes = (th_ack - tp->gput_seq);
5065 	bytes_ps = (uint64_t)bytes;
5066 	/*
5067 	 * Don't measure a b/w for pacing unless we have gotten at least
5068 	 * an initial windows worth of data in this measurement interval.
5069 	 *
5070 	 * Small numbers of bytes get badly influenced by delayed ack and
5071 	 * other artifacts. Note we take the initial window or our
5072 	 * defined minimum GP (defaulting to 10 which hopefully is the
5073 	 * IW).
5074 	 */
5075 	if (rack->rc_gp_filled == 0) {
5076 		/*
5077 		 * The initial estimate is special. We
5078 		 * have blasted out an IW worth of packets
5079 		 * without a real valid ack ts results. We
5080 		 * then setup the app_limited_needs_set flag,
5081 		 * this should get the first ack in (probably 2
5082 		 * MSS worth) to be recorded as the timestamp.
5083 		 * We thus allow a smaller number of bytes i.e.
5084 		 * IW - 2MSS.
5085 		 */
5086 		reqbytes -= (2 * segsiz);
5087 		/* Also lets fill previous for our first measurement to be neutral */
5088 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5089 	}
5090 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
5091 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5092 					   rack->r_ctl.rc_app_limited_cnt,
5093 					   0, 0, 10, __LINE__, NULL, quality);
5094 		goto skip_measurement;
5095 	}
5096 	/*
5097 	 * We now need to calculate the Timely like status so
5098 	 * we can update (possibly) the b/w multipliers.
5099 	 */
5100 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
5101 	if (rack->rc_gp_filled == 0) {
5102 		/* No previous reading */
5103 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
5104 	} else {
5105 		if (rack->measure_saw_probe_rtt == 0) {
5106 			/*
5107 			 * We don't want a probertt to be counted
5108 			 * since it will be negative incorrectly. We
5109 			 * expect to be reducing the RTT when we
5110 			 * pace at a slower rate.
5111 			 */
5112 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
5113 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
5114 		}
5115 	}
5116 	timely_says = rack_make_timely_judgement(rack,
5117 	    rack->r_ctl.rc_gp_srtt,
5118 	    rack->r_ctl.rc_rtt_diff,
5119 	    rack->r_ctl.rc_prev_gp_srtt
5120 	);
5121 	bytes_ps *= HPTS_USEC_IN_SEC;
5122 	bytes_ps /= utim;
5123 	if (bytes_ps > rack->r_ctl.last_max_bw) {
5124 		/*
5125 		 * Something is on path playing
5126 		 * since this b/w is not possible based
5127 		 * on our BDP (highest rwnd and lowest rtt
5128 		 * we saw in the measurement window).
5129 		 *
5130 		 * Another option here would be to
5131 		 * instead skip the measurement.
5132 		 */
5133 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
5134 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
5135 					   11, __LINE__, NULL, quality);
5136 		bytes_ps = rack->r_ctl.last_max_bw;
5137 	}
5138 	/* We store gp for b/w in bytes per second */
5139 	if (rack->rc_gp_filled == 0) {
5140 		/* Initial measurement */
5141 		if (bytes_ps) {
5142 			rack->r_ctl.gp_bw = bytes_ps;
5143 			rack->rc_gp_filled = 1;
5144 			rack->r_ctl.num_measurements = 1;
5145 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5146 		} else {
5147 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5148 						   rack->r_ctl.rc_app_limited_cnt,
5149 						   0, 0, 10, __LINE__, NULL, quality);
5150 		}
5151 		if (tcp_in_hpts(rack->rc_tp) &&
5152 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5153 			/*
5154 			 * Ok we can't trust the pacer in this case
5155 			 * where we transition from un-paced to paced.
5156 			 * Or for that matter when the burst mitigation
5157 			 * was making a wild guess and got it wrong.
5158 			 * Stop the pacer and clear up all the aggregate
5159 			 * delays etc.
5160 			 */
5161 			tcp_hpts_remove(rack->rc_tp);
5162 			rack->r_ctl.rc_hpts_flags = 0;
5163 			rack->r_ctl.rc_last_output_to = 0;
5164 		}
5165 		did_add = 2;
5166 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5167 		/* Still a small number run an average */
5168 		rack->r_ctl.gp_bw += bytes_ps;
5169 		addpart = rack->r_ctl.num_measurements;
5170 		rack->r_ctl.num_measurements++;
5171 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5172 			/* We have collected enough to move forward */
5173 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5174 		}
5175 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5176 		did_add = 3;
5177 	} else {
5178 		/*
5179 		 * We want to take 1/wma of the goodput and add in to 7/8th
5180 		 * of the old value weighted by the srtt. So if your measurement
5181 		 * period is say 2 SRTT's long you would get 1/4 as the
5182 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
5183 		 *
5184 		 * But we must be careful not to take too much i.e. if the
5185 		 * srtt is say 20ms and the measurement is taken over
5186 		 * 400ms our weight would be 400/20 i.e. 20. On the
5187 		 * other hand if we get a measurement over 1ms with a
5188 		 * 10ms rtt we only want to take a much smaller portion.
5189 		 */
5190 		uint8_t meth;
5191 
5192 		if (rack->r_ctl.num_measurements < 0xff) {
5193 			rack->r_ctl.num_measurements++;
5194 		}
5195 		srtt = (uint64_t)tp->t_srtt;
5196 		if (srtt == 0) {
5197 			/*
5198 			 * Strange why did t_srtt go back to zero?
5199 			 */
5200 			if (rack->r_ctl.rc_rack_min_rtt)
5201 				srtt = rack->r_ctl.rc_rack_min_rtt;
5202 			else
5203 				srtt = HPTS_USEC_IN_MSEC;
5204 		}
5205 		/*
5206 		 * XXXrrs: Note for reviewers, in playing with
5207 		 * dynamic pacing I discovered this GP calculation
5208 		 * as done originally leads to some undesired results.
5209 		 * Basically you can get longer measurements contributing
5210 		 * too much to the WMA. Thus I changed it if you are doing
5211 		 * dynamic adjustments to only do the aportioned adjustment
5212 		 * if we have a very small (time wise) measurement. Longer
5213 		 * measurements just get there weight (defaulting to 1/8)
5214 		 * add to the WMA. We may want to think about changing
5215 		 * this to always do that for both sides i.e. dynamic
5216 		 * and non-dynamic... but considering lots of folks
5217 		 * were playing with this I did not want to change the
5218 		 * calculation per.se. without your thoughts.. Lawerence?
5219 		 * Peter??
5220 		 */
5221 		if (rack->rc_gp_dyn_mul == 0) {
5222 			subpart = rack->r_ctl.gp_bw * utim;
5223 			subpart /= (srtt * 8);
5224 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
5225 				/*
5226 				 * The b/w update takes no more
5227 				 * away then 1/2 our running total
5228 				 * so factor it in.
5229 				 */
5230 				addpart = bytes_ps * utim;
5231 				addpart /= (srtt * 8);
5232 				meth = 1;
5233 			} else {
5234 				/*
5235 				 * Don't allow a single measurement
5236 				 * to account for more than 1/2 of the
5237 				 * WMA. This could happen on a retransmission
5238 				 * where utim becomes huge compared to
5239 				 * srtt (multiple retransmissions when using
5240 				 * the sending rate which factors in all the
5241 				 * transmissions from the first one).
5242 				 */
5243 				subpart = rack->r_ctl.gp_bw / 2;
5244 				addpart = bytes_ps / 2;
5245 				meth = 2;
5246 			}
5247 			rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5248 			resid_bw = rack->r_ctl.gp_bw - subpart;
5249 			rack->r_ctl.gp_bw = resid_bw + addpart;
5250 			did_add = 1;
5251 		} else {
5252 			if ((utim / srtt) <= 1) {
5253 				/*
5254 				 * The b/w update was over a small period
5255 				 * of time. The idea here is to prevent a small
5256 				 * measurement time period from counting
5257 				 * too much. So we scale it based on the
5258 				 * time so it attributes less than 1/rack_wma_divisor
5259 				 * of its measurement.
5260 				 */
5261 				subpart = rack->r_ctl.gp_bw * utim;
5262 				subpart /= (srtt * rack_wma_divisor);
5263 				addpart = bytes_ps * utim;
5264 				addpart /= (srtt * rack_wma_divisor);
5265 				meth = 3;
5266 			} else {
5267 				/*
5268 				 * The scaled measurement was long
5269 				 * enough so lets just add in the
5270 				 * portion of the measurement i.e. 1/rack_wma_divisor
5271 				 */
5272 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5273 				addpart = bytes_ps / rack_wma_divisor;
5274 				meth = 4;
5275 			}
5276 			if ((rack->measure_saw_probe_rtt == 0) ||
5277 		            (bytes_ps > rack->r_ctl.gp_bw)) {
5278 				/*
5279 				 * For probe-rtt we only add it in
5280 				 * if its larger, all others we just
5281 				 * add in.
5282 				 */
5283 				did_add = 1;
5284 				rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5285 				resid_bw = rack->r_ctl.gp_bw - subpart;
5286 				rack->r_ctl.gp_bw = resid_bw + addpart;
5287 			}
5288 		}
5289 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5290 	}
5291 	/*
5292 	 * We only watch the growth of the GP during the initial startup
5293 	 * or first-slowstart that ensues. If we ever needed to watch
5294 	 * growth of gp outside of that period all we need to do is
5295 	 * remove the first clause of this if (rc_initial_ss_comp).
5296 	 */
5297 	if ((rack->rc_initial_ss_comp == 0) &&
5298 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG)) {
5299 		uint64_t gp_est;
5300 
5301 		gp_est = bytes_ps;
5302 		if (tcp_bblogging_on(rack->rc_tp)) {
5303 			union tcp_log_stackspecific log;
5304 			struct timeval tv;
5305 
5306 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5307 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5308 			log.u_bbr.flex1 = rack->r_ctl.current_round;
5309 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
5310 			log.u_bbr.delRate = gp_est;
5311 			log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5312 			log.u_bbr.flex8 = 41;
5313 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5314 					    0, &log, false, NULL, __func__, __LINE__,&tv);
5315 		}
5316 		if ((rack->r_ctl.num_measurements == RACK_REQ_AVG) ||
5317 		    (rack->r_ctl.last_gpest == 0)) {
5318 			/*
5319 			 * The round we get our measurement averaging going
5320 			 * is the base round so it always is the source point
5321 			 * for when we had our first increment. From there on
5322 			 * we only record the round that had a rise.
5323 			 */
5324 			rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5325 			rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5326 		} else if (gp_est >= rack->r_ctl.last_gpest) {
5327 			/*
5328 			 * Test to see if its gone up enough
5329 			 * to set the round count up to now. Note
5330 			 * that on the seeding of the 4th measurement we
5331 			 */
5332 			gp_est *= 1000;
5333 			gp_est /= rack->r_ctl.last_gpest;
5334 			if ((uint32_t)gp_est > rack->r_ctl.gp_gain_req) {
5335 				/*
5336 				 * We went up enough to record the round.
5337 				 */
5338 				if (tcp_bblogging_on(rack->rc_tp)) {
5339 					union tcp_log_stackspecific log;
5340 					struct timeval tv;
5341 
5342 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5343 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5344 					log.u_bbr.flex1 = rack->r_ctl.current_round;
5345 					log.u_bbr.flex2 = (uint32_t)gp_est;
5346 					log.u_bbr.flex3 = rack->r_ctl.gp_gain_req;
5347 					log.u_bbr.delRate = gp_est;
5348 					log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5349 					log.u_bbr.flex8 = 42;
5350 					(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5351 							    0, &log, false, NULL, __func__, __LINE__,&tv);
5352 				}
5353 				rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5354 				if (rack->r_ctl.use_gp_not_last == 1)
5355 					rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5356 				else
5357 					rack->r_ctl.last_gpest = bytes_ps;
5358 			}
5359 		}
5360 	}
5361 	if ((rack->gp_ready == 0) &&
5362 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5363 		/* We have enough measurements now */
5364 		rack->gp_ready = 1;
5365 		if (rack->dgp_on ||
5366 		    rack->rack_hibeta)
5367 			rack_set_cc_pacing(rack);
5368 		if (rack->defer_options)
5369 			rack_apply_deferred_options(rack);
5370 	}
5371 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5372 				   rack_get_bw(rack), 22, did_add, NULL, quality);
5373 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
5374 
5375 	if ((rack->measure_saw_probe_rtt == 0) &&
5376 	    rack->rc_gp_rtt_set) {
5377 		if (rack->rc_skip_timely == 0) {
5378 			rack_update_multiplier(rack, timely_says, bytes_ps,
5379 					       rack->r_ctl.rc_gp_srtt,
5380 					       rack->r_ctl.rc_rtt_diff);
5381 		}
5382 	}
5383 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5384 				   rack_get_bw(rack), 3, line, NULL, quality);
5385 	rack_log_pacing_delay_calc(rack,
5386 				   bytes, /* flex2 */
5387 				   tim, /* flex1 */
5388 				   bytes_ps, /* bw_inuse */
5389 				   rack->r_ctl.gp_bw, /* delRate */
5390 				   rack_get_lt_bw(rack), /* rttProp */
5391 				   20, line, NULL, 0);
5392 	/* reset the gp srtt and setup the new prev */
5393 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5394 	/* Record the lost count for the next measurement */
5395 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5396 skip_measurement:
5397 	/*
5398 	 * We restart our diffs based on the gpsrtt in the
5399 	 * measurement window.
5400 	 */
5401 	rack->rc_gp_rtt_set = 0;
5402 	rack->rc_gp_saw_rec = 0;
5403 	rack->rc_gp_saw_ca = 0;
5404 	rack->rc_gp_saw_ss = 0;
5405 	rack->rc_dragged_bottom = 0;
5406 	if (quality == RACK_QUALITY_HIGH) {
5407 		/*
5408 		 * Gput in the stats world is in kbps where bytes_ps is
5409 		 * bytes per second so we do ((x * 8)/ 1000).
5410 		 */
5411 		gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5412 #ifdef STATS
5413 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5414 					 gput);
5415 		/*
5416 		 * XXXLAS: This is a temporary hack, and should be
5417 		 * chained off VOI_TCP_GPUT when stats(9) grows an
5418 		 * API to deal with chained VOIs.
5419 		 */
5420 		if (tp->t_stats_gput_prev > 0)
5421 			stats_voi_update_abs_s32(tp->t_stats,
5422 						 VOI_TCP_GPUT_ND,
5423 						 ((gput - tp->t_stats_gput_prev) * 100) /
5424 						 tp->t_stats_gput_prev);
5425 #endif
5426 		tp->t_stats_gput_prev = gput;
5427 	}
5428 	tp->t_flags &= ~TF_GPUTINPROG;
5429 	/*
5430 	 * Now are we app limited now and there is space from where we
5431 	 * were to where we want to go?
5432 	 *
5433 	 * We don't do the other case i.e. non-applimited here since
5434 	 * the next send will trigger us picking up the missing data.
5435 	 */
5436 	if (rack->r_ctl.rc_first_appl &&
5437 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
5438 	    rack->r_ctl.rc_app_limited_cnt &&
5439 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5440 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5441 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5442 		/*
5443 		 * Yep there is enough outstanding to make a measurement here.
5444 		 */
5445 		struct rack_sendmap *rsm;
5446 
5447 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5448 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5449 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5450 		rack->app_limited_needs_set = 0;
5451 		tp->gput_seq = th_ack;
5452 		if (rack->in_probe_rtt)
5453 			rack->measure_saw_probe_rtt = 1;
5454 		else if ((rack->measure_saw_probe_rtt) &&
5455 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5456 			rack->measure_saw_probe_rtt = 0;
5457 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5458 			/* There is a full window to gain info from */
5459 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5460 		} else {
5461 			/* We can only measure up to the applimited point */
5462 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5463 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5464 				/*
5465 				 * We don't have enough to make a measurement.
5466 				 */
5467 				tp->t_flags &= ~TF_GPUTINPROG;
5468 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5469 							   0, 0, 0, 6, __LINE__, NULL, quality);
5470 				return;
5471 			}
5472 		}
5473 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
5474 			/*
5475 			 * We will get no more data into the SB
5476 			 * this means we need to have the data available
5477 			 * before we start a measurement.
5478 			 */
5479 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5480 				/* Nope not enough data. */
5481 				return;
5482 			}
5483 		}
5484 		tp->t_flags |= TF_GPUTINPROG;
5485 		/*
5486 		 * Now we need to find the timestamp of the send at tp->gput_seq
5487 		 * for the send based measurement.
5488 		 */
5489 		rack->r_ctl.rc_gp_cumack_ts = 0;
5490 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5491 		if (rsm) {
5492 			/* Ok send-based limit is set */
5493 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5494 				/*
5495 				 * Move back to include the earlier part
5496 				 * so our ack time lines up right (this may
5497 				 * make an overlapping measurement but thats
5498 				 * ok).
5499 				 */
5500 				tp->gput_seq = rsm->r_start;
5501 			}
5502 			if (rsm->r_flags & RACK_ACKED) {
5503 				struct rack_sendmap *nrsm;
5504 
5505 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5506 				tp->gput_seq = rsm->r_end;
5507 				nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5508 				if (nrsm)
5509 					rsm = nrsm;
5510 				else {
5511 					rack->app_limited_needs_set = 1;
5512 				}
5513 			} else
5514 				rack->app_limited_needs_set = 1;
5515 			/* We always go from the first send */
5516 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5517 		} else {
5518 			/*
5519 			 * If we don't find the rsm due to some
5520 			 * send-limit set the current time, which
5521 			 * basically disables the send-limit.
5522 			 */
5523 			struct timeval tv;
5524 
5525 			microuptime(&tv);
5526 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5527 		}
5528 		rack_tend_gp_marks(tp, rack);
5529 		rack_log_pacing_delay_calc(rack,
5530 					   tp->gput_seq,
5531 					   tp->gput_ack,
5532 					   (uintptr_t)rsm,
5533 					   tp->gput_ts,
5534 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5535 					   9,
5536 					   __LINE__, rsm, quality);
5537 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5538 	} else {
5539 		/*
5540 		 * To make sure proper timestamp merging occurs, we need to clear
5541 		 * all GP marks if we don't start a measurement.
5542 		 */
5543 		rack_clear_gp_marks(tp, rack);
5544 	}
5545 }
5546 
5547 /*
5548  * CC wrapper hook functions
5549  */
5550 static void
5551 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5552     uint16_t type, int32_t post_recovery)
5553 {
5554 	uint32_t prior_cwnd, acked;
5555 	struct tcp_log_buffer *lgb = NULL;
5556 	uint8_t labc_to_use, quality;
5557 
5558 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5559 	tp->t_ccv.nsegs = nsegs;
5560 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5561 	if ((post_recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5562 		uint32_t max;
5563 
5564 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5565 		if (tp->t_ccv.bytes_this_ack > max) {
5566 			tp->t_ccv.bytes_this_ack = max;
5567 		}
5568 	}
5569 #ifdef STATS
5570 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5571 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5572 #endif
5573 	if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5574 		/*
5575 		 * We will ack all the data, time to end any
5576 		 * lt_bw_up we have running until something
5577 		 * new is sent. Note we need to use the actual
5578 		 * ack_rcv_time which with pacing may be different.
5579 		 */
5580 		uint64_t tmark;
5581 
5582 		rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5583 		rack->r_ctl.lt_seq = tp->snd_max;
5584 		tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
5585 		if (tmark >= rack->r_ctl.lt_timemark) {
5586 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
5587 		}
5588 		rack->r_ctl.lt_timemark = tmark;
5589 		rack->lt_bw_up = 0;
5590 	}
5591 	quality = RACK_QUALITY_NONE;
5592 	if ((tp->t_flags & TF_GPUTINPROG) &&
5593 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5594 		/* Measure the Goodput */
5595 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5596 	}
5597 	/* Which way our we limited, if not cwnd limited no advance in CA */
5598 	if (tp->snd_cwnd <= tp->snd_wnd)
5599 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
5600 	else
5601 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5602 	if (tp->snd_cwnd > tp->snd_ssthresh) {
5603 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5604 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5605 		/* For the setting of a window past use the actual scwnd we are using */
5606 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5607 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5608 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5609 		}
5610 	} else {
5611 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5612 		tp->t_bytes_acked = 0;
5613 	}
5614 	prior_cwnd = tp->snd_cwnd;
5615 	if ((post_recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5616 	    (rack_client_low_buf && rack->client_bufferlvl &&
5617 	    (rack->client_bufferlvl < rack_client_low_buf)))
5618 		labc_to_use = rack->rc_labc;
5619 	else
5620 		labc_to_use = rack_max_abc_post_recovery;
5621 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5622 		union tcp_log_stackspecific log;
5623 		struct timeval tv;
5624 
5625 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5626 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5627 		log.u_bbr.flex1 = th_ack;
5628 		log.u_bbr.flex2 = tp->t_ccv.flags;
5629 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5630 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5631 		log.u_bbr.flex5 = labc_to_use;
5632 		log.u_bbr.flex6 = prior_cwnd;
5633 		log.u_bbr.flex7 = V_tcp_do_newsack;
5634 		log.u_bbr.flex8 = 1;
5635 		lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5636 				     0, &log, false, NULL, __func__, __LINE__,&tv);
5637 	}
5638 	if (CC_ALGO(tp)->ack_received != NULL) {
5639 		/* XXXLAS: Find a way to live without this */
5640 		tp->t_ccv.curack = th_ack;
5641 		tp->t_ccv.labc = labc_to_use;
5642 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5643 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5644 	}
5645 	if (lgb) {
5646 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5647 	}
5648 	if (rack->r_must_retran) {
5649 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5650 			/*
5651 			 * We now are beyond the rxt point so lets disable
5652 			 * the flag.
5653 			 */
5654 			rack->r_ctl.rc_out_at_rto = 0;
5655 			rack->r_must_retran = 0;
5656 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5657 			/*
5658 			 * Only decrement the rc_out_at_rto if the cwnd advances
5659 			 * at least a whole segment. Otherwise next time the peer
5660 			 * acks, we won't be able to send this generaly happens
5661 			 * when we are in Congestion Avoidance.
5662 			 */
5663 			if (acked <= rack->r_ctl.rc_out_at_rto){
5664 				rack->r_ctl.rc_out_at_rto -= acked;
5665 			} else {
5666 				rack->r_ctl.rc_out_at_rto = 0;
5667 			}
5668 		}
5669 	}
5670 #ifdef STATS
5671 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5672 #endif
5673 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5674 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5675 	}
5676 	if ((rack->rc_initial_ss_comp == 0) &&
5677 	    (tp->snd_cwnd >= tp->snd_ssthresh)) {
5678 		/*
5679 		 * The cwnd has grown beyond ssthresh we have
5680 		 * entered ca and completed our first Slowstart.
5681 		 */
5682 		rack->rc_initial_ss_comp = 1;
5683 	}
5684 }
5685 
5686 static void
5687 tcp_rack_partialack(struct tcpcb *tp)
5688 {
5689 	struct tcp_rack *rack;
5690 
5691 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5692 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5693 	/*
5694 	 * If we are doing PRR and have enough
5695 	 * room to send <or> we are pacing and prr
5696 	 * is disabled we will want to see if we
5697 	 * can send data (by setting r_wanted_output to
5698 	 * true).
5699 	 */
5700 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5701 	    rack->rack_no_prr)
5702 		rack->r_wanted_output = 1;
5703 }
5704 
5705 static inline uint64_t
5706 rack_get_rxt_per(uint64_t snds,  uint64_t rxts)
5707 {
5708 	uint64_t rxt_per;
5709 
5710 	if (snds > 0) {
5711 		rxt_per = rxts * 1000;
5712 		rxt_per /= snds;
5713 	} else {
5714 		/* This is an unlikely path */
5715 		if (rxts) {
5716 			/* Its the max it was all re-transmits */
5717 			rxt_per = 0xffffffffffffffff;
5718 		} else {
5719 			rxt_per = 0;
5720 		}
5721 	}
5722 	return (rxt_per);
5723 }
5724 
5725 static void
5726 policer_detection_log(struct tcp_rack *rack, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint32_t flex4, uint8_t flex8)
5727 {
5728 	if (tcp_bblogging_on(rack->rc_tp)) {
5729 		union tcp_log_stackspecific log;
5730 		struct timeval tv;
5731 
5732 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5733 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5734 		log.u_bbr.flex1 = flex1;
5735 		log.u_bbr.flex2 = flex2;
5736 		log.u_bbr.flex3 = flex3;
5737 		log.u_bbr.flex4 = flex4;
5738 		log.u_bbr.flex5 = rack->r_ctl.current_policer_bucket;
5739 		log.u_bbr.flex6 = rack->r_ctl.policer_bucket_size;
5740 		log.u_bbr.flex7 = 0;
5741 		log.u_bbr.flex8 = flex8;
5742 		log.u_bbr.bw_inuse = rack->r_ctl.policer_bw;
5743 		log.u_bbr.applimited = rack->r_ctl.current_round;
5744 		log.u_bbr.epoch = rack->r_ctl.policer_max_seg;
5745 		log.u_bbr.delivered = (uint32_t)rack->r_ctl.bytes_acked_in_recovery;
5746 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
5747 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
5748 		log.u_bbr.rttProp = rack->r_ctl.gp_bw;
5749 		log.u_bbr.bbr_state = rack->rc_policer_detected;
5750 		log.u_bbr.bbr_substate = 0;
5751 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5752 		log.u_bbr.use_lt_bw = rack->policer_detect_on;
5753 		log.u_bbr.lt_epoch = 0;
5754 		log.u_bbr.pkts_out = 0;
5755 		tcp_log_event(rack->rc_tp, NULL, NULL, NULL, TCP_POLICER_DET, 0,
5756 			      0, &log, false, NULL, NULL, 0, &tv);
5757 	}
5758 
5759 }
5760 
5761 static void
5762 policer_detection(struct tcpcb *tp, struct tcp_rack *rack, int post_recovery)
5763 {
5764 	/*
5765 	 * Rack excess rxt accounting is turned on. If we
5766 	 * are above a threshold of rxt's in at least N
5767 	 * rounds, then back off the cwnd and ssthresh
5768 	 * to fit into the long-term b/w.
5769 	 */
5770 
5771 	uint32_t pkts, mid, med, alt_med, avg, segsiz, tot_retran_pkt_count = 0;
5772 	uint32_t cnt_of_mape_rxt = 0;
5773 	uint64_t snds, rxts, rxt_per, tim, del, del_bw;
5774 	int i;
5775 	struct timeval tv;
5776 
5777 
5778 	/*
5779 	 * First is there enough packets delivered during recovery to make
5780 	 * a determiniation of b/w?
5781 	 */
5782 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5783 	if ((rack->rc_policer_detected == 0) &&
5784 	    (rack->r_ctl.policer_del_mss > 0) &&
5785 	    ((uint32_t)rack->r_ctl.policer_del_mss > ((rack->r_ctl.bytes_acked_in_recovery + segsiz - 1)/segsiz))) {
5786 		/*
5787 		 * Not enough data sent in recovery for initial detection. Once
5788 		 * we have deteced a policer we allow less than the threshold (polcer_del_mss)
5789 		 * amount of data in a recovery to let us fall through and double check
5790 		 * our policer settings and possibly expand or collapse the bucket size and
5791 		 * the polcier b/w.
5792 		 *
5793 		 * Once you are declared to be policed. this block of code cannot be
5794 		 * reached, instead blocks further down will re-check the policer detection
5795 		 * triggers and possibly reset the measurements if somehow we have let the
5796 		 * policer bucket size grow too large.
5797 		 */
5798 		if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5799 			policer_detection_log(rack, rack->r_ctl.policer_del_mss,
5800 					      ((rack->r_ctl.bytes_acked_in_recovery + segsiz - 1)/segsiz),
5801 					      rack->r_ctl.bytes_acked_in_recovery, segsiz, 18);
5802 		}
5803 		return;
5804 	}
5805 	tcp_get_usecs(&tv);
5806 	tim = tcp_tv_to_lusectick(&tv) - rack->r_ctl.time_entered_recovery;
5807 	del = rack->r_ctl.bytes_acked_in_recovery;
5808 	if (tim > 0)
5809 		del_bw = (del * (uint64_t)1000000) / tim;
5810 	else
5811 		del_bw = 0;
5812 	/* B/W compensation? */
5813 
5814 	if (rack->r_ctl.pol_bw_comp && ((rack->r_ctl.policer_bw > 0) ||
5815 					(del_bw > 0))) {
5816 		/*
5817 		 * Sanity check now that the data is in. How long does it
5818 		 * take for us to pace out two of our policer_max_seg's?
5819 		 *
5820 		 * If it is longer than the RTT then we are set
5821 		 * too slow, maybe because of not enough data
5822 		 * sent during recovery.
5823 		 */
5824 		uint64_t lentime, res, srtt, max_delbw, alt_bw;
5825 
5826 		srtt = (uint64_t)rack_grab_rtt(tp, rack);
5827 		if ((tp->t_srtt > 0) && (srtt > tp->t_srtt))
5828 			srtt = tp->t_srtt;
5829 		lentime = rack->r_ctl.policer_max_seg * (uint64_t)HPTS_USEC_IN_SEC * 2;
5830 		if (del_bw > rack->r_ctl.policer_bw) {
5831 			max_delbw = del_bw;
5832 		} else {
5833 			max_delbw = rack->r_ctl.policer_bw;
5834 		}
5835 		res = lentime / max_delbw;
5836 		if ((srtt > 0) && (res > srtt)) {
5837 			/*
5838 			 * At this rate we can not get two policer_maxsegs
5839 			 * out before the ack arrives back.
5840 			 *
5841 			 * Lets at least get it raised up so that
5842 			 * we can be a bit faster than that if possible.
5843 			 */
5844 			lentime = (rack->r_ctl.policer_max_seg * 2);
5845 			tim = srtt;
5846 			alt_bw = (lentime * (uint64_t)HPTS_USEC_IN_SEC) / tim;
5847 			if (alt_bw > max_delbw) {
5848 				uint64_t cap_alt_bw;
5849 
5850 				cap_alt_bw = (max_delbw + (max_delbw * rack->r_ctl.pol_bw_comp));
5851 				if ((rack_pol_min_bw > 0) && (cap_alt_bw < rack_pol_min_bw)) {
5852 					/* We place a min on the cap which defaults to 1Mbps */
5853 					cap_alt_bw = rack_pol_min_bw;
5854 				}
5855 				if (alt_bw <= cap_alt_bw) {
5856 					/* It should be */
5857 					del_bw = alt_bw;
5858 					policer_detection_log(rack,
5859 							      (uint32_t)tim,
5860 							      rack->r_ctl.policer_max_seg,
5861 							      0,
5862 							      0,
5863 							      16);
5864 				} else {
5865 					/*
5866 					 * This is an odd case where likely the RTT is very very
5867 					 * low. And yet it is still being policed. We don't want
5868 					 * to get more than (rack_policing_do_bw_comp+1) x del-rate
5869 					 * where del-rate is what we got in recovery for either the
5870 					 * first Policer Detection(PD) or this PD we are on now.
5871 					 */
5872 					del_bw = cap_alt_bw;
5873 					policer_detection_log(rack,
5874 							      (uint32_t)tim,
5875 							      rack->r_ctl.policer_max_seg,
5876 							      (uint32_t)max_delbw,
5877 							      (rack->r_ctl.pol_bw_comp + 1),
5878 							      16);
5879 				}
5880 			}
5881 		}
5882 	}
5883 	snds = tp->t_sndbytes - rack->r_ctl.last_policer_sndbytes;
5884 	rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_policer_snd_rxt_bytes;
5885 	rxt_per = rack_get_rxt_per(snds,  rxts);
5886 	/* Figure up the average  and median */
5887 	for(i = 0; i < RETRAN_CNT_SIZE; i++) {
5888 		if (rack->r_ctl.rc_cnt_of_retran[i] > 0) {
5889 			tot_retran_pkt_count += (i + 1) * rack->r_ctl.rc_cnt_of_retran[i];
5890 			cnt_of_mape_rxt  += rack->r_ctl.rc_cnt_of_retran[i];
5891 		}
5892 	}
5893 	if (cnt_of_mape_rxt)
5894 		avg = (tot_retran_pkt_count * 10)/cnt_of_mape_rxt;
5895 	else
5896 		avg = 0;
5897 	alt_med = med = 0;
5898 	mid = tot_retran_pkt_count/2;
5899 	for(i = 0; i < RETRAN_CNT_SIZE; i++) {
5900 		pkts = (i + 1) * rack->r_ctl.rc_cnt_of_retran[i];
5901 		if (mid > pkts) {
5902 			mid -= pkts;
5903 			continue;
5904 		}
5905 		med = (i + 1);
5906 		break;
5907 	}
5908 	mid = cnt_of_mape_rxt / 2;
5909 	for(i = 0; i < RETRAN_CNT_SIZE; i++) {
5910 		if (mid > rack->r_ctl.rc_cnt_of_retran[i]) {
5911 			mid -= rack->r_ctl.rc_cnt_of_retran[i];
5912 			continue;
5913 		}
5914 		alt_med = (i + 1);
5915 		break;
5916 	}
5917 	if (rack->r_ctl.policer_alt_median) {
5918 		/* Swap the medians */
5919 		uint32_t swap;
5920 
5921 		swap = med;
5922 		med = alt_med;
5923 		alt_med = swap;
5924 	}
5925 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5926 		union tcp_log_stackspecific log;
5927 		struct timeval tv;
5928 
5929 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5930 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5931 		log.u_bbr.flex1 = avg;
5932 		log.u_bbr.flex2 = med;
5933 		log.u_bbr.flex3 = (uint32_t)rxt_per;
5934 		log.u_bbr.flex4 = rack->r_ctl.policer_avg_threshold;
5935 		log.u_bbr.flex5 = rack->r_ctl.policer_med_threshold;
5936 		log.u_bbr.flex6 = rack->r_ctl.policer_rxt_threshold;
5937 		log.u_bbr.flex7 = rack->r_ctl.policer_alt_median;
5938 		log.u_bbr.flex8 = 1;
5939 		log.u_bbr.delivered = rack->r_ctl.policer_bucket_size;
5940 		log.u_bbr.applimited = rack->r_ctl.current_round;
5941 		log.u_bbr.epoch = rack->r_ctl.policer_max_seg;
5942 		log.u_bbr.bw_inuse = del_bw;
5943 		log.u_bbr.cur_del_rate = rxts;
5944 		log.u_bbr.delRate = snds;
5945 		log.u_bbr.rttProp = rack->r_ctl.gp_bw;
5946 		log.u_bbr.bbr_state = rack->rc_policer_detected;
5947 		log.u_bbr.bbr_substate = 0;
5948 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5949 		log.u_bbr.use_lt_bw = rack->policer_detect_on;
5950 		log.u_bbr.lt_epoch = (uint32_t)tim;
5951 		log.u_bbr.pkts_out = rack->r_ctl.bytes_acked_in_recovery;
5952 		tcp_log_event(tp, NULL, NULL, NULL, TCP_POLICER_DET, 0,
5953 			      0, &log, false, NULL, NULL, 0, &tv);
5954 	}
5955 	if (med == RETRAN_CNT_SIZE) {
5956 		/*
5957 		 * If the median is the maximum, then what we
5958 		 * likely have here is a network breakage. Either that
5959 		 * or we are so unlucky that all of our traffic is being
5960 		 * dropped and having to be retransmitted the maximum times
5961 		 * and this just is not how a policer works.
5962 		 *
5963 		 * If it is truely a policer eventually we will come
5964 		 * through and it won't be the maximum.
5965 		 */
5966 		return;
5967 	}
5968 	/* Has enough rounds progressed for us to re-measure? */
5969 	if ((rxt_per >= (uint64_t)rack->r_ctl.policer_rxt_threshold) &&
5970 	    (avg >= rack->r_ctl.policer_avg_threshold) &&
5971 	    (med >= rack->r_ctl.policer_med_threshold)) {
5972 		/*
5973 		 * We hit all thresholds that indicate we are
5974 		 * being policed. Now we may be doing this from a rack timeout
5975 		 * which then means the rest of recovery will hopefully go
5976 		 * smoother as we pace. At the end of recovery we will
5977 		 * fall back in here and reset the values using the
5978 		 * results of the entire recovery episode (we could also
5979 		 * hit this as we exit recovery as well which means only
5980 		 * one time in here).
5981 		 *
5982 		 * This is done explicitly that if we hit the thresholds
5983 		 * again in a second recovery we overwrite the values. We do
5984 		 * that because over time, as we pace the policer_bucket_size may
5985 		 * continue to grow. This then provides more and more times when
5986 		 * we are not pacing to the policer rate. This lets us compensate
5987 		 * for when we hit a false positive and those flows continue to
5988 		 * increase. However if its a real policer we will then get over its
5989 		 * limit, over time, again and thus end up back here hitting the
5990 		 * thresholds again.
5991 		 *
5992 		 * The alternative to this is to instead whenever we pace due to
5993 		 * policing in rack_policed_sending we could add the amount len paced to the
5994 		 * idle_snd_una value (which decreases the amount in last_amount_before_rec
5995 		 * since that is always [th_ack - idle_snd_una]). This would then prevent
5996 		 * the polcier_bucket_size from growing in additional recovery episodes
5997 		 * Which would then mean false  postives would be pretty much stuck
5998 		 * after things got back to normal (assuming that what caused the
5999 		 * false positive was a small network outage).
6000 		 *
6001 		 */
6002 		tcp_trace_point(rack->rc_tp, TCP_TP_POLICER_DET);
6003 		if (rack->rc_policer_detected == 0) {
6004 			/*
6005 			 * Increment the stat that tells us we identified
6006 			 * a policer only once. Note that if we ever allow
6007 			 * the flag to be cleared (reverted) then we need
6008 			 * to adjust this to not do multi-counting.
6009 			 */
6010 			counter_u64_add(tcp_policer_detected, 1);
6011 		}
6012 		rack->r_ctl.last_policer_sndbytes = tp->t_sndbytes;
6013 		rack->r_ctl.last_policer_snd_rxt_bytes = tp->t_snd_rxt_bytes;
6014 		rack->r_ctl.policer_bw = del_bw;
6015 		rack->r_ctl.policer_max_seg = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp,
6016 										  rack->r_ctl.policer_bw,
6017 										  min(ctf_fixed_maxseg(rack->rc_tp),
6018 										      rack->r_ctl.rc_pace_min_segs),
6019 										  0, NULL,
6020 										  NULL, rack->r_ctl.pace_len_divisor);
6021 		/* Now what about the policer bucket size */
6022 		rack->r_ctl.policer_bucket_size = rack->r_ctl.last_amount_before_rec;
6023 		if (rack->r_ctl.policer_bucket_size < rack->r_ctl.policer_max_seg) {
6024 			/* We must be able to send our max-seg or else chaos ensues */
6025 			rack->r_ctl.policer_bucket_size = rack->r_ctl.policer_max_seg * 2;
6026 		}
6027 		if (rack->rc_policer_detected == 0)
6028 			rack->r_ctl.current_policer_bucket = 0;
6029 		if (tcp_bblogging_on(rack->rc_tp)) {
6030 			union tcp_log_stackspecific log;
6031 			struct timeval tv;
6032 
6033 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6034 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6035 			log.u_bbr.flex1 = avg;
6036 			log.u_bbr.flex2 = med;
6037 			log.u_bbr.flex3 = rxt_per;
6038 			log.u_bbr.flex4 = rack->r_ctl.policer_avg_threshold;
6039 			log.u_bbr.flex5 = rack->r_ctl.policer_med_threshold;
6040 			log.u_bbr.flex6 = rack->r_ctl.policer_rxt_threshold;
6041 			log.u_bbr.flex7 = rack->r_ctl.policer_alt_median;
6042 			log.u_bbr.flex8 = 2;
6043 			log.u_bbr.applimited = rack->r_ctl.current_round;
6044 			log.u_bbr.bw_inuse = del_bw;
6045 			log.u_bbr.delivered = rack->r_ctl.policer_bucket_size;
6046 			log.u_bbr.cur_del_rate = rxts;
6047 			log.u_bbr.delRate = snds;
6048 			log.u_bbr.rttProp = rack->r_ctl.gp_bw;
6049 			log.u_bbr.bbr_state = rack->rc_policer_detected;
6050 			log.u_bbr.bbr_substate = 0;
6051 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
6052 			log.u_bbr.use_lt_bw = rack->policer_detect_on;
6053 			log.u_bbr.epoch = rack->r_ctl.policer_max_seg;
6054 			log.u_bbr.lt_epoch = (uint32_t)tim;
6055 			log.u_bbr.pkts_out = rack->r_ctl.bytes_acked_in_recovery;
6056 			tcp_log_event(tp, NULL, NULL, NULL, TCP_POLICER_DET, 0,
6057 				      0, &log, false, NULL, NULL, 0, &tv);
6058 			/*
6059 			 * Put out an added log, 19, for the sole purpose
6060 			 * of getting the txt/rxt so that we can benchmark
6061 			 * in read-bbrlog the ongoing rxt rate after our
6062 			 * policer invocation in the HYSTART announcments.
6063 			 */
6064 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6065 			log.u_bbr.timeStamp = tcp_tv_to_usectick(&tv);
6066 			log.u_bbr.flex1 = alt_med;
6067 			log.u_bbr.flex8 = 19;
6068 			log.u_bbr.cur_del_rate = tp->t_sndbytes;
6069 			log.u_bbr.delRate = tp->t_snd_rxt_bytes;
6070 			tcp_log_event(tp, NULL, NULL, NULL, TCP_POLICER_DET, 0,
6071 				      0, &log, false, NULL, NULL, 0, &tv);
6072 		}
6073 		/* Turn off any fast output, thats ended */
6074 		rack->r_fast_output = 0;
6075 		/* Mark the time for credits */
6076 		rack->r_ctl.last_sendtime = tcp_get_u64_usecs(NULL);
6077 		if (rack->r_rr_config < 2) {
6078 			/*
6079 			 * We need to be stricter on the RR config so
6080 			 * the pacing has priority.
6081 			 */
6082 			rack->r_rr_config = 2;
6083 		}
6084 		policer_detection_log(rack,
6085 				      rack->r_ctl.idle_snd_una,
6086 				      rack->r_ctl.ack_for_idle,
6087 				      0,
6088 				      (uint32_t)tim,
6089 				      14);
6090 		rack->rc_policer_detected = 1;
6091 	} else if ((rack->rc_policer_detected == 1) &&
6092 		   (post_recovery == 1)) {
6093 		/*
6094 		 * If we are exiting recovery and have already detected
6095 		 * we need to possibly update the values.
6096 		 *
6097 		 * First: Update the idle -> recovery sent value.
6098 		 */
6099 		uint32_t srtt;
6100 
6101 		if (rack->r_ctl.last_amount_before_rec > rack->r_ctl.policer_bucket_size) {
6102 			rack->r_ctl.policer_bucket_size = rack->r_ctl.last_amount_before_rec;
6103 		}
6104 		srtt = (uint64_t)rack_grab_rtt(tp, rack);
6105 		if ((tp->t_srtt > 0) && (srtt > tp->t_srtt))
6106 			srtt = tp->t_srtt;
6107 		if ((srtt != 0) &&
6108 		    (tim < (uint64_t)srtt)) {
6109 			/*
6110 			 * Not long enough.
6111 			 */
6112 			if (rack_verbose_logging)
6113 				policer_detection_log(rack,
6114 						      (uint32_t)tim,
6115 						      0,
6116 						      0,
6117 						      0,
6118 						      15);
6119 			return;
6120 		}
6121 		/*
6122 		 * Finally update the b/w if its grown.
6123 		 */
6124 		if (del_bw > rack->r_ctl.policer_bw) {
6125 			rack->r_ctl.policer_bw = del_bw;
6126 			rack->r_ctl.policer_max_seg = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp,
6127 											  rack->r_ctl.policer_bw,
6128 											  min(ctf_fixed_maxseg(rack->rc_tp),
6129 											      rack->r_ctl.rc_pace_min_segs),
6130 											  0, NULL,
6131 											  NULL, rack->r_ctl.pace_len_divisor);
6132 			if (rack->r_ctl.policer_bucket_size < rack->r_ctl.policer_max_seg) {
6133 				/* We must be able to send our max-seg or else chaos ensues */
6134 				rack->r_ctl.policer_bucket_size = rack->r_ctl.policer_max_seg * 2;
6135 			}
6136 		}
6137 		policer_detection_log(rack,
6138 				      rack->r_ctl.idle_snd_una,
6139 				      rack->r_ctl.ack_for_idle,
6140 				      0,
6141 				      (uint32_t)tim,
6142 				      3);
6143 	}
6144 }
6145 
6146 static void
6147 rack_exit_recovery(struct tcpcb *tp, struct tcp_rack *rack, int how)
6148 {
6149 	/* now check with the policer if on */
6150 	if (rack->policer_detect_on == 1) {
6151 		policer_detection(tp, rack, 1);
6152 	}
6153 	/*
6154 	 * Now exit recovery, note we must do the idle set after the policer_detection
6155 	 * to get the amount acked prior to recovery correct.
6156 	 */
6157 	rack->r_ctl.idle_snd_una = tp->snd_una;
6158 	EXIT_RECOVERY(tp->t_flags);
6159 }
6160 
6161 static void
6162 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
6163 {
6164 	struct tcp_rack *rack;
6165 	uint32_t orig_cwnd;
6166 
6167 	orig_cwnd = tp->snd_cwnd;
6168 	INP_WLOCK_ASSERT(tptoinpcb(tp));
6169 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6170 	/* only alert CC if we alerted when we entered */
6171 	if (CC_ALGO(tp)->post_recovery != NULL) {
6172 		tp->t_ccv.curack = th_ack;
6173 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
6174 		if (tp->snd_cwnd < tp->snd_ssthresh) {
6175 			/*
6176 			 * Rack has burst control and pacing
6177 			 * so lets not set this any lower than
6178 			 * snd_ssthresh per RFC-6582 (option 2).
6179 			 */
6180 			tp->snd_cwnd = tp->snd_ssthresh;
6181 		}
6182 	}
6183 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6184 		union tcp_log_stackspecific log;
6185 		struct timeval tv;
6186 
6187 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6188 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6189 		log.u_bbr.flex1 = th_ack;
6190 		log.u_bbr.flex2 = tp->t_ccv.flags;
6191 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
6192 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
6193 		log.u_bbr.flex5 = V_tcp_abc_l_var;
6194 		log.u_bbr.flex6 = orig_cwnd;
6195 		log.u_bbr.flex7 = V_tcp_do_newsack;
6196 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
6197 		log.u_bbr.flex8 = 2;
6198 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
6199 			       0, &log, false, NULL, __func__, __LINE__, &tv);
6200 	}
6201 	if ((rack->rack_no_prr == 0) &&
6202 	    (rack->no_prr_addback == 0) &&
6203 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
6204 		/*
6205 		 * Suck the next prr cnt back into cwnd, but
6206 		 * only do that if we are not application limited.
6207 		 */
6208 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
6209 			/*
6210 			 * We are allowed to add back to the cwnd the amount we did
6211 			 * not get out if:
6212 			 * a) no_prr_addback is off.
6213 			 * b) we are not app limited
6214 			 * c) we are doing prr
6215 			 * <and>
6216 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
6217 			 */
6218 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
6219 					    rack->r_ctl.rc_prr_sndcnt);
6220 		}
6221 		rack->r_ctl.rc_prr_sndcnt = 0;
6222 		rack_log_to_prr(rack, 1, 0, __LINE__);
6223 	}
6224 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
6225 	tp->snd_recover = tp->snd_una;
6226 	if (rack->r_ctl.dsack_persist) {
6227 		rack->r_ctl.dsack_persist--;
6228 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6229 			rack->r_ctl.num_dsack = 0;
6230 		}
6231 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6232 	}
6233 	if (rack->rto_from_rec == 1) {
6234 		rack->rto_from_rec = 0;
6235 		if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
6236 			tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
6237 	}
6238 	rack_exit_recovery(tp, rack, 1);
6239 }
6240 
6241 static void
6242 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
6243 {
6244 	struct tcp_rack *rack;
6245 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
6246 
6247 	INP_WLOCK_ASSERT(tptoinpcb(tp));
6248 #ifdef STATS
6249 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
6250 #endif
6251 	if (IN_RECOVERY(tp->t_flags) == 0) {
6252 		in_rec_at_entry = 0;
6253 		ssthresh_enter = tp->snd_ssthresh;
6254 		cwnd_enter = tp->snd_cwnd;
6255 	} else
6256 		in_rec_at_entry = 1;
6257 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6258 	switch (type) {
6259 	case CC_NDUPACK:
6260 		tp->t_flags &= ~TF_WASFRECOVERY;
6261 		tp->t_flags &= ~TF_WASCRECOVERY;
6262 		if (!IN_FASTRECOVERY(tp->t_flags)) {
6263 			struct rack_sendmap *rsm;
6264 			struct timeval tv;
6265 			uint32_t segsiz;
6266 
6267 			/* Check if this is the end of the initial Start-up i.e. initial slow-start */
6268 			if (rack->rc_initial_ss_comp == 0) {
6269 				/* Yep it is the end of the initial slowstart */
6270 				rack->rc_initial_ss_comp = 1;
6271 			}
6272 			microuptime(&tv);
6273 			rack->r_ctl.time_entered_recovery = tcp_tv_to_lusectick(&tv);
6274 			if (SEQ_GEQ(ack, tp->snd_una)) {
6275 				/*
6276 				 * The ack is above snd_una. Lets see
6277 				 * if we can establish a postive distance from
6278 				 * our idle mark.
6279 				 */
6280 				rack->r_ctl.ack_for_idle = ack;
6281 				if (SEQ_GT(ack, rack->r_ctl.idle_snd_una)) {
6282 					rack->r_ctl.last_amount_before_rec = ack - rack->r_ctl.idle_snd_una;
6283 				} else {
6284 					/* No data thru yet */
6285 					rack->r_ctl.last_amount_before_rec = 0;
6286 				}
6287 			} else if (SEQ_GT(tp->snd_una, rack->r_ctl.idle_snd_una)) {
6288 				/*
6289 				 * The ack is out of order and behind the snd_una. It may
6290 				 * have contained SACK information which we processed else
6291 				 * we would have rejected it.
6292 				 */
6293 				rack->r_ctl.ack_for_idle = tp->snd_una;
6294 				rack->r_ctl.last_amount_before_rec = tp->snd_una - rack->r_ctl.idle_snd_una;
6295 			} else {
6296 				rack->r_ctl.ack_for_idle = ack;
6297 				rack->r_ctl.last_amount_before_rec = 0;
6298 			}
6299 			if (rack->rc_policer_detected) {
6300 				/*
6301 				 * If we are being policed and we have a loss, it
6302 				 * means our bucket is now empty. This can happen
6303 				 * where some other flow on the same host sends
6304 				 * that this connection is not aware of.
6305 				 */
6306 				rack->r_ctl.current_policer_bucket = 0;
6307 				if (rack_verbose_logging)
6308 					policer_detection_log(rack, rack->r_ctl.last_amount_before_rec, 0, 0, 0, 4);
6309 				if (rack->r_ctl.last_amount_before_rec > rack->r_ctl.policer_bucket_size) {
6310 					rack->r_ctl.policer_bucket_size = rack->r_ctl.last_amount_before_rec;
6311 				}
6312 			}
6313 			memset(rack->r_ctl.rc_cnt_of_retran, 0, sizeof(rack->r_ctl.rc_cnt_of_retran));
6314 			segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6315 			TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
6316 				/*
6317 				 * Go through the outstanding and re-peg
6318 				 * any that should have been left in the
6319 				 * retransmit list (on a double recovery).
6320 				 */
6321 				if (rsm->r_act_rxt_cnt > 0) {
6322 					rack_peg_rxt(rack, rsm, segsiz);
6323 				}
6324 			}
6325 			rack->r_ctl.bytes_acked_in_recovery = 0;
6326 			rack->r_ctl.rc_prr_delivered = 0;
6327 			rack->r_ctl.rc_prr_out = 0;
6328 			rack->r_fast_output = 0;
6329 			if (rack->rack_no_prr == 0) {
6330 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
6331 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
6332 			}
6333 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
6334 			tp->snd_recover = tp->snd_max;
6335 			if (tp->t_flags2 & TF2_ECN_PERMIT)
6336 				tp->t_flags2 |= TF2_ECN_SND_CWR;
6337 		}
6338 		break;
6339 	case CC_ECN:
6340 		if (!IN_CONGRECOVERY(tp->t_flags) ||
6341 		    /*
6342 		     * Allow ECN reaction on ACK to CWR, if
6343 		     * that data segment was also CE marked.
6344 		     */
6345 		    SEQ_GEQ(ack, tp->snd_recover)) {
6346 			EXIT_CONGRECOVERY(tp->t_flags);
6347 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
6348 			rack->r_fast_output = 0;
6349 			tp->snd_recover = tp->snd_max + 1;
6350 			if (tp->t_flags2 & TF2_ECN_PERMIT)
6351 				tp->t_flags2 |= TF2_ECN_SND_CWR;
6352 		}
6353 		break;
6354 	case CC_RTO:
6355 		tp->t_dupacks = 0;
6356 		tp->t_bytes_acked = 0;
6357 		rack->r_fast_output = 0;
6358 		if (IN_RECOVERY(tp->t_flags))
6359 			rack_exit_recovery(tp, rack, 2);
6360 		rack->r_ctl.bytes_acked_in_recovery = 0;
6361 		rack->r_ctl.time_entered_recovery = 0;
6362 		orig_cwnd = tp->snd_cwnd;
6363 		rack_log_to_prr(rack, 16, orig_cwnd, line);
6364 		if (CC_ALGO(tp)->cong_signal == NULL) {
6365 			/* TSNH */
6366 			tp->snd_ssthresh = max(2,
6367 			    min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
6368 			    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
6369 			tp->snd_cwnd = ctf_fixed_maxseg(tp);
6370 		}
6371 		if (tp->t_flags2 & TF2_ECN_PERMIT)
6372 			tp->t_flags2 |= TF2_ECN_SND_CWR;
6373 		break;
6374 	case CC_RTO_ERR:
6375 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
6376 		/* RTO was unnecessary, so reset everything. */
6377 		tp->snd_cwnd = tp->snd_cwnd_prev;
6378 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
6379 		tp->snd_recover = tp->snd_recover_prev;
6380 		if (tp->t_flags & TF_WASFRECOVERY) {
6381 			ENTER_FASTRECOVERY(tp->t_flags);
6382 			tp->t_flags &= ~TF_WASFRECOVERY;
6383 		}
6384 		if (tp->t_flags & TF_WASCRECOVERY) {
6385 			ENTER_CONGRECOVERY(tp->t_flags);
6386 			tp->t_flags &= ~TF_WASCRECOVERY;
6387 		}
6388 		tp->snd_nxt = tp->snd_max;
6389 		tp->t_badrxtwin = 0;
6390 		break;
6391 	}
6392 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
6393 	    (type != CC_RTO)){
6394 		tp->t_ccv.curack = ack;
6395 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
6396 	}
6397 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
6398 		rack_log_to_prr(rack, 15, cwnd_enter, line);
6399 		rack->r_ctl.dsack_byte_cnt = 0;
6400 		rack->r_ctl.retran_during_recovery = 0;
6401 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
6402 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
6403 		rack->r_ent_rec_ns = 1;
6404 	}
6405 }
6406 
6407 static inline void
6408 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
6409 {
6410 	uint32_t i_cwnd;
6411 
6412 	INP_WLOCK_ASSERT(tptoinpcb(tp));
6413 
6414 	if (CC_ALGO(tp)->after_idle != NULL)
6415 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
6416 
6417 	if (tp->snd_cwnd == 1)
6418 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
6419 	else
6420 		i_cwnd = rc_init_window(rack);
6421 
6422 	/*
6423 	 * Being idle is no different than the initial window. If the cc
6424 	 * clamps it down below the initial window raise it to the initial
6425 	 * window.
6426 	 */
6427 	if (tp->snd_cwnd < i_cwnd) {
6428 		tp->snd_cwnd = i_cwnd;
6429 	}
6430 }
6431 
6432 /*
6433  * Indicate whether this ack should be delayed.  We can delay the ack if
6434  * following conditions are met:
6435  *	- There is no delayed ack timer in progress.
6436  *	- Our last ack wasn't a 0-sized window. We never want to delay
6437  *	  the ack that opens up a 0-sized window.
6438  *	- LRO wasn't used for this segment. We make sure by checking that the
6439  *	  segment size is not larger than the MSS.
6440  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
6441  *	  connection.
6442  */
6443 #define DELAY_ACK(tp, tlen)			 \
6444 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
6445 	((tp->t_flags & TF_DELACK) == 0) &&	 \
6446 	(tlen <= tp->t_maxseg) &&		 \
6447 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
6448 
6449 static struct rack_sendmap *
6450 rack_find_lowest_rsm(struct tcp_rack *rack)
6451 {
6452 	struct rack_sendmap *rsm;
6453 
6454 	/*
6455 	 * Walk the time-order transmitted list looking for an rsm that is
6456 	 * not acked. This will be the one that was sent the longest time
6457 	 * ago that is still outstanding.
6458 	 */
6459 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
6460 		if (rsm->r_flags & RACK_ACKED) {
6461 			continue;
6462 		}
6463 		goto finish;
6464 	}
6465 finish:
6466 	return (rsm);
6467 }
6468 
6469 static struct rack_sendmap *
6470 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
6471 {
6472 	struct rack_sendmap *prsm;
6473 
6474 	/*
6475 	 * Walk the sequence order list backward until we hit and arrive at
6476 	 * the highest seq not acked. In theory when this is called it
6477 	 * should be the last segment (which it was not).
6478 	 */
6479 	prsm = rsm;
6480 
6481 	TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
6482 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
6483 			continue;
6484 		}
6485 		return (prsm);
6486 	}
6487 	return (NULL);
6488 }
6489 
6490 static uint32_t
6491 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts, int line, int log_allowed)
6492 {
6493 	int32_t lro;
6494 	uint32_t thresh;
6495 
6496 	/*
6497 	 * lro is the flag we use to determine if we have seen reordering.
6498 	 * If it gets set we have seen reordering. The reorder logic either
6499 	 * works in one of two ways:
6500 	 *
6501 	 * If reorder-fade is configured, then we track the last time we saw
6502 	 * re-ordering occur. If we reach the point where enough time as
6503 	 * passed we no longer consider reordering has occuring.
6504 	 *
6505 	 * Or if reorder-face is 0, then once we see reordering we consider
6506 	 * the connection to alway be subject to reordering and just set lro
6507 	 * to 1.
6508 	 *
6509 	 * In the end if lro is non-zero we add the extra time for
6510 	 * reordering in.
6511 	 */
6512 	if (srtt == 0)
6513 		srtt = 1;
6514 	if (rack->r_ctl.rc_reorder_ts) {
6515 		if (rack->r_ctl.rc_reorder_fade) {
6516 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
6517 				lro = cts - rack->r_ctl.rc_reorder_ts;
6518 				if (lro == 0) {
6519 					/*
6520 					 * No time as passed since the last
6521 					 * reorder, mark it as reordering.
6522 					 */
6523 					lro = 1;
6524 				}
6525 			} else {
6526 				/* Negative time? */
6527 				lro = 0;
6528 			}
6529 			if (lro > rack->r_ctl.rc_reorder_fade) {
6530 				/* Turn off reordering seen too */
6531 				rack->r_ctl.rc_reorder_ts = 0;
6532 				lro = 0;
6533 			}
6534 		} else {
6535 			/* Reodering does not fade */
6536 			lro = 1;
6537 		}
6538 	} else {
6539 		lro = 0;
6540 	}
6541 	if (rack->rc_rack_tmr_std_based == 0) {
6542 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
6543 	} else {
6544 		/* Standards based pkt-delay is 1/4 srtt */
6545 		thresh = srtt +  (srtt >> 2);
6546 	}
6547 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
6548 		/* It must be set, if not you get 1/4 rtt */
6549 		if (rack->r_ctl.rc_reorder_shift)
6550 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
6551 		else
6552 			thresh += (srtt >> 2);
6553 	}
6554 	if (rack->rc_rack_use_dsack &&
6555 	    lro &&
6556 	    (rack->r_ctl.num_dsack > 0)) {
6557 		/*
6558 		 * We only increase the reordering window if we
6559 		 * have seen reordering <and> we have a DSACK count.
6560 		 */
6561 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
6562 		if (log_allowed)
6563 			rack_log_dsack_event(rack, 4, line, srtt, thresh);
6564 	}
6565 	/* SRTT * 2 is the ceiling */
6566 	if (thresh > (srtt * 2)) {
6567 		thresh = srtt * 2;
6568 	}
6569 	/* And we don't want it above the RTO max either */
6570 	if (thresh > rack_rto_max) {
6571 		thresh = rack_rto_max;
6572 	}
6573 	if (log_allowed)
6574 		rack_log_dsack_event(rack, 6, line,  srtt, thresh);
6575 	return (thresh);
6576 }
6577 
6578 static uint32_t
6579 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
6580 		     struct rack_sendmap *rsm, uint32_t srtt)
6581 {
6582 	struct rack_sendmap *prsm;
6583 	uint32_t thresh, len;
6584 	int segsiz;
6585 
6586 	if (srtt == 0)
6587 		srtt = 1;
6588 	if (rack->r_ctl.rc_tlp_threshold)
6589 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
6590 	else
6591 		thresh = (srtt * 2);
6592 
6593 	/* Get the previous sent packet, if any */
6594 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6595 	len = rsm->r_end - rsm->r_start;
6596 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
6597 		/* Exactly like the ID */
6598 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6599 			uint32_t alt_thresh;
6600 			/*
6601 			 * Compensate for delayed-ack with the d-ack time.
6602 			 */
6603 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6604 			if (alt_thresh > thresh)
6605 				thresh = alt_thresh;
6606 		}
6607 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6608 		/* 2.1 behavior */
6609 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6610 		if (prsm && (len <= segsiz)) {
6611 			/*
6612 			 * Two packets outstanding, thresh should be (2*srtt) +
6613 			 * possible inter-packet delay (if any).
6614 			 */
6615 			uint32_t inter_gap = 0;
6616 			int idx, nidx;
6617 
6618 			idx = rsm->r_rtr_cnt - 1;
6619 			nidx = prsm->r_rtr_cnt - 1;
6620 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6621 				/* Yes it was sent later (or at the same time) */
6622 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6623 			}
6624 			thresh += inter_gap;
6625 		} else if (len <= segsiz) {
6626 			/*
6627 			 * Possibly compensate for delayed-ack.
6628 			 */
6629 			uint32_t alt_thresh;
6630 
6631 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6632 			if (alt_thresh > thresh)
6633 				thresh = alt_thresh;
6634 		}
6635 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6636 		/* 2.2 behavior */
6637 		if (len <= segsiz) {
6638 			uint32_t alt_thresh;
6639 			/*
6640 			 * Compensate for delayed-ack with the d-ack time.
6641 			 */
6642 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6643 			if (alt_thresh > thresh)
6644 				thresh = alt_thresh;
6645 		}
6646 	}
6647 	/* Not above an RTO */
6648 	if (thresh > tp->t_rxtcur) {
6649 		thresh = tp->t_rxtcur;
6650 	}
6651 	/* Not above a RTO max */
6652 	if (thresh > rack_rto_max) {
6653 		thresh = rack_rto_max;
6654 	}
6655 	/* Apply user supplied min TLP */
6656 	if (thresh < rack_tlp_min) {
6657 		thresh = rack_tlp_min;
6658 	}
6659 	return (thresh);
6660 }
6661 
6662 static uint32_t
6663 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6664 {
6665 	/*
6666 	 * We want the rack_rtt which is the
6667 	 * last rtt we measured. However if that
6668 	 * does not exist we fallback to the srtt (which
6669 	 * we probably will never do) and then as a last
6670 	 * resort we use RACK_INITIAL_RTO if no srtt is
6671 	 * yet set.
6672 	 */
6673 	if (rack->rc_rack_rtt)
6674 		return (rack->rc_rack_rtt);
6675 	else if (tp->t_srtt == 0)
6676 		return (RACK_INITIAL_RTO);
6677 	return (tp->t_srtt);
6678 }
6679 
6680 static struct rack_sendmap *
6681 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6682 {
6683 	/*
6684 	 * Check to see that we don't need to fall into recovery. We will
6685 	 * need to do so if our oldest transmit is past the time we should
6686 	 * have had an ack.
6687 	 */
6688 	struct tcp_rack *rack;
6689 	struct rack_sendmap *rsm;
6690 	int32_t idx;
6691 	uint32_t srtt, thresh;
6692 
6693 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6694 	if (tqhash_empty(rack->r_ctl.tqh)) {
6695 		return (NULL);
6696 	}
6697 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6698 	if (rsm == NULL)
6699 		return (NULL);
6700 
6701 
6702 	if (rsm->r_flags & RACK_ACKED) {
6703 		rsm = rack_find_lowest_rsm(rack);
6704 		if (rsm == NULL)
6705 			return (NULL);
6706 	}
6707 	idx = rsm->r_rtr_cnt - 1;
6708 	srtt = rack_grab_rtt(tp, rack);
6709 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
6710 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6711 		return (NULL);
6712 	}
6713 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6714 		return (NULL);
6715 	}
6716 	/* Ok if we reach here we are over-due and this guy can be sent */
6717 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6718 	return (rsm);
6719 }
6720 
6721 static uint32_t
6722 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6723 {
6724 	int32_t t;
6725 	int32_t tt;
6726 	uint32_t ret_val;
6727 
6728 	t = (tp->t_srtt + (tp->t_rttvar << 2));
6729 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6730  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6731 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6732 	ret_val = (uint32_t)tt;
6733 	return (ret_val);
6734 }
6735 
6736 static uint32_t
6737 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6738 {
6739 	/*
6740 	 * Start the FR timer, we do this based on getting the first one in
6741 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6742 	 * events we need to stop the running timer (if its running) before
6743 	 * starting the new one.
6744 	 */
6745 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6746 	uint32_t srtt_cur;
6747 	int32_t idx;
6748 	int32_t is_tlp_timer = 0;
6749 	struct rack_sendmap *rsm;
6750 
6751 	if (rack->t_timers_stopped) {
6752 		/* All timers have been stopped none are to run */
6753 		return (0);
6754 	}
6755 	if (rack->rc_in_persist) {
6756 		/* We can't start any timer in persists */
6757 		return (rack_get_persists_timer_val(tp, rack));
6758 	}
6759 	rack->rc_on_min_to = 0;
6760 	if ((tp->t_state < TCPS_ESTABLISHED) ||
6761 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6762 		goto activate_rxt;
6763 	}
6764 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6765 	if ((rsm == NULL) || sup_rack) {
6766 		/* Nothing on the send map or no rack */
6767 activate_rxt:
6768 		time_since_sent = 0;
6769 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6770 		if (rsm) {
6771 			/*
6772 			 * Should we discount the RTX timer any?
6773 			 *
6774 			 * We want to discount it the smallest amount.
6775 			 * If a timer (Rack/TLP or RXT) has gone off more
6776 			 * recently thats the discount we want to use (now - timer time).
6777 			 * If the retransmit of the oldest packet was more recent then
6778 			 * we want to use that (now - oldest-packet-last_transmit_time).
6779 			 *
6780 			 */
6781 			idx = rsm->r_rtr_cnt - 1;
6782 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6783 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6784 			else
6785 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6786 			if (TSTMP_GT(cts, tstmp_touse))
6787 			    time_since_sent = cts - tstmp_touse;
6788 		}
6789 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6790 		    sbavail(&tptosocket(tp)->so_snd)) {
6791 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6792 			to = tp->t_rxtcur;
6793 			if (to > time_since_sent)
6794 				to -= time_since_sent;
6795 			else
6796 				to = rack->r_ctl.rc_min_to;
6797 			if (to == 0)
6798 				to = 1;
6799 			/* Special case for KEEPINIT */
6800 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6801 			    (TP_KEEPINIT(tp) != 0) &&
6802 			    rsm) {
6803 				/*
6804 				 * We have to put a ceiling on the rxt timer
6805 				 * of the keep-init timeout.
6806 				 */
6807 				uint32_t max_time, red;
6808 
6809 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6810 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6811 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6812 					if (red < max_time)
6813 						max_time -= red;
6814 					else
6815 						max_time = 1;
6816 				}
6817 				/* Reduce timeout to the keep value if needed */
6818 				if (max_time < to)
6819 					to = max_time;
6820 			}
6821 			return (to);
6822 		}
6823 		return (0);
6824 	}
6825 	if (rsm->r_flags & RACK_ACKED) {
6826 		rsm = rack_find_lowest_rsm(rack);
6827 		if (rsm == NULL) {
6828 			/* No lowest? */
6829 			goto activate_rxt;
6830 		}
6831 	}
6832 	/* Convert from ms to usecs */
6833 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
6834 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6835 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6836 		if ((tp->t_flags & TF_SENTFIN) &&
6837 		    ((tp->snd_max - tp->snd_una) == 1) &&
6838 		    (rsm->r_flags & RACK_HAS_FIN)) {
6839 			/*
6840 			 * We don't start a rack timer if all we have is a
6841 			 * FIN outstanding.
6842 			 */
6843 			goto activate_rxt;
6844 		}
6845 		if ((rack->use_rack_rr == 0) &&
6846 		    (IN_FASTRECOVERY(tp->t_flags)) &&
6847 		    (rack->rack_no_prr == 0) &&
6848 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
6849 			/*
6850 			 * We are not cheating, in recovery  and
6851 			 * not enough ack's to yet get our next
6852 			 * retransmission out.
6853 			 *
6854 			 * Note that classified attackers do not
6855 			 * get to use the rack-cheat.
6856 			 */
6857 			goto activate_tlp;
6858 		}
6859 		srtt = rack_grab_rtt(tp, rack);
6860 		thresh = rack_calc_thresh_rack(rack, srtt, cts, __LINE__, 1);
6861 		idx = rsm->r_rtr_cnt - 1;
6862 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6863 		if (SEQ_GEQ(exp, cts)) {
6864 			to = exp - cts;
6865 			if (to < rack->r_ctl.rc_min_to) {
6866 				to = rack->r_ctl.rc_min_to;
6867 				if (rack->r_rr_config == 3)
6868 					rack->rc_on_min_to = 1;
6869 			}
6870 		} else {
6871 			to = rack->r_ctl.rc_min_to;
6872 			if (rack->r_rr_config == 3)
6873 				rack->rc_on_min_to = 1;
6874 		}
6875 	} else {
6876 		/* Ok we need to do a TLP not RACK */
6877 activate_tlp:
6878 		if ((rack->rc_tlp_in_progress != 0) &&
6879 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6880 			/*
6881 			 * The previous send was a TLP and we have sent
6882 			 * N TLP's without sending new data.
6883 			 */
6884 			goto activate_rxt;
6885 		}
6886 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6887 		if (rsm == NULL) {
6888 			/* We found no rsm to TLP with. */
6889 			goto activate_rxt;
6890 		}
6891 		if (rsm->r_flags & RACK_HAS_FIN) {
6892 			/* If its a FIN we dont do TLP */
6893 			rsm = NULL;
6894 			goto activate_rxt;
6895 		}
6896 		idx = rsm->r_rtr_cnt - 1;
6897 		time_since_sent = 0;
6898 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6899 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6900 		else
6901 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6902 		if (TSTMP_GT(cts, tstmp_touse))
6903 		    time_since_sent = cts - tstmp_touse;
6904 		is_tlp_timer = 1;
6905 		if (tp->t_srtt) {
6906 			if ((rack->rc_srtt_measure_made == 0) &&
6907 			    (tp->t_srtt == 1)) {
6908 				/*
6909 				 * If another stack as run and set srtt to 1,
6910 				 * then the srtt was 0, so lets use the initial.
6911 				 */
6912 				srtt = RACK_INITIAL_RTO;
6913 			} else {
6914 				srtt_cur = tp->t_srtt;
6915 				srtt = srtt_cur;
6916 			}
6917 		} else
6918 			srtt = RACK_INITIAL_RTO;
6919 		/*
6920 		 * If the SRTT is not keeping up and the
6921 		 * rack RTT has spiked we want to use
6922 		 * the last RTT not the smoothed one.
6923 		 */
6924 		if (rack_tlp_use_greater &&
6925 		    tp->t_srtt &&
6926 		    (srtt < rack_grab_rtt(tp, rack))) {
6927 			srtt = rack_grab_rtt(tp, rack);
6928 		}
6929 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6930 		if (thresh > time_since_sent) {
6931 			to = thresh - time_since_sent;
6932 		} else {
6933 			to = rack->r_ctl.rc_min_to;
6934 			rack_log_alt_to_to_cancel(rack,
6935 						  thresh,		/* flex1 */
6936 						  time_since_sent,	/* flex2 */
6937 						  tstmp_touse,		/* flex3 */
6938 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6939 						  (uint32_t)rsm->r_tim_lastsent[idx],
6940 						  srtt,
6941 						  idx, 99);
6942 		}
6943 		if (to < rack_tlp_min) {
6944 			to = rack_tlp_min;
6945 		}
6946 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
6947 			/*
6948 			 * If the TLP time works out to larger than the max
6949 			 * RTO lets not do TLP.. just RTO.
6950 			 */
6951 			goto activate_rxt;
6952 		}
6953 	}
6954 	if (is_tlp_timer == 0) {
6955 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6956 	} else {
6957 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6958 	}
6959 	if (to == 0)
6960 		to = 1;
6961 	return (to);
6962 }
6963 
6964 static void
6965 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6966 {
6967 	if (rack->rc_in_persist == 0) {
6968 		if (tp->t_flags & TF_GPUTINPROG) {
6969 			/*
6970 			 * Stop the goodput now, the calling of the
6971 			 * measurement function clears the flag.
6972 			 */
6973 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6974 						    RACK_QUALITY_PERSIST);
6975 		}
6976 #ifdef NETFLIX_SHARED_CWND
6977 		if (rack->r_ctl.rc_scw) {
6978 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6979 			rack->rack_scwnd_is_idle = 1;
6980 		}
6981 #endif
6982 		rack->r_ctl.rc_went_idle_time = cts;
6983 		if (rack->r_ctl.rc_went_idle_time == 0)
6984 			rack->r_ctl.rc_went_idle_time = 1;
6985 		if (rack->lt_bw_up) {
6986 			/* Suspend our LT BW measurement */
6987 			uint64_t tmark;
6988 
6989 			rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6990 			rack->r_ctl.lt_seq = snd_una;
6991 			tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
6992 			if (tmark >= rack->r_ctl.lt_timemark) {
6993 				rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6994 			}
6995 			rack->r_ctl.lt_timemark = tmark;
6996 			rack->lt_bw_up = 0;
6997 			rack->r_persist_lt_bw_off = 1;
6998 		}
6999 		rack_timer_cancel(tp, rack, cts, __LINE__);
7000 		rack->r_ctl.persist_lost_ends = 0;
7001 		rack->probe_not_answered = 0;
7002 		rack->forced_ack = 0;
7003 		tp->t_rxtshift = 0;
7004 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7005 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7006 		rack->rc_in_persist = 1;
7007 	}
7008 }
7009 
7010 static void
7011 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7012 {
7013 	if (tcp_in_hpts(rack->rc_tp)) {
7014 		tcp_hpts_remove(rack->rc_tp);
7015 		rack->r_ctl.rc_hpts_flags = 0;
7016 	}
7017 #ifdef NETFLIX_SHARED_CWND
7018 	if (rack->r_ctl.rc_scw) {
7019 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
7020 		rack->rack_scwnd_is_idle = 0;
7021 	}
7022 #endif
7023 	if (rack->rc_gp_dyn_mul &&
7024 	    (rack->use_fixed_rate == 0) &&
7025 	    (rack->rc_always_pace)) {
7026 		/*
7027 		 * Do we count this as if a probe-rtt just
7028 		 * finished?
7029 		 */
7030 		uint32_t time_idle, idle_min;
7031 
7032 		time_idle = cts - rack->r_ctl.rc_went_idle_time;
7033 		idle_min = rack_min_probertt_hold;
7034 		if (rack_probertt_gpsrtt_cnt_div) {
7035 			uint64_t extra;
7036 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
7037 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
7038 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
7039 			idle_min += (uint32_t)extra;
7040 		}
7041 		if (time_idle >= idle_min) {
7042 			/* Yes, we count it as a probe-rtt. */
7043 			uint32_t us_cts;
7044 
7045 			us_cts = tcp_get_usecs(NULL);
7046 			if (rack->in_probe_rtt == 0) {
7047 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7048 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
7049 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
7050 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
7051 			} else {
7052 				rack_exit_probertt(rack, us_cts);
7053 			}
7054 		}
7055 	}
7056 	if (rack->r_persist_lt_bw_off) {
7057 		/* Continue where we left off */
7058 		rack->r_ctl.lt_timemark = tcp_get_u64_usecs(NULL);
7059 		rack->lt_bw_up = 1;
7060 		rack->r_persist_lt_bw_off = 0;
7061 	}
7062 	rack->r_ctl.idle_snd_una = tp->snd_una;
7063 	rack->rc_in_persist = 0;
7064 	rack->r_ctl.rc_went_idle_time = 0;
7065 	tp->t_rxtshift = 0;
7066 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7067 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7068 	rack->r_ctl.rc_agg_delayed = 0;
7069 	rack->r_early = 0;
7070 	rack->r_late = 0;
7071 	rack->r_ctl.rc_agg_early = 0;
7072 }
7073 
7074 static void
7075 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
7076 		   struct hpts_diag *diag, struct timeval *tv)
7077 {
7078 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
7079 		union tcp_log_stackspecific log;
7080 
7081 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
7082 		log.u_bbr.flex1 = diag->p_nxt_slot;
7083 		log.u_bbr.flex2 = diag->p_cur_slot;
7084 		log.u_bbr.flex3 = diag->slot_req;
7085 		log.u_bbr.flex4 = diag->inp_hptsslot;
7086 		log.u_bbr.flex5 = diag->slot_remaining;
7087 		log.u_bbr.flex6 = diag->need_new_to;
7088 		log.u_bbr.flex7 = diag->p_hpts_active;
7089 		log.u_bbr.flex8 = diag->p_on_min_sleep;
7090 		/* Hijack other fields as needed */
7091 		log.u_bbr.epoch = diag->have_slept;
7092 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
7093 		log.u_bbr.pkts_out = diag->co_ret;
7094 		log.u_bbr.applimited = diag->hpts_sleep_time;
7095 		log.u_bbr.delivered = diag->p_prev_slot;
7096 		log.u_bbr.inflight = diag->p_runningslot;
7097 		log.u_bbr.bw_inuse = diag->wheel_slot;
7098 		log.u_bbr.rttProp = diag->wheel_cts;
7099 		log.u_bbr.timeStamp = cts;
7100 		log.u_bbr.delRate = diag->maxslots;
7101 		log.u_bbr.cur_del_rate = diag->p_curtick;
7102 		log.u_bbr.cur_del_rate <<= 32;
7103 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
7104 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
7105 		    &rack->rc_inp->inp_socket->so_rcv,
7106 		    &rack->rc_inp->inp_socket->so_snd,
7107 		    BBR_LOG_HPTSDIAG, 0,
7108 		    0, &log, false, tv);
7109 	}
7110 
7111 }
7112 
7113 static void
7114 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
7115 {
7116 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
7117 		union tcp_log_stackspecific log;
7118 		struct timeval tv;
7119 
7120 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
7121 		log.u_bbr.flex1 = sb->sb_flags;
7122 		log.u_bbr.flex2 = len;
7123 		log.u_bbr.flex3 = sb->sb_state;
7124 		log.u_bbr.flex8 = type;
7125 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
7126 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
7127 		    &rack->rc_inp->inp_socket->so_rcv,
7128 		    &rack->rc_inp->inp_socket->so_snd,
7129 		    TCP_LOG_SB_WAKE, 0,
7130 		    len, &log, false, &tv);
7131 	}
7132 }
7133 
7134 static void
7135 rack_start_hpts_timer (struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
7136       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
7137 {
7138 	struct hpts_diag diag;
7139 	struct inpcb *inp = tptoinpcb(tp);
7140 	struct timeval tv;
7141 	uint32_t delayed_ack = 0;
7142 	uint32_t hpts_timeout;
7143 	uint32_t entry_slot = slot;
7144 	uint8_t stopped;
7145 	uint32_t left = 0;
7146 	uint32_t us_cts;
7147 
7148 	if ((tp->t_state == TCPS_CLOSED) ||
7149 	    (tp->t_state == TCPS_LISTEN)) {
7150 		return;
7151 	}
7152 	if (tcp_in_hpts(tp)) {
7153 		/* Already on the pacer */
7154 		return;
7155 	}
7156 	stopped = rack->rc_tmr_stopped;
7157 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
7158 		left = rack->r_ctl.rc_timer_exp - cts;
7159 	}
7160 	rack->r_ctl.rc_timer_exp = 0;
7161 	rack->r_ctl.rc_hpts_flags = 0;
7162 	us_cts = tcp_get_usecs(&tv);
7163 	/* Now early/late accounting */
7164 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
7165 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
7166 		/*
7167 		 * We have a early carry over set,
7168 		 * we can always add more time so we
7169 		 * can always make this compensation.
7170 		 *
7171 		 * Note if ack's are allowed to wake us do not
7172 		 * penalize the next timer for being awoke
7173 		 * by an ack aka the rc_agg_early (non-paced mode).
7174 		 */
7175 		slot += rack->r_ctl.rc_agg_early;
7176 		rack->r_early = 0;
7177 		rack->r_ctl.rc_agg_early = 0;
7178 	}
7179 	if ((rack->r_late) &&
7180 	    ((rack->r_use_hpts_min == 0) || (rack->dgp_on == 0))) {
7181 		/*
7182 		 * This is harder, we can
7183 		 * compensate some but it
7184 		 * really depends on what
7185 		 * the current pacing time is.
7186 		 */
7187 		if (rack->r_ctl.rc_agg_delayed >= slot) {
7188 			/*
7189 			 * We can't compensate for it all.
7190 			 * And we have to have some time
7191 			 * on the clock. We always have a min
7192 			 * 10 slots (10 x 10 i.e. 100 usecs).
7193 			 */
7194 			if (slot <= HPTS_TICKS_PER_SLOT) {
7195 				/* We gain delay */
7196 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
7197 				slot = HPTS_TICKS_PER_SLOT;
7198 			} else {
7199 				/* We take off some */
7200 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
7201 				slot = HPTS_TICKS_PER_SLOT;
7202 			}
7203 		} else {
7204 			slot -= rack->r_ctl.rc_agg_delayed;
7205 			rack->r_ctl.rc_agg_delayed = 0;
7206 			/* Make sure we have 100 useconds at minimum */
7207 			if (slot < HPTS_TICKS_PER_SLOT) {
7208 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
7209 				slot = HPTS_TICKS_PER_SLOT;
7210 			}
7211 			if (rack->r_ctl.rc_agg_delayed == 0)
7212 				rack->r_late = 0;
7213 		}
7214 	} else if (rack->r_late) {
7215 		/* r_use_hpts_min is on and so is DGP */
7216 		uint32_t max_red;
7217 
7218 		max_red = (slot * rack->r_ctl.max_reduction) / 100;
7219 		if (max_red >= rack->r_ctl.rc_agg_delayed) {
7220 			slot -= rack->r_ctl.rc_agg_delayed;
7221 			rack->r_ctl.rc_agg_delayed = 0;
7222 		} else {
7223 			slot -= max_red;
7224 			rack->r_ctl.rc_agg_delayed -= max_red;
7225 		}
7226 	}
7227 	if ((rack->r_use_hpts_min == 1) &&
7228 	    (slot > 0) &&
7229 	    (rack->dgp_on == 1)) {
7230 		/*
7231 		 * We are enforcing a min pacing timer
7232 		 * based on our hpts min timeout.
7233 		 */
7234 		uint32_t min;
7235 
7236 		min = get_hpts_min_sleep_time();
7237 		if (min > slot) {
7238 			slot = min;
7239 		}
7240 	}
7241 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
7242 	if (tp->t_flags & TF_DELACK) {
7243 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
7244 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
7245 	}
7246 	if (delayed_ack && ((hpts_timeout == 0) ||
7247 			    (delayed_ack < hpts_timeout)))
7248 		hpts_timeout = delayed_ack;
7249 	else
7250 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7251 	/*
7252 	 * If no timers are going to run and we will fall off the hptsi
7253 	 * wheel, we resort to a keep-alive timer if its configured.
7254 	 */
7255 	if ((hpts_timeout == 0) &&
7256 	    (slot == 0)) {
7257 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7258 		    (tp->t_state <= TCPS_CLOSING)) {
7259 			/*
7260 			 * Ok we have no timer (persists, rack, tlp, rxt  or
7261 			 * del-ack), we don't have segments being paced. So
7262 			 * all that is left is the keepalive timer.
7263 			 */
7264 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7265 				/* Get the established keep-alive time */
7266 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
7267 			} else {
7268 				/*
7269 				 * Get the initial setup keep-alive time,
7270 				 * note that this is probably not going to
7271 				 * happen, since rack will be running a rxt timer
7272 				 * if a SYN of some sort is outstanding. It is
7273 				 * actually handled in rack_timeout_rxt().
7274 				 */
7275 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
7276 			}
7277 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
7278 			if (rack->in_probe_rtt) {
7279 				/*
7280 				 * We want to instead not wake up a long time from
7281 				 * now but to wake up about the time we would
7282 				 * exit probe-rtt and initiate a keep-alive ack.
7283 				 * This will get us out of probe-rtt and update
7284 				 * our min-rtt.
7285 				 */
7286 				hpts_timeout = rack_min_probertt_hold;
7287 			}
7288 		}
7289 	}
7290 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
7291 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
7292 		/*
7293 		 * RACK, TLP, persists and RXT timers all are restartable
7294 		 * based on actions input .. i.e we received a packet (ack
7295 		 * or sack) and that changes things (rw, or snd_una etc).
7296 		 * Thus we can restart them with a new value. For
7297 		 * keep-alive, delayed_ack we keep track of what was left
7298 		 * and restart the timer with a smaller value.
7299 		 */
7300 		if (left < hpts_timeout)
7301 			hpts_timeout = left;
7302 	}
7303 	if (hpts_timeout) {
7304 		/*
7305 		 * Hack alert for now we can't time-out over 2,147,483
7306 		 * seconds (a bit more than 596 hours), which is probably ok
7307 		 * :).
7308 		 */
7309 		if (hpts_timeout > 0x7ffffffe)
7310 			hpts_timeout = 0x7ffffffe;
7311 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
7312 	}
7313 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
7314 	if ((rack->gp_ready == 0) &&
7315 	    (rack->use_fixed_rate == 0) &&
7316 	    (hpts_timeout < slot) &&
7317 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
7318 		/*
7319 		 * We have no good estimate yet for the
7320 		 * old clunky burst mitigation or the
7321 		 * real pacing. And the tlp or rxt is smaller
7322 		 * than the pacing calculation. Lets not
7323 		 * pace that long since we know the calculation
7324 		 * so far is not accurate.
7325 		 */
7326 		slot = hpts_timeout;
7327 	}
7328 	/**
7329 	 * Turn off all the flags for queuing by default. The
7330 	 * flags have important meanings to what happens when
7331 	 * LRO interacts with the transport. Most likely (by default now)
7332 	 * mbuf_queueing and ack compression are on. So the transport
7333 	 * has a couple of flags that control what happens (if those
7334 	 * are not on then these flags won't have any effect since it
7335 	 * won't go through the queuing LRO path).
7336 	 *
7337 	 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
7338 	 *                        pacing output, so don't disturb. But
7339 	 *                        it also means LRO can wake me if there
7340 	 *                        is a SACK arrival.
7341 	 *
7342 	 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
7343 	 *                       with the above flag (QUEUE_READY) and
7344 	 *                       when present it says don't even wake me
7345 	 *                       if a SACK arrives.
7346 	 *
7347 	 * The idea behind these flags is that if we are pacing we
7348 	 * set the MBUF_QUEUE_READY and only get woken up if
7349 	 * a SACK arrives (which could change things) or if
7350 	 * our pacing timer expires. If, however, we have a rack
7351 	 * timer running, then we don't even want a sack to wake
7352 	 * us since the rack timer has to expire before we can send.
7353 	 *
7354 	 * Other cases should usually have none of the flags set
7355 	 * so LRO can call into us.
7356 	 */
7357 	tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
7358 	if (slot) {
7359 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
7360 		rack->r_ctl.rc_last_output_to = us_cts + slot;
7361 		/*
7362 		 * A pacing timer (slot) is being set, in
7363 		 * such a case we cannot send (we are blocked by
7364 		 * the timer). So lets tell LRO that it should not
7365 		 * wake us unless there is a SACK. Note this only
7366 		 * will be effective if mbuf queueing is on or
7367 		 * compressed acks are being processed.
7368 		 */
7369 		tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
7370 		/*
7371 		 * But wait if we have a Rack timer running
7372 		 * even a SACK should not disturb us (with
7373 		 * the exception of r_rr_config 3).
7374 		 */
7375 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
7376 		    (IN_RECOVERY(tp->t_flags))) {
7377 			if (rack->r_rr_config != 3)
7378 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
7379 			else if (rack->rc_pace_dnd) {
7380 				/*
7381 				 * When DND is on, we only let a sack
7382 				 * interrupt us if we are not in recovery.
7383 				 *
7384 				 * If DND is off, then we never hit here
7385 				 * and let all sacks wake us up.
7386 				 *
7387 				 */
7388 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
7389 			}
7390 		}
7391 		if (rack->rc_ack_can_sendout_data) {
7392 			/*
7393 			 * Ahh but wait, this is that special case
7394 			 * where the pacing timer can be disturbed
7395 			 * backout the changes (used for non-paced
7396 			 * burst limiting).
7397 			 */
7398 			tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
7399 			    TF2_MBUF_QUEUE_READY);
7400 		}
7401 		if ((rack->use_rack_rr) &&
7402 		    (rack->r_rr_config < 2) &&
7403 		    ((hpts_timeout) && (hpts_timeout < slot))) {
7404 			/*
7405 			 * Arrange for the hpts to kick back in after the
7406 			 * t-o if the t-o does not cause a send.
7407 			 */
7408 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7409 						   __LINE__, &diag);
7410 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7411 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7412 		} else {
7413 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
7414 						   __LINE__, &diag);
7415 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7416 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
7417 		}
7418 	} else if (hpts_timeout) {
7419 		/*
7420 		 * With respect to t_flags2(?) here, lets let any new acks wake
7421 		 * us up here. Since we are not pacing (no pacing timer), output
7422 		 * can happen so we should let it. If its a Rack timer, then any inbound
7423 		 * packet probably won't change the sending (we will be blocked)
7424 		 * but it may change the prr stats so letting it in (the set defaults
7425 		 * at the start of this block) are good enough.
7426 		 */
7427 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7428 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7429 					   __LINE__, &diag);
7430 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7431 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7432 	} else {
7433 		/* No timer starting */
7434 #ifdef INVARIANTS
7435 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
7436 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
7437 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
7438 		}
7439 #endif
7440 	}
7441 	rack->rc_tmr_stopped = 0;
7442 	if (slot)
7443 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
7444 }
7445 
7446 static void
7447 rack_mark_lost(struct tcpcb *tp,
7448     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
7449 {
7450 	struct rack_sendmap *nrsm;
7451 	uint32_t thresh,  exp;
7452 
7453 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
7454 	nrsm = rsm;
7455 	TAILQ_FOREACH_FROM(nrsm, &rack->r_ctl.rc_tmap, r_tnext) {
7456 		if ((nrsm->r_flags & RACK_SACK_PASSED) == 0) {
7457 			/* Got up to all that were marked sack-passed */
7458 			break;
7459 		}
7460 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
7461 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
7462 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
7463 				/* We now consider it lost */
7464 				nrsm->r_flags |= RACK_WAS_LOST;
7465 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
7466 			} else {
7467 				/* Past here it won't be lost so stop */
7468 				break;
7469 			}
7470 		}
7471 	}
7472 }
7473 
7474 /*
7475  * RACK Timer, here we simply do logging and house keeping.
7476  * the normal rack_output() function will call the
7477  * appropriate thing to check if we need to do a RACK retransmit.
7478  * We return 1, saying don't proceed with rack_output only
7479  * when all timers have been stopped (destroyed PCB?).
7480  */
7481 static int
7482 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7483 {
7484 	/*
7485 	 * This timer simply provides an internal trigger to send out data.
7486 	 * The check_recovery_mode call will see if there are needed
7487 	 * retransmissions, if so we will enter fast-recovery. The output
7488 	 * call may or may not do the same thing depending on sysctl
7489 	 * settings.
7490 	 */
7491 	struct rack_sendmap *rsm;
7492 
7493 	counter_u64_add(rack_to_tot, 1);
7494 	if (rack->r_state && (rack->r_state != tp->t_state))
7495 		rack_set_state(tp, rack);
7496 	rack->rc_on_min_to = 0;
7497 	rsm = rack_check_recovery_mode(tp, cts);
7498 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
7499 	if (rsm) {
7500 		/* We need to stroke any lost that are now declared as lost */
7501 		rack_mark_lost(tp, rack, rsm, cts);
7502 		rack->r_ctl.rc_resend = rsm;
7503 		rack->r_timer_override = 1;
7504 		if (rack->use_rack_rr) {
7505 			/*
7506 			 * Don't accumulate extra pacing delay
7507 			 * we are allowing the rack timer to
7508 			 * over-ride pacing i.e. rrr takes precedence
7509 			 * if the pacing interval is longer than the rrr
7510 			 * time (in other words we get the min pacing
7511 			 * time versus rrr pacing time).
7512 			 */
7513 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7514 		}
7515 	}
7516 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
7517 	if (rsm == NULL) {
7518 		/* restart a timer and return 1 */
7519 		rack_start_hpts_timer(rack, tp, cts,
7520 				      0, 0, 0);
7521 		return (1);
7522 	}
7523 	if ((rack->policer_detect_on == 1) &&
7524 	    (rack->rc_policer_detected == 0)) {
7525 		/*
7526 		 * We do this early if we have not
7527 		 * deteceted to attempt to detect
7528 		 * quicker. Normally we want to do this
7529 		 * as recovery exits (and we will again).
7530 		 */
7531 		policer_detection(tp, rack, 0);
7532 	}
7533 	return (0);
7534 }
7535 
7536 
7537 
7538 static void
7539 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
7540 {
7541 
7542 	if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
7543 		/*
7544 		 * The trailing space changed, mbufs can grow
7545 		 * at the tail but they can't shrink from
7546 		 * it, KASSERT that. Adjust the orig_m_len to
7547 		 * compensate for this change.
7548 		 */
7549 		KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
7550 			("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
7551 			 rsm->m,
7552 			 rsm,
7553 			 (intmax_t)M_TRAILINGROOM(rsm->m),
7554 			 rsm->orig_t_space,
7555 			 rsm->orig_m_len,
7556 			 rsm->m->m_len));
7557 		rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
7558 		rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7559 	}
7560 	if (rsm->m->m_len < rsm->orig_m_len) {
7561 		/*
7562 		 * Mbuf shrank, trimmed off the top by an ack, our
7563 		 * offset changes.
7564 		 */
7565 		KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
7566 			("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
7567 			 rsm->m, rsm->m->m_len,
7568 			 rsm, rsm->orig_m_len,
7569 			 rsm->soff));
7570 		if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
7571 			rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
7572 		else
7573 			rsm->soff = 0;
7574 		rsm->orig_m_len = rsm->m->m_len;
7575 #ifdef INVARIANTS
7576 	} else if (rsm->m->m_len > rsm->orig_m_len) {
7577 		panic("rsm:%p m:%p m_len grew outside of t_space compensation",
7578 		      rsm, rsm->m);
7579 #endif
7580 	}
7581 }
7582 
7583 static void
7584 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
7585 {
7586 	struct mbuf *m;
7587 	uint32_t soff;
7588 
7589 	if (src_rsm->m &&
7590 	    ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
7591 	     (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
7592 		/* Fix up the orig_m_len and possibly the mbuf offset */
7593 		rack_adjust_orig_mlen(src_rsm);
7594 	}
7595 	m = src_rsm->m;
7596 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
7597 	while (soff >= m->m_len) {
7598 		/* Move out past this mbuf */
7599 		soff -= m->m_len;
7600 		m = m->m_next;
7601 		KASSERT((m != NULL),
7602 			("rsm:%p nrsm:%p hit at soff:%u null m",
7603 			 src_rsm, rsm, soff));
7604 		if (m == NULL) {
7605 			/* This should *not* happen which is why there is a kassert */
7606 			src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7607 					       (src_rsm->r_start - rack->rc_tp->snd_una),
7608 					       &src_rsm->soff);
7609 			src_rsm->orig_m_len = src_rsm->m->m_len;
7610 			src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7611 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7612 					   (rsm->r_start - rack->rc_tp->snd_una),
7613 					   &rsm->soff);
7614 			rsm->orig_m_len = rsm->m->m_len;
7615 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7616 			return;
7617 		}
7618 	}
7619 	rsm->m = m;
7620 	rsm->soff = soff;
7621 	rsm->orig_m_len = m->m_len;
7622 	rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7623 }
7624 
7625 static __inline void
7626 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7627 	       struct rack_sendmap *rsm, uint32_t start)
7628 {
7629 	int idx;
7630 
7631 	nrsm->r_start = start;
7632 	nrsm->r_end = rsm->r_end;
7633 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7634 	nrsm->r_act_rxt_cnt = rsm->r_act_rxt_cnt;
7635 	nrsm->r_flags = rsm->r_flags;
7636 	nrsm->r_dupack = rsm->r_dupack;
7637 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7638 	nrsm->r_rtr_bytes = 0;
7639 	nrsm->r_fas = rsm->r_fas;
7640 	nrsm->r_bas = rsm->r_bas;
7641 	tqhash_update_end(rack->r_ctl.tqh, rsm, nrsm->r_start);
7642 	nrsm->r_just_ret = rsm->r_just_ret;
7643 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7644 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7645 	}
7646 	/* Now if we have SYN flag we keep it on the left edge */
7647 	if (nrsm->r_flags & RACK_HAS_SYN)
7648 		nrsm->r_flags &= ~RACK_HAS_SYN;
7649 	/* Now if we have a FIN flag we keep it on the right edge */
7650 	if (rsm->r_flags & RACK_HAS_FIN)
7651 		rsm->r_flags &= ~RACK_HAS_FIN;
7652 	/* Push bit must go to the right edge as well */
7653 	if (rsm->r_flags & RACK_HAD_PUSH)
7654 		rsm->r_flags &= ~RACK_HAD_PUSH;
7655 	/* Clone over the state of the hw_tls flag */
7656 	nrsm->r_hw_tls = rsm->r_hw_tls;
7657 	/*
7658 	 * Now we need to find nrsm's new location in the mbuf chain
7659 	 * we basically calculate a new offset, which is soff +
7660 	 * how much is left in original rsm. Then we walk out the mbuf
7661 	 * chain to find the righ position, it may be the same mbuf
7662 	 * or maybe not.
7663 	 */
7664 	KASSERT(((rsm->m != NULL) ||
7665 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7666 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7667 	if (rsm->m)
7668 		rack_setup_offset_for_rsm(rack, rsm, nrsm);
7669 }
7670 
7671 static struct rack_sendmap *
7672 rack_merge_rsm(struct tcp_rack *rack,
7673 	       struct rack_sendmap *l_rsm,
7674 	       struct rack_sendmap *r_rsm)
7675 {
7676 	/*
7677 	 * We are merging two ack'd RSM's,
7678 	 * the l_rsm is on the left (lower seq
7679 	 * values) and the r_rsm is on the right
7680 	 * (higher seq value). The simplest way
7681 	 * to merge these is to move the right
7682 	 * one into the left. I don't think there
7683 	 * is any reason we need to try to find
7684 	 * the oldest (or last oldest retransmitted).
7685 	 */
7686 	rack_log_map_chg(rack->rc_tp, rack, NULL,
7687 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7688 	tqhash_update_end(rack->r_ctl.tqh, l_rsm, r_rsm->r_end);
7689 	if (l_rsm->r_dupack < r_rsm->r_dupack)
7690 		l_rsm->r_dupack = r_rsm->r_dupack;
7691 	if (r_rsm->r_rtr_bytes)
7692 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7693 	if (r_rsm->r_in_tmap) {
7694 		/* This really should not happen */
7695 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7696 		r_rsm->r_in_tmap = 0;
7697 	}
7698 
7699 	/* Now the flags */
7700 	if (r_rsm->r_flags & RACK_HAS_FIN)
7701 		l_rsm->r_flags |= RACK_HAS_FIN;
7702 	if (r_rsm->r_flags & RACK_TLP)
7703 		l_rsm->r_flags |= RACK_TLP;
7704 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7705 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7706 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
7707 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7708 		/*
7709 		 * If both are app-limited then let the
7710 		 * free lower the count. If right is app
7711 		 * limited and left is not, transfer.
7712 		 */
7713 		l_rsm->r_flags |= RACK_APP_LIMITED;
7714 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
7715 		if (r_rsm == rack->r_ctl.rc_first_appl)
7716 			rack->r_ctl.rc_first_appl = l_rsm;
7717 	}
7718 	tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7719 	/*
7720 	 * We keep the largest value, which is the newest
7721 	 * send. We do this in case a segment that is
7722 	 * joined together and not part of a GP estimate
7723 	 * later gets expanded into the GP estimate.
7724 	 *
7725 	 * We prohibit the merging of unlike kinds i.e.
7726 	 * all pieces that are in the GP estimate can be
7727 	 * merged and all pieces that are not in a GP estimate
7728 	 * can be merged, but not disimilar pieces. Combine
7729 	 * this with taking the highest here and we should
7730 	 * be ok unless of course the client reneges. Then
7731 	 * all bets are off.
7732 	 */
7733 	if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7734 	   r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7735 		l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7736 	}
7737 	/*
7738 	 * When merging two RSM's we also need to consider the ack time and keep
7739 	 * newest. If the ack gets merged into a measurement then that is the
7740 	 * one we will want to be using.
7741 	 */
7742 	if(l_rsm->r_ack_arrival	 < r_rsm->r_ack_arrival)
7743 		l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7744 
7745 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7746 		/* Transfer the split limit to the map we free */
7747 		r_rsm->r_limit_type = l_rsm->r_limit_type;
7748 		l_rsm->r_limit_type = 0;
7749 	}
7750 	rack_free(rack, r_rsm);
7751 	l_rsm->r_flags |= RACK_MERGED;
7752 	return (l_rsm);
7753 }
7754 
7755 /*
7756  * TLP Timer, here we simply setup what segment we want to
7757  * have the TLP expire on, the normal rack_output() will then
7758  * send it out.
7759  *
7760  * We return 1, saying don't proceed with rack_output only
7761  * when all timers have been stopped (destroyed PCB?).
7762  */
7763 static int
7764 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7765 {
7766 	/*
7767 	 * Tail Loss Probe.
7768 	 */
7769 	struct rack_sendmap *rsm = NULL;
7770 	int insret __diagused;
7771 	struct socket *so = tptosocket(tp);
7772 	uint32_t amm;
7773 	uint32_t out, avail;
7774 	int collapsed_win = 0;
7775 
7776 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7777 		/* Its not time yet */
7778 		return (0);
7779 	}
7780 	if (ctf_progress_timeout_check(tp, true)) {
7781 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7782 		return (-ETIMEDOUT);	/* tcp_drop() */
7783 	}
7784 	/*
7785 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7786 	 * need to figure out how to force a full MSS segment out.
7787 	 */
7788 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7789 	rack->r_ctl.retran_during_recovery = 0;
7790 	rack->r_might_revert = 0;
7791 	rack->r_ctl.dsack_byte_cnt = 0;
7792 	counter_u64_add(rack_tlp_tot, 1);
7793 	if (rack->r_state && (rack->r_state != tp->t_state))
7794 		rack_set_state(tp, rack);
7795 	avail = sbavail(&so->so_snd);
7796 	out = tp->snd_max - tp->snd_una;
7797 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7798 		/* special case, we need a retransmission */
7799 		collapsed_win = 1;
7800 		goto need_retran;
7801 	}
7802 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7803 		rack->r_ctl.dsack_persist--;
7804 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7805 			rack->r_ctl.num_dsack = 0;
7806 		}
7807 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7808 	}
7809 	if ((tp->t_flags & TF_GPUTINPROG) &&
7810 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7811 		/*
7812 		 * If this is the second in a row
7813 		 * TLP and we are doing a measurement
7814 		 * its time to abandon the measurement.
7815 		 * Something is likely broken on
7816 		 * the clients network and measuring a
7817 		 * broken network does us no good.
7818 		 */
7819 		tp->t_flags &= ~TF_GPUTINPROG;
7820 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7821 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7822 					   tp->gput_seq,
7823 					   0, 0, 18, __LINE__, NULL, 0);
7824 	}
7825 	/*
7826 	 * Check our send oldest always settings, and if
7827 	 * there is an oldest to send jump to the need_retran.
7828 	 */
7829 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7830 		goto need_retran;
7831 
7832 	if (avail > out) {
7833 		/* New data is available */
7834 		amm = avail - out;
7835 		if (amm > ctf_fixed_maxseg(tp)) {
7836 			amm = ctf_fixed_maxseg(tp);
7837 			if ((amm + out) > tp->snd_wnd) {
7838 				/* We are rwnd limited */
7839 				goto need_retran;
7840 			}
7841 		} else if (amm < ctf_fixed_maxseg(tp)) {
7842 			/* not enough to fill a MTU */
7843 			goto need_retran;
7844 		}
7845 		if (IN_FASTRECOVERY(tp->t_flags)) {
7846 			/* Unlikely */
7847 			if (rack->rack_no_prr == 0) {
7848 				if (out + amm <= tp->snd_wnd) {
7849 					rack->r_ctl.rc_prr_sndcnt = amm;
7850 					rack->r_ctl.rc_tlp_new_data = amm;
7851 					rack_log_to_prr(rack, 4, 0, __LINE__);
7852 				}
7853 			} else
7854 				goto need_retran;
7855 		} else {
7856 			/* Set the send-new override */
7857 			if (out + amm <= tp->snd_wnd)
7858 				rack->r_ctl.rc_tlp_new_data = amm;
7859 			else
7860 				goto need_retran;
7861 		}
7862 		rack->r_ctl.rc_tlpsend = NULL;
7863 		counter_u64_add(rack_tlp_newdata, 1);
7864 		goto send;
7865 	}
7866 need_retran:
7867 	/*
7868 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
7869 	 * optionally the first un-acked segment.
7870 	 */
7871 	if (collapsed_win == 0) {
7872 		if (rack_always_send_oldest)
7873 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7874 		else {
7875 			rsm = tqhash_max(rack->r_ctl.tqh);
7876 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7877 				rsm = rack_find_high_nonack(rack, rsm);
7878 			}
7879 		}
7880 		if (rsm == NULL) {
7881 #ifdef TCP_BLACKBOX
7882 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7883 #endif
7884 			goto out;
7885 		}
7886 	} else {
7887 		/*
7888 		 * We had a collapsed window, lets find
7889 		 * the point before the collapse.
7890 		 */
7891 		if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7892 			rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7893 		else {
7894 			rsm = tqhash_min(rack->r_ctl.tqh);
7895 		}
7896 		if (rsm == NULL) {
7897 			/* Huh */
7898 			goto out;
7899 		}
7900 	}
7901 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7902 		/*
7903 		 * We need to split this the last segment in two.
7904 		 */
7905 		struct rack_sendmap *nrsm;
7906 
7907 		nrsm = rack_alloc_full_limit(rack);
7908 		if (nrsm == NULL) {
7909 			/*
7910 			 * No memory to split, we will just exit and punt
7911 			 * off to the RXT timer.
7912 			 */
7913 			goto out;
7914 		}
7915 		rack_clone_rsm(rack, nrsm, rsm,
7916 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
7917 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7918 #ifndef INVARIANTS
7919 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7920 #else
7921 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7922 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
7923 			      nrsm, insret, rack, rsm);
7924 		}
7925 #endif
7926 		if (rsm->r_in_tmap) {
7927 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7928 			nrsm->r_in_tmap = 1;
7929 		}
7930 		rsm = nrsm;
7931 	}
7932 	rack->r_ctl.rc_tlpsend = rsm;
7933 send:
7934 	/* Make sure output path knows we are doing a TLP */
7935 	*doing_tlp = 1;
7936 	rack->r_timer_override = 1;
7937 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7938 	return (0);
7939 out:
7940 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7941 	return (0);
7942 }
7943 
7944 /*
7945  * Delayed ack Timer, here we simply need to setup the
7946  * ACK_NOW flag and remove the DELACK flag. From there
7947  * the output routine will send the ack out.
7948  *
7949  * We only return 1, saying don't proceed, if all timers
7950  * are stopped (destroyed PCB?).
7951  */
7952 static int
7953 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7954 {
7955 
7956 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7957 	tp->t_flags &= ~TF_DELACK;
7958 	tp->t_flags |= TF_ACKNOW;
7959 	KMOD_TCPSTAT_INC(tcps_delack);
7960 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7961 	return (0);
7962 }
7963 
7964 static inline int
7965 rack_send_ack_challange(struct tcp_rack *rack)
7966 {
7967 	struct tcptemp *t_template;
7968 
7969 	t_template = tcpip_maketemplate(rack->rc_inp);
7970 	if (t_template) {
7971 		if (rack->forced_ack == 0) {
7972 			rack->forced_ack = 1;
7973 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7974 		} else {
7975 			rack->probe_not_answered = 1;
7976 		}
7977 		tcp_respond(rack->rc_tp, t_template->tt_ipgen,
7978 			    &t_template->tt_t, (struct mbuf *)NULL,
7979 			    rack->rc_tp->rcv_nxt, rack->rc_tp->snd_una - 1, 0);
7980 		free(t_template, M_TEMP);
7981 		/* This does send an ack so kill any D-ack timer */
7982 		if (rack->rc_tp->t_flags & TF_DELACK)
7983 			rack->rc_tp->t_flags &= ~TF_DELACK;
7984 		return(1);
7985 	} else
7986 		return (0);
7987 
7988 }
7989 
7990 /*
7991  * Persists timer, here we simply send the
7992  * same thing as a keepalive will.
7993  * the one byte send.
7994  *
7995  * We only return 1, saying don't proceed, if all timers
7996  * are stopped (destroyed PCB?).
7997  */
7998 static int
7999 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
8000 {
8001 	int32_t retval = 1;
8002 
8003 	if (rack->rc_in_persist == 0)
8004 		return (0);
8005 	if (ctf_progress_timeout_check(tp, false)) {
8006 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
8007 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
8008 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
8009 		return (-ETIMEDOUT);	/* tcp_drop() */
8010 	}
8011 	/*
8012 	 * Persistence timer into zero window. Force a byte to be output, if
8013 	 * possible.
8014 	 */
8015 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
8016 	/*
8017 	 * Hack: if the peer is dead/unreachable, we do not time out if the
8018 	 * window is closed.  After a full backoff, drop the connection if
8019 	 * the idle time (no responses to probes) reaches the maximum
8020 	 * backoff that we would use if retransmitting.
8021 	 */
8022 	if (tp->t_rxtshift >= V_tcp_retries &&
8023 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
8024 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
8025 		KMOD_TCPSTAT_INC(tcps_persistdrop);
8026 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
8027 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
8028 		retval = -ETIMEDOUT;	/* tcp_drop() */
8029 		goto out;
8030 	}
8031 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
8032 	    tp->snd_una == tp->snd_max)
8033 		rack_exit_persist(tp, rack, cts);
8034 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
8035 	/*
8036 	 * If the user has closed the socket then drop a persisting
8037 	 * connection after a much reduced timeout.
8038 	 */
8039 	if (tp->t_state > TCPS_CLOSE_WAIT &&
8040 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
8041 		KMOD_TCPSTAT_INC(tcps_persistdrop);
8042 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
8043 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
8044 		retval = -ETIMEDOUT;	/* tcp_drop() */
8045 		goto out;
8046 	}
8047 	if (rack_send_ack_challange(rack)) {
8048 		/* only set it if we were answered */
8049 		if (rack->probe_not_answered) {
8050 			counter_u64_add(rack_persists_loss, 1);
8051 			rack->r_ctl.persist_lost_ends++;
8052 		}
8053 		counter_u64_add(rack_persists_sends, 1);
8054 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
8055 	}
8056 	if (tp->t_rxtshift < V_tcp_retries)
8057 		tp->t_rxtshift++;
8058 out:
8059 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
8060 	rack_start_hpts_timer(rack, tp, cts,
8061 			      0, 0, 0);
8062 	return (retval);
8063 }
8064 
8065 /*
8066  * If a keepalive goes off, we had no other timers
8067  * happening. We always return 1 here since this
8068  * routine either drops the connection or sends
8069  * out a segment with respond.
8070  */
8071 static int
8072 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
8073 {
8074 	struct inpcb *inp = tptoinpcb(tp);
8075 
8076 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
8077 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
8078 	/*
8079 	 * Keep-alive timer went off; send something or drop connection if
8080 	 * idle for too long.
8081 	 */
8082 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
8083 	if (tp->t_state < TCPS_ESTABLISHED)
8084 		goto dropit;
8085 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
8086 	    tp->t_state <= TCPS_CLOSING) {
8087 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
8088 			goto dropit;
8089 		/*
8090 		 * Send a packet designed to force a response if the peer is
8091 		 * up and reachable: either an ACK if the connection is
8092 		 * still alive, or an RST if the peer has closed the
8093 		 * connection due to timeout or reboot. Using sequence
8094 		 * number tp->snd_una-1 causes the transmitted zero-length
8095 		 * segment to lie outside the receive window; by the
8096 		 * protocol spec, this requires the correspondent TCP to
8097 		 * respond.
8098 		 */
8099 		KMOD_TCPSTAT_INC(tcps_keepprobe);
8100 		rack_send_ack_challange(rack);
8101 	}
8102 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
8103 	return (1);
8104 dropit:
8105 	KMOD_TCPSTAT_INC(tcps_keepdrops);
8106 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
8107 	return (-ETIMEDOUT);	/* tcp_drop() */
8108 }
8109 
8110 /*
8111  * Retransmit helper function, clear up all the ack
8112  * flags and take care of important book keeping.
8113  */
8114 static void
8115 rack_remxt_tmr(struct tcpcb *tp)
8116 {
8117 	/*
8118 	 * The retransmit timer went off, all sack'd blocks must be
8119 	 * un-acked.
8120 	 */
8121 	struct rack_sendmap *rsm, *trsm = NULL;
8122 	struct tcp_rack *rack;
8123 
8124 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8125 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
8126 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
8127 	rack->r_timer_override = 1;
8128 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
8129 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
8130 	rack->r_late = 0;
8131 	rack->r_early = 0;
8132 	rack->r_ctl.rc_agg_delayed = 0;
8133 	rack->r_ctl.rc_agg_early = 0;
8134 	if (rack->r_state && (rack->r_state != tp->t_state))
8135 		rack_set_state(tp, rack);
8136 	if (tp->t_rxtshift <= rack_rxt_scoreboard_clear_thresh) {
8137 		/*
8138 		 * We do not clear the scoreboard until we have had
8139 		 * more than rack_rxt_scoreboard_clear_thresh time-outs.
8140 		 */
8141 		rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
8142 		if (rack->r_ctl.rc_resend != NULL)
8143 			rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
8144 
8145 		return;
8146 	}
8147 	/*
8148 	 * Ideally we would like to be able to
8149 	 * mark SACK-PASS on anything not acked here.
8150 	 *
8151 	 * However, if we do that we would burst out
8152 	 * all that data 1ms apart. This would be unwise,
8153 	 * so for now we will just let the normal rxt timer
8154 	 * and tlp timer take care of it.
8155 	 *
8156 	 * Also we really need to stick them back in sequence
8157 	 * order. This way we send in the proper order and any
8158 	 * sacks that come floating in will "re-ack" the data.
8159 	 * To do this we zap the tmap with an INIT and then
8160 	 * walk through and place every rsm in the tail queue
8161 	 * hash table back in its seq ordered place.
8162 	 */
8163 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
8164 
8165 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8166 		rsm->r_dupack = 0;
8167 		if (rack_verbose_logging)
8168 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8169 		/* We must re-add it back to the tlist */
8170 		if (trsm == NULL) {
8171 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8172 		} else {
8173 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
8174 		}
8175 		rsm->r_in_tmap = 1;
8176 		trsm = rsm;
8177 		if (rsm->r_flags & RACK_ACKED)
8178 			rsm->r_flags |= RACK_WAS_ACKED;
8179 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED | RACK_WAS_LOST);
8180 		rsm->r_flags |= RACK_MUST_RXT;
8181 	}
8182 	/* zero the lost since it's all gone */
8183 	rack->r_ctl.rc_considered_lost = 0;
8184 	/* Clear the count (we just un-acked them) */
8185 	rack->r_ctl.rc_sacked = 0;
8186 	rack->r_ctl.rc_sacklast = NULL;
8187 	/* Clear the tlp rtx mark */
8188 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
8189 	if (rack->r_ctl.rc_resend != NULL)
8190 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
8191 	rack->r_ctl.rc_prr_sndcnt = 0;
8192 	rack_log_to_prr(rack, 6, 0, __LINE__);
8193 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
8194 	if (rack->r_ctl.rc_resend != NULL)
8195 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
8196 	if (((tp->t_flags & TF_SACK_PERMIT) == 0) &&
8197 	    ((tp->t_flags & TF_SENTFIN) == 0)) {
8198 		/*
8199 		 * For non-sack customers new data
8200 		 * needs to go out as retransmits until
8201 		 * we retransmit up to snd_max.
8202 		 */
8203 		rack->r_must_retran = 1;
8204 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
8205 							    rack->r_ctl.rc_sacked);
8206 	}
8207 }
8208 
8209 static void
8210 rack_convert_rtts(struct tcpcb *tp)
8211 {
8212 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
8213 	tp->t_rxtcur = RACK_REXMTVAL(tp);
8214 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
8215 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
8216 	}
8217 	if (tp->t_rxtcur > rack_rto_max) {
8218 		tp->t_rxtcur = rack_rto_max;
8219 	}
8220 }
8221 
8222 static void
8223 rack_cc_conn_init(struct tcpcb *tp)
8224 {
8225 	struct tcp_rack *rack;
8226 	uint32_t srtt;
8227 
8228 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8229 	srtt = tp->t_srtt;
8230 	cc_conn_init(tp);
8231 	/*
8232 	 * Now convert to rack's internal format,
8233 	 * if required.
8234 	 */
8235 	if ((srtt == 0) && (tp->t_srtt != 0))
8236 		rack_convert_rtts(tp);
8237 	/*
8238 	 * We want a chance to stay in slowstart as
8239 	 * we create a connection. TCP spec says that
8240 	 * initially ssthresh is infinite. For our
8241 	 * purposes that is the snd_wnd.
8242 	 */
8243 	if (tp->snd_ssthresh < tp->snd_wnd) {
8244 		tp->snd_ssthresh = tp->snd_wnd;
8245 	}
8246 	/*
8247 	 * We also want to assure a IW worth of
8248 	 * data can get inflight.
8249 	 */
8250 	if (rc_init_window(rack) < tp->snd_cwnd)
8251 		tp->snd_cwnd = rc_init_window(rack);
8252 }
8253 
8254 /*
8255  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
8256  * we will setup to retransmit the lowest seq number outstanding.
8257  */
8258 static int
8259 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
8260 {
8261 	struct inpcb *inp = tptoinpcb(tp);
8262 	int32_t rexmt;
8263 	int32_t retval = 0;
8264 	bool isipv6;
8265 
8266 	if ((tp->t_flags & TF_GPUTINPROG) &&
8267 	    (tp->t_rxtshift)) {
8268 		/*
8269 		 * We have had a second timeout
8270 		 * measurements on successive rxt's are not profitable.
8271 		 * It is unlikely to be of any use (the network is
8272 		 * broken or the client went away).
8273 		 */
8274 		tp->t_flags &= ~TF_GPUTINPROG;
8275 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
8276 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
8277 					   tp->gput_seq,
8278 					   0, 0, 18, __LINE__, NULL, 0);
8279 	}
8280 	if (ctf_progress_timeout_check(tp, false)) {
8281 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
8282 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
8283 		return (-ETIMEDOUT);	/* tcp_drop() */
8284 	}
8285 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
8286 	rack->r_ctl.retran_during_recovery = 0;
8287 	rack->rc_ack_required = 1;
8288 	rack->r_ctl.dsack_byte_cnt = 0;
8289 	if (IN_RECOVERY(tp->t_flags) &&
8290 	    (rack->rto_from_rec == 0)) {
8291 		/*
8292 		 * Mark that we had a rto while in recovery
8293 		 * and save the ssthresh so if we go back
8294 		 * into recovery we will have a chance
8295 		 * to slowstart back to the level.
8296 		 */
8297 		rack->rto_from_rec = 1;
8298 		rack->r_ctl.rto_ssthresh = tp->snd_ssthresh;
8299 	}
8300 	if (IN_FASTRECOVERY(tp->t_flags))
8301 		tp->t_flags |= TF_WASFRECOVERY;
8302 	else
8303 		tp->t_flags &= ~TF_WASFRECOVERY;
8304 	if (IN_CONGRECOVERY(tp->t_flags))
8305 		tp->t_flags |= TF_WASCRECOVERY;
8306 	else
8307 		tp->t_flags &= ~TF_WASCRECOVERY;
8308 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
8309 	    (tp->snd_una == tp->snd_max)) {
8310 		/* Nothing outstanding .. nothing to do */
8311 		return (0);
8312 	}
8313 	if (rack->r_ctl.dsack_persist) {
8314 		rack->r_ctl.dsack_persist--;
8315 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
8316 			rack->r_ctl.num_dsack = 0;
8317 		}
8318 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
8319 	}
8320 	/*
8321 	 * Rack can only run one timer  at a time, so we cannot
8322 	 * run a KEEPINIT (gating SYN sending) and a retransmit
8323 	 * timer for the SYN. So if we are in a front state and
8324 	 * have a KEEPINIT timer we need to check the first transmit
8325 	 * against now to see if we have exceeded the KEEPINIT time
8326 	 * (if one is set).
8327 	 */
8328 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
8329 	    (TP_KEEPINIT(tp) != 0)) {
8330 		struct rack_sendmap *rsm;
8331 
8332 		rsm = tqhash_min(rack->r_ctl.tqh);
8333 		if (rsm) {
8334 			/* Ok we have something outstanding to test keepinit with */
8335 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
8336 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
8337 				/* We have exceeded the KEEPINIT time */
8338 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
8339 				goto drop_it;
8340 			}
8341 		}
8342 	}
8343 	/*
8344 	 * Retransmission timer went off.  Message has not been acked within
8345 	 * retransmit interval.  Back off to a longer retransmit interval
8346 	 * and retransmit one segment.
8347 	 */
8348 	if ((rack->r_ctl.rc_resend == NULL) ||
8349 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
8350 		/*
8351 		 * If the rwnd collapsed on
8352 		 * the one we are retransmitting
8353 		 * it does not count against the
8354 		 * rxt count.
8355 		 */
8356 		tp->t_rxtshift++;
8357 	}
8358 	rack_remxt_tmr(tp);
8359 	if (tp->t_rxtshift > V_tcp_retries) {
8360 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
8361 drop_it:
8362 		tp->t_rxtshift = V_tcp_retries;
8363 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
8364 		/* XXXGL: previously t_softerror was casted to uint16_t */
8365 		MPASS(tp->t_softerror >= 0);
8366 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
8367 		goto out;	/* tcp_drop() */
8368 	}
8369 	if (tp->t_state == TCPS_SYN_SENT) {
8370 		/*
8371 		 * If the SYN was retransmitted, indicate CWND to be limited
8372 		 * to 1 segment in cc_conn_init().
8373 		 */
8374 		tp->snd_cwnd = 1;
8375 	} else if (tp->t_rxtshift == 1) {
8376 		/*
8377 		 * first retransmit; record ssthresh and cwnd so they can be
8378 		 * recovered if this turns out to be a "bad" retransmit. A
8379 		 * retransmit is considered "bad" if an ACK for this segment
8380 		 * is received within RTT/2 interval; the assumption here is
8381 		 * that the ACK was already in flight.  See "On Estimating
8382 		 * End-to-End Network Path Properties" by Allman and Paxson
8383 		 * for more details.
8384 		 */
8385 		tp->snd_cwnd_prev = tp->snd_cwnd;
8386 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
8387 		tp->snd_recover_prev = tp->snd_recover;
8388 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
8389 		tp->t_flags |= TF_PREVVALID;
8390 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
8391 		tp->t_flags &= ~TF_PREVVALID;
8392 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
8393 	if ((tp->t_state == TCPS_SYN_SENT) ||
8394 	    (tp->t_state == TCPS_SYN_RECEIVED))
8395 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
8396 	else
8397 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
8398 
8399 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
8400 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
8401 	/*
8402 	 * We enter the path for PLMTUD if connection is established or, if
8403 	 * connection is FIN_WAIT_1 status, reason for the last is that if
8404 	 * amount of data we send is very small, we could send it in couple
8405 	 * of packets and process straight to FIN. In that case we won't
8406 	 * catch ESTABLISHED state.
8407 	 */
8408 #ifdef INET6
8409 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
8410 #else
8411 	isipv6 = false;
8412 #endif
8413 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
8414 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
8415 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
8416 	    ((tp->t_state == TCPS_ESTABLISHED) ||
8417 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
8418 		/*
8419 		 * Idea here is that at each stage of mtu probe (usually,
8420 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
8421 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
8422 		 * should take care of that.
8423 		 */
8424 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
8425 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
8426 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
8427 		    tp->t_rxtshift % 2 == 0)) {
8428 			/*
8429 			 * Enter Path MTU Black-hole Detection mechanism: -
8430 			 * Disable Path MTU Discovery (IP "DF" bit). -
8431 			 * Reduce MTU to lower value than what we negotiated
8432 			 * with peer.
8433 			 */
8434 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
8435 				/* Record that we may have found a black hole. */
8436 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
8437 				/* Keep track of previous MSS. */
8438 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
8439 			}
8440 
8441 			/*
8442 			 * Reduce the MSS to blackhole value or to the
8443 			 * default in an attempt to retransmit.
8444 			 */
8445 #ifdef INET6
8446 			if (isipv6 &&
8447 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
8448 				/* Use the sysctl tuneable blackhole MSS. */
8449 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
8450 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
8451 			} else if (isipv6) {
8452 				/* Use the default MSS. */
8453 				tp->t_maxseg = V_tcp_v6mssdflt;
8454 				/*
8455 				 * Disable Path MTU Discovery when we switch
8456 				 * to minmss.
8457 				 */
8458 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
8459 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
8460 			}
8461 #endif
8462 #if defined(INET6) && defined(INET)
8463 			else
8464 #endif
8465 #ifdef INET
8466 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
8467 				/* Use the sysctl tuneable blackhole MSS. */
8468 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
8469 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
8470 			} else {
8471 				/* Use the default MSS. */
8472 				tp->t_maxseg = V_tcp_mssdflt;
8473 				/*
8474 				 * Disable Path MTU Discovery when we switch
8475 				 * to minmss.
8476 				 */
8477 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
8478 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
8479 			}
8480 #endif
8481 		} else {
8482 			/*
8483 			 * If further retransmissions are still unsuccessful
8484 			 * with a lowered MTU, maybe this isn't a blackhole
8485 			 * and we restore the previous MSS and blackhole
8486 			 * detection flags. The limit '6' is determined by
8487 			 * giving each probe stage (1448, 1188, 524) 2
8488 			 * chances to recover.
8489 			 */
8490 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
8491 			    (tp->t_rxtshift >= 6)) {
8492 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
8493 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
8494 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
8495 				if (tp->t_maxseg < V_tcp_mssdflt) {
8496 					/*
8497 					 * The MSS is so small we should not
8498 					 * process incoming SACK's since we are
8499 					 * subject to attack in such a case.
8500 					 */
8501 					tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
8502 				} else {
8503 					tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
8504 				}
8505 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
8506 			}
8507 		}
8508 	}
8509 	/*
8510 	 * Disable RFC1323 and SACK if we haven't got any response to
8511 	 * our third SYN to work-around some broken terminal servers
8512 	 * (most of which have hopefully been retired) that have bad VJ
8513 	 * header compression code which trashes TCP segments containing
8514 	 * unknown-to-them TCP options.
8515 	 */
8516 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
8517 	    (tp->t_rxtshift == 3))
8518 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
8519 	/*
8520 	 * If we backed off this far, our srtt estimate is probably bogus.
8521 	 * Clobber it so we'll take the next rtt measurement as our srtt;
8522 	 * move the current srtt into rttvar to keep the current retransmit
8523 	 * times until then.
8524 	 */
8525 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
8526 #ifdef INET6
8527 		if ((inp->inp_vflag & INP_IPV6) != 0)
8528 			in6_losing(inp);
8529 		else
8530 #endif
8531 			in_losing(inp);
8532 		tp->t_rttvar += tp->t_srtt;
8533 		tp->t_srtt = 0;
8534 	}
8535 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
8536 	tp->snd_recover = tp->snd_max;
8537 	tp->t_flags |= TF_ACKNOW;
8538 	tp->t_rtttime = 0;
8539 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
8540 out:
8541 	return (retval);
8542 }
8543 
8544 static int
8545 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
8546 {
8547 	int32_t ret = 0;
8548 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
8549 
8550 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8551 	    (tp->t_flags & TF_GPUTINPROG)) {
8552 		/*
8553 		 * We have a goodput in progress
8554 		 * and we have entered a late state.
8555 		 * Do we have enough data in the sb
8556 		 * to handle the GPUT request?
8557 		 */
8558 		uint32_t bytes;
8559 
8560 		bytes = tp->gput_ack - tp->gput_seq;
8561 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
8562 			bytes += tp->gput_seq - tp->snd_una;
8563 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
8564 			/*
8565 			 * There are not enough bytes in the socket
8566 			 * buffer that have been sent to cover this
8567 			 * measurement. Cancel it.
8568 			 */
8569 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
8570 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
8571 						   tp->gput_seq,
8572 						   0, 0, 18, __LINE__, NULL, 0);
8573 			tp->t_flags &= ~TF_GPUTINPROG;
8574 		}
8575 	}
8576 	if (timers == 0) {
8577 		return (0);
8578 	}
8579 	if (tp->t_state == TCPS_LISTEN) {
8580 		/* no timers on listen sockets */
8581 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
8582 			return (0);
8583 		return (1);
8584 	}
8585 	if ((timers & PACE_TMR_RACK) &&
8586 	    rack->rc_on_min_to) {
8587 		/*
8588 		 * For the rack timer when we
8589 		 * are on a min-timeout (which means rrr_conf = 3)
8590 		 * we don't want to check the timer. It may
8591 		 * be going off for a pace and thats ok we
8592 		 * want to send the retransmit (if its ready).
8593 		 *
8594 		 * If its on a normal rack timer (non-min) then
8595 		 * we will check if its expired.
8596 		 */
8597 		goto skip_time_check;
8598 	}
8599 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
8600 		uint32_t left;
8601 
8602 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
8603 			ret = -1;
8604 			rack_log_to_processing(rack, cts, ret, 0);
8605 			return (0);
8606 		}
8607 		if (hpts_calling == 0) {
8608 			/*
8609 			 * A user send or queued mbuf (sack) has called us? We
8610 			 * return 0 and let the pacing guards
8611 			 * deal with it if they should or
8612 			 * should not cause a send.
8613 			 */
8614 			ret = -2;
8615 			rack_log_to_processing(rack, cts, ret, 0);
8616 			return (0);
8617 		}
8618 		/*
8619 		 * Ok our timer went off early and we are not paced false
8620 		 * alarm, go back to sleep. We make sure we don't have
8621 		 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8622 		 * flag in place.
8623 		 */
8624 		rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8625 		ret = -3;
8626 		left = rack->r_ctl.rc_timer_exp - cts;
8627 		tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8628 		rack_log_to_processing(rack, cts, ret, left);
8629 		return (1);
8630 	}
8631 skip_time_check:
8632 	rack->rc_tmr_stopped = 0;
8633 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8634 	if (timers & PACE_TMR_DELACK) {
8635 		ret = rack_timeout_delack(tp, rack, cts);
8636 	} else if (timers & PACE_TMR_RACK) {
8637 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8638 		rack->r_fast_output = 0;
8639 		ret = rack_timeout_rack(tp, rack, cts);
8640 	} else if (timers & PACE_TMR_TLP) {
8641 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8642 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8643 	} else if (timers & PACE_TMR_RXT) {
8644 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8645 		rack->r_fast_output = 0;
8646 		ret = rack_timeout_rxt(tp, rack, cts);
8647 	} else if (timers & PACE_TMR_PERSIT) {
8648 		ret = rack_timeout_persist(tp, rack, cts);
8649 	} else if (timers & PACE_TMR_KEEP) {
8650 		ret = rack_timeout_keepalive(tp, rack, cts);
8651 	}
8652 	rack_log_to_processing(rack, cts, ret, timers);
8653 	return (ret);
8654 }
8655 
8656 static void
8657 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8658 {
8659 	struct timeval tv;
8660 	uint32_t us_cts, flags_on_entry;
8661 	uint8_t hpts_removed = 0;
8662 
8663 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
8664 	us_cts = tcp_get_usecs(&tv);
8665 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8666 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8667 	     ((tp->snd_max - tp->snd_una) == 0))) {
8668 		tcp_hpts_remove(rack->rc_tp);
8669 		hpts_removed = 1;
8670 		/* If we were not delayed cancel out the flag. */
8671 		if ((tp->snd_max - tp->snd_una) == 0)
8672 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8673 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8674 	}
8675 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8676 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8677 		if (tcp_in_hpts(rack->rc_tp) &&
8678 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8679 			/*
8680 			 * Canceling timer's when we have no output being
8681 			 * paced. We also must remove ourselves from the
8682 			 * hpts.
8683 			 */
8684 			tcp_hpts_remove(rack->rc_tp);
8685 			hpts_removed = 1;
8686 		}
8687 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8688 	}
8689 	if (hpts_removed == 0)
8690 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8691 }
8692 
8693 static int
8694 rack_stopall(struct tcpcb *tp)
8695 {
8696 	struct tcp_rack *rack;
8697 
8698 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8699 	rack->t_timers_stopped = 1;
8700 
8701 	tcp_hpts_remove(tp);
8702 
8703 	return (0);
8704 }
8705 
8706 static void
8707 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8708 {
8709 	/*
8710 	 * Assure no timers are running.
8711 	 */
8712 	if (tcp_timer_active(tp, TT_PERSIST)) {
8713 		/* We enter in persists, set the flag appropriately */
8714 		rack->rc_in_persist = 1;
8715 	}
8716 	if (tcp_in_hpts(rack->rc_tp)) {
8717 		tcp_hpts_remove(rack->rc_tp);
8718 	}
8719 }
8720 
8721 /*
8722  * We maintain an array fo 16 (RETRAN_CNT_SIZE) entries. This
8723  * array is zeroed at the start of recovery. Each time a segment
8724  * is retransmitted, we translate that into a number of packets
8725  * (based on segsiz) and based on how many times its been retransmitted
8726  * increment by the number of packets the counter that represents
8727  * retansmitted N times. Index 0 is retransmitted 1 time, index 1
8728  * is retransmitted 2 times etc.
8729  *
8730  * So for example when we send a 4344 byte transmission with a 1448
8731  * byte segsize, and its the third time we have retransmitted this
8732  * segment, we would add to the rc_cnt_of_retran[2] the value of
8733  * 3. That represents 3 MSS were retransmitted 3 times (index is
8734  * the number of times retranmitted minus 1).
8735  */
8736 static void
8737 rack_peg_rxt(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t segsiz)
8738 {
8739 	int idx;
8740 	uint32_t peg;
8741 
8742 	peg = ((rsm->r_end - rsm->r_start) + segsiz) - 1;
8743 	peg /= segsiz;
8744 	idx = rsm->r_act_rxt_cnt - 1;
8745 	if (idx >= RETRAN_CNT_SIZE)
8746 		idx = RETRAN_CNT_SIZE - 1;
8747 	/* Max of a uint16_t retransmits in a bucket */
8748 	if ((rack->r_ctl.rc_cnt_of_retran[idx] + peg) < 0xffff)
8749 		rack->r_ctl.rc_cnt_of_retran[idx] += peg;
8750 	else
8751 		rack->r_ctl.rc_cnt_of_retran[idx] = 0xffff;
8752 }
8753 
8754 /*
8755  * We maintain an array fo 16 (RETRAN_CNT_SIZE) entries. This
8756  * array is zeroed at the start of recovery. Each time a segment
8757  * is retransmitted, we translate that into a number of packets
8758  * (based on segsiz) and based on how many times its been retransmitted
8759  * increment by the number of packets the counter that represents
8760  * retansmitted N times. Index 0 is retransmitted 1 time, index 1
8761  * is retransmitted 2 times etc.
8762  *
8763  * The rack_unpeg_rxt is used when we go to retransmit a segment
8764  * again. Basically if the segment had previously been retransmitted
8765  * say 3 times (as our previous example illustrated in the comment
8766  * above rack_peg_rxt() prior to calling that and incrementing
8767  * r_ack_rxt_cnt we would have called rack_unpeg_rxt() that would
8768  * subtract back the previous add from its last rxt (in this
8769  * example r_act_cnt would have been 2 for 2 retransmissions. So
8770  * we would have subtracted 3 from rc_cnt_of_reetran[1] to remove
8771  * those 3 segments. You will see this in the rack_update_rsm()
8772  * below where we do:
8773  *	if (rsm->r_act_rxt_cnt > 0) {
8774  *		rack_unpeg_rxt(rack, rsm, segsiz);
8775  *	}
8776  *	rsm->r_act_rxt_cnt++;
8777  *	rack_peg_rxt(rack, rsm, segsiz);
8778  *
8779  * This effectively moves the count from rc_cnt_of_retran[1] to
8780  * rc_cnt_of_retran[2].
8781  */
8782 static void
8783 rack_unpeg_rxt(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t segsiz)
8784 {
8785 	int idx;
8786 	uint32_t peg;
8787 
8788 	idx = rsm->r_act_rxt_cnt - 1;
8789 	if (idx >= RETRAN_CNT_SIZE)
8790 		idx = RETRAN_CNT_SIZE - 1;
8791 	peg = ((rsm->r_end - rsm->r_start) + segsiz) - 1;
8792 	peg /= segsiz;
8793 	if (peg < rack->r_ctl.rc_cnt_of_retran[idx])
8794 		rack->r_ctl.rc_cnt_of_retran[idx] -= peg;
8795 	else {
8796 		/* TSNH */
8797 		rack->r_ctl.rc_cnt_of_retran[idx] = 0;
8798 	}
8799 }
8800 
8801 static void
8802 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8803     struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz)
8804 {
8805 	int32_t idx;
8806 
8807 	rsm->r_rtr_cnt++;
8808 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8809 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8810 		rsm->r_flags |= RACK_OVERMAX;
8811 	}
8812 	if (rsm->r_act_rxt_cnt > 0) {
8813 		/* Drop the count back for this, its retransmitting again */
8814 		rack_unpeg_rxt(rack, rsm, segsiz);
8815 	}
8816 	rsm->r_act_rxt_cnt++;
8817 	/* Peg the count/index */
8818 	rack_peg_rxt(rack, rsm, segsiz);
8819 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8820 	rsm->r_dupack = 0;
8821 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8822 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8823 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8824 	}
8825 	if (rsm->r_flags & RACK_WAS_LOST) {
8826 		/*
8827 		 * We retransmitted it putting it back in flight
8828 		 * remove the lost desgination and reduce the
8829 		 * bytes considered lost.
8830 		 */
8831 		rsm->r_flags  &= ~RACK_WAS_LOST;
8832 		KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
8833 			("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
8834 		if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
8835 			rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
8836 		else
8837 			rack->r_ctl.rc_considered_lost = 0;
8838 	}
8839 	idx = rsm->r_rtr_cnt - 1;
8840 	rsm->r_tim_lastsent[idx] = ts;
8841 	/*
8842 	 * Here we don't add in the len of send, since its already
8843 	 * in snduna <->snd_max.
8844 	 */
8845 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
8846 				     rack->r_ctl.rc_sacked);
8847 	if (rsm->r_flags & RACK_ACKED) {
8848 		/* Problably MTU discovery messing with us */
8849 		rsm->r_flags &= ~RACK_ACKED;
8850 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8851 	}
8852 	if (rsm->r_in_tmap) {
8853 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8854 		rsm->r_in_tmap = 0;
8855 	}
8856 	/* Lets make sure it really is in or not the GP window */
8857 	rack_mark_in_gp_win(tp, rsm);
8858 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8859 	rsm->r_in_tmap = 1;
8860 	rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8861 	/* Take off the must retransmit flag, if its on */
8862 	if (rsm->r_flags & RACK_MUST_RXT) {
8863 		if (rack->r_must_retran)
8864 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8865 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8866 			/*
8867 			 * We have retransmitted all we need. Clear
8868 			 * any must retransmit flags.
8869 			 */
8870 			rack->r_must_retran = 0;
8871 			rack->r_ctl.rc_out_at_rto = 0;
8872 		}
8873 		rsm->r_flags &= ~RACK_MUST_RXT;
8874 	}
8875 	/* Remove any collapsed flag */
8876 	rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8877 	if (rsm->r_flags & RACK_SACK_PASSED) {
8878 		/* We have retransmitted due to the SACK pass */
8879 		rsm->r_flags &= ~RACK_SACK_PASSED;
8880 		rsm->r_flags |= RACK_WAS_SACKPASS;
8881 	}
8882 }
8883 
8884 static uint32_t
8885 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8886     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint32_t add_flag, int segsiz)
8887 {
8888 	/*
8889 	 * We (re-)transmitted starting at rsm->r_start for some length
8890 	 * (possibly less than r_end.
8891 	 */
8892 	struct rack_sendmap *nrsm;
8893 	int insret __diagused;
8894 	uint32_t c_end;
8895 	int32_t len;
8896 
8897 	len = *lenp;
8898 	c_end = rsm->r_start + len;
8899 	if (SEQ_GEQ(c_end, rsm->r_end)) {
8900 		/*
8901 		 * We retransmitted the whole piece or more than the whole
8902 		 * slopping into the next rsm.
8903 		 */
8904 		rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8905 		if (c_end == rsm->r_end) {
8906 			*lenp = 0;
8907 			return (0);
8908 		} else {
8909 			int32_t act_len;
8910 
8911 			/* Hangs over the end return whats left */
8912 			act_len = rsm->r_end - rsm->r_start;
8913 			*lenp = (len - act_len);
8914 			return (rsm->r_end);
8915 		}
8916 		/* We don't get out of this block. */
8917 	}
8918 	/*
8919 	 * Here we retransmitted less than the whole thing which means we
8920 	 * have to split this into what was transmitted and what was not.
8921 	 */
8922 	nrsm = rack_alloc_full_limit(rack);
8923 	if (nrsm == NULL) {
8924 		/*
8925 		 * We can't get memory, so lets not proceed.
8926 		 */
8927 		*lenp = 0;
8928 		return (0);
8929 	}
8930 	/*
8931 	 * So here we are going to take the original rsm and make it what we
8932 	 * retransmitted. nrsm will be the tail portion we did not
8933 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8934 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8935 	 * 1, 6 and the new piece will be 6, 11.
8936 	 */
8937 	rack_clone_rsm(rack, nrsm, rsm, c_end);
8938 	nrsm->r_dupack = 0;
8939 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8940 #ifndef INVARIANTS
8941 	(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8942 #else
8943 	if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8944 		panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8945 		      nrsm, insret, rack, rsm);
8946 	}
8947 #endif
8948 	if (rsm->r_in_tmap) {
8949 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8950 		nrsm->r_in_tmap = 1;
8951 	}
8952 	rsm->r_flags &= (~RACK_HAS_FIN);
8953 	rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8954 	/* Log a split of rsm into rsm and nrsm */
8955 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8956 	*lenp = 0;
8957 	return (0);
8958 }
8959 
8960 static void
8961 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8962 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8963 		struct rack_sendmap *hintrsm, uint32_t add_flag, struct mbuf *s_mb,
8964 		uint32_t s_moff, int hw_tls, int segsiz)
8965 {
8966 	struct tcp_rack *rack;
8967 	struct rack_sendmap *rsm, *nrsm;
8968 	int insret __diagused;
8969 
8970 	register uint32_t snd_max, snd_una;
8971 
8972 	/*
8973 	 * Add to the RACK log of packets in flight or retransmitted. If
8974 	 * there is a TS option we will use the TS echoed, if not we will
8975 	 * grab a TS.
8976 	 *
8977 	 * Retransmissions will increment the count and move the ts to its
8978 	 * proper place. Note that if options do not include TS's then we
8979 	 * won't be able to effectively use the ACK for an RTT on a retran.
8980 	 *
8981 	 * Notes about r_start and r_end. Lets consider a send starting at
8982 	 * sequence 1 for 10 bytes. In such an example the r_start would be
8983 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8984 	 * This means that r_end is actually the first sequence for the next
8985 	 * slot (11).
8986 	 *
8987 	 */
8988 	/*
8989 	 * If err is set what do we do XXXrrs? should we not add the thing?
8990 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
8991 	 * i.e. proceed with add ** do this for now.
8992 	 */
8993 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8994 	if (err)
8995 		/*
8996 		 * We don't log errors -- we could but snd_max does not
8997 		 * advance in this case either.
8998 		 */
8999 		return;
9000 
9001 	if (th_flags & TH_RST) {
9002 		/*
9003 		 * We don't log resets and we return immediately from
9004 		 * sending
9005 		 */
9006 		return;
9007 	}
9008 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9009 	snd_una = tp->snd_una;
9010 	snd_max = tp->snd_max;
9011 	if (th_flags & (TH_SYN | TH_FIN)) {
9012 		/*
9013 		 * The call to rack_log_output is made before bumping
9014 		 * snd_max. This means we can record one extra byte on a SYN
9015 		 * or FIN if seq_out is adding more on and a FIN is present
9016 		 * (and we are not resending).
9017 		 */
9018 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
9019 			len++;
9020 		if (th_flags & TH_FIN)
9021 			len++;
9022 	}
9023 	if (SEQ_LEQ((seq_out + len), snd_una)) {
9024 		/* Are sending an old segment to induce an ack (keep-alive)? */
9025 		return;
9026 	}
9027 	if (SEQ_LT(seq_out, snd_una)) {
9028 		/* huh? should we panic? */
9029 		uint32_t end;
9030 
9031 		end = seq_out + len;
9032 		seq_out = snd_una;
9033 		if (SEQ_GEQ(end, seq_out))
9034 			len = end - seq_out;
9035 		else
9036 			len = 0;
9037 	}
9038 	if (len == 0) {
9039 		/* We don't log zero window probes */
9040 		return;
9041 	}
9042 	if (IN_FASTRECOVERY(tp->t_flags)) {
9043 		rack->r_ctl.rc_prr_out += len;
9044 	}
9045 	/* First question is it a retransmission or new? */
9046 	if (seq_out == snd_max) {
9047 		/* Its new */
9048 		rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
9049 again:
9050 		rsm = rack_alloc(rack);
9051 		if (rsm == NULL) {
9052 			/*
9053 			 * Hmm out of memory and the tcb got destroyed while
9054 			 * we tried to wait.
9055 			 */
9056 			return;
9057 		}
9058 		if (th_flags & TH_FIN) {
9059 			rsm->r_flags = RACK_HAS_FIN|add_flag;
9060 		} else {
9061 			rsm->r_flags = add_flag;
9062 		}
9063 		if (hw_tls)
9064 			rsm->r_hw_tls = 1;
9065 		rsm->r_tim_lastsent[0] = cts;
9066 		rsm->r_rtr_cnt = 1;
9067  		rsm->r_act_rxt_cnt = 0;
9068 		rsm->r_rtr_bytes = 0;
9069 		if (th_flags & TH_SYN) {
9070 			/* The data space is one beyond snd_una */
9071 			rsm->r_flags |= RACK_HAS_SYN;
9072 		}
9073 		rsm->r_start = seq_out;
9074 		rsm->r_end = rsm->r_start + len;
9075 		rack_mark_in_gp_win(tp, rsm);
9076 		rsm->r_dupack = 0;
9077 		/*
9078 		 * save off the mbuf location that
9079 		 * sndmbuf_noadv returned (which is
9080 		 * where we started copying from)..
9081 		 */
9082 		rsm->m = s_mb;
9083 		rsm->soff = s_moff;
9084 		/*
9085 		 * Here we do add in the len of send, since its not yet
9086 		 * reflected in in snduna <->snd_max
9087 		 */
9088 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
9089 					      rack->r_ctl.rc_sacked) +
9090 			      (rsm->r_end - rsm->r_start));
9091 		if ((rack->rc_initial_ss_comp == 0) &&
9092 		    (rack->r_ctl.ss_hi_fs < rsm->r_fas)) {
9093 			   rack->r_ctl.ss_hi_fs = rsm->r_fas;
9094 		}
9095 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
9096 		if (rsm->m) {
9097 			if (rsm->m->m_len <= rsm->soff) {
9098 				/*
9099 				 * XXXrrs Question, will this happen?
9100 				 *
9101 				 * If sbsndptr is set at the correct place
9102 				 * then s_moff should always be somewhere
9103 				 * within rsm->m. But if the sbsndptr was
9104 				 * off then that won't be true. If it occurs
9105 				 * we need to walkout to the correct location.
9106 				 */
9107 				struct mbuf *lm;
9108 
9109 				lm = rsm->m;
9110 				while (lm->m_len <= rsm->soff) {
9111 					rsm->soff -= lm->m_len;
9112 					lm = lm->m_next;
9113 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
9114 							     __func__, rack, s_moff, s_mb, rsm->soff));
9115 				}
9116 				rsm->m = lm;
9117 			}
9118 			rsm->orig_m_len = rsm->m->m_len;
9119 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
9120 		} else {
9121 			rsm->orig_m_len = 0;
9122 			rsm->orig_t_space = 0;
9123 		}
9124 		rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
9125 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9126 		/* Log a new rsm */
9127 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
9128 #ifndef INVARIANTS
9129 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
9130 #else
9131 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
9132 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9133 			      nrsm, insret, rack, rsm);
9134 		}
9135 #endif
9136 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9137 		rsm->r_in_tmap = 1;
9138 		if (rsm->r_flags & RACK_IS_PCM) {
9139 			rack->r_ctl.pcm_i.send_time = cts;
9140 			rack->r_ctl.pcm_i.eseq = rsm->r_end;
9141 			/* First time through we set the start too */
9142 			if (rack->pcm_in_progress == 0)
9143 				rack->r_ctl.pcm_i.sseq = rsm->r_start;
9144 		}
9145 		/*
9146 		 * Special case detection, is there just a single
9147 		 * packet outstanding when we are not in recovery?
9148 		 *
9149 		 * If this is true mark it so.
9150 		 */
9151 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9152 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
9153 			struct rack_sendmap *prsm;
9154 
9155 			prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
9156 			if (prsm)
9157 				prsm->r_one_out_nr = 1;
9158 		}
9159 		return;
9160 	}
9161 	/*
9162 	 * If we reach here its a retransmission and we need to find it.
9163 	 */
9164 more:
9165 	if (hintrsm && (hintrsm->r_start == seq_out)) {
9166 		rsm = hintrsm;
9167 		hintrsm = NULL;
9168 	} else {
9169 		/* No hints sorry */
9170 		rsm = NULL;
9171 	}
9172 	if ((rsm) && (rsm->r_start == seq_out)) {
9173 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
9174 		if (len == 0) {
9175 			return;
9176 		} else {
9177 			goto more;
9178 		}
9179 	}
9180 	/* Ok it was not the last pointer go through it the hard way. */
9181 refind:
9182 	rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
9183 	if (rsm) {
9184 		if (rsm->r_start == seq_out) {
9185 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
9186 			if (len == 0) {
9187 				return;
9188 			} else {
9189 				goto refind;
9190 			}
9191 		}
9192 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
9193 			/* Transmitted within this piece */
9194 			/*
9195 			 * Ok we must split off the front and then let the
9196 			 * update do the rest
9197 			 */
9198 			nrsm = rack_alloc_full_limit(rack);
9199 			if (nrsm == NULL) {
9200 				rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
9201 				return;
9202 			}
9203 			/*
9204 			 * copy rsm to nrsm and then trim the front of rsm
9205 			 * to not include this part.
9206 			 */
9207 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
9208 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
9209 #ifndef INVARIANTS
9210 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9211 #else
9212 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9213 				panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9214 				      nrsm, insret, rack, rsm);
9215 			}
9216 #endif
9217 			if (rsm->r_in_tmap) {
9218 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9219 				nrsm->r_in_tmap = 1;
9220 			}
9221 			rsm->r_flags &= (~RACK_HAS_FIN);
9222 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
9223 			if (len == 0) {
9224 				return;
9225 			} else if (len > 0)
9226 				goto refind;
9227 		}
9228 	}
9229 	/*
9230 	 * Hmm not found in map did they retransmit both old and on into the
9231 	 * new?
9232 	 */
9233 	if (seq_out == tp->snd_max) {
9234 		goto again;
9235 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
9236 #ifdef INVARIANTS
9237 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
9238 		       seq_out, len, tp->snd_una, tp->snd_max);
9239 		printf("Starting Dump of all rack entries\n");
9240 		TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
9241 			printf("rsm:%p start:%u end:%u\n",
9242 			       rsm, rsm->r_start, rsm->r_end);
9243 		}
9244 		printf("Dump complete\n");
9245 		panic("seq_out not found rack:%p tp:%p",
9246 		      rack, tp);
9247 #endif
9248 	} else {
9249 #ifdef INVARIANTS
9250 		/*
9251 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
9252 		 * flag)
9253 		 */
9254 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
9255 		      seq_out, len, tp->snd_max, tp);
9256 #endif
9257 	}
9258 }
9259 
9260 /*
9261  * Record one of the RTT updates from an ack into
9262  * our sample structure.
9263  */
9264 
9265 static void
9266 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
9267 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
9268 {
9269 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
9270 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
9271 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
9272 	}
9273 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
9274 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
9275 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
9276 	}
9277 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
9278 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
9279 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
9280 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
9281 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
9282 	}
9283 	if ((confidence == 1) &&
9284 	    ((rsm == NULL) ||
9285 	     (rsm->r_just_ret) ||
9286 	     (rsm->r_one_out_nr &&
9287 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
9288 		/*
9289 		 * If the rsm had a just return
9290 		 * hit it then we can't trust the
9291 		 * rtt measurement for buffer deterimination
9292 		 * Note that a confidence of 2, indicates
9293 		 * SACK'd which overrides the r_just_ret or
9294 		 * the r_one_out_nr. If it was a CUM-ACK and
9295 		 * we had only two outstanding, but get an
9296 		 * ack for only 1. Then that also lowers our
9297 		 * confidence.
9298 		 */
9299 		confidence = 0;
9300 	}
9301 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
9302 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
9303 		if (rack->r_ctl.rack_rs.confidence == 0) {
9304 			/*
9305 			 * We take anything with no current confidence
9306 			 * saved.
9307 			 */
9308 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
9309 			rack->r_ctl.rack_rs.confidence = confidence;
9310 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
9311 		} else if (confidence != 0) {
9312 			/*
9313 			 * Once we have a confident number,
9314 			 * we can update it with a smaller
9315 			 * value since this confident number
9316 			 * may include the DSACK time until
9317 			 * the next segment (the second one) arrived.
9318 			 */
9319 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
9320 			rack->r_ctl.rack_rs.confidence = confidence;
9321 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
9322 		}
9323 	}
9324 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
9325 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
9326 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
9327 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
9328 }
9329 
9330 /*
9331  * Collect new round-trip time estimate
9332  * and update averages and current timeout.
9333  */
9334 static void
9335 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
9336 {
9337 	int32_t delta;
9338 	int32_t rtt;
9339 
9340 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
9341 		/* No valid sample */
9342 		return;
9343 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
9344 		/* We are to use the lowest RTT seen in a single ack */
9345 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9346 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
9347 		/* We are to use the highest RTT seen in a single ack */
9348 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
9349 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
9350 		/* We are to use the average RTT seen in a single ack */
9351 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
9352 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
9353 	} else {
9354 #ifdef INVARIANTS
9355 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
9356 #endif
9357 		return;
9358 	}
9359 	if (rtt == 0)
9360 		rtt = 1;
9361 	if (rack->rc_gp_rtt_set == 0) {
9362 		/*
9363 		 * With no RTT we have to accept
9364 		 * even one we are not confident of.
9365 		 */
9366 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
9367 		rack->rc_gp_rtt_set = 1;
9368 	} else if (rack->r_ctl.rack_rs.confidence) {
9369 		/* update the running gp srtt */
9370 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
9371 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
9372 	}
9373 	if (rack->r_ctl.rack_rs.confidence) {
9374 		/*
9375 		 * record the low and high for highly buffered path computation,
9376 		 * we only do this if we are confident (not a retransmission).
9377 		 */
9378 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
9379 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9380 		}
9381 		if (rack->rc_highly_buffered == 0) {
9382 			/*
9383 			 * Currently once we declare a path has
9384 			 * highly buffered there is no going
9385 			 * back, which may be a problem...
9386 			 */
9387 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
9388 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
9389 						     rack->r_ctl.rc_highest_us_rtt,
9390 						     rack->r_ctl.rc_lowest_us_rtt,
9391 						     RACK_RTTS_SEEHBP);
9392 				rack->rc_highly_buffered = 1;
9393 			}
9394 		}
9395 	}
9396 	if ((rack->r_ctl.rack_rs.confidence) ||
9397 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
9398 		/*
9399 		 * If we are highly confident of it <or> it was
9400 		 * never retransmitted we accept it as the last us_rtt.
9401 		 */
9402 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9403 		/* The lowest rtt can be set if its was not retransmited */
9404 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
9405 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9406 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
9407 				rack->r_ctl.rc_lowest_us_rtt = 1;
9408 		}
9409 	}
9410 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9411 	if (tp->t_srtt != 0) {
9412 		/*
9413 		 * We keep a simple srtt in microseconds, like our rtt
9414 		 * measurement. We don't need to do any tricks with shifting
9415 		 * etc. Instead we just add in 1/8th of the new measurement
9416 		 * and subtract out 1/8 of the old srtt. We do the same with
9417 		 * the variance after finding the absolute value of the
9418 		 * difference between this sample and the current srtt.
9419 		 */
9420 		delta = tp->t_srtt - rtt;
9421 		/* Take off 1/8th of the current sRTT */
9422 		tp->t_srtt -= (tp->t_srtt >> 3);
9423 		/* Add in 1/8th of the new RTT just measured */
9424 		tp->t_srtt += (rtt >> 3);
9425 		if (tp->t_srtt <= 0)
9426 			tp->t_srtt = 1;
9427 		/* Now lets make the absolute value of the variance */
9428 		if (delta < 0)
9429 			delta = -delta;
9430 		/* Subtract out 1/8th */
9431 		tp->t_rttvar -= (tp->t_rttvar >> 3);
9432 		/* Add in 1/8th of the new variance we just saw */
9433 		tp->t_rttvar += (delta >> 3);
9434 		if (tp->t_rttvar <= 0)
9435 			tp->t_rttvar = 1;
9436 	} else {
9437 		/*
9438 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
9439 		 * variance to half the rtt (so our first retransmit happens
9440 		 * at 3*rtt).
9441 		 */
9442 		tp->t_srtt = rtt;
9443 		tp->t_rttvar = rtt >> 1;
9444 	}
9445 	rack->rc_srtt_measure_made = 1;
9446 	KMOD_TCPSTAT_INC(tcps_rttupdated);
9447 	if (tp->t_rttupdated < UCHAR_MAX)
9448 		tp->t_rttupdated++;
9449 #ifdef STATS
9450 	if (rack_stats_gets_ms_rtt == 0) {
9451 		/* Send in the microsecond rtt used for rxt timeout purposes */
9452 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
9453 	} else if (rack_stats_gets_ms_rtt == 1) {
9454 		/* Send in the millisecond rtt used for rxt timeout purposes */
9455 		int32_t ms_rtt;
9456 
9457 		/* Round up */
9458 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
9459 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
9460 	} else if (rack_stats_gets_ms_rtt == 2) {
9461 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
9462 		int32_t ms_rtt;
9463 
9464 		/* Round up */
9465 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
9466 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
9467 	}  else {
9468 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
9469 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
9470 	}
9471 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
9472 #endif
9473 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
9474 	/*
9475 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
9476 	 * way we do the smoothing, srtt and rttvar will each average +1/2
9477 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
9478 	 * tick of rounding and 1 extra tick because of +-1/2 tick
9479 	 * uncertainty in the firing of the timer.  The bias will give us
9480 	 * exactly the 1.5 tick we need.  But, because the bias is
9481 	 * statistical, we have to test that we don't drop below the minimum
9482 	 * feasible timer (which is 2 ticks).
9483 	 */
9484 	tp->t_rxtshift = 0;
9485 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9486 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
9487 	rack_log_rtt_sample(rack, rtt);
9488 	tp->t_softerror = 0;
9489 }
9490 
9491 
9492 static void
9493 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
9494 {
9495 	/*
9496 	 * Apply to filter the inbound us-rtt at us_cts.
9497 	 */
9498 	uint32_t old_rtt;
9499 
9500 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
9501 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
9502 			       us_rtt, us_cts);
9503 	if (old_rtt > us_rtt) {
9504 		/* We just hit a new lower rtt time */
9505 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
9506 				     __LINE__, RACK_RTTS_NEWRTT);
9507 		/*
9508 		 * Only count it if its lower than what we saw within our
9509 		 * calculated range.
9510 		 */
9511 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
9512 			if (rack_probertt_lower_within &&
9513 			    rack->rc_gp_dyn_mul &&
9514 			    (rack->use_fixed_rate == 0) &&
9515 			    (rack->rc_always_pace)) {
9516 				/*
9517 				 * We are seeing a new lower rtt very close
9518 				 * to the time that we would have entered probe-rtt.
9519 				 * This is probably due to the fact that a peer flow
9520 				 * has entered probe-rtt. Lets go in now too.
9521 				 */
9522 				uint32_t val;
9523 
9524 				val = rack_probertt_lower_within * rack_time_between_probertt;
9525 				val /= 100;
9526 				if ((rack->in_probe_rtt == 0)  &&
9527 				    (rack->rc_skip_timely == 0) &&
9528 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
9529 					rack_enter_probertt(rack, us_cts);
9530 				}
9531 			}
9532 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
9533 		}
9534 	}
9535 }
9536 
9537 static int
9538 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
9539     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
9540 {
9541 	uint32_t us_rtt;
9542 	int32_t i, all;
9543 	uint32_t t, len_acked;
9544 
9545 	if ((rsm->r_flags & RACK_ACKED) ||
9546 	    (rsm->r_flags & RACK_WAS_ACKED))
9547 		/* Already done */
9548 		return (0);
9549 	if (rsm->r_no_rtt_allowed) {
9550 		/* Not allowed */
9551 		return (0);
9552 	}
9553 	if (ack_type == CUM_ACKED) {
9554 		if (SEQ_GT(th_ack, rsm->r_end)) {
9555 			len_acked = rsm->r_end - rsm->r_start;
9556 			all = 1;
9557 		} else {
9558 			len_acked = th_ack - rsm->r_start;
9559 			all = 0;
9560 		}
9561 	} else {
9562 		len_acked = rsm->r_end - rsm->r_start;
9563 		all = 0;
9564 	}
9565 	if (rsm->r_rtr_cnt == 1) {
9566 
9567 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9568 		if ((int)t <= 0)
9569 			t = 1;
9570 		if (!tp->t_rttlow || tp->t_rttlow > t)
9571 			tp->t_rttlow = t;
9572 		if (!rack->r_ctl.rc_rack_min_rtt ||
9573 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9574 			rack->r_ctl.rc_rack_min_rtt = t;
9575 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
9576 				rack->r_ctl.rc_rack_min_rtt = 1;
9577 			}
9578 		}
9579 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
9580 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9581 		else
9582 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9583 		if (us_rtt == 0)
9584 			us_rtt = 1;
9585 		if (CC_ALGO(tp)->rttsample != NULL) {
9586 			/* Kick the RTT to the CC */
9587 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9588 		}
9589 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
9590 		if (ack_type == SACKED) {
9591 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
9592 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
9593 		} else {
9594 			/*
9595 			 * We need to setup what our confidence
9596 			 * is in this ack.
9597 			 *
9598 			 * If the rsm was app limited and it is
9599 			 * less than a mss in length (the end
9600 			 * of the send) then we have a gap. If we
9601 			 * were app limited but say we were sending
9602 			 * multiple MSS's then we are more confident
9603 			 * int it.
9604 			 *
9605 			 * When we are not app-limited then we see if
9606 			 * the rsm is being included in the current
9607 			 * measurement, we tell this by the app_limited_needs_set
9608 			 * flag.
9609 			 *
9610 			 * Note that being cwnd blocked is not applimited
9611 			 * as well as the pacing delay between packets which
9612 			 * are sending only 1 or 2 MSS's also will show up
9613 			 * in the RTT. We probably need to examine this algorithm
9614 			 * a bit more and enhance it to account for the delay
9615 			 * between rsm's. We could do that by saving off the
9616 			 * pacing delay of each rsm (in an rsm) and then
9617 			 * factoring that in somehow though for now I am
9618 			 * not sure how :)
9619 			 */
9620 			int calc_conf = 0;
9621 
9622 			if (rsm->r_flags & RACK_APP_LIMITED) {
9623 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
9624 					calc_conf = 0;
9625 				else
9626 					calc_conf = 1;
9627 			} else if (rack->app_limited_needs_set == 0) {
9628 				calc_conf = 1;
9629 			} else {
9630 				calc_conf = 0;
9631 			}
9632 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
9633 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
9634 					    calc_conf, rsm, rsm->r_rtr_cnt);
9635 		}
9636 		if ((rsm->r_flags & RACK_TLP) &&
9637 		    (!IN_FASTRECOVERY(tp->t_flags))) {
9638 			/* Segment was a TLP and our retrans matched */
9639 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
9640 				rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
9641 			}
9642 		}
9643 		if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9644 		    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9645 			    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9646 			/* New more recent rack_tmit_time */
9647 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9648 			if (rack->r_ctl.rc_rack_tmit_time == 0)
9649 				rack->r_ctl.rc_rack_tmit_time = 1;
9650 			rack->rc_rack_rtt = t;
9651 		}
9652 		return (1);
9653 	}
9654 	/*
9655 	 * We clear the soft/rxtshift since we got an ack.
9656 	 * There is no assurance we will call the commit() function
9657 	 * so we need to clear these to avoid incorrect handling.
9658 	 */
9659 	tp->t_rxtshift = 0;
9660 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9661 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9662 	tp->t_softerror = 0;
9663 	if (to && (to->to_flags & TOF_TS) &&
9664 	    (ack_type == CUM_ACKED) &&
9665 	    (to->to_tsecr) &&
9666 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
9667 		/*
9668 		 * Now which timestamp does it match? In this block the ACK
9669 		 * must be coming from a previous transmission.
9670 		 */
9671 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
9672 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
9673 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9674 				if ((int)t <= 0)
9675 					t = 1;
9676 				if (CC_ALGO(tp)->rttsample != NULL) {
9677 					/*
9678 					 * Kick the RTT to the CC, here
9679 					 * we lie a bit in that we know the
9680 					 * retransmission is correct even though
9681 					 * we retransmitted. This is because
9682 					 * we match the timestamps.
9683 					 */
9684 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
9685 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
9686 					else
9687 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
9688 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9689 				}
9690 				if ((i + 1) < rsm->r_rtr_cnt) {
9691 					/*
9692 					 * The peer ack'd from our previous
9693 					 * transmission. We have a spurious
9694 					 * retransmission and thus we dont
9695 					 * want to update our rack_rtt.
9696 					 *
9697 					 * Hmm should there be a CC revert here?
9698 					 *
9699 					 */
9700 					return (0);
9701 				}
9702 				if (!tp->t_rttlow || tp->t_rttlow > t)
9703 					tp->t_rttlow = t;
9704 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9705 					rack->r_ctl.rc_rack_min_rtt = t;
9706 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
9707 						rack->r_ctl.rc_rack_min_rtt = 1;
9708 					}
9709 				}
9710 				if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9711 				    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9712 					    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9713 					/* New more recent rack_tmit_time */
9714 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9715 					if (rack->r_ctl.rc_rack_tmit_time == 0)
9716 						rack->r_ctl.rc_rack_tmit_time = 1;
9717 					rack->rc_rack_rtt = t;
9718 				}
9719 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9720 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9721 						    rsm->r_rtr_cnt);
9722 				return (1);
9723 			}
9724 		}
9725 		/* If we are logging log out the sendmap */
9726 		if (tcp_bblogging_on(rack->rc_tp)) {
9727 			for (i = 0; i < rsm->r_rtr_cnt; i++) {
9728 				rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9729 			}
9730 		}
9731 		goto ts_not_found;
9732 	} else {
9733 		/*
9734 		 * Ok its a SACK block that we retransmitted. or a windows
9735 		 * machine without timestamps. We can tell nothing from the
9736 		 * time-stamp since its not there or the time the peer last
9737 		 * received a segment that moved forward its cum-ack point.
9738 		 */
9739 ts_not_found:
9740 		i = rsm->r_rtr_cnt - 1;
9741 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9742 		if ((int)t <= 0)
9743 			t = 1;
9744 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9745 			/*
9746 			 * We retransmitted and the ack came back in less
9747 			 * than the smallest rtt we have observed. We most
9748 			 * likely did an improper retransmit as outlined in
9749 			 * 6.2 Step 2 point 2 in the rack-draft so we
9750 			 * don't want to update our rack_rtt. We in
9751 			 * theory (in future) might want to think about reverting our
9752 			 * cwnd state but we won't for now.
9753 			 */
9754 			return (0);
9755 		} else if (rack->r_ctl.rc_rack_min_rtt) {
9756 			/*
9757 			 * We retransmitted it and the retransmit did the
9758 			 * job.
9759 			 */
9760 			if (!rack->r_ctl.rc_rack_min_rtt ||
9761 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9762 				rack->r_ctl.rc_rack_min_rtt = t;
9763 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
9764 					rack->r_ctl.rc_rack_min_rtt = 1;
9765 				}
9766 			}
9767 			if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9768 			    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9769 				    (uint32_t)rsm->r_tim_lastsent[i]))) {
9770 				/* New more recent rack_tmit_time */
9771 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9772 				if (rack->r_ctl.rc_rack_tmit_time == 0)
9773 					rack->r_ctl.rc_rack_tmit_time = 1;
9774 				rack->rc_rack_rtt = t;
9775 			}
9776 			return (1);
9777 		}
9778 	}
9779 	return (0);
9780 }
9781 
9782 /*
9783  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9784  */
9785 static void
9786 rack_log_sack_passed(struct tcpcb *tp,
9787     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
9788 {
9789 	struct rack_sendmap *nrsm;
9790 	uint32_t thresh;
9791 
9792 	/* Get our rxt threshold for lost consideration */
9793 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
9794 	/* Now start looking at rsm's */
9795 	nrsm = rsm;
9796 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9797 	    rack_head, r_tnext) {
9798 		if (nrsm == rsm) {
9799 			/* Skip original segment he is acked */
9800 			continue;
9801 		}
9802 		if (nrsm->r_flags & RACK_ACKED) {
9803 			/*
9804 			 * Skip ack'd segments, though we
9805 			 * should not see these, since tmap
9806 			 * should not have ack'd segments.
9807 			 */
9808 			continue;
9809 		}
9810 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9811 			/*
9812 			 * If the peer dropped the rwnd on
9813 			 * these then we don't worry about them.
9814 			 */
9815 			continue;
9816 		}
9817 		/* Check lost state */
9818 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
9819 			uint32_t exp;
9820 
9821 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
9822 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
9823 				/* We consider it lost */
9824 				nrsm->r_flags |= RACK_WAS_LOST;
9825 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
9826 			}
9827 		}
9828 		if (nrsm->r_flags & RACK_SACK_PASSED) {
9829 			/*
9830 			 * We found one that is already marked
9831 			 * passed, we have been here before and
9832 			 * so all others below this are marked.
9833 			 */
9834 			break;
9835 		}
9836 		nrsm->r_flags |= RACK_SACK_PASSED;
9837 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9838 	}
9839 }
9840 
9841 static void
9842 rack_need_set_test(struct tcpcb *tp,
9843 		   struct tcp_rack *rack,
9844 		   struct rack_sendmap *rsm,
9845 		   tcp_seq th_ack,
9846 		   int line,
9847 		   int use_which)
9848 {
9849 	struct rack_sendmap *s_rsm;
9850 
9851 	if ((tp->t_flags & TF_GPUTINPROG) &&
9852 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9853 		/*
9854 		 * We were app limited, and this ack
9855 		 * butts up or goes beyond the point where we want
9856 		 * to start our next measurement. We need
9857 		 * to record the new gput_ts as here and
9858 		 * possibly update the start sequence.
9859 		 */
9860 		uint32_t seq, ts;
9861 
9862 		if (rsm->r_rtr_cnt > 1) {
9863 			/*
9864 			 * This is a retransmit, can we
9865 			 * really make any assessment at this
9866 			 * point?  We are not really sure of
9867 			 * the timestamp, is it this or the
9868 			 * previous transmission?
9869 			 *
9870 			 * Lets wait for something better that
9871 			 * is not retransmitted.
9872 			 */
9873 			return;
9874 		}
9875 		seq = tp->gput_seq;
9876 		ts = tp->gput_ts;
9877 		rack->app_limited_needs_set = 0;
9878 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9879 		/* Do we start at a new end? */
9880 		if ((use_which == RACK_USE_BEG) &&
9881 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9882 			/*
9883 			 * When we get an ACK that just eats
9884 			 * up some of the rsm, we set RACK_USE_BEG
9885 			 * since whats at r_start (i.e. th_ack)
9886 			 * is left unacked and thats where the
9887 			 * measurement now starts.
9888 			 */
9889 			tp->gput_seq = rsm->r_start;
9890 		}
9891 		if ((use_which == RACK_USE_END) &&
9892 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9893 			/*
9894 			 * We use the end when the cumack
9895 			 * is moving forward and completely
9896 			 * deleting the rsm passed so basically
9897 			 * r_end holds th_ack.
9898 			 *
9899 			 * For SACK's we also want to use the end
9900 			 * since this piece just got sacked and
9901 			 * we want to target anything after that
9902 			 * in our measurement.
9903 			 */
9904 			tp->gput_seq = rsm->r_end;
9905 		}
9906 		if (use_which == RACK_USE_END_OR_THACK) {
9907 			/*
9908 			 * special case for ack moving forward,
9909 			 * not a sack, we need to move all the
9910 			 * way up to where this ack cum-ack moves
9911 			 * to.
9912 			 */
9913 			if (SEQ_GT(th_ack, rsm->r_end))
9914 				tp->gput_seq = th_ack;
9915 			else
9916 				tp->gput_seq = rsm->r_end;
9917 		}
9918 		if (SEQ_LT(tp->gput_seq, tp->snd_max))
9919 			s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9920 		else
9921 			s_rsm = NULL;
9922 		/*
9923 		 * Pick up the correct send time if we can the rsm passed in
9924 		 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9925 		 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9926 		 * find a different seq i.e. the next send up.
9927 		 *
9928 		 * If that has not been sent, s_rsm will be NULL and we must
9929 		 * arrange it so this function will get called again by setting
9930 		 * app_limited_needs_set.
9931 		 */
9932 		if (s_rsm)
9933 			rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9934 		else {
9935 			/* If we hit here we have to have *not* sent tp->gput_seq */
9936 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9937 			/* Set it up so we will go through here again */
9938 			rack->app_limited_needs_set = 1;
9939 		}
9940 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9941 			/*
9942 			 * We moved beyond this guy's range, re-calculate
9943 			 * the new end point.
9944 			 */
9945 			if (rack->rc_gp_filled == 0) {
9946 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9947 			} else {
9948 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9949 			}
9950 		}
9951 		/*
9952 		 * We are moving the goal post, we may be able to clear the
9953 		 * measure_saw_probe_rtt flag.
9954 		 */
9955 		if ((rack->in_probe_rtt == 0) &&
9956 		    (rack->measure_saw_probe_rtt) &&
9957 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9958 			rack->measure_saw_probe_rtt = 0;
9959 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9960 					   seq, tp->gput_seq,
9961 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9962 					    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9963 					   5, line, NULL, 0);
9964 		if (rack->rc_gp_filled &&
9965 		    ((tp->gput_ack - tp->gput_seq) <
9966 		     max(rc_init_window(rack), (MIN_GP_WIN *
9967 						ctf_fixed_maxseg(tp))))) {
9968 			uint32_t ideal_amount;
9969 
9970 			ideal_amount = rack_get_measure_window(tp, rack);
9971 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9972 				/*
9973 				 * There is no sense of continuing this measurement
9974 				 * because its too small to gain us anything we
9975 				 * trust. Skip it and that way we can start a new
9976 				 * measurement quicker.
9977 				 */
9978 				tp->t_flags &= ~TF_GPUTINPROG;
9979 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9980 							   0, 0,
9981 							   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9982 							    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9983 							   6, __LINE__, NULL, 0);
9984 			} else {
9985 				/*
9986 				 * Reset the window further out.
9987 				 */
9988 				tp->gput_ack = tp->gput_seq + ideal_amount;
9989 			}
9990 		}
9991 		rack_tend_gp_marks(tp, rack);
9992 		rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9993 	}
9994 }
9995 
9996 static inline int
9997 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9998 {
9999 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
10000 		/* Behind our TLP definition or right at */
10001 		return (0);
10002 	}
10003 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
10004 		/* The start is beyond or right at our end of TLP definition */
10005 		return (0);
10006 	}
10007 	/* It has to be a sub-part of the original TLP recorded */
10008 	return (1);
10009 }
10010 
10011 static uint32_t
10012 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
10013 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
10014 		   uint32_t segsiz)
10015 {
10016 	uint32_t start, end, changed = 0;
10017 	struct rack_sendmap stack_map;
10018 	struct rack_sendmap *rsm, *nrsm, *prev, *next;
10019 	int insret __diagused;
10020 	int32_t used_ref = 1;
10021 	int can_use_hookery = 0;
10022 
10023 	start = sack->start;
10024 	end = sack->end;
10025 	rsm = *prsm;
10026 
10027 do_rest_ofb:
10028 	if ((rsm == NULL) ||
10029 	    (SEQ_LT(end, rsm->r_start)) ||
10030 	    (SEQ_GEQ(start, rsm->r_end)) ||
10031 	    (SEQ_LT(start, rsm->r_start))) {
10032 		/*
10033 		 * We are not in the right spot,
10034 		 * find the correct spot in the tree.
10035 		 */
10036 		used_ref = 0;
10037 		rsm = tqhash_find(rack->r_ctl.tqh, start);
10038 	}
10039 	if (rsm == NULL) {
10040 		/* TSNH */
10041 		goto out;
10042 	}
10043 	/* Ok we have an ACK for some piece of this rsm */
10044 	if (rsm->r_start != start) {
10045 		if ((rsm->r_flags & RACK_ACKED) == 0) {
10046 			/*
10047 			 * Before any splitting or hookery is
10048 			 * done is it a TLP of interest i.e. rxt?
10049 			 */
10050 			if ((rsm->r_flags & RACK_TLP) &&
10051 			    (rsm->r_rtr_cnt > 1)) {
10052 				/*
10053 				 * We are splitting a rxt TLP, check
10054 				 * if we need to save off the start/end
10055 				 */
10056 				if (rack->rc_last_tlp_acked_set &&
10057 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10058 					/*
10059 					 * We already turned this on since we are inside
10060 					 * the previous one was a partially sack now we
10061 					 * are getting another one (maybe all of it).
10062 					 *
10063 					 */
10064 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10065 					/*
10066 					 * Lets make sure we have all of it though.
10067 					 */
10068 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10069 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10070 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10071 								     rack->r_ctl.last_tlp_acked_end);
10072 					}
10073 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10074 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10075 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10076 								     rack->r_ctl.last_tlp_acked_end);
10077 					}
10078 				} else {
10079 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10080 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10081 					rack->rc_last_tlp_past_cumack = 0;
10082 					rack->rc_last_tlp_acked_set = 1;
10083 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10084 				}
10085 			}
10086 			/**
10087 			 * Need to split this in two pieces the before and after,
10088 			 * the before remains in the map, the after must be
10089 			 * added. In other words we have:
10090 			 * rsm        |--------------|
10091 			 * sackblk        |------->
10092 			 * rsm will become
10093 			 *     rsm    |---|
10094 			 * and nrsm will be  the sacked piece
10095 			 *     nrsm       |----------|
10096 			 *
10097 			 * But before we start down that path lets
10098 			 * see if the sack spans over on top of
10099 			 * the next guy and it is already sacked.
10100 			 *
10101 			 */
10102 			/*
10103 			 * Hookery can only be used if the two entries
10104 			 * are in the same bucket and neither one of
10105 			 * them staddle the bucket line.
10106 			 */
10107 			next = tqhash_next(rack->r_ctl.tqh, rsm);
10108 			if (next &&
10109 			    (rsm->bindex == next->bindex) &&
10110 			    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
10111 			    ((next->r_flags & RACK_STRADDLE) == 0) &&
10112 			    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
10113 			    ((next->r_flags & RACK_IS_PCM) == 0) &&
10114 			    (rsm->r_flags & RACK_IN_GP_WIN) &&
10115 			    (next->r_flags & RACK_IN_GP_WIN))
10116 				can_use_hookery = 1;
10117 			else
10118 				can_use_hookery = 0;
10119 			if (next && can_use_hookery &&
10120 			    (next->r_flags & RACK_ACKED) &&
10121 			    SEQ_GEQ(end, next->r_start)) {
10122 				/**
10123 				 * So the next one is already acked, and
10124 				 * we can thus by hookery use our stack_map
10125 				 * to reflect the piece being sacked and
10126 				 * then adjust the two tree entries moving
10127 				 * the start and ends around. So we start like:
10128 				 *  rsm     |------------|             (not-acked)
10129 				 *  next                 |-----------| (acked)
10130 				 *  sackblk        |-------->
10131 				 *  We want to end like so:
10132 				 *  rsm     |------|                   (not-acked)
10133 				 *  next           |-----------------| (acked)
10134 				 *  nrsm           |-----|
10135 				 * Where nrsm is a temporary stack piece we
10136 				 * use to update all the gizmos.
10137 				 */
10138 				/* Copy up our fudge block */
10139 				nrsm = &stack_map;
10140 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
10141 				/* Now adjust our tree blocks */
10142 				tqhash_update_end(rack->r_ctl.tqh, rsm, start);
10143 				next->r_start = start;
10144  				rsm->r_flags |= RACK_SHUFFLED;
10145 				next->r_flags |= RACK_SHUFFLED;
10146 				/* Now we must adjust back where next->m is */
10147 				rack_setup_offset_for_rsm(rack, rsm, next);
10148 				/*
10149 				 * Which timestamp do we keep? It is rather
10150 				 * important in GP measurements to have the
10151 				 * accurate end of the send window.
10152 				 *
10153 				 * We keep the largest value, which is the newest
10154 				 * send. We do this in case a segment that is
10155 				 * joined together and not part of a GP estimate
10156 				 * later gets expanded into the GP estimate.
10157 				 *
10158 				 * We prohibit the merging of unlike kinds i.e.
10159 				 * all pieces that are in the GP estimate can be
10160 				 * merged and all pieces that are not in a GP estimate
10161 				 * can be merged, but not disimilar pieces. Combine
10162 				 * this with taking the highest here and we should
10163 				 * be ok unless of course the client reneges. Then
10164 				 * all bets are off.
10165 				 */
10166 				if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
10167 				    nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
10168 					next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
10169 				/*
10170 				 * And we must keep the newest ack arrival time.
10171 				 */
10172 				if (next->r_ack_arrival <
10173 				    rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
10174 					next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10175 
10176 
10177 				/* We don't need to adjust rsm, it did not change */
10178 				/* Clear out the dup ack count of the remainder */
10179 				rsm->r_dupack = 0;
10180 				rsm->r_just_ret = 0;
10181 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10182 				/* Now lets make sure our fudge block is right */
10183 				nrsm->r_start = start;
10184 				/* Now lets update all the stats and such */
10185 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
10186 				if (rack->app_limited_needs_set)
10187 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
10188 				changed += (nrsm->r_end - nrsm->r_start);
10189 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
10190 				if (rsm->r_flags & RACK_WAS_LOST) {
10191 					int my_chg;
10192 
10193 					my_chg = (nrsm->r_end - nrsm->r_start);
10194 					KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
10195 						("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10196 					if (my_chg <= rack->r_ctl.rc_considered_lost)
10197 						rack->r_ctl.rc_considered_lost -= my_chg;
10198 					else
10199 						rack->r_ctl.rc_considered_lost = 0;
10200 				}
10201 				if (nrsm->r_flags & RACK_SACK_PASSED) {
10202 					rack->r_ctl.rc_reorder_ts = cts;
10203 					if (rack->r_ctl.rc_reorder_ts == 0)
10204 						rack->r_ctl.rc_reorder_ts = 1;
10205 				}
10206 				/*
10207 				 * Now we want to go up from rsm (the
10208 				 * one left un-acked) to the next one
10209 				 * in the tmap. We do this so when
10210 				 * we walk backwards we include marking
10211 				 * sack-passed on rsm (The one passed in
10212 				 * is skipped since it is generally called
10213 				 * on something sacked before removing it
10214 				 * from the tmap).
10215 				 */
10216 				if (rsm->r_in_tmap) {
10217 					nrsm = TAILQ_NEXT(rsm, r_tnext);
10218 					/*
10219 					 * Now that we have the next
10220 					 * one walk backwards from there.
10221 					 */
10222 					if (nrsm && nrsm->r_in_tmap)
10223 						rack_log_sack_passed(tp, rack, nrsm, cts);
10224 				}
10225 				/* Now are we done? */
10226 				if (SEQ_LT(end, next->r_end) ||
10227 				    (end == next->r_end)) {
10228 					/* Done with block */
10229 					goto out;
10230 				}
10231 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
10232 				counter_u64_add(rack_sack_used_next_merge, 1);
10233 				/* Postion for the next block */
10234 				start = next->r_end;
10235 				rsm = tqhash_next(rack->r_ctl.tqh, next);
10236 				if (rsm == NULL)
10237 					goto out;
10238 			} else {
10239 				/**
10240 				 * We can't use any hookery here, so we
10241 				 * need to split the map. We enter like
10242 				 * so:
10243 				 *  rsm      |--------|
10244 				 *  sackblk       |----->
10245 				 * We will add the new block nrsm and
10246 				 * that will be the new portion, and then
10247 				 * fall through after reseting rsm. So we
10248 				 * split and look like this:
10249 				 *  rsm      |----|
10250 				 *  sackblk       |----->
10251 				 *  nrsm          |---|
10252 				 * We then fall through reseting
10253 				 * rsm to nrsm, so the next block
10254 				 * picks it up.
10255 				 */
10256 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10257 				if (nrsm == NULL) {
10258 					/*
10259 					 * failed XXXrrs what can we do but loose the sack
10260 					 * info?
10261 					 */
10262 					goto out;
10263 				}
10264 				counter_u64_add(rack_sack_splits, 1);
10265 				rack_clone_rsm(rack, nrsm, rsm, start);
10266 				rsm->r_just_ret = 0;
10267 #ifndef INVARIANTS
10268 				(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
10269 #else
10270 				if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
10271 					panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
10272 					      nrsm, insret, rack, rsm);
10273 				}
10274 #endif
10275 				if (rsm->r_in_tmap) {
10276 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10277 					nrsm->r_in_tmap = 1;
10278 				}
10279 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
10280 				rsm->r_flags &= (~RACK_HAS_FIN);
10281 				/* Position us to point to the new nrsm that starts the sack blk */
10282 				rsm = nrsm;
10283 			}
10284 		} else {
10285 			/* Already sacked this piece */
10286 			counter_u64_add(rack_sack_skipped_acked, 1);
10287 			if (end == rsm->r_end) {
10288 				/* Done with block */
10289 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10290 				goto out;
10291 			} else if (SEQ_LT(end, rsm->r_end)) {
10292 				/* A partial sack to a already sacked block */
10293 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10294 				goto out;
10295 			} else {
10296 				/*
10297 				 * The end goes beyond this guy
10298 				 * reposition the start to the
10299 				 * next block.
10300 				 */
10301 				start = rsm->r_end;
10302 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10303 				if (rsm == NULL)
10304 					goto out;
10305 			}
10306 		}
10307 	}
10308 	if (SEQ_GEQ(end, rsm->r_end)) {
10309 		/**
10310 		 * The end of this block is either beyond this guy or right
10311 		 * at this guy. I.e.:
10312 		 *  rsm ---                 |-----|
10313 		 *  end                     |-----|
10314 		 *  <or>
10315 		 *  end                     |---------|
10316 		 */
10317 		if ((rsm->r_flags & RACK_ACKED) == 0) {
10318 			/*
10319 			 * Is it a TLP of interest?
10320 			 */
10321 			if ((rsm->r_flags & RACK_TLP) &&
10322 			    (rsm->r_rtr_cnt > 1)) {
10323 				/*
10324 				 * We are splitting a rxt TLP, check
10325 				 * if we need to save off the start/end
10326 				 */
10327 				if (rack->rc_last_tlp_acked_set &&
10328 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10329 					/*
10330 					 * We already turned this on since we are inside
10331 					 * the previous one was a partially sack now we
10332 					 * are getting another one (maybe all of it).
10333 					 */
10334 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10335 					/*
10336 					 * Lets make sure we have all of it though.
10337 					 */
10338 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10339 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10340 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10341 								     rack->r_ctl.last_tlp_acked_end);
10342 					}
10343 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10344 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10345 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10346 								     rack->r_ctl.last_tlp_acked_end);
10347 					}
10348 				} else {
10349 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10350 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10351 					rack->rc_last_tlp_past_cumack = 0;
10352 					rack->rc_last_tlp_acked_set = 1;
10353 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10354 				}
10355 			}
10356 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
10357 			changed += (rsm->r_end - rsm->r_start);
10358 			/* You get a count for acking a whole segment or more */
10359 			if (rsm->r_flags & RACK_WAS_LOST) {
10360 				int my_chg;
10361 
10362 				my_chg = (rsm->r_end - rsm->r_start);
10363 				rsm->r_flags &= ~RACK_WAS_LOST;
10364 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
10365 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10366 				if (my_chg <= rack->r_ctl.rc_considered_lost)
10367 					rack->r_ctl.rc_considered_lost -= my_chg;
10368 				else
10369 					rack->r_ctl.rc_considered_lost = 0;
10370 			}
10371 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
10372 			if (rsm->r_in_tmap) /* should be true */
10373 				rack_log_sack_passed(tp, rack, rsm, cts);
10374 			/* Is Reordering occuring? */
10375 			if (rsm->r_flags & RACK_SACK_PASSED) {
10376 				rsm->r_flags &= ~RACK_SACK_PASSED;
10377 				rack->r_ctl.rc_reorder_ts = cts;
10378 				if (rack->r_ctl.rc_reorder_ts == 0)
10379 					rack->r_ctl.rc_reorder_ts = 1;
10380 			}
10381 			if (rack->app_limited_needs_set)
10382 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10383 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10384 			rsm->r_flags |= RACK_ACKED;
10385 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
10386 			if (rsm->r_in_tmap) {
10387 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10388 				rsm->r_in_tmap = 0;
10389 			}
10390 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
10391 		} else {
10392 			counter_u64_add(rack_sack_skipped_acked, 1);
10393 		}
10394 		if (end == rsm->r_end) {
10395 			/* This block only - done, setup for next */
10396 			goto out;
10397 		}
10398 		/*
10399 		 * There is more not coverend by this rsm move on
10400 		 * to the next block in the tail queue hash table.
10401 		 */
10402 		nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
10403 		start = rsm->r_end;
10404 		rsm = nrsm;
10405 		if (rsm == NULL)
10406 			goto out;
10407 		goto do_rest_ofb;
10408 	}
10409 	/**
10410 	 * The end of this sack block is smaller than
10411 	 * our rsm i.e.:
10412 	 *  rsm ---                 |-----|
10413 	 *  end                     |--|
10414 	 */
10415 	if ((rsm->r_flags & RACK_ACKED) == 0) {
10416 		/*
10417 		 * Is it a TLP of interest?
10418 		 */
10419 		if ((rsm->r_flags & RACK_TLP) &&
10420 		    (rsm->r_rtr_cnt > 1)) {
10421 			/*
10422 			 * We are splitting a rxt TLP, check
10423 			 * if we need to save off the start/end
10424 			 */
10425 			if (rack->rc_last_tlp_acked_set &&
10426 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10427 				/*
10428 				 * We already turned this on since we are inside
10429 				 * the previous one was a partially sack now we
10430 				 * are getting another one (maybe all of it).
10431 				 */
10432 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10433 				/*
10434 				 * Lets make sure we have all of it though.
10435 				 */
10436 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10437 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10438 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10439 							     rack->r_ctl.last_tlp_acked_end);
10440 				}
10441 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10442 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10443 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10444 							     rack->r_ctl.last_tlp_acked_end);
10445 				}
10446 			} else {
10447 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10448 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10449 				rack->rc_last_tlp_past_cumack = 0;
10450 				rack->rc_last_tlp_acked_set = 1;
10451 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10452 			}
10453 		}
10454 		/*
10455 		 * Hookery can only be used if the two entries
10456 		 * are in the same bucket and neither one of
10457 		 * them staddle the bucket line.
10458 		 */
10459 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10460 		if (prev &&
10461 		    (rsm->bindex == prev->bindex) &&
10462 		    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
10463 		    ((prev->r_flags & RACK_STRADDLE) == 0) &&
10464 		    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
10465 		    ((prev->r_flags & RACK_IS_PCM) == 0) &&
10466 		    (rsm->r_flags & RACK_IN_GP_WIN) &&
10467 		    (prev->r_flags & RACK_IN_GP_WIN))
10468 			can_use_hookery = 1;
10469 		else
10470 			can_use_hookery = 0;
10471 		if (prev && can_use_hookery &&
10472 		    (prev->r_flags & RACK_ACKED)) {
10473 			/**
10474 			 * Goal, we want the right remainder of rsm to shrink
10475 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
10476 			 * We want to expand prev to go all the way
10477 			 * to prev->r_end <- end.
10478 			 * so in the tree we have before:
10479 			 *   prev     |--------|         (acked)
10480 			 *   rsm               |-------| (non-acked)
10481 			 *   sackblk           |-|
10482 			 * We churn it so we end up with
10483 			 *   prev     |----------|       (acked)
10484 			 *   rsm                 |-----| (non-acked)
10485 			 *   nrsm              |-| (temporary)
10486 			 *
10487 			 * Note if either prev/rsm is a TLP we don't
10488 			 * do this.
10489 			 */
10490 			nrsm = &stack_map;
10491 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
10492 			tqhash_update_end(rack->r_ctl.tqh, prev, end);
10493 			rsm->r_start = end;
10494 			rsm->r_flags |= RACK_SHUFFLED;
10495 			prev->r_flags |= RACK_SHUFFLED;
10496 			/* Now adjust nrsm (stack copy) to be
10497 			 * the one that is the small
10498 			 * piece that was "sacked".
10499 			 */
10500 			nrsm->r_end = end;
10501 			rsm->r_dupack = 0;
10502 			/*
10503 			 * Which timestamp do we keep? It is rather
10504 			 * important in GP measurements to have the
10505 			 * accurate end of the send window.
10506 			 *
10507 			 * We keep the largest value, which is the newest
10508 			 * send. We do this in case a segment that is
10509 			 * joined together and not part of a GP estimate
10510 			 * later gets expanded into the GP estimate.
10511 			 *
10512 			 * We prohibit the merging of unlike kinds i.e.
10513 			 * all pieces that are in the GP estimate can be
10514 			 * merged and all pieces that are not in a GP estimate
10515 			 * can be merged, but not disimilar pieces. Combine
10516 			 * this with taking the highest here and we should
10517 			 * be ok unless of course the client reneges. Then
10518 			 * all bets are off.
10519 			 */
10520 			if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
10521 			   nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
10522 				prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
10523 			}
10524 			/*
10525 			 * And we must keep the newest ack arrival time.
10526 			 */
10527 
10528 			if(prev->r_ack_arrival <
10529 			   rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
10530 				prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10531 
10532 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10533 			/*
10534 			 * Now that the rsm has had its start moved forward
10535 			 * lets go ahead and get its new place in the world.
10536 			 */
10537 			rack_setup_offset_for_rsm(rack, prev, rsm);
10538 			/*
10539 			 * Now nrsm is our new little piece
10540 			 * that is acked (which was merged
10541 			 * to prev). Update the rtt and changed
10542 			 * based on that. Also check for reordering.
10543 			 */
10544 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
10545 			if (rack->app_limited_needs_set)
10546 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
10547 			changed += (nrsm->r_end - nrsm->r_start);
10548 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
10549 			if (rsm->r_flags & RACK_WAS_LOST) {
10550 				int my_chg;
10551 
10552 				my_chg = (nrsm->r_end - nrsm->r_start);
10553 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
10554 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10555 				if (my_chg <= rack->r_ctl.rc_considered_lost)
10556 					rack->r_ctl.rc_considered_lost -= my_chg;
10557 				else
10558 					rack->r_ctl.rc_considered_lost = 0;
10559 			}
10560 			if (nrsm->r_flags & RACK_SACK_PASSED) {
10561 				rack->r_ctl.rc_reorder_ts = cts;
10562 				if (rack->r_ctl.rc_reorder_ts == 0)
10563 					rack->r_ctl.rc_reorder_ts = 1;
10564 			}
10565 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
10566 			rsm = prev;
10567 			counter_u64_add(rack_sack_used_prev_merge, 1);
10568 		} else {
10569 			/**
10570 			 * This is the case where our previous
10571 			 * block is not acked either, so we must
10572 			 * split the block in two.
10573 			 */
10574 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10575 			if (nrsm == NULL) {
10576 				/* failed rrs what can we do but loose the sack info? */
10577 				goto out;
10578 			}
10579 			if ((rsm->r_flags & RACK_TLP) &&
10580 			    (rsm->r_rtr_cnt > 1)) {
10581 				/*
10582 				 * We are splitting a rxt TLP, check
10583 				 * if we need to save off the start/end
10584 				 */
10585 				if (rack->rc_last_tlp_acked_set &&
10586 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10587 					/*
10588 					 * We already turned this on since this block is inside
10589 					 * the previous one was a partially sack now we
10590 					 * are getting another one (maybe all of it).
10591 					 */
10592 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10593 					/*
10594 					 * Lets make sure we have all of it though.
10595 					 */
10596 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10597 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10598 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10599 								     rack->r_ctl.last_tlp_acked_end);
10600 					}
10601 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10602 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10603 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10604 								     rack->r_ctl.last_tlp_acked_end);
10605 					}
10606 				} else {
10607 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10608 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10609 					rack->rc_last_tlp_acked_set = 1;
10610 					rack->rc_last_tlp_past_cumack = 0;
10611 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10612 				}
10613 			}
10614 			/**
10615 			 * In this case nrsm becomes
10616 			 * nrsm->r_start = end;
10617 			 * nrsm->r_end = rsm->r_end;
10618 			 * which is un-acked.
10619 			 * <and>
10620 			 * rsm->r_end = nrsm->r_start;
10621 			 * i.e. the remaining un-acked
10622 			 * piece is left on the left
10623 			 * hand side.
10624 			 *
10625 			 * So we start like this
10626 			 * rsm      |----------| (not acked)
10627 			 * sackblk  |---|
10628 			 * build it so we have
10629 			 * rsm      |---|         (acked)
10630 			 * nrsm         |------|  (not acked)
10631 			 */
10632 			counter_u64_add(rack_sack_splits, 1);
10633 			rack_clone_rsm(rack, nrsm, rsm, end);
10634 			rsm->r_flags &= (~RACK_HAS_FIN);
10635 			rsm->r_just_ret = 0;
10636 #ifndef INVARIANTS
10637 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
10638 #else
10639 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
10640 				panic("Insert in tailq_hash of %p fails ret:% rack:%p rsm:%p",
10641 				      nrsm, insret, rack, rsm);
10642 			}
10643 #endif
10644 			if (rsm->r_in_tmap) {
10645 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10646 				nrsm->r_in_tmap = 1;
10647 			}
10648 			nrsm->r_dupack = 0;
10649 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
10650 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
10651 			changed += (rsm->r_end - rsm->r_start);
10652 			if (rsm->r_flags & RACK_WAS_LOST) {
10653 				int my_chg;
10654 
10655 				my_chg = (rsm->r_end - rsm->r_start);
10656 				rsm->r_flags &= ~RACK_WAS_LOST;
10657 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
10658 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10659 				if (my_chg <= rack->r_ctl.rc_considered_lost)
10660 					rack->r_ctl.rc_considered_lost -= my_chg;
10661 				else
10662 					rack->r_ctl.rc_considered_lost = 0;
10663 			}
10664 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
10665 
10666 			if (rsm->r_in_tmap) /* should be true */
10667 				rack_log_sack_passed(tp, rack, rsm, cts);
10668 			/* Is Reordering occuring? */
10669 			if (rsm->r_flags & RACK_SACK_PASSED) {
10670 				rsm->r_flags &= ~RACK_SACK_PASSED;
10671 				rack->r_ctl.rc_reorder_ts = cts;
10672 				if (rack->r_ctl.rc_reorder_ts == 0)
10673 					rack->r_ctl.rc_reorder_ts = 1;
10674 			}
10675 			if (rack->app_limited_needs_set)
10676 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10677 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10678 			rsm->r_flags |= RACK_ACKED;
10679 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
10680 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
10681 			if (rsm->r_in_tmap) {
10682 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10683 				rsm->r_in_tmap = 0;
10684 			}
10685 		}
10686 	} else if (start != end){
10687 		/*
10688 		 * The block was already acked.
10689 		 */
10690 		counter_u64_add(rack_sack_skipped_acked, 1);
10691 	}
10692 out:
10693 	if (rsm &&
10694 	    ((rsm->r_flags & RACK_TLP) == 0) &&
10695 	    (rsm->r_flags & RACK_ACKED)) {
10696 		/*
10697 		 * Now can we merge where we worked
10698 		 * with either the previous or
10699 		 * next block?
10700 		 */
10701 		next = tqhash_next(rack->r_ctl.tqh, rsm);
10702 		while (next) {
10703 			if (next->r_flags & RACK_TLP)
10704 				break;
10705 			/* Only allow merges between ones in or out of GP window */
10706 			if ((next->r_flags & RACK_IN_GP_WIN) &&
10707 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10708 				break;
10709 			}
10710 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10711 			    ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10712 				break;
10713 			}
10714 			if (rsm->bindex != next->bindex)
10715 				break;
10716 			if (rsm->r_flags & RACK_STRADDLE)
10717 				break;
10718 			if (rsm->r_flags & RACK_IS_PCM)
10719 				break;
10720 			if (next->r_flags & RACK_STRADDLE)
10721 				break;
10722 			if (next->r_flags & RACK_IS_PCM)
10723 				break;
10724 			if (next->r_flags & RACK_ACKED) {
10725 				/* yep this and next can be merged */
10726 				rsm = rack_merge_rsm(rack, rsm, next);
10727 				next = tqhash_next(rack->r_ctl.tqh, rsm);
10728 			} else
10729 				break;
10730 		}
10731 		/* Now what about the previous? */
10732 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10733 		while (prev) {
10734 			if (prev->r_flags & RACK_TLP)
10735 				break;
10736 			/* Only allow merges between ones in or out of GP window */
10737 			if ((prev->r_flags & RACK_IN_GP_WIN) &&
10738 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10739 				break;
10740 			}
10741 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10742 			    ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10743 				break;
10744 			}
10745 			if (rsm->bindex != prev->bindex)
10746 				break;
10747 			if (rsm->r_flags & RACK_STRADDLE)
10748 				break;
10749 			if (rsm->r_flags & RACK_IS_PCM)
10750 				break;
10751 			if (prev->r_flags & RACK_STRADDLE)
10752 				break;
10753 			if (prev->r_flags & RACK_IS_PCM)
10754 				break;
10755 			if (prev->r_flags & RACK_ACKED) {
10756 				/* yep the previous and this can be merged */
10757 				rsm = rack_merge_rsm(rack, prev, rsm);
10758 				prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10759 			} else
10760 				break;
10761 		}
10762 	}
10763 	if (used_ref == 0) {
10764 		counter_u64_add(rack_sack_proc_all, 1);
10765 	} else {
10766 		counter_u64_add(rack_sack_proc_short, 1);
10767 	}
10768 	/* Save off the next one for quick reference. */
10769 	nrsm = tqhash_find(rack->r_ctl.tqh, end);
10770 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
10771 	if (IN_RECOVERY(tp->t_flags)) {
10772 		rack->r_ctl.bytes_acked_in_recovery += changed;
10773 	}
10774 	return (changed);
10775 }
10776 
10777 static void inline
10778 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10779 {
10780 	struct rack_sendmap *tmap;
10781 
10782 	tmap = NULL;
10783 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
10784 		/* Its no longer sacked, mark it so */
10785 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10786 #ifdef INVARIANTS
10787 		if (rsm->r_in_tmap) {
10788 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
10789 			      rack, rsm, rsm->r_flags);
10790 		}
10791 #endif
10792 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10793 		/* Rebuild it into our tmap */
10794 		if (tmap == NULL) {
10795 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10796 			tmap = rsm;
10797 		} else {
10798 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10799 			tmap = rsm;
10800 		}
10801 		tmap->r_in_tmap = 1;
10802 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10803 	}
10804 	/*
10805 	 * Now lets possibly clear the sack filter so we start
10806 	 * recognizing sacks that cover this area.
10807 	 */
10808 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10809 
10810 }
10811 
10812 
10813 static void inline
10814 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10815 {
10816 	/*
10817 	 * We look at advancing the end send time for our GP
10818 	 * measurement tracking only as the cumulative acknowledgment
10819 	 * moves forward. You might wonder about this, why not
10820 	 * at every transmission or retransmission within the
10821 	 * GP window update the rc_gp_cumack_ts? Well its rather
10822 	 * nuanced but basically the GP window *may* expand (as
10823 	 * it does below) or worse and harder to track it may shrink.
10824 	 *
10825 	 * This last makes it impossible to track at the time of
10826 	 * the send, since you may set forward your rc_gp_cumack_ts
10827 	 * when you send, because that send *is* in your currently
10828 	 * "guessed" window, but then it shrinks. Now which was
10829 	 * the send time of the last bytes in the window, by the
10830 	 * time you ask that question that part of the sendmap
10831 	 * is freed. So you don't know and you will have too
10832 	 * long of send window. Instead by updating the time
10833 	 * marker only when the cumack advances this assures us
10834 	 * that we will have only the sends in the window of our
10835 	 * GP measurement.
10836 	 *
10837 	 * Another complication from this is the
10838 	 * merging of sendmap entries. During SACK processing this
10839 	 * can happen to conserve the sendmap size. That breaks
10840 	 * everything down in tracking the send window of the GP
10841 	 * estimate. So to prevent that and keep it working with
10842 	 * a tiny bit more limited merging, we only allow like
10843 	 * types to be merged. I.e. if two sends are in the GP window
10844 	 * then its ok to merge them together. If two sends are not
10845 	 * in the GP window its ok to merge them together too. Though
10846 	 * one send in and one send out cannot be merged. We combine
10847 	 * this with never allowing the shrinking of the GP window when
10848 	 * we are in recovery so that we can properly calculate the
10849 	 * sending times.
10850 	 *
10851 	 * This all of course seems complicated, because it is.. :)
10852 	 *
10853 	 * The cum-ack is being advanced upon the sendmap.
10854 	 * If we are not doing a GP estimate don't
10855 	 * proceed.
10856 	 */
10857 	uint64_t ts;
10858 
10859 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
10860 		return;
10861 	/*
10862 	 * If this sendmap entry is going
10863 	 * beyond the measurement window we had picked,
10864 	 * expand the measurement window by that much.
10865 	 */
10866 	if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10867 		tp->gput_ack = rsm->r_end;
10868 	}
10869 	/*
10870 	 * If we have not setup a ack, then we
10871 	 * have no idea if the newly acked pieces
10872 	 * will be "in our seq measurement range". If
10873 	 * it is when we clear the app_limited_needs_set
10874 	 * flag the timestamp will be updated.
10875 	 */
10876 	if (rack->app_limited_needs_set)
10877 		return;
10878 	/*
10879 	 * Finally, we grab out the latest timestamp
10880 	 * that this packet was sent and then see
10881 	 * if:
10882 	 *  a) The packet touches are newly defined GP range.
10883 	 *  b) The time is greater than (newer) than the
10884 	 *     one we currently have. If so we update
10885 	 *     our sending end time window.
10886 	 *
10887 	 * Note we *do not* do this at send time. The reason
10888 	 * is that if you do you *may* pick up a newer timestamp
10889 	 * for a range you are not going to measure. We project
10890 	 * out how far and then sometimes modify that to be
10891 	 * smaller. If that occurs then you will have a send
10892 	 * that does not belong to the range included.
10893 	 */
10894 	if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10895 	    rack->r_ctl.rc_gp_cumack_ts)
10896 		return;
10897 	if (rack_in_gp_window(tp, rsm)) {
10898 		rack->r_ctl.rc_gp_cumack_ts = ts;
10899 		rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10900 			       __LINE__, from, rsm);
10901 	}
10902 }
10903 
10904 static void
10905 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to, uint64_t acktime)
10906 {
10907 	struct rack_sendmap *rsm;
10908 	/*
10909 	 * The ACK point is advancing to th_ack, we must drop off
10910 	 * the packets in the rack log and calculate any eligble
10911 	 * RTT's.
10912 	 */
10913 
10914 	if (sack_filter_blks_used(&rack->r_ctl.rack_sf)) {
10915 		/*
10916 		 * If we have some sack blocks in the filter
10917 		 * lets prune them out by calling sfb with no blocks.
10918 		 */
10919 		sack_filter_blks(tp, &rack->r_ctl.rack_sf, NULL, 0, th_ack);
10920 	}
10921 	if (SEQ_GT(th_ack, tp->snd_una)) {
10922 		/* Clear any app ack remembered settings */
10923 		rack->r_ctl.cleared_app_ack = 0;
10924 	}
10925 	rack->r_wanted_output = 1;
10926 	if (SEQ_GT(th_ack, tp->snd_una))
10927 		rack->r_ctl.last_cumack_advance = acktime;
10928 
10929 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
10930 	if ((rack->rc_last_tlp_acked_set == 1)&&
10931 	    (rack->rc_last_tlp_past_cumack == 1) &&
10932 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10933 		/*
10934 		 * We have reached the point where our last rack
10935 		 * tlp retransmit sequence is ahead of the cum-ack.
10936 		 * This can only happen when the cum-ack moves all
10937 		 * the way around (its been a full 2^^31+1 bytes
10938 		 * or more since we sent a retransmitted TLP). Lets
10939 		 * turn off the valid flag since its not really valid.
10940 		 *
10941 		 * Note since sack's also turn on this event we have
10942 		 * a complication, we have to wait to age it out until
10943 		 * the cum-ack is by the TLP before checking which is
10944 		 * what the next else clause does.
10945 		 */
10946 		rack_log_dsack_event(rack, 9, __LINE__,
10947 				     rack->r_ctl.last_tlp_acked_start,
10948 				     rack->r_ctl.last_tlp_acked_end);
10949 		rack->rc_last_tlp_acked_set = 0;
10950 		rack->rc_last_tlp_past_cumack = 0;
10951 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
10952 		   (rack->rc_last_tlp_past_cumack == 0) &&
10953 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10954 		/*
10955 		 * It is safe to start aging TLP's out.
10956 		 */
10957 		rack->rc_last_tlp_past_cumack = 1;
10958 	}
10959 	/* We do the same for the tlp send seq as well */
10960 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10961 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
10962 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
10963 		rack_log_dsack_event(rack, 9, __LINE__,
10964 				     rack->r_ctl.last_sent_tlp_seq,
10965 				     (rack->r_ctl.last_sent_tlp_seq +
10966 				      rack->r_ctl.last_sent_tlp_len));
10967 		rack->rc_last_sent_tlp_seq_valid = 0;
10968 		rack->rc_last_sent_tlp_past_cumack = 0;
10969 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10970 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
10971 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10972 		/*
10973 		 * It is safe to start aging TLP's send.
10974 		 */
10975 		rack->rc_last_sent_tlp_past_cumack = 1;
10976 	}
10977 more:
10978 	rsm = tqhash_min(rack->r_ctl.tqh);
10979 	if (rsm == NULL) {
10980 		if ((th_ack - 1) == tp->iss) {
10981 			/*
10982 			 * For the SYN incoming case we will not
10983 			 * have called tcp_output for the sending of
10984 			 * the SYN, so there will be no map. All
10985 			 * other cases should probably be a panic.
10986 			 */
10987 			return;
10988 		}
10989 		if (tp->t_flags & TF_SENTFIN) {
10990 			/* if we sent a FIN we often will not have map */
10991 			return;
10992 		}
10993 #ifdef INVARIANTS
10994 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u\n",
10995 		      tp,
10996 		      tp->t_state, th_ack, rack,
10997 		      tp->snd_una, tp->snd_max);
10998 #endif
10999 		return;
11000 	}
11001 	if (SEQ_LT(th_ack, rsm->r_start)) {
11002 		/* Huh map is missing this */
11003 #ifdef INVARIANTS
11004 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
11005 		       rsm->r_start,
11006 		       th_ack, tp->t_state, rack->r_state);
11007 #endif
11008 		return;
11009 	}
11010 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
11011 
11012 	/* Now was it a retransmitted TLP? */
11013 	if ((rsm->r_flags & RACK_TLP) &&
11014 	    (rsm->r_rtr_cnt > 1)) {
11015 		/*
11016 		 * Yes, this rsm was a TLP and retransmitted, remember that
11017 		 * since if a DSACK comes back on this we don't want
11018 		 * to think of it as a reordered segment. This may
11019 		 * get updated again with possibly even other TLPs
11020 		 * in flight, but thats ok. Only when we don't send
11021 		 * a retransmitted TLP for 1/2 the sequences space
11022 		 * will it get turned off (above).
11023 		 */
11024 		if (rack->rc_last_tlp_acked_set &&
11025 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
11026 			/*
11027 			 * We already turned this on since the end matches,
11028 			 * the previous one was a partially ack now we
11029 			 * are getting another one (maybe all of it).
11030 			 */
11031 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
11032 			/*
11033 			 * Lets make sure we have all of it though.
11034 			 */
11035 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
11036 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
11037 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
11038 						     rack->r_ctl.last_tlp_acked_end);
11039 			}
11040 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
11041 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
11042 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
11043 						     rack->r_ctl.last_tlp_acked_end);
11044 			}
11045 		} else {
11046 			rack->rc_last_tlp_past_cumack = 1;
11047 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
11048 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
11049 			rack->rc_last_tlp_acked_set = 1;
11050 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
11051 		}
11052 	}
11053 	/* Now do we consume the whole thing? */
11054 	rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
11055 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
11056 		/* Its all consumed. */
11057 		uint32_t left;
11058 		uint8_t newly_acked;
11059 
11060 		if (rsm->r_flags & RACK_WAS_LOST) {
11061 			/*
11062 			 * This can happen when we marked it as lost
11063 			 * and yet before retransmitting we get an ack
11064 			 * which can happen due to reordering.
11065 			 */
11066 			rsm->r_flags  &= ~RACK_WAS_LOST;
11067 			KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
11068 				("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
11069 			if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
11070 				rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
11071 			else
11072 				rack->r_ctl.rc_considered_lost = 0;
11073 		}
11074 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
11075 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
11076 		rsm->r_rtr_bytes = 0;
11077 		/*
11078 		 * Record the time of highest cumack sent if its in our measurement
11079 		 * window and possibly bump out the end.
11080 		 */
11081 		rack_rsm_sender_update(rack, tp, rsm, 4);
11082 		tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
11083 		if (rsm->r_in_tmap) {
11084 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
11085 			rsm->r_in_tmap = 0;
11086 		}
11087 		newly_acked = 1;
11088 		if (((rsm->r_flags & RACK_ACKED) == 0) &&
11089 		    (IN_RECOVERY(tp->t_flags))) {
11090 			rack->r_ctl.bytes_acked_in_recovery += (rsm->r_end - rsm->r_start);
11091 		}
11092 		if (rsm->r_flags & RACK_ACKED) {
11093 			/*
11094 			 * It was acked on the scoreboard -- remove
11095 			 * it from total
11096 			 */
11097 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
11098 			newly_acked = 0;
11099 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
11100 			/*
11101 			 * There are segments ACKED on the
11102 			 * scoreboard further up. We are seeing
11103 			 * reordering.
11104 			 */
11105 			rsm->r_flags &= ~RACK_SACK_PASSED;
11106 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
11107 			rsm->r_flags |= RACK_ACKED;
11108 			rack->r_ctl.rc_reorder_ts = cts;
11109 			if (rack->r_ctl.rc_reorder_ts == 0)
11110 				rack->r_ctl.rc_reorder_ts = 1;
11111 			if (rack->r_ent_rec_ns) {
11112 				/*
11113 				 * We have sent no more, and we saw an sack
11114 				 * then ack arrive.
11115 				 */
11116 				rack->r_might_revert = 1;
11117 			}
11118 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
11119 		} else {
11120 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
11121 		}
11122 		if ((rsm->r_flags & RACK_TO_REXT) &&
11123 		    (tp->t_flags & TF_RCVD_TSTMP) &&
11124 		    (to->to_flags & TOF_TS) &&
11125 		    (to->to_tsecr != 0) &&
11126 		    (tp->t_flags & TF_PREVVALID)) {
11127 			/*
11128 			 * We can use the timestamp to see
11129 			 * if this retransmission was from the
11130 			 * first transmit. If so we made a mistake.
11131 			 */
11132 			tp->t_flags &= ~TF_PREVVALID;
11133 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
11134 				/* The first transmit is what this ack is for */
11135 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
11136 			}
11137 		}
11138 		left = th_ack - rsm->r_end;
11139 		if (rack->app_limited_needs_set && newly_acked)
11140 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
11141 		/* Free back to zone */
11142 		rack_free(rack, rsm);
11143 		if (left) {
11144 			goto more;
11145 		}
11146 		/* Check for reneging */
11147 		rsm = tqhash_min(rack->r_ctl.tqh);
11148 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
11149 			/*
11150 			 * The peer has moved snd_una up to
11151 			 * the edge of this send, i.e. one
11152 			 * that it had previously acked. The only
11153 			 * way that can be true if the peer threw
11154 			 * away data (space issues) that it had
11155 			 * previously sacked (else it would have
11156 			 * given us snd_una up to (rsm->r_end).
11157 			 * We need to undo the acked markings here.
11158 			 *
11159 			 * Note we have to look to make sure th_ack is
11160 			 * our rsm->r_start in case we get an old ack
11161 			 * where th_ack is behind snd_una.
11162 			 */
11163 			rack_peer_reneges(rack, rsm, th_ack);
11164 		}
11165 		return;
11166 	}
11167 	if (rsm->r_flags & RACK_ACKED) {
11168 		/*
11169 		 * It was acked on the scoreboard -- remove it from
11170 		 * total for the part being cum-acked.
11171 		 */
11172 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
11173 	} else {
11174 		if (((rsm->r_flags & RACK_ACKED) == 0) &&
11175 		    (IN_RECOVERY(tp->t_flags))) {
11176 			rack->r_ctl.bytes_acked_in_recovery += (th_ack - rsm->r_start);
11177 		}
11178 		rack_update_pcm_ack(rack, 1, rsm->r_start, th_ack);
11179 	}
11180 	/* And what about the lost flag? */
11181 	if (rsm->r_flags & RACK_WAS_LOST) {
11182 		/*
11183 		 * This can happen when we marked it as lost
11184 		 * and yet before retransmitting we get an ack
11185 		 * which can happen due to reordering. In this
11186 		 * case its only a partial ack of the send.
11187 		 */
11188 		KASSERT((rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start)),
11189 			("rsm:%p rack:%p rc_considered_lost goes negative th_ack:%u", rsm,  rack, th_ack));
11190 		if (rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start))
11191 			rack->r_ctl.rc_considered_lost -= th_ack - rsm->r_start;
11192 		else
11193 			rack->r_ctl.rc_considered_lost = 0;
11194 	}
11195 	/*
11196 	 * Clear the dup ack count for
11197 	 * the piece that remains.
11198 	 */
11199 	rsm->r_dupack = 0;
11200 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
11201 	if (rsm->r_rtr_bytes) {
11202 		/*
11203 		 * It was retransmitted adjust the
11204 		 * sack holes for what was acked.
11205 		 */
11206 		int ack_am;
11207 
11208 		ack_am = (th_ack - rsm->r_start);
11209 		if (ack_am >= rsm->r_rtr_bytes) {
11210 			rack->r_ctl.rc_holes_rxt -= ack_am;
11211 			rsm->r_rtr_bytes -= ack_am;
11212 		}
11213 	}
11214 	/*
11215 	 * Update where the piece starts and record
11216 	 * the time of send of highest cumack sent if
11217 	 * its in our GP range.
11218 	 */
11219 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
11220 	/* Now we need to move our offset forward too */
11221 	if (rsm->m &&
11222 	    ((rsm->orig_m_len != rsm->m->m_len) ||
11223 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
11224 		/* Fix up the orig_m_len and possibly the mbuf offset */
11225 		rack_adjust_orig_mlen(rsm);
11226 	}
11227 	rsm->soff += (th_ack - rsm->r_start);
11228 	rack_rsm_sender_update(rack, tp, rsm, 5);
11229 	/* The trim will move th_ack into r_start for us */
11230 	tqhash_trim(rack->r_ctl.tqh, th_ack);
11231 	/* Now do we need to move the mbuf fwd too? */
11232 	{
11233 		struct mbuf *m;
11234 		uint32_t soff;
11235 
11236 		m = rsm->m;
11237 		soff = rsm->soff;
11238 		if (m) {
11239 			while (soff >= m->m_len) {
11240 				soff -= m->m_len;
11241 				KASSERT((m->m_next != NULL),
11242 					(" rsm:%p  off:%u soff:%u m:%p",
11243 					 rsm, rsm->soff, soff, m));
11244 				m = m->m_next;
11245 				if (m == NULL) {
11246 					/*
11247 					 * This is a fall-back that prevents a panic. In reality
11248 					 * we should be able to walk the mbuf's and find our place.
11249 					 * At this point snd_una has not been updated with the sbcut() yet
11250 					 * but tqhash_trim did update rsm->r_start so the offset calcuation
11251 					 * should work fine. This is undesirable since we will take cache
11252 					 * hits to access the socket buffer. And even more puzzling is that
11253 					 * it happens occasionally. It should not :(
11254 					 */
11255 					m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
11256 						      (rsm->r_start - tp->snd_una),
11257 						      &soff);
11258 					break;
11259 				}
11260 			}
11261 			/*
11262 			 * Now save in our updated values.
11263 			 */
11264 			rsm->m = m;
11265 			rsm->soff = soff;
11266 			rsm->orig_m_len = rsm->m->m_len;
11267 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11268 		}
11269 	}
11270 	if (rack->app_limited_needs_set &&
11271 	    SEQ_GEQ(th_ack, tp->gput_seq))
11272 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
11273 }
11274 
11275 static void
11276 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
11277 {
11278 	struct rack_sendmap *rsm;
11279 	int sack_pass_fnd = 0;
11280 
11281 	if (rack->r_might_revert) {
11282 		/*
11283 		 * Ok we have reordering, have not sent anything, we
11284 		 * might want to revert the congestion state if nothing
11285 		 * further has SACK_PASSED on it. Lets check.
11286 		 *
11287 		 * We also get here when we have DSACKs come in for
11288 		 * all the data that we FR'd. Note that a rxt or tlp
11289 		 * timer clears this from happening.
11290 		 */
11291 
11292 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
11293 			if (rsm->r_flags & RACK_SACK_PASSED) {
11294 				sack_pass_fnd = 1;
11295 				break;
11296 			}
11297 		}
11298 		if (sack_pass_fnd == 0) {
11299 			/*
11300 			 * We went into recovery
11301 			 * incorrectly due to reordering!
11302 			 */
11303 			int orig_cwnd;
11304 
11305 			rack->r_ent_rec_ns = 0;
11306 			orig_cwnd = tp->snd_cwnd;
11307 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
11308 			tp->snd_recover = tp->snd_una;
11309 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
11310 			if (IN_RECOVERY(tp->t_flags)) {
11311 				rack_exit_recovery(tp, rack, 3);
11312 				if ((rack->rto_from_rec == 1) && (rack_ssthresh_rest_rto_rec != 0) ){
11313 					/*
11314 					 * We were in recovery, had an RTO
11315 					 * and then re-entered recovery (more sack's arrived)
11316 					 * and we have properly recorded the old ssthresh from
11317 					 * the first recovery. We want to be able to slow-start
11318 					 * back to this level. The ssthresh from the timeout
11319 					 * and then back into recovery will end up most likely
11320 					 * to be min(cwnd=1mss, 2mss). Which makes it basically
11321 					 * so we get no slow-start after our RTO.
11322 					 */
11323 					rack->rto_from_rec = 0;
11324 					if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
11325 						tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
11326 				}
11327 			}
11328 			rack->r_ctl.bytes_acked_in_recovery = 0;
11329 			rack->r_ctl.time_entered_recovery = 0;
11330 		}
11331 		rack->r_might_revert = 0;
11332 	}
11333 }
11334 
11335 
11336 static int
11337 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
11338 {
11339 
11340 	uint32_t am, l_end;
11341 	int was_tlp = 0;
11342 
11343 	if (SEQ_GT(end, start))
11344 		am = end - start;
11345 	else
11346 		am = 0;
11347 	if ((rack->rc_last_tlp_acked_set ) &&
11348 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
11349 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
11350 		/*
11351 		 * The DSACK is because of a TLP which we don't
11352 		 * do anything with the reordering window over since
11353 		 * it was not reordering that caused the DSACK but
11354 		 * our previous retransmit TLP.
11355 		 */
11356 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
11357 		was_tlp = 1;
11358 		goto skip_dsack_round;
11359 	}
11360 	if (rack->rc_last_sent_tlp_seq_valid) {
11361 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
11362 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
11363 		    (SEQ_LEQ(end, l_end))) {
11364 			/*
11365 			 * This dsack is from the last sent TLP, ignore it
11366 			 * for reordering purposes.
11367 			 */
11368 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
11369 			was_tlp = 1;
11370 			goto skip_dsack_round;
11371 		}
11372 	}
11373 	if (rack->rc_dsack_round_seen == 0) {
11374 		rack->rc_dsack_round_seen = 1;
11375 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
11376 		rack->r_ctl.num_dsack++;
11377 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
11378 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
11379 	}
11380 skip_dsack_round:
11381 	/*
11382 	 * We keep track of how many DSACK blocks we get
11383 	 * after a recovery incident.
11384 	 */
11385 	rack->r_ctl.dsack_byte_cnt += am;
11386 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
11387 	    rack->r_ctl.retran_during_recovery &&
11388 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
11389 		/*
11390 		 * False recovery most likely culprit is reordering. If
11391 		 * nothing else is missing we need to revert.
11392 		 */
11393 		rack->r_might_revert = 1;
11394 		rack_handle_might_revert(rack->rc_tp, rack);
11395 		rack->r_might_revert = 0;
11396 		rack->r_ctl.retran_during_recovery = 0;
11397 		rack->r_ctl.dsack_byte_cnt = 0;
11398 	}
11399 	return (was_tlp);
11400 }
11401 
11402 static uint32_t
11403 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
11404 {
11405 	return (((tp->snd_max - snd_una) -
11406 		 (rack->r_ctl.rc_sacked + rack->r_ctl.rc_considered_lost)) + rack->r_ctl.rc_holes_rxt);
11407 }
11408 
11409 static int32_t
11410 rack_compute_pipe(struct tcpcb *tp)
11411 {
11412 	return ((int32_t)do_rack_compute_pipe(tp,
11413 					      (struct tcp_rack *)tp->t_fb_ptr,
11414 					      tp->snd_una));
11415 }
11416 
11417 static void
11418 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
11419 {
11420 	/* Deal with changed and PRR here (in recovery only) */
11421 	uint32_t pipe, snd_una;
11422 
11423 	rack->r_ctl.rc_prr_delivered += changed;
11424 
11425 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
11426 		/*
11427 		 * It is all outstanding, we are application limited
11428 		 * and thus we don't need more room to send anything.
11429 		 * Note we use tp->snd_una here and not th_ack because
11430 		 * the data as yet not been cut from the sb.
11431 		 */
11432 		rack->r_ctl.rc_prr_sndcnt = 0;
11433 		return;
11434 	}
11435 	/* Compute prr_sndcnt */
11436 	if (SEQ_GT(tp->snd_una, th_ack)) {
11437 		snd_una = tp->snd_una;
11438 	} else {
11439 		snd_una = th_ack;
11440 	}
11441 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
11442 	if (pipe > tp->snd_ssthresh) {
11443 		long sndcnt;
11444 
11445 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
11446 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
11447 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
11448 		else {
11449 			rack->r_ctl.rc_prr_sndcnt = 0;
11450 			rack_log_to_prr(rack, 9, 0, __LINE__);
11451 			sndcnt = 0;
11452 		}
11453 		sndcnt++;
11454 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
11455 			sndcnt -= rack->r_ctl.rc_prr_out;
11456 		else
11457 			sndcnt = 0;
11458 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
11459 		rack_log_to_prr(rack, 10, 0, __LINE__);
11460 	} else {
11461 		uint32_t limit;
11462 
11463 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
11464 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
11465 		else
11466 			limit = 0;
11467 		if (changed > limit)
11468 			limit = changed;
11469 		limit += ctf_fixed_maxseg(tp);
11470 		if (tp->snd_ssthresh > pipe) {
11471 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
11472 			rack_log_to_prr(rack, 11, 0, __LINE__);
11473 		} else {
11474 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
11475 			rack_log_to_prr(rack, 12, 0, __LINE__);
11476 		}
11477 	}
11478 }
11479 
11480 static void
11481 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
11482 	     int *dsack_seen, int *sacks_seen)
11483 {
11484 	uint32_t changed;
11485 	struct tcp_rack *rack;
11486 	struct rack_sendmap *rsm;
11487 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
11488 	register uint32_t th_ack;
11489 	int32_t i, j, k, num_sack_blks = 0;
11490 	uint32_t cts, acked, ack_point;
11491 	int loop_start = 0;
11492 	uint32_t tsused;
11493 	uint32_t segsiz;
11494 
11495 
11496 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11497 	if (tcp_get_flags(th) & TH_RST) {
11498 		/* We don't log resets */
11499 		return;
11500 	}
11501 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11502 	cts = tcp_get_usecs(NULL);
11503 	rsm = tqhash_min(rack->r_ctl.tqh);
11504 	changed = 0;
11505 	th_ack = th->th_ack;
11506 	segsiz = ctf_fixed_maxseg(rack->rc_tp);
11507 	if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
11508 		/*
11509 		 * You only get credit for
11510 		 * MSS and greater (and you get extra
11511 		 * credit for larger cum-ack moves).
11512 		 */
11513 		int ac;
11514 
11515 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
11516 		counter_u64_add(rack_ack_total, ac);
11517 	}
11518 	if (SEQ_GT(th_ack, tp->snd_una)) {
11519 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
11520 		tp->t_acktime = ticks;
11521 	}
11522 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
11523 		changed = th_ack - rsm->r_start;
11524 	if (changed) {
11525 		rack_process_to_cumack(tp, rack, th_ack, cts, to,
11526 				       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
11527 	}
11528 	if ((to->to_flags & TOF_SACK) == 0) {
11529 		/* We are done nothing left and no sack. */
11530 		rack_handle_might_revert(tp, rack);
11531 		/*
11532 		 * For cases where we struck a dup-ack
11533 		 * with no SACK, add to the changes so
11534 		 * PRR will work right.
11535 		 */
11536 		if (dup_ack_struck && (changed == 0)) {
11537 			changed += ctf_fixed_maxseg(rack->rc_tp);
11538 		}
11539 		goto out;
11540 	}
11541 	/* Sack block processing */
11542 	if (SEQ_GT(th_ack, tp->snd_una))
11543 		ack_point = th_ack;
11544 	else
11545 		ack_point = tp->snd_una;
11546 	for (i = 0; i < to->to_nsacks; i++) {
11547 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
11548 		      &sack, sizeof(sack));
11549 		sack.start = ntohl(sack.start);
11550 		sack.end = ntohl(sack.end);
11551 		if (SEQ_GT(sack.end, sack.start) &&
11552 		    SEQ_GT(sack.start, ack_point) &&
11553 		    SEQ_LT(sack.start, tp->snd_max) &&
11554 		    SEQ_GT(sack.end, ack_point) &&
11555 		    SEQ_LEQ(sack.end, tp->snd_max)) {
11556 			sack_blocks[num_sack_blks] = sack;
11557 			num_sack_blks++;
11558 		} else if (SEQ_LEQ(sack.start, th_ack) &&
11559 			   SEQ_LEQ(sack.end, th_ack)) {
11560 			int was_tlp;
11561 
11562 			if (dsack_seen != NULL)
11563 				*dsack_seen = 1;
11564 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
11565 			/*
11566 			 * Its a D-SACK block.
11567 			 */
11568 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
11569 		}
11570 	}
11571 	if (rack->rc_dsack_round_seen) {
11572 		/* Is the dsack roound over? */
11573 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
11574 			/* Yes it is */
11575 			rack->rc_dsack_round_seen = 0;
11576 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
11577 		}
11578 	}
11579 	/*
11580 	 * Sort the SACK blocks so we can update the rack scoreboard with
11581 	 * just one pass.
11582 	 */
11583 	num_sack_blks = sack_filter_blks(tp, &rack->r_ctl.rack_sf, sack_blocks,
11584 					 num_sack_blks, th->th_ack);
11585 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
11586 	if (sacks_seen != NULL)
11587 		*sacks_seen = num_sack_blks;
11588 	if (num_sack_blks == 0) {
11589 		/* Nothing to sack, but we need to update counts */
11590 		goto out_with_totals;
11591 	}
11592 	/* Its a sack of some sort */
11593 	if (num_sack_blks < 2) {
11594 		/* Only one, we don't need to sort */
11595 		goto do_sack_work;
11596 	}
11597 	/* Sort the sacks */
11598 	for (i = 0; i < num_sack_blks; i++) {
11599 		for (j = i + 1; j < num_sack_blks; j++) {
11600 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
11601 				sack = sack_blocks[i];
11602 				sack_blocks[i] = sack_blocks[j];
11603 				sack_blocks[j] = sack;
11604 			}
11605 		}
11606 	}
11607 	/*
11608 	 * Now are any of the sack block ends the same (yes some
11609 	 * implementations send these)?
11610 	 */
11611 again:
11612 	if (num_sack_blks == 0)
11613 		goto out_with_totals;
11614 	if (num_sack_blks > 1) {
11615 		for (i = 0; i < num_sack_blks; i++) {
11616 			for (j = i + 1; j < num_sack_blks; j++) {
11617 				if (sack_blocks[i].end == sack_blocks[j].end) {
11618 					/*
11619 					 * Ok these two have the same end we
11620 					 * want the smallest end and then
11621 					 * throw away the larger and start
11622 					 * again.
11623 					 */
11624 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
11625 						/*
11626 						 * The second block covers
11627 						 * more area use that
11628 						 */
11629 						sack_blocks[i].start = sack_blocks[j].start;
11630 					}
11631 					/*
11632 					 * Now collapse out the dup-sack and
11633 					 * lower the count
11634 					 */
11635 					for (k = (j + 1); k < num_sack_blks; k++) {
11636 						sack_blocks[j].start = sack_blocks[k].start;
11637 						sack_blocks[j].end = sack_blocks[k].end;
11638 						j++;
11639 					}
11640 					num_sack_blks--;
11641 					goto again;
11642 				}
11643 			}
11644 		}
11645 	}
11646 do_sack_work:
11647 	/*
11648 	 * First lets look to see if
11649 	 * we have retransmitted and
11650 	 * can use the transmit next?
11651 	 */
11652 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11653 	if (rsm &&
11654 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
11655 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
11656 		/*
11657 		 * We probably did the FR and the next
11658 		 * SACK in continues as we would expect.
11659 		 */
11660 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, segsiz);
11661 		if (acked) {
11662 			rack->r_wanted_output = 1;
11663 			changed += acked;
11664 		}
11665 		if (num_sack_blks == 1) {
11666 			/*
11667 			 * This is what we would expect from
11668 			 * a normal implementation to happen
11669 			 * after we have retransmitted the FR,
11670 			 * i.e the sack-filter pushes down
11671 			 * to 1 block and the next to be retransmitted
11672 			 * is the sequence in the sack block (has more
11673 			 * are acked). Count this as ACK'd data to boost
11674 			 * up the chances of recovering any false positives.
11675 			 */
11676 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
11677 			counter_u64_add(rack_express_sack, 1);
11678 			goto out_with_totals;
11679 		} else {
11680 			/*
11681 			 * Start the loop through the
11682 			 * rest of blocks, past the first block.
11683 			 */
11684 			loop_start = 1;
11685 		}
11686 	}
11687 	counter_u64_add(rack_sack_total, 1);
11688 	rsm = rack->r_ctl.rc_sacklast;
11689 	for (i = loop_start; i < num_sack_blks; i++) {
11690 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts,  segsiz);
11691 		if (acked) {
11692 			rack->r_wanted_output = 1;
11693 			changed += acked;
11694 		}
11695 	}
11696 out_with_totals:
11697 	if (num_sack_blks > 1) {
11698 		/*
11699 		 * You get an extra stroke if
11700 		 * you have more than one sack-blk, this
11701 		 * could be where we are skipping forward
11702 		 * and the sack-filter is still working, or
11703 		 * it could be an attacker constantly
11704 		 * moving us.
11705 		 */
11706 		counter_u64_add(rack_move_some, 1);
11707 	}
11708 out:
11709 	if (changed) {
11710 		/* Something changed cancel the rack timer */
11711 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11712 	}
11713 	tsused = tcp_get_usecs(NULL);
11714 	rsm = tcp_rack_output(tp, rack, tsused);
11715 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11716 	    rsm &&
11717 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11718 		/* Enter recovery */
11719 		entered_recovery = 1;
11720 		rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
11721 		/*
11722 		 * When we enter recovery we need to assure we send
11723 		 * one packet.
11724 		 */
11725 		if (rack->rack_no_prr == 0) {
11726 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11727 			rack_log_to_prr(rack, 8, 0, __LINE__);
11728 		}
11729 		rack->r_timer_override = 1;
11730 		rack->r_early = 0;
11731 		rack->r_ctl.rc_agg_early = 0;
11732 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
11733 		   rsm &&
11734 		   (rack->r_rr_config == 3)) {
11735 		/*
11736 		 * Assure we can output and we get no
11737 		 * remembered pace time except the retransmit.
11738 		 */
11739 		rack->r_timer_override = 1;
11740 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11741 		rack->r_ctl.rc_resend = rsm;
11742 	}
11743 	if (IN_FASTRECOVERY(tp->t_flags) &&
11744 	    (rack->rack_no_prr == 0) &&
11745 	    (entered_recovery == 0)) {
11746 		rack_update_prr(tp, rack, changed, th_ack);
11747 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11748 		     ((tcp_in_hpts(rack->rc_tp) == 0) &&
11749 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11750 			/*
11751 			 * If you are pacing output you don't want
11752 			 * to override.
11753 			 */
11754 			rack->r_early = 0;
11755 			rack->r_ctl.rc_agg_early = 0;
11756 			rack->r_timer_override = 1;
11757 		}
11758 	}
11759 }
11760 
11761 static void
11762 rack_strike_dupack(struct tcp_rack *rack, tcp_seq th_ack)
11763 {
11764 	struct rack_sendmap *rsm;
11765 
11766 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11767 	while (rsm) {
11768 		/*
11769 		 * We need to skip anything already set
11770 		 * to be retransmitted.
11771 		 */
11772 		if ((rsm->r_dupack >= DUP_ACK_THRESHOLD)  ||
11773 		    (rsm->r_flags & RACK_MUST_RXT)) {
11774 			rsm = TAILQ_NEXT(rsm, r_tnext);
11775 			continue;
11776 		}
11777 		break;
11778 	}
11779 	if (rsm && (rsm->r_dupack < 0xff)) {
11780 		rsm->r_dupack++;
11781 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11782 			struct timeval tv;
11783 			uint32_t cts;
11784 			/*
11785 			 * Here we see if we need to retransmit. For
11786 			 * a SACK type connection if enough time has passed
11787 			 * we will get a return of the rsm. For a non-sack
11788 			 * connection we will get the rsm returned if the
11789 			 * dupack value is 3 or more.
11790 			 */
11791 			cts = tcp_get_usecs(&tv);
11792 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11793 			if (rack->r_ctl.rc_resend != NULL) {
11794 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11795 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11796 							 th_ack,  __LINE__);
11797 				}
11798 				rack->r_wanted_output = 1;
11799 				rack->r_timer_override = 1;
11800 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11801 			}
11802 		} else {
11803 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11804 		}
11805 	}
11806 }
11807 
11808 static void
11809 rack_check_bottom_drag(struct tcpcb *tp,
11810 		       struct tcp_rack *rack,
11811 		       struct socket *so)
11812 {
11813 	/*
11814 	 * So what is dragging bottom?
11815 	 *
11816 	 * Dragging bottom means you were under pacing and had a
11817 	 * delay in processing inbound acks waiting on our pacing
11818 	 * timer to expire. While you were waiting all of the acknowledgments
11819 	 * for the packets you sent have arrived. This means we are pacing
11820 	 * way underneath the bottleneck to the point where our Goodput
11821 	 * measurements stop working, since they require more than one
11822 	 * ack (usually at least 8 packets worth with multiple acks so we can
11823 	 * gauge the inter-ack times). If that occurs we have a real problem
11824 	 * since we are stuck in a hole that we can't get out of without
11825 	 * something speeding us up.
11826 	 *
11827 	 * We also check to see if we are widdling down to just one segment
11828 	 * outstanding. If this occurs and we have room to send in our cwnd/rwnd
11829 	 * then we are adding the delayed ack interval into our measurments and
11830 	 * we need to speed up slightly.
11831 	 */
11832 	uint32_t segsiz, minseg;
11833 
11834 	segsiz = ctf_fixed_maxseg(tp);
11835 	minseg = segsiz;
11836 	if (tp->snd_max == tp->snd_una) {
11837 		/*
11838 		 * We are doing dynamic pacing and we are way
11839 		 * under. Basically everything got acked while
11840 		 * we were still waiting on the pacer to expire.
11841 		 *
11842 		 * This means we need to boost the b/w in
11843 		 * addition to any earlier boosting of
11844 		 * the multiplier.
11845 		 */
11846 		uint64_t lt_bw;
11847 
11848 		tcp_trace_point(rack->rc_tp, TCP_TP_PACED_BOTTOM);
11849 		lt_bw = rack_get_lt_bw(rack);
11850 		rack->rc_dragged_bottom = 1;
11851 		rack_validate_multipliers_at_or_above100(rack);
11852 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11853 		    (rack->dis_lt_bw == 0) &&
11854 		    (rack->use_lesser_lt_bw == 0) &&
11855 		    (lt_bw > 0)) {
11856 			/*
11857 			 * Lets use the long-term b/w we have
11858 			 * been getting as a base.
11859 			 */
11860 			if (rack->rc_gp_filled == 0) {
11861 				if (lt_bw > ONE_POINT_TWO_MEG) {
11862 					/*
11863 					 * If we have no measurement
11864 					 * don't let us set in more than
11865 					 * 1.2Mbps. If we are still too
11866 					 * low after pacing with this we
11867 					 * will hopefully have a max b/w
11868 					 * available to sanity check things.
11869 					 */
11870 					lt_bw = ONE_POINT_TWO_MEG;
11871 				}
11872 				rack->r_ctl.rc_rtt_diff = 0;
11873 				rack->r_ctl.gp_bw = lt_bw;
11874 				rack->rc_gp_filled = 1;
11875 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11876 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11877 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11878 			} else if (lt_bw > rack->r_ctl.gp_bw) {
11879 				rack->r_ctl.rc_rtt_diff = 0;
11880 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11881 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11882 				rack->r_ctl.gp_bw = lt_bw;
11883 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11884 			} else
11885 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
11886 			if ((rack->gp_ready == 0) &&
11887 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11888 				/* We have enough measurements now */
11889 				rack->gp_ready = 1;
11890 				if (rack->dgp_on ||
11891 				    rack->rack_hibeta)
11892 					rack_set_cc_pacing(rack);
11893 				if (rack->defer_options)
11894 					rack_apply_deferred_options(rack);
11895 			}
11896 		} else {
11897 			/*
11898 			 * zero rtt possibly?, settle for just an old increase.
11899 			 */
11900 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
11901 		}
11902 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11903 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11904 					       minseg)) &&
11905 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11906 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11907 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11908 		    (segsiz * rack_req_segs))) {
11909 		/*
11910 		 * We are doing dynamic GP pacing and
11911 		 * we have everything except 1MSS or less
11912 		 * bytes left out. We are still pacing away.
11913 		 * And there is data that could be sent, This
11914 		 * means we are inserting delayed ack time in
11915 		 * our measurements because we are pacing too slow.
11916 		 */
11917 		rack_validate_multipliers_at_or_above100(rack);
11918 		rack->rc_dragged_bottom = 1;
11919 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
11920 	}
11921 }
11922 
11923 #ifdef TCP_REQUEST_TRK
11924 static void
11925 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11926 		struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11927 {
11928 	int do_log;
11929 
11930 	do_log = tcp_bblogging_on(rack->rc_tp);
11931 	if (do_log == 0) {
11932 		if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11933 			return;
11934 		/* We only allow the three below with point logging on */
11935 		if ((mod != HYBRID_LOG_RULES_APP) &&
11936 		    (mod != HYBRID_LOG_RULES_SET) &&
11937 		    (mod != HYBRID_LOG_REQ_COMP))
11938 			return;
11939 
11940 	}
11941 	if (do_log) {
11942 		union tcp_log_stackspecific log;
11943 		struct timeval tv;
11944 
11945 		/* Convert our ms to a microsecond */
11946 		memset(&log, 0, sizeof(log));
11947 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11948 		log.u_bbr.flex1 = seq;
11949 		log.u_bbr.cwnd_gain = line;
11950 		if (cur != NULL) {
11951 			uint64_t off;
11952 
11953 			log.u_bbr.flex2 = cur->start_seq;
11954 			log.u_bbr.flex3 = cur->end_seq;
11955 			log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11956 			log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11957 			log.u_bbr.flex6 = cur->flags;
11958 			log.u_bbr.pkts_out = cur->hybrid_flags;
11959 			log.u_bbr.rttProp = cur->timestamp;
11960 			log.u_bbr.cur_del_rate = cur->cspr;
11961 			log.u_bbr.bw_inuse = cur->start;
11962 			log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11963 			log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11964 			log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11965 			log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11966 			log.u_bbr.inhpts = 1;
11967 #ifdef TCP_REQUEST_TRK
11968 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11969 			log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11970 #endif
11971 		} else {
11972 			log.u_bbr.flex2 = err;
11973 		}
11974 		/*
11975 		 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11976 		 */
11977 		log.u_bbr.flex7 = rack->rc_catch_up;
11978 		log.u_bbr.flex7 <<= 1;
11979 		log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11980 		log.u_bbr.flex7 <<= 1;
11981 		log.u_bbr.flex7 |= rack->dgp_on;
11982 		/*
11983 		 * Compose bbr_state to be a bit wise 0000ADHF
11984 		 * where A is the always_pace flag
11985 		 * where D is the dgp_on flag
11986 		 * where H is the hybrid_mode on flag
11987 		 * where F is the use_fixed_rate flag.
11988 		 */
11989 		log.u_bbr.bbr_state = rack->rc_always_pace;
11990 		log.u_bbr.bbr_state <<= 1;
11991 		log.u_bbr.bbr_state |= rack->dgp_on;
11992 		log.u_bbr.bbr_state <<= 1;
11993 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
11994 		log.u_bbr.bbr_state <<= 1;
11995 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
11996 		log.u_bbr.flex8 = mod;
11997 		log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11998 		log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11999 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12000 		log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
12001 		log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
12002 		log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
12003 		tcp_log_event(rack->rc_tp, NULL,
12004 		    &rack->rc_inp->inp_socket->so_rcv,
12005 		    &rack->rc_inp->inp_socket->so_snd,
12006 		    TCP_HYBRID_PACING_LOG, 0,
12007 	            0, &log, false, NULL, __func__, __LINE__, &tv);
12008 	}
12009 }
12010 #endif
12011 
12012 #ifdef TCP_REQUEST_TRK
12013 static void
12014 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
12015 {
12016 	struct tcp_sendfile_track *rc_cur, *orig_ent;
12017 	struct tcpcb *tp;
12018 	int err = 0;
12019 
12020 	orig_ent = rack->r_ctl.rc_last_sft;
12021 	rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
12022 	if (rc_cur == NULL) {
12023 		/* If not in the beginning what about the end piece */
12024 		if (rack->rc_hybrid_mode)
12025 			rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
12026 		rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
12027 	} else {
12028 		err = 12345;
12029 	}
12030 	/* If we find no parameters we are in straight DGP mode */
12031 	if(rc_cur == NULL) {
12032 		/* None found for this seq, just DGP for now */
12033 		if (rack->rc_hybrid_mode) {
12034 			rack->r_ctl.client_suggested_maxseg = 0;
12035 			rack->rc_catch_up = 0;
12036 			if (rack->cspr_is_fcc == 0)
12037 				rack->r_ctl.bw_rate_cap = 0;
12038 			else
12039 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
12040 		}
12041 		if (rack->rc_hybrid_mode) {
12042 			rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
12043 		}
12044 		if (rack->r_ctl.rc_last_sft) {
12045 			rack->r_ctl.rc_last_sft = NULL;
12046 		}
12047 		return;
12048 	}
12049 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
12050 		/* This entry was never setup for hybrid pacing on/off etc */
12051 		if (rack->rc_hybrid_mode) {
12052 			rack->r_ctl.client_suggested_maxseg = 0;
12053 			rack->rc_catch_up = 0;
12054 			rack->r_ctl.bw_rate_cap = 0;
12055 		}
12056 		if (rack->r_ctl.rc_last_sft) {
12057 			rack->r_ctl.rc_last_sft = NULL;
12058 		}
12059 		if ((rc_cur->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
12060 			rc_cur->flags |= TCP_TRK_TRACK_FLG_FSND;
12061 			rc_cur->first_send = cts;
12062 			rc_cur->sent_at_fs = rack->rc_tp->t_sndbytes;
12063 			rc_cur->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
12064 		}
12065 		return;
12066 	}
12067 	/*
12068 	 * Ok if we have a new entry *or* have never
12069 	 * set up an entry we need to proceed. If
12070 	 * we have already set it up this entry we
12071 	 * just continue along with what we already
12072 	 * setup.
12073 	 */
12074 	tp = rack->rc_tp;
12075 	if ((rack->r_ctl.rc_last_sft != NULL) &&
12076 	    (rack->r_ctl.rc_last_sft == rc_cur)) {
12077 		/* Its already in place */
12078 		if (rack->rc_hybrid_mode)
12079 			rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
12080 		return;
12081 	}
12082 	if (rack->rc_hybrid_mode == 0) {
12083 		rack->r_ctl.rc_last_sft = rc_cur;
12084 		if (orig_ent) {
12085 			orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
12086 			orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
12087 			orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
12088 		}
12089 		rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
12090 		return;
12091 	}
12092 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
12093 		/* Compensate for all the header overhead's */
12094 		if (rack->cspr_is_fcc == 0)
12095 			rack->r_ctl.bw_rate_cap	= rack_compensate_for_linerate(rack, rc_cur->cspr);
12096 		else
12097 			rack->r_ctl.fillcw_cap =  rack_compensate_for_linerate(rack, rc_cur->cspr);
12098 	} else {
12099 		if (rack->rc_hybrid_mode) {
12100 			if (rack->cspr_is_fcc == 0)
12101 				rack->r_ctl.bw_rate_cap = 0;
12102 			else
12103 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
12104 		}
12105 	}
12106 	if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
12107 		rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
12108 	else
12109 		rack->r_ctl.client_suggested_maxseg = 0;
12110 	if (rc_cur->timestamp == rack->r_ctl.last_tm_mark) {
12111 		/*
12112 		 * It is the same timestamp as the previous one
12113 		 * add the hybrid flag that will indicate we use
12114 		 * sendtime not arrival time for catch-up mode.
12115 		 */
12116 		rc_cur->hybrid_flags |= TCP_HYBRID_PACING_SENDTIME;
12117 	}
12118 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
12119 	    (rc_cur->cspr > 0)) {
12120 		uint64_t len;
12121 
12122 		rack->rc_catch_up = 1;
12123 		/*
12124 		 * Calculate the deadline time, first set the
12125 		 * time to when the request arrived.
12126 		 */
12127 		if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_SENDTIME) {
12128 			/*
12129 			 * For cases where its a duplicate tm (we received more
12130 			 * than one request for a tm) we want to use now, the point
12131 			 * where we are just sending the first bit of the request.
12132 			 */
12133 			rc_cur->deadline = cts;
12134 		} else {
12135 			/*
12136 			 * Here we have a different tm from the last request
12137 			 * so we want to use arrival time as our base.
12138 			 */
12139 			rc_cur->deadline = rc_cur->localtime;
12140 		}
12141 		/*
12142 		 * Next calculate the length and compensate for
12143 		 * TLS if need be.
12144 		 */
12145 		len = rc_cur->end - rc_cur->start;
12146 		if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
12147 			/*
12148 			 * This session is doing TLS. Take a swag guess
12149 			 * at the overhead.
12150 			 */
12151 			len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
12152 		}
12153 		/*
12154 		 * Now considering the size, and the cspr, what is the time that
12155 		 * would be required at the cspr rate. Here we use the raw
12156 		 * cspr value since the client only looks at the raw data. We
12157 		 * do use len which includes TLS overhead, but not the TCP/IP etc.
12158 		 * That will get made up for in the CU pacing rate set.
12159 		 */
12160 		len *= HPTS_USEC_IN_SEC;
12161 		len /= rc_cur->cspr;
12162 		rc_cur->deadline += len;
12163 	} else {
12164 		rack->rc_catch_up = 0;
12165 		rc_cur->deadline = 0;
12166 	}
12167 	if (rack->r_ctl.client_suggested_maxseg != 0) {
12168 		/*
12169 		 * We need to reset the max pace segs if we have a
12170 		 * client_suggested_maxseg.
12171 		 */
12172 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12173 	}
12174 	if (orig_ent) {
12175 		orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
12176 		orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
12177 		orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
12178 	}
12179 	rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
12180 	/* Remember it for next time and for CU mode */
12181 	rack->r_ctl.rc_last_sft = rc_cur;
12182 	rack->r_ctl.last_tm_mark = rc_cur->timestamp;
12183 }
12184 #endif
12185 
12186 static void
12187 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
12188 {
12189 #ifdef TCP_REQUEST_TRK
12190 	struct tcp_sendfile_track *ent;
12191 
12192 	ent = rack->r_ctl.rc_last_sft;
12193 	if ((ent == NULL) ||
12194 	    (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
12195 	    (SEQ_GEQ(seq, ent->end_seq))) {
12196 		/* Time to update the track. */
12197 		rack_set_dgp_hybrid_mode(rack, seq, len, cts);
12198 		ent = rack->r_ctl.rc_last_sft;
12199 	}
12200 	/* Out of all */
12201 	if (ent == NULL) {
12202 		return;
12203 	}
12204 	if (SEQ_LT(ent->end_seq, (seq + len))) {
12205 		/*
12206 		 * This is the case where our end_seq guess
12207 		 * was wrong. This is usually due to TLS having
12208 		 * more bytes then our guess. It could also be the
12209 		 * case that the client sent in two requests closely
12210 		 * and the SB is full of both so we are sending part
12211 		 * of each (end|beg). In such a case lets move this
12212 		 * guys end to match the end of this send. That
12213 		 * way it will complete when all of it is acked.
12214 		 */
12215 		ent->end_seq = (seq + len);
12216 		if (rack->rc_hybrid_mode)
12217 			rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
12218 	}
12219 	/* Now validate we have set the send time of this one */
12220 	if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
12221 		ent->flags |= TCP_TRK_TRACK_FLG_FSND;
12222 		ent->first_send = cts;
12223 		ent->sent_at_fs = rack->rc_tp->t_sndbytes;
12224 		ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
12225 	}
12226 #endif
12227 }
12228 
12229 static void
12230 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
12231 {
12232 	/*
12233 	 * The fast output path is enabled and we
12234 	 * have moved the cumack forward. Lets see if
12235 	 * we can expand forward the fast path length by
12236 	 * that amount. What we would ideally like to
12237 	 * do is increase the number of bytes in the
12238 	 * fast path block (left_to_send) by the
12239 	 * acked amount. However we have to gate that
12240 	 * by two factors:
12241 	 * 1) The amount outstanding and the rwnd of the peer
12242 	 *    (i.e. we don't want to exceed the rwnd of the peer).
12243 	 *    <and>
12244 	 * 2) The amount of data left in the socket buffer (i.e.
12245 	 *    we can't send beyond what is in the buffer).
12246 	 *
12247 	 * Note that this does not take into account any increase
12248 	 * in the cwnd. We will only extend the fast path by
12249 	 * what was acked.
12250 	 */
12251 	uint32_t new_total, gating_val;
12252 
12253 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
12254 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
12255 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
12256 	if (new_total <= gating_val) {
12257 		/* We can increase left_to_send by the acked amount */
12258 		counter_u64_add(rack_extended_rfo, 1);
12259 		rack->r_ctl.fsb.left_to_send = new_total;
12260 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
12261 			("rack:%p left_to_send:%u sbavail:%u out:%u",
12262 			 rack, rack->r_ctl.fsb.left_to_send,
12263 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
12264 			 (tp->snd_max - tp->snd_una)));
12265 
12266 	}
12267 }
12268 
12269 static void
12270 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
12271 {
12272 	/*
12273 	 * Here any sendmap entry that points to the
12274 	 * beginning mbuf must be adjusted to the correct
12275 	 * offset. This must be called with:
12276 	 * 1) The socket buffer locked
12277 	 * 2) snd_una adjusted to its new position.
12278 	 *
12279 	 * Note that (2) implies rack_ack_received has also
12280 	 * been called and all the sbcut's have been done.
12281 	 *
12282 	 * We grab the first mbuf in the socket buffer and
12283 	 * then go through the front of the sendmap, recalculating
12284 	 * the stored offset for any sendmap entry that has
12285 	 * that mbuf. We must use the sb functions to do this
12286 	 * since its possible an add was done has well as
12287 	 * the subtraction we may have just completed. This should
12288 	 * not be a penalty though, since we just referenced the sb
12289 	 * to go in and trim off the mbufs that we freed (of course
12290 	 * there will be a penalty for the sendmap references though).
12291 	 *
12292 	 * Note also with INVARIANT on, we validate with a KASSERT
12293 	 * that the first sendmap entry has a soff of 0.
12294 	 *
12295 	 */
12296 	struct mbuf *m;
12297 	struct rack_sendmap *rsm;
12298 	tcp_seq snd_una;
12299 #ifdef INVARIANTS
12300 	int first_processed = 0;
12301 #endif
12302 
12303 	snd_una = rack->rc_tp->snd_una;
12304 	SOCKBUF_LOCK_ASSERT(sb);
12305 	m = sb->sb_mb;
12306 	rsm = tqhash_min(rack->r_ctl.tqh);
12307 	if ((rsm == NULL) || (m == NULL)) {
12308 		/* Nothing outstanding */
12309 		return;
12310 	}
12311 	/* The very first RSM's mbuf must point to the head mbuf in the sb */
12312 	KASSERT((rsm->m == m),
12313 		("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
12314 		 rack, sb, rsm));
12315 	while (rsm->m && (rsm->m == m)) {
12316 		/* one to adjust */
12317 #ifdef INVARIANTS
12318 		struct mbuf *tm;
12319 		uint32_t soff;
12320 
12321 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
12322 		if ((rsm->orig_m_len != m->m_len) ||
12323 		    (rsm->orig_t_space != M_TRAILINGROOM(m))){
12324 			rack_adjust_orig_mlen(rsm);
12325 		}
12326 		if (first_processed == 0) {
12327 			KASSERT((rsm->soff == 0),
12328 				("Rack:%p rsm:%p -- rsm at head but soff not zero",
12329 				 rack, rsm));
12330 			first_processed = 1;
12331 		}
12332 		if ((rsm->soff != soff) || (rsm->m != tm)) {
12333 			/*
12334 			 * This is not a fatal error, we anticipate it
12335 			 * might happen (the else code), so we count it here
12336 			 * so that under invariant we can see that it really
12337 			 * does happen.
12338 			 */
12339 			counter_u64_add(rack_adjust_map_bw, 1);
12340 		}
12341 		rsm->m = tm;
12342 		rsm->soff = soff;
12343 		if (tm) {
12344 			rsm->orig_m_len = rsm->m->m_len;
12345 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12346 		} else {
12347 			rsm->orig_m_len = 0;
12348 			rsm->orig_t_space = 0;
12349 		}
12350 #else
12351 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
12352 		if (rsm->m) {
12353 			rsm->orig_m_len = rsm->m->m_len;
12354 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12355 		} else {
12356 			rsm->orig_m_len = 0;
12357 			rsm->orig_t_space = 0;
12358 		}
12359 #endif
12360 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
12361 		if (rsm == NULL)
12362 			break;
12363 	}
12364 }
12365 
12366 #ifdef TCP_REQUEST_TRK
12367 static inline void
12368 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
12369 {
12370 	struct tcp_sendfile_track *ent;
12371 	int i;
12372 
12373 	if ((rack->rc_hybrid_mode == 0) &&
12374 	    (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
12375 		/*
12376 		 * Just do normal completions hybrid pacing is not on
12377 		 * and CLDL is off as well.
12378 		 */
12379 		tcp_req_check_for_comp(rack->rc_tp, th_ack);
12380 		return;
12381 	}
12382 	/*
12383 	 * Originally I was just going to find the th_ack associated
12384 	 * with an entry. But then I realized a large strech ack could
12385 	 * in theory ack two or more requests at once. So instead we
12386 	 * need to find all entries that are completed by th_ack not
12387 	 * just a single entry and do our logging.
12388 	 */
12389 	ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
12390 	while (ent != NULL) {
12391 		/*
12392 		 * We may be doing hybrid pacing or CLDL and need more details possibly
12393 		 * so we do it manually instead of calling
12394 		 * tcp_req_check_for_comp()
12395 		 */
12396 		uint64_t laa, tim, data, cbw, ftim;
12397 
12398 		/* Ok this ack frees it */
12399 		rack_log_hybrid(rack, th_ack,
12400 				ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
12401 		rack_log_hybrid_sends(rack, ent, __LINE__);
12402 		/* calculate the time based on the ack arrival */
12403 		data = ent->end - ent->start;
12404 		laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
12405 		if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
12406 			if (ent->first_send > ent->localtime)
12407 				ftim = ent->first_send;
12408 			else
12409 				ftim = ent->localtime;
12410 		} else {
12411 			/* TSNH */
12412 			ftim = ent->localtime;
12413 		}
12414 		if (laa > ent->localtime)
12415 			tim = laa - ftim;
12416 		else
12417 			tim = 0;
12418 		cbw = data * HPTS_USEC_IN_SEC;
12419 		if (tim > 0)
12420 			cbw /= tim;
12421 		else
12422 			cbw = 0;
12423 		rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
12424 		/*
12425 		 * Check to see if we are freeing what we are pointing to send wise
12426 		 * if so be sure to NULL the pointer so we know we are no longer
12427 		 * set to anything.
12428 		 */
12429 		if (ent == rack->r_ctl.rc_last_sft) {
12430 			rack->r_ctl.rc_last_sft = NULL;
12431 			if (rack->rc_hybrid_mode) {
12432 				rack->rc_catch_up = 0;
12433 				if (rack->cspr_is_fcc == 0)
12434 					rack->r_ctl.bw_rate_cap = 0;
12435 				else
12436 					rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
12437 				rack->r_ctl.client_suggested_maxseg = 0;
12438 			}
12439 		}
12440 		/* Generate the log that the tcp_netflix call would have */
12441 		tcp_req_log_req_info(rack->rc_tp, ent,
12442 				      i, TCP_TRK_REQ_LOG_FREED, 0, 0);
12443 		/* Free it and see if there is another one */
12444 		tcp_req_free_a_slot(rack->rc_tp, ent);
12445 		ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
12446 	}
12447 }
12448 #endif
12449 
12450 
12451 /*
12452  * Return value of 1, we do not need to call rack_process_data().
12453  * return value of 0, rack_process_data can be called.
12454  * For ret_val if its 0 the TCP is locked, if its non-zero
12455  * its unlocked and probably unsafe to touch the TCB.
12456  */
12457 static int
12458 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12459     struct tcpcb *tp, struct tcpopt *to,
12460     uint32_t tiwin, int32_t tlen,
12461     int32_t * ofia, int32_t thflags, int32_t *ret_val, int32_t orig_tlen)
12462 {
12463 	int32_t ourfinisacked = 0;
12464 	int32_t nsegs, acked_amount;
12465 	int32_t acked;
12466 	struct mbuf *mfree;
12467 	struct tcp_rack *rack;
12468 	int32_t under_pacing = 0;
12469 	int32_t post_recovery = 0;
12470 	uint32_t p_cwnd;
12471 
12472 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12473 
12474 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12475 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
12476 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
12477 				      &rack->r_ctl.challenge_ack_ts,
12478 				      &rack->r_ctl.challenge_ack_cnt);
12479 		rack->r_wanted_output = 1;
12480 		return (1);
12481 	}
12482 	if (rack->gp_ready &&
12483 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12484 		under_pacing = 1;
12485 	}
12486 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
12487 		int in_rec, dup_ack_struck = 0;
12488 		int dsack_seen = 0, sacks_seen = 0;
12489 
12490 		in_rec = IN_FASTRECOVERY(tp->t_flags);
12491 		if (rack->rc_in_persist) {
12492 			tp->t_rxtshift = 0;
12493 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12494 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12495 		}
12496 
12497 		if ((th->th_ack == tp->snd_una) &&
12498 		    (tiwin == tp->snd_wnd) &&
12499 		    (orig_tlen == 0) &&
12500 		    ((to->to_flags & TOF_SACK) == 0)) {
12501 			rack_strike_dupack(rack, th->th_ack);
12502 			dup_ack_struck = 1;
12503 		}
12504 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
12505 			     dup_ack_struck, &dsack_seen, &sacks_seen);
12506 
12507 	}
12508 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12509 		/*
12510 		 * Old ack, behind (or duplicate to) the last one rcv'd
12511 		 * Note: We mark reordering is occuring if its
12512 		 * less than and we have not closed our window.
12513 		 */
12514 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
12515 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12516 			if (rack->r_ctl.rc_reorder_ts == 0)
12517 				rack->r_ctl.rc_reorder_ts = 1;
12518 		}
12519 		return (0);
12520 	}
12521 	/*
12522 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
12523 	 * something we sent.
12524 	 */
12525 	if (tp->t_flags & TF_NEEDSYN) {
12526 		/*
12527 		 * T/TCP: Connection was half-synchronized, and our SYN has
12528 		 * been ACK'd (so connection is now fully synchronized).  Go
12529 		 * to non-starred state, increment snd_una for ACK of SYN,
12530 		 * and check if we can do window scaling.
12531 		 */
12532 		tp->t_flags &= ~TF_NEEDSYN;
12533 		tp->snd_una++;
12534 		/* Do window scaling? */
12535 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12536 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12537 			tp->rcv_scale = tp->request_r_scale;
12538 			/* Send window already scaled. */
12539 		}
12540 	}
12541 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12542 
12543 	acked = BYTES_THIS_ACK(tp, th);
12544 	if (acked) {
12545 		/*
12546 		 * Any time we move the cum-ack forward clear
12547 		 * keep-alive tied probe-not-answered. The
12548 		 * persists clears its own on entry.
12549 		 */
12550 		rack->probe_not_answered = 0;
12551 	}
12552 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12553 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12554 	/*
12555 	 * If we just performed our first retransmit, and the ACK arrives
12556 	 * within our recovery window, then it was a mistake to do the
12557 	 * retransmit in the first place.  Recover our original cwnd and
12558 	 * ssthresh, and proceed to transmit where we left off.
12559 	 */
12560 	if ((tp->t_flags & TF_PREVVALID) &&
12561 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12562 		tp->t_flags &= ~TF_PREVVALID;
12563 		if (tp->t_rxtshift == 1 &&
12564 		    (int)(ticks - tp->t_badrxtwin) < 0)
12565 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12566 	}
12567 	if (acked) {
12568 		/* assure we are not backed off */
12569 		tp->t_rxtshift = 0;
12570 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12571 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12572 		rack->rc_tlp_in_progress = 0;
12573 		rack->r_ctl.rc_tlp_cnt_out = 0;
12574 		/*
12575 		 * If it is the RXT timer we want to
12576 		 * stop it, so we can restart a TLP.
12577 		 */
12578 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12579 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12580 #ifdef TCP_REQUEST_TRK
12581 		rack_req_check_for_comp(rack, th->th_ack);
12582 #endif
12583 	}
12584 	/*
12585 	 * If we have a timestamp reply, update smoothed round trip time. If
12586 	 * no timestamp is present but transmit timer is running and timed
12587 	 * sequence number was acked, update smoothed round trip time. Since
12588 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
12589 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
12590 	 * timer.
12591 	 *
12592 	 * Some boxes send broken timestamp replies during the SYN+ACK
12593 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
12594 	 * and blow up the retransmit timer.
12595 	 */
12596 	/*
12597 	 * If all outstanding data is acked, stop retransmit timer and
12598 	 * remember to restart (more output or persist). If there is more
12599 	 * data to be acked, restart retransmit timer, using current
12600 	 * (possibly backed-off) value.
12601 	 */
12602 	if (acked == 0) {
12603 		if (ofia)
12604 			*ofia = ourfinisacked;
12605 		return (0);
12606 	}
12607 	if (IN_RECOVERY(tp->t_flags)) {
12608 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
12609 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
12610 			tcp_rack_partialack(tp);
12611 		} else {
12612 			rack_post_recovery(tp, th->th_ack);
12613 			post_recovery = 1;
12614 			/*
12615 			 * Grab the segsiz, multiply by 2 and add the snd_cwnd
12616 			 * that is the max the CC should add if we are exiting
12617 			 * recovery and doing a late add.
12618 			 */
12619 			p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
12620 			p_cwnd <<= 1;
12621 			p_cwnd += tp->snd_cwnd;
12622 		}
12623 	} else if ((rack->rto_from_rec == 1) &&
12624 		   SEQ_GEQ(th->th_ack, tp->snd_recover)) {
12625 		/*
12626 		 * We were in recovery, hit a rxt timeout
12627 		 * and never re-entered recovery. The timeout(s)
12628 		 * made up all the lost data. In such a case
12629 		 * we need to clear the rto_from_rec flag.
12630 		 */
12631 		rack->rto_from_rec = 0;
12632 	}
12633 	/*
12634 	 * Let the congestion control algorithm update congestion control
12635 	 * related information. This typically means increasing the
12636 	 * congestion window.
12637 	 */
12638 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, post_recovery);
12639 	if (post_recovery &&
12640 	    (tp->snd_cwnd > p_cwnd)) {
12641 		/* Must be non-newreno (cubic) getting too ahead of itself */
12642 		tp->snd_cwnd = p_cwnd;
12643 	}
12644 	SOCKBUF_LOCK(&so->so_snd);
12645 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
12646 	tp->snd_wnd -= acked_amount;
12647 	mfree = sbcut_locked(&so->so_snd, acked_amount);
12648 	if ((sbused(&so->so_snd) == 0) &&
12649 	    (acked > acked_amount) &&
12650 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
12651 	    (tp->t_flags & TF_SENTFIN)) {
12652 		/*
12653 		 * We must be sure our fin
12654 		 * was sent and acked (we can be
12655 		 * in FIN_WAIT_1 without having
12656 		 * sent the fin).
12657 		 */
12658 		ourfinisacked = 1;
12659 	}
12660 	tp->snd_una = th->th_ack;
12661 	/* wakeups? */
12662 	if (acked_amount && sbavail(&so->so_snd))
12663 		rack_adjust_sendmap_head(rack, &so->so_snd);
12664 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12665 	/* NB: sowwakeup_locked() does an implicit unlock. */
12666 	sowwakeup_locked(so);
12667 	m_freem(mfree);
12668 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
12669 		tp->snd_recover = tp->snd_una;
12670 
12671 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12672 		tp->snd_nxt = tp->snd_max;
12673 	}
12674 	if (under_pacing &&
12675 	    (rack->use_fixed_rate == 0) &&
12676 	    (rack->in_probe_rtt == 0) &&
12677 	    rack->rc_gp_dyn_mul &&
12678 	    rack->rc_always_pace) {
12679 		/* Check if we are dragging bottom */
12680 		rack_check_bottom_drag(tp, rack, so);
12681 	}
12682 	if (tp->snd_una == tp->snd_max) {
12683 		/* Nothing left outstanding */
12684 		tp->t_flags &= ~TF_PREVVALID;
12685 		rack->r_ctl.idle_snd_una = tp->snd_una;
12686 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12687 		if (rack->r_ctl.rc_went_idle_time == 0)
12688 			rack->r_ctl.rc_went_idle_time = 1;
12689 		rack->r_ctl.retran_during_recovery = 0;
12690 		rack->r_ctl.dsack_byte_cnt = 0;
12691 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12692 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12693 			tp->t_acktime = 0;
12694 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12695 		rack->rc_suspicious = 0;
12696 		/* Set need output so persist might get set */
12697 		rack->r_wanted_output = 1;
12698 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12699 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12700 		    (sbavail(&so->so_snd) == 0) &&
12701 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12702 			/*
12703 			 * The socket was gone and the
12704 			 * peer sent data (now or in the past), time to
12705 			 * reset him.
12706 			 */
12707 			*ret_val = 1;
12708 			/* tcp_close will kill the inp pre-log the Reset */
12709 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12710 			tp = tcp_close(tp);
12711 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12712 			return (1);
12713 		}
12714 	}
12715 	if (ofia)
12716 		*ofia = ourfinisacked;
12717 	return (0);
12718 }
12719 
12720 
12721 static void
12722 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12723 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
12724 {
12725 	if (tcp_bblogging_on(rack->rc_tp)) {
12726 		union tcp_log_stackspecific log;
12727 		struct timeval tv;
12728 
12729 		memset(&log, 0, sizeof(log));
12730 		log.u_bbr.flex1 = cnt;
12731 		log.u_bbr.flex2 = split;
12732 		log.u_bbr.flex3 = out;
12733 		log.u_bbr.flex4 = line;
12734 		log.u_bbr.flex5 = rack->r_must_retran;
12735 		log.u_bbr.flex6 = flags;
12736 		log.u_bbr.flex7 = rack->rc_has_collapsed;
12737 		log.u_bbr.flex8 = dir;	/*
12738 					 * 1 is collapsed, 0 is uncollapsed,
12739 					 * 2 is log of a rsm being marked, 3 is a split.
12740 					 */
12741 		if (rsm == NULL)
12742 			log.u_bbr.rttProp = 0;
12743 		else
12744 			log.u_bbr.rttProp = (uintptr_t)rsm;
12745 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12746 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12747 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
12748 		    &rack->rc_inp->inp_socket->so_rcv,
12749 		    &rack->rc_inp->inp_socket->so_snd,
12750 		    TCP_RACK_LOG_COLLAPSE, 0,
12751 		    0, &log, false, &tv);
12752 	}
12753 }
12754 
12755 static void
12756 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12757 {
12758 	/*
12759 	 * Here all we do is mark the collapsed point and set the flag.
12760 	 * This may happen again and again, but there is no
12761 	 * sense splitting our map until we know where the
12762 	 * peer finally lands in the collapse.
12763 	 */
12764 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12765 	if ((rack->rc_has_collapsed == 0) ||
12766 	    (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12767 		counter_u64_add(rack_collapsed_win_seen, 1);
12768 	rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12769 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12770 	rack->rc_has_collapsed = 1;
12771 	rack->r_collapse_point_valid = 1;
12772 	rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12773 }
12774 
12775 static void
12776 rack_un_collapse_window(struct tcp_rack *rack, int line)
12777 {
12778 	struct rack_sendmap *nrsm, *rsm;
12779 	int cnt = 0, split = 0;
12780 	int insret __diagused;
12781 
12782 
12783 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12784 	rack->rc_has_collapsed = 0;
12785 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12786 	if (rsm == NULL) {
12787 		/* Nothing to do maybe the peer ack'ed it all */
12788 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12789 		return;
12790 	}
12791 	/* Now do we need to split this one? */
12792 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12793 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12794 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12795 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12796 		if (nrsm == NULL) {
12797 			/* We can't get a rsm, mark all? */
12798 			nrsm = rsm;
12799 			goto no_split;
12800 		}
12801 		/* Clone it */
12802 		split = 1;
12803 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12804 #ifndef INVARIANTS
12805 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12806 #else
12807 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12808 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
12809 			      nrsm, insret, rack, rsm);
12810 		}
12811 #endif
12812 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12813 				 rack->r_ctl.last_collapse_point, __LINE__);
12814 		if (rsm->r_in_tmap) {
12815 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12816 			nrsm->r_in_tmap = 1;
12817 		}
12818 		/*
12819 		 * Set in the new RSM as the
12820 		 * collapsed starting point
12821 		 */
12822 		rsm = nrsm;
12823 	}
12824 
12825 no_split:
12826 	TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
12827 		cnt++;
12828 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
12829 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12830 		cnt++;
12831 	}
12832 	if (cnt) {
12833 		counter_u64_add(rack_collapsed_win, 1);
12834 	}
12835 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12836 }
12837 
12838 static void
12839 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12840 			int32_t tlen, int32_t tfo_syn)
12841 {
12842 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
12843 		rack_timer_cancel(tp, rack,
12844 				  rack->r_ctl.rc_rcvtime, __LINE__);
12845 		tp->t_flags |= TF_DELACK;
12846 	} else {
12847 		rack->r_wanted_output = 1;
12848 		tp->t_flags |= TF_ACKNOW;
12849 	}
12850 }
12851 
12852 static void
12853 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12854 {
12855 	/*
12856 	 * If fast output is in progress, lets validate that
12857 	 * the new window did not shrink on us and make it
12858 	 * so fast output should end.
12859 	 */
12860 	if (rack->r_fast_output) {
12861 		uint32_t out;
12862 
12863 		/*
12864 		 * Calculate what we will send if left as is
12865 		 * and compare that to our send window.
12866 		 */
12867 		out = ctf_outstanding(tp);
12868 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12869 			/* ok we have an issue */
12870 			if (out >= tp->snd_wnd) {
12871 				/* Turn off fast output the window is met or collapsed */
12872 				rack->r_fast_output = 0;
12873 			} else {
12874 				/* we have some room left */
12875 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12876 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12877 					/* If not at least 1 full segment never mind */
12878 					rack->r_fast_output = 0;
12879 				}
12880 			}
12881 		}
12882 	}
12883 }
12884 
12885 /*
12886  * Return value of 1, the TCB is unlocked and most
12887  * likely gone, return value of 0, the TCP is still
12888  * locked.
12889  */
12890 static int
12891 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12892     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12893     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12894 {
12895 	/*
12896 	 * Update window information. Don't look at window if no ACK: TAC's
12897 	 * send garbage on first SYN.
12898 	 */
12899 	int32_t nsegs;
12900 	int32_t tfo_syn;
12901 	struct tcp_rack *rack;
12902 
12903 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12904 
12905 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12906 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12907 	if ((thflags & TH_ACK) &&
12908 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12909 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12910 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12911 		/* keep track of pure window updates */
12912 		if (tlen == 0 &&
12913 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12914 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12915 		tp->snd_wnd = tiwin;
12916 		rack_validate_fo_sendwin_up(tp, rack);
12917 		tp->snd_wl1 = th->th_seq;
12918 		tp->snd_wl2 = th->th_ack;
12919 		if (tp->snd_wnd > tp->max_sndwnd)
12920 			tp->max_sndwnd = tp->snd_wnd;
12921 		rack->r_wanted_output = 1;
12922 	} else if (thflags & TH_ACK) {
12923 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12924 			tp->snd_wnd = tiwin;
12925 			rack_validate_fo_sendwin_up(tp, rack);
12926 			tp->snd_wl1 = th->th_seq;
12927 			tp->snd_wl2 = th->th_ack;
12928 		}
12929 	}
12930 	if (tp->snd_wnd < ctf_outstanding(tp))
12931 		/* The peer collapsed the window */
12932 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12933 	else if (rack->rc_has_collapsed)
12934 		rack_un_collapse_window(rack, __LINE__);
12935 	if ((rack->r_collapse_point_valid) &&
12936 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12937 		rack->r_collapse_point_valid = 0;
12938 	/* Was persist timer active and now we have window space? */
12939 	if ((rack->rc_in_persist != 0) &&
12940 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12941 				rack->r_ctl.rc_pace_min_segs))) {
12942 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12943 		tp->snd_nxt = tp->snd_max;
12944 		/* Make sure we output to start the timer */
12945 		rack->r_wanted_output = 1;
12946 	}
12947 	/* Do we enter persists? */
12948 	if ((rack->rc_in_persist == 0) &&
12949 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12950 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12951 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12952 	    sbavail(&tptosocket(tp)->so_snd) &&
12953 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12954 		/*
12955 		 * Here the rwnd is less than
12956 		 * the pacing size, we are established,
12957 		 * nothing is outstanding, and there is
12958 		 * data to send. Enter persists.
12959 		 */
12960 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12961 	}
12962 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12963 		m_freem(m);
12964 		return (0);
12965 	}
12966 	/*
12967 	 * don't process the URG bit, ignore them drag
12968 	 * along the up.
12969 	 */
12970 	tp->rcv_up = tp->rcv_nxt;
12971 
12972 	/*
12973 	 * Process the segment text, merging it into the TCP sequencing
12974 	 * queue, and arranging for acknowledgment of receipt if necessary.
12975 	 * This process logically involves adjusting tp->rcv_wnd as data is
12976 	 * presented to the user (this happens in tcp_usrreq.c, case
12977 	 * PRU_RCVD).  If a FIN has already been received on this connection
12978 	 * then we just ignore the text.
12979 	 */
12980 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12981 	    (tp->t_flags & TF_FASTOPEN));
12982 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12983 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12984 		tcp_seq save_start = th->th_seq;
12985 		tcp_seq save_rnxt  = tp->rcv_nxt;
12986 		int     save_tlen  = tlen;
12987 
12988 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12989 		/*
12990 		 * Insert segment which includes th into TCP reassembly
12991 		 * queue with control block tp.  Set thflags to whether
12992 		 * reassembly now includes a segment with FIN.  This handles
12993 		 * the common case inline (segment is the next to be
12994 		 * received on an established connection, and the queue is
12995 		 * empty), avoiding linkage into and removal from the queue
12996 		 * and repetition of various conversions. Set DELACK for
12997 		 * segments received in order, but ack immediately when
12998 		 * segments are out of order (so fast retransmit can work).
12999 		 */
13000 		if (th->th_seq == tp->rcv_nxt &&
13001 		    SEGQ_EMPTY(tp) &&
13002 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
13003 		    tfo_syn)) {
13004 #ifdef NETFLIX_SB_LIMITS
13005 			u_int mcnt, appended;
13006 
13007 			if (so->so_rcv.sb_shlim) {
13008 				mcnt = m_memcnt(m);
13009 				appended = 0;
13010 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
13011 				    CFO_NOSLEEP, NULL) == false) {
13012 					counter_u64_add(tcp_sb_shlim_fails, 1);
13013 					m_freem(m);
13014 					return (0);
13015 				}
13016 			}
13017 #endif
13018 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
13019 			tp->rcv_nxt += tlen;
13020 			if (tlen &&
13021 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
13022 			    (tp->t_fbyte_in == 0)) {
13023 				tp->t_fbyte_in = ticks;
13024 				if (tp->t_fbyte_in == 0)
13025 					tp->t_fbyte_in = 1;
13026 				if (tp->t_fbyte_out && tp->t_fbyte_in)
13027 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
13028 			}
13029 			thflags = tcp_get_flags(th) & TH_FIN;
13030 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
13031 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
13032 			SOCKBUF_LOCK(&so->so_rcv);
13033 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13034 				m_freem(m);
13035 			} else {
13036 				int32_t newsize;
13037 
13038 				if (tlen > 0) {
13039 					newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
13040 					if (newsize)
13041 						if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
13042 							so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
13043 				}
13044 #ifdef NETFLIX_SB_LIMITS
13045 				appended =
13046 #endif
13047 					sbappendstream_locked(&so->so_rcv, m, 0);
13048 			}
13049 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
13050 			/* NB: sorwakeup_locked() does an implicit unlock. */
13051 			sorwakeup_locked(so);
13052 #ifdef NETFLIX_SB_LIMITS
13053 			if (so->so_rcv.sb_shlim && appended != mcnt)
13054 				counter_fo_release(so->so_rcv.sb_shlim,
13055 				    mcnt - appended);
13056 #endif
13057 		} else {
13058 			/*
13059 			 * XXX: Due to the header drop above "th" is
13060 			 * theoretically invalid by now.  Fortunately
13061 			 * m_adj() doesn't actually frees any mbufs when
13062 			 * trimming from the head.
13063 			 */
13064 			tcp_seq temp = save_start;
13065 
13066 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
13067 			tp->t_flags |= TF_ACKNOW;
13068 			if (tp->t_flags & TF_WAKESOR) {
13069 				tp->t_flags &= ~TF_WAKESOR;
13070 				/* NB: sorwakeup_locked() does an implicit unlock. */
13071 				sorwakeup_locked(so);
13072 			}
13073 		}
13074 		if ((tp->t_flags & TF_SACK_PERMIT) &&
13075 		    (save_tlen > 0) &&
13076 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
13077 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
13078 				/*
13079 				 * DSACK actually handled in the fastpath
13080 				 * above.
13081 				 */
13082 				tcp_update_sack_list(tp, save_start,
13083 				    save_start + save_tlen);
13084 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
13085 				if ((tp->rcv_numsacks >= 1) &&
13086 				    (tp->sackblks[0].end == save_start)) {
13087 					/*
13088 					 * Partial overlap, recorded at todrop
13089 					 * above.
13090 					 */
13091 					tcp_update_sack_list(tp,
13092 					    tp->sackblks[0].start,
13093 					    tp->sackblks[0].end);
13094 				} else {
13095 					tcp_update_dsack_list(tp, save_start,
13096 					    save_start + save_tlen);
13097 				}
13098 			} else if (tlen >= save_tlen) {
13099 				/* Update of sackblks. */
13100 				tcp_update_dsack_list(tp, save_start,
13101 				    save_start + save_tlen);
13102 			} else if (tlen > 0) {
13103 				tcp_update_dsack_list(tp, save_start,
13104 				    save_start + tlen);
13105 			}
13106 		}
13107 	} else {
13108 		m_freem(m);
13109 		thflags &= ~TH_FIN;
13110 	}
13111 
13112 	/*
13113 	 * If FIN is received ACK the FIN and let the user know that the
13114 	 * connection is closing.
13115 	 */
13116 	if (thflags & TH_FIN) {
13117 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
13118 			/* The socket upcall is handled by socantrcvmore. */
13119 			socantrcvmore(so);
13120 			/*
13121 			 * If connection is half-synchronized (ie NEEDSYN
13122 			 * flag on) then delay ACK, so it may be piggybacked
13123 			 * when SYN is sent. Otherwise, since we received a
13124 			 * FIN then no more input can be expected, send ACK
13125 			 * now.
13126 			 */
13127 			if (tp->t_flags & TF_NEEDSYN) {
13128 				rack_timer_cancel(tp, rack,
13129 				    rack->r_ctl.rc_rcvtime, __LINE__);
13130 				tp->t_flags |= TF_DELACK;
13131 			} else {
13132 				tp->t_flags |= TF_ACKNOW;
13133 			}
13134 			tp->rcv_nxt++;
13135 		}
13136 		switch (tp->t_state) {
13137 			/*
13138 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
13139 			 * CLOSE_WAIT state.
13140 			 */
13141 		case TCPS_SYN_RECEIVED:
13142 			tp->t_starttime = ticks;
13143 			/* FALLTHROUGH */
13144 		case TCPS_ESTABLISHED:
13145 			rack_timer_cancel(tp, rack,
13146 			    rack->r_ctl.rc_rcvtime, __LINE__);
13147 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
13148 			break;
13149 
13150 			/*
13151 			 * If still in FIN_WAIT_1 STATE FIN has not been
13152 			 * acked so enter the CLOSING state.
13153 			 */
13154 		case TCPS_FIN_WAIT_1:
13155 			rack_timer_cancel(tp, rack,
13156 			    rack->r_ctl.rc_rcvtime, __LINE__);
13157 			tcp_state_change(tp, TCPS_CLOSING);
13158 			break;
13159 
13160 			/*
13161 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
13162 			 * starting the time-wait timer, turning off the
13163 			 * other standard timers.
13164 			 */
13165 		case TCPS_FIN_WAIT_2:
13166 			rack_timer_cancel(tp, rack,
13167 			    rack->r_ctl.rc_rcvtime, __LINE__);
13168 			tcp_twstart(tp);
13169 			return (1);
13170 		}
13171 	}
13172 	/*
13173 	 * Return any desired output.
13174 	 */
13175 	if ((tp->t_flags & TF_ACKNOW) ||
13176 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
13177 		rack->r_wanted_output = 1;
13178 	}
13179 	return (0);
13180 }
13181 
13182 /*
13183  * Here nothing is really faster, its just that we
13184  * have broken out the fast-data path also just like
13185  * the fast-ack.
13186  */
13187 static int
13188 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
13189     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13190     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
13191 {
13192 	int32_t nsegs;
13193 	int32_t newsize = 0;	/* automatic sockbuf scaling */
13194 	struct tcp_rack *rack;
13195 #ifdef NETFLIX_SB_LIMITS
13196 	u_int mcnt, appended;
13197 #endif
13198 
13199 	/*
13200 	 * If last ACK falls within this segment's sequence numbers, record
13201 	 * the timestamp. NOTE that the test is modified according to the
13202 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
13203 	 */
13204 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
13205 		return (0);
13206 	}
13207 	if (tiwin && tiwin != tp->snd_wnd) {
13208 		return (0);
13209 	}
13210 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
13211 		return (0);
13212 	}
13213 	if (__predict_false((to->to_flags & TOF_TS) &&
13214 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
13215 		return (0);
13216 	}
13217 	if (__predict_false((th->th_ack != tp->snd_una))) {
13218 		return (0);
13219 	}
13220 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
13221 		return (0);
13222 	}
13223 	if ((to->to_flags & TOF_TS) != 0 &&
13224 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
13225 		tp->ts_recent_age = tcp_ts_getticks();
13226 		tp->ts_recent = to->to_tsval;
13227 	}
13228 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13229 	/*
13230 	 * This is a pure, in-sequence data packet with nothing on the
13231 	 * reassembly queue and we have enough buffer space to take it.
13232 	 */
13233 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
13234 
13235 #ifdef NETFLIX_SB_LIMITS
13236 	if (so->so_rcv.sb_shlim) {
13237 		mcnt = m_memcnt(m);
13238 		appended = 0;
13239 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
13240 		    CFO_NOSLEEP, NULL) == false) {
13241 			counter_u64_add(tcp_sb_shlim_fails, 1);
13242 			m_freem(m);
13243 			return (1);
13244 		}
13245 	}
13246 #endif
13247 	/* Clean receiver SACK report if present */
13248 	if (tp->rcv_numsacks)
13249 		tcp_clean_sackreport(tp);
13250 	KMOD_TCPSTAT_INC(tcps_preddat);
13251 	tp->rcv_nxt += tlen;
13252 	if (tlen &&
13253 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
13254 	    (tp->t_fbyte_in == 0)) {
13255 		tp->t_fbyte_in = ticks;
13256 		if (tp->t_fbyte_in == 0)
13257 			tp->t_fbyte_in = 1;
13258 		if (tp->t_fbyte_out && tp->t_fbyte_in)
13259 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
13260 	}
13261 	/*
13262 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
13263 	 */
13264 	tp->snd_wl1 = th->th_seq;
13265 	/*
13266 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
13267 	 */
13268 	tp->rcv_up = tp->rcv_nxt;
13269 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
13270 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
13271 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
13272 
13273 	/* Add data to socket buffer. */
13274 	SOCKBUF_LOCK(&so->so_rcv);
13275 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13276 		m_freem(m);
13277 	} else {
13278 		/*
13279 		 * Set new socket buffer size. Give up when limit is
13280 		 * reached.
13281 		 */
13282 		if (newsize)
13283 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
13284 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
13285 		m_adj(m, drop_hdrlen);	/* delayed header drop */
13286 #ifdef NETFLIX_SB_LIMITS
13287 		appended =
13288 #endif
13289 			sbappendstream_locked(&so->so_rcv, m, 0);
13290 		ctf_calc_rwin(so, tp);
13291 	}
13292 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
13293 	/* NB: sorwakeup_locked() does an implicit unlock. */
13294 	sorwakeup_locked(so);
13295 #ifdef NETFLIX_SB_LIMITS
13296 	if (so->so_rcv.sb_shlim && mcnt != appended)
13297 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
13298 #endif
13299 	rack_handle_delayed_ack(tp, rack, tlen, 0);
13300 	if (tp->snd_una == tp->snd_max)
13301 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13302 	return (1);
13303 }
13304 
13305 /*
13306  * This subfunction is used to try to highly optimize the
13307  * fast path. We again allow window updates that are
13308  * in sequence to remain in the fast-path. We also add
13309  * in the __predict's to attempt to help the compiler.
13310  * Note that if we return a 0, then we can *not* process
13311  * it and the caller should push the packet into the
13312  * slow-path.
13313  */
13314 static int
13315 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13316     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13317     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
13318 {
13319 	int32_t acked;
13320 	int32_t nsegs;
13321 	int32_t under_pacing = 0;
13322 	struct tcp_rack *rack;
13323 
13324 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
13325 		/* Old ack, behind (or duplicate to) the last one rcv'd */
13326 		return (0);
13327 	}
13328 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
13329 		/* Above what we have sent? */
13330 		return (0);
13331 	}
13332 	if (__predict_false(tiwin == 0)) {
13333 		/* zero window */
13334 		return (0);
13335 	}
13336 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
13337 		/* We need a SYN or a FIN, unlikely.. */
13338 		return (0);
13339 	}
13340 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
13341 		/* Timestamp is behind .. old ack with seq wrap? */
13342 		return (0);
13343 	}
13344 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
13345 		/* Still recovering */
13346 		return (0);
13347 	}
13348 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13349 	if (rack->r_ctl.rc_sacked) {
13350 		/* We have sack holes on our scoreboard */
13351 		return (0);
13352 	}
13353 	/* Ok if we reach here, we can process a fast-ack */
13354 	if (rack->gp_ready &&
13355 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13356 		under_pacing = 1;
13357 	}
13358 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
13359 	rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
13360 	/* Did the window get updated? */
13361 	if (tiwin != tp->snd_wnd) {
13362 		tp->snd_wnd = tiwin;
13363 		rack_validate_fo_sendwin_up(tp, rack);
13364 		tp->snd_wl1 = th->th_seq;
13365 		if (tp->snd_wnd > tp->max_sndwnd)
13366 			tp->max_sndwnd = tp->snd_wnd;
13367 	}
13368 	/* Do we exit persists? */
13369 	if ((rack->rc_in_persist != 0) &&
13370 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13371 			       rack->r_ctl.rc_pace_min_segs))) {
13372 		rack_exit_persist(tp, rack, cts);
13373 	}
13374 	/* Do we enter persists? */
13375 	if ((rack->rc_in_persist == 0) &&
13376 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13377 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13378 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13379 	    sbavail(&tptosocket(tp)->so_snd) &&
13380 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13381 		/*
13382 		 * Here the rwnd is less than
13383 		 * the pacing size, we are established,
13384 		 * nothing is outstanding, and there is
13385 		 * data to send. Enter persists.
13386 		 */
13387 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
13388 	}
13389 	/*
13390 	 * If last ACK falls within this segment's sequence numbers, record
13391 	 * the timestamp. NOTE that the test is modified according to the
13392 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
13393 	 */
13394 	if ((to->to_flags & TOF_TS) != 0 &&
13395 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
13396 		tp->ts_recent_age = tcp_ts_getticks();
13397 		tp->ts_recent = to->to_tsval;
13398 	}
13399 	/*
13400 	 * This is a pure ack for outstanding data.
13401 	 */
13402 	KMOD_TCPSTAT_INC(tcps_predack);
13403 
13404 	/*
13405 	 * "bad retransmit" recovery.
13406 	 */
13407 	if ((tp->t_flags & TF_PREVVALID) &&
13408 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13409 		tp->t_flags &= ~TF_PREVVALID;
13410 		if (tp->t_rxtshift == 1 &&
13411 		    (int)(ticks - tp->t_badrxtwin) < 0)
13412 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
13413 	}
13414 	/*
13415 	 * Recalculate the transmit timer / rtt.
13416 	 *
13417 	 * Some boxes send broken timestamp replies during the SYN+ACK
13418 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
13419 	 * and blow up the retransmit timer.
13420 	 */
13421 	acked = BYTES_THIS_ACK(tp, th);
13422 
13423 #ifdef TCP_HHOOK
13424 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
13425 	hhook_run_tcp_est_in(tp, th, to);
13426 #endif
13427 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
13428 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13429 	if (acked) {
13430 		struct mbuf *mfree;
13431 
13432 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
13433 		SOCKBUF_LOCK(&so->so_snd);
13434 		mfree = sbcut_locked(&so->so_snd, acked);
13435 		tp->snd_una = th->th_ack;
13436 		/* Note we want to hold the sb lock through the sendmap adjust */
13437 		rack_adjust_sendmap_head(rack, &so->so_snd);
13438 		/* Wake up the socket if we have room to write more */
13439 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13440 		sowwakeup_locked(so);
13441 		m_freem(mfree);
13442 		tp->t_rxtshift = 0;
13443 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13444 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13445 		rack->rc_tlp_in_progress = 0;
13446 		rack->r_ctl.rc_tlp_cnt_out = 0;
13447 		/*
13448 		 * If it is the RXT timer we want to
13449 		 * stop it, so we can restart a TLP.
13450 		 */
13451 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13452 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13453 
13454 #ifdef TCP_REQUEST_TRK
13455 		rack_req_check_for_comp(rack, th->th_ack);
13456 #endif
13457 	}
13458 	/*
13459 	 * Let the congestion control algorithm update congestion control
13460 	 * related information. This typically means increasing the
13461 	 * congestion window.
13462 	 */
13463 	if (tp->snd_wnd < ctf_outstanding(tp)) {
13464 		/* The peer collapsed the window */
13465 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
13466 	} else if (rack->rc_has_collapsed)
13467 		rack_un_collapse_window(rack, __LINE__);
13468 	if ((rack->r_collapse_point_valid) &&
13469 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
13470 		rack->r_collapse_point_valid = 0;
13471 	/*
13472 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
13473 	 */
13474 	tp->snd_wl2 = th->th_ack;
13475 	tp->t_dupacks = 0;
13476 	m_freem(m);
13477 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
13478 
13479 	/*
13480 	 * If all outstanding data are acked, stop retransmit timer,
13481 	 * otherwise restart timer using current (possibly backed-off)
13482 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
13483 	 * If data are ready to send, let tcp_output decide between more
13484 	 * output or persist.
13485 	 */
13486 	if (under_pacing &&
13487 	    (rack->use_fixed_rate == 0) &&
13488 	    (rack->in_probe_rtt == 0) &&
13489 	    rack->rc_gp_dyn_mul &&
13490 	    rack->rc_always_pace) {
13491 		/* Check if we are dragging bottom */
13492 		rack_check_bottom_drag(tp, rack, so);
13493 	}
13494 	if (tp->snd_una == tp->snd_max) {
13495 		tp->t_flags &= ~TF_PREVVALID;
13496 		rack->r_ctl.retran_during_recovery = 0;
13497 		rack->rc_suspicious = 0;
13498 		rack->r_ctl.dsack_byte_cnt = 0;
13499 		rack->r_ctl.idle_snd_una = tp->snd_una;
13500 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13501 		if (rack->r_ctl.rc_went_idle_time == 0)
13502 			rack->r_ctl.rc_went_idle_time = 1;
13503 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13504 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
13505 			tp->t_acktime = 0;
13506 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13507 	}
13508 	if (acked && rack->r_fast_output)
13509 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
13510 	if (sbavail(&so->so_snd)) {
13511 		rack->r_wanted_output = 1;
13512 	}
13513 	return (1);
13514 }
13515 
13516 /*
13517  * Return value of 1, the TCB is unlocked and most
13518  * likely gone, return value of 0, the TCP is still
13519  * locked.
13520  */
13521 static int
13522 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
13523     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13524     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13525 {
13526 	int32_t ret_val = 0;
13527 	int32_t orig_tlen = tlen;
13528 	int32_t todrop;
13529 	int32_t ourfinisacked = 0;
13530 	struct tcp_rack *rack;
13531 
13532 	INP_WLOCK_ASSERT(tptoinpcb(tp));
13533 
13534 	ctf_calc_rwin(so, tp);
13535 	/*
13536 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
13537 	 * SYN, drop the input. if seg contains a RST, then drop the
13538 	 * connection. if seg does not contain SYN, then drop it. Otherwise
13539 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
13540 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
13541 	 * contains an ECE and ECN support is enabled, the stream is ECN
13542 	 * capable. if SYN has been acked change to ESTABLISHED else
13543 	 * SYN_RCVD state arrange for segment to be acked (eventually)
13544 	 * continue processing rest of data/controls.
13545 	 */
13546 	if ((thflags & TH_ACK) &&
13547 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
13548 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13549 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13550 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13551 		return (1);
13552 	}
13553 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
13554 		TCP_PROBE5(connect__refused, NULL, tp,
13555 		    mtod(m, const char *), tp, th);
13556 		tp = tcp_drop(tp, ECONNREFUSED);
13557 		ctf_do_drop(m, tp);
13558 		return (1);
13559 	}
13560 	if (thflags & TH_RST) {
13561 		ctf_do_drop(m, tp);
13562 		return (1);
13563 	}
13564 	if (!(thflags & TH_SYN)) {
13565 		ctf_do_drop(m, tp);
13566 		return (1);
13567 	}
13568 	tp->irs = th->th_seq;
13569 	tcp_rcvseqinit(tp);
13570 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13571 	if (thflags & TH_ACK) {
13572 		int tfo_partial = 0;
13573 
13574 		KMOD_TCPSTAT_INC(tcps_connects);
13575 		soisconnected(so);
13576 #ifdef MAC
13577 		mac_socketpeer_set_from_mbuf(m, so);
13578 #endif
13579 		/* Do window scaling on this connection? */
13580 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13581 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13582 			tp->rcv_scale = tp->request_r_scale;
13583 		}
13584 		tp->rcv_adv += min(tp->rcv_wnd,
13585 		    TCP_MAXWIN << tp->rcv_scale);
13586 		/*
13587 		 * If not all the data that was sent in the TFO SYN
13588 		 * has been acked, resend the remainder right away.
13589 		 */
13590 		if ((tp->t_flags & TF_FASTOPEN) &&
13591 		    (tp->snd_una != tp->snd_max)) {
13592 			/* Was it a partial ack? */
13593 			if (SEQ_LT(th->th_ack, tp->snd_max))
13594 				tfo_partial = 1;
13595 		}
13596 		/*
13597 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
13598 		 * will be turned on later.
13599 		 */
13600 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
13601 			rack_timer_cancel(tp, rack,
13602 					  rack->r_ctl.rc_rcvtime, __LINE__);
13603 			tp->t_flags |= TF_DELACK;
13604 		} else {
13605 			rack->r_wanted_output = 1;
13606 			tp->t_flags |= TF_ACKNOW;
13607 		}
13608 
13609 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
13610 
13611 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
13612 			/*
13613 			 * We advance snd_una for the
13614 			 * fast open case. If th_ack is
13615 			 * acknowledging data beyond
13616 			 * snd_una we can't just call
13617 			 * ack-processing since the
13618 			 * data stream in our send-map
13619 			 * will start at snd_una + 1 (one
13620 			 * beyond the SYN). If its just
13621 			 * equal we don't need to do that
13622 			 * and there is no send_map.
13623 			 */
13624 			tp->snd_una++;
13625 			if (tfo_partial && (SEQ_GT(tp->snd_max, tp->snd_una))) {
13626 				/*
13627 				 * We sent a SYN with data, and thus have a
13628 				 * sendmap entry with a SYN set. Lets find it
13629 				 * and take off the send bit and the byte and
13630 				 * set it up to be what we send (send it next).
13631 				 */
13632 				struct rack_sendmap *rsm;
13633 
13634 				rsm = tqhash_min(rack->r_ctl.tqh);
13635 				if (rsm) {
13636 					if (rsm->r_flags & RACK_HAS_SYN) {
13637 						rsm->r_flags &= ~RACK_HAS_SYN;
13638 						rsm->r_start++;
13639 					}
13640 					rack->r_ctl.rc_resend = rsm;
13641 				}
13642 			}
13643 		}
13644 		/*
13645 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
13646 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
13647 		 */
13648 		tp->t_starttime = ticks;
13649 		if (tp->t_flags & TF_NEEDFIN) {
13650 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
13651 			tp->t_flags &= ~TF_NEEDFIN;
13652 			thflags &= ~TH_SYN;
13653 		} else {
13654 			tcp_state_change(tp, TCPS_ESTABLISHED);
13655 			TCP_PROBE5(connect__established, NULL, tp,
13656 			    mtod(m, const char *), tp, th);
13657 			rack_cc_conn_init(tp);
13658 		}
13659 	} else {
13660 		/*
13661 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
13662 		 * open.  If segment contains CC option and there is a
13663 		 * cached CC, apply TAO test. If it succeeds, connection is *
13664 		 * half-synchronized. Otherwise, do 3-way handshake:
13665 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
13666 		 * there was no CC option, clear cached CC value.
13667 		 */
13668 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
13669 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
13670 	}
13671 	/*
13672 	 * Advance th->th_seq to correspond to first data byte. If data,
13673 	 * trim to stay within window, dropping FIN if necessary.
13674 	 */
13675 	th->th_seq++;
13676 	if (tlen > tp->rcv_wnd) {
13677 		todrop = tlen - tp->rcv_wnd;
13678 		m_adj(m, -todrop);
13679 		tlen = tp->rcv_wnd;
13680 		thflags &= ~TH_FIN;
13681 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13682 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13683 	}
13684 	tp->snd_wl1 = th->th_seq - 1;
13685 	tp->rcv_up = th->th_seq;
13686 	/*
13687 	 * Client side of transaction: already sent SYN and data. If the
13688 	 * remote host used T/TCP to validate the SYN, our data will be
13689 	 * ACK'd; if so, enter normal data segment processing in the middle
13690 	 * of step 5, ack processing. Otherwise, goto step 6.
13691 	 */
13692 	if (thflags & TH_ACK) {
13693 		/* For syn-sent we need to possibly update the rtt */
13694 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13695 			uint32_t t, mcts;
13696 
13697 			mcts = tcp_ts_getticks();
13698 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13699 			if (!tp->t_rttlow || tp->t_rttlow > t)
13700 				tp->t_rttlow = t;
13701 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13702 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13703 			tcp_rack_xmit_timer_commit(rack, tp);
13704 		}
13705 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen))
13706 			return (ret_val);
13707 		/* We may have changed to FIN_WAIT_1 above */
13708 		if (tp->t_state == TCPS_FIN_WAIT_1) {
13709 			/*
13710 			 * In FIN_WAIT_1 STATE in addition to the processing
13711 			 * for the ESTABLISHED state if our FIN is now
13712 			 * acknowledged then enter FIN_WAIT_2.
13713 			 */
13714 			if (ourfinisacked) {
13715 				/*
13716 				 * If we can't receive any more data, then
13717 				 * closing user can proceed. Starting the
13718 				 * timer is contrary to the specification,
13719 				 * but if we don't get a FIN we'll hang
13720 				 * forever.
13721 				 *
13722 				 * XXXjl: we should release the tp also, and
13723 				 * use a compressed state.
13724 				 */
13725 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13726 					soisdisconnected(so);
13727 					tcp_timer_activate(tp, TT_2MSL,
13728 					    (tcp_fast_finwait2_recycle ?
13729 					    tcp_finwait2_timeout :
13730 					    TP_MAXIDLE(tp)));
13731 				}
13732 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13733 			}
13734 		}
13735 	}
13736 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13737 	   tiwin, thflags, nxt_pkt));
13738 }
13739 
13740 /*
13741  * Return value of 1, the TCB is unlocked and most
13742  * likely gone, return value of 0, the TCP is still
13743  * locked.
13744  */
13745 static int
13746 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13747     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13748     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13749 {
13750 	struct tcp_rack *rack;
13751 	int32_t orig_tlen = tlen;
13752 	int32_t ret_val = 0;
13753 	int32_t ourfinisacked = 0;
13754 
13755 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13756 	ctf_calc_rwin(so, tp);
13757 	if ((thflags & TH_RST) ||
13758 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13759 		return (__ctf_process_rst(m, th, so, tp,
13760 					  &rack->r_ctl.challenge_ack_ts,
13761 					  &rack->r_ctl.challenge_ack_cnt));
13762 	if ((thflags & TH_ACK) &&
13763 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13764 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13765 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13766 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13767 		return (1);
13768 	}
13769 	if (tp->t_flags & TF_FASTOPEN) {
13770 		/*
13771 		 * When a TFO connection is in SYN_RECEIVED, the
13772 		 * only valid packets are the initial SYN, a
13773 		 * retransmit/copy of the initial SYN (possibly with
13774 		 * a subset of the original data), a valid ACK, a
13775 		 * FIN, or a RST.
13776 		 */
13777 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13778 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13779 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13780 			return (1);
13781 		} else if (thflags & TH_SYN) {
13782 			/* non-initial SYN is ignored */
13783 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13784 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13785 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13786 				ctf_do_drop(m, NULL);
13787 				return (0);
13788 			}
13789 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13790 			ctf_do_drop(m, NULL);
13791 			return (0);
13792 		}
13793 	}
13794 
13795 	/*
13796 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13797 	 * it's less than ts_recent, drop it.
13798 	 */
13799 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13800 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13801 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13802 			return (ret_val);
13803 	}
13804 	/*
13805 	 * In the SYN-RECEIVED state, validate that the packet belongs to
13806 	 * this connection before trimming the data to fit the receive
13807 	 * window.  Check the sequence number versus IRS since we know the
13808 	 * sequence numbers haven't wrapped.  This is a partial fix for the
13809 	 * "LAND" DoS attack.
13810 	 */
13811 	if (SEQ_LT(th->th_seq, tp->irs)) {
13812 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13813 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13814 		return (1);
13815 	}
13816 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13817 			      &rack->r_ctl.challenge_ack_ts,
13818 			      &rack->r_ctl.challenge_ack_cnt)) {
13819 		return (ret_val);
13820 	}
13821 	/*
13822 	 * If last ACK falls within this segment's sequence numbers, record
13823 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13824 	 * from the latest proposal of the tcplw@cray.com list (Braden
13825 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13826 	 * with our earlier PAWS tests, so this check should be solely
13827 	 * predicated on the sequence space of this segment. 3) That we
13828 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13829 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13830 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13831 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13832 	 * p.869. In such cases, we can still calculate the RTT correctly
13833 	 * when RCV.NXT == Last.ACK.Sent.
13834 	 */
13835 	if ((to->to_flags & TOF_TS) != 0 &&
13836 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13837 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13838 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13839 		tp->ts_recent_age = tcp_ts_getticks();
13840 		tp->ts_recent = to->to_tsval;
13841 	}
13842 	tp->snd_wnd = tiwin;
13843 	rack_validate_fo_sendwin_up(tp, rack);
13844 	/*
13845 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13846 	 * is on (half-synchronized state), then queue data for later
13847 	 * processing; else drop segment and return.
13848 	 */
13849 	if ((thflags & TH_ACK) == 0) {
13850 		if (tp->t_flags & TF_FASTOPEN) {
13851 			rack_cc_conn_init(tp);
13852 		}
13853 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13854 		    tiwin, thflags, nxt_pkt));
13855 	}
13856 	KMOD_TCPSTAT_INC(tcps_connects);
13857 	if (tp->t_flags & TF_SONOTCONN) {
13858 		tp->t_flags &= ~TF_SONOTCONN;
13859 		soisconnected(so);
13860 	}
13861 	/* Do window scaling? */
13862 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13863 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13864 		tp->rcv_scale = tp->request_r_scale;
13865 	}
13866 	/*
13867 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
13868 	 * FIN-WAIT-1
13869 	 */
13870 	tp->t_starttime = ticks;
13871 	if ((tp->t_flags & TF_FASTOPEN) && tp->t_tfo_pending) {
13872 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13873 		tp->t_tfo_pending = NULL;
13874 	}
13875 	if (tp->t_flags & TF_NEEDFIN) {
13876 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
13877 		tp->t_flags &= ~TF_NEEDFIN;
13878 	} else {
13879 		tcp_state_change(tp, TCPS_ESTABLISHED);
13880 		TCP_PROBE5(accept__established, NULL, tp,
13881 		    mtod(m, const char *), tp, th);
13882 		/*
13883 		 * TFO connections call cc_conn_init() during SYN
13884 		 * processing.  Calling it again here for such connections
13885 		 * is not harmless as it would undo the snd_cwnd reduction
13886 		 * that occurs when a TFO SYN|ACK is retransmitted.
13887 		 */
13888 		if (!(tp->t_flags & TF_FASTOPEN))
13889 			rack_cc_conn_init(tp);
13890 	}
13891 	/*
13892 	 * Account for the ACK of our SYN prior to
13893 	 * regular ACK processing below, except for
13894 	 * simultaneous SYN, which is handled later.
13895 	 */
13896 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13897 		tp->snd_una++;
13898 	/*
13899 	 * If segment contains data or ACK, will call tcp_reass() later; if
13900 	 * not, do so now to pass queued data to user.
13901 	 */
13902 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
13903 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13904 		    (struct mbuf *)0);
13905 		if (tp->t_flags & TF_WAKESOR) {
13906 			tp->t_flags &= ~TF_WAKESOR;
13907 			/* NB: sorwakeup_locked() does an implicit unlock. */
13908 			sorwakeup_locked(so);
13909 		}
13910 	}
13911 	tp->snd_wl1 = th->th_seq - 1;
13912 	/* For syn-recv we need to possibly update the rtt */
13913 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13914 		uint32_t t, mcts;
13915 
13916 		mcts = tcp_ts_getticks();
13917 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13918 		if (!tp->t_rttlow || tp->t_rttlow > t)
13919 			tp->t_rttlow = t;
13920 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13921 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13922 		tcp_rack_xmit_timer_commit(rack, tp);
13923 	}
13924 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13925 		return (ret_val);
13926 	}
13927 	if (tp->t_state == TCPS_FIN_WAIT_1) {
13928 		/* We could have went to FIN_WAIT_1 (or EST) above */
13929 		/*
13930 		 * In FIN_WAIT_1 STATE in addition to the processing for the
13931 		 * ESTABLISHED state if our FIN is now acknowledged then
13932 		 * enter FIN_WAIT_2.
13933 		 */
13934 		if (ourfinisacked) {
13935 			/*
13936 			 * If we can't receive any more data, then closing
13937 			 * user can proceed. Starting the timer is contrary
13938 			 * to the specification, but if we don't get a FIN
13939 			 * we'll hang forever.
13940 			 *
13941 			 * XXXjl: we should release the tp also, and use a
13942 			 * compressed state.
13943 			 */
13944 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13945 				soisdisconnected(so);
13946 				tcp_timer_activate(tp, TT_2MSL,
13947 				    (tcp_fast_finwait2_recycle ?
13948 				    tcp_finwait2_timeout :
13949 				    TP_MAXIDLE(tp)));
13950 			}
13951 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
13952 		}
13953 	}
13954 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13955 	    tiwin, thflags, nxt_pkt));
13956 }
13957 
13958 /*
13959  * Return value of 1, the TCB is unlocked and most
13960  * likely gone, return value of 0, the TCP is still
13961  * locked.
13962  */
13963 static int
13964 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13965     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13966     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13967 {
13968 	int32_t ret_val = 0;
13969 	int32_t orig_tlen = tlen;
13970 	struct tcp_rack *rack;
13971 
13972 	/*
13973 	 * Header prediction: check for the two common cases of a
13974 	 * uni-directional data xfer.  If the packet has no control flags,
13975 	 * is in-sequence, the window didn't change and we're not
13976 	 * retransmitting, it's a candidate.  If the length is zero and the
13977 	 * ack moved forward, we're the sender side of the xfer.  Just free
13978 	 * the data acked & wake any higher level process that was blocked
13979 	 * waiting for space.  If the length is non-zero and the ack didn't
13980 	 * move, we're the receiver side.  If we're getting packets in-order
13981 	 * (the reassembly queue is empty), add the data toc The socket
13982 	 * buffer and note that we need a delayed ack. Make sure that the
13983 	 * hidden state-flags are also off. Since we check for
13984 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13985 	 */
13986 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13987 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13988 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13989 	    __predict_true(SEGQ_EMPTY(tp)) &&
13990 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
13991 		if (tlen == 0) {
13992 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13993 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13994 				return (0);
13995 			}
13996 		} else {
13997 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13998 			    tiwin, nxt_pkt, iptos)) {
13999 				return (0);
14000 			}
14001 		}
14002 	}
14003 	ctf_calc_rwin(so, tp);
14004 
14005 	if ((thflags & TH_RST) ||
14006 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14007 		return (__ctf_process_rst(m, th, so, tp,
14008 					  &rack->r_ctl.challenge_ack_ts,
14009 					  &rack->r_ctl.challenge_ack_cnt));
14010 
14011 	/*
14012 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14013 	 * synchronized state.
14014 	 */
14015 	if (thflags & TH_SYN) {
14016 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14017 		return (ret_val);
14018 	}
14019 	/*
14020 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14021 	 * it's less than ts_recent, drop it.
14022 	 */
14023 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14024 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14025 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14026 			return (ret_val);
14027 	}
14028 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14029 			      &rack->r_ctl.challenge_ack_ts,
14030 			      &rack->r_ctl.challenge_ack_cnt)) {
14031 		return (ret_val);
14032 	}
14033 	/*
14034 	 * If last ACK falls within this segment's sequence numbers, record
14035 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14036 	 * from the latest proposal of the tcplw@cray.com list (Braden
14037 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14038 	 * with our earlier PAWS tests, so this check should be solely
14039 	 * predicated on the sequence space of this segment. 3) That we
14040 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14041 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14042 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14043 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14044 	 * p.869. In such cases, we can still calculate the RTT correctly
14045 	 * when RCV.NXT == Last.ACK.Sent.
14046 	 */
14047 	if ((to->to_flags & TOF_TS) != 0 &&
14048 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14049 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14050 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14051 		tp->ts_recent_age = tcp_ts_getticks();
14052 		tp->ts_recent = to->to_tsval;
14053 	}
14054 	/*
14055 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14056 	 * is on (half-synchronized state), then queue data for later
14057 	 * processing; else drop segment and return.
14058 	 */
14059 	if ((thflags & TH_ACK) == 0) {
14060 		if (tp->t_flags & TF_NEEDSYN) {
14061 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14062 			    tiwin, thflags, nxt_pkt));
14063 
14064 		} else if (tp->t_flags & TF_ACKNOW) {
14065 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14066 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14067 			return (ret_val);
14068 		} else {
14069 			ctf_do_drop(m, NULL);
14070 			return (0);
14071 		}
14072 	}
14073 	/*
14074 	 * Ack processing.
14075 	 */
14076 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
14077 		return (ret_val);
14078 	}
14079 	if (sbavail(&so->so_snd)) {
14080 		if (ctf_progress_timeout_check(tp, true)) {
14081 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
14082 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14083 			return (1);
14084 		}
14085 	}
14086 	/* State changes only happen in rack_process_data() */
14087 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14088 	    tiwin, thflags, nxt_pkt));
14089 }
14090 
14091 /*
14092  * Return value of 1, the TCB is unlocked and most
14093  * likely gone, return value of 0, the TCP is still
14094  * locked.
14095  */
14096 static int
14097 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
14098     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14099     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14100 {
14101 	int32_t ret_val = 0;
14102 	int32_t orig_tlen = tlen;
14103 	struct tcp_rack *rack;
14104 
14105 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14106 	ctf_calc_rwin(so, tp);
14107 	if ((thflags & TH_RST) ||
14108 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14109 		return (__ctf_process_rst(m, th, so, tp,
14110 					  &rack->r_ctl.challenge_ack_ts,
14111 					  &rack->r_ctl.challenge_ack_cnt));
14112 	/*
14113 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14114 	 * synchronized state.
14115 	 */
14116 	if (thflags & TH_SYN) {
14117 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14118 		return (ret_val);
14119 	}
14120 	/*
14121 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14122 	 * it's less than ts_recent, drop it.
14123 	 */
14124 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14125 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14126 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14127 			return (ret_val);
14128 	}
14129 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14130 			      &rack->r_ctl.challenge_ack_ts,
14131 			      &rack->r_ctl.challenge_ack_cnt)) {
14132 		return (ret_val);
14133 	}
14134 	/*
14135 	 * If last ACK falls within this segment's sequence numbers, record
14136 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14137 	 * from the latest proposal of the tcplw@cray.com list (Braden
14138 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14139 	 * with our earlier PAWS tests, so this check should be solely
14140 	 * predicated on the sequence space of this segment. 3) That we
14141 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14142 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14143 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14144 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14145 	 * p.869. In such cases, we can still calculate the RTT correctly
14146 	 * when RCV.NXT == Last.ACK.Sent.
14147 	 */
14148 	if ((to->to_flags & TOF_TS) != 0 &&
14149 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14150 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14151 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14152 		tp->ts_recent_age = tcp_ts_getticks();
14153 		tp->ts_recent = to->to_tsval;
14154 	}
14155 	/*
14156 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14157 	 * is on (half-synchronized state), then queue data for later
14158 	 * processing; else drop segment and return.
14159 	 */
14160 	if ((thflags & TH_ACK) == 0) {
14161 		if (tp->t_flags & TF_NEEDSYN) {
14162 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14163 			    tiwin, thflags, nxt_pkt));
14164 
14165 		} else if (tp->t_flags & TF_ACKNOW) {
14166 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14167 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14168 			return (ret_val);
14169 		} else {
14170 			ctf_do_drop(m, NULL);
14171 			return (0);
14172 		}
14173 	}
14174 	/*
14175 	 * Ack processing.
14176 	 */
14177 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
14178 		return (ret_val);
14179 	}
14180 	if (sbavail(&so->so_snd)) {
14181 		if (ctf_progress_timeout_check(tp, true)) {
14182 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14183 						tp, tick, PROGRESS_DROP, __LINE__);
14184 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14185 			return (1);
14186 		}
14187 	}
14188 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14189 	    tiwin, thflags, nxt_pkt));
14190 }
14191 
14192 static int
14193 rack_check_data_after_close(struct mbuf *m,
14194     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
14195 {
14196 	struct tcp_rack *rack;
14197 
14198 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14199 	if (rack->rc_allow_data_af_clo == 0) {
14200 	close_now:
14201 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
14202 		/* tcp_close will kill the inp pre-log the Reset */
14203 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
14204 		tp = tcp_close(tp);
14205 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
14206 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
14207 		return (1);
14208 	}
14209 	if (sbavail(&so->so_snd) == 0)
14210 		goto close_now;
14211 	/* Ok we allow data that is ignored and a followup reset */
14212 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
14213 	tp->rcv_nxt = th->th_seq + *tlen;
14214 	tp->t_flags2 |= TF2_DROP_AF_DATA;
14215 	rack->r_wanted_output = 1;
14216 	*tlen = 0;
14217 	return (0);
14218 }
14219 
14220 /*
14221  * Return value of 1, the TCB is unlocked and most
14222  * likely gone, return value of 0, the TCP is still
14223  * locked.
14224  */
14225 static int
14226 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
14227     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14228     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14229 {
14230 	int32_t ret_val = 0;
14231 	int32_t orig_tlen = tlen;
14232 	int32_t ourfinisacked = 0;
14233 	struct tcp_rack *rack;
14234 
14235 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14236 	ctf_calc_rwin(so, tp);
14237 
14238 	if ((thflags & TH_RST) ||
14239 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14240 		return (__ctf_process_rst(m, th, so, tp,
14241 					  &rack->r_ctl.challenge_ack_ts,
14242 					  &rack->r_ctl.challenge_ack_cnt));
14243 	/*
14244 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14245 	 * synchronized state.
14246 	 */
14247 	if (thflags & TH_SYN) {
14248 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14249 		return (ret_val);
14250 	}
14251 	/*
14252 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14253 	 * it's less than ts_recent, drop it.
14254 	 */
14255 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14256 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14257 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14258 			return (ret_val);
14259 	}
14260 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14261 			      &rack->r_ctl.challenge_ack_ts,
14262 			      &rack->r_ctl.challenge_ack_cnt)) {
14263 		return (ret_val);
14264 	}
14265 	/*
14266 	 * If new data are received on a connection after the user processes
14267 	 * are gone, then RST the other end.
14268 	 */
14269 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14270 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14271 		return (1);
14272 	/*
14273 	 * If last ACK falls within this segment's sequence numbers, record
14274 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14275 	 * from the latest proposal of the tcplw@cray.com list (Braden
14276 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14277 	 * with our earlier PAWS tests, so this check should be solely
14278 	 * predicated on the sequence space of this segment. 3) That we
14279 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14280 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14281 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14282 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14283 	 * p.869. In such cases, we can still calculate the RTT correctly
14284 	 * when RCV.NXT == Last.ACK.Sent.
14285 	 */
14286 	if ((to->to_flags & TOF_TS) != 0 &&
14287 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14288 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14289 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14290 		tp->ts_recent_age = tcp_ts_getticks();
14291 		tp->ts_recent = to->to_tsval;
14292 	}
14293 	/*
14294 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14295 	 * is on (half-synchronized state), then queue data for later
14296 	 * processing; else drop segment and return.
14297 	 */
14298 	if ((thflags & TH_ACK) == 0) {
14299 		if (tp->t_flags & TF_NEEDSYN) {
14300 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14301 			    tiwin, thflags, nxt_pkt));
14302 		} else if (tp->t_flags & TF_ACKNOW) {
14303 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14304 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14305 			return (ret_val);
14306 		} else {
14307 			ctf_do_drop(m, NULL);
14308 			return (0);
14309 		}
14310 	}
14311 	/*
14312 	 * Ack processing.
14313 	 */
14314 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
14315 		return (ret_val);
14316 	}
14317 	if (ourfinisacked) {
14318 		/*
14319 		 * If we can't receive any more data, then closing user can
14320 		 * proceed. Starting the timer is contrary to the
14321 		 * specification, but if we don't get a FIN we'll hang
14322 		 * forever.
14323 		 *
14324 		 * XXXjl: we should release the tp also, and use a
14325 		 * compressed state.
14326 		 */
14327 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
14328 			soisdisconnected(so);
14329 			tcp_timer_activate(tp, TT_2MSL,
14330 			    (tcp_fast_finwait2_recycle ?
14331 			    tcp_finwait2_timeout :
14332 			    TP_MAXIDLE(tp)));
14333 		}
14334 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
14335 	}
14336 	if (sbavail(&so->so_snd)) {
14337 		if (ctf_progress_timeout_check(tp, true)) {
14338 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14339 						tp, tick, PROGRESS_DROP, __LINE__);
14340 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14341 			return (1);
14342 		}
14343 	}
14344 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14345 	    tiwin, thflags, nxt_pkt));
14346 }
14347 
14348 /*
14349  * Return value of 1, the TCB is unlocked and most
14350  * likely gone, return value of 0, the TCP is still
14351  * locked.
14352  */
14353 static int
14354 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
14355     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14356     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14357 {
14358 	int32_t ret_val = 0;
14359 	int32_t orig_tlen = tlen;
14360 	int32_t ourfinisacked = 0;
14361 	struct tcp_rack *rack;
14362 
14363 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14364 	ctf_calc_rwin(so, tp);
14365 
14366 	if ((thflags & TH_RST) ||
14367 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14368 		return (__ctf_process_rst(m, th, so, tp,
14369 					  &rack->r_ctl.challenge_ack_ts,
14370 					  &rack->r_ctl.challenge_ack_cnt));
14371 	/*
14372 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14373 	 * synchronized state.
14374 	 */
14375 	if (thflags & TH_SYN) {
14376 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14377 		return (ret_val);
14378 	}
14379 	/*
14380 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14381 	 * it's less than ts_recent, drop it.
14382 	 */
14383 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14384 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14385 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14386 			return (ret_val);
14387 	}
14388 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14389 			      &rack->r_ctl.challenge_ack_ts,
14390 			      &rack->r_ctl.challenge_ack_cnt)) {
14391 		return (ret_val);
14392 	}
14393 	/*
14394 	 * If last ACK falls within this segment's sequence numbers, record
14395 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14396 	 * from the latest proposal of the tcplw@cray.com list (Braden
14397 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14398 	 * with our earlier PAWS tests, so this check should be solely
14399 	 * predicated on the sequence space of this segment. 3) That we
14400 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14401 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14402 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14403 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14404 	 * p.869. In such cases, we can still calculate the RTT correctly
14405 	 * when RCV.NXT == Last.ACK.Sent.
14406 	 */
14407 	if ((to->to_flags & TOF_TS) != 0 &&
14408 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14409 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14410 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14411 		tp->ts_recent_age = tcp_ts_getticks();
14412 		tp->ts_recent = to->to_tsval;
14413 	}
14414 	/*
14415 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14416 	 * is on (half-synchronized state), then queue data for later
14417 	 * processing; else drop segment and return.
14418 	 */
14419 	if ((thflags & TH_ACK) == 0) {
14420 		if (tp->t_flags & TF_NEEDSYN) {
14421 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14422 			    tiwin, thflags, nxt_pkt));
14423 		} else if (tp->t_flags & TF_ACKNOW) {
14424 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14425 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14426 			return (ret_val);
14427 		} else {
14428 			ctf_do_drop(m, NULL);
14429 			return (0);
14430 		}
14431 	}
14432 	/*
14433 	 * Ack processing.
14434 	 */
14435 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
14436 		return (ret_val);
14437 	}
14438 	if (ourfinisacked) {
14439 		tcp_twstart(tp);
14440 		m_freem(m);
14441 		return (1);
14442 	}
14443 	if (sbavail(&so->so_snd)) {
14444 		if (ctf_progress_timeout_check(tp, true)) {
14445 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14446 						tp, tick, PROGRESS_DROP, __LINE__);
14447 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14448 			return (1);
14449 		}
14450 	}
14451 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14452 	    tiwin, thflags, nxt_pkt));
14453 }
14454 
14455 /*
14456  * Return value of 1, the TCB is unlocked and most
14457  * likely gone, return value of 0, the TCP is still
14458  * locked.
14459  */
14460 static int
14461 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
14462     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14463     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14464 {
14465 	int32_t ret_val = 0;
14466 	int32_t orig_tlen;
14467 	int32_t ourfinisacked = 0;
14468 	struct tcp_rack *rack;
14469 
14470 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14471 	ctf_calc_rwin(so, tp);
14472 
14473 	if ((thflags & TH_RST) ||
14474 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14475 		return (__ctf_process_rst(m, th, so, tp,
14476 					  &rack->r_ctl.challenge_ack_ts,
14477 					  &rack->r_ctl.challenge_ack_cnt));
14478 	/*
14479 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14480 	 * synchronized state.
14481 	 */
14482 	if (thflags & TH_SYN) {
14483 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14484 		return (ret_val);
14485 	}
14486 	/*
14487 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14488 	 * it's less than ts_recent, drop it.
14489 	 */
14490 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14491 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14492 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14493 			return (ret_val);
14494 	}
14495 	orig_tlen = tlen;
14496 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14497 			      &rack->r_ctl.challenge_ack_ts,
14498 			      &rack->r_ctl.challenge_ack_cnt)) {
14499 		return (ret_val);
14500 	}
14501 	/*
14502 	 * If last ACK falls within this segment's sequence numbers, record
14503 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14504 	 * from the latest proposal of the tcplw@cray.com list (Braden
14505 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14506 	 * with our earlier PAWS tests, so this check should be solely
14507 	 * predicated on the sequence space of this segment. 3) That we
14508 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14509 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14510 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14511 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14512 	 * p.869. In such cases, we can still calculate the RTT correctly
14513 	 * when RCV.NXT == Last.ACK.Sent.
14514 	 */
14515 	if ((to->to_flags & TOF_TS) != 0 &&
14516 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14517 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14518 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14519 		tp->ts_recent_age = tcp_ts_getticks();
14520 		tp->ts_recent = to->to_tsval;
14521 	}
14522 	/*
14523 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14524 	 * is on (half-synchronized state), then queue data for later
14525 	 * processing; else drop segment and return.
14526 	 */
14527 	if ((thflags & TH_ACK) == 0) {
14528 		if (tp->t_flags & TF_NEEDSYN) {
14529 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14530 			    tiwin, thflags, nxt_pkt));
14531 		} else if (tp->t_flags & TF_ACKNOW) {
14532 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14533 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14534 			return (ret_val);
14535 		} else {
14536 			ctf_do_drop(m, NULL);
14537 			return (0);
14538 		}
14539 	}
14540 	/*
14541 	 * case TCPS_LAST_ACK: Ack processing.
14542 	 */
14543 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
14544 		return (ret_val);
14545 	}
14546 	if (ourfinisacked) {
14547 		tp = tcp_close(tp);
14548 		ctf_do_drop(m, tp);
14549 		return (1);
14550 	}
14551 	if (sbavail(&so->so_snd)) {
14552 		if (ctf_progress_timeout_check(tp, true)) {
14553 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14554 						tp, tick, PROGRESS_DROP, __LINE__);
14555 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14556 			return (1);
14557 		}
14558 	}
14559 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14560 	    tiwin, thflags, nxt_pkt));
14561 }
14562 
14563 /*
14564  * Return value of 1, the TCB is unlocked and most
14565  * likely gone, return value of 0, the TCP is still
14566  * locked.
14567  */
14568 static int
14569 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
14570     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14571     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14572 {
14573 	int32_t ret_val = 0;
14574 	int32_t orig_tlen = tlen;
14575 	int32_t ourfinisacked = 0;
14576 	struct tcp_rack *rack;
14577 
14578 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14579 	ctf_calc_rwin(so, tp);
14580 
14581 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
14582 	if ((thflags & TH_RST) ||
14583 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14584 		return (__ctf_process_rst(m, th, so, tp,
14585 					  &rack->r_ctl.challenge_ack_ts,
14586 					  &rack->r_ctl.challenge_ack_cnt));
14587 	/*
14588 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14589 	 * synchronized state.
14590 	 */
14591 	if (thflags & TH_SYN) {
14592 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14593 		return (ret_val);
14594 	}
14595 	/*
14596 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14597 	 * it's less than ts_recent, drop it.
14598 	 */
14599 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14600 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14601 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14602 			return (ret_val);
14603 	}
14604 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14605 			      &rack->r_ctl.challenge_ack_ts,
14606 			      &rack->r_ctl.challenge_ack_cnt)) {
14607 		return (ret_val);
14608 	}
14609 	/*
14610 	 * If new data are received on a connection after the user processes
14611 	 * are gone, then RST the other end.
14612 	 */
14613 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14614 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14615 		return (1);
14616 	/*
14617 	 * If last ACK falls within this segment's sequence numbers, record
14618 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14619 	 * from the latest proposal of the tcplw@cray.com list (Braden
14620 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14621 	 * with our earlier PAWS tests, so this check should be solely
14622 	 * predicated on the sequence space of this segment. 3) That we
14623 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14624 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14625 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14626 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14627 	 * p.869. In such cases, we can still calculate the RTT correctly
14628 	 * when RCV.NXT == Last.ACK.Sent.
14629 	 */
14630 	if ((to->to_flags & TOF_TS) != 0 &&
14631 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14632 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14633 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14634 		tp->ts_recent_age = tcp_ts_getticks();
14635 		tp->ts_recent = to->to_tsval;
14636 	}
14637 	/*
14638 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14639 	 * is on (half-synchronized state), then queue data for later
14640 	 * processing; else drop segment and return.
14641 	 */
14642 	if ((thflags & TH_ACK) == 0) {
14643 		if (tp->t_flags & TF_NEEDSYN) {
14644 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14645 			    tiwin, thflags, nxt_pkt));
14646 		} else if (tp->t_flags & TF_ACKNOW) {
14647 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14648 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14649 			return (ret_val);
14650 		} else {
14651 			ctf_do_drop(m, NULL);
14652 			return (0);
14653 		}
14654 	}
14655 	/*
14656 	 * Ack processing.
14657 	 */
14658 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
14659 		return (ret_val);
14660 	}
14661 	if (sbavail(&so->so_snd)) {
14662 		if (ctf_progress_timeout_check(tp, true)) {
14663 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14664 						tp, tick, PROGRESS_DROP, __LINE__);
14665 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14666 			return (1);
14667 		}
14668 	}
14669 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14670 	    tiwin, thflags, nxt_pkt));
14671 }
14672 
14673 static void inline
14674 rack_clear_rate_sample(struct tcp_rack *rack)
14675 {
14676 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
14677 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
14678 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
14679 }
14680 
14681 static void
14682 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
14683 {
14684 	uint64_t bw_est, rate_wanted;
14685 	int chged = 0;
14686 	uint32_t user_max, orig_min, orig_max;
14687 
14688 #ifdef TCP_REQUEST_TRK
14689 	if (rack->rc_hybrid_mode &&
14690 	    (rack->r_ctl.rc_pace_max_segs != 0) &&
14691 	    (rack_hybrid_allow_set_maxseg == 1) &&
14692 	    (rack->r_ctl.rc_last_sft != NULL)) {
14693 		rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
14694 		return;
14695 	}
14696 #endif
14697 	orig_min = rack->r_ctl.rc_pace_min_segs;
14698 	orig_max = rack->r_ctl.rc_pace_max_segs;
14699 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
14700 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
14701 		chged = 1;
14702 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
14703 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
14704 		if (user_max != rack->r_ctl.rc_pace_max_segs)
14705 			chged = 1;
14706 	}
14707 	if (rack->rc_force_max_seg) {
14708 		rack->r_ctl.rc_pace_max_segs = user_max;
14709 	} else if (rack->use_fixed_rate) {
14710 		bw_est = rack_get_bw(rack);
14711 		if ((rack->r_ctl.crte == NULL) ||
14712 		    (bw_est != rack->r_ctl.crte->rate)) {
14713 			rack->r_ctl.rc_pace_max_segs = user_max;
14714 		} else {
14715 			/* We are pacing right at the hardware rate */
14716 			uint32_t segsiz, pace_one;
14717 
14718 			if (rack_pace_one_seg ||
14719 			    (rack->r_ctl.rc_user_set_min_segs == 1))
14720 				pace_one = 1;
14721 			else
14722 				pace_one = 0;
14723 			segsiz = min(ctf_fixed_maxseg(tp),
14724 				     rack->r_ctl.rc_pace_min_segs);
14725 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14726 				tp, bw_est, segsiz, pace_one,
14727 				rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14728 		}
14729 	} else if (rack->rc_always_pace) {
14730 		if (rack->r_ctl.gp_bw ||
14731 		    rack->r_ctl.init_rate) {
14732 			/* We have a rate of some sort set */
14733 			uint32_t  orig;
14734 
14735 			bw_est = rack_get_bw(rack);
14736 			orig = rack->r_ctl.rc_pace_max_segs;
14737 			if (fill_override)
14738 				rate_wanted = *fill_override;
14739 			else
14740 				rate_wanted = rack_get_gp_est(rack);
14741 			if (rate_wanted) {
14742 				/* We have something */
14743 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14744 										   rate_wanted,
14745 										   ctf_fixed_maxseg(rack->rc_tp));
14746 			} else
14747 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14748 			if (orig != rack->r_ctl.rc_pace_max_segs)
14749 				chged = 1;
14750 		} else if ((rack->r_ctl.gp_bw == 0) &&
14751 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
14752 			/*
14753 			 * If we have nothing limit us to bursting
14754 			 * out IW sized pieces.
14755 			 */
14756 			chged = 1;
14757 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14758 		}
14759 	}
14760 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14761 		chged = 1;
14762 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14763 	}
14764 	if (chged)
14765 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14766 }
14767 
14768 
14769 static void
14770 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14771 {
14772 #ifdef INET6
14773 	struct ip6_hdr *ip6 = NULL;
14774 #endif
14775 #ifdef INET
14776 	struct ip *ip = NULL;
14777 #endif
14778 	struct udphdr *udp = NULL;
14779 
14780 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
14781 #ifdef INET6
14782 	if (rack->r_is_v6) {
14783 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14784 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14785 		if (tp->t_port) {
14786 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14787 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14788 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14789 			udp->uh_dport = tp->t_port;
14790 			rack->r_ctl.fsb.udp = udp;
14791 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14792 		} else
14793 		{
14794 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14795 			rack->r_ctl.fsb.udp = NULL;
14796 		}
14797 		tcpip_fillheaders(rack->rc_inp,
14798 				  tp->t_port,
14799 				  ip6, rack->r_ctl.fsb.th);
14800 		rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14801 	} else
14802 #endif				/* INET6 */
14803 #ifdef INET
14804 	{
14805 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14806 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14807 		if (tp->t_port) {
14808 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14809 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14810 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14811 			udp->uh_dport = tp->t_port;
14812 			rack->r_ctl.fsb.udp = udp;
14813 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14814 		} else
14815 		{
14816 			rack->r_ctl.fsb.udp = NULL;
14817 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14818 		}
14819 		tcpip_fillheaders(rack->rc_inp,
14820 				  tp->t_port,
14821 				  ip, rack->r_ctl.fsb.th);
14822 		rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14823 	}
14824 #endif
14825 	rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14826 	    (long)TCP_MAXWIN << tp->rcv_scale);
14827 	rack->r_fsb_inited = 1;
14828 }
14829 
14830 static int
14831 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14832 {
14833 	/*
14834 	 * Allocate the larger of spaces V6 if available else just
14835 	 * V4 and include udphdr (overbook)
14836 	 */
14837 #ifdef INET6
14838 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14839 #else
14840 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14841 #endif
14842 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14843 					    M_TCPFSB, M_NOWAIT|M_ZERO);
14844 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14845 		return (ENOMEM);
14846 	}
14847 	rack->r_fsb_inited = 0;
14848 	return (0);
14849 }
14850 
14851 static void
14852 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14853 {
14854 	/*
14855 	 * Types of logs (mod value)
14856 	 * 20 - Initial round setup
14857 	 * 21 - Rack declares a new round.
14858 	 */
14859 	struct tcpcb *tp;
14860 
14861 	tp = rack->rc_tp;
14862 	if (tcp_bblogging_on(tp)) {
14863 		union tcp_log_stackspecific log;
14864 		struct timeval tv;
14865 
14866 		memset(&log, 0, sizeof(log));
14867 		log.u_bbr.flex1 = rack->r_ctl.current_round;
14868 		log.u_bbr.flex2 = rack->r_ctl.roundends;
14869 		log.u_bbr.flex3 = high_seq;
14870 		log.u_bbr.flex4 = tp->snd_max;
14871 		log.u_bbr.flex8 = mod;
14872 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14873 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14874 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14875 		TCP_LOG_EVENTP(tp, NULL,
14876 		    &tptosocket(tp)->so_rcv,
14877 		    &tptosocket(tp)->so_snd,
14878 		    TCP_HYSTART, 0,
14879 		    0, &log, false, &tv);
14880 	}
14881 }
14882 
14883 static void
14884 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14885 {
14886 	rack->rack_deferred_inited = 1;
14887 	rack->r_ctl.roundends = tp->snd_max;
14888 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14889 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14890 }
14891 
14892 static void
14893 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14894 {
14895 	/* Retransmit bit controls.
14896 	 *
14897 	 * The setting of these values control one of
14898 	 * three settings you can have and dictate
14899 	 * how rack does retransmissions. Note this
14900 	 * is in *any* mode i.e. pacing on or off DGP
14901 	 * fixed rate pacing, or just bursting rack.
14902 	 *
14903 	 * 1 - Use full sized retransmits i.e. limit
14904 	 *     the size to whatever the pace_max_segments
14905 	 *     size is.
14906 	 *
14907 	 * 2 - Use pacer min granularity as a guide to
14908 	 *     the size combined with the current calculated
14909 	 *     goodput b/w measurement. So for example if
14910 	 *     the goodput is measured at 20Mbps we would
14911 	 *     calculate 8125 (pacer minimum 250usec in
14912 	 *     that b/w) and then round it up to the next
14913 	 *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14914 	 *
14915 	 * 0 - The rack default 1 MSS (anything not 0/1/2
14916 	 *     fall here too if we are setting via rack_init()).
14917 	 *
14918 	 */
14919 	if (ctl == 1) {
14920 		rack->full_size_rxt = 1;
14921 		rack->shape_rxt_to_pacing_min  = 0;
14922 	} else if (ctl == 2) {
14923 		rack->full_size_rxt = 0;
14924 		rack->shape_rxt_to_pacing_min  = 1;
14925 	} else {
14926 		rack->full_size_rxt = 0;
14927 		rack->shape_rxt_to_pacing_min  = 0;
14928 	}
14929 }
14930 
14931 static void
14932 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14933 		  uint32_t flex1,
14934 		  uint32_t flex2,
14935 		  uint32_t flex3)
14936 {
14937 	if (tcp_bblogging_on(rack->rc_tp)) {
14938 		union tcp_log_stackspecific log;
14939 		struct timeval tv;
14940 
14941 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14942 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14943 		log.u_bbr.flex8 = mod;
14944 		log.u_bbr.flex1 = flex1;
14945 		log.u_bbr.flex2 = flex2;
14946 		log.u_bbr.flex3 = flex3;
14947 		tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14948 			       0, &log, false, NULL, __func__, __LINE__, &tv);
14949 	}
14950 }
14951 
14952 static int
14953 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14954 {
14955 	struct tcp_rack *rack;
14956 	struct rack_sendmap *rsm;
14957 	int i;
14958 
14959 
14960 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14961 	switch (reqr->req) {
14962 	case TCP_QUERY_SENDMAP:
14963 		if ((reqr->req_param == tp->snd_max) ||
14964 		    (tp->snd_max == tp->snd_una)){
14965 			/* Unlikely */
14966 			return (0);
14967 		}
14968 		rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14969 		if (rsm == NULL) {
14970 			/* Can't find that seq -- unlikely */
14971 			return (0);
14972 		}
14973 		reqr->sendmap_start = rsm->r_start;
14974 		reqr->sendmap_end = rsm->r_end;
14975 		reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14976 		reqr->sendmap_fas = rsm->r_fas;
14977 		if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14978 			reqr->sendmap_send_cnt = SNDMAP_NRTX;
14979 		for(i=0; i<reqr->sendmap_send_cnt; i++)
14980 			reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14981 		reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14982 		reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14983 		reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14984 		reqr->sendmap_dupacks = rsm->r_dupack;
14985 		rack_log_chg_info(tp, rack, 1,
14986 				  rsm->r_start,
14987 				  rsm->r_end,
14988 				  rsm->r_flags);
14989 		return(1);
14990 		break;
14991 	case TCP_QUERY_TIMERS_UP:
14992 		if (rack->r_ctl.rc_hpts_flags == 0) {
14993 			/* no timers up */
14994 			return (0);
14995 		}
14996 		reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14997 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14998 			reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14999 		}
15000 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
15001 			reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
15002 		}
15003 		rack_log_chg_info(tp, rack, 2,
15004 				  rack->r_ctl.rc_hpts_flags,
15005 				  rack->r_ctl.rc_last_output_to,
15006 				  rack->r_ctl.rc_timer_exp);
15007 		return (1);
15008 		break;
15009 	case TCP_QUERY_RACK_TIMES:
15010 		/* Reordering items */
15011 		reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
15012 		reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
15013 		/* Timerstamps and timers */
15014 		reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
15015 		reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
15016 		reqr->rack_rtt = rack->rc_rack_rtt;
15017 		reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
15018 		reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
15019 		/* PRR data */
15020 		reqr->rack_sacked = rack->r_ctl.rc_sacked;
15021 		reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
15022 		reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
15023 		reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
15024 		reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
15025 		reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
15026 		/* TLP and persists info */
15027 		reqr->rack_tlp_out = rack->rc_tlp_in_progress;
15028 		reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
15029 		if (rack->rc_in_persist) {
15030 			reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
15031 			reqr->rack_in_persist = 1;
15032 		} else {
15033 			reqr->rack_time_went_idle = 0;
15034 			reqr->rack_in_persist = 0;
15035 		}
15036 		if (rack->r_wanted_output)
15037 			reqr->rack_wanted_output = 1;
15038 		else
15039 			reqr->rack_wanted_output = 0;
15040 		return (1);
15041 		break;
15042 	default:
15043 		return (-EINVAL);
15044 	}
15045 }
15046 
15047 static void
15048 rack_switch_failed(struct tcpcb *tp)
15049 {
15050 	/*
15051 	 * This method gets called if a stack switch was
15052 	 * attempted and it failed. We are left
15053 	 * but our hpts timers were stopped and we
15054 	 * need to validate time units and t_flags2.
15055 	 */
15056 	struct tcp_rack *rack;
15057 	struct timeval tv;
15058 	uint32_t cts;
15059 	uint32_t toval;
15060 	struct hpts_diag diag;
15061 
15062 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15063 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
15064 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
15065 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
15066 	else
15067 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
15068 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15069 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
15070 	if (tp->t_in_hpts > IHPTS_NONE) {
15071 		/* Strange */
15072 		return;
15073 	}
15074 	cts = tcp_get_usecs(&tv);
15075 	if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15076 		if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
15077 			toval = rack->r_ctl.rc_last_output_to - cts;
15078 		} else {
15079 			/* one slot please */
15080 			toval = HPTS_TICKS_PER_SLOT;
15081 		}
15082 	} else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
15083 		if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
15084 			toval = rack->r_ctl.rc_timer_exp - cts;
15085 		} else {
15086 			/* one slot please */
15087 			toval = HPTS_TICKS_PER_SLOT;
15088 		}
15089 	} else
15090 		toval = HPTS_TICKS_PER_SLOT;
15091 	(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
15092 				   __LINE__, &diag);
15093 	rack_log_hpts_diag(rack, cts, &diag, &tv);
15094 }
15095 
15096 static int
15097 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
15098 {
15099 	struct rack_sendmap *rsm, *ersm;
15100 	int insret __diagused;
15101 	/*
15102 	 * When initing outstanding, we must be quite careful
15103 	 * to not refer to tp->t_fb_ptr. This has the old rack
15104 	 * pointer in it, not the "new" one (when we are doing
15105 	 * a stack switch).
15106 	 */
15107 
15108 
15109 	if (tp->t_fb->tfb_chg_query == NULL) {
15110 		/* Create a send map for the current outstanding data */
15111 
15112 		rsm = rack_alloc(rack);
15113 		if (rsm == NULL) {
15114 			uma_zfree(rack_pcb_zone, ptr);
15115 			return (ENOMEM);
15116 		}
15117 		rsm->r_no_rtt_allowed = 1;
15118 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
15119 		rsm->r_rtr_cnt = 1;
15120 		rsm->r_rtr_bytes = 0;
15121 		if (tp->t_flags & TF_SENTFIN)
15122 			rsm->r_flags |= RACK_HAS_FIN;
15123 		rsm->r_end = tp->snd_max;
15124 		if (tp->snd_una == tp->iss) {
15125 			/* The data space is one beyond snd_una */
15126 			rsm->r_flags |= RACK_HAS_SYN;
15127 			rsm->r_start = tp->iss;
15128 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
15129 		} else
15130 			rsm->r_start = tp->snd_una;
15131 		rsm->r_dupack = 0;
15132 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
15133 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
15134 			if (rsm->m) {
15135 				rsm->orig_m_len = rsm->m->m_len;
15136 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
15137 			} else {
15138 				rsm->orig_m_len = 0;
15139 				rsm->orig_t_space = 0;
15140 			}
15141 		} else {
15142 			/*
15143 			 * This can happen if we have a stand-alone FIN or
15144 			 *  SYN.
15145 			 */
15146 			rsm->m = NULL;
15147 			rsm->orig_m_len = 0;
15148 			rsm->orig_t_space = 0;
15149 			rsm->soff = 0;
15150 		}
15151 #ifdef INVARIANTS
15152 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
15153 			panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
15154 			      insret, rack, rsm);
15155 		}
15156 #else
15157 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
15158 #endif
15159 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
15160 		rsm->r_in_tmap = 1;
15161 	} else {
15162 		/* We have a query mechanism, lets use it */
15163 		struct tcp_query_resp qr;
15164 		int i;
15165 		tcp_seq at;
15166 
15167 		at = tp->snd_una;
15168 		while (at != tp->snd_max) {
15169 			memset(&qr, 0, sizeof(qr));
15170 			qr.req = TCP_QUERY_SENDMAP;
15171 			qr.req_param = at;
15172 			if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
15173 				break;
15174 			/* Move forward */
15175 			at = qr.sendmap_end;
15176 			/* Now lets build the entry for this one */
15177 			rsm = rack_alloc(rack);
15178 			if (rsm == NULL) {
15179 				uma_zfree(rack_pcb_zone, ptr);
15180 				return (ENOMEM);
15181 			}
15182 			memset(rsm, 0, sizeof(struct rack_sendmap));
15183 			/* Now configure the rsm and insert it */
15184 			rsm->r_dupack = qr.sendmap_dupacks;
15185 			rsm->r_start = qr.sendmap_start;
15186 			rsm->r_end = qr.sendmap_end;
15187 			if (qr.sendmap_fas)
15188 				rsm->r_fas = qr.sendmap_end;
15189 			else
15190 				rsm->r_fas = rsm->r_start - tp->snd_una;
15191 			/*
15192 			 * We have carefully aligned the bits
15193 			 * so that all we have to do is copy over
15194 			 * the bits with the mask.
15195 			 */
15196 			rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
15197 			rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
15198 			rsm->r_rtr_cnt = qr.sendmap_send_cnt;
15199 			rsm->r_ack_arrival = qr.sendmap_ack_arrival;
15200 			for (i=0 ; i<rsm->r_rtr_cnt; i++)
15201 				rsm->r_tim_lastsent[i]	= qr.sendmap_time[i];
15202 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
15203 					   (rsm->r_start - tp->snd_una), &rsm->soff);
15204 			if (rsm->m) {
15205 				rsm->orig_m_len = rsm->m->m_len;
15206 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
15207 			} else {
15208 				rsm->orig_m_len = 0;
15209 				rsm->orig_t_space = 0;
15210 			}
15211 #ifdef INVARIANTS
15212 			if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
15213 				panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
15214 				      insret, rack, rsm);
15215 			}
15216 #else
15217 			(void)tqhash_insert(rack->r_ctl.tqh, rsm);
15218 #endif
15219 			if ((rsm->r_flags & RACK_ACKED) == 0)  {
15220 				TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
15221 					if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
15222 					    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
15223 						/*
15224 						 * If the existing ersm was sent at
15225 						 * a later time than the new one, then
15226 						 * the new one should appear ahead of this
15227 						 * ersm.
15228 						 */
15229 						rsm->r_in_tmap = 1;
15230 						TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
15231 						break;
15232 					}
15233 				}
15234 				if (rsm->r_in_tmap == 0) {
15235 					/*
15236 					 * Not found so shove it on the tail.
15237 					 */
15238 					TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
15239 					rsm->r_in_tmap = 1;
15240 				}
15241  			} else {
15242 				if ((rack->r_ctl.rc_sacklast == NULL) ||
15243 				    (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
15244 					rack->r_ctl.rc_sacklast = rsm;
15245 				}
15246 			}
15247 			rack_log_chg_info(tp, rack, 3,
15248 					  rsm->r_start,
15249 					  rsm->r_end,
15250 					  rsm->r_flags);
15251 		}
15252 	}
15253 	return (0);
15254 }
15255 
15256 static void
15257 rack_translate_policer_detect(struct tcp_rack *rack, uint32_t optval)
15258 {
15259 	/*
15260 	 * P = Percent of retransmits 499 = 49.9%
15261 	 * A = Average number 1 (.1%) -> 169 (16.9%)
15262 	 * M = Median number of retrans 1 - 16
15263 	 * MMMM MMMM AAAA AAAA PPPP PPPP PPPP PPPP
15264 	 *
15265 	 */
15266 	uint16_t per, upp;
15267 
15268 	per = optval & 0x0000ffff;
15269 	rack->r_ctl.policer_rxt_threshold = (uint32_t)(per & 0xffff);
15270 	upp = ((optval & 0xffff0000) >> 16);
15271 	rack->r_ctl.policer_avg_threshold = (0x00ff & upp);
15272 	rack->r_ctl.policer_med_threshold = ((upp >> 8) & 0x00ff);
15273 	if ((rack->r_ctl.policer_rxt_threshold > 0) &&
15274 	    (rack->r_ctl.policer_avg_threshold > 0) &&
15275 	    (rack->r_ctl.policer_med_threshold > 0)) {
15276 		rack->policer_detect_on = 1;
15277 	} else {
15278 		rack->policer_detect_on = 0;
15279 	}
15280 	rack->r_ctl.saved_policer_val = optval;
15281 	policer_detection_log(rack, optval,
15282 			      rack->r_ctl.policer_avg_threshold,
15283 			      rack->r_ctl.policer_med_threshold,
15284 			      rack->r_ctl.policer_rxt_threshold, 11);
15285 }
15286 
15287 static int32_t
15288 rack_init(struct tcpcb *tp, void **ptr)
15289 {
15290 	struct inpcb *inp = tptoinpcb(tp);
15291 	struct tcp_rack *rack = NULL;
15292 	uint32_t iwin, snt, us_cts;
15293 	size_t sz;
15294 	int err, no_query;
15295 
15296 	tcp_hpts_init(tp);
15297 
15298 	/*
15299 	 * First are we the initial or are we a switched stack?
15300 	 * If we are initing via tcp_newtcppcb the ptr passed
15301 	 * will be tp->t_fb_ptr. If its a stack switch that
15302 	 * has a previous stack we can query it will be a local
15303 	 * var that will in the end be set into t_fb_ptr.
15304 	 */
15305 	if (ptr == &tp->t_fb_ptr)
15306 		no_query = 1;
15307 	else
15308 		no_query = 0;
15309 	*ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
15310 	if (*ptr == NULL) {
15311 		/*
15312 		 * We need to allocate memory but cant. The INP and INP_INFO
15313 		 * locks and they are recursive (happens during setup. So a
15314 		 * scheme to drop the locks fails :(
15315 		 *
15316 		 */
15317 		return(ENOMEM);
15318 	}
15319 	memset(*ptr, 0, sizeof(struct tcp_rack));
15320 	rack = (struct tcp_rack *)*ptr;
15321 	rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
15322 	if (rack->r_ctl.tqh == NULL) {
15323 		uma_zfree(rack_pcb_zone, rack);
15324 		return(ENOMEM);
15325 	}
15326 	tqhash_init(rack->r_ctl.tqh);
15327 	TAILQ_INIT(&rack->r_ctl.rc_free);
15328 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
15329 	rack->rc_tp = tp;
15330 	rack->rc_inp = inp;
15331 	/* Set the flag */
15332 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
15333 	/* Probably not needed but lets be sure */
15334 	rack_clear_rate_sample(rack);
15335 	/*
15336 	 * Save off the default values, socket options will poke
15337 	 * at these if pacing is not on or we have not yet
15338 	 * reached where pacing is on (gp_ready/fixed enabled).
15339 	 * When they get set into the CC module (when gp_ready
15340 	 * is enabled or we enable fixed) then we will set these
15341 	 * values into the CC and place in here the old values
15342 	 * so we have a restoral. Then we will set the flag
15343 	 * rc_pacing_cc_set. That way whenever we turn off pacing
15344 	 * or switch off this stack, we will know to go restore
15345 	 * the saved values.
15346 	 *
15347 	 * We specifically put into the beta the ecn value for pacing.
15348 	 */
15349 	rack->rc_new_rnd_needed = 1;
15350 	rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
15351 	/* We want abe like behavior as well */
15352 
15353 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
15354 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
15355 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
15356 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
15357 	rack->r_ctl.policer_del_mss = rack_req_del_mss;
15358 	if ((rack_policer_rxt_thresh > 0) &&
15359 	    (rack_policer_avg_thresh > 0) &&
15360 	    (rack_policer_med_thresh > 0)) {
15361 		rack->r_ctl.policer_rxt_threshold = rack_policer_rxt_thresh;
15362 		rack->r_ctl.policer_avg_threshold = rack_policer_avg_thresh;
15363 		rack->r_ctl.policer_med_threshold = rack_policer_med_thresh;
15364 		rack->policer_detect_on = 1;
15365 	} else {
15366 		rack->policer_detect_on = 0;
15367 	}
15368 	if (rack_fill_cw_state)
15369 		rack->rc_pace_to_cwnd = 1;
15370 	if (rack_pacing_min_seg)
15371 		rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
15372 	if (use_rack_rr)
15373 		rack->use_rack_rr = 1;
15374 	if (rack_dnd_default) {
15375 		rack->rc_pace_dnd = 1;
15376 	}
15377 	if (V_tcp_delack_enabled)
15378 		tp->t_delayed_ack = 1;
15379 	else
15380 		tp->t_delayed_ack = 0;
15381 #ifdef TCP_ACCOUNTING
15382 	if (rack_tcp_accounting) {
15383 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
15384 	}
15385 #endif
15386 	rack->r_ctl.pcm_i.cnt_alloc = RACK_DEFAULT_PCM_ARRAY;
15387 	sz = (sizeof(struct rack_pcm_stats) * rack->r_ctl.pcm_i.cnt_alloc);
15388 	rack->r_ctl.pcm_s = malloc(sz,M_TCPPCM, M_NOWAIT);
15389 	if (rack->r_ctl.pcm_s == NULL) {
15390 		rack->r_ctl.pcm_i.cnt_alloc = 0;
15391 	}
15392 #ifdef NETFLIX_STATS
15393 	rack->r_ctl.side_chan_dis_mask = tcp_sidechannel_disable_mask;
15394 #endif
15395 	rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
15396 	rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
15397 	if (rack_enable_shared_cwnd)
15398 		rack->rack_enable_scwnd = 1;
15399 	rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
15400 	rack->rc_user_set_max_segs = rack_hptsi_segments;
15401 	rack->r_ctl.max_reduction = rack_max_reduce;
15402 	rack->rc_force_max_seg = 0;
15403 	TAILQ_INIT(&rack->r_ctl.opt_list);
15404 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
15405 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
15406 	if (rack_hibeta_setting) {
15407 		rack->rack_hibeta = 1;
15408 		if ((rack_hibeta_setting >= 50) &&
15409 		    (rack_hibeta_setting <= 100)) {
15410 			rack->r_ctl.rc_saved_beta.beta = rack_hibeta_setting;
15411 			rack->r_ctl.saved_hibeta = rack_hibeta_setting;
15412 		}
15413 	} else {
15414 		rack->r_ctl.saved_hibeta = 50;
15415 	}
15416 	/*
15417 	 * We initialize to all ones so we never match 0
15418 	 * just in case the client sends in 0, it hopefully
15419 	 * will never have all 1's in ms :-)
15420 	 */
15421 	rack->r_ctl.last_tm_mark = 0xffffffffffffffff;
15422 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
15423 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
15424 	rack->r_ctl.pol_bw_comp = rack_policing_do_bw_comp;
15425 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
15426 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
15427 	rack->r_ctl.rc_highest_us_rtt = 0;
15428 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
15429 	rack->pcm_enabled = rack_pcm_is_enabled;
15430 	if (rack_fillcw_bw_cap)
15431 		rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
15432 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
15433 	if (rack_use_cmp_acks)
15434 		rack->r_use_cmp_ack = 1;
15435 	if (rack_disable_prr)
15436 		rack->rack_no_prr = 1;
15437 	if (rack_gp_no_rec_chg)
15438 		rack->rc_gp_no_rec_chg = 1;
15439 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
15440 		rack->r_ctl.pacing_method |= RACK_REG_PACING;
15441 		rack->rc_always_pace = 1;
15442 		if (rack->rack_hibeta)
15443 			rack_set_cc_pacing(rack);
15444 	} else
15445 		rack->rc_always_pace = 0;
15446 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
15447 		rack->r_mbuf_queue = 1;
15448 	else
15449 		rack->r_mbuf_queue = 0;
15450 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
15451 	if (rack_limits_scwnd)
15452 		rack->r_limit_scw = 1;
15453 	else
15454 		rack->r_limit_scw = 0;
15455 	rack_init_retransmit_value(rack, rack_rxt_controls);
15456 	rack->rc_labc = V_tcp_abc_l_var;
15457 	if (rack_honors_hpts_min_to)
15458 		rack->r_use_hpts_min = 1;
15459 	if (tp->snd_una != 0) {
15460 		rack->r_ctl.idle_snd_una = tp->snd_una;
15461 		rack->rc_sendvars_notset = 0;
15462 		/*
15463 		 * Make sure any TCP timers are not running.
15464 		 */
15465 		tcp_timer_stop(tp);
15466 	} else {
15467 		/*
15468 		 * Server side, we are called from the
15469 		 * syn-cache. This means none of the
15470 		 * snd_una/max are set yet so we have
15471 		 * to defer this until the first send.
15472 		 */
15473 		rack->rc_sendvars_notset = 1;
15474 	}
15475 
15476 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
15477 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
15478 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
15479 	rack->r_ctl.rc_min_to = rack_min_to;
15480 	microuptime(&rack->r_ctl.act_rcv_time);
15481 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
15482 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
15483 	if (rack_hw_up_only)
15484 		rack->r_up_only = 1;
15485 	if (rack_do_dyn_mul) {
15486 		/* When dynamic adjustment is on CA needs to start at 100% */
15487 		rack->rc_gp_dyn_mul = 1;
15488 		if (rack_do_dyn_mul >= 100)
15489 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
15490 	} else
15491 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
15492 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
15493 	if (rack_timely_off) {
15494 		rack->rc_skip_timely = 1;
15495 	}
15496 	if (rack->rc_skip_timely) {
15497 		rack->r_ctl.rack_per_of_gp_rec = 90;
15498 		rack->r_ctl.rack_per_of_gp_ca = 100;
15499 		rack->r_ctl.rack_per_of_gp_ss = 250;
15500 	}
15501 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
15502 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
15503 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
15504 
15505 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
15506 				rack_probertt_filter_life);
15507 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15508 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
15509 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
15510 	rack->r_ctl.rc_went_idle_time = us_cts;
15511 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks() - (tcp_ack_war_time_window + 1);
15512 	rack->r_ctl.rc_time_probertt_starts = 0;
15513 
15514 	rack->r_ctl.gp_rnd_thresh = rack_rnd_cnt_req & 0xff;
15515 	if (rack_rnd_cnt_req  & 0x10000)
15516 		rack->r_ctl.gate_to_fs = 1;
15517 	rack->r_ctl.gp_gain_req = rack_gp_gain_req;
15518 	if ((rack_rnd_cnt_req & 0x100) > 0) {
15519 
15520 	}
15521 	if (rack_dsack_std_based & 0x1) {
15522 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
15523 		rack->rc_rack_tmr_std_based = 1;
15524 	}
15525 	if (rack_dsack_std_based & 0x2) {
15526 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
15527 		rack->rc_rack_use_dsack = 1;
15528 	}
15529 	/* We require at least one measurement, even if the sysctl is 0 */
15530 	if (rack_req_measurements)
15531 		rack->r_ctl.req_measurements = rack_req_measurements;
15532 	else
15533 		rack->r_ctl.req_measurements = 1;
15534 	if (rack_enable_hw_pacing)
15535 		rack->rack_hdw_pace_ena = 1;
15536 	if (rack_hw_rate_caps)
15537 		rack->r_rack_hw_rate_caps = 1;
15538 	if (rack_non_rxt_use_cr)
15539 		rack->rack_rec_nonrxt_use_cr = 1;
15540 	/* Lets setup the fsb block */
15541 	err = rack_init_fsb(tp, rack);
15542 	if (err) {
15543 		uma_zfree(rack_pcb_zone, *ptr);
15544 		*ptr = NULL;
15545 		return (err);
15546 	}
15547 	if (rack_do_hystart) {
15548 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
15549 		if (rack_do_hystart > 1)
15550 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
15551 		if (rack_do_hystart > 2)
15552 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
15553 	}
15554 	/* Log what we will do with queries */
15555 	rack_log_chg_info(tp, rack, 7,
15556 			  no_query, 0, 0);
15557 	if (rack_def_profile)
15558 		rack_set_profile(rack, rack_def_profile);
15559 	/* Cancel the GP measurement in progress */
15560 	tp->t_flags &= ~TF_GPUTINPROG;
15561 	if ((tp->t_state != TCPS_CLOSED) &&
15562 	    (tp->t_state != TCPS_TIME_WAIT)) {
15563 		/*
15564 		 * We are already open, we may
15565 		 * need to adjust a few things.
15566 		 */
15567 		if (SEQ_GT(tp->snd_max, tp->iss))
15568 			snt = tp->snd_max - tp->iss;
15569 		else
15570 			snt = 0;
15571 		iwin = rc_init_window(rack);
15572 		if ((snt < iwin) &&
15573 		    (no_query == 1)) {
15574 			/* We are not past the initial window
15575 			 * on the first init (i.e. a stack switch
15576 			 * has not yet occured) so we need to make
15577 			 * sure cwnd and ssthresh is correct.
15578 			 */
15579 			if (tp->snd_cwnd < iwin)
15580 				tp->snd_cwnd = iwin;
15581 			/*
15582 			 * If we are within the initial window
15583 			 * we want ssthresh to be unlimited. Setting
15584 			 * it to the rwnd (which the default stack does
15585 			 * and older racks) is not really a good idea
15586 			 * since we want to be in SS and grow both the
15587 			 * cwnd and the rwnd (via dynamic rwnd growth). If
15588 			 * we set it to the rwnd then as the peer grows its
15589 			 * rwnd we will be stuck in CA and never hit SS.
15590 			 *
15591 			 * Its far better to raise it up high (this takes the
15592 			 * risk that there as been a loss already, probably
15593 			 * we should have an indicator in all stacks of loss
15594 			 * but we don't), but considering the normal use this
15595 			 * is a risk worth taking. The consequences of not
15596 			 * hitting SS are far worse than going one more time
15597 			 * into it early on (before we have sent even a IW).
15598 			 * It is highly unlikely that we will have had a loss
15599 			 * before getting the IW out.
15600 			 */
15601 			tp->snd_ssthresh = 0xffffffff;
15602 		}
15603 		/*
15604 		 * Any init based on sequence numbers
15605 		 * should be done in the deferred init path
15606 		 * since we can be CLOSED and not have them
15607 		 * inited when rack_init() is called. We
15608 		 * are not closed so lets call it.
15609 		 */
15610 		rack_deferred_init(tp, rack);
15611 	}
15612 	if ((tp->t_state != TCPS_CLOSED) &&
15613 	    (tp->t_state != TCPS_TIME_WAIT) &&
15614 	    (no_query == 0) &&
15615 	    (tp->snd_una != tp->snd_max))  {
15616 		err = rack_init_outstanding(tp, rack, us_cts, *ptr);
15617 		if (err) {
15618 			*ptr = NULL;
15619 			return(err);
15620 		}
15621 	}
15622 	rack_stop_all_timers(tp, rack);
15623 	/* Setup all the t_flags2 */
15624 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
15625 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
15626 	else
15627 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
15628 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15629 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
15630 	/*
15631 	 * Timers in Rack are kept in microseconds so lets
15632 	 * convert any initial incoming variables
15633 	 * from ticks into usecs. Note that we
15634 	 * also change the values of t_srtt and t_rttvar, if
15635 	 * they are non-zero. They are kept with a 5
15636 	 * bit decimal so we have to carefully convert
15637 	 * these to get the full precision.
15638 	 */
15639 	rack_convert_rtts(tp);
15640 	rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
15641 	if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
15642 		/* We do not start any timers on DROPPED connections */
15643 		if (tp->t_fb->tfb_chg_query == NULL) {
15644 			rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15645 		} else {
15646 			struct tcp_query_resp qr;
15647 			int ret;
15648 
15649 			memset(&qr, 0, sizeof(qr));
15650 
15651 			/* Get the misc time stamps and such for rack */
15652 			qr.req = TCP_QUERY_RACK_TIMES;
15653 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15654 			if (ret == 1) {
15655 				rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
15656 				rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
15657 				rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
15658 				rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
15659 				rack->rc_rack_rtt = qr.rack_rtt;
15660 				rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
15661 				rack->r_ctl.rc_sacked = qr.rack_sacked;
15662 				rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
15663 				rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
15664 				rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
15665 				rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
15666 				rack->r_ctl.rc_prr_out = qr.rack_prr_out;
15667 				if (qr.rack_tlp_out) {
15668 					rack->rc_tlp_in_progress = 1;
15669 					rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
15670 				} else {
15671 					rack->rc_tlp_in_progress = 0;
15672 					rack->r_ctl.rc_tlp_cnt_out = 0;
15673 				}
15674 				if (qr.rack_srtt_measured)
15675 					rack->rc_srtt_measure_made = 1;
15676 				if (qr.rack_in_persist == 1) {
15677 					rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
15678 #ifdef NETFLIX_SHARED_CWND
15679 					if (rack->r_ctl.rc_scw) {
15680 						tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
15681 						rack->rack_scwnd_is_idle = 1;
15682 					}
15683 #endif
15684 					rack->r_ctl.persist_lost_ends = 0;
15685 					rack->probe_not_answered = 0;
15686 					rack->forced_ack = 0;
15687 					tp->t_rxtshift = 0;
15688 					rack->rc_in_persist = 1;
15689 					RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
15690 							   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
15691 				}
15692 				if (qr.rack_wanted_output)
15693 					rack->r_wanted_output = 1;
15694 				rack_log_chg_info(tp, rack, 6,
15695 						  qr.rack_min_rtt,
15696 						  qr.rack_rtt,
15697 						  qr.rack_reorder_ts);
15698 			}
15699 			/* Get the old stack timers */
15700 			qr.req_param = 0;
15701 			qr.req = TCP_QUERY_TIMERS_UP;
15702 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15703 			if (ret) {
15704 				/*
15705 				 * non-zero return means we have a timer('s)
15706 				 * to start. Zero means no timer (no keepalive
15707 				 * I suppose).
15708 				 */
15709 				uint32_t tov = 0;
15710 
15711 				rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
15712 				if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
15713 					rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
15714 					if (TSTMP_GT(qr.timer_pacing_to, us_cts))
15715 						tov = qr.timer_pacing_to - us_cts;
15716 					else
15717 						tov = HPTS_TICKS_PER_SLOT;
15718 				}
15719 				if (qr.timer_hpts_flags & PACE_TMR_MASK) {
15720 					rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
15721 					if (tov == 0) {
15722 						if (TSTMP_GT(qr.timer_timer_exp, us_cts))
15723 							tov = qr.timer_timer_exp - us_cts;
15724 						else
15725 							tov = HPTS_TICKS_PER_SLOT;
15726 					}
15727 				}
15728 				rack_log_chg_info(tp, rack, 4,
15729 						  rack->r_ctl.rc_hpts_flags,
15730 						  rack->r_ctl.rc_last_output_to,
15731 						  rack->r_ctl.rc_timer_exp);
15732 				if (tov) {
15733 					struct hpts_diag diag;
15734 
15735 					(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
15736 								   __LINE__, &diag);
15737 					rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
15738 				}
15739 			}
15740 		}
15741 		rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
15742 				     __LINE__, RACK_RTTS_INIT);
15743 	}
15744 	return (0);
15745 }
15746 
15747 static int
15748 rack_handoff_ok(struct tcpcb *tp)
15749 {
15750 	if ((tp->t_state == TCPS_CLOSED) ||
15751 	    (tp->t_state == TCPS_LISTEN)) {
15752 		/* Sure no problem though it may not stick */
15753 		return (0);
15754 	}
15755 	if ((tp->t_state == TCPS_SYN_SENT) ||
15756 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
15757 		/*
15758 		 * We really don't know if you support sack,
15759 		 * you have to get to ESTAB or beyond to tell.
15760 		 */
15761 		return (EAGAIN);
15762 	}
15763 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15764 		/*
15765 		 * Rack will only send a FIN after all data is acknowledged.
15766 		 * So in this case we have more data outstanding. We can't
15767 		 * switch stacks until either all data and only the FIN
15768 		 * is left (in which case rack_init() now knows how
15769 		 * to deal with that) <or> all is acknowledged and we
15770 		 * are only left with incoming data, though why you
15771 		 * would want to switch to rack after all data is acknowledged
15772 		 * I have no idea (rrs)!
15773 		 */
15774 		return (EAGAIN);
15775 	}
15776 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15777 		return (0);
15778 	}
15779 	/*
15780 	 * If we reach here we don't do SACK on this connection so we can
15781 	 * never do rack.
15782 	 */
15783 	return (EINVAL);
15784 }
15785 
15786 static void
15787 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15788 {
15789 
15790 	if (tp->t_fb_ptr) {
15791 		uint32_t cnt_free = 0;
15792 		struct tcp_rack *rack;
15793 		struct rack_sendmap *rsm;
15794 
15795 		tcp_handle_orphaned_packets(tp);
15796 		tp->t_flags &= ~TF_FORCEDATA;
15797 		rack = (struct tcp_rack *)tp->t_fb_ptr;
15798 		rack_log_pacing_delay_calc(rack,
15799 					   0,
15800 					   0,
15801 					   0,
15802 					   rack_get_gp_est(rack), /* delRate */
15803 					   rack_get_lt_bw(rack), /* rttProp */
15804 					   20, __LINE__, NULL, 0);
15805 #ifdef NETFLIX_SHARED_CWND
15806 		if (rack->r_ctl.rc_scw) {
15807 			uint32_t limit;
15808 
15809 			if (rack->r_limit_scw)
15810 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15811 			else
15812 				limit = 0;
15813 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15814 						  rack->r_ctl.rc_scw_index,
15815 						  limit);
15816 			rack->r_ctl.rc_scw = NULL;
15817 		}
15818 #endif
15819 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
15820 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15821 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15822 			rack->r_ctl.fsb.th = NULL;
15823 		}
15824 		if (rack->rc_always_pace == 1) {
15825 			rack_remove_pacing(rack);
15826 		}
15827 		/* Clean up any options if they were not applied */
15828 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15829 			struct deferred_opt_list *dol;
15830 
15831 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15832 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15833 			free(dol, M_TCPDO);
15834 		}
15835 		/* rack does not use force data but other stacks may clear it */
15836 		if (rack->r_ctl.crte != NULL) {
15837 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15838 			rack->rack_hdrw_pacing = 0;
15839 			rack->r_ctl.crte = NULL;
15840 		}
15841 #ifdef TCP_BLACKBOX
15842 		tcp_log_flowend(tp);
15843 #endif
15844 		/*
15845 		 * Lets take a different approach to purging just
15846 		 * get each one and free it like a cum-ack would and
15847 		 * not use a foreach loop.
15848 		 */
15849 		rsm = tqhash_min(rack->r_ctl.tqh);
15850 		while (rsm) {
15851 			tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15852 			rack->r_ctl.rc_num_maps_alloced--;
15853 			uma_zfree(rack_zone, rsm);
15854 			rsm = tqhash_min(rack->r_ctl.tqh);
15855 		}
15856 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15857 		while (rsm) {
15858 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15859 			rack->r_ctl.rc_num_maps_alloced--;
15860 			rack->rc_free_cnt--;
15861 			cnt_free++;
15862 			uma_zfree(rack_zone, rsm);
15863 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15864 		}
15865 		if (rack->r_ctl.pcm_s != NULL) {
15866 			free(rack->r_ctl.pcm_s, M_TCPPCM);
15867 			rack->r_ctl.pcm_s = NULL;
15868 			rack->r_ctl.pcm_i.cnt_alloc = 0;
15869 			rack->r_ctl.pcm_i.cnt = 0;
15870 		}
15871 		if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15872 		    (tcp_bblogging_on(tp))) {
15873 			union tcp_log_stackspecific log;
15874 			struct timeval tv;
15875 
15876 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15877 			log.u_bbr.flex8 = 10;
15878 			log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15879 			log.u_bbr.flex2 = rack->rc_free_cnt;
15880 			log.u_bbr.flex3 = cnt_free;
15881 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15882 			rsm = tqhash_min(rack->r_ctl.tqh);
15883 			log.u_bbr.delRate = (uintptr_t)rsm;
15884 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15885 			log.u_bbr.cur_del_rate = (uintptr_t)rsm;
15886 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15887 			log.u_bbr.pkt_epoch = __LINE__;
15888 			(void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15889 					     0, &log, false, NULL, NULL, 0, &tv);
15890 		}
15891 		KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15892 			("rack:%p num_aloc:%u after freeing all?",
15893 			 rack,
15894 			 rack->r_ctl.rc_num_maps_alloced));
15895 		rack->rc_free_cnt = 0;
15896 		free(rack->r_ctl.tqh, M_TCPFSB);
15897 		rack->r_ctl.tqh = NULL;
15898 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15899 		tp->t_fb_ptr = NULL;
15900 	}
15901 	/* Make sure snd_nxt is correctly set */
15902 	tp->snd_nxt = tp->snd_max;
15903 }
15904 
15905 static void
15906 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15907 {
15908 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15909 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15910 	}
15911 	switch (tp->t_state) {
15912 	case TCPS_SYN_SENT:
15913 		rack->r_state = TCPS_SYN_SENT;
15914 		rack->r_substate = rack_do_syn_sent;
15915 		break;
15916 	case TCPS_SYN_RECEIVED:
15917 		rack->r_state = TCPS_SYN_RECEIVED;
15918 		rack->r_substate = rack_do_syn_recv;
15919 		break;
15920 	case TCPS_ESTABLISHED:
15921 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15922 		rack->r_state = TCPS_ESTABLISHED;
15923 		rack->r_substate = rack_do_established;
15924 		break;
15925 	case TCPS_CLOSE_WAIT:
15926 		rack->r_state = TCPS_CLOSE_WAIT;
15927 		rack->r_substate = rack_do_close_wait;
15928 		break;
15929 	case TCPS_FIN_WAIT_1:
15930 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15931 		rack->r_state = TCPS_FIN_WAIT_1;
15932 		rack->r_substate = rack_do_fin_wait_1;
15933 		break;
15934 	case TCPS_CLOSING:
15935 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15936 		rack->r_state = TCPS_CLOSING;
15937 		rack->r_substate = rack_do_closing;
15938 		break;
15939 	case TCPS_LAST_ACK:
15940 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15941 		rack->r_state = TCPS_LAST_ACK;
15942 		rack->r_substate = rack_do_lastack;
15943 		break;
15944 	case TCPS_FIN_WAIT_2:
15945 		rack->r_state = TCPS_FIN_WAIT_2;
15946 		rack->r_substate = rack_do_fin_wait_2;
15947 		break;
15948 	case TCPS_LISTEN:
15949 	case TCPS_CLOSED:
15950 	case TCPS_TIME_WAIT:
15951 	default:
15952 		break;
15953 	};
15954 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15955 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15956 
15957 }
15958 
15959 static void
15960 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15961 {
15962 	/*
15963 	 * We received an ack, and then did not
15964 	 * call send or were bounced out due to the
15965 	 * hpts was running. Now a timer is up as well, is
15966 	 * it the right timer?
15967 	 */
15968 	struct rack_sendmap *rsm;
15969 	int tmr_up;
15970 
15971 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15972 	if (tcp_in_hpts(rack->rc_tp) == 0) {
15973 		/*
15974 		 * Ok we probably need some timer up, but no
15975 		 * matter what the mask we are not in hpts. We
15976 		 * may have received an old ack and thus did nothing.
15977 		 */
15978 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15979 		rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15980 		return;
15981 	}
15982 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15983 		return;
15984 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15985 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15986 	    (tmr_up == PACE_TMR_RXT)) {
15987 		/* Should be an RXT */
15988 		return;
15989 	}
15990 	if (rsm == NULL) {
15991 		/* Nothing outstanding? */
15992 		if (tp->t_flags & TF_DELACK) {
15993 			if (tmr_up == PACE_TMR_DELACK)
15994 				/* We are supposed to have delayed ack up and we do */
15995 				return;
15996 		} else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
15997 			/*
15998 			 * if we hit enobufs then we would expect the possibility
15999 			 * of nothing outstanding and the RXT up (and the hptsi timer).
16000 			 */
16001 			return;
16002 		} else if (((V_tcp_always_keepalive ||
16003 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
16004 			    (tp->t_state <= TCPS_CLOSING)) &&
16005 			   (tmr_up == PACE_TMR_KEEP) &&
16006 			   (tp->snd_max == tp->snd_una)) {
16007 			/* We should have keep alive up and we do */
16008 			return;
16009 		}
16010 	}
16011 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
16012 		   ((tmr_up == PACE_TMR_TLP) ||
16013 		    (tmr_up == PACE_TMR_RACK) ||
16014 		    (tmr_up == PACE_TMR_RXT))) {
16015 		/*
16016 		 * Either a Rack, TLP or RXT is fine if  we
16017 		 * have outstanding data.
16018 		 */
16019 		return;
16020 	} else if (tmr_up == PACE_TMR_DELACK) {
16021 		/*
16022 		 * If the delayed ack was going to go off
16023 		 * before the rtx/tlp/rack timer were going to
16024 		 * expire, then that would be the timer in control.
16025 		 * Note we don't check the time here trusting the
16026 		 * code is correct.
16027 		 */
16028 		return;
16029 	}
16030 	/*
16031 	 * Ok the timer originally started is not what we want now.
16032 	 * We will force the hpts to be stopped if any, and restart
16033 	 * with the slot set to what was in the saved slot.
16034 	 */
16035 	if (tcp_in_hpts(rack->rc_tp)) {
16036 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
16037 			uint32_t us_cts;
16038 
16039 			us_cts = tcp_get_usecs(NULL);
16040 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
16041 				rack->r_early = 1;
16042 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
16043 			}
16044 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16045 		}
16046 		tcp_hpts_remove(rack->rc_tp);
16047 	}
16048 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16049 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
16050 }
16051 
16052 
16053 static void
16054 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
16055 {
16056 	if ((SEQ_LT(tp->snd_wl1, seq) ||
16057 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
16058 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
16059 		/* keep track of pure window updates */
16060 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
16061 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
16062 		tp->snd_wnd = tiwin;
16063 		rack_validate_fo_sendwin_up(tp, rack);
16064 		tp->snd_wl1 = seq;
16065 		tp->snd_wl2 = ack;
16066 		if (tp->snd_wnd > tp->max_sndwnd)
16067 			tp->max_sndwnd = tp->snd_wnd;
16068 	    rack->r_wanted_output = 1;
16069 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
16070 		tp->snd_wnd = tiwin;
16071 		rack_validate_fo_sendwin_up(tp, rack);
16072 		tp->snd_wl1 = seq;
16073 		tp->snd_wl2 = ack;
16074 	} else {
16075 		/* Not a valid win update */
16076 		return;
16077 	}
16078 	if (tp->snd_wnd > tp->max_sndwnd)
16079 		tp->max_sndwnd = tp->snd_wnd;
16080 	/* Do we exit persists? */
16081 	if ((rack->rc_in_persist != 0) &&
16082 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
16083 				rack->r_ctl.rc_pace_min_segs))) {
16084 		rack_exit_persist(tp, rack, cts);
16085 	}
16086 	/* Do we enter persists? */
16087 	if ((rack->rc_in_persist == 0) &&
16088 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
16089 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
16090 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
16091 	    sbavail(&tptosocket(tp)->so_snd) &&
16092 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
16093 		/*
16094 		 * Here the rwnd is less than
16095 		 * the pacing size, we are established,
16096 		 * nothing is outstanding, and there is
16097 		 * data to send. Enter persists.
16098 		 */
16099 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
16100 	}
16101 }
16102 
16103 static void
16104 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
16105 {
16106 
16107 	if (tcp_bblogging_on(rack->rc_tp)) {
16108 		struct inpcb *inp = tptoinpcb(tp);
16109 		union tcp_log_stackspecific log;
16110 		struct timeval ltv;
16111 		char tcp_hdr_buf[60];
16112 		struct tcphdr *th;
16113 		struct timespec ts;
16114 		uint32_t orig_snd_una;
16115 		uint8_t xx = 0;
16116 
16117 #ifdef TCP_REQUEST_TRK
16118 		struct tcp_sendfile_track *tcp_req;
16119 
16120 		if (SEQ_GT(ae->ack, tp->snd_una)) {
16121 			tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
16122 		} else {
16123 			tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
16124 		}
16125 #endif
16126 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16127 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16128 		if (rack->rack_no_prr == 0)
16129 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16130 		else
16131 			log.u_bbr.flex1 = 0;
16132 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16133 		log.u_bbr.use_lt_bw <<= 1;
16134 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
16135 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16136 		log.u_bbr.bbr_state = rack->rc_free_cnt;
16137 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
16138 		log.u_bbr.pkts_out = tp->t_maxseg;
16139 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16140 		log.u_bbr.flex7 = 1;
16141 		log.u_bbr.lost = ae->flags;
16142 		log.u_bbr.cwnd_gain = ackval;
16143 		log.u_bbr.pacing_gain = 0x2;
16144 		if (ae->flags & TSTMP_HDWR) {
16145 			/* Record the hardware timestamp if present */
16146 			log.u_bbr.flex3 = M_TSTMP;
16147 			ts.tv_sec = ae->timestamp / 1000000000;
16148 			ts.tv_nsec = ae->timestamp % 1000000000;
16149 			ltv.tv_sec = ts.tv_sec;
16150 			ltv.tv_usec = ts.tv_nsec / 1000;
16151 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
16152 		} else if (ae->flags & TSTMP_LRO) {
16153 			/* Record the LRO the arrival timestamp */
16154 			log.u_bbr.flex3 = M_TSTMP_LRO;
16155 			ts.tv_sec = ae->timestamp / 1000000000;
16156 			ts.tv_nsec = ae->timestamp % 1000000000;
16157 			ltv.tv_sec = ts.tv_sec;
16158 			ltv.tv_usec = ts.tv_nsec / 1000;
16159 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
16160 		}
16161 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16162 		/* Log the rcv time */
16163 		log.u_bbr.delRate = ae->timestamp;
16164 #ifdef TCP_REQUEST_TRK
16165 		log.u_bbr.applimited = tp->t_tcpreq_closed;
16166 		log.u_bbr.applimited <<= 8;
16167 		log.u_bbr.applimited |= tp->t_tcpreq_open;
16168 		log.u_bbr.applimited <<= 8;
16169 		log.u_bbr.applimited |= tp->t_tcpreq_req;
16170 		if (tcp_req) {
16171 			/* Copy out any client req info */
16172 			/* seconds */
16173 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16174 			/* useconds */
16175 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16176 			log.u_bbr.rttProp = tcp_req->timestamp;
16177 			log.u_bbr.cur_del_rate = tcp_req->start;
16178 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16179 				log.u_bbr.flex8 |= 1;
16180 			} else {
16181 				log.u_bbr.flex8 |= 2;
16182 				log.u_bbr.bw_inuse = tcp_req->end;
16183 			}
16184 			log.u_bbr.flex6 = tcp_req->start_seq;
16185 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16186 				log.u_bbr.flex8 |= 4;
16187 				log.u_bbr.epoch = tcp_req->end_seq;
16188 			}
16189 		}
16190 #endif
16191 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
16192 		th = (struct tcphdr *)tcp_hdr_buf;
16193 		th->th_seq = ae->seq;
16194 		th->th_ack = ae->ack;
16195 		th->th_win = ae->win;
16196 		/* Now fill in the ports */
16197 		th->th_sport = inp->inp_fport;
16198 		th->th_dport = inp->inp_lport;
16199 		tcp_set_flags(th, ae->flags);
16200 		/* Now do we have a timestamp option? */
16201 		if (ae->flags & HAS_TSTMP) {
16202 			u_char *cp;
16203 			uint32_t val;
16204 
16205 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
16206 			cp = (u_char *)(th + 1);
16207 			*cp = TCPOPT_NOP;
16208 			cp++;
16209 			*cp = TCPOPT_NOP;
16210 			cp++;
16211 			*cp = TCPOPT_TIMESTAMP;
16212 			cp++;
16213 			*cp = TCPOLEN_TIMESTAMP;
16214 			cp++;
16215 			val = htonl(ae->ts_value);
16216 			bcopy((char *)&val,
16217 			      (char *)cp, sizeof(uint32_t));
16218 			val = htonl(ae->ts_echo);
16219 			bcopy((char *)&val,
16220 			      (char *)(cp + 4), sizeof(uint32_t));
16221 		} else
16222 			th->th_off = (sizeof(struct tcphdr) >> 2);
16223 
16224 		/*
16225 		 * For sane logging we need to play a little trick.
16226 		 * If the ack were fully processed we would have moved
16227 		 * snd_una to high_seq, but since compressed acks are
16228 		 * processed in two phases, at this point (logging) snd_una
16229 		 * won't be advanced. So we would see multiple acks showing
16230 		 * the advancement. We can prevent that by "pretending" that
16231 		 * snd_una was advanced and then un-advancing it so that the
16232 		 * logging code has the right value for tlb_snd_una.
16233 		 */
16234 		if (tp->snd_una != high_seq) {
16235 			orig_snd_una = tp->snd_una;
16236 			tp->snd_una = high_seq;
16237 			xx = 1;
16238 		} else
16239 			xx = 0;
16240 		TCP_LOG_EVENTP(tp, th,
16241 			       &tptosocket(tp)->so_rcv,
16242 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
16243 			       0, &log, true, &ltv);
16244 		if (xx) {
16245 			tp->snd_una = orig_snd_una;
16246 		}
16247 	}
16248 
16249 }
16250 
16251 static void
16252 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
16253 {
16254 	uint32_t us_rtt;
16255 	/*
16256 	 * A persist or keep-alive was forced out, update our
16257 	 * min rtt time. Note now worry about lost responses.
16258 	 * When a subsequent keep-alive or persist times out
16259 	 * and forced_ack is still on, then the last probe
16260 	 * was not responded to. In such cases we have a
16261 	 * sysctl that controls the behavior. Either we apply
16262 	 * the rtt but with reduced confidence (0). Or we just
16263 	 * plain don't apply the rtt estimate. Having data flow
16264 	 * will clear the probe_not_answered flag i.e. cum-ack
16265 	 * move forward <or> exiting and reentering persists.
16266 	 */
16267 
16268 	rack->forced_ack = 0;
16269 	rack->rc_tp->t_rxtshift = 0;
16270 	if ((rack->rc_in_persist &&
16271 	     (tiwin == rack->rc_tp->snd_wnd)) ||
16272 	    (rack->rc_in_persist == 0)) {
16273 		/*
16274 		 * In persists only apply the RTT update if this is
16275 		 * a response to our window probe. And that
16276 		 * means the rwnd sent must match the current
16277 		 * snd_wnd. If it does not, then we got a
16278 		 * window update ack instead. For keepalive
16279 		 * we allow the answer no matter what the window.
16280 		 *
16281 		 * Note that if the probe_not_answered is set then
16282 		 * the forced_ack_ts is the oldest one i.e. the first
16283 		 * probe sent that might have been lost. This assures
16284 		 * us that if we do calculate an RTT it is longer not
16285 		 * some short thing.
16286 		 */
16287 		if (rack->rc_in_persist)
16288 			counter_u64_add(rack_persists_acks, 1);
16289 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
16290 		if (us_rtt == 0)
16291 			us_rtt = 1;
16292 		if (rack->probe_not_answered == 0) {
16293 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
16294 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
16295 		} else {
16296 			/* We have a retransmitted probe here too */
16297 			if (rack_apply_rtt_with_reduced_conf) {
16298 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
16299 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
16300 			}
16301 		}
16302 	}
16303 }
16304 
16305 static void
16306 rack_new_round_starts(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
16307 {
16308 	/*
16309 	 * The next send has occurred mark the end of the round
16310 	 * as when that data gets acknowledged. We can
16311 	 * also do common things we might need to do when
16312 	 * a round begins.
16313 	 */
16314 	rack->r_ctl.roundends = tp->snd_max;
16315 	rack->rc_new_rnd_needed = 0;
16316 	rack_log_hystart_event(rack, tp->snd_max, 4);
16317 }
16318 
16319 
16320 static void
16321 rack_log_pcm(struct tcp_rack *rack, uint8_t mod, uint32_t flex1, uint32_t flex2,
16322 	     uint32_t flex3)
16323 {
16324 	if (tcp_bblogging_on(rack->rc_tp)) {
16325 		union tcp_log_stackspecific log;
16326 		struct timeval tv;
16327 
16328 		(void)tcp_get_usecs(&tv);
16329 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16330 		log.u_bbr.timeStamp = tcp_tv_to_usectick(&tv);
16331 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16332 		log.u_bbr.flex8 = mod;
16333 		log.u_bbr.flex1 = flex1;
16334 		log.u_bbr.flex2 = flex2;
16335 		log.u_bbr.flex3 = flex3;
16336 		log.u_bbr.flex4 = rack_pcm_every_n_rounds;
16337 		log.u_bbr.flex5 = rack->r_ctl.pcm_idle_rounds;
16338 		log.u_bbr.bbr_substate = rack->pcm_needed;
16339 		log.u_bbr.bbr_substate <<= 1;
16340 		log.u_bbr.bbr_substate |= rack->pcm_in_progress;
16341 		log.u_bbr.bbr_substate <<= 1;
16342 		log.u_bbr.bbr_substate |= rack->pcm_enabled; /* bits are NIE for Needed, Inprogress, Enabled */
16343 		(void)tcp_log_event(rack->rc_tp, NULL, NULL, NULL, TCP_PCM_MEASURE, ERRNO_UNK,
16344 				    0, &log, false, NULL, NULL, 0, &tv);
16345 	}
16346 }
16347 
16348 static void
16349 rack_new_round_setup(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
16350 {
16351 	/*
16352 	 * The round (current_round) has ended. We now
16353 	 * setup for the next round by incrementing the
16354 	 * round numnber and doing any round specific
16355 	 * things.
16356 	 */
16357 	rack_log_hystart_event(rack, high_seq, 21);
16358 	rack->r_ctl.current_round++;
16359 	/* New round (current_round) begins at next send */
16360 	rack->rc_new_rnd_needed = 1;
16361 	if ((rack->pcm_enabled == 1) &&
16362 	    (rack->pcm_needed == 0) &&
16363 	    (rack->pcm_in_progress == 0)) {
16364 		/*
16365 		 * If we have enabled PCM, then we need to
16366 		 * check if the round has adanced to the state
16367 		 * where one is required.
16368 		 */
16369 		int rnds;
16370 
16371 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
16372 		if ((rnds + rack->r_ctl.pcm_idle_rounds) >= rack_pcm_every_n_rounds) {
16373 			rack->pcm_needed = 1;
16374 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
16375 		} else if (rack_verbose_logging) {
16376 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
16377 		}
16378 	}
16379 	if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
16380 		/* We have hystart enabled send the round info in */
16381 		if (CC_ALGO(tp)->newround != NULL) {
16382 			CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
16383 		}
16384 	}
16385 	/*
16386 	 * For DGP an initial startup check. We want to validate
16387 	 * that we are not just pushing on slow-start and just
16388 	 * not gaining.. i.e. filling buffers without getting any
16389 	 * boost in b/w during the inital slow-start.
16390 	 */
16391 	if (rack->dgp_on &&
16392 	    (rack->rc_initial_ss_comp == 0) &&
16393 	    (tp->snd_cwnd < tp->snd_ssthresh) &&
16394 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG) &&
16395 	    (rack->r_ctl.gp_rnd_thresh > 0) &&
16396 	    ((rack->r_ctl.current_round - rack->r_ctl.last_rnd_of_gp_rise) >= rack->r_ctl.gp_rnd_thresh)) {
16397 
16398 		/*
16399 		 * We are in the initial SS and we have hd rack_rnd_cnt_req rounds(def:5) where
16400 		 * we have not gained the required amount in the gp_est (120.0% aka 1200). Lets
16401 		 * exit SS.
16402 		 *
16403 		 * Pick up the flight size now as we enter slowstart (not the
16404 		 * cwnd which may be inflated).
16405 		 */
16406 		rack->rc_initial_ss_comp = 1;
16407 
16408 		if (tcp_bblogging_on(rack->rc_tp)) {
16409 			union tcp_log_stackspecific log;
16410 			struct timeval tv;
16411 
16412 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16413 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
16414 			log.u_bbr.flex1 = rack->r_ctl.current_round;
16415 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
16416 			log.u_bbr.flex3 = rack->r_ctl.gp_rnd_thresh;
16417 			log.u_bbr.flex4 = rack->r_ctl.gate_to_fs;
16418 			log.u_bbr.flex5 = rack->r_ctl.ss_hi_fs;
16419 			log.u_bbr.flex8 = 40;
16420 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
16421 					    0, &log, false, NULL, __func__, __LINE__,&tv);
16422 		}
16423 		if ((rack->r_ctl.gate_to_fs == 1) &&
16424 		     (tp->snd_cwnd > rack->r_ctl.ss_hi_fs)) {
16425 			tp->snd_cwnd = rack->r_ctl.ss_hi_fs;
16426 		}
16427 		tp->snd_ssthresh = tp->snd_cwnd - 1;
16428 		/* Turn off any fast output running */
16429 		rack->r_fast_output = 0;
16430 	}
16431 }
16432 
16433 static int
16434 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
16435 {
16436 	/*
16437 	 * Handle a "special" compressed ack mbuf. Each incoming
16438 	 * ack has only four possible dispositions:
16439 	 *
16440 	 * A) It moves the cum-ack forward
16441 	 * B) It is behind the cum-ack.
16442 	 * C) It is a window-update ack.
16443 	 * D) It is a dup-ack.
16444 	 *
16445 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
16446 	 * in the incoming mbuf. We also need to still pay attention
16447 	 * to nxt_pkt since there may be another packet after this
16448 	 * one.
16449 	 */
16450 #ifdef TCP_ACCOUNTING
16451 	uint64_t ts_val;
16452 	uint64_t rdstc;
16453 #endif
16454 	int segsiz;
16455 	struct timespec ts;
16456 	struct tcp_rack *rack;
16457 	struct tcp_ackent *ae;
16458 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
16459 	int cnt, i, did_out, ourfinisacked = 0;
16460 	struct tcpopt to_holder, *to = NULL;
16461 #ifdef TCP_ACCOUNTING
16462 	int win_up_req = 0;
16463 #endif
16464 	int nsegs = 0;
16465 	int under_pacing = 0;
16466 	int post_recovery = 0;
16467 #ifdef TCP_ACCOUNTING
16468 	sched_pin();
16469 #endif
16470 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16471 	if (rack->gp_ready &&
16472 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
16473 		under_pacing = 1;
16474 
16475 	if (rack->r_state != tp->t_state)
16476 		rack_set_state(tp, rack);
16477 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16478 	    (tp->t_flags & TF_GPUTINPROG)) {
16479 		/*
16480 		 * We have a goodput in progress
16481 		 * and we have entered a late state.
16482 		 * Do we have enough data in the sb
16483 		 * to handle the GPUT request?
16484 		 */
16485 		uint32_t bytes;
16486 
16487 		bytes = tp->gput_ack - tp->gput_seq;
16488 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
16489 			bytes += tp->gput_seq - tp->snd_una;
16490 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16491 			/*
16492 			 * There are not enough bytes in the socket
16493 			 * buffer that have been sent to cover this
16494 			 * measurement. Cancel it.
16495 			 */
16496 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16497 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
16498 						   tp->gput_seq,
16499 						   0, 0, 18, __LINE__, NULL, 0);
16500 			tp->t_flags &= ~TF_GPUTINPROG;
16501 		}
16502 	}
16503 	to = &to_holder;
16504 	to->to_flags = 0;
16505 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
16506 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
16507 	cnt = m->m_len / sizeof(struct tcp_ackent);
16508 	counter_u64_add(rack_multi_single_eq, cnt);
16509 	high_seq = tp->snd_una;
16510 	the_win = tp->snd_wnd;
16511 	win_seq = tp->snd_wl1;
16512 	win_upd_ack = tp->snd_wl2;
16513 	cts = tcp_tv_to_usectick(tv);
16514 	ms_cts = tcp_tv_to_mssectick(tv);
16515 	rack->r_ctl.rc_rcvtime = cts;
16516 	segsiz = ctf_fixed_maxseg(tp);
16517 	if ((rack->rc_gp_dyn_mul) &&
16518 	    (rack->use_fixed_rate == 0) &&
16519 	    (rack->rc_always_pace)) {
16520 		/* Check in on probertt */
16521 		rack_check_probe_rtt(rack, cts);
16522 	}
16523 	for (i = 0; i < cnt; i++) {
16524 #ifdef TCP_ACCOUNTING
16525 		ts_val = get_cyclecount();
16526 #endif
16527 		rack_clear_rate_sample(rack);
16528 		ae = ((mtod(m, struct tcp_ackent *)) + i);
16529 		if (ae->flags & TH_FIN)
16530 			rack_log_pacing_delay_calc(rack,
16531 						   0,
16532 						   0,
16533 						   0,
16534 						   rack_get_gp_est(rack), /* delRate */
16535 						   rack_get_lt_bw(rack), /* rttProp */
16536 						   20, __LINE__, NULL, 0);
16537 		/* Setup the window */
16538 		tiwin = ae->win << tp->snd_scale;
16539 		if (tiwin > rack->r_ctl.rc_high_rwnd)
16540 			rack->r_ctl.rc_high_rwnd = tiwin;
16541 		/* figure out the type of ack */
16542 		if (SEQ_LT(ae->ack, high_seq)) {
16543 			/* Case B*/
16544 			ae->ack_val_set = ACK_BEHIND;
16545 		} else if (SEQ_GT(ae->ack, high_seq)) {
16546 			/* Case A */
16547 			ae->ack_val_set = ACK_CUMACK;
16548 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
16549 			/* Case D */
16550 			ae->ack_val_set = ACK_DUPACK;
16551 		} else {
16552 			/* Case C */
16553 			ae->ack_val_set = ACK_RWND;
16554 		}
16555 		rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16556 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
16557 		/* Validate timestamp */
16558 		if (ae->flags & HAS_TSTMP) {
16559 			/* Setup for a timestamp */
16560 			to->to_flags = TOF_TS;
16561 			ae->ts_echo -= tp->ts_offset;
16562 			to->to_tsecr = ae->ts_echo;
16563 			to->to_tsval = ae->ts_value;
16564 			/*
16565 			 * If echoed timestamp is later than the current time, fall back to
16566 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16567 			 * were used when this connection was established.
16568 			 */
16569 			if (TSTMP_GT(ae->ts_echo, ms_cts))
16570 				to->to_tsecr = 0;
16571 			if (tp->ts_recent &&
16572 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
16573 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
16574 #ifdef TCP_ACCOUNTING
16575 					rdstc = get_cyclecount();
16576 					if (rdstc > ts_val) {
16577 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16578 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16579 						}
16580 					}
16581 #endif
16582 					continue;
16583 				}
16584 			}
16585 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
16586 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
16587 				tp->ts_recent_age = tcp_ts_getticks();
16588 				tp->ts_recent = ae->ts_value;
16589 			}
16590 		} else {
16591 			/* Setup for a no options */
16592 			to->to_flags = 0;
16593 		}
16594 		/* Update the rcv time and perform idle reduction possibly */
16595 		if  (tp->t_idle_reduce &&
16596 		     (tp->snd_max == tp->snd_una) &&
16597 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16598 			counter_u64_add(rack_input_idle_reduces, 1);
16599 			rack_cc_after_idle(rack, tp);
16600 		}
16601 		tp->t_rcvtime = ticks;
16602 		/* Now what about ECN of a chain of pure ACKs? */
16603 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
16604 			tcp_packets_this_ack(tp, ae->ack),
16605 			ae->codepoint))
16606 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
16607 #ifdef TCP_ACCOUNTING
16608 		/* Count for the specific type of ack in */
16609 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16610 			tp->tcp_cnt_counters[ae->ack_val_set]++;
16611 		}
16612 #endif
16613 		/*
16614 		 * Note how we could move up these in the determination
16615 		 * above, but we don't so that way the timestamp checks (and ECN)
16616 		 * is done first before we do any processing on the ACK.
16617 		 * The non-compressed path through the code has this
16618 		 * weakness (noted by @jtl) that it actually does some
16619 		 * processing before verifying the timestamp information.
16620 		 * We don't take that path here which is why we set
16621 		 * the ack_val_set first, do the timestamp and ecn
16622 		 * processing, and then look at what we have setup.
16623 		 */
16624 		if (ae->ack_val_set == ACK_BEHIND) {
16625 			/*
16626 			 * Case B flag reordering, if window is not closed
16627 			 * or it could be a keep-alive or persists
16628 			 */
16629 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
16630 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16631 				if (rack->r_ctl.rc_reorder_ts == 0)
16632 					rack->r_ctl.rc_reorder_ts = 1;
16633 			}
16634 		} else if (ae->ack_val_set == ACK_DUPACK) {
16635 			/* Case D */
16636 			rack_strike_dupack(rack, ae->ack);
16637 		} else if (ae->ack_val_set == ACK_RWND) {
16638 			/* Case C */
16639 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16640 				ts.tv_sec = ae->timestamp / 1000000000;
16641 				ts.tv_nsec = ae->timestamp % 1000000000;
16642 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16643 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16644 			} else {
16645 				rack->r_ctl.act_rcv_time = *tv;
16646 			}
16647 			if (rack->forced_ack) {
16648 				rack_handle_probe_response(rack, tiwin,
16649 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
16650 			}
16651 #ifdef TCP_ACCOUNTING
16652 			win_up_req = 1;
16653 #endif
16654 			win_upd_ack = ae->ack;
16655 			win_seq = ae->seq;
16656 			the_win = tiwin;
16657 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16658 		} else {
16659 			/* Case A */
16660 			if (SEQ_GT(ae->ack, tp->snd_max)) {
16661 				/*
16662 				 * We just send an ack since the incoming
16663 				 * ack is beyond the largest seq we sent.
16664 				 */
16665 				if ((tp->t_flags & TF_ACKNOW) == 0) {
16666 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
16667 					if (tp->t_flags && TF_ACKNOW)
16668 						rack->r_wanted_output = 1;
16669 				}
16670 			} else {
16671 				nsegs++;
16672 				/* If the window changed setup to update */
16673 				if (tiwin != tp->snd_wnd) {
16674 					win_upd_ack = ae->ack;
16675 					win_seq = ae->seq;
16676 					the_win = tiwin;
16677 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16678 				}
16679 #ifdef TCP_ACCOUNTING
16680 				/* Account for the acks */
16681 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16682 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
16683 				}
16684 #endif
16685 				high_seq = ae->ack;
16686 				/* Setup our act_rcv_time */
16687 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16688 					ts.tv_sec = ae->timestamp / 1000000000;
16689 					ts.tv_nsec = ae->timestamp % 1000000000;
16690 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16691 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16692 				} else {
16693 					rack->r_ctl.act_rcv_time = *tv;
16694 				}
16695 				rack_process_to_cumack(tp, rack, ae->ack, cts, to,
16696 						       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
16697 #ifdef TCP_REQUEST_TRK
16698 				rack_req_check_for_comp(rack, high_seq);
16699 #endif
16700 				if (rack->rc_dsack_round_seen) {
16701 					/* Is the dsack round over? */
16702 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
16703 						/* Yes it is */
16704 						rack->rc_dsack_round_seen = 0;
16705 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
16706 					}
16707 				}
16708 			}
16709 		}
16710 		/* And lets be sure to commit the rtt measurements for this ack */
16711 		tcp_rack_xmit_timer_commit(rack, tp);
16712 #ifdef TCP_ACCOUNTING
16713 		rdstc = get_cyclecount();
16714 		if (rdstc > ts_val) {
16715 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16716 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16717 				if (ae->ack_val_set == ACK_CUMACK)
16718 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
16719 			}
16720 		}
16721 #endif
16722 	}
16723 #ifdef TCP_ACCOUNTING
16724 	ts_val = get_cyclecount();
16725 #endif
16726 	/* Tend to any collapsed window */
16727 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
16728 		/* The peer collapsed the window */
16729 		rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
16730 	} else if (rack->rc_has_collapsed)
16731 		rack_un_collapse_window(rack, __LINE__);
16732 	if ((rack->r_collapse_point_valid) &&
16733 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
16734 		rack->r_collapse_point_valid = 0;
16735 	acked_amount = acked = (high_seq - tp->snd_una);
16736 	if (acked) {
16737 		/*
16738 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
16739 		 * causes issues when we are just going app limited. Lets
16740 		 * instead use SEQ_GT <or> where its equal but more data
16741 		 * is outstanding.
16742 		 *
16743 		 * Also make sure we are on the last ack of a series. We
16744 		 * have to have all the ack's processed in queue to know
16745 		 * if there is something left outstanding.
16746 		 *
16747 		 */
16748 		if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
16749 		    (rack->rc_new_rnd_needed == 0) &&
16750 		    (nxt_pkt == 0)) {
16751 			/*
16752 			 * We have crossed into a new round with
16753 			 * this th_ack value.
16754 			 */
16755 			rack_new_round_setup(tp, rack, high_seq);
16756 		}
16757 		/*
16758 		 * Clear the probe not answered flag
16759 		 * since cum-ack moved forward.
16760 		 */
16761 		rack->probe_not_answered = 0;
16762 		if (tp->t_flags & TF_NEEDSYN) {
16763 			/*
16764 			 * T/TCP: Connection was half-synchronized, and our SYN has
16765 			 * been ACK'd (so connection is now fully synchronized).  Go
16766 			 * to non-starred state, increment snd_una for ACK of SYN,
16767 			 * and check if we can do window scaling.
16768 			 */
16769 			tp->t_flags &= ~TF_NEEDSYN;
16770 			tp->snd_una++;
16771 			acked_amount = acked = (high_seq - tp->snd_una);
16772 		}
16773 		if (acked > sbavail(&so->so_snd))
16774 			acked_amount = sbavail(&so->so_snd);
16775 		if (IN_FASTRECOVERY(tp->t_flags) &&
16776 		    (rack->rack_no_prr == 0))
16777 			rack_update_prr(tp, rack, acked_amount, high_seq);
16778 		if (IN_RECOVERY(tp->t_flags)) {
16779 			if (SEQ_LT(high_seq, tp->snd_recover) &&
16780 			    (SEQ_LT(high_seq, tp->snd_max))) {
16781 				tcp_rack_partialack(tp);
16782 			} else {
16783 				rack_post_recovery(tp, high_seq);
16784 				post_recovery = 1;
16785 			}
16786 		}  else if ((rack->rto_from_rec == 1) &&
16787 			    SEQ_GEQ(high_seq, tp->snd_recover)) {
16788 			/*
16789 			 * We were in recovery, hit a rxt timeout
16790 			 * and never re-entered recovery. The timeout(s)
16791 			 * made up all the lost data. In such a case
16792 			 * we need to clear the rto_from_rec flag.
16793 			 */
16794 			rack->rto_from_rec = 0;
16795 		}
16796 		/* Handle the rack-log-ack part (sendmap) */
16797 		if ((sbused(&so->so_snd) == 0) &&
16798 		    (acked > acked_amount) &&
16799 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16800 		    (tp->t_flags & TF_SENTFIN)) {
16801 			/*
16802 			 * We must be sure our fin
16803 			 * was sent and acked (we can be
16804 			 * in FIN_WAIT_1 without having
16805 			 * sent the fin).
16806 			 */
16807 			ourfinisacked = 1;
16808 			/*
16809 			 * Lets make sure snd_una is updated
16810 			 * since most likely acked_amount = 0 (it
16811 			 * should be).
16812 			 */
16813 			tp->snd_una = high_seq;
16814 		}
16815 		/* Did we make a RTO error? */
16816 		if ((tp->t_flags & TF_PREVVALID) &&
16817 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16818 			tp->t_flags &= ~TF_PREVVALID;
16819 			if (tp->t_rxtshift == 1 &&
16820 			    (int)(ticks - tp->t_badrxtwin) < 0)
16821 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16822 		}
16823 		/* Handle the data in the socket buffer */
16824 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16825 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16826 		if (acked_amount > 0) {
16827 			uint32_t p_cwnd;
16828 			struct mbuf *mfree;
16829 
16830 			if (post_recovery) {
16831 				/*
16832 				 * Grab the segsiz, multiply by 2 and add the snd_cwnd
16833 				 * that is the max the CC should add if we are exiting
16834 				 * recovery and doing a late add.
16835 				 */
16836 				p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16837 				p_cwnd <<= 1;
16838 				p_cwnd += tp->snd_cwnd;
16839 			}
16840 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, post_recovery);
16841 			if (post_recovery && (tp->snd_cwnd > p_cwnd)) {
16842 				/* Must be non-newreno (cubic) getting too ahead of itself */
16843 				tp->snd_cwnd = p_cwnd;
16844 			}
16845 			SOCKBUF_LOCK(&so->so_snd);
16846 			mfree = sbcut_locked(&so->so_snd, acked_amount);
16847 			tp->snd_una = high_seq;
16848 			/* Note we want to hold the sb lock through the sendmap adjust */
16849 			rack_adjust_sendmap_head(rack, &so->so_snd);
16850 			/* Wake up the socket if we have room to write more */
16851 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16852 			sowwakeup_locked(so);
16853 			m_freem(mfree);
16854 		}
16855 		/* update progress */
16856 		tp->t_acktime = ticks;
16857 		rack_log_progress_event(rack, tp, tp->t_acktime,
16858 					PROGRESS_UPDATE, __LINE__);
16859 		/* Clear out shifts and such */
16860 		tp->t_rxtshift = 0;
16861 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16862 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16863 		rack->rc_tlp_in_progress = 0;
16864 		rack->r_ctl.rc_tlp_cnt_out = 0;
16865 		/* Send recover and snd_nxt must be dragged along */
16866 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
16867 			tp->snd_recover = tp->snd_una;
16868 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
16869 			tp->snd_nxt = tp->snd_max;
16870 		/*
16871 		 * If the RXT timer is running we want to
16872 		 * stop it, so we can restart a TLP (or new RXT).
16873 		 */
16874 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16875 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16876 		tp->snd_wl2 = high_seq;
16877 		tp->t_dupacks = 0;
16878 		if (under_pacing &&
16879 		    (rack->use_fixed_rate == 0) &&
16880 		    (rack->in_probe_rtt == 0) &&
16881 		    rack->rc_gp_dyn_mul &&
16882 		    rack->rc_always_pace) {
16883 			/* Check if we are dragging bottom */
16884 			rack_check_bottom_drag(tp, rack, so);
16885 		}
16886 		if (tp->snd_una == tp->snd_max) {
16887 			tp->t_flags &= ~TF_PREVVALID;
16888 			rack->r_ctl.retran_during_recovery = 0;
16889 			rack->rc_suspicious = 0;
16890 			rack->r_ctl.dsack_byte_cnt = 0;
16891 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16892 			if (rack->r_ctl.rc_went_idle_time == 0)
16893 				rack->r_ctl.rc_went_idle_time = 1;
16894 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16895 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
16896 				tp->t_acktime = 0;
16897 			/* Set so we might enter persists... */
16898 			rack->r_wanted_output = 1;
16899 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16900 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16901 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16902 			    (sbavail(&so->so_snd) == 0) &&
16903 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16904 				/*
16905 				 * The socket was gone and the
16906 				 * peer sent data (not now in the past), time to
16907 				 * reset him.
16908 				 */
16909 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16910 				/* tcp_close will kill the inp pre-log the Reset */
16911 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16912 #ifdef TCP_ACCOUNTING
16913 				rdstc = get_cyclecount();
16914 				if (rdstc > ts_val) {
16915 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16916 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16917 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16918 					}
16919 				}
16920 #endif
16921 				m_freem(m);
16922 				tp = tcp_close(tp);
16923 				if (tp == NULL) {
16924 #ifdef TCP_ACCOUNTING
16925 					sched_unpin();
16926 #endif
16927 					return (1);
16928 				}
16929 				/*
16930 				 * We would normally do drop-with-reset which would
16931 				 * send back a reset. We can't since we don't have
16932 				 * all the needed bits. Instead lets arrange for
16933 				 * a call to tcp_output(). That way since we
16934 				 * are in the closed state we will generate a reset.
16935 				 *
16936 				 * Note if tcp_accounting is on we don't unpin since
16937 				 * we do that after the goto label.
16938 				 */
16939 				goto send_out_a_rst;
16940 			}
16941 			if ((sbused(&so->so_snd) == 0) &&
16942 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16943 			    (tp->t_flags & TF_SENTFIN)) {
16944 				/*
16945 				 * If we can't receive any more data, then closing user can
16946 				 * proceed. Starting the timer is contrary to the
16947 				 * specification, but if we don't get a FIN we'll hang
16948 				 * forever.
16949 				 *
16950 				 */
16951 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16952 					soisdisconnected(so);
16953 					tcp_timer_activate(tp, TT_2MSL,
16954 							   (tcp_fast_finwait2_recycle ?
16955 							    tcp_finwait2_timeout :
16956 							    TP_MAXIDLE(tp)));
16957 				}
16958 				if (ourfinisacked == 0) {
16959 					/*
16960 					 * We don't change to fin-wait-2 if we have our fin acked
16961 					 * which means we are probably in TCPS_CLOSING.
16962 					 */
16963 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
16964 				}
16965 			}
16966 		}
16967 		/* Wake up the socket if we have room to write more */
16968 		if (sbavail(&so->so_snd)) {
16969 			rack->r_wanted_output = 1;
16970 			if (ctf_progress_timeout_check(tp, true)) {
16971 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16972 							tp, tick, PROGRESS_DROP, __LINE__);
16973 				/*
16974 				 * We cheat here and don't send a RST, we should send one
16975 				 * when the pacer drops the connection.
16976 				 */
16977 #ifdef TCP_ACCOUNTING
16978 				rdstc = get_cyclecount();
16979 				if (rdstc > ts_val) {
16980 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16981 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16982 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16983 					}
16984 				}
16985 				sched_unpin();
16986 #endif
16987 				(void)tcp_drop(tp, ETIMEDOUT);
16988 				m_freem(m);
16989 				return (1);
16990 			}
16991 		}
16992 		if (ourfinisacked) {
16993 			switch(tp->t_state) {
16994 			case TCPS_CLOSING:
16995 #ifdef TCP_ACCOUNTING
16996 				rdstc = get_cyclecount();
16997 				if (rdstc > ts_val) {
16998 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16999 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17000 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17001 					}
17002 				}
17003 				sched_unpin();
17004 #endif
17005 				tcp_twstart(tp);
17006 				m_freem(m);
17007 				return (1);
17008 				break;
17009 			case TCPS_LAST_ACK:
17010 #ifdef TCP_ACCOUNTING
17011 				rdstc = get_cyclecount();
17012 				if (rdstc > ts_val) {
17013 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17014 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17015 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17016 					}
17017 				}
17018 				sched_unpin();
17019 #endif
17020 				tp = tcp_close(tp);
17021 				ctf_do_drop(m, tp);
17022 				return (1);
17023 				break;
17024 			case TCPS_FIN_WAIT_1:
17025 #ifdef TCP_ACCOUNTING
17026 				rdstc = get_cyclecount();
17027 				if (rdstc > ts_val) {
17028 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17029 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17030 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17031 					}
17032 				}
17033 #endif
17034 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
17035 					soisdisconnected(so);
17036 					tcp_timer_activate(tp, TT_2MSL,
17037 							   (tcp_fast_finwait2_recycle ?
17038 							    tcp_finwait2_timeout :
17039 							    TP_MAXIDLE(tp)));
17040 				}
17041 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
17042 				break;
17043 			default:
17044 				break;
17045 			}
17046 		}
17047 		if (rack->r_fast_output) {
17048 			/*
17049 			 * We re doing fast output.. can we expand that?
17050 			 */
17051 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
17052 		}
17053 #ifdef TCP_ACCOUNTING
17054 		rdstc = get_cyclecount();
17055 		if (rdstc > ts_val) {
17056 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17057 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17058 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17059 			}
17060 		}
17061 
17062 	} else if (win_up_req) {
17063 		rdstc = get_cyclecount();
17064 		if (rdstc > ts_val) {
17065 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17066 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
17067 			}
17068 		}
17069 #endif
17070 	}
17071 	/* Now is there a next packet, if so we are done */
17072 	m_freem(m);
17073 	did_out = 0;
17074 	if (nxt_pkt) {
17075 #ifdef TCP_ACCOUNTING
17076 		sched_unpin();
17077 #endif
17078 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
17079 		return (0);
17080 	}
17081 	rack_handle_might_revert(tp, rack);
17082 	ctf_calc_rwin(so, tp);
17083 	if ((rack->r_wanted_output != 0) ||
17084 	    (rack->r_fast_output != 0) ||
17085 	    (tp->t_flags & TF_ACKNOW )) {
17086 	send_out_a_rst:
17087 		if (tcp_output(tp) < 0) {
17088 #ifdef TCP_ACCOUNTING
17089 			sched_unpin();
17090 #endif
17091 			return (1);
17092 		}
17093 		did_out = 1;
17094 	}
17095 	if (tp->t_flags2 & TF2_HPTS_CALLS)
17096 		tp->t_flags2 &= ~TF2_HPTS_CALLS;
17097 	rack_free_trim(rack);
17098 #ifdef TCP_ACCOUNTING
17099 	sched_unpin();
17100 #endif
17101 	rack_timer_audit(tp, rack, &so->so_snd);
17102 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
17103 	return (0);
17104 }
17105 
17106 #define	TCP_LRO_TS_OPTION \
17107     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
17108 	  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
17109 
17110 static int
17111 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17112     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
17113     struct timeval *tv)
17114 {
17115 	struct inpcb *inp = tptoinpcb(tp);
17116 	struct socket *so = tptosocket(tp);
17117 #ifdef TCP_ACCOUNTING
17118 	uint64_t ts_val;
17119 #endif
17120 	int32_t thflags, retval, did_out = 0;
17121 	int32_t way_out = 0;
17122 	/*
17123 	 * cts - is the current time from tv (caller gets ts) in microseconds.
17124 	 * ms_cts - is the current time from tv in milliseconds.
17125 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
17126 	 */
17127 	uint32_t cts, us_cts, ms_cts;
17128 	uint32_t tiwin;
17129 	struct timespec ts;
17130 	struct tcpopt to;
17131 	struct tcp_rack *rack;
17132 	struct rack_sendmap *rsm;
17133 	int32_t prev_state = 0;
17134 	int no_output = 0;
17135 	int slot_remaining = 0;
17136 #ifdef TCP_ACCOUNTING
17137 	int ack_val_set = 0xf;
17138 #endif
17139 	int nsegs;
17140 
17141 	NET_EPOCH_ASSERT();
17142 	INP_WLOCK_ASSERT(inp);
17143 
17144 	/*
17145 	 * tv passed from common code is from either M_TSTMP_LRO or
17146 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
17147 	 */
17148 	rack = (struct tcp_rack *)tp->t_fb_ptr;
17149 	if (rack->rack_deferred_inited == 0) {
17150 		/*
17151 		 * If we are the connecting socket we will
17152 		 * hit rack_init() when no sequence numbers
17153 		 * are setup. This makes it so we must defer
17154 		 * some initialization. Call that now.
17155 		 */
17156 		rack_deferred_init(tp, rack);
17157 	}
17158 	/*
17159 	 * Check to see if we need to skip any output plans. This
17160 	 * can happen in the non-LRO path where we are pacing and
17161 	 * must process the ack coming in but need to defer sending
17162 	 * anything becase a pacing timer is running.
17163 	 */
17164 	us_cts = tcp_tv_to_usectick(tv);
17165 	if (m->m_flags & M_ACKCMP) {
17166 		/*
17167 		 * All compressed ack's are ack's by definition so
17168 		 * remove any ack required flag and then do the processing.
17169 		 */
17170 		rack->rc_ack_required = 0;
17171 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
17172 	}
17173 	thflags = tcp_get_flags(th);
17174 	if ((rack->rc_always_pace == 1) &&
17175 	    (rack->rc_ack_can_sendout_data == 0) &&
17176 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
17177 	    (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
17178 		/*
17179 		 * Ok conditions are right for queuing the packets
17180 		 * but we do have to check the flags in the inp, it
17181 		 * could be, if a sack is present, we want to be awoken and
17182 		 * so should process the packets.
17183 		 */
17184 		slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
17185 		if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
17186 			no_output = 1;
17187 		} else {
17188 			/*
17189 			 * If there is no options, or just a
17190 			 * timestamp option, we will want to queue
17191 			 * the packets. This is the same that LRO does
17192 			 * and will need to change with accurate ECN.
17193 			 */
17194 			uint32_t *ts_ptr;
17195 			int optlen;
17196 
17197 			optlen = (th->th_off << 2) - sizeof(struct tcphdr);
17198 			ts_ptr = (uint32_t *)(th + 1);
17199 			if ((optlen == 0) ||
17200 			    ((optlen == TCPOLEN_TSTAMP_APPA) &&
17201 			     (*ts_ptr == TCP_LRO_TS_OPTION)))
17202 				no_output = 1;
17203 		}
17204 		if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
17205 			/*
17206 			 * It is unrealistic to think we can pace in less than
17207 			 * the minimum granularity of the pacer (def:250usec). So
17208 			 * if we have less than that time remaining we should go
17209 			 * ahead and allow output to be "early". We will attempt to
17210 			 * make up for it in any pacing time we try to apply on
17211 			 * the outbound packet.
17212 			 */
17213 			no_output = 0;
17214 		}
17215 	}
17216 	/*
17217 	 * If there is a RST or FIN lets dump out the bw
17218 	 * with a FIN the connection may go on but we
17219 	 * may not.
17220 	 */
17221 	if ((thflags & TH_FIN) || (thflags & TH_RST))
17222 		rack_log_pacing_delay_calc(rack,
17223 					   rack->r_ctl.gp_bw,
17224 					   0,
17225 					   0,
17226 					   rack_get_gp_est(rack), /* delRate */
17227 					   rack_get_lt_bw(rack), /* rttProp */
17228 					   20, __LINE__, NULL, 0);
17229 	if (m->m_flags & M_ACKCMP) {
17230 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
17231 	}
17232 	cts = tcp_tv_to_usectick(tv);
17233 	ms_cts =  tcp_tv_to_mssectick(tv);
17234 	nsegs = m->m_pkthdr.lro_nsegs;
17235 	counter_u64_add(rack_proc_non_comp_ack, 1);
17236 #ifdef TCP_ACCOUNTING
17237 	sched_pin();
17238 	if (thflags & TH_ACK)
17239 		ts_val = get_cyclecount();
17240 #endif
17241 	if ((m->m_flags & M_TSTMP) ||
17242 	    (m->m_flags & M_TSTMP_LRO)) {
17243 		mbuf_tstmp2timespec(m, &ts);
17244 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
17245 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
17246 	} else
17247 		rack->r_ctl.act_rcv_time = *tv;
17248 	kern_prefetch(rack, &prev_state);
17249 	prev_state = 0;
17250 	/*
17251 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
17252 	 * the scale is zero.
17253 	 */
17254 	tiwin = th->th_win << tp->snd_scale;
17255 #ifdef TCP_ACCOUNTING
17256 	if (thflags & TH_ACK) {
17257 		/*
17258 		 * We have a tradeoff here. We can either do what we are
17259 		 * doing i.e. pinning to this CPU and then doing the accounting
17260 		 * <or> we could do a critical enter, setup the rdtsc and cpu
17261 		 * as in below, and then validate we are on the same CPU on
17262 		 * exit. I have choosen to not do the critical enter since
17263 		 * that often will gain you a context switch, and instead lock
17264 		 * us (line above this if) to the same CPU with sched_pin(). This
17265 		 * means we may be context switched out for a higher priority
17266 		 * interupt but we won't be moved to another CPU.
17267 		 *
17268 		 * If this occurs (which it won't very often since we most likely
17269 		 * are running this code in interupt context and only a higher
17270 		 * priority will bump us ... clock?) we will falsely add in
17271 		 * to the time the interupt processing time plus the ack processing
17272 		 * time. This is ok since its a rare event.
17273 		 */
17274 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
17275 						    ctf_fixed_maxseg(tp));
17276 	}
17277 #endif
17278 	/*
17279 	 * Parse options on any incoming segment.
17280 	 */
17281 	memset(&to, 0, sizeof(to));
17282 	tcp_dooptions(&to, (u_char *)(th + 1),
17283 	    (th->th_off << 2) - sizeof(struct tcphdr),
17284 	    (thflags & TH_SYN) ? TO_SYN : 0);
17285 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
17286 	    __func__));
17287 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
17288 	    __func__));
17289 	if (tp->t_flags2 & TF2_PROC_SACK_PROHIBIT) {
17290 		/*
17291 		 * We don't look at sack's from the
17292 		 * peer because the MSS is too small which
17293 		 * can subject us to an attack.
17294 		 */
17295 		to.to_flags &= ~TOF_SACK;
17296 	}
17297 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
17298 	    (tp->t_flags & TF_GPUTINPROG)) {
17299 		/*
17300 		 * We have a goodput in progress
17301 		 * and we have entered a late state.
17302 		 * Do we have enough data in the sb
17303 		 * to handle the GPUT request?
17304 		 */
17305 		uint32_t bytes;
17306 
17307 		bytes = tp->gput_ack - tp->gput_seq;
17308 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
17309 			bytes += tp->gput_seq - tp->snd_una;
17310 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
17311 			/*
17312 			 * There are not enough bytes in the socket
17313 			 * buffer that have been sent to cover this
17314 			 * measurement. Cancel it.
17315 			 */
17316 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17317 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
17318 						   tp->gput_seq,
17319 						   0, 0, 18, __LINE__, NULL, 0);
17320 			tp->t_flags &= ~TF_GPUTINPROG;
17321 		}
17322 	}
17323 	if (tcp_bblogging_on(rack->rc_tp)) {
17324 		union tcp_log_stackspecific log;
17325 		struct timeval ltv;
17326 #ifdef TCP_REQUEST_TRK
17327 		struct tcp_sendfile_track *tcp_req;
17328 
17329 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
17330 			tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
17331 		} else {
17332 			tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
17333 		}
17334 #endif
17335 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17336 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
17337 		if (rack->rack_no_prr == 0)
17338 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
17339 		else
17340 			log.u_bbr.flex1 = 0;
17341 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
17342 		log.u_bbr.use_lt_bw <<= 1;
17343 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
17344 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
17345 		log.u_bbr.bbr_state = rack->rc_free_cnt;
17346 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17347 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
17348 		log.u_bbr.flex3 = m->m_flags;
17349 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
17350 		log.u_bbr.lost = thflags;
17351 		log.u_bbr.pacing_gain = 0x1;
17352 #ifdef TCP_ACCOUNTING
17353 		log.u_bbr.cwnd_gain = ack_val_set;
17354 #endif
17355 		log.u_bbr.flex7 = 2;
17356 		if (m->m_flags & M_TSTMP) {
17357 			/* Record the hardware timestamp if present */
17358 			mbuf_tstmp2timespec(m, &ts);
17359 			ltv.tv_sec = ts.tv_sec;
17360 			ltv.tv_usec = ts.tv_nsec / 1000;
17361 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
17362 		} else if (m->m_flags & M_TSTMP_LRO) {
17363 			/* Record the LRO the arrival timestamp */
17364 			mbuf_tstmp2timespec(m, &ts);
17365 			ltv.tv_sec = ts.tv_sec;
17366 			ltv.tv_usec = ts.tv_nsec / 1000;
17367 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
17368 		}
17369 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
17370 		/* Log the rcv time */
17371 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
17372 #ifdef TCP_REQUEST_TRK
17373 		log.u_bbr.applimited = tp->t_tcpreq_closed;
17374 		log.u_bbr.applimited <<= 8;
17375 		log.u_bbr.applimited |= tp->t_tcpreq_open;
17376 		log.u_bbr.applimited <<= 8;
17377 		log.u_bbr.applimited |= tp->t_tcpreq_req;
17378 		if (tcp_req) {
17379 			/* Copy out any client req info */
17380 			/* seconds */
17381 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
17382 			/* useconds */
17383 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
17384 			log.u_bbr.rttProp = tcp_req->timestamp;
17385 			log.u_bbr.cur_del_rate = tcp_req->start;
17386 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
17387 				log.u_bbr.flex8 |= 1;
17388 			} else {
17389 				log.u_bbr.flex8 |= 2;
17390 				log.u_bbr.bw_inuse = tcp_req->end;
17391 			}
17392 			log.u_bbr.flex6 = tcp_req->start_seq;
17393 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
17394 				log.u_bbr.flex8 |= 4;
17395 				log.u_bbr.epoch = tcp_req->end_seq;
17396 			}
17397 		}
17398 #endif
17399 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
17400 		    tlen, &log, true, &ltv);
17401 	}
17402 	/* Remove ack required flag if set, we have one  */
17403 	if (thflags & TH_ACK)
17404 		rack->rc_ack_required = 0;
17405 	rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
17406 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
17407 		way_out = 4;
17408 		retval = 0;
17409 		m_freem(m);
17410 		goto done_with_input;
17411 	}
17412 	/*
17413 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
17414 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
17415 	 */
17416 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
17417 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
17418 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
17419 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
17420 #ifdef TCP_ACCOUNTING
17421 		sched_unpin();
17422 #endif
17423 		return (1);
17424 	}
17425 	/*
17426 	 * If timestamps were negotiated during SYN/ACK and a
17427 	 * segment without a timestamp is received, silently drop
17428 	 * the segment, unless it is a RST segment or missing timestamps are
17429 	 * tolerated.
17430 	 * See section 3.2 of RFC 7323.
17431 	 */
17432 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
17433 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
17434 		way_out = 5;
17435 		retval = 0;
17436 		m_freem(m);
17437 		goto done_with_input;
17438 	}
17439 	/*
17440 	 * Segment received on connection. Reset idle time and keep-alive
17441 	 * timer. XXX: This should be done after segment validation to
17442 	 * ignore broken/spoofed segs.
17443 	 */
17444 	if  (tp->t_idle_reduce &&
17445 	     (tp->snd_max == tp->snd_una) &&
17446 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
17447 		counter_u64_add(rack_input_idle_reduces, 1);
17448 		rack_cc_after_idle(rack, tp);
17449 	}
17450 	tp->t_rcvtime = ticks;
17451 #ifdef STATS
17452 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
17453 #endif
17454 	if (tiwin > rack->r_ctl.rc_high_rwnd)
17455 		rack->r_ctl.rc_high_rwnd = tiwin;
17456 	/*
17457 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
17458 	 * this to occur after we've validated the segment.
17459 	 */
17460 	if (tcp_ecn_input_segment(tp, thflags, tlen,
17461 	    tcp_packets_this_ack(tp, th->th_ack),
17462 	    iptos))
17463 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
17464 
17465 	/*
17466 	 * If echoed timestamp is later than the current time, fall back to
17467 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
17468 	 * were used when this connection was established.
17469 	 */
17470 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
17471 		to.to_tsecr -= tp->ts_offset;
17472 		if (TSTMP_GT(to.to_tsecr, ms_cts))
17473 			to.to_tsecr = 0;
17474 	}
17475 	if ((rack->r_rcvpath_rtt_up == 1) &&
17476 	    (to.to_flags & TOF_TS) &&
17477 	    (TSTMP_GEQ(to.to_tsecr, rack->r_ctl.last_rcv_tstmp_for_rtt))) {
17478 		uint32_t rtt = 0;
17479 
17480 		/*
17481 		 * We are receiving only and thus not sending
17482 		 * data to do an RTT. We set a flag when we first
17483 		 * sent this TS to the peer. We now have it back
17484 		 * and have an RTT to share. We log it as a conf
17485 		 * 4, we are not so sure about it.. since we
17486 		 * may have lost an ack.
17487 		 */
17488 		if (TSTMP_GT(cts, rack->r_ctl.last_time_of_arm_rcv))
17489 		    rtt = (cts - rack->r_ctl.last_time_of_arm_rcv);
17490 		rack->r_rcvpath_rtt_up = 0;
17491 		/* Submit and commit the timer */
17492 		if (rtt > 0) {
17493 			tcp_rack_xmit_timer(rack, rtt, 0, rtt, 4, NULL, 1);
17494 			tcp_rack_xmit_timer_commit(rack, tp);
17495 		}
17496 	}
17497 	/*
17498 	 * If its the first time in we need to take care of options and
17499 	 * verify we can do SACK for rack!
17500 	 */
17501 	if (rack->r_state == 0) {
17502 		/* Should be init'd by rack_init() */
17503 		KASSERT(rack->rc_inp != NULL,
17504 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
17505 		if (rack->rc_inp == NULL) {
17506 			rack->rc_inp = inp;
17507 		}
17508 
17509 		/*
17510 		 * Process options only when we get SYN/ACK back. The SYN
17511 		 * case for incoming connections is handled in tcp_syncache.
17512 		 * According to RFC1323 the window field in a SYN (i.e., a
17513 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
17514 		 * this is traditional behavior, may need to be cleaned up.
17515 		 */
17516 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
17517 			/* Handle parallel SYN for ECN */
17518 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
17519 			if ((to.to_flags & TOF_SCALE) &&
17520 			    (tp->t_flags & TF_REQ_SCALE)) {
17521 				tp->t_flags |= TF_RCVD_SCALE;
17522 				tp->snd_scale = to.to_wscale;
17523 			} else
17524 				tp->t_flags &= ~TF_REQ_SCALE;
17525 			/*
17526 			 * Initial send window.  It will be updated with the
17527 			 * next incoming segment to the scaled value.
17528 			 */
17529 			tp->snd_wnd = th->th_win;
17530 			rack_validate_fo_sendwin_up(tp, rack);
17531 			if ((to.to_flags & TOF_TS) &&
17532 			    (tp->t_flags & TF_REQ_TSTMP)) {
17533 				tp->t_flags |= TF_RCVD_TSTMP;
17534 				tp->ts_recent = to.to_tsval;
17535 				tp->ts_recent_age = cts;
17536 			} else
17537 				tp->t_flags &= ~TF_REQ_TSTMP;
17538 			if (to.to_flags & TOF_MSS) {
17539 				tcp_mss(tp, to.to_mss);
17540 			}
17541 			if ((tp->t_flags & TF_SACK_PERMIT) &&
17542 			    (to.to_flags & TOF_SACKPERM) == 0)
17543 				tp->t_flags &= ~TF_SACK_PERMIT;
17544 			if (tp->t_flags & TF_FASTOPEN) {
17545 				if (to.to_flags & TOF_FASTOPEN) {
17546 					uint16_t mss;
17547 
17548 					if (to.to_flags & TOF_MSS)
17549 						mss = to.to_mss;
17550 					else
17551 						if ((inp->inp_vflag & INP_IPV6) != 0)
17552 							mss = TCP6_MSS;
17553 						else
17554 							mss = TCP_MSS;
17555 					tcp_fastopen_update_cache(tp, mss,
17556 					    to.to_tfo_len, to.to_tfo_cookie);
17557 				} else
17558 					tcp_fastopen_disable_path(tp);
17559 			}
17560 		}
17561 		/*
17562 		 * At this point we are at the initial call. Here we decide
17563 		 * if we are doing RACK or not. We do this by seeing if
17564 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
17565 		 * The code now does do dup-ack counting so if you don't
17566 		 * switch back you won't get rack & TLP, but you will still
17567 		 * get this stack.
17568 		 */
17569 
17570 		if ((rack_sack_not_required == 0) &&
17571 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
17572 			tcp_switch_back_to_default(tp);
17573 			(*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
17574 			    tlen, iptos);
17575 #ifdef TCP_ACCOUNTING
17576 			sched_unpin();
17577 #endif
17578 			return (1);
17579 		}
17580 		tcp_set_hpts(tp);
17581 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
17582 	}
17583 	if (thflags & TH_FIN)
17584 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
17585 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
17586 	if ((rack->rc_gp_dyn_mul) &&
17587 	    (rack->use_fixed_rate == 0) &&
17588 	    (rack->rc_always_pace)) {
17589 		/* Check in on probertt */
17590 		rack_check_probe_rtt(rack, cts);
17591 	}
17592 	rack_clear_rate_sample(rack);
17593 	if ((rack->forced_ack) &&
17594 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
17595 		rack_handle_probe_response(rack, tiwin, us_cts);
17596 	}
17597 	/*
17598 	 * This is the one exception case where we set the rack state
17599 	 * always. All other times (timers etc) we must have a rack-state
17600 	 * set (so we assure we have done the checks above for SACK).
17601 	 */
17602 	rack->r_ctl.rc_rcvtime = cts;
17603 	if (rack->r_state != tp->t_state)
17604 		rack_set_state(tp, rack);
17605 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
17606 	    (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
17607 		kern_prefetch(rsm, &prev_state);
17608 	prev_state = rack->r_state;
17609 	if ((thflags & TH_RST) &&
17610 	    ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
17611 	      SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
17612 	     (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
17613 		/* The connection will be killed by a reset check the tracepoint */
17614 		tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
17615 	}
17616 	retval = (*rack->r_substate) (m, th, so,
17617 	    tp, &to, drop_hdrlen,
17618 	    tlen, tiwin, thflags, nxt_pkt, iptos);
17619 	if (retval == 0) {
17620 		/*
17621 		 * If retval is 1 the tcb is unlocked and most likely the tp
17622 		 * is gone.
17623 		 */
17624 		INP_WLOCK_ASSERT(inp);
17625 		if ((rack->rc_gp_dyn_mul) &&
17626 		    (rack->rc_always_pace) &&
17627 		    (rack->use_fixed_rate == 0) &&
17628 		    rack->in_probe_rtt &&
17629 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
17630 			/*
17631 			 * If we are going for target, lets recheck before
17632 			 * we output.
17633 			 */
17634 			rack_check_probe_rtt(rack, cts);
17635 		}
17636 		if (rack->set_pacing_done_a_iw == 0) {
17637 			/* How much has been acked? */
17638 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
17639 				/* We have enough to set in the pacing segment size */
17640 				rack->set_pacing_done_a_iw = 1;
17641 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
17642 			}
17643 		}
17644 		tcp_rack_xmit_timer_commit(rack, tp);
17645 #ifdef TCP_ACCOUNTING
17646 		/*
17647 		 * If we set the ack_val_se to what ack processing we are doing
17648 		 * we also want to track how many cycles we burned. Note
17649 		 * the bits after tcp_output we let be "free". This is because
17650 		 * we are also tracking the tcp_output times as well. Note the
17651 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
17652 		 * 0xf cannot be returned and is what we initialize it too to
17653 		 * indicate we are not doing the tabulations.
17654 		 */
17655 		if (ack_val_set != 0xf) {
17656 			uint64_t crtsc;
17657 
17658 			crtsc = get_cyclecount();
17659 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17660 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
17661 			}
17662 		}
17663 #endif
17664 		if ((nxt_pkt == 0) && (no_output == 0)) {
17665 			if ((rack->r_wanted_output != 0) ||
17666 			    (tp->t_flags & TF_ACKNOW) ||
17667 			    (rack->r_fast_output != 0)) {
17668 
17669 do_output_now:
17670 				if (tcp_output(tp) < 0) {
17671 #ifdef TCP_ACCOUNTING
17672 					sched_unpin();
17673 #endif
17674 					return (1);
17675 				}
17676 				did_out = 1;
17677 			}
17678 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
17679 			rack_free_trim(rack);
17680 		} else if ((nxt_pkt == 0) && (tp->t_flags & TF_ACKNOW)) {
17681 			goto do_output_now;
17682 		} else if ((no_output == 1) &&
17683 			   (nxt_pkt == 0)  &&
17684 			   (tcp_in_hpts(rack->rc_tp) == 0)) {
17685 			/*
17686 			 * We are not in hpts and we had a pacing timer up. Use
17687 			 * the remaining time (slot_remaining) to restart the timer.
17688 			 */
17689 			KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
17690 			rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
17691 			rack_free_trim(rack);
17692 		}
17693 		/* Clear the flag, it may have been cleared by output but we may not have  */
17694 		if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
17695 			tp->t_flags2 &= ~TF2_HPTS_CALLS;
17696 		/*
17697 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
17698 		 * causes issues when we are just going app limited. Lets
17699 		 * instead use SEQ_GT <or> where its equal but more data
17700 		 * is outstanding.
17701 		 *
17702 		 * Also make sure we are on the last ack of a series. We
17703 		 * have to have all the ack's processed in queue to know
17704 		 * if there is something left outstanding.
17705 		 */
17706 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
17707 		    (rack->rc_new_rnd_needed == 0) &&
17708 		    (nxt_pkt == 0)) {
17709 			/*
17710 			 * We have crossed into a new round with
17711 			 * the new snd_unae.
17712 			 */
17713 			rack_new_round_setup(tp, rack, tp->snd_una);
17714 		}
17715 		if ((nxt_pkt == 0) &&
17716 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
17717 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
17718 		     (tp->t_flags & TF_DELACK) ||
17719 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
17720 		      (tp->t_state <= TCPS_CLOSING)))) {
17721 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
17722 			if ((tp->snd_max == tp->snd_una) &&
17723 			    ((tp->t_flags & TF_DELACK) == 0) &&
17724 			    (tcp_in_hpts(rack->rc_tp)) &&
17725 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
17726 				/* keep alive not needed if we are hptsi output yet */
17727 				;
17728 			} else {
17729 				int late = 0;
17730 				if (tcp_in_hpts(tp)) {
17731 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
17732 						us_cts = tcp_get_usecs(NULL);
17733 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
17734 							rack->r_early = 1;
17735 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
17736 						} else
17737 							late = 1;
17738 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
17739 					}
17740 					tcp_hpts_remove(tp);
17741 				}
17742 				if (late && (did_out == 0)) {
17743 					/*
17744 					 * We are late in the sending
17745 					 * and we did not call the output
17746 					 * (this probably should not happen).
17747 					 */
17748 					goto do_output_now;
17749 				}
17750 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
17751 			}
17752 			way_out = 1;
17753 		} else if (nxt_pkt == 0) {
17754 			/* Do we have the correct timer running? */
17755 			rack_timer_audit(tp, rack, &so->so_snd);
17756 			way_out = 2;
17757 		}
17758 	done_with_input:
17759 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
17760 		if (did_out)
17761 			rack->r_wanted_output = 0;
17762 	}
17763 
17764 #ifdef TCP_ACCOUNTING
17765 	sched_unpin();
17766 #endif
17767 	return (retval);
17768 }
17769 
17770 static void
17771 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17772     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17773 {
17774 	struct timeval tv;
17775 
17776 	/* First lets see if we have old packets */
17777 	if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17778 		if (ctf_do_queued_segments(tp, 1)) {
17779 			m_freem(m);
17780 			return;
17781 		}
17782 	}
17783 	if (m->m_flags & M_TSTMP_LRO) {
17784 		mbuf_tstmp2timeval(m, &tv);
17785 	} else {
17786 		/* Should not be should we kassert instead? */
17787 		tcp_get_usecs(&tv);
17788 	}
17789 	if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17790 	    &tv) == 0) {
17791 		INP_WUNLOCK(tptoinpcb(tp));
17792 	}
17793 }
17794 
17795 struct rack_sendmap *
17796 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17797 {
17798 	struct rack_sendmap *rsm = NULL;
17799 	int32_t idx;
17800 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
17801 
17802 	/* Return the next guy to be re-transmitted */
17803 	if (tqhash_empty(rack->r_ctl.tqh)) {
17804 		return (NULL);
17805 	}
17806 	if (tp->t_flags & TF_SENTFIN) {
17807 		/* retran the end FIN? */
17808 		return (NULL);
17809 	}
17810 	/* ok lets look at this one */
17811 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17812 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17813 		return (rsm);
17814 	}
17815 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17816 		goto check_it;
17817 	}
17818 	rsm = rack_find_lowest_rsm(rack);
17819 	if (rsm == NULL) {
17820 		return (NULL);
17821 	}
17822 check_it:
17823 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
17824 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17825 		/*
17826 		 * No sack so we automatically do the 3 strikes and
17827 		 * retransmit (no rack timer would be started).
17828 		 */
17829 		return (rsm);
17830 	}
17831 	if (rsm->r_flags & RACK_ACKED) {
17832 		return (NULL);
17833 	}
17834 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17835 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17836 		/* Its not yet ready */
17837 		return (NULL);
17838 	}
17839 	srtt = rack_grab_rtt(tp, rack);
17840 	idx = rsm->r_rtr_cnt - 1;
17841 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17842 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
17843 	if ((tsused == ts_low) ||
17844 	    (TSTMP_LT(tsused, ts_low))) {
17845 		/* No time since sending */
17846 		return (NULL);
17847 	}
17848 	if ((tsused - ts_low) < thresh) {
17849 		/* It has not been long enough yet */
17850 		return (NULL);
17851 	}
17852 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17853 	    ((rsm->r_flags & RACK_SACK_PASSED))) {
17854 		/*
17855 		 * We have passed the dup-ack threshold <or>
17856 		 * a SACK has indicated this is missing.
17857 		 * Note that if you are a declared attacker
17858 		 * it is only the dup-ack threshold that
17859 		 * will cause retransmits.
17860 		 */
17861 		/* log retransmit reason */
17862 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17863 		rack->r_fast_output = 0;
17864 		return (rsm);
17865 	}
17866 	return (NULL);
17867 }
17868 
17869 static void
17870 rack_log_pacing_delay_calc (struct tcp_rack *rack, uint32_t len, uint32_t slot,
17871 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17872 			   int line, struct rack_sendmap *rsm, uint8_t quality)
17873 {
17874 	if (tcp_bblogging_on(rack->rc_tp)) {
17875 		union tcp_log_stackspecific log;
17876 		struct timeval tv;
17877 
17878 		if (rack_verbose_logging == 0) {
17879 			/*
17880 			 * We are not verbose screen out all but
17881 			 * ones we always want.
17882 			 */
17883 			if ((method != 2) &&
17884 			    (method != 3) &&
17885 			    (method != 7) &&
17886 			    (method != 89) &&
17887 			    (method != 14) &&
17888 			    (method != 20)) {
17889 				return;
17890 			}
17891 		}
17892 		memset(&log, 0, sizeof(log));
17893 		log.u_bbr.flex1 = slot;
17894 		log.u_bbr.flex2 = len;
17895 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17896 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17897 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17898 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17899 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17900 		log.u_bbr.use_lt_bw <<= 1;
17901 		log.u_bbr.use_lt_bw |= rack->r_late;
17902 		log.u_bbr.use_lt_bw <<= 1;
17903 		log.u_bbr.use_lt_bw |= rack->r_early;
17904 		log.u_bbr.use_lt_bw <<= 1;
17905 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17906 		log.u_bbr.use_lt_bw <<= 1;
17907 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17908 		log.u_bbr.use_lt_bw <<= 1;
17909 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17910 		log.u_bbr.use_lt_bw <<= 1;
17911 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17912 		log.u_bbr.use_lt_bw <<= 1;
17913 		log.u_bbr.use_lt_bw |= rack->gp_ready;
17914 		log.u_bbr.pkt_epoch = line;
17915 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17916 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17917 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17918 		log.u_bbr.bw_inuse = bw_est;
17919 		log.u_bbr.delRate = bw;
17920 		if (rack->r_ctl.gp_bw == 0)
17921 			log.u_bbr.cur_del_rate = 0;
17922 		else
17923 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
17924 		log.u_bbr.rttProp = len_time;
17925 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17926 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17927 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17928 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17929 			/* We are in slow start */
17930 			log.u_bbr.flex7 = 1;
17931 		} else {
17932 			/* we are on congestion avoidance */
17933 			log.u_bbr.flex7 = 0;
17934 		}
17935 		log.u_bbr.flex8 = method;
17936 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17937 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17938 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17939 		log.u_bbr.cwnd_gain <<= 1;
17940 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17941 		log.u_bbr.cwnd_gain <<= 1;
17942 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17943 		log.u_bbr.bbr_substate = quality;
17944 		log.u_bbr.bbr_state = rack->dgp_on;
17945 		log.u_bbr.bbr_state <<= 1;
17946 		log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17947 		log.u_bbr.bbr_state <<= 2;
17948 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
17949 		    &rack->rc_inp->inp_socket->so_rcv,
17950 		    &rack->rc_inp->inp_socket->so_snd,
17951 		    BBR_LOG_HPTSI_CALC, 0,
17952 		    0, &log, false, &tv);
17953 	}
17954 }
17955 
17956 static uint32_t
17957 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17958 {
17959 	uint32_t new_tso, user_max, pace_one;
17960 
17961 	user_max = rack->rc_user_set_max_segs * mss;
17962 	if (rack->rc_force_max_seg) {
17963 		return (user_max);
17964 	}
17965 	if (rack->use_fixed_rate &&
17966 	    ((rack->r_ctl.crte == NULL) ||
17967 	     (bw != rack->r_ctl.crte->rate))) {
17968 		/* Use the user mss since we are not exactly matched */
17969 		return (user_max);
17970 	}
17971 	if (rack_pace_one_seg ||
17972 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17973 		pace_one = 1;
17974 	else
17975 		pace_one = 0;
17976 
17977 	new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17978 		     pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17979 	if (new_tso > user_max)
17980 		new_tso = user_max;
17981 	if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17982 		if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17983 			new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17984 	}
17985 	if (rack->r_ctl.rc_user_set_min_segs &&
17986 	    ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17987 	    new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17988 	return (new_tso);
17989 }
17990 
17991 static uint64_t
17992 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
17993 {
17994 	uint64_t reduced_win;
17995 	uint32_t gain;
17996 
17997 	if (window_input < rc_init_window(rack)) {
17998 		/*
17999 		 * The cwnd is collapsed to
18000 		 * nearly zero, maybe because of a time-out?
18001 		 * Lets drop back to the lt-bw.
18002 		 */
18003 		reduced_win = rack_get_lt_bw(rack);
18004 		/* Set the flag so the caller knows its a rate and not a reduced window */
18005 		*rate_set = 1;
18006 		gain = 100;
18007 	} else if  (IN_RECOVERY(rack->rc_tp->t_flags)) {
18008 		/*
18009 		 * If we are in recover our cwnd needs to be less for
18010 		 * our pacing consideration.
18011 		 */
18012 		if (rack->rack_hibeta == 0) {
18013 			reduced_win = window_input / 2;
18014 			gain = 50;
18015 		} else {
18016 			reduced_win = window_input * rack->r_ctl.saved_hibeta;
18017 			reduced_win /= 100;
18018 			gain = rack->r_ctl.saved_hibeta;
18019 		}
18020 	} else {
18021 		/*
18022 		 * Apply Timely factor to increase/decrease the
18023 		 * amount we are pacing at.
18024 		 */
18025 		gain = rack_get_output_gain(rack, NULL);
18026 		if (gain > rack_gain_p5_ub) {
18027 			gain = rack_gain_p5_ub;
18028 		}
18029 		reduced_win = window_input * gain;
18030 		reduced_win /= 100;
18031 	}
18032 	if (gain_b != NULL)
18033 		*gain_b = gain;
18034 	/*
18035 	 * What is being returned here is a trimmed down
18036 	 * window values in all cases where rate_set is left
18037 	 * at 0. In one case we actually return the rate (lt_bw).
18038 	 * the "reduced_win" is returned as a slimmed down cwnd that
18039 	 * is then calculated by the caller into a rate when rate_set
18040 	 * is 0.
18041 	 */
18042 	return (reduced_win);
18043 }
18044 
18045 static int32_t
18046 pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced)
18047 {
18048 	uint64_t lentim, fill_bw;
18049 
18050 	rack->r_via_fill_cw = 0;
18051 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
18052 		return (slot);
18053 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
18054 		return (slot);
18055 	if (rack->r_ctl.rc_last_us_rtt == 0)
18056 		return (slot);
18057 	if (rack->rc_pace_fill_if_rttin_range &&
18058 	    (rack->r_ctl.rc_last_us_rtt >=
18059 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
18060 		/* The rtt is huge, N * smallest, lets not fill */
18061 		return (slot);
18062 	}
18063 	if (rack->r_ctl.fillcw_cap && *rate_wanted >= rack->r_ctl.fillcw_cap)
18064 		return (slot);
18065 	/*
18066 	 * first lets calculate the b/w based on the last us-rtt
18067 	 * and the the smallest send window.
18068 	 */
18069 	fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
18070 	if (rack->rc_fillcw_apply_discount) {
18071 		uint32_t rate_set = 0;
18072 
18073 		fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
18074 		if (rate_set) {
18075 			goto at_lt_bw;
18076 		}
18077 	}
18078 	/* Take the rwnd if its smaller */
18079 	if (fill_bw > rack->rc_tp->snd_wnd)
18080 		fill_bw = rack->rc_tp->snd_wnd;
18081 	/* Now lets make it into a b/w */
18082 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
18083 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
18084 	/* Adjust to any cap */
18085 	if (rack->r_ctl.fillcw_cap && fill_bw >= rack->r_ctl.fillcw_cap)
18086 		fill_bw = rack->r_ctl.fillcw_cap;
18087 
18088 at_lt_bw:
18089 	if (rack_bw_multipler > 0) {
18090 		/*
18091 		 * We want to limit fill-cw to the some multiplier
18092 		 * of the max(lt_bw, gp_est). The normal default
18093 		 * is 0 for off, so a sysctl has enabled it.
18094 		 */
18095 		uint64_t lt_bw, gp, rate;
18096 
18097 		gp = rack_get_gp_est(rack);
18098 		lt_bw = rack_get_lt_bw(rack);
18099 		if (lt_bw > gp)
18100 			rate = lt_bw;
18101 		else
18102 			rate = gp;
18103 		rate *= rack_bw_multipler;
18104 		rate /= 100;
18105 		if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18106 			union tcp_log_stackspecific log;
18107 			struct timeval tv;
18108 
18109 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18110 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18111 			log.u_bbr.flex1 = rack_bw_multipler;
18112 			log.u_bbr.flex2 = len;
18113 			log.u_bbr.cur_del_rate = gp;
18114 			log.u_bbr.delRate = lt_bw;
18115 			log.u_bbr.bw_inuse = rate;
18116 			log.u_bbr.rttProp = fill_bw;
18117 			log.u_bbr.flex8 = 44;
18118 			tcp_log_event(rack->rc_tp, NULL, NULL, NULL,
18119 				      BBR_LOG_CWND, 0,
18120 				      0, &log, false, NULL,
18121 				      __func__, __LINE__, &tv);
18122 		}
18123 		if (fill_bw > rate)
18124 			fill_bw = rate;
18125 	}
18126 	/* We are below the min b/w */
18127 	if (non_paced)
18128 		*rate_wanted = fill_bw;
18129 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
18130 		return (slot);
18131 	rack->r_via_fill_cw = 1;
18132 	if (rack->r_rack_hw_rate_caps &&
18133 	    (rack->r_ctl.crte != NULL)) {
18134 		uint64_t high_rate;
18135 
18136 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
18137 		if (fill_bw > high_rate) {
18138 			/* We are capping bw at the highest rate table entry */
18139 			if (*rate_wanted > high_rate) {
18140 				/* The original rate was also capped */
18141 				rack->r_via_fill_cw = 0;
18142 			}
18143 			rack_log_hdwr_pacing(rack,
18144 					     fill_bw, high_rate, __LINE__,
18145 					     0, 3);
18146 			fill_bw = high_rate;
18147 			if (capped)
18148 				*capped = 1;
18149 		}
18150 	} else if ((rack->r_ctl.crte == NULL) &&
18151 		   (rack->rack_hdrw_pacing == 0) &&
18152 		   (rack->rack_hdw_pace_ena) &&
18153 		   rack->r_rack_hw_rate_caps &&
18154 		   (rack->rack_attempt_hdwr_pace == 0) &&
18155 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
18156 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
18157 		/*
18158 		 * Ok we may have a first attempt that is greater than our top rate
18159 		 * lets check.
18160 		 */
18161 		uint64_t high_rate;
18162 
18163 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
18164 		if (high_rate) {
18165 			if (fill_bw > high_rate) {
18166 				fill_bw = high_rate;
18167 				if (capped)
18168 					*capped = 1;
18169 			}
18170 		}
18171 	}
18172 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
18173 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
18174 				   fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
18175 		fill_bw = rack->r_ctl.bw_rate_cap;
18176 	}
18177 	/*
18178 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
18179 	 * in an rtt (unless it was capped), what does that
18180 	 * time wise equate too?
18181 	 */
18182 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
18183 	lentim /= fill_bw;
18184 	*rate_wanted = fill_bw;
18185 	if (non_paced || (lentim < slot)) {
18186 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
18187 					   0, lentim, 12, __LINE__, NULL, 0);
18188 		return ((int32_t)lentim);
18189 	} else
18190 		return (slot);
18191 }
18192 
18193 static uint32_t
18194 rack_policer_check_send(struct tcp_rack *rack, uint32_t len, uint32_t segsiz, uint32_t *needs)
18195 {
18196 	uint64_t calc;
18197 
18198 	rack->rc_policer_should_pace = 0;
18199 	calc = rack_policer_bucket_reserve * rack->r_ctl.policer_bucket_size;
18200 	calc /= 100;
18201 	/*
18202 	 * Now lets look at if we want more than is in the bucket <or>
18203 	 * we want more than is reserved in the bucket.
18204 	 */
18205 	if (rack_verbose_logging > 0)
18206 		policer_detection_log(rack, len, segsiz, calc, rack->r_ctl.current_policer_bucket, 8);
18207 	if ((calc > rack->r_ctl.current_policer_bucket) ||
18208 	    (len >= (rack->r_ctl.current_policer_bucket - calc))) {
18209 		/*
18210 		 * We may want to pace depending on if we are going
18211 		 * into the reserve or not.
18212 		 */
18213 		uint32_t newlen;
18214 
18215 		if (calc > rack->r_ctl.current_policer_bucket) {
18216 			/*
18217 			 * This will eat into the reserve if we
18218 			 * don't have room at all some lines
18219 			 * below will catch it.
18220 			 */
18221 			newlen = rack->r_ctl.policer_max_seg;
18222 			rack->rc_policer_should_pace = 1;
18223 		} else {
18224 			/*
18225 			 * We have all of the reserve plus something in the bucket
18226 			 * that we can give out.
18227 			 */
18228 			newlen = rack->r_ctl.current_policer_bucket - calc;
18229 			if (newlen < rack->r_ctl.policer_max_seg) {
18230 				/*
18231 				 * Into the reserve to get a full policer_max_seg
18232 				 * so we set the len to that and eat into
18233 				 * the reserve. If we go over the code
18234 				 * below will make us wait.
18235 				 */
18236 				newlen = rack->r_ctl.policer_max_seg;
18237 				rack->rc_policer_should_pace = 1;
18238 			}
18239 		}
18240 		if (newlen > rack->r_ctl.current_policer_bucket) {
18241 			/* We have to wait some */
18242 			*needs = newlen - rack->r_ctl.current_policer_bucket;
18243 			return (0);
18244 		}
18245 		if (rack_verbose_logging > 0)
18246 			policer_detection_log(rack, len, segsiz, newlen, 0, 9);
18247 		len = newlen;
18248 	} /* else we have all len available above the reserve */
18249 	if (rack_verbose_logging > 0)
18250 		policer_detection_log(rack, len, segsiz, calc, 0, 10);
18251 	return (len);
18252 }
18253 
18254 static uint32_t
18255 rack_policed_sending(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, uint32_t segsiz, int call_line)
18256 {
18257 	/*
18258 	 * Given a send of len, and a token bucket set at current_policer_bucket_size
18259 	 * are we close enough to the end of the bucket that we need to pace? If so
18260 	 * calculate out a time and return it. Otherwise subtract the tokens from
18261 	 * the bucket.
18262 	 */
18263 	uint64_t calc;
18264 
18265 	if ((rack->r_ctl.policer_bw == 0) ||
18266 	    (rack->r_ctl.policer_bucket_size < segsiz)) {
18267 		/*
18268 		 * We should have an estimate here...
18269 		 */
18270 		return (0);
18271 	}
18272 	calc = (uint64_t)rack_policer_bucket_reserve * (uint64_t)rack->r_ctl.policer_bucket_size;
18273 	calc /= 100;
18274 	if ((rack->r_ctl.current_policer_bucket < len) ||
18275 	    (rack->rc_policer_should_pace == 1) ||
18276 	    ((rack->r_ctl.current_policer_bucket - len) <= (uint32_t)calc)) {
18277 		/* we need to pace */
18278 		uint64_t lentim, res;
18279 		uint32_t slot;
18280 
18281 		lentim = (uint64_t)len * (uint64_t)HPTS_USEC_IN_SEC;
18282 		res = lentim / rack->r_ctl.policer_bw;
18283 		slot = (uint32_t)res;
18284 		if (rack->r_ctl.current_policer_bucket > len)
18285 			rack->r_ctl.current_policer_bucket -= len;
18286 		else
18287 			rack->r_ctl.current_policer_bucket = 0;
18288 		policer_detection_log(rack, len, slot, (uint32_t)rack_policer_bucket_reserve, call_line, 5);
18289 		rack->rc_policer_should_pace = 0;
18290 		return(slot);
18291 	}
18292 	/* Just take tokens out of the bucket and let rack do whatever it would have */
18293 	policer_detection_log(rack, len, 0, (uint32_t)rack_policer_bucket_reserve, call_line, 6);
18294 	if (len < rack->r_ctl.current_policer_bucket) {
18295 		rack->r_ctl.current_policer_bucket -= len;
18296 	} else {
18297 		rack->r_ctl.current_policer_bucket = 0;
18298 	}
18299 	return (0);
18300 }
18301 
18302 
18303 static int32_t
18304 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz, int line)
18305 {
18306 	uint64_t srtt;
18307 	int32_t slot = 0;
18308 	int32_t minslot = 0;
18309 	int can_start_hw_pacing = 1;
18310 	int err;
18311 	int pace_one;
18312 
18313 	if (rack_pace_one_seg ||
18314 	    (rack->r_ctl.rc_user_set_min_segs == 1))
18315 		pace_one = 1;
18316 	else
18317 		pace_one = 0;
18318 	if (rack->rc_policer_detected == 1) {
18319 		/*
18320 		 * A policer has been detected and we
18321 		 * have all of our data (policer-bw and
18322 		 * policer bucket size) calculated. Call
18323 		 * into the function to find out if we are
18324 		 * overriding the time.
18325 		 */
18326 		slot = rack_policed_sending(rack, tp, len, segsiz, line);
18327 		if (slot) {
18328 			uint64_t logbw;
18329 
18330 			logbw = rack->r_ctl.current_policer_bucket;
18331 			logbw <<= 32;
18332 			logbw |= rack->r_ctl.policer_bucket_size;
18333 			rack_log_pacing_delay_calc(rack, len, slot, rack->r_ctl.policer_bw, logbw, 0, 89, __LINE__, NULL, 0);
18334 			return(slot);
18335 		}
18336 	}
18337 	if (rack->rc_always_pace == 0) {
18338 		/*
18339 		 * We use the most optimistic possible cwnd/srtt for
18340 		 * sending calculations. This will make our
18341 		 * calculation anticipate getting more through
18342 		 * quicker then possible. But thats ok we don't want
18343 		 * the peer to have a gap in data sending.
18344 		 */
18345 		uint64_t cwnd, tr_perms = 0;
18346 		int32_t reduce = 0;
18347 
18348 	old_method:
18349 		/*
18350 		 * We keep no precise pacing with the old method
18351 		 * instead we use the pacer to mitigate bursts.
18352 		 */
18353 		if (rack->r_ctl.rc_rack_min_rtt)
18354 			srtt = rack->r_ctl.rc_rack_min_rtt;
18355 		else
18356 			srtt = max(tp->t_srtt, 1);
18357 		if (rack->r_ctl.rc_rack_largest_cwnd)
18358 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
18359 		else
18360 			cwnd = rack->r_ctl.cwnd_to_use;
18361 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
18362 		tr_perms = (cwnd * 1000) / srtt;
18363 		if (tr_perms == 0) {
18364 			tr_perms = ctf_fixed_maxseg(tp);
18365 		}
18366 		/*
18367 		 * Calculate how long this will take to drain, if
18368 		 * the calculation comes out to zero, thats ok we
18369 		 * will use send_a_lot to possibly spin around for
18370 		 * more increasing tot_len_this_send to the point
18371 		 * that its going to require a pace, or we hit the
18372 		 * cwnd. Which in that case we are just waiting for
18373 		 * a ACK.
18374 		 */
18375 		slot = len / tr_perms;
18376 		/* Now do we reduce the time so we don't run dry? */
18377 		if (slot && rack_slot_reduction) {
18378 			reduce = (slot / rack_slot_reduction);
18379 			if (reduce < slot) {
18380 				slot -= reduce;
18381 			} else
18382 				slot = 0;
18383 		}
18384 		slot *= HPTS_USEC_IN_MSEC;
18385 		if (rack->rc_pace_to_cwnd) {
18386 			uint64_t rate_wanted = 0;
18387 
18388 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
18389 			rack->rc_ack_can_sendout_data = 1;
18390 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
18391 		} else
18392 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
18393 		/*******************************************************/
18394 		/* RRS: We insert non-paced call to stats here for len */
18395 		/*******************************************************/
18396 	} else {
18397 		uint64_t bw_est, res, lentim, rate_wanted;
18398 		uint32_t segs, oh;
18399 		int capped = 0;
18400 		int prev_fill;
18401 
18402 		if ((rack->r_rr_config == 1) && rsm) {
18403 			return (rack->r_ctl.rc_min_to);
18404 		}
18405 		if (rack->use_fixed_rate) {
18406 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
18407 		} else if ((rack->r_ctl.init_rate == 0) &&
18408 			   (rack->r_ctl.gp_bw == 0)) {
18409 			/* no way to yet do an estimate */
18410 			bw_est = rate_wanted = 0;
18411 		} else if (rack->dgp_on) {
18412 			bw_est = rack_get_bw(rack);
18413 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
18414 		} else {
18415 			uint32_t gain, rate_set = 0;
18416 
18417 			rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
18418 			rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
18419 			if (rate_set == 0) {
18420 				if (rate_wanted > rack->rc_tp->snd_wnd)
18421 					rate_wanted = rack->rc_tp->snd_wnd;
18422 				/* Now lets make it into a b/w */
18423 				rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
18424 				rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
18425 			}
18426 			bw_est = rate_wanted;
18427 			rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
18428 						   rack->r_ctl.cwnd_to_use,
18429 						   rate_wanted, bw_est,
18430 						   rack->r_ctl.rc_last_us_rtt,
18431 						   88, __LINE__, NULL, gain);
18432 		}
18433 		if ((bw_est == 0) || (rate_wanted == 0) ||
18434 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
18435 			/*
18436 			 * No way yet to make a b/w estimate or
18437 			 * our raise is set incorrectly.
18438 			 */
18439 			goto old_method;
18440 		}
18441 		rack_rate_cap_bw(rack, &rate_wanted, &capped);
18442 		/* We need to account for all the overheads */
18443 		segs = (len + segsiz - 1) / segsiz;
18444 		/*
18445 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
18446 		 * and how much data we put in each packet. Yes this
18447 		 * means we may be off if we are larger than 1500 bytes
18448 		 * or smaller. But this just makes us more conservative.
18449 		 */
18450 
18451 		oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
18452 		if (rack->r_is_v6) {
18453 #ifdef INET6
18454 			oh += sizeof(struct ip6_hdr);
18455 #endif
18456 		} else {
18457 #ifdef INET
18458 			oh += sizeof(struct ip);
18459 #endif
18460 		}
18461 		/* We add a fixed 14 for the ethernet header */
18462 		oh += 14;
18463 		segs *= oh;
18464 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
18465 		res = lentim / rate_wanted;
18466 		slot = (uint32_t)res;
18467 		if (rack_hw_rate_min &&
18468 		    (rate_wanted < rack_hw_rate_min)) {
18469 			can_start_hw_pacing = 0;
18470 			if (rack->r_ctl.crte) {
18471 				/*
18472 				 * Ok we need to release it, we
18473 				 * have fallen too low.
18474 				 */
18475 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
18476 				rack->r_ctl.crte = NULL;
18477 				rack->rack_attempt_hdwr_pace = 0;
18478 				rack->rack_hdrw_pacing = 0;
18479 			}
18480 		}
18481 		if (rack->r_ctl.crte &&
18482 		    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
18483 			/*
18484 			 * We want more than the hardware can give us,
18485 			 * don't start any hw pacing.
18486 			 */
18487 			can_start_hw_pacing = 0;
18488 			if (rack->r_rack_hw_rate_caps == 0) {
18489 				/*
18490 				 * Ok we need to release it, we
18491 				 * want more than the card can give us and
18492 				 * no rate cap is in place. Set it up so
18493 				 * when we want less we can retry.
18494 				 */
18495 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
18496 				rack->r_ctl.crte = NULL;
18497 				rack->rack_attempt_hdwr_pace = 0;
18498 				rack->rack_hdrw_pacing = 0;
18499 			}
18500 		}
18501 		if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
18502 			/*
18503 			 * We lost our rate somehow, this can happen
18504 			 * if the interface changed underneath us.
18505 			 */
18506 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
18507 			rack->r_ctl.crte = NULL;
18508 			/* Lets re-allow attempting to setup pacing */
18509 			rack->rack_hdrw_pacing = 0;
18510 			rack->rack_attempt_hdwr_pace = 0;
18511 			rack_log_hdwr_pacing(rack,
18512 					     rate_wanted, bw_est, __LINE__,
18513 					     0, 6);
18514 		}
18515 		prev_fill = rack->r_via_fill_cw;
18516 		if ((rack->rc_pace_to_cwnd) &&
18517 		    (capped == 0) &&
18518 		    (rack->dgp_on == 1) &&
18519 		    (rack->use_fixed_rate == 0) &&
18520 		    (rack->in_probe_rtt == 0) &&
18521 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
18522 			/*
18523 			 * We want to pace at our rate *or* faster to
18524 			 * fill the cwnd to the max if its not full.
18525 			 */
18526 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
18527 			/* Re-check to make sure we are not exceeding our max b/w */
18528 			if ((rack->r_ctl.crte != NULL) &&
18529 			    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
18530 				/*
18531 				 * We want more than the hardware can give us,
18532 				 * don't start any hw pacing.
18533 				 */
18534 				can_start_hw_pacing = 0;
18535 				if (rack->r_rack_hw_rate_caps == 0) {
18536 					/*
18537 					 * Ok we need to release it, we
18538 					 * want more than the card can give us and
18539 					 * no rate cap is in place. Set it up so
18540 					 * when we want less we can retry.
18541 					 */
18542 					tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
18543 					rack->r_ctl.crte = NULL;
18544 					rack->rack_attempt_hdwr_pace = 0;
18545 					rack->rack_hdrw_pacing = 0;
18546 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
18547 				}
18548 			}
18549 		}
18550 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
18551 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
18552 			if ((rack->rack_hdw_pace_ena) &&
18553 			    (can_start_hw_pacing > 0) &&
18554 			    (rack->rack_hdrw_pacing == 0) &&
18555 			    (rack->rack_attempt_hdwr_pace == 0)) {
18556 				/*
18557 				 * Lets attempt to turn on hardware pacing
18558 				 * if we can.
18559 				 */
18560 				rack->rack_attempt_hdwr_pace = 1;
18561 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
18562 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
18563 								       rate_wanted,
18564 								       RS_PACING_GEQ,
18565 								       &err, &rack->r_ctl.crte_prev_rate);
18566 				if (rack->r_ctl.crte) {
18567 					rack->rack_hdrw_pacing = 1;
18568 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
18569 													   pace_one, rack->r_ctl.crte,
18570 													   NULL, rack->r_ctl.pace_len_divisor);
18571 					rack_log_hdwr_pacing(rack,
18572 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
18573 							     err, 0);
18574 					rack->r_ctl.last_hw_bw_req = rate_wanted;
18575 				} else {
18576 					counter_u64_add(rack_hw_pace_init_fail, 1);
18577 				}
18578 			} else if (rack->rack_hdrw_pacing &&
18579 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
18580 				/* Do we need to adjust our rate? */
18581 				const struct tcp_hwrate_limit_table *nrte;
18582 
18583 				if (rack->r_up_only &&
18584 				    (rate_wanted < rack->r_ctl.crte->rate)) {
18585 					/**
18586 					 * We have four possible states here
18587 					 * having to do with the previous time
18588 					 * and this time.
18589 					 *   previous  |  this-time
18590 					 * A)     0      |     0   -- fill_cw not in the picture
18591 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
18592 					 * C)     1      |     1   -- all rates from fill_cw
18593 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
18594 					 *
18595 					 * For case A, C and D we don't allow a drop. But for
18596 					 * case B where we now our on our steady rate we do
18597 					 * allow a drop.
18598 					 *
18599 					 */
18600 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
18601 						goto done_w_hdwr;
18602 				}
18603 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
18604 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
18605 					if (rack_hw_rate_to_low &&
18606 					    (bw_est < rack_hw_rate_to_low)) {
18607 						/*
18608 						 * The pacing rate is too low for hardware, but
18609 						 * do allow hardware pacing to be restarted.
18610 						 */
18611 						rack_log_hdwr_pacing(rack,
18612 								     bw_est, rack->r_ctl.crte->rate, __LINE__,
18613 								     0, 5);
18614 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
18615 						rack->r_ctl.crte = NULL;
18616 						rack->rack_attempt_hdwr_pace = 0;
18617 						rack->rack_hdrw_pacing = 0;
18618 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
18619 						goto done_w_hdwr;
18620 					}
18621 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
18622 								   rack->rc_tp,
18623 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
18624 								   rate_wanted,
18625 								   RS_PACING_GEQ,
18626 								   &err, &rack->r_ctl.crte_prev_rate);
18627 					if (nrte == NULL) {
18628 						/*
18629 						 * Lost the rate, lets drop hardware pacing
18630 						 * period.
18631 						 */
18632 						rack->rack_hdrw_pacing = 0;
18633 						rack->r_ctl.crte = NULL;
18634 						rack_log_hdwr_pacing(rack,
18635 								     rate_wanted, 0, __LINE__,
18636 								     err, 1);
18637 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
18638 						counter_u64_add(rack_hw_pace_lost, 1);
18639 					} else if (nrte != rack->r_ctl.crte) {
18640 						rack->r_ctl.crte = nrte;
18641 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
18642 														   segsiz, pace_one, rack->r_ctl.crte,
18643 														   NULL, rack->r_ctl.pace_len_divisor);
18644 						rack_log_hdwr_pacing(rack,
18645 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
18646 								     err, 2);
18647 						rack->r_ctl.last_hw_bw_req = rate_wanted;
18648 					}
18649 				} else {
18650 					/* We just need to adjust the segment size */
18651 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
18652 					rack_log_hdwr_pacing(rack,
18653 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
18654 							     0, 4);
18655 					rack->r_ctl.last_hw_bw_req = rate_wanted;
18656 				}
18657 			}
18658 		}
18659 		if (minslot && (minslot > slot)) {
18660 			rack_log_pacing_delay_calc(rack, minslot, slot, rack->r_ctl.crte->rate, bw_est, lentim,
18661 						   98, __LINE__, NULL, 0);
18662 			slot = minslot;
18663 		}
18664 	done_w_hdwr:
18665 		if (rack_limit_time_with_srtt &&
18666 		    (rack->use_fixed_rate == 0) &&
18667 		    (rack->rack_hdrw_pacing == 0)) {
18668 			/*
18669 			 * Sanity check, we do not allow the pacing delay
18670 			 * to be longer than the SRTT of the path. If it is
18671 			 * a slow path, then adding a packet should increase
18672 			 * the RTT and compensate for this i.e. the srtt will
18673 			 * be greater so the allowed pacing time will be greater.
18674 			 *
18675 			 * Note this restriction is not for where a peak rate
18676 			 * is set, we are doing fixed pacing or hardware pacing.
18677 			 */
18678 			if (rack->rc_tp->t_srtt)
18679 				srtt = rack->rc_tp->t_srtt;
18680 			else
18681 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
18682 			if (srtt < (uint64_t)slot) {
18683 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
18684 				slot = srtt;
18685 			}
18686 		}
18687 		/*******************************************************************/
18688 		/* RRS: We insert paced call to stats here for len and rate_wanted */
18689 		/*******************************************************************/
18690 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
18691 	}
18692 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
18693 		/*
18694 		 * If this rate is seeing enobufs when it
18695 		 * goes to send then either the nic is out
18696 		 * of gas or we are mis-estimating the time
18697 		 * somehow and not letting the queue empty
18698 		 * completely. Lets add to the pacing time.
18699 		 */
18700 		int hw_boost_delay;
18701 
18702 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
18703 		if (hw_boost_delay > rack_enobuf_hw_max)
18704 			hw_boost_delay = rack_enobuf_hw_max;
18705 		else if (hw_boost_delay < rack_enobuf_hw_min)
18706 			hw_boost_delay = rack_enobuf_hw_min;
18707 		slot += hw_boost_delay;
18708 	}
18709 	return (slot);
18710 }
18711 
18712 static void
18713 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
18714     tcp_seq startseq, uint32_t sb_offset)
18715 {
18716 	struct rack_sendmap *my_rsm = NULL;
18717 
18718 	if (tp->t_state < TCPS_ESTABLISHED) {
18719 		/*
18720 		 * We don't start any measurements if we are
18721 		 * not at least established.
18722 		 */
18723 		return;
18724 	}
18725 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
18726 		/*
18727 		 * We will get no more data into the SB
18728 		 * this means we need to have the data available
18729 		 * before we start a measurement.
18730 		 */
18731 
18732 		if (sbavail(&tptosocket(tp)->so_snd) <
18733 		    max(rc_init_window(rack),
18734 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
18735 			/* Nope not enough data */
18736 			return;
18737 		}
18738 	}
18739 	tp->t_flags |= TF_GPUTINPROG;
18740 	rack->r_ctl.rc_gp_cumack_ts = 0;
18741 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
18742 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
18743 	tp->gput_seq = startseq;
18744 	rack->app_limited_needs_set = 0;
18745 	if (rack->in_probe_rtt)
18746 		rack->measure_saw_probe_rtt = 1;
18747 	else if ((rack->measure_saw_probe_rtt) &&
18748 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
18749 		rack->measure_saw_probe_rtt = 0;
18750 	if (rack->rc_gp_filled)
18751 		tp->gput_ts = rack->r_ctl.last_cumack_advance;
18752 	else {
18753 		/* Special case initial measurement */
18754 		struct timeval tv;
18755 
18756 		tp->gput_ts = tcp_get_usecs(&tv);
18757 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18758 	}
18759 	/*
18760 	 * We take a guess out into the future,
18761 	 * if we have no measurement and no
18762 	 * initial rate, we measure the first
18763 	 * initial-windows worth of data to
18764 	 * speed up getting some GP measurement and
18765 	 * thus start pacing.
18766 	 */
18767 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
18768 		rack->app_limited_needs_set = 1;
18769 		tp->gput_ack = startseq + max(rc_init_window(rack),
18770 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
18771 		rack_log_pacing_delay_calc(rack,
18772 					   tp->gput_seq,
18773 					   tp->gput_ack,
18774 					   0,
18775 					   tp->gput_ts,
18776 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18777 					   9,
18778 					   __LINE__, NULL, 0);
18779 		rack_tend_gp_marks(tp, rack);
18780 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18781 		return;
18782 	}
18783 	if (sb_offset) {
18784 		/*
18785 		 * We are out somewhere in the sb
18786 		 * can we use the already outstanding data?
18787 		 */
18788 
18789 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
18790 			/*
18791 			 * Yes first one is good and in this case
18792 			 * the tp->gput_ts is correctly set based on
18793 			 * the last ack that arrived (no need to
18794 			 * set things up when an ack comes in).
18795 			 */
18796 			my_rsm = tqhash_min(rack->r_ctl.tqh);
18797 			if ((my_rsm == NULL) ||
18798 			    (my_rsm->r_rtr_cnt != 1)) {
18799 				/* retransmission? */
18800 				goto use_latest;
18801 			}
18802 		} else {
18803 			if (rack->r_ctl.rc_first_appl == NULL) {
18804 				/*
18805 				 * If rc_first_appl is NULL
18806 				 * then the cnt should be 0.
18807 				 * This is probably an error, maybe
18808 				 * a KASSERT would be approprate.
18809 				 */
18810 				goto use_latest;
18811 			}
18812 			/*
18813 			 * If we have a marker pointer to the last one that is
18814 			 * app limited we can use that, but we need to set
18815 			 * things up so that when it gets ack'ed we record
18816 			 * the ack time (if its not already acked).
18817 			 */
18818 			rack->app_limited_needs_set = 1;
18819 			/*
18820 			 * We want to get to the rsm that is either
18821 			 * next with space i.e. over 1 MSS or the one
18822 			 * after that (after the app-limited).
18823 			 */
18824 			my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
18825 			if (my_rsm) {
18826 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
18827 					/* Have to use the next one */
18828 					my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
18829 				else {
18830 					/* Use after the first MSS of it is acked */
18831 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
18832 					goto start_set;
18833 				}
18834 			}
18835 			if ((my_rsm == NULL) ||
18836 			    (my_rsm->r_rtr_cnt != 1)) {
18837 				/*
18838 				 * Either its a retransmit or
18839 				 * the last is the app-limited one.
18840 				 */
18841 				goto use_latest;
18842 			}
18843 		}
18844 		tp->gput_seq = my_rsm->r_start;
18845 start_set:
18846 		if (my_rsm->r_flags & RACK_ACKED) {
18847 			/*
18848 			 * This one has been acked use the arrival ack time
18849 			 */
18850 			struct rack_sendmap *nrsm;
18851 
18852 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18853 			rack->app_limited_needs_set = 0;
18854 			/*
18855 			 * Ok in this path we need to use the r_end now
18856 			 * since this guy is the starting ack.
18857 			 */
18858 			tp->gput_seq = my_rsm->r_end;
18859 			/*
18860 			 * We also need to adjust up the sendtime
18861 			 * to the send of the next data after my_rsm.
18862 			 */
18863 			nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
18864 			if (nrsm != NULL)
18865 				my_rsm = nrsm;
18866 			else {
18867 				/*
18868 				 * The next as not been sent, thats the
18869 				 * case for using the latest.
18870 				 */
18871 				goto use_latest;
18872 			}
18873 		}
18874 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18875 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
18876 		rack->r_ctl.rc_gp_cumack_ts = 0;
18877 		if ((rack->r_ctl.cleared_app_ack == 1) &&
18878 		    (SEQ_GEQ(rack->r_ctl.cleared_app_ack, tp->gput_seq))) {
18879 			/*
18880 			 * We just cleared an application limited period
18881 			 * so the next seq out needs to skip the first
18882 			 * ack.
18883 			 */
18884 			rack->app_limited_needs_set = 1;
18885 			rack->r_ctl.cleared_app_ack = 0;
18886 		}
18887 		rack_log_pacing_delay_calc(rack,
18888 					   tp->gput_seq,
18889 					   tp->gput_ack,
18890 					   (uintptr_t)my_rsm,
18891 					   tp->gput_ts,
18892 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18893 					   9,
18894 					   __LINE__, my_rsm, 0);
18895 		/* Now lets make sure all are marked as they should be */
18896 		rack_tend_gp_marks(tp, rack);
18897 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18898 		return;
18899 	}
18900 
18901 use_latest:
18902 	/*
18903 	 * We don't know how long we may have been
18904 	 * idle or if this is the first-send. Lets
18905 	 * setup the flag so we will trim off
18906 	 * the first ack'd data so we get a true
18907 	 * measurement.
18908 	 */
18909 	rack->app_limited_needs_set = 1;
18910 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
18911 	rack->r_ctl.rc_gp_cumack_ts = 0;
18912 	/* Find this guy so we can pull the send time */
18913 	my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
18914 	if (my_rsm) {
18915 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18916 		if (my_rsm->r_flags & RACK_ACKED) {
18917 			/*
18918 			 * Unlikely since its probably what was
18919 			 * just transmitted (but I am paranoid).
18920 			 */
18921 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18922 			rack->app_limited_needs_set = 0;
18923 		}
18924 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18925 			/* This also is unlikely */
18926 			tp->gput_seq = my_rsm->r_start;
18927 		}
18928 	} else {
18929 		/*
18930 		 * TSNH unless we have some send-map limit,
18931 		 * and even at that it should not be hitting
18932 		 * that limit (we should have stopped sending).
18933 		 */
18934 		struct timeval tv;
18935 
18936 		microuptime(&tv);
18937 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18938 	}
18939 	rack_tend_gp_marks(tp, rack);
18940 	rack_log_pacing_delay_calc(rack,
18941 				   tp->gput_seq,
18942 				   tp->gput_ack,
18943 				   (uintptr_t)my_rsm,
18944 				   tp->gput_ts,
18945 				   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18946 				   9, __LINE__, NULL, 0);
18947 	rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18948 }
18949 
18950 static inline uint32_t
18951 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
18952     uint32_t avail, int32_t sb_offset)
18953 {
18954 	uint32_t len;
18955 	uint32_t sendwin;
18956 
18957 	if (tp->snd_wnd > cwnd_to_use)
18958 		sendwin = cwnd_to_use;
18959 	else
18960 		sendwin = tp->snd_wnd;
18961 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
18962 		/* We never want to go over our peers rcv-window */
18963 		len = 0;
18964 	} else {
18965 		uint32_t flight;
18966 
18967 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18968 		if (flight >= sendwin) {
18969 			/*
18970 			 * We have in flight what we are allowed by cwnd (if
18971 			 * it was rwnd blocking it would have hit above out
18972 			 * >= tp->snd_wnd).
18973 			 */
18974 			return (0);
18975 		}
18976 		len = sendwin - flight;
18977 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18978 			/* We would send too much (beyond the rwnd) */
18979 			len = tp->snd_wnd - ctf_outstanding(tp);
18980 		}
18981 		if ((len + sb_offset) > avail) {
18982 			/*
18983 			 * We don't have that much in the SB, how much is
18984 			 * there?
18985 			 */
18986 			len = avail - sb_offset;
18987 		}
18988 	}
18989 	return (len);
18990 }
18991 
18992 static void
18993 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18994 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18995 	     int rsm_is_null, int optlen, int line, uint16_t mode)
18996 {
18997 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18998 		union tcp_log_stackspecific log;
18999 		struct timeval tv;
19000 
19001 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19002 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19003 		log.u_bbr.flex1 = error;
19004 		log.u_bbr.flex2 = flags;
19005 		log.u_bbr.flex3 = rsm_is_null;
19006 		log.u_bbr.flex4 = ipoptlen;
19007 		log.u_bbr.flex5 = tp->rcv_numsacks;
19008 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19009 		log.u_bbr.flex7 = optlen;
19010 		log.u_bbr.flex8 = rack->r_fsb_inited;
19011 		log.u_bbr.applimited = rack->r_fast_output;
19012 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19013 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19014 		log.u_bbr.cwnd_gain = mode;
19015 		log.u_bbr.pkts_out = orig_len;
19016 		log.u_bbr.lt_epoch = len;
19017 		log.u_bbr.delivered = line;
19018 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
19019 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19020 		tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
19021 			       len, &log, false, NULL, __func__, __LINE__, &tv);
19022 	}
19023 }
19024 
19025 
19026 static struct mbuf *
19027 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
19028 		   struct rack_fast_send_blk *fsb,
19029 		   int32_t seglimit, int32_t segsize, int hw_tls)
19030 {
19031 #ifdef KERN_TLS
19032 	struct ktls_session *tls, *ntls;
19033 #ifdef INVARIANTS
19034 	struct mbuf *start;
19035 #endif
19036 #endif
19037 	struct mbuf *m, *n, **np, *smb;
19038 	struct mbuf *top;
19039 	int32_t off, soff;
19040 	int32_t len = *plen;
19041 	int32_t fragsize;
19042 	int32_t len_cp = 0;
19043 	uint32_t mlen, frags;
19044 
19045 	soff = off = the_off;
19046 	smb = m = the_m;
19047 	np = &top;
19048 	top = NULL;
19049 #ifdef KERN_TLS
19050 	if (hw_tls && (m->m_flags & M_EXTPG))
19051 		tls = m->m_epg_tls;
19052 	else
19053 		tls = NULL;
19054 #ifdef INVARIANTS
19055 	start = m;
19056 #endif
19057 #endif
19058 	while (len > 0) {
19059 		if (m == NULL) {
19060 			*plen = len_cp;
19061 			break;
19062 		}
19063 #ifdef KERN_TLS
19064 		if (hw_tls) {
19065 			if (m->m_flags & M_EXTPG)
19066 				ntls = m->m_epg_tls;
19067 			else
19068 				ntls = NULL;
19069 
19070 			/*
19071 			 * Avoid mixing TLS records with handshake
19072 			 * data or TLS records from different
19073 			 * sessions.
19074 			 */
19075 			if (tls != ntls) {
19076 				MPASS(m != start);
19077 				*plen = len_cp;
19078 				break;
19079 			}
19080 		}
19081 #endif
19082 		mlen = min(len, m->m_len - off);
19083 		if (seglimit) {
19084 			/*
19085 			 * For M_EXTPG mbufs, add 3 segments
19086 			 * + 1 in case we are crossing page boundaries
19087 			 * + 2 in case the TLS hdr/trailer are used
19088 			 * It is cheaper to just add the segments
19089 			 * than it is to take the cache miss to look
19090 			 * at the mbuf ext_pgs state in detail.
19091 			 */
19092 			if (m->m_flags & M_EXTPG) {
19093 				fragsize = min(segsize, PAGE_SIZE);
19094 				frags = 3;
19095 			} else {
19096 				fragsize = segsize;
19097 				frags = 0;
19098 			}
19099 
19100 			/* Break if we really can't fit anymore. */
19101 			if ((frags + 1) >= seglimit) {
19102 				*plen =	len_cp;
19103 				break;
19104 			}
19105 
19106 			/*
19107 			 * Reduce size if you can't copy the whole
19108 			 * mbuf. If we can't copy the whole mbuf, also
19109 			 * adjust len so the loop will end after this
19110 			 * mbuf.
19111 			 */
19112 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
19113 				mlen = (seglimit - frags - 1) * fragsize;
19114 				len = mlen;
19115 				*plen = len_cp + len;
19116 			}
19117 			frags += howmany(mlen, fragsize);
19118 			if (frags == 0)
19119 				frags++;
19120 			seglimit -= frags;
19121 			KASSERT(seglimit > 0,
19122 			    ("%s: seglimit went too low", __func__));
19123 		}
19124 		n = m_get(M_NOWAIT, m->m_type);
19125 		*np = n;
19126 		if (n == NULL)
19127 			goto nospace;
19128 		n->m_len = mlen;
19129 		soff += mlen;
19130 		len_cp += n->m_len;
19131 		if (m->m_flags & (M_EXT | M_EXTPG)) {
19132 			n->m_data = m->m_data + off;
19133 			mb_dupcl(n, m);
19134 		} else {
19135 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
19136 			    (u_int)n->m_len);
19137 		}
19138 		len -= n->m_len;
19139 		off = 0;
19140 		m = m->m_next;
19141 		np = &n->m_next;
19142 		if (len || (soff == smb->m_len)) {
19143 			/*
19144 			 * We have more so we move forward  or
19145 			 * we have consumed the entire mbuf and
19146 			 * len has fell to 0.
19147 			 */
19148 			soff = 0;
19149 			smb = m;
19150 		}
19151 
19152 	}
19153 	if (fsb != NULL) {
19154 		fsb->m = smb;
19155 		fsb->off = soff;
19156 		if (smb) {
19157 			/*
19158 			 * Save off the size of the mbuf. We do
19159 			 * this so that we can recognize when it
19160 			 * has been trimmed by sbcut() as acks
19161 			 * come in.
19162 			 */
19163 			fsb->o_m_len = smb->m_len;
19164 			fsb->o_t_len = M_TRAILINGROOM(smb);
19165 		} else {
19166 			/*
19167 			 * This is the case where the next mbuf went to NULL. This
19168 			 * means with this copy we have sent everything in the sb.
19169 			 * In theory we could clear the fast_output flag, but lets
19170 			 * not since its possible that we could get more added
19171 			 * and acks that call the extend function which would let
19172 			 * us send more.
19173 			 */
19174 			fsb->o_m_len = 0;
19175 			fsb->o_t_len = 0;
19176 		}
19177 	}
19178 	return (top);
19179 nospace:
19180 	if (top)
19181 		m_freem(top);
19182 	return (NULL);
19183 
19184 }
19185 
19186 /*
19187  * This is a copy of m_copym(), taking the TSO segment size/limit
19188  * constraints into account, and advancing the sndptr as it goes.
19189  */
19190 static struct mbuf *
19191 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
19192 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
19193 {
19194 	struct mbuf *m, *n;
19195 	int32_t soff;
19196 
19197 	m = rack->r_ctl.fsb.m;
19198 	if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
19199 		/*
19200 		 * The trailing space changed, mbufs can grow
19201 		 * at the tail but they can't shrink from
19202 		 * it, KASSERT that. Adjust the orig_m_len to
19203 		 * compensate for this change.
19204 		 */
19205 		KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
19206 			("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
19207 			 m,
19208 			 rack,
19209 			 (intmax_t)M_TRAILINGROOM(m),
19210 			 rack->r_ctl.fsb.o_t_len,
19211 			 rack->r_ctl.fsb.o_m_len,
19212 			 m->m_len));
19213 		rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
19214 		rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
19215 	}
19216 	if (m->m_len < rack->r_ctl.fsb.o_m_len) {
19217 		/*
19218 		 * Mbuf shrank, trimmed off the top by an ack, our
19219 		 * offset changes.
19220 		 */
19221 		KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
19222 			("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
19223 			 m, m->m_len,
19224 			 rack, rack->r_ctl.fsb.o_m_len,
19225 			 rack->r_ctl.fsb.off));
19226 
19227 		if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
19228 			rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
19229 		else
19230 			rack->r_ctl.fsb.off = 0;
19231 		rack->r_ctl.fsb.o_m_len = m->m_len;
19232 #ifdef INVARIANTS
19233 	} else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
19234 		panic("rack:%p m:%p m_len grew outside of t_space compensation",
19235 		      rack, m);
19236 #endif
19237 	}
19238 	soff = rack->r_ctl.fsb.off;
19239 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
19240 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
19241 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
19242 				 __FUNCTION__,
19243 				 rack, *plen, m, m->m_len));
19244 	/* Save off the right location before we copy and advance */
19245 	*s_soff = soff;
19246 	*s_mb = rack->r_ctl.fsb.m;
19247 	n = rack_fo_base_copym(m, soff, plen,
19248 			       &rack->r_ctl.fsb,
19249 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
19250 	return (n);
19251 }
19252 
19253 /* Log the buffer level */
19254 static void
19255 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
19256 		     int len, struct timeval *tv,
19257 		     uint32_t cts)
19258 {
19259 	uint32_t p_rate = 0, p_queue = 0, err = 0;
19260 	union tcp_log_stackspecific log;
19261 
19262 #ifdef RATELIMIT
19263 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
19264 	err = in_pcbquery_txrtlmt(rack->rc_inp,	&p_rate);
19265 #endif
19266 	memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19267 	log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19268 	log.u_bbr.flex1 = p_rate;
19269 	log.u_bbr.flex2 = p_queue;
19270 	log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
19271 	log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
19272 	log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
19273 	log.u_bbr.flex7 = 99;
19274 	log.u_bbr.flex8 = 0;
19275 	log.u_bbr.pkts_out = err;
19276 	log.u_bbr.delRate = rack->r_ctl.crte->rate;
19277 	log.u_bbr.timeStamp = cts;
19278 	log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19279 	tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
19280 		       len, &log, false, NULL, __func__, __LINE__, tv);
19281 
19282 }
19283 
19284 static uint32_t
19285 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
19286 		       struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
19287 {
19288 	uint64_t lentime = 0;
19289 #ifdef RATELIMIT
19290 	uint32_t p_rate = 0, p_queue = 0, err;
19291 	union tcp_log_stackspecific log;
19292 	uint64_t bw;
19293 
19294 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
19295 	/* Failed or queue is zero */
19296 	if (err || (p_queue == 0)) {
19297 		lentime = 0;
19298 		goto out;
19299 	}
19300 	err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
19301 	if (err) {
19302 		lentime = 0;
19303 		goto out;
19304 	}
19305 	/*
19306 	 * If we reach here we have some bytes in
19307 	 * the queue. The number returned is a value
19308 	 * between 0 and 0xffff where ffff is full
19309 	 * and 0 is empty. So how best to make this into
19310 	 * something usable?
19311 	 *
19312 	 * The "safer" way is lets take the b/w gotten
19313 	 * from the query (which should be our b/w rate)
19314 	 * and pretend that a full send (our rc_pace_max_segs)
19315 	 * is outstanding. We factor it so its as if a full
19316 	 * number of our MSS segment is terms of full
19317 	 * ethernet segments are outstanding.
19318 	 */
19319 	bw = p_rate / 8;
19320 	if (bw) {
19321 		lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
19322 		lentime *= ETHERNET_SEGMENT_SIZE;
19323 		lentime *= (uint64_t)HPTS_USEC_IN_SEC;
19324 		lentime /= bw;
19325 	} else {
19326 		/* TSNH -- KASSERT? */
19327 		lentime = 0;
19328 	}
19329 out:
19330 	if (tcp_bblogging_on(tp)) {
19331 		memset(&log, 0, sizeof(log));
19332 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19333 		log.u_bbr.flex1 = p_rate;
19334 		log.u_bbr.flex2 = p_queue;
19335 		log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
19336 		log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
19337 		log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
19338 		log.u_bbr.flex7 = 99;
19339 		log.u_bbr.flex8 = 0;
19340 		log.u_bbr.pkts_out = err;
19341 		log.u_bbr.delRate = rack->r_ctl.crte->rate;
19342 		log.u_bbr.cur_del_rate = lentime;
19343 		log.u_bbr.timeStamp = cts;
19344 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19345 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
19346 			       len, &log, false, NULL, __func__, __LINE__,tv);
19347 	}
19348 #endif
19349 	return ((uint32_t)lentime);
19350 }
19351 
19352 static int
19353 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
19354 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
19355 {
19356 	/*
19357 	 * Enter the fast retransmit path. We are given that a sched_pin is
19358 	 * in place (if accounting is compliled in) and the cycle count taken
19359 	 * at the entry is in the ts_val. The concept her is that the rsm
19360 	 * now holds the mbuf offsets and such so we can directly transmit
19361 	 * without a lot of overhead, the len field is already set for
19362 	 * us to prohibit us from sending too much (usually its 1MSS).
19363 	 */
19364 	struct ip *ip = NULL;
19365 	struct udphdr *udp = NULL;
19366 	struct tcphdr *th = NULL;
19367 	struct mbuf *m = NULL;
19368 	struct inpcb *inp;
19369 	uint8_t *cpto;
19370 	struct tcp_log_buffer *lgb;
19371 #ifdef TCP_ACCOUNTING
19372 	uint64_t crtsc;
19373 	int cnt_thru = 1;
19374 #endif
19375 	struct tcpopt to;
19376 	u_char opt[TCP_MAXOLEN];
19377 	uint32_t hdrlen, optlen;
19378 	int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
19379 	uint16_t flags;
19380 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
19381 	uint32_t if_hw_tsomaxsegsize;
19382 	int32_t ip_sendflag = IP_NO_SND_TAG_RL;
19383 
19384 #ifdef INET6
19385 	struct ip6_hdr *ip6 = NULL;
19386 
19387 	if (rack->r_is_v6) {
19388 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19389 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19390 	} else
19391 #endif				/* INET6 */
19392 	{
19393 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19394 		hdrlen = sizeof(struct tcpiphdr);
19395 	}
19396 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19397 		goto failed;
19398 	}
19399 	if (doing_tlp) {
19400 		/* Its a TLP add the flag, it may already be there but be sure */
19401 		rsm->r_flags |= RACK_TLP;
19402 	} else {
19403 		/* If it was a TLP it is not not on this retransmit */
19404 		rsm->r_flags &= ~RACK_TLP;
19405 	}
19406 	startseq = rsm->r_start;
19407 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19408 	inp = rack->rc_inp;
19409 	to.to_flags = 0;
19410 	flags = tcp_outflags[tp->t_state];
19411 	if (flags & (TH_SYN|TH_RST)) {
19412 		goto failed;
19413 	}
19414 	if (rsm->r_flags & RACK_HAS_FIN) {
19415 		/* We can't send a FIN here */
19416 		goto failed;
19417 	}
19418 	if (flags & TH_FIN) {
19419 		/* We never send a FIN */
19420 		flags &= ~TH_FIN;
19421 	}
19422 	if (tp->t_flags & TF_RCVD_TSTMP) {
19423 		to.to_tsval = ms_cts + tp->ts_offset;
19424 		to.to_tsecr = tp->ts_recent;
19425 		to.to_flags = TOF_TS;
19426 	}
19427 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19428 	/* TCP-MD5 (RFC2385). */
19429 	if (tp->t_flags & TF_SIGNATURE)
19430 		to.to_flags |= TOF_SIGNATURE;
19431 #endif
19432 	optlen = tcp_addoptions(&to, opt);
19433 	hdrlen += optlen;
19434 	udp = rack->r_ctl.fsb.udp;
19435 	if (udp)
19436 		hdrlen += sizeof(struct udphdr);
19437 	if (rack->r_ctl.rc_pace_max_segs)
19438 		max_val = rack->r_ctl.rc_pace_max_segs;
19439 	else if (rack->rc_user_set_max_segs)
19440 		max_val = rack->rc_user_set_max_segs * segsiz;
19441 	else
19442 		max_val = len;
19443 	if ((tp->t_flags & TF_TSO) &&
19444 	    V_tcp_do_tso &&
19445 	    (len > segsiz) &&
19446 	    (tp->t_port == 0))
19447 		tso = 1;
19448 #ifdef INET6
19449 	if (MHLEN < hdrlen + max_linkhdr)
19450 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19451 	else
19452 #endif
19453 		m = m_gethdr(M_NOWAIT, MT_DATA);
19454 	if (m == NULL)
19455 		goto failed;
19456 	m->m_data += max_linkhdr;
19457 	m->m_len = hdrlen;
19458 	th = rack->r_ctl.fsb.th;
19459 	/* Establish the len to send */
19460 	if (len > max_val)
19461 		len = max_val;
19462 	if ((tso) && (len + optlen > segsiz)) {
19463 		uint32_t if_hw_tsomax;
19464 		int32_t max_len;
19465 
19466 		/* extract TSO information */
19467 		if_hw_tsomax = tp->t_tsomax;
19468 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19469 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19470 		/*
19471 		 * Check if we should limit by maximum payload
19472 		 * length:
19473 		 */
19474 		if (if_hw_tsomax != 0) {
19475 			/* compute maximum TSO length */
19476 			max_len = (if_hw_tsomax - hdrlen -
19477 				   max_linkhdr);
19478 			if (max_len <= 0) {
19479 				goto failed;
19480 			} else if (len > max_len) {
19481 				len = max_len;
19482 			}
19483 		}
19484 		if (len <= segsiz) {
19485 			/*
19486 			 * In case there are too many small fragments don't
19487 			 * use TSO:
19488 			 */
19489 			tso = 0;
19490 		}
19491 	} else {
19492 		tso = 0;
19493 	}
19494 	if ((tso == 0) && (len > segsiz))
19495 		len = segsiz;
19496 	(void)tcp_get_usecs(tv);
19497 	if ((len == 0) ||
19498 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
19499 		goto failed;
19500 	}
19501 	th->th_seq = htonl(rsm->r_start);
19502 	th->th_ack = htonl(tp->rcv_nxt);
19503 	/*
19504 	 * The PUSH bit should only be applied
19505 	 * if the full retransmission is made. If
19506 	 * we are sending less than this is the
19507 	 * left hand edge and should not have
19508 	 * the PUSH bit.
19509 	 */
19510 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
19511 	    (len == (rsm->r_end - rsm->r_start)))
19512 		flags |= TH_PUSH;
19513 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19514 	if (th->th_win == 0) {
19515 		tp->t_sndzerowin++;
19516 		tp->t_flags |= TF_RXWIN0SENT;
19517 	} else
19518 		tp->t_flags &= ~TF_RXWIN0SENT;
19519 	if (rsm->r_flags & RACK_TLP) {
19520 		/*
19521 		 * TLP should not count in retran count, but
19522 		 * in its own bin
19523 		 */
19524 		counter_u64_add(rack_tlp_retran, 1);
19525 		counter_u64_add(rack_tlp_retran_bytes, len);
19526 	} else {
19527 		tp->t_sndrexmitpack++;
19528 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
19529 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
19530 	}
19531 #ifdef STATS
19532 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
19533 				 len);
19534 #endif
19535 	if (rsm->m == NULL)
19536 		goto failed;
19537 	if (rsm->m &&
19538 	    ((rsm->orig_m_len != rsm->m->m_len) ||
19539 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
19540 		/* Fix up the orig_m_len and possibly the mbuf offset */
19541 		rack_adjust_orig_mlen(rsm);
19542 	}
19543 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
19544 	if (len <= segsiz) {
19545 		/*
19546 		 * Must have ran out of mbufs for the copy
19547 		 * shorten it to no longer need tso. Lets
19548 		 * not put on sendalot since we are low on
19549 		 * mbufs.
19550 		 */
19551 		tso = 0;
19552 	}
19553 	if ((m->m_next == NULL) || (len <= 0)){
19554 		goto failed;
19555 	}
19556 	if (udp) {
19557 		if (rack->r_is_v6)
19558 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
19559 		else
19560 			ulen = hdrlen + len - sizeof(struct ip);
19561 		udp->uh_ulen = htons(ulen);
19562 	}
19563 	m->m_pkthdr.rcvif = (struct ifnet *)0;
19564 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
19565 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19566 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
19567 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19568 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
19569 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19570 #ifdef INET6
19571 		if (rack->r_is_v6) {
19572 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19573 		    ip6->ip6_flow |= htonl(ect << 20);
19574 		}
19575 		else
19576 #endif
19577 		{
19578 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
19579 		    ip->ip_tos |= ect;
19580 		}
19581 	}
19582 	if (rack->r_ctl.crte != NULL) {
19583 		/* See if we can send via the hw queue */
19584 		slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
19585 		/* If there is nothing in queue (no pacing time) we can send via the hw queue */
19586 		if (slot == 0)
19587 			ip_sendflag = 0;
19588 	}
19589 	tcp_set_flags(th, flags);
19590 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
19591 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19592 	if (to.to_flags & TOF_SIGNATURE) {
19593 		/*
19594 		 * Calculate MD5 signature and put it into the place
19595 		 * determined before.
19596 		 * NOTE: since TCP options buffer doesn't point into
19597 		 * mbuf's data, calculate offset and use it.
19598 		 */
19599 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
19600 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
19601 			/*
19602 			 * Do not send segment if the calculation of MD5
19603 			 * digest has failed.
19604 			 */
19605 			goto failed;
19606 		}
19607 	}
19608 #endif
19609 #ifdef INET6
19610 	if (rack->r_is_v6) {
19611 		if (tp->t_port) {
19612 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19613 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19614 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19615 			th->th_sum = htons(0);
19616 			UDPSTAT_INC(udps_opackets);
19617 		} else {
19618 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19619 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19620 			th->th_sum = in6_cksum_pseudo(ip6,
19621 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19622 						      0);
19623 		}
19624 	}
19625 #endif
19626 #if defined(INET6) && defined(INET)
19627 	else
19628 #endif
19629 #ifdef INET
19630 	{
19631 		if (tp->t_port) {
19632 			m->m_pkthdr.csum_flags = CSUM_UDP;
19633 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19634 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19635 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19636 			th->th_sum = htons(0);
19637 			UDPSTAT_INC(udps_opackets);
19638 		} else {
19639 			m->m_pkthdr.csum_flags = CSUM_TCP;
19640 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19641 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
19642 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19643 									IPPROTO_TCP + len + optlen));
19644 		}
19645 		/* IP version must be set here for ipv4/ipv6 checking later */
19646 		KASSERT(ip->ip_v == IPVERSION,
19647 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
19648 	}
19649 #endif
19650 	if (tso) {
19651 		/*
19652 		 * Here we use segsiz since we have no added options besides
19653 		 * any standard timestamp options (no DSACKs or SACKS are sent
19654 		 * via either fast-path).
19655 		 */
19656 		KASSERT(len > segsiz,
19657 			("%s: len <= tso_segsz tp:%p", __func__, tp));
19658 		m->m_pkthdr.csum_flags |= CSUM_TSO;
19659 		m->m_pkthdr.tso_segsz = segsiz;
19660 	}
19661 #ifdef INET6
19662 	if (rack->r_is_v6) {
19663 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19664 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19665 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19666 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19667 		else
19668 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19669 	}
19670 #endif
19671 #if defined(INET) && defined(INET6)
19672 	else
19673 #endif
19674 #ifdef INET
19675 	{
19676 		ip->ip_len = htons(m->m_pkthdr.len);
19677 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19678 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19679 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19680 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19681 				ip->ip_off |= htons(IP_DF);
19682 			}
19683 		} else {
19684 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19685 		}
19686 	}
19687 #endif
19688 	if (doing_tlp == 0) {
19689 		/* Set we retransmitted */
19690 		rack->rc_gp_saw_rec = 1;
19691 	} else {
19692 		/* Its a TLP set ca or ss */
19693 		if (tp->snd_cwnd > tp->snd_ssthresh) {
19694 			/* Set we sent in CA */
19695 			rack->rc_gp_saw_ca = 1;
19696 		} else {
19697 			/* Set we sent in SS */
19698 			rack->rc_gp_saw_ss = 1;
19699 		}
19700 	}
19701 	/* Time to copy in our header */
19702 	cpto = mtod(m, uint8_t *);
19703 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19704 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19705 	if (optlen) {
19706 		bcopy(opt, th + 1, optlen);
19707 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19708 	} else {
19709 		th->th_off = sizeof(struct tcphdr) >> 2;
19710 	}
19711 	if (tcp_bblogging_on(rack->rc_tp)) {
19712 		union tcp_log_stackspecific log;
19713 
19714 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
19715 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
19716 			counter_u64_add(rack_collapsed_win_rxt, 1);
19717 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
19718 		}
19719 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19720 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19721 		if (rack->rack_no_prr)
19722 			log.u_bbr.flex1 = 0;
19723 		else
19724 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19725 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19726 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19727 		log.u_bbr.flex4 = max_val;
19728 		/* Save off the early/late values */
19729 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19730 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19731 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19732 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19733 		if (doing_tlp == 0)
19734 			log.u_bbr.flex8 = 1;
19735 		else
19736 			log.u_bbr.flex8 = 2;
19737 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19738 		log.u_bbr.flex7 = 55;
19739 		log.u_bbr.pkts_out = tp->t_maxseg;
19740 		log.u_bbr.timeStamp = cts;
19741 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19742 		if (rsm && (rsm->r_rtr_cnt > 0)) {
19743 			/*
19744 			 * When we have a retransmit we want to log the
19745 			 * burst at send and flight at send from before.
19746 			 */
19747 			log.u_bbr.flex5 = rsm->r_fas;
19748 			log.u_bbr.bbr_substate = rsm->r_bas;
19749 		} else {
19750 			/*
19751 			 * This is currently unlikely until we do the
19752 			 * packet pair probes but I will add it for completeness.
19753 			 */
19754 			log.u_bbr.flex5 = log.u_bbr.inflight;
19755 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19756 		}
19757 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19758 		log.u_bbr.delivered = 0;
19759 		log.u_bbr.rttProp = (uintptr_t)rsm;
19760 		log.u_bbr.delRate = rsm->r_flags;
19761 		log.u_bbr.delRate <<= 31;
19762 		log.u_bbr.delRate |= rack->r_must_retran;
19763 		log.u_bbr.delRate <<= 1;
19764 		log.u_bbr.delRate |= 1;
19765 		log.u_bbr.pkt_epoch = __LINE__;
19766 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19767 				     len, &log, false, NULL, __func__, __LINE__, tv);
19768 	} else
19769 		lgb = NULL;
19770 	if ((rack->r_ctl.crte != NULL) &&
19771 	    tcp_bblogging_on(tp)) {
19772 		rack_log_queue_level(tp, rack, len, tv, cts);
19773 	}
19774 #ifdef INET6
19775 	if (rack->r_is_v6) {
19776 		error = ip6_output(m, inp->in6p_outputopts,
19777 				   &inp->inp_route6,
19778 				   ip_sendflag, NULL, NULL, inp);
19779 	}
19780 	else
19781 #endif
19782 #ifdef INET
19783 	{
19784 		error = ip_output(m, NULL,
19785 				  &inp->inp_route,
19786 				  ip_sendflag, 0, inp);
19787 	}
19788 #endif
19789 	m = NULL;
19790 	if (lgb) {
19791 		lgb->tlb_errno = error;
19792 		lgb = NULL;
19793 	}
19794 	/* Move snd_nxt to snd_max so we don't have false retransmissions */
19795 	tp->snd_nxt = tp->snd_max;
19796 	if (error) {
19797 		goto failed;
19798 	} else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
19799 		rack->rc_hw_nobuf = 0;
19800 		rack->r_ctl.rc_agg_delayed = 0;
19801 		rack->r_early = 0;
19802 		rack->r_late = 0;
19803 		rack->r_ctl.rc_agg_early = 0;
19804 	}
19805 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
19806 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
19807 	if (doing_tlp) {
19808 		rack->rc_tlp_in_progress = 1;
19809 		rack->r_ctl.rc_tlp_cnt_out++;
19810 	}
19811 	if (error == 0) {
19812 		counter_u64_add(rack_total_bytes, len);
19813 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
19814 		if (doing_tlp) {
19815 			rack->rc_last_sent_tlp_past_cumack = 0;
19816 			rack->rc_last_sent_tlp_seq_valid = 1;
19817 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
19818 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
19819 		}
19820 		if (rack->r_ctl.rc_prr_sndcnt >= len)
19821 			rack->r_ctl.rc_prr_sndcnt -= len;
19822 		else
19823 			rack->r_ctl.rc_prr_sndcnt = 0;
19824 	}
19825 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19826 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19827 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19828 		rack->r_ctl.retran_during_recovery += len;
19829 	{
19830 		int idx;
19831 
19832 		idx = (len / segsiz) + 3;
19833 		if (idx >= TCP_MSS_ACCT_ATIMER)
19834 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19835 		else
19836 			counter_u64_add(rack_out_size[idx], 1);
19837 	}
19838 	if (tp->t_rtttime == 0) {
19839 		tp->t_rtttime = ticks;
19840 		tp->t_rtseq = startseq;
19841 		KMOD_TCPSTAT_INC(tcps_segstimed);
19842 	}
19843 	counter_u64_add(rack_fto_rsm_send, 1);
19844 	if (error && (error == ENOBUFS)) {
19845 		if (rack->r_ctl.crte != NULL) {
19846 			tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
19847 			if (tcp_bblogging_on(rack->rc_tp))
19848 				rack_log_queue_level(tp, rack, len, tv, cts);
19849 		} else
19850 			tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
19851 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
19852 		if (rack->rc_enobuf < 0x7f)
19853 			rack->rc_enobuf++;
19854 		if (slot < (10 * HPTS_USEC_IN_MSEC))
19855 			slot = 10 * HPTS_USEC_IN_MSEC;
19856 		if (rack->r_ctl.crte != NULL) {
19857 			counter_u64_add(rack_saw_enobuf_hw, 1);
19858 			tcp_rl_log_enobuf(rack->r_ctl.crte);
19859 		}
19860 		counter_u64_add(rack_saw_enobuf, 1);
19861 	} else {
19862 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz, __LINE__);
19863 	}
19864 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
19865 #ifdef TCP_ACCOUNTING
19866 	crtsc = get_cyclecount();
19867 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19868 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19869 	}
19870 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19871 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19872 	}
19873 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19874 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
19875 	}
19876 	sched_unpin();
19877 #endif
19878 	return (0);
19879 failed:
19880 	if (m)
19881 		m_free(m);
19882 	return (-1);
19883 }
19884 
19885 static void
19886 rack_sndbuf_autoscale(struct tcp_rack *rack)
19887 {
19888 	/*
19889 	 * Automatic sizing of send socket buffer.  Often the send buffer
19890 	 * size is not optimally adjusted to the actual network conditions
19891 	 * at hand (delay bandwidth product).  Setting the buffer size too
19892 	 * small limits throughput on links with high bandwidth and high
19893 	 * delay (eg. trans-continental/oceanic links).  Setting the
19894 	 * buffer size too big consumes too much real kernel memory,
19895 	 * especially with many connections on busy servers.
19896 	 *
19897 	 * The criteria to step up the send buffer one notch are:
19898 	 *  1. receive window of remote host is larger than send buffer
19899 	 *     (with a fudge factor of 5/4th);
19900 	 *  2. send buffer is filled to 7/8th with data (so we actually
19901 	 *     have data to make use of it);
19902 	 *  3. send buffer fill has not hit maximal automatic size;
19903 	 *  4. our send window (slow start and cogestion controlled) is
19904 	 *     larger than sent but unacknowledged data in send buffer.
19905 	 *
19906 	 * Note that the rack version moves things much faster since
19907 	 * we want to avoid hitting cache lines in the rack_fast_output()
19908 	 * path so this is called much less often and thus moves
19909 	 * the SB forward by a percentage.
19910 	 */
19911 	struct socket *so;
19912 	struct tcpcb *tp;
19913 	uint32_t sendwin, scaleup;
19914 
19915 	tp = rack->rc_tp;
19916 	so = rack->rc_inp->inp_socket;
19917 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
19918 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
19919 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
19920 		    sbused(&so->so_snd) >=
19921 		    (so->so_snd.sb_hiwat / 8 * 7) &&
19922 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
19923 		    sendwin >= (sbused(&so->so_snd) -
19924 		    (tp->snd_max - tp->snd_una))) {
19925 			if (rack_autosndbuf_inc)
19926 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
19927 			else
19928 				scaleup = V_tcp_autosndbuf_inc;
19929 			if (scaleup < V_tcp_autosndbuf_inc)
19930 				scaleup = V_tcp_autosndbuf_inc;
19931 			scaleup += so->so_snd.sb_hiwat;
19932 			if (scaleup > V_tcp_autosndbuf_max)
19933 				scaleup = V_tcp_autosndbuf_max;
19934 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
19935 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
19936 		}
19937 	}
19938 }
19939 
19940 static int
19941 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
19942 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
19943 {
19944 	/*
19945 	 * Enter to do fast output. We are given that the sched_pin is
19946 	 * in place (if accounting is compiled in) and the cycle count taken
19947 	 * at entry is in place in ts_val. The idea here is that
19948 	 * we know how many more bytes needs to be sent (presumably either
19949 	 * during pacing or to fill the cwnd and that was greater than
19950 	 * the max-burst). We have how much to send and all the info we
19951 	 * need to just send.
19952 	 */
19953 #ifdef INET
19954 	struct ip *ip = NULL;
19955 #endif
19956 	struct udphdr *udp = NULL;
19957 	struct tcphdr *th = NULL;
19958 	struct mbuf *m, *s_mb;
19959 	struct inpcb *inp;
19960 	uint8_t *cpto;
19961 	struct tcp_log_buffer *lgb;
19962 #ifdef TCP_ACCOUNTING
19963 	uint64_t crtsc;
19964 #endif
19965 	struct tcpopt to;
19966 	u_char opt[TCP_MAXOLEN];
19967 	uint32_t hdrlen, optlen;
19968 #ifdef TCP_ACCOUNTING
19969 	int cnt_thru = 1;
19970 #endif
19971 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19972 	uint16_t flags;
19973 	uint32_t s_soff;
19974 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
19975 	uint32_t if_hw_tsomaxsegsize;
19976 	uint32_t add_flag = RACK_SENT_FP;
19977 #ifdef INET6
19978 	struct ip6_hdr *ip6 = NULL;
19979 
19980 	if (rack->r_is_v6) {
19981 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19982 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19983 	} else
19984 #endif				/* INET6 */
19985 	{
19986 #ifdef INET
19987 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19988 		hdrlen = sizeof(struct tcpiphdr);
19989 #endif
19990 	}
19991 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19992 		m = NULL;
19993 		goto failed;
19994 	}
19995 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19996 	startseq = tp->snd_max;
19997 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19998 	inp = rack->rc_inp;
19999 	len = rack->r_ctl.fsb.left_to_send;
20000 	to.to_flags = 0;
20001 	flags = rack->r_ctl.fsb.tcp_flags;
20002 	if (tp->t_flags & TF_RCVD_TSTMP) {
20003 		to.to_tsval = ms_cts + tp->ts_offset;
20004 		to.to_tsecr = tp->ts_recent;
20005 		to.to_flags = TOF_TS;
20006 	}
20007 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
20008 	/* TCP-MD5 (RFC2385). */
20009 	if (tp->t_flags & TF_SIGNATURE)
20010 		to.to_flags |= TOF_SIGNATURE;
20011 #endif
20012 	optlen = tcp_addoptions(&to, opt);
20013 	hdrlen += optlen;
20014 	udp = rack->r_ctl.fsb.udp;
20015 	if (udp)
20016 		hdrlen += sizeof(struct udphdr);
20017 	if (rack->r_ctl.rc_pace_max_segs)
20018 		max_val = rack->r_ctl.rc_pace_max_segs;
20019 	else if (rack->rc_user_set_max_segs)
20020 		max_val = rack->rc_user_set_max_segs * segsiz;
20021 	else
20022 		max_val = len;
20023 	if ((tp->t_flags & TF_TSO) &&
20024 	    V_tcp_do_tso &&
20025 	    (len > segsiz) &&
20026 	    (tp->t_port == 0))
20027 		tso = 1;
20028 again:
20029 #ifdef INET6
20030 	if (MHLEN < hdrlen + max_linkhdr)
20031 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
20032 	else
20033 #endif
20034 		m = m_gethdr(M_NOWAIT, MT_DATA);
20035 	if (m == NULL)
20036 		goto failed;
20037 	m->m_data += max_linkhdr;
20038 	m->m_len = hdrlen;
20039 	th = rack->r_ctl.fsb.th;
20040 	/* Establish the len to send */
20041 	if (len > max_val)
20042 		len = max_val;
20043 	if ((tso) && (len + optlen > segsiz)) {
20044 		uint32_t if_hw_tsomax;
20045 		int32_t max_len;
20046 
20047 		/* extract TSO information */
20048 		if_hw_tsomax = tp->t_tsomax;
20049 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
20050 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
20051 		/*
20052 		 * Check if we should limit by maximum payload
20053 		 * length:
20054 		 */
20055 		if (if_hw_tsomax != 0) {
20056 			/* compute maximum TSO length */
20057 			max_len = (if_hw_tsomax - hdrlen -
20058 				   max_linkhdr);
20059 			if (max_len <= 0) {
20060 				goto failed;
20061 			} else if (len > max_len) {
20062 				len = max_len;
20063 			}
20064 		}
20065 		if (len <= segsiz) {
20066 			/*
20067 			 * In case there are too many small fragments don't
20068 			 * use TSO:
20069 			 */
20070 			tso = 0;
20071 		}
20072 	} else {
20073 		tso = 0;
20074 	}
20075 	if ((tso == 0) && (len > segsiz))
20076 		len = segsiz;
20077 	(void)tcp_get_usecs(tv);
20078 	if ((len == 0) ||
20079 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
20080 		goto failed;
20081 	}
20082 	sb_offset = tp->snd_max - tp->snd_una;
20083 	th->th_seq = htonl(tp->snd_max);
20084 	th->th_ack = htonl(tp->rcv_nxt);
20085 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
20086 	if (th->th_win == 0) {
20087 		tp->t_sndzerowin++;
20088 		tp->t_flags |= TF_RXWIN0SENT;
20089 	} else
20090 		tp->t_flags &= ~TF_RXWIN0SENT;
20091 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
20092 	KMOD_TCPSTAT_INC(tcps_sndpack);
20093 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
20094 #ifdef STATS
20095 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
20096 				 len);
20097 #endif
20098 	if (rack->r_ctl.fsb.m == NULL)
20099 		goto failed;
20100 
20101 	/* s_mb and s_soff are saved for rack_log_output */
20102 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
20103 				    &s_mb, &s_soff);
20104 	if (len <= segsiz) {
20105 		/*
20106 		 * Must have ran out of mbufs for the copy
20107 		 * shorten it to no longer need tso. Lets
20108 		 * not put on sendalot since we are low on
20109 		 * mbufs.
20110 		 */
20111 		tso = 0;
20112 	}
20113 	if (rack->r_ctl.fsb.rfo_apply_push &&
20114 	    (len == rack->r_ctl.fsb.left_to_send)) {
20115 		tcp_set_flags(th, flags | TH_PUSH);
20116 		add_flag |= RACK_HAD_PUSH;
20117 	}
20118 	if ((m->m_next == NULL) || (len <= 0)){
20119 		goto failed;
20120 	}
20121 	if (udp) {
20122 		if (rack->r_is_v6)
20123 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
20124 		else
20125 			ulen = hdrlen + len - sizeof(struct ip);
20126 		udp->uh_ulen = htons(ulen);
20127 	}
20128 	m->m_pkthdr.rcvif = (struct ifnet *)0;
20129 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
20130 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
20131 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
20132 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
20133 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
20134 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
20135 #ifdef INET6
20136 		if (rack->r_is_v6) {
20137 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
20138 			ip6->ip6_flow |= htonl(ect << 20);
20139 		}
20140 		else
20141 #endif
20142 		{
20143 #ifdef INET
20144 			ip->ip_tos &= ~IPTOS_ECN_MASK;
20145 			ip->ip_tos |= ect;
20146 #endif
20147 		}
20148 	}
20149 	tcp_set_flags(th, flags);
20150 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
20151 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
20152 	if (to.to_flags & TOF_SIGNATURE) {
20153 		/*
20154 		 * Calculate MD5 signature and put it into the place
20155 		 * determined before.
20156 		 * NOTE: since TCP options buffer doesn't point into
20157 		 * mbuf's data, calculate offset and use it.
20158 		 */
20159 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
20160 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
20161 			/*
20162 			 * Do not send segment if the calculation of MD5
20163 			 * digest has failed.
20164 			 */
20165 			goto failed;
20166 		}
20167 	}
20168 #endif
20169 #ifdef INET6
20170 	if (rack->r_is_v6) {
20171 		if (tp->t_port) {
20172 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
20173 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
20174 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
20175 			th->th_sum = htons(0);
20176 			UDPSTAT_INC(udps_opackets);
20177 		} else {
20178 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
20179 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
20180 			th->th_sum = in6_cksum_pseudo(ip6,
20181 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
20182 						      0);
20183 		}
20184 	}
20185 #endif
20186 #if defined(INET6) && defined(INET)
20187 	else
20188 #endif
20189 #ifdef INET
20190 	{
20191 		if (tp->t_port) {
20192 			m->m_pkthdr.csum_flags = CSUM_UDP;
20193 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
20194 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
20195 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
20196 			th->th_sum = htons(0);
20197 			UDPSTAT_INC(udps_opackets);
20198 		} else {
20199 			m->m_pkthdr.csum_flags = CSUM_TCP;
20200 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
20201 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
20202 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
20203 									IPPROTO_TCP + len + optlen));
20204 		}
20205 		/* IP version must be set here for ipv4/ipv6 checking later */
20206 		KASSERT(ip->ip_v == IPVERSION,
20207 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
20208 	}
20209 #endif
20210 	if (tso) {
20211 		/*
20212 		 * Here we use segsiz since we have no added options besides
20213 		 * any standard timestamp options (no DSACKs or SACKS are sent
20214 		 * via either fast-path).
20215 		 */
20216 		KASSERT(len > segsiz,
20217 			("%s: len <= tso_segsz tp:%p", __func__, tp));
20218 		m->m_pkthdr.csum_flags |= CSUM_TSO;
20219 		m->m_pkthdr.tso_segsz = segsiz;
20220 	}
20221 #ifdef INET6
20222 	if (rack->r_is_v6) {
20223 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
20224 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
20225 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
20226 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
20227 		else
20228 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
20229 	}
20230 #endif
20231 #if defined(INET) && defined(INET6)
20232 	else
20233 #endif
20234 #ifdef INET
20235 	{
20236 		ip->ip_len = htons(m->m_pkthdr.len);
20237 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
20238 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
20239 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
20240 			if (tp->t_port == 0 || len < V_tcp_minmss) {
20241 				ip->ip_off |= htons(IP_DF);
20242 			}
20243 		} else {
20244 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
20245 		}
20246 	}
20247 #endif
20248 	if (tp->snd_cwnd > tp->snd_ssthresh) {
20249 		/* Set we sent in CA */
20250 		rack->rc_gp_saw_ca = 1;
20251 	} else {
20252 		/* Set we sent in SS */
20253 		rack->rc_gp_saw_ss = 1;
20254 	}
20255 	/* Time to copy in our header */
20256 	cpto = mtod(m, uint8_t *);
20257 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
20258 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
20259 	if (optlen) {
20260 		bcopy(opt, th + 1, optlen);
20261 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
20262 	} else {
20263 		th->th_off = sizeof(struct tcphdr) >> 2;
20264 	}
20265 	if ((rack->r_ctl.crte != NULL) &&
20266 	    tcp_bblogging_on(tp)) {
20267 		rack_log_queue_level(tp, rack, len, tv, cts);
20268 	}
20269 	if (tcp_bblogging_on(rack->rc_tp)) {
20270 		union tcp_log_stackspecific log;
20271 
20272 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
20273 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
20274 		if (rack->rack_no_prr)
20275 			log.u_bbr.flex1 = 0;
20276 		else
20277 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
20278 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
20279 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
20280 		log.u_bbr.flex4 = max_val;
20281 		/* Save off the early/late values */
20282 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
20283 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
20284 		log.u_bbr.bw_inuse = rack_get_bw(rack);
20285 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
20286 		log.u_bbr.flex8 = 0;
20287 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
20288 		log.u_bbr.flex7 = 44;
20289 		log.u_bbr.pkts_out = tp->t_maxseg;
20290 		log.u_bbr.timeStamp = cts;
20291 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20292 		log.u_bbr.flex5 = log.u_bbr.inflight;
20293 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
20294 		log.u_bbr.delivered = 0;
20295 		log.u_bbr.rttProp = 0;
20296 		log.u_bbr.delRate = rack->r_must_retran;
20297 		log.u_bbr.delRate <<= 1;
20298 		log.u_bbr.pkt_epoch = __LINE__;
20299 		/* For fast output no retrans so just inflight and how many mss we send */
20300 		log.u_bbr.flex5 = log.u_bbr.inflight;
20301 		log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
20302 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
20303 				     len, &log, false, NULL, __func__, __LINE__, tv);
20304 	} else
20305 		lgb = NULL;
20306 #ifdef INET6
20307 	if (rack->r_is_v6) {
20308 		error = ip6_output(m, inp->in6p_outputopts,
20309 				   &inp->inp_route6,
20310 				   0, NULL, NULL, inp);
20311 	}
20312 #endif
20313 #if defined(INET) && defined(INET6)
20314 	else
20315 #endif
20316 #ifdef INET
20317 	{
20318 		error = ip_output(m, NULL,
20319 				  &inp->inp_route,
20320 				  0, 0, inp);
20321 	}
20322 #endif
20323 	if (lgb) {
20324 		lgb->tlb_errno = error;
20325 		lgb = NULL;
20326 	}
20327 	if (error) {
20328 		*send_err = error;
20329 		m = NULL;
20330 		goto failed;
20331 	} else if (rack->rc_hw_nobuf) {
20332 		rack->rc_hw_nobuf = 0;
20333 		rack->r_ctl.rc_agg_delayed = 0;
20334 		rack->r_early = 0;
20335 		rack->r_late = 0;
20336 		rack->r_ctl.rc_agg_early = 0;
20337 	}
20338 	if ((error == 0) && (rack->lt_bw_up == 0)) {
20339 		/* Unlikely */
20340 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
20341 		rack->r_ctl.lt_seq = tp->snd_una;
20342 		rack->lt_bw_up = 1;
20343 	} else if ((error == 0) &&
20344 		   (((tp->snd_max + len) - rack->r_ctl.lt_seq) > 0x7fffffff)) {
20345 		/*
20346 		 * Need to record what we have since we are
20347 		 * approaching seq wrap.
20348 		 */
20349 		struct timeval tv;
20350 		uint64_t tmark;
20351 
20352 		rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
20353 		rack->r_ctl.lt_seq = tp->snd_una;
20354 		tmark = tcp_get_u64_usecs(&tv);
20355 		if (tmark > rack->r_ctl.lt_timemark) {
20356 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
20357 			rack->r_ctl.lt_timemark = tmark;
20358 		}
20359 	}
20360 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
20361 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
20362 	m = NULL;
20363 	if (tp->snd_una == tp->snd_max) {
20364 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
20365 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
20366 		tp->t_acktime = ticks;
20367 	}
20368 	counter_u64_add(rack_total_bytes, len);
20369 	tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
20370 
20371 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
20372 	tot_len += len;
20373 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
20374 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
20375 	tp->snd_max += len;
20376 	tp->snd_nxt = tp->snd_max;
20377 	if (rack->rc_new_rnd_needed) {
20378 		rack_new_round_starts(tp, rack, tp->snd_max);
20379 	}
20380 	{
20381 		int idx;
20382 
20383 		idx = (len / segsiz) + 3;
20384 		if (idx >= TCP_MSS_ACCT_ATIMER)
20385 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
20386 		else
20387 			counter_u64_add(rack_out_size[idx], 1);
20388 	}
20389 	if (len <= rack->r_ctl.fsb.left_to_send)
20390 		rack->r_ctl.fsb.left_to_send -= len;
20391 	else
20392 		rack->r_ctl.fsb.left_to_send = 0;
20393 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
20394 		rack->r_fast_output = 0;
20395 		rack->r_ctl.fsb.left_to_send = 0;
20396 		/* At the end of fast_output scale up the sb */
20397 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
20398 		rack_sndbuf_autoscale(rack);
20399 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
20400 	}
20401 	if (tp->t_rtttime == 0) {
20402 		tp->t_rtttime = ticks;
20403 		tp->t_rtseq = startseq;
20404 		KMOD_TCPSTAT_INC(tcps_segstimed);
20405 	}
20406 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
20407 	    (max_val > len) &&
20408 	    (tso == 0)) {
20409 		max_val -= len;
20410 		len = segsiz;
20411 		th = rack->r_ctl.fsb.th;
20412 #ifdef TCP_ACCOUNTING
20413 		cnt_thru++;
20414 #endif
20415 		goto again;
20416 	}
20417 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
20418 	counter_u64_add(rack_fto_send, 1);
20419 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz, __LINE__);
20420 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
20421 #ifdef TCP_ACCOUNTING
20422 	crtsc = get_cyclecount();
20423 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20424 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
20425 	}
20426 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20427 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
20428 	}
20429 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20430 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
20431 	}
20432 	sched_unpin();
20433 #endif
20434 	return (0);
20435 failed:
20436 	if (m)
20437 		m_free(m);
20438 	rack->r_fast_output = 0;
20439 	return (-1);
20440 }
20441 
20442 static inline void
20443 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
20444 		       struct sockbuf *sb,
20445 		       int len, int orig_len, int segsiz, uint32_t pace_max_seg,
20446 		       bool hw_tls,
20447 		       uint16_t flags)
20448 {
20449 	rack->r_fast_output = 1;
20450 	rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
20451 	rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
20452 	rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
20453 	rack->r_ctl.fsb.tcp_flags = flags;
20454 	rack->r_ctl.fsb.left_to_send = orig_len - len;
20455 	if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
20456 		/* Less than a full sized pace, lets not  */
20457 		rack->r_fast_output = 0;
20458 		return;
20459 	} else {
20460 		/* Round down to the nearest pace_max_seg */
20461 		rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
20462 	}
20463 	if (hw_tls)
20464 		rack->r_ctl.fsb.hw_tls = 1;
20465 	else
20466 		rack->r_ctl.fsb.hw_tls = 0;
20467 	KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
20468 		("rack:%p left_to_send:%u sbavail:%u out:%u",
20469 		 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
20470 		 (tp->snd_max - tp->snd_una)));
20471 	if (rack->r_ctl.fsb.left_to_send < segsiz)
20472 		rack->r_fast_output = 0;
20473 	else {
20474 		if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
20475 			rack->r_ctl.fsb.rfo_apply_push = 1;
20476 		else
20477 			rack->r_ctl.fsb.rfo_apply_push = 0;
20478 	}
20479 }
20480 
20481 static uint32_t
20482 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
20483 {
20484 	uint64_t min_time;
20485 	uint32_t maxlen;
20486 
20487 	min_time = (uint64_t)get_hpts_min_sleep_time();
20488 	maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
20489 	maxlen = roundup(maxlen, segsiz);
20490 	return (maxlen);
20491 }
20492 
20493 static struct rack_sendmap *
20494 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
20495 {
20496 	struct rack_sendmap *rsm = NULL;
20497 	int thresh;
20498 
20499 restart:
20500 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
20501 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
20502 		/* Nothing, strange turn off validity  */
20503 		rack->r_collapse_point_valid = 0;
20504 		return (NULL);
20505 	}
20506 	/* Can we send it yet? */
20507 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
20508 		/*
20509 		 * Receiver window has not grown enough for
20510 		 * the segment to be put on the wire.
20511 		 */
20512 		return (NULL);
20513 	}
20514 	if (rsm->r_flags & RACK_ACKED) {
20515 		/*
20516 		 * It has been sacked, lets move to the
20517 		 * next one if possible.
20518 		 */
20519 		rack->r_ctl.last_collapse_point = rsm->r_end;
20520 		/* Are we done? */
20521 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
20522 			    rack->r_ctl.high_collapse_point)) {
20523 			rack->r_collapse_point_valid = 0;
20524 			return (NULL);
20525 		}
20526 		goto restart;
20527 	}
20528 	/* Now has it been long enough ? */
20529 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts, __LINE__, 1);
20530 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
20531 		rack_log_collapse(rack, rsm->r_start,
20532 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
20533 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
20534 		return (rsm);
20535 	}
20536 	/* Not enough time */
20537 	rack_log_collapse(rack, rsm->r_start,
20538 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
20539 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
20540 	return (NULL);
20541 }
20542 
20543 static void
20544 rack_credit_back_policer_idle_time(struct tcp_rack *rack, uint64_t idle_t, int line)
20545 {
20546 	/*
20547 	 * We were idle some time (idle_t) and so our policer bucket
20548 	 * needs to grow. It can go no higher than policer_bucket_size.
20549 	 */
20550 	uint64_t len;
20551 
20552 	len = idle_t * rack->r_ctl.policer_bw;
20553 	len /= HPTS_USEC_IN_SEC;
20554 	rack->r_ctl.current_policer_bucket += (uint32_t)len;
20555 	if (rack->r_ctl.policer_bucket_size < rack->r_ctl.current_policer_bucket) {
20556 		rack->r_ctl.current_policer_bucket = rack->r_ctl.policer_bucket_size;
20557 	}
20558 	if (rack_verbose_logging > 0)
20559 		policer_detection_log(rack, (uint32_t)len, line, (uint32_t)idle_t, 0, 7);
20560 }
20561 
20562 static inline void
20563 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
20564 {
20565 	if ((rack->full_size_rxt == 0) &&
20566 	    (rack->shape_rxt_to_pacing_min == 0) &&
20567 	    (*len >= segsiz)) {
20568 		*len = segsiz;
20569 	} else if (rack->shape_rxt_to_pacing_min &&
20570 		 rack->gp_ready) {
20571 		/* We use pacing min as shaping len req */
20572 		uint32_t maxlen;
20573 
20574 		maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20575 		if (*len > maxlen)
20576 			*len = maxlen;
20577 	} else {
20578 		/*
20579 		 * The else is full_size_rxt is on so send it all
20580 		 * note we do need to check this for exceeding
20581 		 * our max segment size due to the fact that
20582 		 * we do sometimes merge chunks together i.e.
20583 		 * we cannot just assume that we will never have
20584 		 * a chunk greater than pace_max_seg
20585 		 */
20586 		if (*len > pace_max_seg)
20587 			*len = pace_max_seg;
20588 	}
20589 }
20590 
20591 static int
20592 rack_output(struct tcpcb *tp)
20593 {
20594 	struct socket *so;
20595 	uint32_t recwin;
20596 	uint32_t sb_offset, s_moff = 0;
20597 	int32_t len, error = 0;
20598 	uint16_t flags;
20599 	struct mbuf *m, *s_mb = NULL;
20600 	struct mbuf *mb;
20601 	uint32_t if_hw_tsomaxsegcount = 0;
20602 	uint32_t if_hw_tsomaxsegsize;
20603 	int32_t segsiz, minseg;
20604 	long tot_len_this_send = 0;
20605 #ifdef INET
20606 	struct ip *ip = NULL;
20607 #endif
20608 	struct udphdr *udp = NULL;
20609 	struct tcp_rack *rack;
20610 	struct tcphdr *th;
20611 	uint8_t pass = 0;
20612 	uint8_t mark = 0;
20613 	uint8_t check_done = 0;
20614 	uint8_t wanted_cookie = 0;
20615 	u_char opt[TCP_MAXOLEN];
20616 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
20617 	uint32_t rack_seq;
20618 
20619 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20620 	unsigned ipsec_optlen = 0;
20621 
20622 #endif
20623 	int32_t idle, sendalot;
20624 	uint32_t tot_idle;
20625 	int32_t sub_from_prr = 0;
20626 	volatile int32_t sack_rxmit;
20627 	struct rack_sendmap *rsm = NULL;
20628 	int32_t tso, mtu;
20629 	struct tcpopt to;
20630 	int32_t slot = 0;
20631 	int32_t sup_rack = 0;
20632 	uint32_t cts, ms_cts, delayed, early;
20633 	uint32_t add_flag = RACK_SENT_SP;
20634 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
20635 	uint8_t doing_tlp = 0;
20636 	uint32_t cwnd_to_use, pace_max_seg;
20637 	int32_t do_a_prefetch = 0;
20638 	int32_t prefetch_rsm = 0;
20639 	int32_t orig_len = 0;
20640 	struct timeval tv;
20641 	int32_t prefetch_so_done = 0;
20642 	struct tcp_log_buffer *lgb;
20643 	struct inpcb *inp = tptoinpcb(tp);
20644 	struct sockbuf *sb;
20645 	uint64_t ts_val = 0;
20646 #ifdef TCP_ACCOUNTING
20647 	uint64_t crtsc;
20648 #endif
20649 #ifdef INET6
20650 	struct ip6_hdr *ip6 = NULL;
20651 	int32_t isipv6;
20652 #endif
20653 	bool hpts_calling, hw_tls = false;
20654 
20655 	NET_EPOCH_ASSERT();
20656 	INP_WLOCK_ASSERT(inp);
20657 
20658 	/* setup and take the cache hits here */
20659 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20660 #ifdef TCP_ACCOUNTING
20661 	sched_pin();
20662 	ts_val = get_cyclecount();
20663 #endif
20664 	hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
20665 	tp->t_flags2 &= ~TF2_HPTS_CALLS;
20666 #ifdef TCP_OFFLOAD
20667 	if (tp->t_flags & TF_TOE) {
20668 #ifdef TCP_ACCOUNTING
20669 		sched_unpin();
20670 #endif
20671 		return (tcp_offload_output(tp));
20672 	}
20673 #endif
20674 	if (rack->rack_deferred_inited == 0) {
20675 		/*
20676 		 * If we are the connecting socket we will
20677 		 * hit rack_init() when no sequence numbers
20678 		 * are setup. This makes it so we must defer
20679 		 * some initialization. Call that now.
20680 		 */
20681 		rack_deferred_init(tp, rack);
20682 	}
20683 	/*
20684 	 * For TFO connections in SYN_RECEIVED, only allow the initial
20685 	 * SYN|ACK and those sent by the retransmit timer.
20686 	 */
20687 	if ((tp->t_flags & TF_FASTOPEN) &&
20688 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
20689 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
20690 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
20691 #ifdef TCP_ACCOUNTING
20692 		sched_unpin();
20693 #endif
20694 		return (0);
20695 	}
20696 #ifdef INET6
20697 	if (rack->r_state) {
20698 		/* Use the cache line loaded if possible */
20699 		isipv6 = rack->r_is_v6;
20700 	} else {
20701 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
20702 	}
20703 #endif
20704 	early = 0;
20705 	cts = tcp_get_usecs(&tv);
20706 	ms_cts = tcp_tv_to_mssectick(&tv);
20707 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
20708 	    tcp_in_hpts(rack->rc_tp)) {
20709 		/*
20710 		 * We are on the hpts for some timer but not hptsi output.
20711 		 * Remove from the hpts unconditionally.
20712 		 */
20713 		rack_timer_cancel(tp, rack, cts, __LINE__);
20714 	}
20715 	/* Are we pacing and late? */
20716 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
20717 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
20718 		/* We are delayed */
20719 		delayed = cts - rack->r_ctl.rc_last_output_to;
20720 	} else {
20721 		delayed = 0;
20722 	}
20723 	/* Do the timers, which may override the pacer */
20724 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
20725 		int retval;
20726 
20727 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
20728 					     &doing_tlp);
20729 		if (retval != 0) {
20730 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
20731 #ifdef TCP_ACCOUNTING
20732 			sched_unpin();
20733 #endif
20734 			/*
20735 			 * If timers want tcp_drop(), then pass error out,
20736 			 * otherwise suppress it.
20737 			 */
20738 			return (retval < 0 ? retval : 0);
20739 		}
20740 	}
20741 	if (rack->rc_in_persist) {
20742 		if (tcp_in_hpts(rack->rc_tp) == 0) {
20743 			/* Timer is not running */
20744 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
20745 		}
20746 #ifdef TCP_ACCOUNTING
20747 		sched_unpin();
20748 #endif
20749 		return (0);
20750 	}
20751 	if ((rack->rc_ack_required == 1) &&
20752 	    (rack->r_timer_override == 0)){
20753 		/* A timeout occurred and no ack has arrived */
20754 		if (tcp_in_hpts(rack->rc_tp) == 0) {
20755 			/* Timer is not running */
20756 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
20757 		}
20758 #ifdef TCP_ACCOUNTING
20759 		sched_unpin();
20760 #endif
20761 		return (0);
20762 	}
20763 	if ((rack->r_timer_override) ||
20764 	    (rack->rc_ack_can_sendout_data) ||
20765 	    (delayed) ||
20766 	    (tp->t_state < TCPS_ESTABLISHED)) {
20767 		rack->rc_ack_can_sendout_data = 0;
20768 		if (tcp_in_hpts(rack->rc_tp))
20769 			tcp_hpts_remove(rack->rc_tp);
20770 	} else if (tcp_in_hpts(rack->rc_tp)) {
20771 		/*
20772 		 * On the hpts you can't pass even if ACKNOW is on, we will
20773 		 * when the hpts fires.
20774 		 */
20775 #ifdef TCP_ACCOUNTING
20776 		crtsc = get_cyclecount();
20777 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20778 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
20779 		}
20780 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20781 			tp->tcp_cnt_counters[SND_BLOCKED]++;
20782 		}
20783 		sched_unpin();
20784 #endif
20785 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
20786 		return (0);
20787 	}
20788 	/* Finish out both pacing early and late accounting */
20789 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
20790 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
20791 		early = rack->r_ctl.rc_last_output_to - cts;
20792 	} else
20793 		early = 0;
20794 	if (delayed && (rack->rc_always_pace == 1)) {
20795 		rack->r_ctl.rc_agg_delayed += delayed;
20796 		rack->r_late = 1;
20797 	} else if (early && (rack->rc_always_pace == 1)) {
20798 		rack->r_ctl.rc_agg_early += early;
20799 		rack->r_early = 1;
20800 	} else if (rack->rc_always_pace == 0) {
20801 		/* Non-paced we are not late */
20802 		rack->r_ctl.rc_agg_delayed = rack->r_ctl.rc_agg_early = 0;
20803 		rack->r_early = rack->r_late = 0;
20804 	}
20805 	/* Now that early/late accounting is done turn off the flag */
20806 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
20807 	rack->r_wanted_output = 0;
20808 	rack->r_timer_override = 0;
20809 	if ((tp->t_state != rack->r_state) &&
20810 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
20811 		rack_set_state(tp, rack);
20812 	}
20813 	if ((rack->r_fast_output) &&
20814 	    (doing_tlp == 0) &&
20815 	    (tp->rcv_numsacks == 0)) {
20816 		int ret;
20817 
20818 		error = 0;
20819 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
20820 		if (ret >= 0)
20821 			return(ret);
20822 		else if (error) {
20823 			inp = rack->rc_inp;
20824 			so = inp->inp_socket;
20825 			sb = &so->so_snd;
20826 			goto nomore;
20827 		}
20828 	}
20829 	inp = rack->rc_inp;
20830 	/*
20831 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
20832 	 * only allow the initial SYN or SYN|ACK and those sent
20833 	 * by the retransmit timer.
20834 	 */
20835 	if ((tp->t_flags & TF_FASTOPEN) &&
20836 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
20837 	     (tp->t_state == TCPS_SYN_SENT)) &&
20838 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
20839 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
20840 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20841 		so = inp->inp_socket;
20842 		sb = &so->so_snd;
20843 		goto just_return_nolock;
20844 	}
20845 	/*
20846 	 * Determine length of data that should be transmitted, and flags
20847 	 * that will be used. If there is some data or critical controls
20848 	 * (SYN, RST) to send, then transmit; otherwise, investigate
20849 	 * further.
20850 	 */
20851 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
20852 	if (tp->t_idle_reduce) {
20853 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
20854 			rack_cc_after_idle(rack, tp);
20855 	}
20856 	tp->t_flags &= ~TF_LASTIDLE;
20857 	if (idle) {
20858 		if (tp->t_flags & TF_MORETOCOME) {
20859 			tp->t_flags |= TF_LASTIDLE;
20860 			idle = 0;
20861 		}
20862 	}
20863 	if ((tp->snd_una == tp->snd_max) &&
20864 	    rack->r_ctl.rc_went_idle_time &&
20865 	    (cts > rack->r_ctl.rc_went_idle_time)) {
20866 		tot_idle = (cts - rack->r_ctl.rc_went_idle_time);
20867 		if (tot_idle > rack_min_probertt_hold) {
20868 			/* Count as a probe rtt */
20869 			if (rack->in_probe_rtt == 0) {
20870 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
20871 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
20872 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
20873 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
20874 			} else {
20875 				rack_exit_probertt(rack, cts);
20876 			}
20877 		}
20878 	}
20879 	if(rack->policer_detect_on) {
20880 		/*
20881 		 * If we are doing policer detetion we at a minium
20882 		 * record the time but if possible add back to
20883 		 * the bucket based on the idle time.
20884 		 */
20885 		uint64_t idle_t, u64_cts;
20886 
20887 		segsiz = min(ctf_fixed_maxseg(tp),
20888 			     rack->r_ctl.rc_pace_min_segs);
20889 		u64_cts = tcp_tv_to_lusectick(&tv);
20890 		if ((rack->rc_policer_detected == 1) &&
20891 		    (rack->r_ctl.policer_bucket_size > segsiz) &&
20892 		    (rack->r_ctl.policer_bw > 0) &&
20893 		    (u64_cts > rack->r_ctl.last_sendtime)) {
20894 			/* We are being policed add back the time */
20895 			idle_t = u64_cts - rack->r_ctl.last_sendtime;
20896 			rack_credit_back_policer_idle_time(rack, idle_t, __LINE__);
20897 		}
20898 		rack->r_ctl.last_sendtime = u64_cts;
20899 	}
20900 	if (rack_use_fsb &&
20901 	    (rack->r_ctl.fsb.tcp_ip_hdr) &&
20902 	    (rack->r_fsb_inited == 0) &&
20903 	    (rack->r_state != TCPS_CLOSED))
20904 		rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
20905 	if (rack->rc_sendvars_notset == 1) {
20906 		rack->r_ctl.idle_snd_una = tp->snd_una;
20907 		rack->rc_sendvars_notset = 0;
20908 		/*
20909 		 * Make sure any TCP timers (keep-alive) is not running.
20910 		 */
20911 		tcp_timer_stop(tp);
20912 	}
20913 	if ((rack->rack_no_prr == 1) &&
20914 	    (rack->rc_always_pace == 0)) {
20915 		/*
20916 		 * Sanity check before sending, if we have
20917 		 * no-pacing enabled and prr is turned off that
20918 		 * is a logistics error. Correct this by turnning
20919 		 * prr back on. A user *must* set some form of
20920 		 * pacing in order to turn PRR off. We do this
20921 		 * in the output path so that we can avoid socket
20922 		 * option ordering issues that would occur if we
20923 		 * tried to do it while setting rack_no_prr on.
20924 		 */
20925 		rack->rack_no_prr = 0;
20926 	}
20927 	if ((rack->pcm_enabled == 1) &&
20928 	    (rack->pcm_needed == 0) &&
20929 	    (tot_idle > 0)) {
20930 		/*
20931 		 * We have been idle some micro seconds. We need
20932 		 * to factor this in to see if a PCM is needed.
20933 		 */
20934 		uint32_t rtts_idle, rnds;
20935 
20936 		if (tp->t_srtt)
20937 			rtts_idle = tot_idle / tp->t_srtt;
20938 		else
20939 			rtts_idle = 0;
20940 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
20941 		rack->r_ctl.pcm_idle_rounds += rtts_idle;
20942 		if ((rnds + rack->r_ctl.pcm_idle_rounds)  >= rack_pcm_every_n_rounds) {
20943 			rack->pcm_needed = 1;
20944 			rack_log_pcm(rack, 8, rack->r_ctl.last_pcm_round, rtts_idle, rack->r_ctl.current_round );
20945 		}
20946 	}
20947 again:
20948 	sendalot = 0;
20949 	cts = tcp_get_usecs(&tv);
20950 	ms_cts = tcp_tv_to_mssectick(&tv);
20951 	tso = 0;
20952 	mtu = 0;
20953 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
20954 	minseg = segsiz;
20955 	if (rack->r_ctl.rc_pace_max_segs == 0)
20956 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
20957 	else
20958 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
20959 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
20960 	    (rack->r_ctl.pcm_max_seg == 0)) {
20961 		/*
20962 		 * We set in our first send so we know that the ctf_fixed_maxseg
20963 		 * has been fully set. If we do it in rack_init() we most likely
20964 		 * see 512 bytes so we end up at 5120, not desirable.
20965 		 */
20966 		rack->r_ctl.pcm_max_seg = rc_init_window(rack);
20967 		if (rack->r_ctl.pcm_max_seg < (ctf_fixed_maxseg(tp) * 10)) {
20968 			/*
20969 			 * Assure our initial PCM probe is at least 10 MSS.
20970 			 */
20971 			rack->r_ctl.pcm_max_seg = ctf_fixed_maxseg(tp) * 10;
20972 		}
20973 	}
20974 	if ((rack->r_ctl.pcm_max_seg != 0)  && (rack->pcm_needed == 1)) {
20975 		uint32_t rw_avail, cwa;
20976 
20977 		if (tp->snd_wnd > ctf_outstanding(tp))
20978 			rw_avail = tp->snd_wnd - ctf_outstanding(tp);
20979 		else
20980 			rw_avail = 0;
20981 		if (tp->snd_cwnd > ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked))
20982 			cwa = tp->snd_cwnd -ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20983 		else
20984 			cwa = 0;
20985 		if ((cwa >= rack->r_ctl.pcm_max_seg) &&
20986 		    (rw_avail > rack->r_ctl.pcm_max_seg)) {
20987 			/* Raise up the max seg for this trip through */
20988 			pace_max_seg = rack->r_ctl.pcm_max_seg;
20989 			/* Disable any fast output */
20990 			rack->r_fast_output = 0;
20991 		}
20992 		if (rack_verbose_logging) {
20993 			rack_log_pcm(rack, 4,
20994 				     cwa, rack->r_ctl.pcm_max_seg, rw_avail);
20995 		}
20996 	}
20997 	sb_offset = tp->snd_max - tp->snd_una;
20998 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20999 	flags = tcp_outflags[tp->t_state];
21000 	while (rack->rc_free_cnt < rack_free_cache) {
21001 		rsm = rack_alloc(rack);
21002 		if (rsm == NULL) {
21003 			if (hpts_calling)
21004 				/* Retry in a ms */
21005 				slot = (1 * HPTS_USEC_IN_MSEC);
21006 			so = inp->inp_socket;
21007 			sb = &so->so_snd;
21008 			goto just_return_nolock;
21009 		}
21010 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
21011 		rack->rc_free_cnt++;
21012 		rsm = NULL;
21013 	}
21014 	sack_rxmit = 0;
21015 	len = 0;
21016 	rsm = NULL;
21017 	if (flags & TH_RST) {
21018 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
21019 		so = inp->inp_socket;
21020 		sb = &so->so_snd;
21021 		goto send;
21022 	}
21023 	if (rack->r_ctl.rc_resend) {
21024 		/* Retransmit timer */
21025 		rsm = rack->r_ctl.rc_resend;
21026 		rack->r_ctl.rc_resend = NULL;
21027 		len = rsm->r_end - rsm->r_start;
21028 		sack_rxmit = 1;
21029 		sendalot = 0;
21030 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
21031 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
21032 			 __func__, __LINE__,
21033 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
21034 		sb_offset = rsm->r_start - tp->snd_una;
21035 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
21036 	} else if (rack->r_collapse_point_valid &&
21037 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
21038 		/*
21039 		 * If an RSM is returned then enough time has passed
21040 		 * for us to retransmit it. Move up the collapse point,
21041 		 * since this rsm has its chance to retransmit now.
21042 		 */
21043 		tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
21044 		rack->r_ctl.last_collapse_point = rsm->r_end;
21045 		/* Are we done? */
21046 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
21047 			    rack->r_ctl.high_collapse_point))
21048 			rack->r_collapse_point_valid = 0;
21049 		sack_rxmit = 1;
21050 		/* We are not doing a TLP */
21051 		doing_tlp = 0;
21052 		len = rsm->r_end - rsm->r_start;
21053 		sb_offset = rsm->r_start - tp->snd_una;
21054 		sendalot = 0;
21055 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
21056 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
21057 		/* We have a retransmit that takes precedence */
21058 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
21059 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
21060 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
21061 			/* Enter recovery if not induced by a time-out */
21062 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
21063 		}
21064 #ifdef INVARIANTS
21065 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
21066 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
21067 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
21068 		}
21069 #endif
21070 		len = rsm->r_end - rsm->r_start;
21071 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
21072 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
21073 			 __func__, __LINE__,
21074 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
21075 		sb_offset = rsm->r_start - tp->snd_una;
21076 		sendalot = 0;
21077 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
21078 		if (len > 0) {
21079 			sack_rxmit = 1;
21080 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
21081 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
21082 					 min(len, segsiz));
21083 		}
21084 	} else if (rack->r_ctl.rc_tlpsend) {
21085 		/* Tail loss probe */
21086 		long cwin;
21087 		long tlen;
21088 
21089 		/*
21090 		 * Check if we can do a TLP with a RACK'd packet
21091 		 * this can happen if we are not doing the rack
21092 		 * cheat and we skipped to a TLP and it
21093 		 * went off.
21094 		 */
21095 		rsm = rack->r_ctl.rc_tlpsend;
21096 		/* We are doing a TLP make sure the flag is preent */
21097 		rsm->r_flags |= RACK_TLP;
21098 		rack->r_ctl.rc_tlpsend = NULL;
21099 		sack_rxmit = 1;
21100 		tlen = rsm->r_end - rsm->r_start;
21101 		if (tlen > segsiz)
21102 			tlen = segsiz;
21103 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
21104 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
21105 			 __func__, __LINE__,
21106 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
21107 		sb_offset = rsm->r_start - tp->snd_una;
21108 		cwin = min(tp->snd_wnd, tlen);
21109 		len = cwin;
21110 	}
21111 	if (rack->r_must_retran &&
21112 	    (doing_tlp == 0) &&
21113 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
21114 	    (rsm == NULL)) {
21115 		/*
21116 		 * There are two different ways that we
21117 		 * can get into this block:
21118 		 * a) This is a non-sack connection, we had a time-out
21119 		 *    and thus r_must_retran was set and everything
21120 		 *    left outstanding as been marked for retransmit.
21121 		 * b) The MTU of the path shrank, so that everything
21122 		 *    was marked to be retransmitted with the smaller
21123 		 *    mtu and r_must_retran was set.
21124 		 *
21125 		 * This means that we expect the sendmap (outstanding)
21126 		 * to all be marked must. We can use the tmap to
21127 		 * look at them.
21128 		 *
21129 		 */
21130 		int sendwin, flight;
21131 
21132 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
21133 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
21134 		if (flight >= sendwin) {
21135 			/*
21136 			 * We can't send yet.
21137 			 */
21138 			so = inp->inp_socket;
21139 			sb = &so->so_snd;
21140 			goto just_return_nolock;
21141 		}
21142 		/*
21143 		 * This is the case a/b mentioned above. All
21144 		 * outstanding/not-acked should be marked.
21145 		 * We can use the tmap to find them.
21146 		 */
21147 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
21148 		if (rsm == NULL) {
21149 			/* TSNH */
21150 			rack->r_must_retran = 0;
21151 			rack->r_ctl.rc_out_at_rto = 0;
21152 			so = inp->inp_socket;
21153 			sb = &so->so_snd;
21154 			goto just_return_nolock;
21155 		}
21156 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
21157 			/*
21158 			 * The first one does not have the flag, did we collapse
21159 			 * further up in our list?
21160 			 */
21161 			rack->r_must_retran = 0;
21162 			rack->r_ctl.rc_out_at_rto = 0;
21163 			rsm = NULL;
21164 			sack_rxmit = 0;
21165 		} else {
21166 			sack_rxmit = 1;
21167 			len = rsm->r_end - rsm->r_start;
21168 			sb_offset = rsm->r_start - tp->snd_una;
21169 			sendalot = 0;
21170 			if ((rack->full_size_rxt == 0) &&
21171 			    (rack->shape_rxt_to_pacing_min == 0) &&
21172 			    (len >= segsiz))
21173 				len = segsiz;
21174 			else if (rack->shape_rxt_to_pacing_min &&
21175 				 rack->gp_ready) {
21176 				/* We use pacing min as shaping len req */
21177 				uint32_t maxlen;
21178 
21179 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
21180 				if (len > maxlen)
21181 					len = maxlen;
21182 			}
21183 			/*
21184 			 * Delay removing the flag RACK_MUST_RXT so
21185 			 * that the fastpath for retransmit will
21186 			 * work with this rsm.
21187 			 */
21188 		}
21189 	}
21190 	/*
21191 	 * Enforce a connection sendmap count limit if set
21192 	 * as long as we are not retransmiting.
21193 	 */
21194 	if ((rsm == NULL) &&
21195 	    (V_tcp_map_entries_limit > 0) &&
21196 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
21197 		counter_u64_add(rack_to_alloc_limited, 1);
21198 		if (!rack->alloc_limit_reported) {
21199 			rack->alloc_limit_reported = 1;
21200 			counter_u64_add(rack_alloc_limited_conns, 1);
21201 		}
21202 		so = inp->inp_socket;
21203 		sb = &so->so_snd;
21204 		goto just_return_nolock;
21205 	}
21206 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
21207 		/* we are retransmitting the fin */
21208 		len--;
21209 		if (len) {
21210 			/*
21211 			 * When retransmitting data do *not* include the
21212 			 * FIN. This could happen from a TLP probe.
21213 			 */
21214 			flags &= ~TH_FIN;
21215 		}
21216 	}
21217 	if (rsm && rack->r_fsb_inited &&
21218 	    rack_use_rsm_rfo &&
21219 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
21220 		int ret;
21221 
21222 		if ((rack->rc_policer_detected == 1) &&
21223 		    (rack->r_ctl.policer_bucket_size > segsiz) &&
21224 		    (rack->r_ctl.policer_bw > 0)) {
21225 			/* Check to see if there is room */
21226 			if (rack->r_ctl.current_policer_bucket < len) {
21227 				goto skip_fast_output;
21228 			}
21229 		}
21230 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
21231 		if (ret == 0)
21232 			return (0);
21233 	}
21234 skip_fast_output:
21235 	so = inp->inp_socket;
21236 	sb = &so->so_snd;
21237 	if (do_a_prefetch == 0) {
21238 		kern_prefetch(sb, &do_a_prefetch);
21239 		do_a_prefetch = 1;
21240 	}
21241 #ifdef NETFLIX_SHARED_CWND
21242 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
21243 	    rack->rack_enable_scwnd) {
21244 		/* We are doing cwnd sharing */
21245 		if (rack->gp_ready &&
21246 		    (rack->rack_attempted_scwnd == 0) &&
21247 		    (rack->r_ctl.rc_scw == NULL) &&
21248 		    tp->t_lib) {
21249 			/* The pcbid is in, lets make an attempt */
21250 			counter_u64_add(rack_try_scwnd, 1);
21251 			rack->rack_attempted_scwnd = 1;
21252 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
21253 								   &rack->r_ctl.rc_scw_index,
21254 								   segsiz);
21255 		}
21256 		if (rack->r_ctl.rc_scw &&
21257 		    (rack->rack_scwnd_is_idle == 1) &&
21258 		    sbavail(&so->so_snd)) {
21259 			/* we are no longer out of data */
21260 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21261 			rack->rack_scwnd_is_idle = 0;
21262 		}
21263 		if (rack->r_ctl.rc_scw) {
21264 			/* First lets update and get the cwnd */
21265 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
21266 										       rack->r_ctl.rc_scw_index,
21267 										       tp->snd_cwnd, tp->snd_wnd, segsiz);
21268 		}
21269 	}
21270 #endif
21271 	/*
21272 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
21273 	 * state flags.
21274 	 */
21275 	if (tp->t_flags & TF_NEEDFIN)
21276 		flags |= TH_FIN;
21277 	if (tp->t_flags & TF_NEEDSYN)
21278 		flags |= TH_SYN;
21279 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
21280 		void *end_rsm;
21281 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
21282 		if (end_rsm)
21283 			kern_prefetch(end_rsm, &prefetch_rsm);
21284 		prefetch_rsm = 1;
21285 	}
21286 	SOCKBUF_LOCK(sb);
21287 	if ((sack_rxmit == 0) &&
21288 	    (TCPS_HAVEESTABLISHED(tp->t_state) ||
21289 	    (tp->t_flags & TF_FASTOPEN))) {
21290 		/*
21291 		 * We are not retransmitting (sack_rxmit is 0) so we
21292 		 * are sending new data. This is always based on snd_max.
21293 		 * Now in theory snd_max may be equal to snd_una, if so
21294 		 * then nothing is outstanding and the offset would be 0.
21295 		 */
21296 		uint32_t avail;
21297 
21298 		avail = sbavail(sb);
21299 		if (SEQ_GT(tp->snd_max, tp->snd_una) && avail)
21300 			sb_offset = tp->snd_max - tp->snd_una;
21301 		else
21302 			sb_offset = 0;
21303 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
21304 			if (rack->r_ctl.rc_tlp_new_data) {
21305 				/* TLP is forcing out new data */
21306 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
21307 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
21308 				}
21309 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
21310 					if (tp->snd_wnd > sb_offset)
21311 						len = tp->snd_wnd - sb_offset;
21312 					else
21313 						len = 0;
21314 				} else {
21315 					len = rack->r_ctl.rc_tlp_new_data;
21316 				}
21317 				rack->r_ctl.rc_tlp_new_data = 0;
21318 			}  else {
21319 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
21320 			}
21321 			if ((rack->r_ctl.crte == NULL) &&
21322 			    IN_FASTRECOVERY(tp->t_flags) &&
21323 			    (rack->full_size_rxt == 0) &&
21324 			    (rack->shape_rxt_to_pacing_min == 0) &&
21325 			    (len > segsiz)) {
21326 				/*
21327 				 * For prr=off, we need to send only 1 MSS
21328 				 * at a time. We do this because another sack could
21329 				 * be arriving that causes us to send retransmits and
21330 				 * we don't want to be on a long pace due to a larger send
21331 				 * that keeps us from sending out the retransmit.
21332 				 */
21333 				len = segsiz;
21334 			} else if (rack->shape_rxt_to_pacing_min &&
21335 				   rack->gp_ready) {
21336 				/* We use pacing min as shaping len req */
21337 				uint32_t maxlen;
21338 
21339 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
21340 				if (len > maxlen)
21341 					len = maxlen;
21342 			}/* The else is full_size_rxt is on so send it all */
21343 		} else {
21344 			uint32_t outstanding;
21345 			/*
21346 			 * We are inside of a Fast recovery episode, this
21347 			 * is caused by a SACK or 3 dup acks. At this point
21348 			 * we have sent all the retransmissions and we rely
21349 			 * on PRR to dictate what we will send in the form of
21350 			 * new data.
21351 			 */
21352 
21353 			outstanding = tp->snd_max - tp->snd_una;
21354 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
21355 				if (tp->snd_wnd > outstanding) {
21356 					len = tp->snd_wnd - outstanding;
21357 					/* Check to see if we have the data */
21358 					if ((sb_offset + len) > avail) {
21359 						/* It does not all fit */
21360 						if (avail > sb_offset)
21361 							len = avail - sb_offset;
21362 						else
21363 							len = 0;
21364 					}
21365 				} else {
21366 					len = 0;
21367 				}
21368 			} else if (avail > sb_offset) {
21369 				len = avail - sb_offset;
21370 			} else {
21371 				len = 0;
21372 			}
21373 			if (len > 0) {
21374 				if (len > rack->r_ctl.rc_prr_sndcnt) {
21375 					len = rack->r_ctl.rc_prr_sndcnt;
21376 				}
21377 				if (len > 0) {
21378 					sub_from_prr = 1;
21379 				}
21380 			}
21381 			if (len > segsiz) {
21382 				/*
21383 				 * We should never send more than a MSS when
21384 				 * retransmitting or sending new data in prr
21385 				 * mode unless the override flag is on. Most
21386 				 * likely the PRR algorithm is not going to
21387 				 * let us send a lot as well :-)
21388 				 */
21389 				if (rack->r_ctl.rc_prr_sendalot == 0) {
21390 					len = segsiz;
21391 				}
21392 			} else if (len < segsiz) {
21393 				/*
21394 				 * Do we send any? The idea here is if the
21395 				 * send empty's the socket buffer we want to
21396 				 * do it. However if not then lets just wait
21397 				 * for our prr_sndcnt to get bigger.
21398 				 */
21399 				long leftinsb;
21400 
21401 				leftinsb = sbavail(sb) - sb_offset;
21402 				if (leftinsb > len) {
21403 					/* This send does not empty the sb */
21404 					len = 0;
21405 				}
21406 			}
21407 		}
21408 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
21409 		/*
21410 		 * If you have not established
21411 		 * and are not doing FAST OPEN
21412 		 * no data please.
21413 		 */
21414 		if ((sack_rxmit == 0) &&
21415 		    !(tp->t_flags & TF_FASTOPEN)) {
21416 			len = 0;
21417 			sb_offset = 0;
21418 		}
21419 	}
21420 	if (prefetch_so_done == 0) {
21421 		kern_prefetch(so, &prefetch_so_done);
21422 		prefetch_so_done = 1;
21423 	}
21424 	orig_len = len;
21425 	if ((rack->rc_policer_detected == 1) &&
21426 	    (rack->r_ctl.policer_bucket_size > segsiz) &&
21427 	    (rack->r_ctl.policer_bw > 0) &&
21428 	    (len > 0)) {
21429 		/*
21430 		 * Ok we believe we have a policer watching
21431 		 * what we send, can we send len? If not can
21432 		 * we tune it down to a smaller value?
21433 		 */
21434 		uint32_t plen, buck_needs;
21435 
21436 		plen = rack_policer_check_send(rack, len, segsiz, &buck_needs);
21437 		if (plen == 0) {
21438 			/*
21439 			 * We are not allowed to send. How long
21440 			 * do we need to pace for i.e. how long
21441 			 * before len is available to send?
21442 			 */
21443 			uint64_t lentime;
21444 
21445 			lentime = buck_needs;
21446 			lentime *= HPTS_USEC_IN_SEC;
21447 			lentime /= rack->r_ctl.policer_bw;
21448 			slot = (uint32_t)lentime;
21449 			tot_len_this_send = 0;
21450 			SOCKBUF_UNLOCK(sb);
21451 			if (rack_verbose_logging > 0)
21452 				policer_detection_log(rack, len, slot, buck_needs, 0, 12);
21453 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
21454 			rack_log_type_just_return(rack, cts, 0, slot, hpts_calling, 0, cwnd_to_use);
21455 			goto just_return_clean;
21456 		}
21457 		if (plen < len) {
21458 			sendalot = 0;
21459 			len = plen;
21460 		}
21461 	}
21462 	/*
21463 	 * Lop off SYN bit if it has already been sent.  However, if this is
21464 	 * SYN-SENT state and if segment contains data and if we don't know
21465 	 * that foreign host supports TAO, suppress sending segment.
21466 	 */
21467 	if ((flags & TH_SYN) &&
21468 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
21469 	    ((sack_rxmit == 0) &&
21470 	     (tp->t_rxtshift == 0))) {
21471 		/*
21472 		 * When sending additional segments following a TFO SYN|ACK,
21473 		 * do not include the SYN bit.
21474 		 */
21475 		if ((tp->t_flags & TF_FASTOPEN) &&
21476 		    (tp->t_state == TCPS_SYN_RECEIVED))
21477 			flags &= ~TH_SYN;
21478 	}
21479 	/*
21480 	 * Be careful not to send data and/or FIN on SYN segments. This
21481 	 * measure is needed to prevent interoperability problems with not
21482 	 * fully conformant TCP implementations.
21483 	 */
21484 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
21485 		len = 0;
21486 		flags &= ~TH_FIN;
21487 	}
21488 	/*
21489 	 * On TFO sockets, ensure no data is sent in the following cases:
21490 	 *
21491 	 *  - When retransmitting SYN|ACK on a passively-created socket
21492 	 *
21493 	 *  - When retransmitting SYN on an actively created socket
21494 	 *
21495 	 *  - When sending a zero-length cookie (cookie request) on an
21496 	 *    actively created socket
21497 	 *
21498 	 *  - When the socket is in the CLOSED state (RST is being sent)
21499 	 */
21500 	if ((tp->t_flags & TF_FASTOPEN) &&
21501 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
21502 	     ((tp->t_state == TCPS_SYN_SENT) &&
21503 	      (tp->t_tfo_client_cookie_len == 0)) ||
21504 	     (flags & TH_RST))) {
21505 		sack_rxmit = 0;
21506 		len = 0;
21507 	}
21508 	/* Without fast-open there should never be data sent on a SYN */
21509 	if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
21510 		len = 0;
21511 	}
21512 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
21513 		/* We only send 1 MSS if we have a DSACK block */
21514 		add_flag |= RACK_SENT_W_DSACK;
21515 		len = segsiz;
21516 	}
21517 	if (len <= 0) {
21518 		/*
21519 		 * We have nothing to send, or the window shrank, or
21520 		 * is closed, do we need to go into persists?
21521 		 */
21522 		len = 0;
21523 		if ((tp->snd_wnd == 0) &&
21524 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
21525 		    (tp->snd_una == tp->snd_max) &&
21526 		    (sb_offset < (int)sbavail(sb))) {
21527 			rack_enter_persist(tp, rack, cts, tp->snd_una);
21528 		}
21529 	} else if ((rsm == NULL) &&
21530 		   (doing_tlp == 0) &&
21531 		   (len < pace_max_seg)) {
21532 		/*
21533 		 * We are not sending a maximum sized segment for
21534 		 * some reason. Should we not send anything (think
21535 		 * sws or persists)?
21536 		 */
21537 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
21538 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
21539 		    (len < minseg) &&
21540 		    (len < (int)(sbavail(sb) - sb_offset))) {
21541 			/*
21542 			 * Here the rwnd is less than
21543 			 * the minimum pacing size, this is not a retransmit,
21544 			 * we are established and
21545 			 * the send is not the last in the socket buffer
21546 			 * we send nothing, and we may enter persists
21547 			 * if nothing is outstanding.
21548 			 */
21549 			len = 0;
21550 			if (tp->snd_max == tp->snd_una) {
21551 				/*
21552 				 * Nothing out we can
21553 				 * go into persists.
21554 				 */
21555 				rack_enter_persist(tp, rack, cts, tp->snd_una);
21556 			}
21557 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
21558 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
21559 			   (len < (int)(sbavail(sb) - sb_offset)) &&
21560 			   (len < minseg)) {
21561 			/*
21562 			 * Here we are not retransmitting, and
21563 			 * the cwnd is not so small that we could
21564 			 * not send at least a min size (rxt timer
21565 			 * not having gone off), We have 2 segments or
21566 			 * more already in flight, its not the tail end
21567 			 * of the socket buffer  and the cwnd is blocking
21568 			 * us from sending out a minimum pacing segment size.
21569 			 * Lets not send anything.
21570 			 */
21571 			len = 0;
21572 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
21573 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
21574 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
21575 			   (len < (int)(sbavail(sb) - sb_offset)) &&
21576 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
21577 			/*
21578 			 * Here we have a send window but we have
21579 			 * filled it up and we can't send another pacing segment.
21580 			 * We also have in flight more than 2 segments
21581 			 * and we are not completing the sb i.e. we allow
21582 			 * the last bytes of the sb to go out even if
21583 			 * its not a full pacing segment.
21584 			 */
21585 			len = 0;
21586 		} else if ((rack->r_ctl.crte != NULL) &&
21587 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
21588 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
21589 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
21590 			   (len < (int)(sbavail(sb) - sb_offset))) {
21591 			/*
21592 			 * Here we are doing hardware pacing, this is not a TLP,
21593 			 * we are not sending a pace max segment size, there is rwnd
21594 			 * room to send at least N pace_max_seg, the cwnd is greater
21595 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
21596 			 * more segments in flight and its not the tail of the socket buffer.
21597 			 *
21598 			 * We don't want to send instead we need to get more ack's in to
21599 			 * allow us to send a full pacing segment. Normally, if we are pacing
21600 			 * about the right speed, we should have finished our pacing
21601 			 * send as most of the acks have come back if we are at the
21602 			 * right rate. This is a bit fuzzy since return path delay
21603 			 * can delay the acks, which is why we want to make sure we
21604 			 * have cwnd space to have a bit more than a max pace segments in flight.
21605 			 *
21606 			 * If we have not gotten our acks back we are pacing at too high a
21607 			 * rate delaying will not hurt and will bring our GP estimate down by
21608 			 * injecting the delay. If we don't do this we will send
21609 			 * 2 MSS out in response to the acks being clocked in which
21610 			 * defeats the point of hw-pacing (i.e. to help us get
21611 			 * larger TSO's out).
21612 			 */
21613 			len = 0;
21614 		}
21615 
21616 	}
21617 	/* len will be >= 0 after this point. */
21618 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21619 	rack_sndbuf_autoscale(rack);
21620 	/*
21621 	 * Decide if we can use TCP Segmentation Offloading (if supported by
21622 	 * hardware).
21623 	 *
21624 	 * TSO may only be used if we are in a pure bulk sending state.  The
21625 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
21626 	 * options prevent using TSO.  With TSO the TCP header is the same
21627 	 * (except for the sequence number) for all generated packets.  This
21628 	 * makes it impossible to transmit any options which vary per
21629 	 * generated segment or packet.
21630 	 *
21631 	 * IPv4 handling has a clear separation of ip options and ip header
21632 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
21633 	 * the right thing below to provide length of just ip options and thus
21634 	 * checking for ipoptlen is enough to decide if ip options are present.
21635 	 */
21636 	ipoptlen = 0;
21637 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21638 	/*
21639 	 * Pre-calculate here as we save another lookup into the darknesses
21640 	 * of IPsec that way and can actually decide if TSO is ok.
21641 	 */
21642 #ifdef INET6
21643 	if (isipv6 && IPSEC_ENABLED(ipv6))
21644 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
21645 #ifdef INET
21646 	else
21647 #endif
21648 #endif				/* INET6 */
21649 #ifdef INET
21650 		if (IPSEC_ENABLED(ipv4))
21651 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
21652 #endif				/* INET */
21653 #endif
21654 
21655 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21656 	ipoptlen += ipsec_optlen;
21657 #endif
21658 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
21659 	    (tp->t_port == 0) &&
21660 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
21661 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
21662 	    ipoptlen == 0)
21663 		tso = 1;
21664 	{
21665 		uint32_t outstanding __unused;
21666 
21667 		outstanding = tp->snd_max - tp->snd_una;
21668 		if (tp->t_flags & TF_SENTFIN) {
21669 			/*
21670 			 * If we sent a fin, snd_max is 1 higher than
21671 			 * snd_una
21672 			 */
21673 			outstanding--;
21674 		}
21675 		if (sack_rxmit) {
21676 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
21677 				flags &= ~TH_FIN;
21678 		}
21679 	}
21680 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
21681 		      (long)TCP_MAXWIN << tp->rcv_scale);
21682 
21683 	/*
21684 	 * Sender silly window avoidance.   We transmit under the following
21685 	 * conditions when len is non-zero:
21686 	 *
21687 	 * - We have a full segment (or more with TSO) - This is the last
21688 	 * buffer in a write()/send() and we are either idle or running
21689 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
21690 	 * then 1/2 the maximum send window's worth of data (receiver may be
21691 	 * limited the window size) - we need to retransmit
21692 	 */
21693 	if (len) {
21694 		if (len >= segsiz) {
21695 			goto send;
21696 		}
21697 		/*
21698 		 * NOTE! on localhost connections an 'ack' from the remote
21699 		 * end may occur synchronously with the output and cause us
21700 		 * to flush a buffer queued with moretocome.  XXX
21701 		 *
21702 		 */
21703 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
21704 		    (idle || (tp->t_flags & TF_NODELAY)) &&
21705 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
21706 		    (tp->t_flags & TF_NOPUSH) == 0) {
21707 			pass = 2;
21708 			goto send;
21709 		}
21710 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
21711 			pass = 22;
21712 			goto send;
21713 		}
21714 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
21715 			pass = 4;
21716 			goto send;
21717 		}
21718 		if (sack_rxmit) {
21719 			pass = 6;
21720 			goto send;
21721 		}
21722 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
21723 		    (ctf_outstanding(tp) < (segsiz * 2))) {
21724 			/*
21725 			 * We have less than two MSS outstanding (delayed ack)
21726 			 * and our rwnd will not let us send a full sized
21727 			 * MSS. Lets go ahead and let this small segment
21728 			 * out because we want to try to have at least two
21729 			 * packets inflight to not be caught by delayed ack.
21730 			 */
21731 			pass = 12;
21732 			goto send;
21733 		}
21734 	}
21735 	/*
21736 	 * Sending of standalone window updates.
21737 	 *
21738 	 * Window updates are important when we close our window due to a
21739 	 * full socket buffer and are opening it again after the application
21740 	 * reads data from it.  Once the window has opened again and the
21741 	 * remote end starts to send again the ACK clock takes over and
21742 	 * provides the most current window information.
21743 	 *
21744 	 * We must avoid the silly window syndrome whereas every read from
21745 	 * the receive buffer, no matter how small, causes a window update
21746 	 * to be sent.  We also should avoid sending a flurry of window
21747 	 * updates when the socket buffer had queued a lot of data and the
21748 	 * application is doing small reads.
21749 	 *
21750 	 * Prevent a flurry of pointless window updates by only sending an
21751 	 * update when we can increase the advertized window by more than
21752 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
21753 	 * full or is very small be more aggressive and send an update
21754 	 * whenever we can increase by two mss sized segments. In all other
21755 	 * situations the ACK's to new incoming data will carry further
21756 	 * window increases.
21757 	 *
21758 	 * Don't send an independent window update if a delayed ACK is
21759 	 * pending (it will get piggy-backed on it) or the remote side
21760 	 * already has done a half-close and won't send more data.  Skip
21761 	 * this if the connection is in T/TCP half-open state.
21762 	 */
21763 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
21764 	    !(tp->t_flags & TF_DELACK) &&
21765 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
21766 		/*
21767 		 * "adv" is the amount we could increase the window, taking
21768 		 * into account that we are limited by TCP_MAXWIN <<
21769 		 * tp->rcv_scale.
21770 		 */
21771 		int32_t adv;
21772 		int oldwin;
21773 
21774 		adv = recwin;
21775 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
21776 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
21777 			if (adv > oldwin)
21778 				adv -= oldwin;
21779 			else {
21780 				/* We can't increase the window */
21781 				adv = 0;
21782 			}
21783 		} else
21784 			oldwin = 0;
21785 
21786 		/*
21787 		 * If the new window size ends up being the same as or less
21788 		 * than the old size when it is scaled, then don't force
21789 		 * a window update.
21790 		 */
21791 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
21792 			goto dontupdate;
21793 
21794 		if (adv >= (int32_t)(2 * segsiz) &&
21795 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
21796 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
21797 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
21798 			pass = 7;
21799 			goto send;
21800 		}
21801 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
21802 			pass = 23;
21803 			goto send;
21804 		}
21805 	}
21806 dontupdate:
21807 
21808 	/*
21809 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
21810 	 * is also a catch-all for the retransmit timer timeout case.
21811 	 */
21812 	if (tp->t_flags & TF_ACKNOW) {
21813 		pass = 8;
21814 		goto send;
21815 	}
21816 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
21817 		pass = 9;
21818 		goto send;
21819 	}
21820 	/*
21821 	 * If our state indicates that FIN should be sent and we have not
21822 	 * yet done so, then we need to send.
21823 	 */
21824 	if ((flags & TH_FIN) &&
21825 	    (tp->snd_max == tp->snd_una)) {
21826 		pass = 11;
21827 		goto send;
21828 	}
21829 	/*
21830 	 * No reason to send a segment, just return.
21831 	 */
21832 just_return:
21833 	SOCKBUF_UNLOCK(sb);
21834 just_return_nolock:
21835 	{
21836 		int app_limited = CTF_JR_SENT_DATA;
21837 
21838 		if ((tp->t_flags & TF_FASTOPEN) == 0 &&
21839 		    (flags & TH_FIN) &&
21840 		    (len == 0) &&
21841 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21842 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
21843 			/*
21844 			 * Ok less than or right at a MSS is
21845 			 * outstanding. The original FreeBSD stack would
21846 			 * have sent a FIN, which can speed things up for
21847 			 * a transactional application doing a MSG_WAITALL.
21848 			 * To speed things up since we do *not* send a FIN
21849 			 * if data is outstanding, we send a "challenge ack".
21850 			 * The idea behind that is instead of having to have
21851 			 * the peer wait for the delayed-ack timer to run off
21852 			 * we send an ack that makes the peer send us an ack.
21853 			 */
21854 			rack_send_ack_challange(rack);
21855 		}
21856 		if (tot_len_this_send > 0) {
21857 			rack->r_ctl.fsb.recwin = recwin;
21858 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz, __LINE__);
21859 			if ((error == 0) &&
21860 			    (rack->rc_policer_detected == 0)  &&
21861 			    rack_use_rfo &&
21862 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
21863 			    (ipoptlen == 0) &&
21864 			    (tp->rcv_numsacks == 0) &&
21865 			    rack->r_fsb_inited &&
21866 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
21867 			    ((IN_RECOVERY(tp->t_flags)) == 0) &&
21868 			    (rack->r_must_retran == 0) &&
21869 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
21870 			    (len > 0) && (orig_len > 0) &&
21871 			    (orig_len > len) &&
21872 			    ((orig_len - len) >= segsiz) &&
21873 			    ((optlen == 0) ||
21874 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
21875 				/* We can send at least one more MSS using our fsb */
21876 				rack_setup_fast_output(tp, rack, sb, len, orig_len,
21877 						       segsiz, pace_max_seg, hw_tls, flags);
21878 			} else
21879 				rack->r_fast_output = 0;
21880 			rack_log_fsb(rack, tp, so, flags,
21881 				     ipoptlen, orig_len, len, 0,
21882 				     1, optlen, __LINE__, 1);
21883 			/* Assure when we leave that snd_nxt will point to top */
21884 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
21885 				tp->snd_nxt = tp->snd_max;
21886 		} else {
21887 			int end_window = 0;
21888 			uint32_t seq = tp->gput_ack;
21889 
21890 			rsm = tqhash_max(rack->r_ctl.tqh);
21891 			if (rsm) {
21892 				/*
21893 				 * Mark the last sent that we just-returned (hinting
21894 				 * that delayed ack may play a role in any rtt measurement).
21895 				 */
21896 				rsm->r_just_ret = 1;
21897 			}
21898 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
21899 			rack->r_ctl.rc_agg_delayed = 0;
21900 			rack->r_early = 0;
21901 			rack->r_late = 0;
21902 			rack->r_ctl.rc_agg_early = 0;
21903 			if ((ctf_outstanding(tp) +
21904 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
21905 				 minseg)) >= tp->snd_wnd) {
21906 				/* We are limited by the rwnd */
21907 				app_limited = CTF_JR_RWND_LIMITED;
21908 				if (IN_FASTRECOVERY(tp->t_flags))
21909 					rack->r_ctl.rc_prr_sndcnt = 0;
21910 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
21911 				/* We are limited by whats available -- app limited */
21912 				app_limited = CTF_JR_APP_LIMITED;
21913 				if (IN_FASTRECOVERY(tp->t_flags))
21914 					rack->r_ctl.rc_prr_sndcnt = 0;
21915 			} else if ((idle == 0) &&
21916 				   ((tp->t_flags & TF_NODELAY) == 0) &&
21917 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
21918 				   (len < segsiz)) {
21919 				/*
21920 				 * No delay is not on and the
21921 				 * user is sending less than 1MSS. This
21922 				 * brings out SWS avoidance so we
21923 				 * don't send. Another app-limited case.
21924 				 */
21925 				app_limited = CTF_JR_APP_LIMITED;
21926 			} else if (tp->t_flags & TF_NOPUSH) {
21927 				/*
21928 				 * The user has requested no push of
21929 				 * the last segment and we are
21930 				 * at the last segment. Another app
21931 				 * limited case.
21932 				 */
21933 				app_limited = CTF_JR_APP_LIMITED;
21934 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
21935 				/* Its the cwnd */
21936 				app_limited = CTF_JR_CWND_LIMITED;
21937 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
21938 				   (rack->rack_no_prr == 0) &&
21939 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
21940 				app_limited = CTF_JR_PRR;
21941 			} else {
21942 				/* Now why here are we not sending? */
21943 #ifdef NOW
21944 #ifdef INVARIANTS
21945 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
21946 #endif
21947 #endif
21948 				app_limited = CTF_JR_ASSESSING;
21949 			}
21950 			/*
21951 			 * App limited in some fashion, for our pacing GP
21952 			 * measurements we don't want any gap (even cwnd).
21953 			 * Close  down the measurement window.
21954 			 */
21955 			if (rack_cwnd_block_ends_measure &&
21956 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
21957 			     (app_limited == CTF_JR_PRR))) {
21958 				/*
21959 				 * The reason we are not sending is
21960 				 * the cwnd (or prr). We have been configured
21961 				 * to end the measurement window in
21962 				 * this case.
21963 				 */
21964 				end_window = 1;
21965 			} else if (rack_rwnd_block_ends_measure &&
21966 				   (app_limited == CTF_JR_RWND_LIMITED)) {
21967 				/*
21968 				 * We are rwnd limited and have been
21969 				 * configured to end the measurement
21970 				 * window in this case.
21971 				 */
21972 				end_window = 1;
21973 			} else if (app_limited == CTF_JR_APP_LIMITED) {
21974 				/*
21975 				 * A true application limited period, we have
21976 				 * ran out of data.
21977 				 */
21978 				end_window = 1;
21979 			} else if (app_limited == CTF_JR_ASSESSING) {
21980 				/*
21981 				 * In the assessing case we hit the end of
21982 				 * the if/else and had no known reason
21983 				 * This will panic us under invariants..
21984 				 *
21985 				 * If we get this out in logs we need to
21986 				 * investagate which reason we missed.
21987 				 */
21988 				end_window = 1;
21989 			}
21990 			if (end_window) {
21991 				uint8_t log = 0;
21992 
21993 				/* Adjust the Gput measurement */
21994 				if ((tp->t_flags & TF_GPUTINPROG) &&
21995 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
21996 					tp->gput_ack = tp->snd_max;
21997 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
21998 						/*
21999 						 * There is not enough to measure.
22000 						 */
22001 						tp->t_flags &= ~TF_GPUTINPROG;
22002 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
22003 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
22004 									   tp->gput_seq,
22005 									   0, 0, 18, __LINE__, NULL, 0);
22006 					} else
22007 						log = 1;
22008 				}
22009 				/* Mark the last packet has app limited */
22010 				rsm = tqhash_max(rack->r_ctl.tqh);
22011 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
22012 					if (rack->r_ctl.rc_app_limited_cnt == 0)
22013 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
22014 					else {
22015 						/*
22016 						 * Go out to the end app limited and mark
22017 						 * this new one as next and move the end_appl up
22018 						 * to this guy.
22019 						 */
22020 						if (rack->r_ctl.rc_end_appl)
22021 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
22022 						rack->r_ctl.rc_end_appl = rsm;
22023 					}
22024 					rsm->r_flags |= RACK_APP_LIMITED;
22025 					rack->r_ctl.rc_app_limited_cnt++;
22026 				}
22027 				if (log)
22028 					rack_log_pacing_delay_calc(rack,
22029 								   rack->r_ctl.rc_app_limited_cnt, seq,
22030 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
22031 			}
22032 		}
22033 		/* Check if we need to go into persists or not */
22034 		if ((tp->snd_max == tp->snd_una) &&
22035 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22036 		    sbavail(sb) &&
22037 		    (sbavail(sb) > tp->snd_wnd) &&
22038 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
22039 			/* Yes lets make sure to move to persist before timer-start */
22040 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
22041 		}
22042 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
22043 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
22044 	}
22045 just_return_clean:
22046 #ifdef NETFLIX_SHARED_CWND
22047 	if ((sbavail(sb) == 0) &&
22048 	    rack->r_ctl.rc_scw) {
22049 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
22050 		rack->rack_scwnd_is_idle = 1;
22051 	}
22052 #endif
22053 #ifdef TCP_ACCOUNTING
22054 	if (tot_len_this_send > 0) {
22055 		crtsc = get_cyclecount();
22056 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22057 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
22058 		}
22059 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22060 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
22061 		}
22062 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22063 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
22064 		}
22065 	} else {
22066 		crtsc = get_cyclecount();
22067 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22068 			tp->tcp_cnt_counters[SND_LIMITED]++;
22069 		}
22070 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22071 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
22072 		}
22073 	}
22074 	sched_unpin();
22075 #endif
22076 	return (0);
22077 
22078 send:
22079 	if ((rack->r_ctl.crte != NULL) &&
22080 	    (rsm == NULL) &&
22081 	    ((rack->rc_hw_nobuf == 1) ||
22082 	     (rack_hw_check_queue && (check_done == 0)))) {
22083 		/*
22084 		 * We only want to do this once with the hw_check_queue,
22085 		 * for the enobuf case we would only do it once if
22086 		 * we come around to again, the flag will be clear.
22087 		 */
22088 		check_done = 1;
22089 		slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
22090 		if (slot) {
22091 			rack->r_ctl.rc_agg_delayed = 0;
22092 			rack->r_ctl.rc_agg_early = 0;
22093 			rack->r_early = 0;
22094 			rack->r_late = 0;
22095 			SOCKBUF_UNLOCK(&so->so_snd);
22096 			goto skip_all_send;
22097 		}
22098 	}
22099 	if (rsm || sack_rxmit)
22100 		counter_u64_add(rack_nfto_resend, 1);
22101 	else
22102 		counter_u64_add(rack_non_fto_send, 1);
22103 	if ((flags & TH_FIN) &&
22104 	    sbavail(sb)) {
22105 		/*
22106 		 * We do not transmit a FIN
22107 		 * with data outstanding. We
22108 		 * need to make it so all data
22109 		 * is acked first.
22110 		 */
22111 		flags &= ~TH_FIN;
22112 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22113 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
22114 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
22115 			/*
22116 			 * Ok less than or right at a MSS is
22117 			 * outstanding. The original FreeBSD stack would
22118 			 * have sent a FIN, which can speed things up for
22119 			 * a transactional application doing a MSG_WAITALL.
22120 			 * To speed things up since we do *not* send a FIN
22121 			 * if data is outstanding, we send a "challenge ack".
22122 			 * The idea behind that is instead of having to have
22123 			 * the peer wait for the delayed-ack timer to run off
22124 			 * we send an ack that makes the peer send us an ack.
22125 			 */
22126 			rack_send_ack_challange(rack);
22127 		}
22128 	}
22129 	/* Enforce stack imposed max seg size if we have one */
22130 	if (pace_max_seg &&
22131 	    (len > pace_max_seg)) {
22132 		mark = 1;
22133 		len = pace_max_seg;
22134 	}
22135 	if ((rsm == NULL) &&
22136 	    (rack->pcm_in_progress == 0) &&
22137 	    (rack->r_ctl.pcm_max_seg > 0) &&
22138 	    (len >= rack->r_ctl.pcm_max_seg)) {
22139 		/* It is large enough for a measurement */
22140 		add_flag |= RACK_IS_PCM;
22141 		rack_log_pcm(rack, 5, len, rack->r_ctl.pcm_max_seg,  add_flag);
22142 	} else if (rack_verbose_logging) {
22143 		rack_log_pcm(rack, 6, len, rack->r_ctl.pcm_max_seg,  add_flag);
22144 	}
22145 
22146 	SOCKBUF_LOCK_ASSERT(sb);
22147 	if (len > 0) {
22148 		if (len >= segsiz)
22149 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
22150 		else
22151 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
22152 	}
22153 	/*
22154 	 * Before ESTABLISHED, force sending of initial options unless TCP
22155 	 * set not to do any options. NOTE: we assume that the IP/TCP header
22156 	 * plus TCP options always fit in a single mbuf, leaving room for a
22157 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
22158 	 * + optlen <= MCLBYTES
22159 	 */
22160 	optlen = 0;
22161 #ifdef INET6
22162 	if (isipv6)
22163 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
22164 	else
22165 #endif
22166 		hdrlen = sizeof(struct tcpiphdr);
22167 
22168 	/*
22169 	 * Ok what seq are we sending from. If we have
22170 	 * no rsm to use, then we look at various bits,
22171 	 * if we are putting out a SYN it will be ISS.
22172 	 * If we are retransmitting a FIN it will
22173 	 * be snd_max-1 else its snd_max.
22174 	 */
22175 	if (rsm == NULL) {
22176 		if (flags & TH_SYN)
22177 			rack_seq = tp->iss;
22178 		else if ((flags & TH_FIN) &&
22179 			 (tp->t_flags & TF_SENTFIN))
22180 			rack_seq = tp->snd_max - 1;
22181 		else
22182 			rack_seq = tp->snd_max;
22183 	} else {
22184 		rack_seq = rsm->r_start;
22185 	}
22186 	/*
22187 	 * Compute options for segment. We only have to care about SYN and
22188 	 * established connection segments.  Options for SYN-ACK segments
22189 	 * are handled in TCP syncache.
22190 	 */
22191 	to.to_flags = 0;
22192 	if ((tp->t_flags & TF_NOOPT) == 0) {
22193 		/* Maximum segment size. */
22194 		if (flags & TH_SYN) {
22195 			to.to_mss = tcp_mssopt(&inp->inp_inc);
22196 			if (tp->t_port)
22197 				to.to_mss -= V_tcp_udp_tunneling_overhead;
22198 			to.to_flags |= TOF_MSS;
22199 
22200 			/*
22201 			 * On SYN or SYN|ACK transmits on TFO connections,
22202 			 * only include the TFO option if it is not a
22203 			 * retransmit, as the presence of the TFO option may
22204 			 * have caused the original SYN or SYN|ACK to have
22205 			 * been dropped by a middlebox.
22206 			 */
22207 			if ((tp->t_flags & TF_FASTOPEN) &&
22208 			    (tp->t_rxtshift == 0)) {
22209 				if (tp->t_state == TCPS_SYN_RECEIVED) {
22210 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
22211 					to.to_tfo_cookie =
22212 						(u_int8_t *)&tp->t_tfo_cookie.server;
22213 					to.to_flags |= TOF_FASTOPEN;
22214 					wanted_cookie = 1;
22215 				} else if (tp->t_state == TCPS_SYN_SENT) {
22216 					to.to_tfo_len =
22217 						tp->t_tfo_client_cookie_len;
22218 					to.to_tfo_cookie =
22219 						tp->t_tfo_cookie.client;
22220 					to.to_flags |= TOF_FASTOPEN;
22221 					wanted_cookie = 1;
22222 					/*
22223 					 * If we wind up having more data to
22224 					 * send with the SYN than can fit in
22225 					 * one segment, don't send any more
22226 					 * until the SYN|ACK comes back from
22227 					 * the other end.
22228 					 */
22229 					sendalot = 0;
22230 				}
22231 			}
22232 		}
22233 		/* Window scaling. */
22234 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
22235 			to.to_wscale = tp->request_r_scale;
22236 			to.to_flags |= TOF_SCALE;
22237 		}
22238 		/* Timestamps. */
22239 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
22240 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
22241 			uint32_t ts_to_use;
22242 
22243 			if ((rack->r_rcvpath_rtt_up == 1) &&
22244 			    (ms_cts == rack->r_ctl.last_rcv_tstmp_for_rtt)) {
22245 				/*
22246 				 * When we are doing a rcv_rtt probe all
22247 				 * other timestamps use the next msec. This
22248 				 * is safe since our previous ack is in the
22249 				 * air and we will just have a few more
22250 				 * on the next ms. This assures that only
22251 				 * the one ack has the ms_cts that was on
22252 				 * our ack-probe.
22253 				 */
22254 				ts_to_use = ms_cts + 1;
22255 			} else {
22256 				ts_to_use = ms_cts;
22257 			}
22258 			to.to_tsval = ts_to_use + tp->ts_offset;
22259 			to.to_tsecr = tp->ts_recent;
22260 			to.to_flags |= TOF_TS;
22261 			if ((len == 0) &&
22262 			    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
22263 			    ((ms_cts - rack->r_ctl.last_rcv_tstmp_for_rtt) > RCV_PATH_RTT_MS) &&
22264 			    (tp->snd_una == tp->snd_max) &&
22265 			    (flags & TH_ACK) &&
22266 			    (sbavail(sb) == 0) &&
22267 			    (rack->r_ctl.current_round != 0) &&
22268 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22269 			    (rack->r_rcvpath_rtt_up == 0)) {
22270 				rack->r_ctl.last_rcv_tstmp_for_rtt = ms_cts;
22271 				rack->r_ctl.last_time_of_arm_rcv = cts;
22272 				rack->r_rcvpath_rtt_up = 1;
22273 				/* Subtract 1 from seq to force a response */
22274 				rack_seq--;
22275 			}
22276 		}
22277 		/* Set receive buffer autosizing timestamp. */
22278 		if (tp->rfbuf_ts == 0 &&
22279 		    (so->so_rcv.sb_flags & SB_AUTOSIZE)) {
22280 			tp->rfbuf_ts = ms_cts;
22281 		}
22282 		/* Selective ACK's. */
22283 		if (tp->t_flags & TF_SACK_PERMIT) {
22284 			if (flags & TH_SYN)
22285 				to.to_flags |= TOF_SACKPERM;
22286 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22287 				 tp->rcv_numsacks > 0) {
22288 				to.to_flags |= TOF_SACK;
22289 				to.to_nsacks = tp->rcv_numsacks;
22290 				to.to_sacks = (u_char *)tp->sackblks;
22291 			}
22292 		}
22293 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
22294 		/* TCP-MD5 (RFC2385). */
22295 		if (tp->t_flags & TF_SIGNATURE)
22296 			to.to_flags |= TOF_SIGNATURE;
22297 #endif
22298 
22299 		/* Processing the options. */
22300 		hdrlen += optlen = tcp_addoptions(&to, opt);
22301 		/*
22302 		 * If we wanted a TFO option to be added, but it was unable
22303 		 * to fit, ensure no data is sent.
22304 		 */
22305 		if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
22306 		    !(to.to_flags & TOF_FASTOPEN))
22307 			len = 0;
22308 	}
22309 	if (tp->t_port) {
22310 		if (V_tcp_udp_tunneling_port == 0) {
22311 			/* The port was removed?? */
22312 			SOCKBUF_UNLOCK(&so->so_snd);
22313 #ifdef TCP_ACCOUNTING
22314 			crtsc = get_cyclecount();
22315 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22316 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22317 			}
22318 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22319 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22320 			}
22321 			sched_unpin();
22322 #endif
22323 			return (EHOSTUNREACH);
22324 		}
22325 		hdrlen += sizeof(struct udphdr);
22326 	}
22327 #ifdef INET6
22328 	if (isipv6)
22329 		ipoptlen = ip6_optlen(inp);
22330 	else
22331 #endif
22332 		if (inp->inp_options)
22333 			ipoptlen = inp->inp_options->m_len -
22334 				offsetof(struct ipoption, ipopt_list);
22335 		else
22336 			ipoptlen = 0;
22337 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22338 	ipoptlen += ipsec_optlen;
22339 #endif
22340 
22341 	/*
22342 	 * Adjust data length if insertion of options will bump the packet
22343 	 * length beyond the t_maxseg length. Clear the FIN bit because we
22344 	 * cut off the tail of the segment.
22345 	 */
22346 	if (len + optlen + ipoptlen > tp->t_maxseg) {
22347 		if (tso) {
22348 			uint32_t if_hw_tsomax;
22349 			uint32_t moff;
22350 			int32_t max_len;
22351 
22352 			/* extract TSO information */
22353 			if_hw_tsomax = tp->t_tsomax;
22354 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
22355 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
22356 			KASSERT(ipoptlen == 0,
22357 				("%s: TSO can't do IP options", __func__));
22358 
22359 			/*
22360 			 * Check if we should limit by maximum payload
22361 			 * length:
22362 			 */
22363 			if (if_hw_tsomax != 0) {
22364 				/* compute maximum TSO length */
22365 				max_len = (if_hw_tsomax - hdrlen -
22366 					   max_linkhdr);
22367 				if (max_len <= 0) {
22368 					len = 0;
22369 				} else if (len > max_len) {
22370 					sendalot = 1;
22371 					len = max_len;
22372 					mark = 2;
22373 				}
22374 			}
22375 			/*
22376 			 * Prevent the last segment from being fractional
22377 			 * unless the send sockbuf can be emptied:
22378 			 */
22379 			max_len = (tp->t_maxseg - optlen);
22380 			if ((sb_offset + len) < sbavail(sb)) {
22381 				moff = len % (u_int)max_len;
22382 				if (moff != 0) {
22383 					mark = 3;
22384 					len -= moff;
22385 				}
22386 			}
22387 			/*
22388 			 * In case there are too many small fragments don't
22389 			 * use TSO:
22390 			 */
22391 			if (len <= max_len) {
22392 				mark = 4;
22393 				tso = 0;
22394 			}
22395 			/*
22396 			 * Send the FIN in a separate segment after the bulk
22397 			 * sending is done. We don't trust the TSO
22398 			 * implementations to clear the FIN flag on all but
22399 			 * the last segment.
22400 			 */
22401 			if (tp->t_flags & TF_NEEDFIN) {
22402 				sendalot = 4;
22403 			}
22404 		} else {
22405 			mark = 5;
22406 			if (optlen + ipoptlen >= tp->t_maxseg) {
22407 				/*
22408 				 * Since we don't have enough space to put
22409 				 * the IP header chain and the TCP header in
22410 				 * one packet as required by RFC 7112, don't
22411 				 * send it. Also ensure that at least one
22412 				 * byte of the payload can be put into the
22413 				 * TCP segment.
22414 				 */
22415 				SOCKBUF_UNLOCK(&so->so_snd);
22416 				error = EMSGSIZE;
22417 				sack_rxmit = 0;
22418 				goto out;
22419 			}
22420 			len = tp->t_maxseg - optlen - ipoptlen;
22421 			sendalot = 5;
22422 		}
22423 	} else {
22424 		tso = 0;
22425 		mark = 6;
22426 	}
22427 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
22428 		("%s: len > IP_MAXPACKET", __func__));
22429 #ifdef DIAGNOSTIC
22430 #ifdef INET6
22431 	if (max_linkhdr + hdrlen > MCLBYTES)
22432 #else
22433 		if (max_linkhdr + hdrlen > MHLEN)
22434 #endif
22435 			panic("tcphdr too big");
22436 #endif
22437 
22438 	/*
22439 	 * This KASSERT is here to catch edge cases at a well defined place.
22440 	 * Before, those had triggered (random) panic conditions further
22441 	 * down.
22442 	 */
22443 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
22444 	if ((len == 0) &&
22445 	    (flags & TH_FIN) &&
22446 	    (sbused(sb))) {
22447 		/*
22448 		 * We have outstanding data, don't send a fin by itself!.
22449 		 *
22450 		 * Check to see if we need to send a challenge ack.
22451 		 */
22452 		if ((sbused(sb) == (tp->snd_max - tp->snd_una)) &&
22453 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
22454 			/*
22455 			 * Ok less than or right at a MSS is
22456 			 * outstanding. The original FreeBSD stack would
22457 			 * have sent a FIN, which can speed things up for
22458 			 * a transactional application doing a MSG_WAITALL.
22459 			 * To speed things up since we do *not* send a FIN
22460 			 * if data is outstanding, we send a "challenge ack".
22461 			 * The idea behind that is instead of having to have
22462 			 * the peer wait for the delayed-ack timer to run off
22463 			 * we send an ack that makes the peer send us an ack.
22464 			 */
22465 			rack_send_ack_challange(rack);
22466 		}
22467 		goto just_return;
22468 	}
22469 	/*
22470 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
22471 	 * and initialize the header from the template for sends on this
22472 	 * connection.
22473 	 */
22474 	hw_tls = tp->t_nic_ktls_xmit != 0;
22475 	if (len) {
22476 		uint32_t max_val;
22477 		uint32_t moff;
22478 
22479 		if (pace_max_seg)
22480 			max_val = pace_max_seg;
22481 		else
22482 			max_val = len;
22483 		/*
22484 		 * We allow a limit on sending with hptsi.
22485 		 */
22486 		if (len > max_val) {
22487 			mark = 7;
22488 			len = max_val;
22489 		}
22490 #ifdef INET6
22491 		if (MHLEN < hdrlen + max_linkhdr)
22492 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
22493 		else
22494 #endif
22495 			m = m_gethdr(M_NOWAIT, MT_DATA);
22496 
22497 		if (m == NULL) {
22498 			SOCKBUF_UNLOCK(sb);
22499 			error = ENOBUFS;
22500 			sack_rxmit = 0;
22501 			goto out;
22502 		}
22503 		m->m_data += max_linkhdr;
22504 		m->m_len = hdrlen;
22505 
22506 		/*
22507 		 * Start the m_copy functions from the closest mbuf to the
22508 		 * sb_offset in the socket buffer chain.
22509 		 */
22510 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
22511 		s_mb = mb;
22512 		s_moff = moff;
22513 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
22514 			m_copydata(mb, moff, (int)len,
22515 				   mtod(m, caddr_t)+hdrlen);
22516 			/*
22517 			 * If we are not retransmitting advance the
22518 			 * sndptr to help remember the next place in
22519 			 * the sb.
22520 			 */
22521 			if (rsm == NULL)
22522 				sbsndptr_adv(sb, mb, len);
22523 			m->m_len += len;
22524 		} else {
22525 			struct sockbuf *msb;
22526 
22527 			/*
22528 			 * If we are not retransmitting pass in msb so
22529 			 * the socket buffer can be advanced. Otherwise
22530 			 * set it to NULL if its a retransmission since
22531 			 * we don't want to change the sb remembered
22532 			 * location.
22533 			 */
22534 			if (rsm == NULL)
22535 				msb = sb;
22536 			else
22537 				msb = NULL;
22538 			m->m_next = tcp_m_copym(
22539 				mb, moff, &len,
22540 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
22541 				((rsm == NULL) ? hw_tls : 0)
22542 #ifdef NETFLIX_COPY_ARGS
22543 				, &s_mb, &s_moff
22544 #endif
22545 				);
22546 			if (len <= (tp->t_maxseg - optlen)) {
22547 				/*
22548 				 * Must have ran out of mbufs for the copy
22549 				 * shorten it to no longer need tso. Lets
22550 				 * not put on sendalot since we are low on
22551 				 * mbufs.
22552 				 */
22553 				tso = 0;
22554 			}
22555 			if (m->m_next == NULL) {
22556 				SOCKBUF_UNLOCK(sb);
22557 				(void)m_free(m);
22558 				error = ENOBUFS;
22559 				sack_rxmit = 0;
22560 				goto out;
22561 			}
22562 		}
22563 		if (sack_rxmit) {
22564 			if (rsm && (rsm->r_flags & RACK_TLP)) {
22565 				/*
22566 				 * TLP should not count in retran count, but
22567 				 * in its own bin
22568 				 */
22569 				counter_u64_add(rack_tlp_retran, 1);
22570 				counter_u64_add(rack_tlp_retran_bytes, len);
22571 			} else {
22572 				tp->t_sndrexmitpack++;
22573 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
22574 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
22575 			}
22576 #ifdef STATS
22577 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
22578 						 len);
22579 #endif
22580 		} else {
22581 			KMOD_TCPSTAT_INC(tcps_sndpack);
22582 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
22583 #ifdef STATS
22584 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
22585 						 len);
22586 #endif
22587 		}
22588 		/*
22589 		 * If we're sending everything we've got, set PUSH. (This
22590 		 * will keep happy those implementations which only give
22591 		 * data to the user when a buffer fills or a PUSH comes in.)
22592 		 */
22593 		if (sb_offset + len == sbused(sb) &&
22594 		    sbused(sb) &&
22595 		    !(flags & TH_SYN)) {
22596 			flags |= TH_PUSH;
22597 			add_flag |= RACK_HAD_PUSH;
22598 		}
22599 
22600 		SOCKBUF_UNLOCK(sb);
22601 	} else {
22602 		SOCKBUF_UNLOCK(sb);
22603 		if (tp->t_flags & TF_ACKNOW)
22604 			KMOD_TCPSTAT_INC(tcps_sndacks);
22605 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
22606 			KMOD_TCPSTAT_INC(tcps_sndctrl);
22607 		else
22608 			KMOD_TCPSTAT_INC(tcps_sndwinup);
22609 
22610 		m = m_gethdr(M_NOWAIT, MT_DATA);
22611 		if (m == NULL) {
22612 			error = ENOBUFS;
22613 			sack_rxmit = 0;
22614 			goto out;
22615 		}
22616 #ifdef INET6
22617 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
22618 		    MHLEN >= hdrlen) {
22619 			M_ALIGN(m, hdrlen);
22620 		} else
22621 #endif
22622 			m->m_data += max_linkhdr;
22623 		m->m_len = hdrlen;
22624 	}
22625 	SOCKBUF_UNLOCK_ASSERT(sb);
22626 	m->m_pkthdr.rcvif = (struct ifnet *)0;
22627 #ifdef MAC
22628 	mac_inpcb_create_mbuf(inp, m);
22629 #endif
22630 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
22631 #ifdef INET6
22632 		if (isipv6)
22633 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
22634 		else
22635 #endif				/* INET6 */
22636 #ifdef INET
22637 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
22638 #endif
22639 		th = rack->r_ctl.fsb.th;
22640 		udp = rack->r_ctl.fsb.udp;
22641 		if (udp) {
22642 #ifdef INET6
22643 			if (isipv6)
22644 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
22645 			else
22646 #endif				/* INET6 */
22647 				ulen = hdrlen + len - sizeof(struct ip);
22648 			udp->uh_ulen = htons(ulen);
22649 		}
22650 	} else {
22651 #ifdef INET6
22652 		if (isipv6) {
22653 			ip6 = mtod(m, struct ip6_hdr *);
22654 			if (tp->t_port) {
22655 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
22656 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
22657 				udp->uh_dport = tp->t_port;
22658 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
22659 				udp->uh_ulen = htons(ulen);
22660 				th = (struct tcphdr *)(udp + 1);
22661 			} else
22662 				th = (struct tcphdr *)(ip6 + 1);
22663 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
22664 		} else
22665 #endif				/* INET6 */
22666 		{
22667 #ifdef INET
22668 			ip = mtod(m, struct ip *);
22669 			if (tp->t_port) {
22670 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
22671 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
22672 				udp->uh_dport = tp->t_port;
22673 				ulen = hdrlen + len - sizeof(struct ip);
22674 				udp->uh_ulen = htons(ulen);
22675 				th = (struct tcphdr *)(udp + 1);
22676 			} else
22677 				th = (struct tcphdr *)(ip + 1);
22678 			tcpip_fillheaders(inp, tp->t_port, ip, th);
22679 #endif
22680 		}
22681 	}
22682 	/*
22683 	 * If we are starting a connection, send ECN setup SYN packet. If we
22684 	 * are on a retransmit, we may resend those bits a number of times
22685 	 * as per RFC 3168.
22686 	 */
22687 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
22688 		flags |= tcp_ecn_output_syn_sent(tp);
22689 	}
22690 	/* Also handle parallel SYN for ECN */
22691 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
22692 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
22693 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
22694 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
22695 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
22696 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
22697 #ifdef INET6
22698 		if (isipv6) {
22699 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
22700 			ip6->ip6_flow |= htonl(ect << 20);
22701 		}
22702 		else
22703 #endif
22704 		{
22705 #ifdef INET
22706 			ip->ip_tos &= ~IPTOS_ECN_MASK;
22707 			ip->ip_tos |= ect;
22708 #endif
22709 		}
22710 	}
22711 	th->th_seq = htonl(rack_seq);
22712 	th->th_ack = htonl(tp->rcv_nxt);
22713 	tcp_set_flags(th, flags);
22714 	/*
22715 	 * Calculate receive window.  Don't shrink window, but avoid silly
22716 	 * window syndrome.
22717 	 * If a RST segment is sent, advertise a window of zero.
22718 	 */
22719 	if (flags & TH_RST) {
22720 		recwin = 0;
22721 	} else {
22722 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
22723 		    recwin < (long)segsiz) {
22724 			recwin = 0;
22725 		}
22726 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
22727 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
22728 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
22729 	}
22730 
22731 	/*
22732 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
22733 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
22734 	 * handled in syncache.
22735 	 */
22736 	if (flags & TH_SYN)
22737 		th->th_win = htons((u_short)
22738 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
22739 	else {
22740 		/* Avoid shrinking window with window scaling. */
22741 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
22742 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
22743 	}
22744 	/*
22745 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
22746 	 * window.  This may cause the remote transmitter to stall.  This
22747 	 * flag tells soreceive() to disable delayed acknowledgements when
22748 	 * draining the buffer.  This can occur if the receiver is
22749 	 * attempting to read more data than can be buffered prior to
22750 	 * transmitting on the connection.
22751 	 */
22752 	if (th->th_win == 0) {
22753 		tp->t_sndzerowin++;
22754 		tp->t_flags |= TF_RXWIN0SENT;
22755 	} else
22756 		tp->t_flags &= ~TF_RXWIN0SENT;
22757 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
22758 	/* Now are we using fsb?, if so copy the template data to the mbuf */
22759 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
22760 		uint8_t *cpto;
22761 
22762 		cpto = mtod(m, uint8_t *);
22763 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
22764 		/*
22765 		 * We have just copied in:
22766 		 * IP/IP6
22767 		 * <optional udphdr>
22768 		 * tcphdr (no options)
22769 		 *
22770 		 * We need to grab the correct pointers into the mbuf
22771 		 * for both the tcp header, and possibly the udp header (if tunneling).
22772 		 * We do this by using the offset in the copy buffer and adding it
22773 		 * to the mbuf base pointer (cpto).
22774 		 */
22775 #ifdef INET6
22776 		if (isipv6)
22777 			ip6 = mtod(m, struct ip6_hdr *);
22778 		else
22779 #endif				/* INET6 */
22780 #ifdef INET
22781 			ip = mtod(m, struct ip *);
22782 #endif
22783 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
22784 		/* If we have a udp header lets set it into the mbuf as well */
22785 		if (udp)
22786 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
22787 	}
22788 	if (optlen) {
22789 		bcopy(opt, th + 1, optlen);
22790 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
22791 	}
22792 	/*
22793 	 * Put TCP length in extended header, and then checksum extended
22794 	 * header and data.
22795 	 */
22796 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
22797 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
22798 	if (to.to_flags & TOF_SIGNATURE) {
22799 		/*
22800 		 * Calculate MD5 signature and put it into the place
22801 		 * determined before.
22802 		 * NOTE: since TCP options buffer doesn't point into
22803 		 * mbuf's data, calculate offset and use it.
22804 		 */
22805 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
22806 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
22807 			/*
22808 			 * Do not send segment if the calculation of MD5
22809 			 * digest has failed.
22810 			 */
22811 			goto out;
22812 		}
22813 	}
22814 #endif
22815 #ifdef INET6
22816 	if (isipv6) {
22817 		/*
22818 		 * ip6_plen is not need to be filled now, and will be filled
22819 		 * in ip6_output.
22820 		 */
22821 		if (tp->t_port) {
22822 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
22823 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
22824 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
22825 			th->th_sum = htons(0);
22826 			UDPSTAT_INC(udps_opackets);
22827 		} else {
22828 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
22829 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
22830 			th->th_sum = in6_cksum_pseudo(ip6,
22831 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
22832 						      0);
22833 		}
22834 	}
22835 #endif
22836 #if defined(INET6) && defined(INET)
22837 	else
22838 #endif
22839 #ifdef INET
22840 	{
22841 		if (tp->t_port) {
22842 			m->m_pkthdr.csum_flags = CSUM_UDP;
22843 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
22844 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
22845 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
22846 			th->th_sum = htons(0);
22847 			UDPSTAT_INC(udps_opackets);
22848 		} else {
22849 			m->m_pkthdr.csum_flags = CSUM_TCP;
22850 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
22851 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
22852 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
22853 									IPPROTO_TCP + len + optlen));
22854 		}
22855 		/* IP version must be set here for ipv4/ipv6 checking later */
22856 		KASSERT(ip->ip_v == IPVERSION,
22857 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
22858 	}
22859 #endif
22860 	/*
22861 	 * Enable TSO and specify the size of the segments. The TCP pseudo
22862 	 * header checksum is always provided. XXX: Fixme: This is currently
22863 	 * not the case for IPv6.
22864 	 */
22865 	if (tso) {
22866 		/*
22867 		 * Here we must use t_maxseg and the optlen since
22868 		 * the optlen may include SACK's (or DSACK).
22869 		 */
22870 		KASSERT(len > tp->t_maxseg - optlen,
22871 			("%s: len <= tso_segsz", __func__));
22872 		m->m_pkthdr.csum_flags |= CSUM_TSO;
22873 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
22874 	}
22875 	KASSERT(len + hdrlen == m_length(m, NULL),
22876 		("%s: mbuf chain different than expected: %d + %u != %u",
22877 		 __func__, len, hdrlen, m_length(m, NULL)));
22878 
22879 #ifdef TCP_HHOOK
22880 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
22881 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
22882 #endif
22883 	if ((rack->r_ctl.crte != NULL) &&
22884 	    (rack->rc_hw_nobuf == 0) &&
22885 	    tcp_bblogging_on(tp)) {
22886 		rack_log_queue_level(tp, rack, len, &tv, cts);
22887 	}
22888 	/* We're getting ready to send; log now. */
22889 	if (tcp_bblogging_on(rack->rc_tp)) {
22890 		union tcp_log_stackspecific log;
22891 
22892 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
22893 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
22894 		if (rack->rack_no_prr)
22895 			log.u_bbr.flex1 = 0;
22896 		else
22897 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
22898 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
22899 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
22900 		log.u_bbr.flex4 = orig_len;
22901 		/* Save off the early/late values */
22902 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
22903 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
22904 		log.u_bbr.bw_inuse = rack_get_bw(rack);
22905 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
22906 		log.u_bbr.flex8 = 0;
22907 		if (rsm) {
22908 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
22909 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
22910 				counter_u64_add(rack_collapsed_win_rxt, 1);
22911 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
22912 			}
22913 			if (doing_tlp)
22914 				log.u_bbr.flex8 = 2;
22915 			else
22916 				log.u_bbr.flex8 = 1;
22917 		} else {
22918 			if (doing_tlp)
22919 				log.u_bbr.flex8 = 3;
22920 		}
22921 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
22922 		log.u_bbr.flex7 = mark;
22923 		log.u_bbr.flex7 <<= 8;
22924 		log.u_bbr.flex7 |= pass;
22925 		log.u_bbr.pkts_out = tp->t_maxseg;
22926 		log.u_bbr.timeStamp = cts;
22927 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
22928 		if (rsm && (rsm->r_rtr_cnt > 0)) {
22929 			/*
22930 			 * When we have a retransmit we want to log the
22931 			 * burst at send and flight at send from before.
22932 			 */
22933 			log.u_bbr.flex5 = rsm->r_fas;
22934 			log.u_bbr.bbr_substate = rsm->r_bas;
22935 		} else {
22936 			/*
22937 			 * New transmits we log in flex5 the inflight again as
22938 			 * well as the number of segments in our send in the
22939 			 * substate field.
22940 			 */
22941 			log.u_bbr.flex5 = log.u_bbr.inflight;
22942 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
22943 		}
22944 		log.u_bbr.lt_epoch = cwnd_to_use;
22945 		log.u_bbr.delivered = sendalot;
22946 		log.u_bbr.rttProp = (uintptr_t)rsm;
22947 		log.u_bbr.pkt_epoch = __LINE__;
22948 		if (rsm) {
22949 			log.u_bbr.delRate = rsm->r_flags;
22950 			log.u_bbr.delRate <<= 31;
22951 			log.u_bbr.delRate |= rack->r_must_retran;
22952 			log.u_bbr.delRate <<= 1;
22953 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22954 		} else {
22955 			log.u_bbr.delRate = rack->r_must_retran;
22956 			log.u_bbr.delRate <<= 1;
22957 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22958 		}
22959 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
22960 				    len, &log, false, NULL, __func__, __LINE__, &tv);
22961 	} else
22962 		lgb = NULL;
22963 
22964 	/*
22965 	 * Fill in IP length and desired time to live and send to IP level.
22966 	 * There should be a better way to handle ttl and tos; we could keep
22967 	 * them in the template, but need a way to checksum without them.
22968 	 */
22969 	/*
22970 	 * m->m_pkthdr.len should have been set before cksum calcuration,
22971 	 * because in6_cksum() need it.
22972 	 */
22973 #ifdef INET6
22974 	if (isipv6) {
22975 		/*
22976 		 * we separately set hoplimit for every segment, since the
22977 		 * user might want to change the value via setsockopt. Also,
22978 		 * desired default hop limit might be changed via Neighbor
22979 		 * Discovery.
22980 		 */
22981 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
22982 
22983 		/*
22984 		 * Set the packet size here for the benefit of DTrace
22985 		 * probes. ip6_output() will set it properly; it's supposed
22986 		 * to include the option header lengths as well.
22987 		 */
22988 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
22989 
22990 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
22991 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22992 		else
22993 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22994 
22995 		if (tp->t_state == TCPS_SYN_SENT)
22996 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
22997 
22998 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
22999 		/* TODO: IPv6 IP6TOS_ECT bit on */
23000 		error = ip6_output(m,
23001 				   inp->in6p_outputopts,
23002 				   &inp->inp_route6,
23003 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
23004 				   NULL, NULL, inp);
23005 
23006 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
23007 			mtu = inp->inp_route6.ro_nh->nh_mtu;
23008 	}
23009 #endif				/* INET6 */
23010 #if defined(INET) && defined(INET6)
23011 	else
23012 #endif
23013 #ifdef INET
23014 	{
23015 		ip->ip_len = htons(m->m_pkthdr.len);
23016 #ifdef INET6
23017 		if (inp->inp_vflag & INP_IPV6PROTO)
23018 			ip->ip_ttl = in6_selecthlim(inp, NULL);
23019 #endif				/* INET6 */
23020 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
23021 		/*
23022 		 * If we do path MTU discovery, then we set DF on every
23023 		 * packet. This might not be the best thing to do according
23024 		 * to RFC3390 Section 2. However the tcp hostcache migitates
23025 		 * the problem so it affects only the first tcp connection
23026 		 * with a host.
23027 		 *
23028 		 * NB: Don't set DF on small MTU/MSS to have a safe
23029 		 * fallback.
23030 		 */
23031 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
23032 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
23033 			if (tp->t_port == 0 || len < V_tcp_minmss) {
23034 				ip->ip_off |= htons(IP_DF);
23035 			}
23036 		} else {
23037 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
23038 		}
23039 
23040 		if (tp->t_state == TCPS_SYN_SENT)
23041 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
23042 
23043 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
23044 
23045 		error = ip_output(m,
23046 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
23047 				  inp->inp_options,
23048 #else
23049 				  NULL,
23050 #endif
23051 				  &inp->inp_route,
23052 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
23053 				  inp);
23054 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
23055 			mtu = inp->inp_route.ro_nh->nh_mtu;
23056 	}
23057 #endif				/* INET */
23058 	if (lgb) {
23059 		lgb->tlb_errno = error;
23060 		lgb = NULL;
23061 	}
23062 
23063 out:
23064 	/*
23065 	 * In transmit state, time the transmission and arrange for the
23066 	 * retransmit.  In persist state, just set snd_max.
23067 	 */
23068 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
23069 			rack_to_usec_ts(&tv),
23070 			rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
23071 	if (error == 0) {
23072 		if (add_flag & RACK_IS_PCM) {
23073 			/* We just launched a PCM */
23074 			/* rrs here log */
23075 			rack->pcm_in_progress = 1;
23076 			rack->pcm_needed = 0;
23077 			rack_log_pcm(rack, 7, len, rack->r_ctl.pcm_max_seg,  add_flag);
23078 		}
23079 		if (rsm == NULL) {
23080 			if (rack->lt_bw_up == 0) {
23081 				rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
23082 				rack->r_ctl.lt_seq = tp->snd_una;
23083 				rack->lt_bw_up = 1;
23084 			} else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
23085 				/*
23086 				 * Need to record what we have since we are
23087 				 * approaching seq wrap.
23088 				 */
23089 				uint64_t tmark;
23090 
23091 				rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
23092 				rack->r_ctl.lt_seq = tp->snd_una;
23093 				tmark = tcp_get_u64_usecs(&tv);
23094 				if (tmark > rack->r_ctl.lt_timemark) {
23095 					rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
23096 					rack->r_ctl.lt_timemark = tmark;
23097 				}
23098 			}
23099 		}
23100 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
23101 		counter_u64_add(rack_total_bytes, len);
23102 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
23103 		if (rsm && doing_tlp) {
23104 			rack->rc_last_sent_tlp_past_cumack = 0;
23105 			rack->rc_last_sent_tlp_seq_valid = 1;
23106 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
23107 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
23108 		}
23109 		if (rack->rc_hw_nobuf) {
23110 			rack->rc_hw_nobuf = 0;
23111 			rack->r_ctl.rc_agg_delayed = 0;
23112 			rack->r_early = 0;
23113 			rack->r_late = 0;
23114 			rack->r_ctl.rc_agg_early = 0;
23115 		}
23116 		if (rsm && (doing_tlp == 0)) {
23117 			/* Set we retransmitted */
23118 			rack->rc_gp_saw_rec = 1;
23119 		} else {
23120 			if (cwnd_to_use > tp->snd_ssthresh) {
23121 				/* Set we sent in CA */
23122 				rack->rc_gp_saw_ca = 1;
23123 			} else {
23124 				/* Set we sent in SS */
23125 				rack->rc_gp_saw_ss = 1;
23126 			}
23127 		}
23128 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
23129 		    (tp->t_flags & TF_SACK_PERMIT) &&
23130 		    tp->rcv_numsacks > 0)
23131 			tcp_clean_dsack_blocks(tp);
23132 		tot_len_this_send += len;
23133 		if (len == 0) {
23134 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
23135 		} else {
23136 			int idx;
23137 
23138 			idx = (len / segsiz) + 3;
23139 			if (idx >= TCP_MSS_ACCT_ATIMER)
23140 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
23141 			else
23142 				counter_u64_add(rack_out_size[idx], 1);
23143 		}
23144 	}
23145 	if ((rack->rack_no_prr == 0) &&
23146 	    sub_from_prr &&
23147 	    (error == 0)) {
23148 		if (rack->r_ctl.rc_prr_sndcnt >= len)
23149 			rack->r_ctl.rc_prr_sndcnt -= len;
23150 		else
23151 			rack->r_ctl.rc_prr_sndcnt = 0;
23152 	}
23153 	sub_from_prr = 0;
23154 	if (doing_tlp) {
23155 		/* Make sure the TLP is added */
23156 		add_flag |= RACK_TLP;
23157 	} else if (rsm) {
23158 		/* If its a resend without TLP then it must not have the flag */
23159 		rsm->r_flags &= ~RACK_TLP;
23160 	}
23161 
23162 
23163 	if ((error == 0) &&
23164 	    (len > 0) &&
23165 	    (tp->snd_una == tp->snd_max))
23166 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
23167 
23168 	{
23169 		/*
23170 		 * This block is not associated with the above error == 0 test.
23171 		 * It is used to advance snd_max if we have a new transmit.
23172 		 */
23173 		tcp_seq startseq = tp->snd_max;
23174 
23175 
23176 		if (rsm && (doing_tlp == 0))
23177 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
23178 		if (error)
23179 			/* We don't log or do anything with errors */
23180 			goto nomore;
23181 		if (doing_tlp == 0) {
23182 			if (rsm == NULL) {
23183 				/*
23184 				 * Not a retransmission of some
23185 				 * sort, new data is going out so
23186 				 * clear our TLP count and flag.
23187 				 */
23188 				rack->rc_tlp_in_progress = 0;
23189 				rack->r_ctl.rc_tlp_cnt_out = 0;
23190 			}
23191 		} else {
23192 			/*
23193 			 * We have just sent a TLP, mark that it is true
23194 			 * and make sure our in progress is set so we
23195 			 * continue to check the count.
23196 			 */
23197 			rack->rc_tlp_in_progress = 1;
23198 			rack->r_ctl.rc_tlp_cnt_out++;
23199 		}
23200 		/*
23201 		 * If we are retransmitting we are done, snd_max
23202 		 * does not get updated.
23203 		 */
23204 		if (sack_rxmit)
23205 			goto nomore;
23206 		if ((tp->snd_una == tp->snd_max) && (len > 0)) {
23207 			/*
23208 			 * Update the time we just added data since
23209 			 * nothing was outstanding.
23210 			 */
23211 			rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
23212 			tp->t_acktime = ticks;
23213 		}
23214 		/*
23215 		 * Now for special SYN/FIN handling.
23216 		 */
23217 		if (flags & (TH_SYN | TH_FIN)) {
23218 			if ((flags & TH_SYN) &&
23219 			    ((tp->t_flags & TF_SENTSYN) == 0)) {
23220 				tp->snd_max++;
23221 				tp->t_flags |= TF_SENTSYN;
23222 			}
23223 			if ((flags & TH_FIN) &&
23224 			    ((tp->t_flags & TF_SENTFIN) == 0)) {
23225 				tp->snd_max++;
23226 				tp->t_flags |= TF_SENTFIN;
23227 			}
23228 		}
23229 		tp->snd_max += len;
23230 		if (rack->rc_new_rnd_needed) {
23231 			rack_new_round_starts(tp, rack, tp->snd_max);
23232 		}
23233 		/*
23234 		 * Time this transmission if not a retransmission and
23235 		 * not currently timing anything.
23236 		 * This is only relevant in case of switching back to
23237 		 * the base stack.
23238 		 */
23239 		if (tp->t_rtttime == 0) {
23240 			tp->t_rtttime = ticks;
23241 			tp->t_rtseq = startseq;
23242 			KMOD_TCPSTAT_INC(tcps_segstimed);
23243 		}
23244 		if (len &&
23245 		    ((tp->t_flags & TF_GPUTINPROG) == 0))
23246 			rack_start_gp_measurement(tp, rack, startseq, sb_offset);
23247 		/*
23248 		 * If we are doing FO we need to update the mbuf position and subtract
23249 		 * this happens when the peer sends us duplicate information and
23250 		 * we thus want to send a DSACK.
23251 		 *
23252 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
23253 		 * turned off? If not then we are going to echo multiple DSACK blocks
23254 		 * out (with the TSO), which we should not be doing.
23255 		 */
23256 		if (rack->r_fast_output && len) {
23257 			if (rack->r_ctl.fsb.left_to_send > len)
23258 				rack->r_ctl.fsb.left_to_send -= len;
23259 			else
23260 				rack->r_ctl.fsb.left_to_send = 0;
23261 			if (rack->r_ctl.fsb.left_to_send < segsiz)
23262 				rack->r_fast_output = 0;
23263 			if (rack->r_fast_output) {
23264 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
23265 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
23266 				rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
23267 			}
23268 		}
23269 		if (rack_pcm_blast == 0) {
23270 			if ((orig_len > len) &&
23271 			    (add_flag & RACK_IS_PCM) &&
23272 			    (len < pace_max_seg) &&
23273 			    ((pace_max_seg - len) > segsiz)) {
23274 				/*
23275 				 * We are doing a PCM measurement and we did
23276 				 * not get enough data in the TSO to meet the
23277 				 * burst requirement.
23278 				 */
23279 				uint32_t n_len;
23280 
23281 				n_len = (orig_len - len);
23282 				orig_len -= len;
23283 				pace_max_seg -= len;
23284 				len = n_len;
23285 				sb_offset = tp->snd_max - tp->snd_una;
23286 				/* Re-lock for the next spin */
23287 				SOCKBUF_LOCK(sb);
23288 				goto send;
23289 			}
23290 		} else {
23291 			if ((orig_len > len) &&
23292 			    (add_flag & RACK_IS_PCM) &&
23293 			    ((orig_len - len) > segsiz)) {
23294 				/*
23295 				 * We are doing a PCM measurement and we did
23296 				 * not get enough data in the TSO to meet the
23297 				 * burst requirement.
23298 				 */
23299 				uint32_t n_len;
23300 
23301 				n_len = (orig_len - len);
23302 				orig_len -= len;
23303 				len = n_len;
23304 				sb_offset = tp->snd_max - tp->snd_una;
23305 				/* Re-lock for the next spin */
23306 				SOCKBUF_LOCK(sb);
23307 				goto send;
23308 			}
23309 		}
23310 	}
23311 nomore:
23312 	if (error) {
23313 		rack->r_ctl.rc_agg_delayed = 0;
23314 		rack->r_early = 0;
23315 		rack->r_late = 0;
23316 		rack->r_ctl.rc_agg_early = 0;
23317 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
23318 		/*
23319 		 * Failures do not advance the seq counter above. For the
23320 		 * case of ENOBUFS we will fall out and retry in 1ms with
23321 		 * the hpts. Everything else will just have to retransmit
23322 		 * with the timer.
23323 		 *
23324 		 * In any case, we do not want to loop around for another
23325 		 * send without a good reason.
23326 		 */
23327 		sendalot = 0;
23328 		switch (error) {
23329 		case EPERM:
23330 		case EACCES:
23331 			tp->t_softerror = error;
23332 #ifdef TCP_ACCOUNTING
23333 			crtsc = get_cyclecount();
23334 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23335 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
23336 			}
23337 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23338 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
23339 			}
23340 			sched_unpin();
23341 #endif
23342 			return (error);
23343 		case ENOBUFS:
23344 			/*
23345 			 * Pace us right away to retry in a some
23346 			 * time
23347 			 */
23348 			if (rack->r_ctl.crte != NULL) {
23349 				tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
23350 				if (tcp_bblogging_on(rack->rc_tp))
23351 					rack_log_queue_level(tp, rack, len, &tv, cts);
23352 			} else
23353 				tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
23354 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
23355 			if (rack->rc_enobuf < 0x7f)
23356 				rack->rc_enobuf++;
23357 			if (slot < (10 * HPTS_USEC_IN_MSEC))
23358 				slot = 10 * HPTS_USEC_IN_MSEC;
23359 			if (rack->r_ctl.crte != NULL) {
23360 				counter_u64_add(rack_saw_enobuf_hw, 1);
23361 				tcp_rl_log_enobuf(rack->r_ctl.crte);
23362 			}
23363 			counter_u64_add(rack_saw_enobuf, 1);
23364 			goto enobufs;
23365 		case EMSGSIZE:
23366 			/*
23367 			 * For some reason the interface we used initially
23368 			 * to send segments changed to another or lowered
23369 			 * its MTU. If TSO was active we either got an
23370 			 * interface without TSO capabilits or TSO was
23371 			 * turned off. If we obtained mtu from ip_output()
23372 			 * then update it and try again.
23373 			 */
23374 			if (tso)
23375 				tp->t_flags &= ~TF_TSO;
23376 			if (mtu != 0) {
23377 				int saved_mtu;
23378 
23379 				saved_mtu = tp->t_maxseg;
23380 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
23381 				if (saved_mtu > tp->t_maxseg) {
23382 					goto again;
23383 				}
23384 			}
23385 			slot = 10 * HPTS_USEC_IN_MSEC;
23386 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
23387 #ifdef TCP_ACCOUNTING
23388 			crtsc = get_cyclecount();
23389 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23390 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
23391 			}
23392 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23393 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
23394 			}
23395 			sched_unpin();
23396 #endif
23397 			return (error);
23398 		case ENETUNREACH:
23399 			counter_u64_add(rack_saw_enetunreach, 1);
23400 		case EHOSTDOWN:
23401 		case EHOSTUNREACH:
23402 		case ENETDOWN:
23403 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
23404 				tp->t_softerror = error;
23405 			}
23406 			/* FALLTHROUGH */
23407 		default:
23408 			slot = 10 * HPTS_USEC_IN_MSEC;
23409 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
23410 #ifdef TCP_ACCOUNTING
23411 			crtsc = get_cyclecount();
23412 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23413 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
23414 			}
23415 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23416 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
23417 			}
23418 			sched_unpin();
23419 #endif
23420 			return (error);
23421 		}
23422 	} else {
23423 		rack->rc_enobuf = 0;
23424 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
23425 			rack->r_ctl.retran_during_recovery += len;
23426 	}
23427 	KMOD_TCPSTAT_INC(tcps_sndtotal);
23428 
23429 	/*
23430 	 * Data sent (as far as we can tell). If this advertises a larger
23431 	 * window than any other segment, then remember the size of the
23432 	 * advertised window. Any pending ACK has now been sent.
23433 	 */
23434 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
23435 		tp->rcv_adv = tp->rcv_nxt + recwin;
23436 
23437 	tp->last_ack_sent = tp->rcv_nxt;
23438 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
23439 enobufs:
23440 	if (sendalot) {
23441 		/* Do we need to turn off sendalot? */
23442 		if (pace_max_seg &&
23443 		    (tot_len_this_send >= pace_max_seg)) {
23444 			/* We hit our max. */
23445 			sendalot = 0;
23446 		}
23447 	}
23448 	if ((error == 0) && (flags & TH_FIN))
23449 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
23450 	if (flags & TH_RST) {
23451 		/*
23452 		 * We don't send again after sending a RST.
23453 		 */
23454 		slot = 0;
23455 		sendalot = 0;
23456 		if (error == 0)
23457 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
23458 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
23459 		/*
23460 		 * Get our pacing rate, if an error
23461 		 * occurred in sending (ENOBUF) we would
23462 		 * hit the else if with slot preset. Other
23463 		 * errors return.
23464 		 */
23465 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz, __LINE__);
23466 	}
23467 	/* We have sent clear the flag */
23468 	rack->r_ent_rec_ns = 0;
23469 	if (rack->r_must_retran) {
23470 		if (rsm) {
23471 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
23472 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
23473 				/*
23474 				 * We have retransmitted all.
23475 				 */
23476 				rack->r_must_retran = 0;
23477 				rack->r_ctl.rc_out_at_rto = 0;
23478 			}
23479 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
23480 			/*
23481 			 * Sending new data will also kill
23482 			 * the loop.
23483 			 */
23484 			rack->r_must_retran = 0;
23485 			rack->r_ctl.rc_out_at_rto = 0;
23486 		}
23487 	}
23488 	rack->r_ctl.fsb.recwin = recwin;
23489 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
23490 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
23491 		/*
23492 		 * We hit an RTO and now have past snd_max at the RTO
23493 		 * clear all the WAS flags.
23494 		 */
23495 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
23496 	}
23497 	if (slot) {
23498 		/* set the rack tcb into the slot N */
23499 		if ((error == 0) &&
23500 		    rack_use_rfo &&
23501 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
23502 		    (rsm == NULL) &&
23503 		    (ipoptlen == 0) &&
23504 		    (tp->rcv_numsacks == 0) &&
23505 		    (rack->rc_policer_detected == 0)  &&
23506 		    rack->r_fsb_inited &&
23507 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
23508 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
23509 		    (rack->r_must_retran == 0) &&
23510 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
23511 		    (len > 0) && (orig_len > 0) &&
23512 		    (orig_len > len) &&
23513 		    ((orig_len - len) >= segsiz) &&
23514 		    ((optlen == 0) ||
23515 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
23516 			/* We can send at least one more MSS using our fsb */
23517 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
23518 					       segsiz, pace_max_seg, hw_tls, flags);
23519 		} else
23520 			rack->r_fast_output = 0;
23521 		rack_log_fsb(rack, tp, so, flags,
23522 			     ipoptlen, orig_len, len, error,
23523 			     (rsm == NULL), optlen, __LINE__, 2);
23524 	} else if (sendalot) {
23525 		int ret;
23526 
23527 		sack_rxmit = 0;
23528 		if ((error == 0) &&
23529 		    rack_use_rfo &&
23530 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
23531 		    (rsm == NULL) &&
23532 		    (ipoptlen == 0) &&
23533 		    (tp->rcv_numsacks == 0) &&
23534 		    (rack->r_must_retran == 0) &&
23535 		    rack->r_fsb_inited &&
23536 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
23537 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
23538 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
23539 		    (len > 0) && (orig_len > 0) &&
23540 		    (orig_len > len) &&
23541 		    ((orig_len - len) >= segsiz) &&
23542 		    ((optlen == 0) ||
23543 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
23544 			/* we can use fast_output for more */
23545 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
23546 					       segsiz, pace_max_seg, hw_tls, flags);
23547 			if (rack->r_fast_output) {
23548 				error = 0;
23549 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
23550 				if (ret >= 0)
23551 					return (ret);
23552 			        else if (error)
23553 					goto nomore;
23554 
23555 			}
23556 		}
23557 		goto again;
23558 	}
23559 skip_all_send:
23560 	/* Assure when we leave that snd_nxt will point to top */
23561 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
23562 		tp->snd_nxt = tp->snd_max;
23563 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
23564 #ifdef TCP_ACCOUNTING
23565 	crtsc = get_cyclecount() - ts_val;
23566 	if (tot_len_this_send) {
23567 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23568 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
23569 		}
23570 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23571 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
23572 		}
23573 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23574 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
23575 		}
23576 	} else {
23577 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23578 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
23579 		}
23580 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
23581 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
23582 		}
23583 	}
23584 	sched_unpin();
23585 #endif
23586 	if (error == ENOBUFS)
23587 		error = 0;
23588 	return (error);
23589 }
23590 
23591 static void
23592 rack_update_seg(struct tcp_rack *rack)
23593 {
23594 	uint32_t orig_val;
23595 
23596 	orig_val = rack->r_ctl.rc_pace_max_segs;
23597 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
23598 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
23599 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
23600 }
23601 
23602 static void
23603 rack_mtu_change(struct tcpcb *tp)
23604 {
23605 	/*
23606 	 * The MSS may have changed
23607 	 */
23608 	struct tcp_rack *rack;
23609 	struct rack_sendmap *rsm;
23610 
23611 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23612 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
23613 		/*
23614 		 * The MTU has changed we need to resend everything
23615 		 * since all we have sent is lost. We first fix
23616 		 * up the mtu though.
23617 		 */
23618 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23619 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
23620 		rack_remxt_tmr(tp);
23621 		rack->r_fast_output = 0;
23622 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
23623 						rack->r_ctl.rc_sacked);
23624 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
23625 		rack->r_must_retran = 1;
23626 		/* Mark all inflight to needing to be rxt'd */
23627 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
23628 			rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
23629 		}
23630 	}
23631 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
23632 	/* We don't use snd_nxt to retransmit */
23633 	tp->snd_nxt = tp->snd_max;
23634 }
23635 
23636 static int
23637 rack_set_dgp(struct tcp_rack *rack)
23638 {
23639 	if (rack->dgp_on == 1)
23640 		return(0);
23641 	if ((rack->use_fixed_rate == 1) &&
23642 	    (rack->rc_always_pace == 1)) {
23643 		/*
23644 		 * We are already pacing another
23645 		 * way.
23646 		 */
23647 		return (EBUSY);
23648 	}
23649 	if (rack->rc_always_pace == 1) {
23650 		rack_remove_pacing(rack);
23651 	}
23652 	if (tcp_incr_dgp_pacing_cnt() == 0)
23653 		return (ENOSPC);
23654 	rack->r_ctl.pacing_method |= RACK_DGP_PACING;
23655 	rack->rc_fillcw_apply_discount = 0;
23656 	rack->dgp_on = 1;
23657 	rack->rc_always_pace = 1;
23658 	rack->rc_pace_dnd = 1;
23659 	rack->use_fixed_rate = 0;
23660 	if (rack->gp_ready)
23661 		rack_set_cc_pacing(rack);
23662 	rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23663 	rack->rack_attempt_hdwr_pace = 0;
23664 	/* rxt settings */
23665 	rack->full_size_rxt = 1;
23666 	rack->shape_rxt_to_pacing_min  = 0;
23667 	/* cmpack=1 */
23668 	rack->r_use_cmp_ack = 1;
23669 	if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
23670 	    rack->r_use_cmp_ack)
23671 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
23672 	/* scwnd=1 */
23673 	rack->rack_enable_scwnd = 1;
23674 	/* dynamic=100 */
23675 	rack->rc_gp_dyn_mul = 1;
23676 	/* gp_inc_ca */
23677 	rack->r_ctl.rack_per_of_gp_ca = 100;
23678 	/* rrr_conf=3 */
23679 	rack->r_rr_config = 3;
23680 	/* npush=2 */
23681 	rack->r_ctl.rc_no_push_at_mrtt = 2;
23682 	/* fillcw=1 */
23683 	rack->rc_pace_to_cwnd = 1;
23684 	rack->rc_pace_fill_if_rttin_range = 0;
23685 	rack->rtt_limit_mul = 0;
23686 	/* noprr=1 */
23687 	rack->rack_no_prr = 1;
23688 	/* lscwnd=1 */
23689 	rack->r_limit_scw = 1;
23690 	/* gp_inc_rec */
23691 	rack->r_ctl.rack_per_of_gp_rec = 90;
23692 	return (0);
23693 }
23694 
23695 static int
23696 rack_set_profile(struct tcp_rack *rack, int prof)
23697 {
23698 	int err = EINVAL;
23699 	if (prof == 1) {
23700 		/*
23701 		 * Profile 1 is "standard" DGP. It ignores
23702 		 * client buffer level.
23703 		 */
23704 		err = rack_set_dgp(rack);
23705 		if (err)
23706 			return (err);
23707 	} else if (prof == 6) {
23708 		err = rack_set_dgp(rack);
23709 		if (err)
23710 			return (err);
23711 		/*
23712 		 * Profile 6 tweaks DGP so that it will apply to
23713 		 * fill-cw the same settings that profile5 does
23714 		 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
23715 		 */
23716 		rack->rc_fillcw_apply_discount = 1;
23717 	} else if (prof == 0) {
23718 		/* This changes things back to the default settings */
23719 		if (rack->rc_always_pace == 1) {
23720 			rack_remove_pacing(rack);
23721 		} else {
23722 			/* Make sure any stray flags are off */
23723 			rack->dgp_on = 0;
23724 			rack->rc_hybrid_mode = 0;
23725 			rack->use_fixed_rate = 0;
23726 		}
23727 		err = 0;
23728 		if (rack_fill_cw_state)
23729 			rack->rc_pace_to_cwnd = 1;
23730 		else
23731 			rack->rc_pace_to_cwnd = 0;
23732 
23733 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
23734 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
23735 			rack->rc_always_pace = 1;
23736 			if (rack->rack_hibeta)
23737 				rack_set_cc_pacing(rack);
23738 		} else
23739 			rack->rc_always_pace = 0;
23740 		if (rack_dsack_std_based & 0x1) {
23741 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
23742 			rack->rc_rack_tmr_std_based = 1;
23743 		}
23744 		if (rack_dsack_std_based & 0x2) {
23745 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
23746 			rack->rc_rack_use_dsack = 1;
23747 		}
23748 		if (rack_use_cmp_acks)
23749 			rack->r_use_cmp_ack = 1;
23750 		else
23751 			rack->r_use_cmp_ack = 0;
23752 		if (rack_disable_prr)
23753 			rack->rack_no_prr = 1;
23754 		else
23755 			rack->rack_no_prr = 0;
23756 		if (rack_gp_no_rec_chg)
23757 			rack->rc_gp_no_rec_chg = 1;
23758 		else
23759 			rack->rc_gp_no_rec_chg = 0;
23760 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
23761 			rack->r_mbuf_queue = 1;
23762 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
23763 				rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
23764 			rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23765 		} else {
23766 			rack->r_mbuf_queue = 0;
23767 			rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23768 		}
23769 		if (rack_enable_shared_cwnd)
23770 			rack->rack_enable_scwnd = 1;
23771 		else
23772 			rack->rack_enable_scwnd = 0;
23773 		if (rack_do_dyn_mul) {
23774 			/* When dynamic adjustment is on CA needs to start at 100% */
23775 			rack->rc_gp_dyn_mul = 1;
23776 			if (rack_do_dyn_mul >= 100)
23777 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
23778 		} else {
23779 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
23780 			rack->rc_gp_dyn_mul = 0;
23781 		}
23782 		rack->r_rr_config = 0;
23783 		rack->r_ctl.rc_no_push_at_mrtt = 0;
23784 		rack->rc_pace_fill_if_rttin_range = 0;
23785 		rack->rtt_limit_mul = 0;
23786 
23787 		if (rack_enable_hw_pacing)
23788 			rack->rack_hdw_pace_ena = 1;
23789 		else
23790 			rack->rack_hdw_pace_ena = 0;
23791 		if (rack_disable_prr)
23792 			rack->rack_no_prr = 1;
23793 		else
23794 			rack->rack_no_prr = 0;
23795 		if (rack_limits_scwnd)
23796 			rack->r_limit_scw  = 1;
23797 		else
23798 			rack->r_limit_scw  = 0;
23799 		rack_init_retransmit_value(rack, rack_rxt_controls);
23800 		err = 0;
23801 	}
23802 	return (err);
23803 }
23804 
23805 static int
23806 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
23807 {
23808 	struct deferred_opt_list *dol;
23809 
23810 	dol = malloc(sizeof(struct deferred_opt_list),
23811 		     M_TCPDO, M_NOWAIT|M_ZERO);
23812 	if (dol == NULL) {
23813 		/*
23814 		 * No space yikes -- fail out..
23815 		 */
23816 		return (0);
23817 	}
23818 	dol->optname = sopt_name;
23819 	dol->optval = loptval;
23820 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
23821 	return (1);
23822 }
23823 
23824 static int
23825 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
23826 {
23827 #ifdef TCP_REQUEST_TRK
23828 	struct tcp_sendfile_track *sft;
23829 	struct timeval tv;
23830 	tcp_seq seq;
23831 	int err;
23832 
23833 	microuptime(&tv);
23834 
23835 	/* Make sure no fixed rate is on */
23836 	rack->use_fixed_rate = 0;
23837 	rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
23838 	rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
23839 	rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
23840 	/* Now allocate or find our entry that will have these settings */
23841 	sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
23842 	if (sft == NULL) {
23843 		rack->rc_tp->tcp_hybrid_error++;
23844 		/* no space, where would it have gone? */
23845 		seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
23846 		rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
23847 		return (ENOSPC);
23848 	}
23849 	/* mask our internal flags */
23850 	hybrid->hybrid_flags &= TCP_HYBRID_PACING_USER_MASK;
23851 	/* The seq will be snd_una + everything in the buffer */
23852 	seq = sft->start_seq;
23853 	if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
23854 		/* Disabling hybrid pacing */
23855 		if (rack->rc_hybrid_mode) {
23856 			rack_set_profile(rack, 0);
23857 			rack->rc_tp->tcp_hybrid_stop++;
23858 		}
23859 		rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
23860 		return (0);
23861 	}
23862 	if (rack->dgp_on == 0) {
23863 		/*
23864 		 * If we have not yet turned DGP on, do so
23865 		 * now setting pure DGP mode, no buffer level
23866 		 * response.
23867 		 */
23868 		if ((err = rack_set_profile(rack, 1)) != 0){
23869 			/* Failed to turn pacing on */
23870 			rack->rc_tp->tcp_hybrid_error++;
23871 			rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
23872 			return (err);
23873 		}
23874 	}
23875 	/*
23876 	 * Now we must switch to hybrid mode as well which also
23877 	 * means moving to regular pacing.
23878 	 */
23879 	if (rack->rc_hybrid_mode == 0) {
23880 		/* First time */
23881 		if (tcp_can_enable_pacing()) {
23882 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
23883 			rack->rc_hybrid_mode = 1;
23884 		} else {
23885 			return (ENOSPC);
23886 		}
23887 		if (rack->r_ctl.pacing_method & RACK_DGP_PACING) {
23888 			/*
23889 			 * This should be true.
23890 			 */
23891 			tcp_dec_dgp_pacing_cnt();
23892 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23893 		}
23894 	}
23895 	/* Now set in our flags */
23896 	sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
23897 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
23898 		sft->cspr = hybrid->cspr;
23899 	else
23900 		sft->cspr = 0;
23901 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
23902 		sft->hint_maxseg = hybrid->hint_maxseg;
23903 	else
23904 		sft->hint_maxseg = 0;
23905 	rack->rc_tp->tcp_hybrid_start++;
23906 	rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
23907 	return (0);
23908 #else
23909 	return (ENOTSUP);
23910 #endif
23911 }
23912 
23913 static int
23914 rack_stack_information(struct tcpcb *tp, struct stack_specific_info *si)
23915 {
23916 	/*
23917 	 * Gather rack specific information.
23918 	 */
23919 	struct tcp_rack *rack;
23920 
23921 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23922 	/* We pulled a SSI info log out what was there */
23923 	policer_detection_log(rack, rack->rc_highly_buffered, 0, 0, 0, 20);
23924 	if (rack->policer_detect_on) {
23925 		si->policer_detection_enabled = 1;
23926 		if (rack->rc_policer_detected) {
23927 			si->policer_detected = 1;
23928 			si->policer_bucket_size = rack->r_ctl.policer_bucket_size;
23929 			si->policer_last_bw = rack->r_ctl.policer_bw;
23930 		} else {
23931 			si->policer_detected = 0;
23932 			si->policer_bucket_size = 0;
23933 			si->policer_last_bw = 0;
23934 		}
23935 		si->current_round = rack->r_ctl.current_round;
23936 		si->highly_buffered = rack->rc_highly_buffered;
23937 	}
23938 	si->bytes_transmitted = tp->t_sndbytes;
23939 	si->bytes_retransmitted = tp->t_snd_rxt_bytes;
23940 	return (0);
23941 }
23942 
23943 static int
23944 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
23945 		    uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
23946 
23947 {
23948 	struct epoch_tracker et;
23949 	struct sockopt sopt;
23950 	struct cc_newreno_opts opt;
23951 	uint64_t val;
23952 	int error = 0;
23953 	uint16_t ca, ss;
23954 
23955 	switch (sopt_name) {
23956 	case TCP_RACK_SET_RXT_OPTIONS:
23957 		if ((optval >= 0) && (optval <= 2)) {
23958 			rack_init_retransmit_value(rack, optval);
23959 		} else {
23960 			/*
23961 			 * You must send in 0, 1 or 2 all else is
23962 			 * invalid.
23963 			 */
23964 			error = EINVAL;
23965 		}
23966 		break;
23967 	case TCP_RACK_DSACK_OPT:
23968 		RACK_OPTS_INC(tcp_rack_dsack_opt);
23969 		if (optval & 0x1) {
23970 			rack->rc_rack_tmr_std_based = 1;
23971 		} else {
23972 			rack->rc_rack_tmr_std_based = 0;
23973 		}
23974 		if (optval & 0x2) {
23975 			rack->rc_rack_use_dsack = 1;
23976 		} else {
23977 			rack->rc_rack_use_dsack = 0;
23978 		}
23979 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
23980 		break;
23981 	case TCP_RACK_PACING_DIVISOR:
23982 		RACK_OPTS_INC(tcp_rack_pacing_divisor);
23983 		if (optval == 0) {
23984 			rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
23985 		} else {
23986 			if (optval < RL_MIN_DIVISOR)
23987 				rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
23988 			else
23989 				rack->r_ctl.pace_len_divisor = optval;
23990 		}
23991 		break;
23992 	case TCP_RACK_HI_BETA:
23993 		RACK_OPTS_INC(tcp_rack_hi_beta);
23994 		if (optval > 0) {
23995 			rack->rack_hibeta = 1;
23996 			if ((optval >= 50) &&
23997 			    (optval <= 100)) {
23998 				/*
23999 				 * User wants to set a custom beta.
24000 				 */
24001 				rack->r_ctl.saved_hibeta = optval;
24002 				if (rack->rc_pacing_cc_set)
24003 					rack_undo_cc_pacing(rack);
24004 				rack->r_ctl.rc_saved_beta.beta = optval;
24005 			}
24006 			if (rack->rc_pacing_cc_set == 0)
24007 				rack_set_cc_pacing(rack);
24008 		} else {
24009 			rack->rack_hibeta = 0;
24010 			if (rack->rc_pacing_cc_set)
24011 				rack_undo_cc_pacing(rack);
24012 		}
24013 		break;
24014 	case TCP_RACK_PACING_BETA:
24015 		error = EINVAL;
24016 		break;
24017 	case TCP_RACK_TIMER_SLOP:
24018 		RACK_OPTS_INC(tcp_rack_timer_slop);
24019 		rack->r_ctl.timer_slop = optval;
24020 		if (rack->rc_tp->t_srtt) {
24021 			/*
24022 			 * If we have an SRTT lets update t_rxtcur
24023 			 * to have the new slop.
24024 			 */
24025 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
24026 					   rack_rto_min, rack_rto_max,
24027 					   rack->r_ctl.timer_slop);
24028 		}
24029 		break;
24030 	case TCP_RACK_PACING_BETA_ECN:
24031 		RACK_OPTS_INC(tcp_rack_beta_ecn);
24032 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
24033 			/* This only works for newreno. */
24034 			error = EINVAL;
24035 			break;
24036 		}
24037 		if (rack->rc_pacing_cc_set) {
24038 			/*
24039 			 * Set them into the real CC module
24040 			 * whats in the rack pcb is the old values
24041 			 * to be used on restoral/
24042 			 */
24043 			sopt.sopt_dir = SOPT_SET;
24044 			opt.name = CC_NEWRENO_BETA_ECN;
24045 			opt.val = optval;
24046 			if (CC_ALGO(tp)->ctl_output != NULL)
24047 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
24048 			else
24049 				error = ENOENT;
24050 		} else {
24051 			/*
24052 			 * Not pacing yet so set it into our local
24053 			 * rack pcb storage.
24054 			 */
24055 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
24056 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
24057 		}
24058 		break;
24059 	case TCP_DEFER_OPTIONS:
24060 		RACK_OPTS_INC(tcp_defer_opt);
24061 		if (optval) {
24062 			if (rack->gp_ready) {
24063 				/* Too late */
24064 				error = EINVAL;
24065 				break;
24066 			}
24067 			rack->defer_options = 1;
24068 		} else
24069 			rack->defer_options = 0;
24070 		break;
24071 	case TCP_RACK_MEASURE_CNT:
24072 		RACK_OPTS_INC(tcp_rack_measure_cnt);
24073 		if (optval && (optval <= 0xff)) {
24074 			rack->r_ctl.req_measurements = optval;
24075 		} else
24076 			error = EINVAL;
24077 		break;
24078 	case TCP_REC_ABC_VAL:
24079 		RACK_OPTS_INC(tcp_rec_abc_val);
24080 		if (optval > 0)
24081 			rack->r_use_labc_for_rec = 1;
24082 		else
24083 			rack->r_use_labc_for_rec = 0;
24084 		break;
24085 	case TCP_RACK_ABC_VAL:
24086 		RACK_OPTS_INC(tcp_rack_abc_val);
24087 		if ((optval > 0) && (optval < 255))
24088 			rack->rc_labc = optval;
24089 		else
24090 			error = EINVAL;
24091 		break;
24092 	case TCP_HDWR_UP_ONLY:
24093 		RACK_OPTS_INC(tcp_pacing_up_only);
24094 		if (optval)
24095 			rack->r_up_only = 1;
24096 		else
24097 			rack->r_up_only = 0;
24098 		break;
24099 	case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
24100 		RACK_OPTS_INC(tcp_fillcw_rate_cap);
24101 		rack->r_ctl.fillcw_cap = loptval;
24102 		break;
24103 	case TCP_PACING_RATE_CAP:
24104 		RACK_OPTS_INC(tcp_pacing_rate_cap);
24105 		if ((rack->dgp_on == 1) &&
24106 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
24107 			/*
24108 			 * If we are doing DGP we need to switch
24109 			 * to using the pacing limit.
24110 			 */
24111 			if (tcp_can_enable_pacing() == 0) {
24112 				error = ENOSPC;
24113 				break;
24114 			}
24115 			/*
24116 			 * Now change up the flags and counts to be correct.
24117 			 */
24118 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
24119 			tcp_dec_dgp_pacing_cnt();
24120 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
24121 		}
24122 		rack->r_ctl.bw_rate_cap = loptval;
24123 		break;
24124 	case TCP_HYBRID_PACING:
24125 		if (hybrid == NULL) {
24126 			error = EINVAL;
24127 			break;
24128 		}
24129 		if (rack->r_ctl.side_chan_dis_mask & HYBRID_DIS_MASK) {
24130 			error = EPERM;
24131 			break;
24132 		}
24133 		error = process_hybrid_pacing(rack, hybrid);
24134 		break;
24135 	case TCP_SIDECHAN_DIS:			/*  URL:scodm */
24136 		if (optval)
24137 			rack->r_ctl.side_chan_dis_mask = optval;
24138 		else
24139 			rack->r_ctl.side_chan_dis_mask = 0;
24140 		break;
24141 	case TCP_RACK_PROFILE:
24142 		RACK_OPTS_INC(tcp_profile);
24143 		error = rack_set_profile(rack, optval);
24144 		break;
24145 	case TCP_USE_CMP_ACKS:
24146 		RACK_OPTS_INC(tcp_use_cmp_acks);
24147 		if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
24148 			/* You can't turn it off once its on! */
24149 			error = EINVAL;
24150 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
24151 			rack->r_use_cmp_ack = 1;
24152 			rack->r_mbuf_queue = 1;
24153 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
24154 		}
24155 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
24156 			tp->t_flags2 |= TF2_MBUF_ACKCMP;
24157 		break;
24158 	case TCP_SHARED_CWND_TIME_LIMIT:
24159 		RACK_OPTS_INC(tcp_lscwnd);
24160 		if (optval)
24161 			rack->r_limit_scw = 1;
24162 		else
24163 			rack->r_limit_scw = 0;
24164 		break;
24165 	case TCP_RACK_DGP_IN_REC:
24166 		error = EINVAL;
24167 		break;
24168 	case TCP_POLICER_DETECT:		/*  URL:pol_det */
24169 		RACK_OPTS_INC(tcp_pol_detect);
24170 		rack_translate_policer_detect(rack, optval);
24171 		break;
24172 	case TCP_POLICER_MSS:
24173 		RACK_OPTS_INC(tcp_pol_mss);
24174 		rack->r_ctl.policer_del_mss = (uint8_t)optval;
24175 		if (optval & 0x00000100) {
24176 			/*
24177 			 * Value is setup like so:
24178 			 * VVVV VVVV VVVV VVVV VVVV VVAI MMMM MMMM
24179 			 * Where MMMM MMMM is MSS setting
24180 			 * I (9th bit) is the Postive value that
24181 			 * says it is being set (if its 0 then the
24182 			 * upper bits 11 - 32 have no meaning.
24183 			 * This allows setting it off with
24184 			 * 0x000001MM.
24185 			 *
24186 			 * The 10th bit is used to turn on the
24187 			 * alternate median (not the expanded one).
24188 			 *
24189 			 */
24190 			rack->r_ctl.pol_bw_comp = (optval >> 10);
24191 		}
24192 		if (optval & 0x00000200) {
24193 			rack->r_ctl.policer_alt_median = 1;
24194 		} else {
24195 			rack->r_ctl.policer_alt_median = 0;
24196 		}
24197 		break;
24198  	case TCP_RACK_PACE_TO_FILL:
24199 		RACK_OPTS_INC(tcp_fillcw);
24200 		if (optval == 0)
24201 			rack->rc_pace_to_cwnd = 0;
24202 		else {
24203 			rack->rc_pace_to_cwnd = 1;
24204 		}
24205 		if ((optval >= rack_gp_rtt_maxmul) &&
24206 		    rack_gp_rtt_maxmul &&
24207 		    (optval < 0xf)) {
24208 			rack->rc_pace_fill_if_rttin_range = 1;
24209 			rack->rtt_limit_mul = optval;
24210 		} else {
24211 			rack->rc_pace_fill_if_rttin_range = 0;
24212 			rack->rtt_limit_mul = 0;
24213 		}
24214 		break;
24215 	case TCP_RACK_NO_PUSH_AT_MAX:
24216 		RACK_OPTS_INC(tcp_npush);
24217 		if (optval == 0)
24218 			rack->r_ctl.rc_no_push_at_mrtt = 0;
24219 		else if (optval < 0xff)
24220 			rack->r_ctl.rc_no_push_at_mrtt = optval;
24221 		else
24222 			error = EINVAL;
24223 		break;
24224 	case TCP_SHARED_CWND_ENABLE:
24225 		RACK_OPTS_INC(tcp_rack_scwnd);
24226 		if (optval == 0)
24227 			rack->rack_enable_scwnd = 0;
24228 		else
24229 			rack->rack_enable_scwnd = 1;
24230 		break;
24231 	case TCP_RACK_MBUF_QUEUE:
24232 		/* Now do we use the LRO mbuf-queue feature */
24233 		RACK_OPTS_INC(tcp_rack_mbufq);
24234 		if (optval || rack->r_use_cmp_ack)
24235 			rack->r_mbuf_queue = 1;
24236 		else
24237 			rack->r_mbuf_queue = 0;
24238 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
24239 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
24240 		else
24241 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
24242 		break;
24243 	case TCP_RACK_NONRXT_CFG_RATE:
24244 		RACK_OPTS_INC(tcp_rack_cfg_rate);
24245 		if (optval == 0)
24246 			rack->rack_rec_nonrxt_use_cr = 0;
24247 		else
24248 			rack->rack_rec_nonrxt_use_cr = 1;
24249 		break;
24250 	case TCP_NO_PRR:
24251 		RACK_OPTS_INC(tcp_rack_noprr);
24252 		if (optval == 0)
24253 			rack->rack_no_prr = 0;
24254 		else if (optval == 1)
24255 			rack->rack_no_prr = 1;
24256 		else if (optval == 2)
24257 			rack->no_prr_addback = 1;
24258 		else
24259 			error = EINVAL;
24260 		break;
24261 	case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
24262 		if (optval > 0)
24263 			rack->cspr_is_fcc = 1;
24264 		else
24265 			rack->cspr_is_fcc = 0;
24266 		break;
24267 	case TCP_TIMELY_DYN_ADJ:
24268 		RACK_OPTS_INC(tcp_timely_dyn);
24269 		if (optval == 0)
24270 			rack->rc_gp_dyn_mul = 0;
24271 		else {
24272 			rack->rc_gp_dyn_mul = 1;
24273 			if (optval >= 100) {
24274 				/*
24275 				 * If the user sets something 100 or more
24276 				 * its the gp_ca value.
24277 				 */
24278 				rack->r_ctl.rack_per_of_gp_ca  = optval;
24279 			}
24280 		}
24281 		break;
24282 	case TCP_RACK_DO_DETECTION:
24283 		error = EINVAL;
24284 		break;
24285 	case TCP_RACK_TLP_USE:
24286 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
24287 			error = EINVAL;
24288 			break;
24289 		}
24290 		RACK_OPTS_INC(tcp_tlp_use);
24291 		rack->rack_tlp_threshold_use = optval;
24292 		break;
24293 	case TCP_RACK_TLP_REDUCE:
24294 		/* RACK TLP cwnd reduction (bool) */
24295 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
24296 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
24297 		break;
24298 		/*  Pacing related ones */
24299 	case TCP_RACK_PACE_ALWAYS:
24300 		/*
24301 		 * zero is old rack method, 1 is new
24302 		 * method using a pacing rate.
24303 		 */
24304 		RACK_OPTS_INC(tcp_rack_pace_always);
24305 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
24306 			error = EPERM;
24307 			break;
24308 		}
24309 		if (optval > 0) {
24310 			if (rack->rc_always_pace) {
24311 				error = EALREADY;
24312 				break;
24313 			} else if (tcp_can_enable_pacing()) {
24314 				rack->r_ctl.pacing_method |= RACK_REG_PACING;
24315 				rack->rc_always_pace = 1;
24316 				if (rack->rack_hibeta)
24317 					rack_set_cc_pacing(rack);
24318 			}
24319 			else {
24320 				error = ENOSPC;
24321 				break;
24322 			}
24323 		} else {
24324 			if (rack->rc_always_pace == 1) {
24325 				rack_remove_pacing(rack);
24326 			}
24327 		}
24328 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
24329 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
24330 		else
24331 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
24332 		/* A rate may be set irate or other, if so set seg size */
24333 		rack_update_seg(rack);
24334 		break;
24335 	case TCP_BBR_RACK_INIT_RATE:
24336 		RACK_OPTS_INC(tcp_initial_rate);
24337 		val = optval;
24338 		/* Change from kbits per second to bytes per second */
24339 		val *= 1000;
24340 		val /= 8;
24341 		rack->r_ctl.init_rate = val;
24342 		if (rack->rc_always_pace)
24343 			rack_update_seg(rack);
24344 		break;
24345 	case TCP_BBR_IWINTSO:
24346 		error = EINVAL;
24347 		break;
24348 	case TCP_RACK_FORCE_MSEG:
24349 		RACK_OPTS_INC(tcp_rack_force_max_seg);
24350 		if (optval)
24351 			rack->rc_force_max_seg = 1;
24352 		else
24353 			rack->rc_force_max_seg = 0;
24354 		break;
24355 	case TCP_RACK_PACE_MIN_SEG:
24356 		RACK_OPTS_INC(tcp_rack_min_seg);
24357 		rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
24358 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
24359 		break;
24360 	case TCP_RACK_PACE_MAX_SEG:
24361 		/* Max segments size in a pace in bytes */
24362 		RACK_OPTS_INC(tcp_rack_max_seg);
24363 		if ((rack->dgp_on == 1) &&
24364 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
24365 			/*
24366 			 * If we set a max-seg and are doing DGP then
24367 			 * we now fall under the pacing limits not the
24368 			 * DGP ones.
24369 			 */
24370 			if (tcp_can_enable_pacing() == 0) {
24371 				error = ENOSPC;
24372 				break;
24373 			}
24374 			/*
24375 			 * Now change up the flags and counts to be correct.
24376 			 */
24377 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
24378 			tcp_dec_dgp_pacing_cnt();
24379 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
24380 		}
24381 		if (optval <= MAX_USER_SET_SEG)
24382 			rack->rc_user_set_max_segs = optval;
24383 		else
24384 			rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
24385 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
24386 		break;
24387 	case TCP_RACK_PACE_RATE_REC:
24388 		/* Set the fixed pacing rate in Bytes per second ca */
24389 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
24390 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
24391 			error = EPERM;
24392 			break;
24393 		}
24394 		if (rack->dgp_on) {
24395 			/*
24396 			 * We are already pacing another
24397 			 * way.
24398 			 */
24399 			error = EBUSY;
24400 			break;
24401 		}
24402 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
24403 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
24404 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
24405 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
24406 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
24407 		rack->use_fixed_rate = 1;
24408 		if (rack->rack_hibeta)
24409 			rack_set_cc_pacing(rack);
24410 		rack_log_pacing_delay_calc(rack,
24411 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
24412 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
24413 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
24414 					   __LINE__, NULL,0);
24415 		break;
24416 
24417 	case TCP_RACK_PACE_RATE_SS:
24418 		/* Set the fixed pacing rate in Bytes per second ca */
24419 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
24420 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
24421 			error = EPERM;
24422 			break;
24423 		}
24424 		if (rack->dgp_on) {
24425 			/*
24426 			 * We are already pacing another
24427 			 * way.
24428 			 */
24429 			error = EBUSY;
24430 			break;
24431 		}
24432 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
24433 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
24434 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
24435 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
24436 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
24437 		rack->use_fixed_rate = 1;
24438 		if (rack->rack_hibeta)
24439 			rack_set_cc_pacing(rack);
24440 		rack_log_pacing_delay_calc(rack,
24441 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
24442 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
24443 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
24444 					   __LINE__, NULL, 0);
24445 		break;
24446 
24447 	case TCP_RACK_PACE_RATE_CA:
24448 		/* Set the fixed pacing rate in Bytes per second ca */
24449 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
24450 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
24451 			error = EPERM;
24452 			break;
24453 		}
24454 		if (rack->dgp_on) {
24455 			/*
24456 			 * We are already pacing another
24457 			 * way.
24458 			 */
24459 			error = EBUSY;
24460 			break;
24461 		}
24462 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
24463 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
24464 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
24465 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
24466 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
24467 		rack->use_fixed_rate = 1;
24468 		if (rack->rack_hibeta)
24469 			rack_set_cc_pacing(rack);
24470 		rack_log_pacing_delay_calc(rack,
24471 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
24472 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
24473 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
24474 					   __LINE__, NULL, 0);
24475 		break;
24476 	case TCP_RACK_GP_INCREASE_REC:
24477 		RACK_OPTS_INC(tcp_gp_inc_rec);
24478 		rack->r_ctl.rack_per_of_gp_rec = optval;
24479 		rack_log_pacing_delay_calc(rack,
24480 					   rack->r_ctl.rack_per_of_gp_ss,
24481 					   rack->r_ctl.rack_per_of_gp_ca,
24482 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
24483 					   __LINE__, NULL, 0);
24484 		break;
24485 	case TCP_RACK_GP_INCREASE_CA:
24486 		RACK_OPTS_INC(tcp_gp_inc_ca);
24487 		ca = optval;
24488 		if (ca < 100) {
24489 			/*
24490 			 * We don't allow any reduction
24491 			 * over the GP b/w.
24492 			 */
24493 			error = EINVAL;
24494 			break;
24495 		}
24496 		rack->r_ctl.rack_per_of_gp_ca = ca;
24497 		rack_log_pacing_delay_calc(rack,
24498 					   rack->r_ctl.rack_per_of_gp_ss,
24499 					   rack->r_ctl.rack_per_of_gp_ca,
24500 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
24501 					   __LINE__, NULL, 0);
24502 		break;
24503 	case TCP_RACK_GP_INCREASE_SS:
24504 		RACK_OPTS_INC(tcp_gp_inc_ss);
24505 		ss = optval;
24506 		if (ss < 100) {
24507 			/*
24508 			 * We don't allow any reduction
24509 			 * over the GP b/w.
24510 			 */
24511 			error = EINVAL;
24512 			break;
24513 		}
24514 		rack->r_ctl.rack_per_of_gp_ss = ss;
24515 		rack_log_pacing_delay_calc(rack,
24516 					   rack->r_ctl.rack_per_of_gp_ss,
24517 					   rack->r_ctl.rack_per_of_gp_ca,
24518 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
24519 					   __LINE__, NULL, 0);
24520 		break;
24521 	case TCP_RACK_RR_CONF:
24522 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
24523 		if (optval && optval <= 3)
24524 			rack->r_rr_config = optval;
24525 		else
24526 			rack->r_rr_config = 0;
24527 		break;
24528 	case TCP_PACING_DND:			/*  URL:dnd */
24529 		if (optval > 0)
24530 			rack->rc_pace_dnd = 1;
24531 		else
24532 			rack->rc_pace_dnd = 0;
24533 		break;
24534 	case TCP_HDWR_RATE_CAP:
24535 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
24536 		if (optval) {
24537 			if (rack->r_rack_hw_rate_caps == 0)
24538 				rack->r_rack_hw_rate_caps = 1;
24539 			else
24540 				error = EALREADY;
24541 		} else {
24542 			rack->r_rack_hw_rate_caps = 0;
24543 		}
24544 		break;
24545 	case TCP_DGP_UPPER_BOUNDS:
24546 	{
24547 		uint8_t val;
24548 		val = optval & 0x0000ff;
24549 		rack->r_ctl.rack_per_upper_bound_ca = val;
24550 		val = (optval >> 16) & 0x0000ff;
24551 		rack->r_ctl.rack_per_upper_bound_ss = val;
24552 		break;
24553 	}
24554 	case TCP_SS_EEXIT:			/*  URL:eexit */
24555 		if (optval > 0) {
24556 			rack->r_ctl.gp_rnd_thresh =  optval & 0x0ff;
24557 			if (optval & 0x10000) {
24558 				rack->r_ctl.gate_to_fs = 1;
24559 			} else {
24560 				rack->r_ctl.gate_to_fs = 0;
24561 			}
24562 			if (optval & 0x20000) {
24563 				rack->r_ctl.use_gp_not_last = 1;
24564 			} else {
24565 				rack->r_ctl.use_gp_not_last = 0;
24566 			}
24567 			if (optval & 0xfffc0000) {
24568 				uint32_t v;
24569 
24570 				v = (optval >> 18) & 0x00003fff;
24571 				if (v >= 1000)
24572 					rack->r_ctl.gp_gain_req = v;
24573 			}
24574 		} else {
24575 			/* We do not do ss early exit at all */
24576 			rack->rc_initial_ss_comp = 1;
24577 			rack->r_ctl.gp_rnd_thresh = 0;
24578 		}
24579 		break;
24580 	case TCP_RACK_SPLIT_LIMIT:
24581 		RACK_OPTS_INC(tcp_split_limit);
24582 		rack->r_ctl.rc_split_limit = optval;
24583 		break;
24584 	case TCP_BBR_HDWR_PACE:
24585 		RACK_OPTS_INC(tcp_hdwr_pacing);
24586 		if (optval){
24587 			if (rack->rack_hdrw_pacing == 0) {
24588 				rack->rack_hdw_pace_ena = 1;
24589 				rack->rack_attempt_hdwr_pace = 0;
24590 			} else
24591 				error = EALREADY;
24592 		} else {
24593 			rack->rack_hdw_pace_ena = 0;
24594 #ifdef RATELIMIT
24595 			if (rack->r_ctl.crte != NULL) {
24596 				rack->rack_hdrw_pacing = 0;
24597 				rack->rack_attempt_hdwr_pace = 0;
24598 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
24599 				rack->r_ctl.crte = NULL;
24600 			}
24601 #endif
24602 		}
24603 		break;
24604 		/*  End Pacing related ones */
24605 	case TCP_RACK_PRR_SENDALOT:
24606 		/* Allow PRR to send more than one seg */
24607 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
24608 		rack->r_ctl.rc_prr_sendalot = optval;
24609 		break;
24610 	case TCP_RACK_MIN_TO:
24611 		/* Minimum time between rack t-o's in ms */
24612 		RACK_OPTS_INC(tcp_rack_min_to);
24613 		rack->r_ctl.rc_min_to = optval;
24614 		break;
24615 	case TCP_RACK_EARLY_SEG:
24616 		/* If early recovery max segments */
24617 		RACK_OPTS_INC(tcp_rack_early_seg);
24618 		rack->r_ctl.rc_early_recovery_segs = optval;
24619 		break;
24620 	case TCP_RACK_ENABLE_HYSTART:
24621 	{
24622 		if (optval) {
24623 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
24624 			if (rack_do_hystart > RACK_HYSTART_ON)
24625 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
24626 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
24627 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
24628 		} else {
24629 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
24630 		}
24631 	}
24632 	break;
24633 	case TCP_RACK_REORD_THRESH:
24634 		/* RACK reorder threshold (shift amount) */
24635 		RACK_OPTS_INC(tcp_rack_reord_thresh);
24636 		if ((optval > 0) && (optval < 31))
24637 			rack->r_ctl.rc_reorder_shift = optval;
24638 		else
24639 			error = EINVAL;
24640 		break;
24641 	case TCP_RACK_REORD_FADE:
24642 		/* Does reordering fade after ms time */
24643 		RACK_OPTS_INC(tcp_rack_reord_fade);
24644 		rack->r_ctl.rc_reorder_fade = optval;
24645 		break;
24646 	case TCP_RACK_TLP_THRESH:
24647 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
24648 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
24649 		if (optval)
24650 			rack->r_ctl.rc_tlp_threshold = optval;
24651 		else
24652 			error = EINVAL;
24653 		break;
24654 	case TCP_BBR_USE_RACK_RR:
24655 		RACK_OPTS_INC(tcp_rack_rr);
24656 		if (optval)
24657 			rack->use_rack_rr = 1;
24658 		else
24659 			rack->use_rack_rr = 0;
24660 		break;
24661 	case TCP_RACK_PKT_DELAY:
24662 		/* RACK added ms i.e. rack-rtt + reord + N */
24663 		RACK_OPTS_INC(tcp_rack_pkt_delay);
24664 		rack->r_ctl.rc_pkt_delay = optval;
24665 		break;
24666 	case TCP_DELACK:
24667 		RACK_OPTS_INC(tcp_rack_delayed_ack);
24668 		if (optval == 0)
24669 			tp->t_delayed_ack = 0;
24670 		else
24671 			tp->t_delayed_ack = 1;
24672 		if (tp->t_flags & TF_DELACK) {
24673 			tp->t_flags &= ~TF_DELACK;
24674 			tp->t_flags |= TF_ACKNOW;
24675 			NET_EPOCH_ENTER(et);
24676 			rack_output(tp);
24677 			NET_EPOCH_EXIT(et);
24678 		}
24679 		break;
24680 
24681 	case TCP_BBR_RACK_RTT_USE:
24682 		RACK_OPTS_INC(tcp_rack_rtt_use);
24683 		if ((optval != USE_RTT_HIGH) &&
24684 		    (optval != USE_RTT_LOW) &&
24685 		    (optval != USE_RTT_AVG))
24686 			error = EINVAL;
24687 		else
24688 			rack->r_ctl.rc_rate_sample_method = optval;
24689 		break;
24690 	case TCP_HONOR_HPTS_MIN:
24691 		RACK_OPTS_INC(tcp_honor_hpts);
24692 		if (optval) {
24693 			rack->r_use_hpts_min = 1;
24694 			/*
24695 			 * Must be between 2 - 80% to be a reduction else
24696 			 * we keep the default (10%).
24697 			 */
24698 			if ((optval > 1) && (optval <= 80)) {
24699 				rack->r_ctl.max_reduction = optval;
24700 			}
24701 		} else
24702 			rack->r_use_hpts_min = 0;
24703 		break;
24704 	case TCP_REC_IS_DYN:			/*  URL:dynrec */
24705 		RACK_OPTS_INC(tcp_dyn_rec);
24706 		if (optval)
24707 			rack->rc_gp_no_rec_chg = 1;
24708 		else
24709 			rack->rc_gp_no_rec_chg = 0;
24710 		break;
24711 	case TCP_NO_TIMELY:
24712 		RACK_OPTS_INC(tcp_notimely);
24713 		if (optval) {
24714 			rack->rc_skip_timely = 1;
24715 			rack->r_ctl.rack_per_of_gp_rec = 90;
24716 			rack->r_ctl.rack_per_of_gp_ca = 100;
24717 			rack->r_ctl.rack_per_of_gp_ss = 250;
24718 		} else {
24719 			rack->rc_skip_timely = 0;
24720 		}
24721 		break;
24722 	case TCP_GP_USE_LTBW:
24723 		if (optval == 0) {
24724 			rack->use_lesser_lt_bw = 0;
24725 			rack->dis_lt_bw = 1;
24726 		} else if (optval == 1) {
24727 			rack->use_lesser_lt_bw = 1;
24728 			rack->dis_lt_bw = 0;
24729 		} else if (optval == 2) {
24730 			rack->use_lesser_lt_bw = 0;
24731 			rack->dis_lt_bw = 0;
24732 		}
24733 		break;
24734 	case TCP_DATA_AFTER_CLOSE:
24735 		RACK_OPTS_INC(tcp_data_after_close);
24736 		if (optval)
24737 			rack->rc_allow_data_af_clo = 1;
24738 		else
24739 			rack->rc_allow_data_af_clo = 0;
24740 		break;
24741 	default:
24742 		break;
24743 	}
24744 	tcp_log_socket_option(tp, sopt_name, optval, error);
24745 	return (error);
24746 }
24747 
24748 static void
24749 rack_inherit(struct tcpcb *tp, struct inpcb *parent)
24750 {
24751 	/*
24752 	 * A new connection has been created (tp) and
24753 	 * the parent is the inpcb given. We want to
24754 	 * apply a read-lock to the parent (we are already
24755 	 * holding a write lock on the tp) and copy anything
24756 	 * out of the rack specific data as long as its tfb is
24757 	 * the same as ours i.e. we are the same stack. Otherwise
24758 	 * we just return.
24759 	 */
24760 	struct tcpcb *par;
24761 	struct tcp_rack *dest, *src;
24762 	int cnt = 0;
24763 
24764 	par = intotcpcb(parent);
24765 	if (par->t_fb != tp->t_fb) {
24766 		/* Not the same stack */
24767 		tcp_log_socket_option(tp, 0, 0, 1);
24768 		return;
24769 	}
24770 	/* Ok if we reach here lets setup the two rack pointers */
24771 	dest = (struct tcp_rack *)tp->t_fb_ptr;
24772 	src = (struct tcp_rack *)par->t_fb_ptr;
24773 	if ((src == NULL) || (dest == NULL)) {
24774 		/* Huh? */
24775 		tcp_log_socket_option(tp, 0, 0, 2);
24776 		return;
24777 	}
24778 	/* Now copy out anything we wish to inherit i.e. things in socket-options */
24779 	/* TCP_RACK_PROFILE we can't know but we can set DGP if its on */
24780 	if ((src->dgp_on) && (dest->dgp_on == 0)) {
24781 		/* Profile 1 had to be set via sock opt */
24782 		rack_set_dgp(dest);
24783 		cnt++;
24784 	}
24785 	/* TCP_RACK_SET_RXT_OPTIONS */
24786 	if (dest->full_size_rxt != src->full_size_rxt) {
24787 		dest->full_size_rxt = src->full_size_rxt;
24788 		cnt++;
24789 	}
24790 	if (dest->shape_rxt_to_pacing_min  != src->shape_rxt_to_pacing_min) {
24791 		dest->shape_rxt_to_pacing_min = src->shape_rxt_to_pacing_min;
24792 		cnt++;
24793 	}
24794 	/* TCP_RACK_DSACK_OPT */
24795 	if (dest->rc_rack_tmr_std_based != src->rc_rack_tmr_std_based) {
24796 		dest->rc_rack_tmr_std_based = src->rc_rack_tmr_std_based;
24797 		cnt++;
24798 	}
24799 	if (dest->rc_rack_use_dsack != src->rc_rack_use_dsack) {
24800 		dest->rc_rack_use_dsack = src->rc_rack_use_dsack;
24801 		cnt++;
24802 	}
24803 	/* TCP_RACK_PACING_DIVISOR */
24804 	if (dest->r_ctl.pace_len_divisor != src->r_ctl.pace_len_divisor) {
24805 		dest->r_ctl.pace_len_divisor = src->r_ctl.pace_len_divisor;
24806 		cnt++;
24807 	}
24808 	/* TCP_RACK_HI_BETA */
24809 	if (src->rack_hibeta != dest->rack_hibeta) {
24810 		cnt++;
24811 		if (src->rack_hibeta) {
24812 			dest->r_ctl.rc_saved_beta.beta = src->r_ctl.rc_saved_beta.beta;
24813 			dest->rack_hibeta = 1;
24814 		} else {
24815 			dest->rack_hibeta = 0;
24816 		}
24817 	}
24818 	/* TCP_RACK_TIMER_SLOP */
24819 	if (dest->r_ctl.timer_slop != src->r_ctl.timer_slop) {
24820 		dest->r_ctl.timer_slop = src->r_ctl.timer_slop;
24821 		cnt++;
24822 	}
24823 	/* TCP_RACK_PACING_BETA_ECN */
24824 	if (dest->r_ctl.rc_saved_beta.beta_ecn != src->r_ctl.rc_saved_beta.beta_ecn) {
24825 		dest->r_ctl.rc_saved_beta.beta_ecn = src->r_ctl.rc_saved_beta.beta_ecn;
24826 		cnt++;
24827 	}
24828 	if (dest->r_ctl.rc_saved_beta.newreno_flags != src->r_ctl.rc_saved_beta.newreno_flags) {
24829 		dest->r_ctl.rc_saved_beta.newreno_flags = src->r_ctl.rc_saved_beta.newreno_flags;
24830 		cnt++;
24831 	}
24832 	/* We do not do TCP_DEFER_OPTIONS */
24833 	/* TCP_RACK_MEASURE_CNT */
24834 	if (dest->r_ctl.req_measurements != src->r_ctl.req_measurements) {
24835 		dest->r_ctl.req_measurements = src->r_ctl.req_measurements;
24836 		cnt++;
24837 	}
24838 	/* TCP_HDWR_UP_ONLY */
24839 	if (dest->r_up_only != src->r_up_only) {
24840 		dest->r_up_only = src->r_up_only;
24841 		cnt++;
24842 	}
24843 	/* TCP_FILLCW_RATE_CAP */
24844 	if (dest->r_ctl.fillcw_cap != src->r_ctl.fillcw_cap) {
24845 		dest->r_ctl.fillcw_cap = src->r_ctl.fillcw_cap;
24846 		cnt++;
24847 	}
24848 	/* TCP_PACING_RATE_CAP */
24849 	if (dest->r_ctl.bw_rate_cap != src->r_ctl.bw_rate_cap) {
24850 		dest->r_ctl.bw_rate_cap = src->r_ctl.bw_rate_cap;
24851 		cnt++;
24852 	}
24853 	/* A listener can't set TCP_HYBRID_PACING */
24854 	/* TCP_SIDECHAN_DIS */
24855 	if (dest->r_ctl.side_chan_dis_mask != src->r_ctl.side_chan_dis_mask) {
24856 		dest->r_ctl.side_chan_dis_mask = src->r_ctl.side_chan_dis_mask;
24857 		cnt++;
24858 	}
24859 	/* TCP_SHARED_CWND_TIME_LIMIT */
24860 	if (dest->r_limit_scw != src->r_limit_scw) {
24861 		dest->r_limit_scw = src->r_limit_scw;
24862 		cnt++;
24863 	}
24864 	/* TCP_POLICER_DETECT */
24865 	if (dest->r_ctl.policer_rxt_threshold != src->r_ctl.policer_rxt_threshold) {
24866 		dest->r_ctl.policer_rxt_threshold = src->r_ctl.policer_rxt_threshold;
24867 		cnt++;
24868 	}
24869 	if (dest->r_ctl.policer_avg_threshold != src->r_ctl.policer_avg_threshold) {
24870 		dest->r_ctl.policer_avg_threshold = src->r_ctl.policer_avg_threshold;
24871 		cnt++;
24872 	}
24873 	if (dest->r_ctl.policer_med_threshold != src->r_ctl.policer_med_threshold) {
24874 		dest->r_ctl.policer_med_threshold = src->r_ctl.policer_med_threshold;
24875 		cnt++;
24876 	}
24877 	if (dest->policer_detect_on != src->policer_detect_on) {
24878 		dest->policer_detect_on = src->policer_detect_on;
24879 		cnt++;
24880 	}
24881 
24882 	if (dest->r_ctl.saved_policer_val != src->r_ctl.saved_policer_val) {
24883 		dest->r_ctl.saved_policer_val = src->r_ctl.saved_policer_val;
24884 		cnt++;
24885 	}
24886 	/* TCP_POLICER_MSS */
24887 	if (dest->r_ctl.policer_del_mss != src->r_ctl.policer_del_mss) {
24888 		dest->r_ctl.policer_del_mss = src->r_ctl.policer_del_mss;
24889 		cnt++;
24890 	}
24891 
24892 	if (dest->r_ctl.pol_bw_comp != src->r_ctl.pol_bw_comp) {
24893 		dest->r_ctl.pol_bw_comp = src->r_ctl.pol_bw_comp;
24894 		cnt++;
24895 	}
24896 
24897 	if (dest->r_ctl.policer_alt_median != src->r_ctl.policer_alt_median) {
24898 		dest->r_ctl.policer_alt_median = src->r_ctl.policer_alt_median;
24899 		cnt++;
24900 	}
24901 	/* TCP_RACK_PACE_TO_FILL */
24902 	if (dest->rc_pace_to_cwnd != src->rc_pace_to_cwnd) {
24903 		dest->rc_pace_to_cwnd = src->rc_pace_to_cwnd;
24904 		cnt++;
24905 	}
24906 	if (dest->rc_pace_fill_if_rttin_range != src->rc_pace_fill_if_rttin_range) {
24907 		dest->rc_pace_fill_if_rttin_range = src->rc_pace_fill_if_rttin_range;
24908 		cnt++;
24909 	}
24910 	if (dest->rtt_limit_mul != src->rtt_limit_mul) {
24911 		dest->rtt_limit_mul = src->rtt_limit_mul;
24912 		cnt++;
24913 	}
24914 	/* TCP_RACK_NO_PUSH_AT_MAX */
24915 	if (dest->r_ctl.rc_no_push_at_mrtt != src->r_ctl.rc_no_push_at_mrtt) {
24916 		dest->r_ctl.rc_no_push_at_mrtt = src->r_ctl.rc_no_push_at_mrtt;
24917 		cnt++;
24918 	}
24919 	/* TCP_SHARED_CWND_ENABLE */
24920 	if (dest->rack_enable_scwnd != src->rack_enable_scwnd) {
24921 		dest->rack_enable_scwnd = src->rack_enable_scwnd;
24922 		cnt++;
24923 	}
24924 	/* TCP_USE_CMP_ACKS */
24925 	if (dest->r_use_cmp_ack != src->r_use_cmp_ack) {
24926 		dest->r_use_cmp_ack = src->r_use_cmp_ack;
24927 		cnt++;
24928 	}
24929 
24930 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
24931 		dest->r_mbuf_queue = src->r_mbuf_queue;
24932 		cnt++;
24933 	}
24934 	/* TCP_RACK_MBUF_QUEUE */
24935 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
24936 		dest->r_mbuf_queue = src->r_mbuf_queue;
24937 		cnt++;
24938 	}
24939 	if  (dest->r_mbuf_queue || dest->rc_always_pace || dest->r_use_cmp_ack) {
24940 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
24941 	} else {
24942 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
24943 	}
24944 	if (dest->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) {
24945 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
24946 	}
24947 	/* TCP_RACK_NONRXT_CFG_RATE */
24948 	if (dest->rack_rec_nonrxt_use_cr != src->rack_rec_nonrxt_use_cr) {
24949 		dest->rack_rec_nonrxt_use_cr = src->rack_rec_nonrxt_use_cr;
24950 		cnt++;
24951 	}
24952 	/* TCP_NO_PRR */
24953 	if (dest->rack_no_prr != src->rack_no_prr) {
24954 		dest->rack_no_prr = src->rack_no_prr;
24955 		cnt++;
24956 	}
24957 	if (dest->no_prr_addback != src->no_prr_addback) {
24958 		dest->no_prr_addback = src->no_prr_addback;
24959 		cnt++;
24960 	}
24961 	/* RACK_CSPR_IS_FCC */
24962 	if (dest->cspr_is_fcc != src->cspr_is_fcc) {
24963 		dest->cspr_is_fcc = src->cspr_is_fcc;
24964 		cnt++;
24965 	}
24966 	/* TCP_TIMELY_DYN_ADJ */
24967 	if (dest->rc_gp_dyn_mul != src->rc_gp_dyn_mul) {
24968 		dest->rc_gp_dyn_mul = src->rc_gp_dyn_mul;
24969 		cnt++;
24970 	}
24971 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
24972 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
24973 		cnt++;
24974 	}
24975 	/* TCP_RACK_TLP_USE */
24976 	if (dest->rack_tlp_threshold_use != src->rack_tlp_threshold_use) {
24977 		dest->rack_tlp_threshold_use = src->rack_tlp_threshold_use;
24978 		cnt++;
24979 	}
24980 	/* we don't allow inheritence of TCP_RACK_PACE_ALWAYS */
24981 	/* TCP_BBR_RACK_INIT_RATE */
24982 	if (dest->r_ctl.init_rate != src->r_ctl.init_rate) {
24983 		dest->r_ctl.init_rate = src->r_ctl.init_rate;
24984 		cnt++;
24985 	}
24986 	/* TCP_RACK_FORCE_MSEG */
24987 	if (dest->rc_force_max_seg != src->rc_force_max_seg) {
24988 		dest->rc_force_max_seg = src->rc_force_max_seg;
24989 		cnt++;
24990 	}
24991 	/* TCP_RACK_PACE_MIN_SEG */
24992 	if (dest->r_ctl.rc_user_set_min_segs != src->r_ctl.rc_user_set_min_segs) {
24993 		dest->r_ctl.rc_user_set_min_segs = src->r_ctl.rc_user_set_min_segs;
24994 		cnt++;
24995 	}
24996 	/* we don't allow TCP_RACK_PACE_MAX_SEG */
24997 	/* TCP_RACK_PACE_RATE_REC, TCP_RACK_PACE_RATE_SS,  TCP_RACK_PACE_RATE_CA */
24998 	if (dest->r_ctl.rc_fixed_pacing_rate_ca != src->r_ctl.rc_fixed_pacing_rate_ca) {
24999 		dest->r_ctl.rc_fixed_pacing_rate_ca = src->r_ctl.rc_fixed_pacing_rate_ca;
25000 		cnt++;
25001 	}
25002 	if (dest->r_ctl.rc_fixed_pacing_rate_ss != src->r_ctl.rc_fixed_pacing_rate_ss) {
25003 		dest->r_ctl.rc_fixed_pacing_rate_ss = src->r_ctl.rc_fixed_pacing_rate_ss;
25004 		cnt++;
25005 	}
25006 	if (dest->r_ctl.rc_fixed_pacing_rate_rec != src->r_ctl.rc_fixed_pacing_rate_rec) {
25007 		dest->r_ctl.rc_fixed_pacing_rate_rec = src->r_ctl.rc_fixed_pacing_rate_rec;
25008 		cnt++;
25009 	}
25010 	/* TCP_RACK_GP_INCREASE_REC, TCP_RACK_GP_INCREASE_CA, TCP_RACK_GP_INCREASE_SS */
25011 	if (dest->r_ctl.rack_per_of_gp_rec != src->r_ctl.rack_per_of_gp_rec) {
25012 		dest->r_ctl.rack_per_of_gp_rec = src->r_ctl.rack_per_of_gp_rec;
25013 		cnt++;
25014 	}
25015 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
25016 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
25017 		cnt++;
25018 	}
25019 
25020 	if (dest->r_ctl.rack_per_of_gp_ss != src->r_ctl.rack_per_of_gp_ss) {
25021 		dest->r_ctl.rack_per_of_gp_ss = src->r_ctl.rack_per_of_gp_ss;
25022 		cnt++;
25023 	}
25024 	/* TCP_RACK_RR_CONF */
25025 	if (dest->r_rr_config != src->r_rr_config) {
25026 		dest->r_rr_config = src->r_rr_config;
25027 		cnt++;
25028 	}
25029 	/* TCP_PACING_DND */
25030 	if (dest->rc_pace_dnd != src->rc_pace_dnd) {
25031 		dest->rc_pace_dnd = src->rc_pace_dnd;
25032 		cnt++;
25033 	}
25034 	/* TCP_HDWR_RATE_CAP */
25035 	if (dest->r_rack_hw_rate_caps != src->r_rack_hw_rate_caps) {
25036 		dest->r_rack_hw_rate_caps = src->r_rack_hw_rate_caps;
25037 		cnt++;
25038 	}
25039 	/* TCP_DGP_UPPER_BOUNDS */
25040 	if (dest->r_ctl.rack_per_upper_bound_ca != src->r_ctl.rack_per_upper_bound_ca) {
25041 		dest->r_ctl.rack_per_upper_bound_ca = src->r_ctl.rack_per_upper_bound_ca;
25042 		cnt++;
25043 	}
25044 	if (dest->r_ctl.rack_per_upper_bound_ss != src->r_ctl.rack_per_upper_bound_ss) {
25045 		dest->r_ctl.rack_per_upper_bound_ss = src->r_ctl.rack_per_upper_bound_ss;
25046 		cnt++;
25047 	}
25048 	/* TCP_SS_EEXIT */
25049 	if (dest->r_ctl.gp_rnd_thresh != src->r_ctl.gp_rnd_thresh) {
25050 		dest->r_ctl.gp_rnd_thresh = src->r_ctl.gp_rnd_thresh;
25051 		cnt++;
25052 	}
25053 	if (dest->r_ctl.gate_to_fs != src->r_ctl.gate_to_fs) {
25054 		dest->r_ctl.gate_to_fs = src->r_ctl.gate_to_fs;
25055 		cnt++;
25056 	}
25057 	if (dest->r_ctl.use_gp_not_last != src->r_ctl.use_gp_not_last) {
25058 		dest->r_ctl.use_gp_not_last = src->r_ctl.use_gp_not_last;
25059 		cnt++;
25060 	}
25061 	if (dest->r_ctl.gp_gain_req != src->r_ctl.gp_gain_req) {
25062 		dest->r_ctl.gp_gain_req = src->r_ctl.gp_gain_req;
25063 		cnt++;
25064 	}
25065 	/* TCP_BBR_HDWR_PACE */
25066 	if (dest->rack_hdw_pace_ena != src->rack_hdw_pace_ena) {
25067 		dest->rack_hdw_pace_ena = src->rack_hdw_pace_ena;
25068 		cnt++;
25069 	}
25070 	if (dest->rack_attempt_hdwr_pace != src->rack_attempt_hdwr_pace) {
25071 		dest->rack_attempt_hdwr_pace = src->rack_attempt_hdwr_pace;
25072 		cnt++;
25073 	}
25074 	/* TCP_RACK_PRR_SENDALOT */
25075 	if (dest->r_ctl.rc_prr_sendalot != src->r_ctl.rc_prr_sendalot) {
25076 		dest->r_ctl.rc_prr_sendalot = src->r_ctl.rc_prr_sendalot;
25077 		cnt++;
25078 	}
25079 	/* TCP_RACK_MIN_TO */
25080 	if (dest->r_ctl.rc_min_to != src->r_ctl.rc_min_to) {
25081 		dest->r_ctl.rc_min_to = src->r_ctl.rc_min_to;
25082 		cnt++;
25083 	}
25084 	/* TCP_RACK_EARLY_SEG */
25085 	if (dest->r_ctl.rc_early_recovery_segs != src->r_ctl.rc_early_recovery_segs) {
25086 		dest->r_ctl.rc_early_recovery_segs = src->r_ctl.rc_early_recovery_segs;
25087 		cnt++;
25088 	}
25089 	/* TCP_RACK_ENABLE_HYSTART */
25090 	if (par->t_ccv.flags != tp->t_ccv.flags) {
25091 		cnt++;
25092 		if (par->t_ccv.flags & CCF_HYSTART_ALLOWED) {
25093 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
25094 			if (rack_do_hystart > RACK_HYSTART_ON)
25095 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
25096 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
25097 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
25098 		} else {
25099 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
25100 		}
25101 	}
25102 	/* TCP_RACK_REORD_THRESH */
25103 	if (dest->r_ctl.rc_reorder_shift != src->r_ctl.rc_reorder_shift) {
25104 		dest->r_ctl.rc_reorder_shift = src->r_ctl.rc_reorder_shift;
25105 		cnt++;
25106 	}
25107 	/* TCP_RACK_REORD_FADE */
25108 	if (dest->r_ctl.rc_reorder_fade != src->r_ctl.rc_reorder_fade) {
25109 		dest->r_ctl.rc_reorder_fade = src->r_ctl.rc_reorder_fade;
25110 		cnt++;
25111 	}
25112 	/* TCP_RACK_TLP_THRESH */
25113 	if (dest->r_ctl.rc_tlp_threshold != src->r_ctl.rc_tlp_threshold) {
25114 		dest->r_ctl.rc_tlp_threshold = src->r_ctl.rc_tlp_threshold;
25115 		cnt++;
25116 	}
25117 	/* TCP_BBR_USE_RACK_RR */
25118 	if (dest->use_rack_rr != src->use_rack_rr) {
25119 		dest->use_rack_rr = src->use_rack_rr;
25120 		cnt++;
25121 	}
25122 	/* TCP_RACK_PKT_DELAY */
25123 	if (dest->r_ctl.rc_pkt_delay != src->r_ctl.rc_pkt_delay) {
25124 		dest->r_ctl.rc_pkt_delay = src->r_ctl.rc_pkt_delay;
25125 		cnt++;
25126 	}
25127 	/* TCP_DELACK will get copied via the main code if applicable */
25128 	/* TCP_BBR_RACK_RTT_USE */
25129 	if (dest->r_ctl.rc_rate_sample_method != src->r_ctl.rc_rate_sample_method) {
25130 		dest->r_ctl.rc_rate_sample_method = src->r_ctl.rc_rate_sample_method;
25131 		cnt++;
25132 	}
25133 	/* TCP_HONOR_HPTS_MIN */
25134 	if (dest->r_use_hpts_min != src->r_use_hpts_min) {
25135 		dest->r_use_hpts_min = src->r_use_hpts_min;
25136 		cnt++;
25137 	}
25138 	if (dest->r_ctl.max_reduction != src->r_ctl.max_reduction) {
25139 		dest->r_ctl.max_reduction = src->r_ctl.max_reduction;
25140 		cnt++;
25141 	}
25142 	/* TCP_REC_IS_DYN */
25143 	if (dest->rc_gp_no_rec_chg != src->rc_gp_no_rec_chg) {
25144 		dest->rc_gp_no_rec_chg = src->rc_gp_no_rec_chg;
25145 		cnt++;
25146 	}
25147 	if (dest->rc_skip_timely != src->rc_skip_timely) {
25148 		dest->rc_skip_timely = src->rc_skip_timely;
25149 		cnt++;
25150 	}
25151 	/* TCP_DATA_AFTER_CLOSE */
25152 	if (dest->rc_allow_data_af_clo != src->rc_allow_data_af_clo) {
25153 		dest->rc_allow_data_af_clo = src->rc_allow_data_af_clo;
25154 		cnt++;
25155 	}
25156 	/* TCP_GP_USE_LTBW */
25157 	if (src->use_lesser_lt_bw != dest->use_lesser_lt_bw) {
25158 		dest->use_lesser_lt_bw = src->use_lesser_lt_bw;
25159 		cnt++;
25160 	}
25161 	if (dest->dis_lt_bw != src->dis_lt_bw) {
25162 		dest->dis_lt_bw = src->dis_lt_bw;
25163 		cnt++;
25164 	}
25165 	tcp_log_socket_option(tp, 0, cnt, 0);
25166 }
25167 
25168 
25169 static void
25170 rack_apply_deferred_options(struct tcp_rack *rack)
25171 {
25172 	struct deferred_opt_list *dol, *sdol;
25173 	uint32_t s_optval;
25174 
25175 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
25176 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
25177 		/* Disadvantage of deferal is you loose the error return */
25178 		s_optval = (uint32_t)dol->optval;
25179 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
25180 		free(dol, M_TCPDO);
25181 	}
25182 }
25183 
25184 static void
25185 rack_hw_tls_change(struct tcpcb *tp, int chg)
25186 {
25187 	/* Update HW tls state */
25188 	struct tcp_rack *rack;
25189 
25190 	rack = (struct tcp_rack *)tp->t_fb_ptr;
25191 	if (chg)
25192 		rack->r_ctl.fsb.hw_tls = 1;
25193 	else
25194 		rack->r_ctl.fsb.hw_tls = 0;
25195 }
25196 
25197 static int
25198 rack_pru_options(struct tcpcb *tp, int flags)
25199 {
25200 	if (flags & PRUS_OOB)
25201 		return (EOPNOTSUPP);
25202 	return (0);
25203 }
25204 
25205 static bool
25206 rack_wake_check(struct tcpcb *tp)
25207 {
25208 	struct tcp_rack *rack;
25209 	struct timeval tv;
25210 	uint32_t cts;
25211 
25212 	rack = (struct tcp_rack *)tp->t_fb_ptr;
25213 	if (rack->r_ctl.rc_hpts_flags) {
25214 		cts = tcp_get_usecs(&tv);
25215 		if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
25216 			/*
25217 			 * Pacing timer is up, check if we are ready.
25218 			 */
25219 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
25220 				return (true);
25221 		} else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
25222 			/*
25223 			 * A timer is up, check if we are ready.
25224 			 */
25225 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
25226 				return (true);
25227 		}
25228 	}
25229 	return (false);
25230 }
25231 
25232 static struct tcp_function_block __tcp_rack = {
25233 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
25234 	.tfb_tcp_output = rack_output,
25235 	.tfb_do_queued_segments = ctf_do_queued_segments,
25236 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
25237 	.tfb_tcp_do_segment = rack_do_segment,
25238 	.tfb_tcp_ctloutput = rack_ctloutput,
25239 	.tfb_tcp_fb_init = rack_init,
25240 	.tfb_tcp_fb_fini = rack_fini,
25241 	.tfb_tcp_timer_stop_all = rack_stopall,
25242 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
25243 	.tfb_tcp_handoff_ok = rack_handoff_ok,
25244 	.tfb_tcp_mtu_chg = rack_mtu_change,
25245 	.tfb_pru_options = rack_pru_options,
25246 	.tfb_hwtls_change = rack_hw_tls_change,
25247 	.tfb_chg_query = rack_chg_query,
25248 	.tfb_switch_failed = rack_switch_failed,
25249 	.tfb_early_wake_check = rack_wake_check,
25250 	.tfb_compute_pipe = rack_compute_pipe,
25251 	.tfb_stack_info = rack_stack_information,
25252 	.tfb_inherit = rack_inherit,
25253 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
25254 
25255 };
25256 
25257 /*
25258  * rack_ctloutput() must drop the inpcb lock before performing copyin on
25259  * socket option arguments.  When it re-acquires the lock after the copy, it
25260  * has to revalidate that the connection is still valid for the socket
25261  * option.
25262  */
25263 static int
25264 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
25265 {
25266 	struct inpcb *inp = tptoinpcb(tp);
25267 #ifdef INET
25268 	struct ip *ip;
25269 #endif
25270 	struct tcp_rack *rack;
25271 	struct tcp_hybrid_req hybrid;
25272 	uint64_t loptval;
25273 	int32_t error = 0, optval;
25274 
25275 	rack = (struct tcp_rack *)tp->t_fb_ptr;
25276 	if (rack == NULL) {
25277 		INP_WUNLOCK(inp);
25278 		return (EINVAL);
25279 	}
25280 #ifdef INET
25281 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
25282 #endif
25283 
25284 	switch (sopt->sopt_level) {
25285 #ifdef INET6
25286 	case IPPROTO_IPV6:
25287 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
25288 		switch (sopt->sopt_name) {
25289 		case IPV6_USE_MIN_MTU:
25290 			tcp6_use_min_mtu(tp);
25291 			break;
25292 		}
25293 		INP_WUNLOCK(inp);
25294 		return (0);
25295 #endif
25296 #ifdef INET
25297 	case IPPROTO_IP:
25298 		switch (sopt->sopt_name) {
25299 		case IP_TOS:
25300 			/*
25301 			 * The DSCP codepoint has changed, update the fsb.
25302 			 */
25303 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
25304 			break;
25305 		case IP_TTL:
25306 			/*
25307 			 * The TTL has changed, update the fsb.
25308 			 */
25309 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
25310 			break;
25311 		}
25312 		INP_WUNLOCK(inp);
25313 		return (0);
25314 #endif
25315 #ifdef SO_PEERPRIO
25316 	case SOL_SOCKET:
25317 		switch (sopt->sopt_name) {
25318 		case SO_PEERPRIO:			/*  SC-URL:bs */
25319 			/* Already read in and sanity checked in sosetopt(). */
25320 			if (inp->inp_socket) {
25321 				rack->client_bufferlvl = inp->inp_socket->so_peerprio;
25322 			}
25323 			break;
25324 		}
25325 		INP_WUNLOCK(inp);
25326 		return (0);
25327 #endif
25328 	case IPPROTO_TCP:
25329 		switch (sopt->sopt_name) {
25330 		case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
25331 		/*  Pacing related ones */
25332 		case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
25333 		case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
25334 		case TCP_RACK_PACE_MIN_SEG:		/*  URL:pace_min_seg */
25335 		case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
25336 		case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
25337 		case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
25338 		case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
25339 		case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
25340 		case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
25341 		case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
25342 		case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
25343 		case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
25344 		case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
25345 		case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
25346 		case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
25347 		case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
25348 		case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
25349 		case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
25350 		case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
25351 			/* End pacing related */
25352 		case TCP_POLICER_DETECT:		/*  URL:pol_det */
25353 		case TCP_POLICER_MSS:			/*  URL:pol_mss */
25354 		case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
25355 		case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
25356 		case TCP_RACK_MIN_TO:			/*  URL:min_to */
25357 		case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
25358 		case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
25359 		case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
25360 		case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
25361 		case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
25362 		case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
25363 		case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
25364 		case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
25365 		case TCP_NO_PRR:			/*  URL:noprr */
25366 		case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
25367 		case TCP_DATA_AFTER_CLOSE:		/*  no URL */
25368 		case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
25369 		case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
25370 		case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
25371 		case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
25372 		case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
25373 		case TCP_RACK_PROFILE:			/*  URL:profile */
25374 		case TCP_SIDECHAN_DIS:			/*  URL:scodm */
25375 		case TCP_HYBRID_PACING:			/*  URL:pacing=hybrid */
25376 		case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
25377 		case TCP_RACK_ABC_VAL:			/*  URL:labc */
25378 		case TCP_REC_ABC_VAL:			/*  URL:reclabc */
25379 		case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
25380 		case TCP_DEFER_OPTIONS:			/*  URL:defer */
25381 		case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
25382 		case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
25383 		case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
25384 		case TCP_RACK_SET_RXT_OPTIONS:		/*  URL:rxtsz */
25385 		case TCP_RACK_HI_BETA:			/*  URL:hibeta */
25386 		case TCP_RACK_SPLIT_LIMIT:		/*  URL:split */
25387 		case TCP_SS_EEXIT:			/*  URL:eexit */
25388 		case TCP_DGP_UPPER_BOUNDS:		/*  URL:upper */
25389 		case TCP_RACK_PACING_DIVISOR:		/*  URL:divisor */
25390 		case TCP_PACING_DND:			/*  URL:dnd */
25391 		case TCP_NO_TIMELY:			/*  URL:notimely */
25392 		case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
25393 		case TCP_HONOR_HPTS_MIN:		/*  URL:hptsmin */
25394 		case TCP_REC_IS_DYN:			/*  URL:dynrec */
25395 		case TCP_GP_USE_LTBW:			/*  URL:useltbw */
25396 			goto process_opt;
25397 			break;
25398 		default:
25399 			/* Filter off all unknown options to the base stack */
25400 			return (tcp_default_ctloutput(tp, sopt));
25401 			break;
25402 		}
25403 	default:
25404 		INP_WUNLOCK(inp);
25405 		return (0);
25406 	}
25407 process_opt:
25408 	INP_WUNLOCK(inp);
25409 	if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
25410 	    (sopt->sopt_name == TCP_FILLCW_RATE_CAP)) {
25411 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
25412 		/*
25413 		 * We truncate it down to 32 bits for the socket-option trace this
25414 		 * means rates > 34Gbps won't show right, but thats probably ok.
25415 		 */
25416 		optval = (uint32_t)loptval;
25417 	} else if (sopt->sopt_name == TCP_HYBRID_PACING) {
25418 		error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
25419 	} else {
25420 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
25421 		/* Save it in 64 bit form too */
25422 		loptval = optval;
25423 	}
25424 	if (error)
25425 		return (error);
25426 	INP_WLOCK(inp);
25427 	if (tp->t_fb != &__tcp_rack) {
25428 		INP_WUNLOCK(inp);
25429 		return (ENOPROTOOPT);
25430 	}
25431 	if (rack->defer_options && (rack->gp_ready == 0) &&
25432 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
25433 	    (sopt->sopt_name != TCP_HYBRID_PACING) &&
25434 	    (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
25435 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
25436 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
25437 		/* Options are being deferred */
25438 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
25439 			INP_WUNLOCK(inp);
25440 			return (0);
25441 		} else {
25442 			/* No memory to defer, fail */
25443 			INP_WUNLOCK(inp);
25444 			return (ENOMEM);
25445 		}
25446 	}
25447 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
25448 	INP_WUNLOCK(inp);
25449 	return (error);
25450 }
25451 
25452 static void
25453 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
25454 {
25455 
25456 	INP_WLOCK_ASSERT(tptoinpcb(tp));
25457 	bzero(ti, sizeof(*ti));
25458 
25459 	ti->tcpi_state = tp->t_state;
25460 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
25461 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
25462 	if (tp->t_flags & TF_SACK_PERMIT)
25463 		ti->tcpi_options |= TCPI_OPT_SACK;
25464 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
25465 		ti->tcpi_options |= TCPI_OPT_WSCALE;
25466 		ti->tcpi_snd_wscale = tp->snd_scale;
25467 		ti->tcpi_rcv_wscale = tp->rcv_scale;
25468 	}
25469 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
25470 		ti->tcpi_options |= TCPI_OPT_ECN;
25471 	if (tp->t_flags & TF_FASTOPEN)
25472 		ti->tcpi_options |= TCPI_OPT_TFO;
25473 	/* still kept in ticks is t_rcvtime */
25474 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
25475 	/* Since we hold everything in precise useconds this is easy */
25476 	ti->tcpi_rtt = tp->t_srtt;
25477 	ti->tcpi_rttvar = tp->t_rttvar;
25478 	ti->tcpi_rto = tp->t_rxtcur;
25479 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
25480 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
25481 	/*
25482 	 * FreeBSD-specific extension fields for tcp_info.
25483 	 */
25484 	ti->tcpi_rcv_space = tp->rcv_wnd;
25485 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
25486 	ti->tcpi_snd_wnd = tp->snd_wnd;
25487 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
25488 	ti->tcpi_snd_nxt = tp->snd_nxt;
25489 	ti->tcpi_snd_mss = tp->t_maxseg;
25490 	ti->tcpi_rcv_mss = tp->t_maxseg;
25491 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
25492 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
25493 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
25494 	ti->tcpi_total_tlp = tp->t_sndtlppack;
25495 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
25496 	ti->tcpi_rttmin = tp->t_rttlow;
25497 #ifdef NETFLIX_STATS
25498 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
25499 #endif
25500 #ifdef TCP_OFFLOAD
25501 	if (tp->t_flags & TF_TOE) {
25502 		ti->tcpi_options |= TCPI_OPT_TOE;
25503 		tcp_offload_tcp_info(tp, ti);
25504 	}
25505 #endif
25506 }
25507 
25508 static int
25509 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
25510 {
25511 	struct inpcb *inp = tptoinpcb(tp);
25512 	struct tcp_rack *rack;
25513 	int32_t error, optval;
25514 	uint64_t val, loptval;
25515 	struct	tcp_info ti;
25516 	/*
25517 	 * Because all our options are either boolean or an int, we can just
25518 	 * pull everything into optval and then unlock and copy. If we ever
25519 	 * add a option that is not a int, then this will have quite an
25520 	 * impact to this routine.
25521 	 */
25522 	error = 0;
25523 	rack = (struct tcp_rack *)tp->t_fb_ptr;
25524 	if (rack == NULL) {
25525 		INP_WUNLOCK(inp);
25526 		return (EINVAL);
25527 	}
25528 	switch (sopt->sopt_name) {
25529 	case TCP_INFO:
25530 		/* First get the info filled */
25531 		rack_fill_info(tp, &ti);
25532 		/* Fix up the rtt related fields if needed */
25533 		INP_WUNLOCK(inp);
25534 		error = sooptcopyout(sopt, &ti, sizeof ti);
25535 		return (error);
25536 	/*
25537 	 * Beta is the congestion control value for NewReno that influences how
25538 	 * much of a backoff happens when loss is detected. It is normally set
25539 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
25540 	 * when you exit recovery.
25541 	 */
25542 	case TCP_RACK_PACING_BETA:
25543 		break;
25544 		/*
25545 		 * Beta_ecn is the congestion control value for NewReno that influences how
25546 		 * much of a backoff happens when a ECN mark is detected. It is normally set
25547 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
25548 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
25549 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
25550 		 */
25551 
25552 	case TCP_RACK_PACING_BETA_ECN:
25553 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
25554 			error = EINVAL;
25555 		else if (rack->rc_pacing_cc_set == 0)
25556 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
25557 		else {
25558 			/*
25559 			 * Reach out into the CC data and report back what
25560 			 * I have previously set. Yeah it looks hackish but
25561 			 * we don't want to report the saved values.
25562 			 */
25563 			if (tp->t_ccv.cc_data)
25564 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
25565 			else
25566 				error = EINVAL;
25567 		}
25568 		break;
25569 	case TCP_RACK_DSACK_OPT:
25570 		optval = 0;
25571 		if (rack->rc_rack_tmr_std_based) {
25572 			optval |= 1;
25573 		}
25574 		if (rack->rc_rack_use_dsack) {
25575 			optval |= 2;
25576 		}
25577 		break;
25578 	case TCP_RACK_ENABLE_HYSTART:
25579 	{
25580 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
25581 			optval = RACK_HYSTART_ON;
25582 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
25583 				optval = RACK_HYSTART_ON_W_SC;
25584 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
25585 				optval = RACK_HYSTART_ON_W_SC_C;
25586 		} else {
25587 			optval = RACK_HYSTART_OFF;
25588 		}
25589 	}
25590 	break;
25591 	case TCP_RACK_DGP_IN_REC:
25592 		error = EINVAL;
25593 		break;
25594 	case TCP_RACK_HI_BETA:
25595 		optval = rack->rack_hibeta;
25596 		break;
25597 	case TCP_POLICER_MSS:
25598 		optval = rack->r_ctl.policer_del_mss;
25599 		break;
25600 	case TCP_POLICER_DETECT:
25601 		optval = rack->r_ctl.saved_policer_val;
25602 		break;
25603 	case TCP_DEFER_OPTIONS:
25604 		optval = rack->defer_options;
25605 		break;
25606 	case TCP_RACK_MEASURE_CNT:
25607 		optval = rack->r_ctl.req_measurements;
25608 		break;
25609 	case TCP_REC_ABC_VAL:
25610 		optval = rack->r_use_labc_for_rec;
25611 		break;
25612 	case TCP_RACK_ABC_VAL:
25613 		optval = rack->rc_labc;
25614 		break;
25615 	case TCP_HDWR_UP_ONLY:
25616 		optval= rack->r_up_only;
25617 		break;
25618 	case TCP_FILLCW_RATE_CAP:
25619 		loptval = rack->r_ctl.fillcw_cap;
25620 		break;
25621 	case TCP_PACING_RATE_CAP:
25622 		loptval = rack->r_ctl.bw_rate_cap;
25623 		break;
25624 	case TCP_RACK_PROFILE:
25625 		/* You cannot retrieve a profile, its write only */
25626 		error = EINVAL;
25627 		break;
25628 	case TCP_SIDECHAN_DIS:
25629 		optval = rack->r_ctl.side_chan_dis_mask;
25630 		break;
25631 	case TCP_HYBRID_PACING:
25632 		/* You cannot retrieve hybrid pacing information, its write only */
25633 		error = EINVAL;
25634 		break;
25635 	case TCP_USE_CMP_ACKS:
25636 		optval = rack->r_use_cmp_ack;
25637 		break;
25638 	case TCP_RACK_PACE_TO_FILL:
25639 		optval = rack->rc_pace_to_cwnd;
25640 		break;
25641 	case TCP_RACK_NO_PUSH_AT_MAX:
25642 		optval = rack->r_ctl.rc_no_push_at_mrtt;
25643 		break;
25644 	case TCP_SHARED_CWND_ENABLE:
25645 		optval = rack->rack_enable_scwnd;
25646 		break;
25647 	case TCP_RACK_NONRXT_CFG_RATE:
25648 		optval = rack->rack_rec_nonrxt_use_cr;
25649 		break;
25650 	case TCP_NO_PRR:
25651 		if (rack->rack_no_prr  == 1)
25652 			optval = 1;
25653 		else if (rack->no_prr_addback == 1)
25654 			optval = 2;
25655 		else
25656 			optval = 0;
25657 		break;
25658 	case TCP_GP_USE_LTBW:
25659 		if (rack->dis_lt_bw) {
25660 			/* It is not used */
25661 			optval = 0;
25662 		} else if (rack->use_lesser_lt_bw) {
25663 			/* we use min() */
25664 			optval = 1;
25665 		} else {
25666 			/* we use max() */
25667 			optval = 2;
25668 		}
25669 		break;
25670 	case TCP_RACK_DO_DETECTION:
25671 		error = EINVAL;
25672 		break;
25673 	case TCP_RACK_MBUF_QUEUE:
25674 		/* Now do we use the LRO mbuf-queue feature */
25675 		optval = rack->r_mbuf_queue;
25676 		break;
25677 	case RACK_CSPR_IS_FCC:
25678 		optval = rack->cspr_is_fcc;
25679 		break;
25680 	case TCP_TIMELY_DYN_ADJ:
25681 		optval = rack->rc_gp_dyn_mul;
25682 		break;
25683 	case TCP_BBR_IWINTSO:
25684 		error = EINVAL;
25685 		break;
25686 	case TCP_RACK_TLP_REDUCE:
25687 		/* RACK TLP cwnd reduction (bool) */
25688 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
25689 		break;
25690 	case TCP_BBR_RACK_INIT_RATE:
25691 		val = rack->r_ctl.init_rate;
25692 		/* convert to kbits per sec */
25693 		val *= 8;
25694 		val /= 1000;
25695 		optval = (uint32_t)val;
25696 		break;
25697 	case TCP_RACK_FORCE_MSEG:
25698 		optval = rack->rc_force_max_seg;
25699 		break;
25700 	case TCP_RACK_PACE_MIN_SEG:
25701 		optval = rack->r_ctl.rc_user_set_min_segs;
25702 		break;
25703 	case TCP_RACK_PACE_MAX_SEG:
25704 		/* Max segments in a pace */
25705 		optval = rack->rc_user_set_max_segs;
25706 		break;
25707 	case TCP_RACK_PACE_ALWAYS:
25708 		/* Use the always pace method */
25709 		optval = rack->rc_always_pace;
25710 		break;
25711 	case TCP_RACK_PRR_SENDALOT:
25712 		/* Allow PRR to send more than one seg */
25713 		optval = rack->r_ctl.rc_prr_sendalot;
25714 		break;
25715 	case TCP_RACK_MIN_TO:
25716 		/* Minimum time between rack t-o's in ms */
25717 		optval = rack->r_ctl.rc_min_to;
25718 		break;
25719 	case TCP_RACK_SPLIT_LIMIT:
25720 		optval = rack->r_ctl.rc_split_limit;
25721 		break;
25722 	case TCP_RACK_EARLY_SEG:
25723 		/* If early recovery max segments */
25724 		optval = rack->r_ctl.rc_early_recovery_segs;
25725 		break;
25726 	case TCP_RACK_REORD_THRESH:
25727 		/* RACK reorder threshold (shift amount) */
25728 		optval = rack->r_ctl.rc_reorder_shift;
25729 		break;
25730 	case TCP_SS_EEXIT:
25731 		if (rack->r_ctl.gp_rnd_thresh) {
25732 			uint32_t v;
25733 
25734 			v = rack->r_ctl.gp_gain_req;
25735 			v <<= 17;
25736 			optval = v | (rack->r_ctl.gp_rnd_thresh & 0xff);
25737 			if (rack->r_ctl.gate_to_fs == 1)
25738 				optval |= 0x10000;
25739 		} else
25740 			optval = 0;
25741 		break;
25742 	case TCP_RACK_REORD_FADE:
25743 		/* Does reordering fade after ms time */
25744 		optval = rack->r_ctl.rc_reorder_fade;
25745 		break;
25746 	case TCP_BBR_USE_RACK_RR:
25747 		/* Do we use the rack cheat for rxt */
25748 		optval = rack->use_rack_rr;
25749 		break;
25750 	case TCP_RACK_RR_CONF:
25751 		optval = rack->r_rr_config;
25752 		break;
25753 	case TCP_HDWR_RATE_CAP:
25754 		optval = rack->r_rack_hw_rate_caps;
25755 		break;
25756 	case TCP_BBR_HDWR_PACE:
25757 		optval = rack->rack_hdw_pace_ena;
25758 		break;
25759 	case TCP_RACK_TLP_THRESH:
25760 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
25761 		optval = rack->r_ctl.rc_tlp_threshold;
25762 		break;
25763 	case TCP_RACK_PKT_DELAY:
25764 		/* RACK added ms i.e. rack-rtt + reord + N */
25765 		optval = rack->r_ctl.rc_pkt_delay;
25766 		break;
25767 	case TCP_RACK_TLP_USE:
25768 		optval = rack->rack_tlp_threshold_use;
25769 		break;
25770 	case TCP_PACING_DND:
25771 		optval = rack->rc_pace_dnd;
25772 		break;
25773 	case TCP_RACK_PACE_RATE_CA:
25774 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
25775 		break;
25776 	case TCP_RACK_PACE_RATE_SS:
25777 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
25778 		break;
25779 	case TCP_RACK_PACE_RATE_REC:
25780 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
25781 		break;
25782 	case TCP_DGP_UPPER_BOUNDS:
25783 		optval = rack->r_ctl.rack_per_upper_bound_ss;
25784 		optval <<= 16;
25785 		optval |= rack->r_ctl.rack_per_upper_bound_ca;
25786 		break;
25787 	case TCP_RACK_GP_INCREASE_SS:
25788 		optval = rack->r_ctl.rack_per_of_gp_ca;
25789 		break;
25790 	case TCP_RACK_GP_INCREASE_CA:
25791 		optval = rack->r_ctl.rack_per_of_gp_ss;
25792 		break;
25793 	case TCP_RACK_PACING_DIVISOR:
25794 		optval = rack->r_ctl.pace_len_divisor;
25795 		break;
25796 	case TCP_BBR_RACK_RTT_USE:
25797 		optval = rack->r_ctl.rc_rate_sample_method;
25798 		break;
25799 	case TCP_DELACK:
25800 		optval = tp->t_delayed_ack;
25801 		break;
25802 	case TCP_DATA_AFTER_CLOSE:
25803 		optval = rack->rc_allow_data_af_clo;
25804 		break;
25805 	case TCP_SHARED_CWND_TIME_LIMIT:
25806 		optval = rack->r_limit_scw;
25807 		break;
25808 	case TCP_HONOR_HPTS_MIN:
25809 		if (rack->r_use_hpts_min)
25810 			optval = rack->r_ctl.max_reduction;
25811 		else
25812 			optval = 0;
25813 		break;
25814 	case TCP_REC_IS_DYN:
25815 		optval = rack->rc_gp_no_rec_chg;
25816 		break;
25817 	case TCP_NO_TIMELY:
25818 		optval = rack->rc_skip_timely;
25819 		break;
25820 	case TCP_RACK_TIMER_SLOP:
25821 		optval = rack->r_ctl.timer_slop;
25822 		break;
25823 	default:
25824 		return (tcp_default_ctloutput(tp, sopt));
25825 		break;
25826 	}
25827 	INP_WUNLOCK(inp);
25828 	if (error == 0) {
25829 		if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
25830 		    (sopt->sopt_name == TCP_FILLCW_RATE_CAP))
25831 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
25832 		else
25833 			error = sooptcopyout(sopt, &optval, sizeof optval);
25834 	}
25835 	return (error);
25836 }
25837 
25838 static int
25839 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
25840 {
25841 	if (sopt->sopt_dir == SOPT_SET) {
25842 		return (rack_set_sockopt(tp, sopt));
25843 	} else if (sopt->sopt_dir == SOPT_GET) {
25844 		return (rack_get_sockopt(tp, sopt));
25845 	} else {
25846 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
25847 	}
25848 }
25849 
25850 static const char *rack_stack_names[] = {
25851 	__XSTRING(STACKNAME),
25852 #ifdef STACKALIAS
25853 	__XSTRING(STACKALIAS),
25854 #endif
25855 };
25856 
25857 static int
25858 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
25859 {
25860 	memset(mem, 0, size);
25861 	return (0);
25862 }
25863 
25864 static void
25865 rack_dtor(void *mem, int32_t size, void *arg)
25866 {
25867 
25868 }
25869 
25870 static bool rack_mod_inited = false;
25871 
25872 static int
25873 tcp_addrack(module_t mod, int32_t type, void *data)
25874 {
25875 	int32_t err = 0;
25876 	int num_stacks;
25877 
25878 	switch (type) {
25879 	case MOD_LOAD:
25880 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
25881 		    sizeof(struct rack_sendmap),
25882 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
25883 
25884 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
25885 		    sizeof(struct tcp_rack),
25886 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
25887 
25888 		sysctl_ctx_init(&rack_sysctl_ctx);
25889 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
25890 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
25891 		    OID_AUTO,
25892 #ifdef STACKALIAS
25893 		    __XSTRING(STACKALIAS),
25894 #else
25895 		    __XSTRING(STACKNAME),
25896 #endif
25897 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
25898 		    "");
25899 		if (rack_sysctl_root == NULL) {
25900 			printf("Failed to add sysctl node\n");
25901 			err = EFAULT;
25902 			goto free_uma;
25903 		}
25904 		rack_init_sysctls();
25905 		num_stacks = nitems(rack_stack_names);
25906 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
25907 		    rack_stack_names, &num_stacks);
25908 		if (err) {
25909 			printf("Failed to register %s stack name for "
25910 			    "%s module\n", rack_stack_names[num_stacks],
25911 			    __XSTRING(MODNAME));
25912 			sysctl_ctx_free(&rack_sysctl_ctx);
25913 free_uma:
25914 			uma_zdestroy(rack_zone);
25915 			uma_zdestroy(rack_pcb_zone);
25916 			rack_counter_destroy();
25917 			printf("Failed to register rack module -- err:%d\n", err);
25918 			return (err);
25919 		}
25920 		tcp_lro_reg_mbufq();
25921 		rack_mod_inited = true;
25922 		break;
25923 	case MOD_QUIESCE:
25924 		err = deregister_tcp_functions(&__tcp_rack, true, false);
25925 		break;
25926 	case MOD_UNLOAD:
25927 		err = deregister_tcp_functions(&__tcp_rack, false, true);
25928 		if (err == EBUSY)
25929 			break;
25930 		if (rack_mod_inited) {
25931 			uma_zdestroy(rack_zone);
25932 			uma_zdestroy(rack_pcb_zone);
25933 			sysctl_ctx_free(&rack_sysctl_ctx);
25934 			rack_counter_destroy();
25935 			rack_mod_inited = false;
25936 		}
25937 		tcp_lro_dereg_mbufq();
25938 		err = 0;
25939 		break;
25940 	default:
25941 		return (EOPNOTSUPP);
25942 	}
25943 	return (err);
25944 }
25945 
25946 static moduledata_t tcp_rack = {
25947 	.name = __XSTRING(MODNAME),
25948 	.evhand = tcp_addrack,
25949 	.priv = 0
25950 };
25951 
25952 MODULE_VERSION(MODNAME, 1);
25953 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
25954 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
25955 
25956 #endif /* #if !defined(INET) && !defined(INET6) */
25957