xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 849f9ac370bd66993ce5cc6fca0d2ef9bd03c2c9) 
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 uint32_t rack_pcm_every_n_rounds = 100;
197 static uint32_t rack_pcm_blast = 0;
198 static uint32_t rack_pcm_is_enabled = 1;
199 static uint8_t rack_ssthresh_rest_rto_rec = 0; /* Do we restore ssthresh when we have rec -> rto -> rec */
200 
201 static uint32_t rack_gp_gain_req = 1200;		/* Amount percent wise required to gain to record a round has "gaining" */
202 static uint32_t rack_rnd_cnt_req = 0x10005;		/* Default number of rounds if we are below rack_gp_gain_req where we exit ss */
203 
204 
205 static int32_t rack_rxt_scoreboard_clear_thresh = 2;
206 static int32_t rack_dnd_default = 0;		/* For rr_conf = 3, what is the default for dnd */
207 static int32_t rack_rxt_controls = 0;
208 static int32_t rack_fill_cw_state = 0;
209 static uint8_t rack_req_measurements = 1;
210 /* Attack threshold detections */
211 static uint32_t rack_highest_sack_thresh_seen = 0;
212 static uint32_t rack_highest_move_thresh_seen = 0;
213 static uint32_t rack_merge_out_sacks_on_attack = 0;
214 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
215 static int32_t rack_hw_pace_extra_slots = 0;	/* 2 extra MSS time betweens */
216 static int32_t rack_hw_rate_caps = 0; /* 1; */
217 static int32_t rack_hw_rate_cap_per = 0;	/* 0 -- off  */
218 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
219 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
220 static int32_t rack_hw_up_only = 0;
221 static int32_t rack_stats_gets_ms_rtt = 1;
222 static int32_t rack_prr_addbackmax = 2;
223 static int32_t rack_do_hystart = 0;
224 static int32_t rack_apply_rtt_with_reduced_conf = 0;
225 static int32_t rack_hibeta_setting = 0;
226 static int32_t rack_default_pacing_divisor = 250;
227 static uint16_t rack_pacing_min_seg = 0;
228 static int32_t rack_timely_off = 0;
229 
230 static uint32_t sad_seg_size_per = 800;	/* 80.0 % */
231 static int32_t rack_pkt_delay = 1000;
232 static int32_t rack_send_a_lot_in_prr = 1;
233 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
234 static int32_t rack_verbose_logging = 0;
235 static int32_t rack_ignore_data_after_close = 1;
236 static int32_t rack_enable_shared_cwnd = 1;
237 static int32_t rack_use_cmp_acks = 1;
238 static int32_t rack_use_fsb = 1;
239 static int32_t rack_use_rfo = 1;
240 static int32_t rack_use_rsm_rfo = 1;
241 static int32_t rack_max_abc_post_recovery = 2;
242 static int32_t rack_client_low_buf = 0;
243 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 */
244 static int32_t rack_bw_multipler = 0;		/* Limit on fill cw's jump up to be this x gp_est */
245 #ifdef TCP_ACCOUNTING
246 static int32_t rack_tcp_accounting = 0;
247 #endif
248 static int32_t rack_limits_scwnd = 1;
249 static int32_t rack_enable_mqueue_for_nonpaced = 0;
250 static int32_t rack_hybrid_allow_set_maxseg = 0;
251 static int32_t rack_disable_prr = 0;
252 static int32_t use_rack_rr = 1;
253 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
254 static int32_t rack_persist_min = 250000;	/* 250usec */
255 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
256 static int32_t rack_honors_hpts_min_to =  1;	/* Do we honor the hpts minimum time out for pacing timers */
257 static uint32_t rack_max_reduce = 10;		/* Percent we can reduce slot by */
258 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
259 static int32_t rack_limit_time_with_srtt = 0;
260 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
261 static int32_t rack_enobuf_hw_boost_mult = 0;	/* How many times the hw rate we boost slot using time_between */
262 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
263 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
264 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
265 static int32_t rack_hw_check_queue = 0;		/* Do we always pre-check queue depth of a hw queue */
266 static int32_t rack_full_buffer_discount = 10;
267 /*
268  * Currently regular tcp has a rto_min of 30ms
269  * the backoff goes 12 times so that ends up
270  * being a total of 122.850 seconds before a
271  * connection is killed.
272  */
273 static uint32_t rack_def_data_window = 20;
274 static uint32_t rack_goal_bdp = 2;
275 static uint32_t rack_min_srtts = 1;
276 static uint32_t rack_min_measure_usec = 0;
277 static int32_t rack_tlp_min = 10000;	/* 10ms */
278 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
279 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
280 static const int32_t rack_free_cache = 2;
281 static int32_t rack_hptsi_segments = 40;
282 static int32_t rack_rate_sample_method = USE_RTT_LOW;
283 static int32_t rack_pace_every_seg = 0;
284 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
285 static int32_t rack_slot_reduction = 4;
286 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
287 static int32_t rack_cwnd_block_ends_measure = 0;
288 static int32_t rack_rwnd_block_ends_measure = 0;
289 static int32_t rack_def_profile = 0;
290 
291 static int32_t rack_lower_cwnd_at_tlp = 0;
292 static int32_t rack_always_send_oldest = 0;
293 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
294 
295 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
296 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
297 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
298 
299 /* Probertt */
300 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
301 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
302 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
303 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
304 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
305 
306 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
307 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
308 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
309 static uint32_t rack_probertt_use_min_rtt_exit = 0;
310 static uint32_t rack_probe_rtt_sets_cwnd = 0;
311 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
312 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
313 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
314 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
315 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
316 static uint32_t rack_probertt_filter_life = 10000000;
317 static uint32_t rack_probertt_lower_within = 10;
318 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
319 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
320 static int32_t rack_probertt_clear_is = 1;
321 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
322 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
323 
324 /* Part of pacing */
325 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
326 
327 /* Timely information:
328  *
329  * Here we have various control parameters on how
330  * timely may change the multiplier. rack_gain_p5_ub
331  * is associated with timely but not directly influencing
332  * the rate decision like the other variables. It controls
333  * the way fill-cw interacts with timely and caps how much
334  * timely can boost the fill-cw b/w.
335  *
336  * The other values are various boost/shrink numbers as well
337  * as potential caps when adjustments are made to the timely
338  * gain (returned by rack_get_output_gain(). Remember too that
339  * the gain returned can be overriden by other factors such as
340  * probeRTT as well as fixed-rate-pacing.
341  */
342 static int32_t rack_gain_p5_ub = 250;
343 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
344 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
345 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
346 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
347 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
348 static int32_t rack_gp_decrease_per = 80;	/* Beta value of timely decrease (.8) = 80 */
349 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
350 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
351 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
352 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
353 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
354 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
355 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
356 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
357 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
358 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
359 static int32_t rack_use_max_for_nobackoff = 0;
360 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
361 static int32_t rack_timely_no_stopping = 0;
362 static int32_t rack_down_raise_thresh = 100;
363 static int32_t rack_req_segs = 1;
364 static uint64_t rack_bw_rate_cap = 0;
365 static uint64_t rack_fillcw_bw_cap = 3750000;	/* Cap fillcw at 30Mbps */
366 
367 
368 /* Rack specific counters */
369 counter_u64_t rack_saw_enobuf;
370 counter_u64_t rack_saw_enobuf_hw;
371 counter_u64_t rack_saw_enetunreach;
372 counter_u64_t rack_persists_sends;
373 counter_u64_t rack_persists_acks;
374 counter_u64_t rack_persists_loss;
375 counter_u64_t rack_persists_lost_ends;
376 counter_u64_t rack_total_bytes;
377 #ifdef INVARIANTS
378 counter_u64_t rack_adjust_map_bw;
379 #endif
380 /* Tail loss probe counters */
381 counter_u64_t rack_tlp_tot;
382 counter_u64_t rack_tlp_newdata;
383 counter_u64_t rack_tlp_retran;
384 counter_u64_t rack_tlp_retran_bytes;
385 counter_u64_t rack_to_tot;
386 counter_u64_t rack_hot_alloc;
387 counter_u64_t rack_to_alloc;
388 counter_u64_t rack_to_alloc_hard;
389 counter_u64_t rack_to_alloc_emerg;
390 counter_u64_t rack_to_alloc_limited;
391 counter_u64_t rack_alloc_limited_conns;
392 counter_u64_t rack_split_limited;
393 counter_u64_t rack_rxt_clamps_cwnd;
394 counter_u64_t rack_rxt_clamps_cwnd_uniq;
395 
396 counter_u64_t rack_multi_single_eq;
397 counter_u64_t rack_proc_non_comp_ack;
398 
399 counter_u64_t rack_fto_send;
400 counter_u64_t rack_fto_rsm_send;
401 counter_u64_t rack_nfto_resend;
402 counter_u64_t rack_non_fto_send;
403 counter_u64_t rack_extended_rfo;
404 
405 counter_u64_t rack_sack_proc_all;
406 counter_u64_t rack_sack_proc_short;
407 counter_u64_t rack_sack_proc_restart;
408 counter_u64_t rack_sack_attacks_detected;
409 counter_u64_t rack_sack_attacks_reversed;
410 counter_u64_t rack_sack_attacks_suspect;
411 counter_u64_t rack_sack_used_next_merge;
412 counter_u64_t rack_sack_splits;
413 counter_u64_t rack_sack_used_prev_merge;
414 counter_u64_t rack_sack_skipped_acked;
415 counter_u64_t rack_ack_total;
416 counter_u64_t rack_express_sack;
417 counter_u64_t rack_sack_total;
418 counter_u64_t rack_move_none;
419 counter_u64_t rack_move_some;
420 
421 counter_u64_t rack_input_idle_reduces;
422 counter_u64_t rack_collapsed_win;
423 counter_u64_t rack_collapsed_win_seen;
424 counter_u64_t rack_collapsed_win_rxt;
425 counter_u64_t rack_collapsed_win_rxt_bytes;
426 counter_u64_t rack_try_scwnd;
427 counter_u64_t rack_hw_pace_init_fail;
428 counter_u64_t rack_hw_pace_lost;
429 
430 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
431 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
432 
433 
434 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
435 
436 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
437 	(tv) = (value) + slop;	 \
438 	if ((u_long)(tv) < (u_long)(tvmin)) \
439 		(tv) = (tvmin); \
440 	if ((u_long)(tv) > (u_long)(tvmax)) \
441 		(tv) = (tvmax); \
442 } while (0)
443 
444 static void
445 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
446 
447 static int
448 rack_process_ack(struct mbuf *m, struct tcphdr *th,
449     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
450     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val, int32_t orig_tlen);
451 static int
452 rack_process_data(struct mbuf *m, struct tcphdr *th,
453     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
454     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
455 static void
456 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
457    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
458 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
459 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
460     uint8_t limit_type);
461 static struct rack_sendmap *
462 rack_check_recovery_mode(struct tcpcb *tp,
463     uint32_t tsused);
464 static uint32_t
465 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack);
466 static void
467 rack_cong_signal(struct tcpcb *tp,
468 		 uint32_t type, uint32_t ack, int );
469 static void rack_counter_destroy(void);
470 static int
471 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
472 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
473 static void
474 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
475 static void
476 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
477     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
478 static void rack_dtor(void *mem, int32_t size, void *arg);
479 static void
480 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
481     uint32_t flex1, uint32_t flex2,
482     uint32_t flex3, uint32_t flex4,
483     uint32_t flex5, uint32_t flex6,
484     uint16_t flex7, uint8_t mod);
485 
486 static void
487 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
488    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
489    struct rack_sendmap *rsm, uint8_t quality);
490 static struct rack_sendmap *
491 rack_find_high_nonack(struct tcp_rack *rack,
492     struct rack_sendmap *rsm);
493 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
494 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
495 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
496 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
497 static void
498 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
499 			    tcp_seq th_ack, int line, uint8_t quality);
500 static void
501 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
502 
503 static uint32_t
504 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
505 static int32_t rack_handoff_ok(struct tcpcb *tp);
506 static int32_t rack_init(struct tcpcb *tp, void **ptr);
507 static void rack_init_sysctls(void);
508 
509 static void
510 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
511     struct tcphdr *th, int entered_rec, int dup_ack_struck,
512     int *dsack_seen, int *sacks_seen);
513 static void
514 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
515     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
516     struct rack_sendmap *hintrsm, uint32_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
517 
518 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
519 
520 
521 static void
522 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
523     struct rack_sendmap *rsm, uint32_t cts);
524 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
525 static int32_t rack_output(struct tcpcb *tp);
526 
527 static uint32_t
528 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
529     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
530     uint32_t cts, uint32_t segsiz);
531 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
532 static void rack_remxt_tmr(struct tcpcb *tp);
533 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
534 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
535 static int32_t rack_stopall(struct tcpcb *tp);
536 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
537 static uint32_t
538 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
539     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint32_t add_flag, int segsiz);
540 static void
541 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
542     struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz);
543 static int
544 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
545     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
546 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
547 static int
548 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
549     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
550     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
551 
552 static int
553 rack_do_closing(struct mbuf *m, struct tcphdr *th,
554     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
555     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
556 static int
557 rack_do_established(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 static int
561 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
562     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
563     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
564 static int
565 rack_do_fin_wait_1(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_fin_wait_2(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_lastack(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 thflags, int32_t nxt_pkt, uint8_t iptos);
576 static int
577 rack_do_syn_recv(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_syn_sent(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 void rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
585 struct rack_sendmap *
586 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
587     uint32_t tsused);
588 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
589     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
590 static void
591      tcp_rack_partialack(struct tcpcb *tp);
592 static int
593 rack_set_profile(struct tcp_rack *rack, int prof);
594 static void
595 rack_apply_deferred_options(struct tcp_rack *rack);
596 
597 int32_t rack_clear_counter=0;
598 
599 static uint64_t
600 rack_get_lt_bw(struct tcp_rack *rack)
601 {
602 	struct timeval tv;
603 	uint64_t tim, bytes;
604 
605 	tim = rack->r_ctl.lt_bw_time;
606 	bytes = rack->r_ctl.lt_bw_bytes;
607 	if (rack->lt_bw_up) {
608 		/* Include all the current bytes too */
609 		microuptime(&tv);
610 		bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
611 		tim += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
612 	}
613 	if ((bytes != 0) && (tim != 0))
614 		return ((bytes * (uint64_t)1000000) / tim);
615 	else
616 		return (0);
617 }
618 
619 static void
620 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
621 {
622 	struct sockopt sopt;
623 	struct cc_newreno_opts opt;
624 	struct newreno old;
625 	struct tcpcb *tp;
626 	int error, failed = 0;
627 
628 	tp = rack->rc_tp;
629 	if (tp->t_cc == NULL) {
630 		/* Tcb is leaving */
631 		return;
632 	}
633 	rack->rc_pacing_cc_set = 1;
634 	if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
635 		/* Not new-reno we can't play games with beta! */
636 		failed = 1;
637 		goto out;
638 
639 	}
640 	if (CC_ALGO(tp)->ctl_output == NULL)  {
641 		/* Huh, not using new-reno so no swaps.? */
642 		failed = 2;
643 		goto out;
644 	}
645 	/* Get the current values out */
646 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
647 	sopt.sopt_dir = SOPT_GET;
648 	opt.name = CC_NEWRENO_BETA;
649 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
650 	if (error)  {
651 		failed = 3;
652 		goto out;
653 	}
654 	old.beta = opt.val;
655 	opt.name = CC_NEWRENO_BETA_ECN;
656 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
657 	if (error)  {
658 		failed = 4;
659 		goto out;
660 	}
661 	old.beta_ecn = opt.val;
662 
663 	/* Now lets set in the values we have stored */
664 	sopt.sopt_dir = SOPT_SET;
665 	opt.name = CC_NEWRENO_BETA;
666 	opt.val = rack->r_ctl.rc_saved_beta.beta;
667 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
668 	if (error)  {
669 		failed = 5;
670 		goto out;
671 	}
672 	opt.name = CC_NEWRENO_BETA_ECN;
673 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
674 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
675 	if (error) {
676 		failed = 6;
677 		goto out;
678 	}
679 	/* Save off the values for restoral */
680 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
681 out:
682 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
683 		union tcp_log_stackspecific log;
684 		struct timeval tv;
685 		struct newreno *ptr;
686 
687 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
688 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
689 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
690 		log.u_bbr.flex1 = ptr->beta;
691 		log.u_bbr.flex2 = ptr->beta_ecn;
692 		log.u_bbr.flex3 = ptr->newreno_flags;
693 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
694 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
695 		log.u_bbr.flex6 = failed;
696 		log.u_bbr.flex7 = rack->gp_ready;
697 		log.u_bbr.flex7 <<= 1;
698 		log.u_bbr.flex7 |= rack->use_fixed_rate;
699 		log.u_bbr.flex7 <<= 1;
700 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
701 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
702 		log.u_bbr.flex8 = flex8;
703 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
704 			       0, &log, false, NULL, NULL, 0, &tv);
705 	}
706 }
707 
708 static void
709 rack_set_cc_pacing(struct tcp_rack *rack)
710 {
711 	if (rack->rc_pacing_cc_set)
712 		return;
713 	/*
714 	 * Use the swap utility placing in 3 for flex8 to id a
715 	 * set of a new set of values.
716 	 */
717 	rack->rc_pacing_cc_set = 1;
718 	rack_swap_beta_values(rack, 3);
719 }
720 
721 static void
722 rack_undo_cc_pacing(struct tcp_rack *rack)
723 {
724 	if (rack->rc_pacing_cc_set == 0)
725 		return;
726 	/*
727 	 * Use the swap utility placing in 4 for flex8 to id a
728 	 * restoral of the old values.
729 	 */
730 	rack->rc_pacing_cc_set = 0;
731 	rack_swap_beta_values(rack, 4);
732 }
733 
734 static void
735 rack_remove_pacing(struct tcp_rack *rack)
736 {
737 	if (rack->rc_pacing_cc_set)
738 		rack_undo_cc_pacing(rack);
739 	if (rack->r_ctl.pacing_method & RACK_REG_PACING)
740 		tcp_decrement_paced_conn();
741 	if (rack->r_ctl.pacing_method & RACK_DGP_PACING)
742 		tcp_dec_dgp_pacing_cnt();
743 	rack->rc_always_pace = 0;
744 	rack->r_ctl.pacing_method = RACK_PACING_NONE;
745 	rack->dgp_on = 0;
746 	rack->rc_hybrid_mode = 0;
747 	rack->use_fixed_rate = 0;
748 }
749 
750 static void
751 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
752 	       uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
753 {
754 	if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
755 		union tcp_log_stackspecific log;
756 		struct timeval tv;
757 
758 		memset(&log, 0, sizeof(log));
759 		log.u_bbr.flex1 = seq_end;
760 		log.u_bbr.flex2 = rack->rc_tp->gput_seq;
761 		log.u_bbr.flex3 = ack_end_t;
762 		log.u_bbr.flex4 = rack->rc_tp->gput_ts;
763 		log.u_bbr.flex5 = send_end_t;
764 		log.u_bbr.flex6 = rack->rc_tp->gput_ack;
765 		log.u_bbr.flex7 = mode;
766 		log.u_bbr.flex8 = 69;
767 		log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
768 		log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
769 		log.u_bbr.pkts_out = line;
770 		log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
771 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
772 		log.u_bbr.epoch = rack->r_ctl.current_round;
773 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
774 		if (rsm != NULL) {
775 			log.u_bbr.applimited = rsm->r_start;
776 			log.u_bbr.delivered = rsm->r_end;
777 			log.u_bbr.epoch = rsm->r_flags;
778 		}
779 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
780 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
781 		    &rack->rc_inp->inp_socket->so_rcv,
782 		    &rack->rc_inp->inp_socket->so_snd,
783 		    BBR_LOG_HPTSI_CALC, 0,
784 		    0, &log, false, &tv);
785 	}
786 }
787 
788 static int
789 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
790 {
791 	uint32_t stat;
792 	int32_t error;
793 
794 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
795 	if (error || req->newptr == NULL)
796 		return error;
797 
798 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
799 	if (error)
800 		return (error);
801 	if (stat == 1) {
802 #ifdef INVARIANTS
803 		printf("Clearing RACK counters\n");
804 #endif
805 		counter_u64_zero(rack_tlp_tot);
806 		counter_u64_zero(rack_tlp_newdata);
807 		counter_u64_zero(rack_tlp_retran);
808 		counter_u64_zero(rack_tlp_retran_bytes);
809 		counter_u64_zero(rack_to_tot);
810 		counter_u64_zero(rack_saw_enobuf);
811 		counter_u64_zero(rack_saw_enobuf_hw);
812 		counter_u64_zero(rack_saw_enetunreach);
813 		counter_u64_zero(rack_persists_sends);
814 		counter_u64_zero(rack_total_bytes);
815 		counter_u64_zero(rack_persists_acks);
816 		counter_u64_zero(rack_persists_loss);
817 		counter_u64_zero(rack_persists_lost_ends);
818 #ifdef INVARIANTS
819 		counter_u64_zero(rack_adjust_map_bw);
820 #endif
821 		counter_u64_zero(rack_to_alloc_hard);
822 		counter_u64_zero(rack_to_alloc_emerg);
823 		counter_u64_zero(rack_sack_proc_all);
824 		counter_u64_zero(rack_fto_send);
825 		counter_u64_zero(rack_fto_rsm_send);
826 		counter_u64_zero(rack_extended_rfo);
827 		counter_u64_zero(rack_hw_pace_init_fail);
828 		counter_u64_zero(rack_hw_pace_lost);
829 		counter_u64_zero(rack_non_fto_send);
830 		counter_u64_zero(rack_nfto_resend);
831 		counter_u64_zero(rack_sack_proc_short);
832 		counter_u64_zero(rack_sack_proc_restart);
833 		counter_u64_zero(rack_to_alloc);
834 		counter_u64_zero(rack_to_alloc_limited);
835 		counter_u64_zero(rack_alloc_limited_conns);
836 		counter_u64_zero(rack_split_limited);
837 		counter_u64_zero(rack_rxt_clamps_cwnd);
838 		counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
839 		counter_u64_zero(rack_multi_single_eq);
840 		counter_u64_zero(rack_proc_non_comp_ack);
841 		counter_u64_zero(rack_sack_attacks_detected);
842 		counter_u64_zero(rack_sack_attacks_reversed);
843 		counter_u64_zero(rack_sack_attacks_suspect);
844 		counter_u64_zero(rack_sack_used_next_merge);
845 		counter_u64_zero(rack_sack_used_prev_merge);
846 		counter_u64_zero(rack_sack_splits);
847 		counter_u64_zero(rack_sack_skipped_acked);
848 		counter_u64_zero(rack_ack_total);
849 		counter_u64_zero(rack_express_sack);
850 		counter_u64_zero(rack_sack_total);
851 		counter_u64_zero(rack_move_none);
852 		counter_u64_zero(rack_move_some);
853 		counter_u64_zero(rack_try_scwnd);
854 		counter_u64_zero(rack_collapsed_win);
855 		counter_u64_zero(rack_collapsed_win_rxt);
856 		counter_u64_zero(rack_collapsed_win_seen);
857 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
858 	} else if (stat == 2) {
859 #ifdef INVARIANTS
860 		printf("Clearing RACK option array\n");
861 #endif
862 		COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
863 	} else if (stat == 3) {
864 		printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
865 	} else if (stat == 4) {
866 #ifdef INVARIANTS
867 		printf("Clearing RACK out size array\n");
868 #endif
869 		COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
870 	}
871 	rack_clear_counter = 0;
872 	return (0);
873 }
874 
875 static void
876 rack_init_sysctls(void)
877 {
878 	struct sysctl_oid *rack_counters;
879 	struct sysctl_oid *rack_attack;
880 	struct sysctl_oid *rack_pacing;
881 	struct sysctl_oid *rack_timely;
882 	struct sysctl_oid *rack_timers;
883 	struct sysctl_oid *rack_tlp;
884 	struct sysctl_oid *rack_misc;
885 	struct sysctl_oid *rack_features;
886 	struct sysctl_oid *rack_measure;
887 	struct sysctl_oid *rack_probertt;
888 	struct sysctl_oid *rack_hw_pacing;
889 
890 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
891 	    SYSCTL_CHILDREN(rack_sysctl_root),
892 	    OID_AUTO,
893 	    "sack_attack",
894 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
895 	    "Rack Sack Attack Counters and Controls");
896 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
897 	    SYSCTL_CHILDREN(rack_sysctl_root),
898 	    OID_AUTO,
899 	    "stats",
900 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
901 	    "Rack Counters");
902 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
903 	    SYSCTL_CHILDREN(rack_sysctl_root),
904 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
905 	    &rack_rate_sample_method , USE_RTT_LOW,
906 	    "What method should we use for rate sampling 0=high, 1=low ");
907 	/* Probe rtt related controls */
908 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
909 	    SYSCTL_CHILDREN(rack_sysctl_root),
910 	    OID_AUTO,
911 	    "probertt",
912 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
913 	    "ProbeRTT related Controls");
914 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
915 	    SYSCTL_CHILDREN(rack_probertt),
916 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
917 	    &rack_atexit_prtt_hbp, 130,
918 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
919 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
920 	    SYSCTL_CHILDREN(rack_probertt),
921 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
922 	    &rack_atexit_prtt, 130,
923 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
924 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
925 	    SYSCTL_CHILDREN(rack_probertt),
926 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
927 	    &rack_per_of_gp_probertt, 60,
928 	    "What percentage of goodput do we pace at in probertt");
929 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
930 	    SYSCTL_CHILDREN(rack_probertt),
931 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
932 	    &rack_per_of_gp_probertt_reduce, 10,
933 	    "What percentage of goodput do we reduce every gp_srtt");
934 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
935 	    SYSCTL_CHILDREN(rack_probertt),
936 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
937 	    &rack_per_of_gp_lowthresh, 40,
938 	    "What percentage of goodput do we allow the multiplier to fall to");
939 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
940 	    SYSCTL_CHILDREN(rack_probertt),
941 	    OID_AUTO, "time_between", CTLFLAG_RW,
942 	    & rack_time_between_probertt, 96000000,
943 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
944 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
945 	    SYSCTL_CHILDREN(rack_probertt),
946 	    OID_AUTO, "safety", CTLFLAG_RW,
947 	    &rack_probe_rtt_safety_val, 2000000,
948 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
949 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
950 	    SYSCTL_CHILDREN(rack_probertt),
951 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
952 	    &rack_probe_rtt_sets_cwnd, 0,
953 	    "Do we set the cwnd too (if always_lower is on)");
954 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
955 	    SYSCTL_CHILDREN(rack_probertt),
956 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
957 	    &rack_max_drain_wait, 2,
958 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
959 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
960 	    SYSCTL_CHILDREN(rack_probertt),
961 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
962 	    &rack_must_drain, 1,
963 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
964 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
965 	    SYSCTL_CHILDREN(rack_probertt),
966 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
967 	    &rack_probertt_use_min_rtt_entry, 1,
968 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
969 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
970 	    SYSCTL_CHILDREN(rack_probertt),
971 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
972 	    &rack_probertt_use_min_rtt_exit, 0,
973 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
974 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
975 	    SYSCTL_CHILDREN(rack_probertt),
976 	    OID_AUTO, "length_div", CTLFLAG_RW,
977 	    &rack_probertt_gpsrtt_cnt_div, 0,
978 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
979 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
980 	    SYSCTL_CHILDREN(rack_probertt),
981 	    OID_AUTO, "length_mul", CTLFLAG_RW,
982 	    &rack_probertt_gpsrtt_cnt_mul, 0,
983 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
984 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
985 	    SYSCTL_CHILDREN(rack_probertt),
986 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
987 	    &rack_min_probertt_hold, 200000,
988 	    "What is the minimum time we hold probertt at target");
989 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
990 	    SYSCTL_CHILDREN(rack_probertt),
991 	    OID_AUTO, "filter_life", CTLFLAG_RW,
992 	    &rack_probertt_filter_life, 10000000,
993 	    "What is the time for the filters life in useconds");
994 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
995 	    SYSCTL_CHILDREN(rack_probertt),
996 	    OID_AUTO, "lower_within", CTLFLAG_RW,
997 	    &rack_probertt_lower_within, 10,
998 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
999 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1000 	    SYSCTL_CHILDREN(rack_probertt),
1001 	    OID_AUTO, "must_move", CTLFLAG_RW,
1002 	    &rack_min_rtt_movement, 250,
1003 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
1004 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1005 	    SYSCTL_CHILDREN(rack_probertt),
1006 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
1007 	    &rack_probertt_clear_is, 1,
1008 	    "Do we clear I/S counts on exiting probe-rtt");
1009 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1010 	    SYSCTL_CHILDREN(rack_probertt),
1011 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
1012 	    &rack_max_drain_hbp, 1,
1013 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
1014 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1015 	    SYSCTL_CHILDREN(rack_probertt),
1016 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
1017 	    &rack_hbp_thresh, 3,
1018 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
1019 	/* Pacing related sysctls */
1020 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1021 	    SYSCTL_CHILDREN(rack_sysctl_root),
1022 	    OID_AUTO,
1023 	    "pacing",
1024 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1025 	    "Pacing related Controls");
1026 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1027 	    SYSCTL_CHILDREN(rack_pacing),
1028 	    OID_AUTO, "pcm_enabled", CTLFLAG_RW,
1029 	    &rack_pcm_is_enabled, 1,
1030 	    "Do we by default do PCM measurements?");
1031 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1032 	    SYSCTL_CHILDREN(rack_pacing),
1033 	    OID_AUTO, "pcm_rnds", CTLFLAG_RW,
1034 	    &rack_pcm_every_n_rounds, 100,
1035 	    "How many rounds before we need to do a PCM measurement");
1036 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1037 	    SYSCTL_CHILDREN(rack_pacing),
1038 	    OID_AUTO, "pcm_blast", CTLFLAG_RW,
1039 	    &rack_pcm_blast, 0,
1040 	    "Blast out the full cwnd/rwnd when doing a PCM measurement");
1041 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1042 	    SYSCTL_CHILDREN(rack_pacing),
1043 	    OID_AUTO, "rnd_gp_gain", CTLFLAG_RW,
1044 	    &rack_gp_gain_req, 1200,
1045 	    "How much do we have to increase the GP to record the round 1200 = 120.0");
1046 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1047 	    SYSCTL_CHILDREN(rack_pacing),
1048 	    OID_AUTO, "dgp_out_of_ss_at", CTLFLAG_RW,
1049 	    &rack_rnd_cnt_req, 0x10005,
1050 	    "How many rounds less than rnd_gp_gain will drop us out of SS");
1051 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1052 	    SYSCTL_CHILDREN(rack_pacing),
1053 	    OID_AUTO, "no_timely", CTLFLAG_RW,
1054 	    &rack_timely_off, 0,
1055 	    "Do we not use timely in DGP?");
1056 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1057 	    SYSCTL_CHILDREN(rack_pacing),
1058 	    OID_AUTO, "fullbufdisc", CTLFLAG_RW,
1059 	    &rack_full_buffer_discount, 10,
1060 	    "What percentage b/w reduction over the GP estimate for a full buffer (default=0 off)?");
1061 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1062 	    SYSCTL_CHILDREN(rack_pacing),
1063 	    OID_AUTO, "fillcw", CTLFLAG_RW,
1064 	    &rack_fill_cw_state, 0,
1065 	    "Enable fillcw on new connections (default=0 off)?");
1066 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1067 	    SYSCTL_CHILDREN(rack_pacing),
1068 	    OID_AUTO, "min_burst", CTLFLAG_RW,
1069 	    &rack_pacing_min_seg, 0,
1070 	    "What is the min burst size for pacing (0 disables)?");
1071 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1072 	    SYSCTL_CHILDREN(rack_pacing),
1073 	    OID_AUTO, "divisor", CTLFLAG_RW,
1074 	    &rack_default_pacing_divisor, 250,
1075 	    "What is the default divisor given to the rl code?");
1076 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1077 	    SYSCTL_CHILDREN(rack_pacing),
1078 	    OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1079 	    &rack_bw_multipler, 0,
1080 	    "What is the limit multiplier of the current gp_est that fillcw can increase the b/w too, 200 == 200% (0 = off)?");
1081 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1082 	    SYSCTL_CHILDREN(rack_pacing),
1083 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
1084 	    &rack_max_per_above, 30,
1085 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1086 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1087 	    SYSCTL_CHILDREN(rack_pacing),
1088 	    OID_AUTO, "allow1mss", CTLFLAG_RW,
1089 	    &rack_pace_one_seg, 0,
1090 	    "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1091 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1092 	    SYSCTL_CHILDREN(rack_pacing),
1093 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1094 	    &rack_limit_time_with_srtt, 0,
1095 	    "Do we limit pacing time based on srtt");
1096 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1097 	    SYSCTL_CHILDREN(rack_pacing),
1098 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1099 	    &rack_per_of_gp_ss, 250,
1100 	    "If non zero, what percentage of goodput to pace at in slow start");
1101 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1102 	    SYSCTL_CHILDREN(rack_pacing),
1103 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1104 	    &rack_per_of_gp_ca, 150,
1105 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1106 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1107 	    SYSCTL_CHILDREN(rack_pacing),
1108 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1109 	    &rack_per_of_gp_rec, 200,
1110 	    "If non zero, what percentage of goodput to pace at in recovery");
1111 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1112 	    SYSCTL_CHILDREN(rack_pacing),
1113 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1114 	    &rack_hptsi_segments, 40,
1115 	    "What size is the max for TSO segments in pacing and burst mitigation");
1116 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1117 	    SYSCTL_CHILDREN(rack_pacing),
1118 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1119 	    &rack_slot_reduction, 4,
1120 	    "When doing only burst mitigation what is the reduce divisor");
1121 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1122 	    SYSCTL_CHILDREN(rack_sysctl_root),
1123 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1124 	    &rack_pace_every_seg, 0,
1125 	    "If set we use pacing, if clear we use only the original burst mitigation");
1126 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1127 	    SYSCTL_CHILDREN(rack_pacing),
1128 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1129 	    &rack_bw_rate_cap, 0,
1130 	    "If set we apply this value to the absolute rate cap used by pacing");
1131 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1132 	    SYSCTL_CHILDREN(rack_pacing),
1133 	    OID_AUTO, "fillcw_cap", CTLFLAG_RW,
1134 	    &rack_fillcw_bw_cap, 3750000,
1135 	    "Do we have an absolute cap on the amount of b/w fillcw can specify (0 = no)?");
1136 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1137 	    SYSCTL_CHILDREN(rack_sysctl_root),
1138 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1139 	    &rack_req_measurements, 1,
1140 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1141 	/* Hardware pacing */
1142 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1143 	    SYSCTL_CHILDREN(rack_sysctl_root),
1144 	    OID_AUTO,
1145 	    "hdwr_pacing",
1146 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1147 	    "Pacing related Controls");
1148 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1149 	    SYSCTL_CHILDREN(rack_hw_pacing),
1150 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1151 	    &rack_hw_rwnd_factor, 2,
1152 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1153 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1154 	    SYSCTL_CHILDREN(rack_hw_pacing),
1155 	    OID_AUTO, "precheck", CTLFLAG_RW,
1156 	    &rack_hw_check_queue, 0,
1157 	    "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1158 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1159 	    SYSCTL_CHILDREN(rack_hw_pacing),
1160 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1161 	    &rack_enobuf_hw_boost_mult, 0,
1162 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1163 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1164 	    SYSCTL_CHILDREN(rack_hw_pacing),
1165 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1166 	    &rack_enobuf_hw_max, 2,
1167 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1168 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1169 	    SYSCTL_CHILDREN(rack_hw_pacing),
1170 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1171 	    &rack_enobuf_hw_min, 2,
1172 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1173 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1174 	    SYSCTL_CHILDREN(rack_hw_pacing),
1175 	    OID_AUTO, "enable", CTLFLAG_RW,
1176 	    &rack_enable_hw_pacing, 0,
1177 	    "Should RACK attempt to use hw pacing?");
1178 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1179 	    SYSCTL_CHILDREN(rack_hw_pacing),
1180 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1181 	    &rack_hw_rate_caps, 0,
1182 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1183 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1184 	    SYSCTL_CHILDREN(rack_hw_pacing),
1185 	    OID_AUTO, "uncap_per", CTLFLAG_RW,
1186 	    &rack_hw_rate_cap_per, 0,
1187 	    "If you go over b/w by this amount you will be uncapped (0 = never)");
1188 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1189 	    SYSCTL_CHILDREN(rack_hw_pacing),
1190 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1191 	    &rack_hw_rate_min, 0,
1192 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1193 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1194 	    SYSCTL_CHILDREN(rack_hw_pacing),
1195 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1196 	    &rack_hw_rate_to_low, 0,
1197 	    "If we fall below this rate, dis-engage hw pacing?");
1198 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1199 	    SYSCTL_CHILDREN(rack_hw_pacing),
1200 	    OID_AUTO, "up_only", CTLFLAG_RW,
1201 	    &rack_hw_up_only, 0,
1202 	    "Do we allow hw pacing to lower the rate selected?");
1203 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1204 	    SYSCTL_CHILDREN(rack_hw_pacing),
1205 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1206 	    &rack_hw_pace_extra_slots, 0,
1207 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1208 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1209 	    SYSCTL_CHILDREN(rack_sysctl_root),
1210 	    OID_AUTO,
1211 	    "timely",
1212 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1213 	    "Rack Timely RTT Controls");
1214 	/* Timely based GP dynmics */
1215 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1216 	    SYSCTL_CHILDREN(rack_timely),
1217 	    OID_AUTO, "upper", CTLFLAG_RW,
1218 	    &rack_gp_per_bw_mul_up, 2,
1219 	    "Rack timely upper range for equal b/w (in percentage)");
1220 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1221 	    SYSCTL_CHILDREN(rack_timely),
1222 	    OID_AUTO, "lower", CTLFLAG_RW,
1223 	    &rack_gp_per_bw_mul_down, 4,
1224 	    "Rack timely lower range for equal b/w (in percentage)");
1225 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1226 	    SYSCTL_CHILDREN(rack_timely),
1227 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1228 	    &rack_gp_rtt_maxmul, 3,
1229 	    "Rack timely multiplier of lowest rtt for rtt_max");
1230 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1231 	    SYSCTL_CHILDREN(rack_timely),
1232 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1233 	    &rack_gp_rtt_mindiv, 4,
1234 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1235 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1236 	    SYSCTL_CHILDREN(rack_timely),
1237 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1238 	    &rack_gp_rtt_minmul, 1,
1239 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1240 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1241 	    SYSCTL_CHILDREN(rack_timely),
1242 	    OID_AUTO, "decrease", CTLFLAG_RW,
1243 	    &rack_gp_decrease_per, 80,
1244 	    "Rack timely Beta value 80 = .8 (scaled by 100)");
1245 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1246 	    SYSCTL_CHILDREN(rack_timely),
1247 	    OID_AUTO, "increase", CTLFLAG_RW,
1248 	    &rack_gp_increase_per, 2,
1249 	    "Rack timely increase perentage of our GP multiplication factor");
1250 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1251 	    SYSCTL_CHILDREN(rack_timely),
1252 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1253 	    &rack_per_lower_bound, 50,
1254 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1255 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1256 	    SYSCTL_CHILDREN(rack_timely),
1257 	    OID_AUTO, "p5_upper", CTLFLAG_RW,
1258 	    &rack_gain_p5_ub, 250,
1259 	    "Profile 5 upper bound to timely gain");
1260 
1261 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1262 	    SYSCTL_CHILDREN(rack_timely),
1263 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1264 	    &rack_per_upper_bound_ss, 0,
1265 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1266 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1267 	    SYSCTL_CHILDREN(rack_timely),
1268 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1269 	    &rack_per_upper_bound_ca, 0,
1270 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1271 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1272 	    SYSCTL_CHILDREN(rack_timely),
1273 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1274 	    &rack_do_dyn_mul, 0,
1275 	    "Rack timely do we enable dynmaic timely goodput by default");
1276 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1277 	    SYSCTL_CHILDREN(rack_timely),
1278 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1279 	    &rack_gp_no_rec_chg, 1,
1280 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1281 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1282 	    SYSCTL_CHILDREN(rack_timely),
1283 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1284 	    &rack_timely_dec_clear, 6,
1285 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1286 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1287 	    SYSCTL_CHILDREN(rack_timely),
1288 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1289 	    &rack_timely_max_push_rise, 3,
1290 	    "Rack timely how many times do we push up with b/w increase");
1291 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1292 	    SYSCTL_CHILDREN(rack_timely),
1293 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1294 	    &rack_timely_max_push_drop, 3,
1295 	    "Rack timely how many times do we push back on b/w decent");
1296 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1297 	    SYSCTL_CHILDREN(rack_timely),
1298 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1299 	    &rack_timely_min_segs, 4,
1300 	    "Rack timely when setting the cwnd what is the min num segments");
1301 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1302 	    SYSCTL_CHILDREN(rack_timely),
1303 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1304 	    &rack_use_max_for_nobackoff, 0,
1305 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1306 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1307 	    SYSCTL_CHILDREN(rack_timely),
1308 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1309 	    &rack_timely_int_timely_only, 0,
1310 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1311 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1312 	    SYSCTL_CHILDREN(rack_timely),
1313 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1314 	    &rack_timely_no_stopping, 0,
1315 	    "Rack timely don't stop increase");
1316 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1317 	    SYSCTL_CHILDREN(rack_timely),
1318 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1319 	    &rack_down_raise_thresh, 100,
1320 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1321 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1322 	    SYSCTL_CHILDREN(rack_timely),
1323 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1324 	    &rack_req_segs, 1,
1325 	    "Bottom dragging if not these many segments outstanding and room");
1326 
1327 	/* TLP and Rack related parameters */
1328 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1329 	    SYSCTL_CHILDREN(rack_sysctl_root),
1330 	    OID_AUTO,
1331 	    "tlp",
1332 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1333 	    "TLP and Rack related Controls");
1334 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1335 	    SYSCTL_CHILDREN(rack_tlp),
1336 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1337 	    &use_rack_rr, 1,
1338 	    "Do we use Rack Rapid Recovery");
1339 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1340 	    SYSCTL_CHILDREN(rack_tlp),
1341 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1342 	    &rack_max_abc_post_recovery, 2,
1343 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1344 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1345 	    SYSCTL_CHILDREN(rack_tlp),
1346 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1347 	    &rack_non_rxt_use_cr, 0,
1348 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1349 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1350 	    SYSCTL_CHILDREN(rack_tlp),
1351 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1352 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1353 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1354 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1355 	    SYSCTL_CHILDREN(rack_tlp),
1356 	    OID_AUTO, "limit", CTLFLAG_RW,
1357 	    &rack_tlp_limit, 2,
1358 	    "How many TLP's can be sent without sending new data");
1359 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1360 	    SYSCTL_CHILDREN(rack_tlp),
1361 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1362 	    &rack_tlp_use_greater, 1,
1363 	    "Should we use the rack_rtt time if its greater than srtt");
1364 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1365 	    SYSCTL_CHILDREN(rack_tlp),
1366 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1367 	    &rack_tlp_min, 10000,
1368 	    "TLP minimum timeout per the specification (in microseconds)");
1369 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1370 	    SYSCTL_CHILDREN(rack_tlp),
1371 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1372 	    &rack_always_send_oldest, 0,
1373 	    "Should we always send the oldest TLP and RACK-TLP");
1374 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1375 	    SYSCTL_CHILDREN(rack_tlp),
1376 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1377 	    &rack_lower_cwnd_at_tlp, 0,
1378 	    "When a TLP completes a retran should we enter recovery");
1379 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1380 	    SYSCTL_CHILDREN(rack_tlp),
1381 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1382 	    &rack_reorder_thresh, 2,
1383 	    "What factor for rack will be added when seeing reordering (shift right)");
1384 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1385 	    SYSCTL_CHILDREN(rack_tlp),
1386 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1387 	    &rack_tlp_thresh, 1,
1388 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1389 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1390 	    SYSCTL_CHILDREN(rack_tlp),
1391 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1392 	    &rack_reorder_fade, 60000000,
1393 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1394 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1395 	    SYSCTL_CHILDREN(rack_tlp),
1396 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1397 	    &rack_pkt_delay, 1000,
1398 	    "Extra RACK time (in microseconds) besides reordering thresh");
1399 
1400 	/* Timer related controls */
1401 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1402 	    SYSCTL_CHILDREN(rack_sysctl_root),
1403 	    OID_AUTO,
1404 	    "timers",
1405 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1406 	    "Timer related controls");
1407 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1408 	    SYSCTL_CHILDREN(rack_timers),
1409 	    OID_AUTO, "reset_ssth_rec_rto", CTLFLAG_RW,
1410 	    &rack_ssthresh_rest_rto_rec, 0,
1411 	    "When doing recovery -> rto -> recovery do we reset SSthresh?");
1412 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1413 	    SYSCTL_CHILDREN(rack_timers),
1414 	    OID_AUTO, "scoreboard_thresh", CTLFLAG_RW,
1415 	    &rack_rxt_scoreboard_clear_thresh, 2,
1416 	    "How many RTO's are allowed before we clear the scoreboard");
1417 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1418 	    SYSCTL_CHILDREN(rack_timers),
1419 	    OID_AUTO, "honor_hpts_min", CTLFLAG_RW,
1420 	    &rack_honors_hpts_min_to, 1,
1421 	    "Do rack pacing timers honor hpts min timeout");
1422 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1423 	    SYSCTL_CHILDREN(rack_timers),
1424 	    OID_AUTO, "hpts_max_reduce", CTLFLAG_RW,
1425 	    &rack_max_reduce, 10,
1426 	    "Max percentage we will reduce slot by for pacing when we are behind");
1427 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1428 	    SYSCTL_CHILDREN(rack_timers),
1429 	    OID_AUTO, "persmin", CTLFLAG_RW,
1430 	    &rack_persist_min, 250000,
1431 	    "What is the minimum time in microseconds between persists");
1432 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1433 	    SYSCTL_CHILDREN(rack_timers),
1434 	    OID_AUTO, "persmax", CTLFLAG_RW,
1435 	    &rack_persist_max, 2000000,
1436 	    "What is the largest delay in microseconds between persists");
1437 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1438 	    SYSCTL_CHILDREN(rack_timers),
1439 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1440 	    &rack_delayed_ack_time, 40000,
1441 	    "Delayed ack time (40ms in microseconds)");
1442 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1443 	    SYSCTL_CHILDREN(rack_timers),
1444 	    OID_AUTO, "minrto", CTLFLAG_RW,
1445 	    &rack_rto_min, 30000,
1446 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1447 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1448 	    SYSCTL_CHILDREN(rack_timers),
1449 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1450 	    &rack_rto_max, 4000000,
1451 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1452 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1453 	    SYSCTL_CHILDREN(rack_timers),
1454 	    OID_AUTO, "minto", CTLFLAG_RW,
1455 	    &rack_min_to, 1000,
1456 	    "Minimum rack timeout in microseconds");
1457 	/* Measure controls */
1458 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1459 	    SYSCTL_CHILDREN(rack_sysctl_root),
1460 	    OID_AUTO,
1461 	    "measure",
1462 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1463 	    "Measure related controls");
1464 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1465 	    SYSCTL_CHILDREN(rack_measure),
1466 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1467 	    &rack_wma_divisor, 8,
1468 	    "When doing b/w calculation what is the  divisor for the WMA");
1469 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1470 	    SYSCTL_CHILDREN(rack_measure),
1471 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1472 	    &rack_cwnd_block_ends_measure, 0,
1473 	    "Does a cwnd just-return end the measurement window (app limited)");
1474 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1475 	    SYSCTL_CHILDREN(rack_measure),
1476 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1477 	    &rack_rwnd_block_ends_measure, 0,
1478 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1479 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1480 	    SYSCTL_CHILDREN(rack_measure),
1481 	    OID_AUTO, "min_target", CTLFLAG_RW,
1482 	    &rack_def_data_window, 20,
1483 	    "What is the minimum target window (in mss) for a GP measurements");
1484 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1485 	    SYSCTL_CHILDREN(rack_measure),
1486 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1487 	    &rack_goal_bdp, 2,
1488 	    "What is the goal BDP to measure");
1489 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1490 	    SYSCTL_CHILDREN(rack_measure),
1491 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1492 	    &rack_min_srtts, 1,
1493 	    "What is the goal BDP to measure");
1494 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1495 	    SYSCTL_CHILDREN(rack_measure),
1496 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1497 	    &rack_min_measure_usec, 0,
1498 	    "What is the Minimum time time for a measurement if 0, this is off");
1499 	/* Features */
1500 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1501 	    SYSCTL_CHILDREN(rack_sysctl_root),
1502 	    OID_AUTO,
1503 	    "features",
1504 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1505 	    "Feature controls");
1506 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1507 	    SYSCTL_CHILDREN(rack_features),
1508 	    OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1509 	    &rack_hybrid_allow_set_maxseg, 0,
1510 	    "Should hybrid pacing allow the setmss command");
1511 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1512 	    SYSCTL_CHILDREN(rack_features),
1513 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1514 	    &rack_use_cmp_acks, 1,
1515 	    "Should RACK have LRO send compressed acks");
1516 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1517 	    SYSCTL_CHILDREN(rack_features),
1518 	    OID_AUTO, "fsb", CTLFLAG_RW,
1519 	    &rack_use_fsb, 1,
1520 	    "Should RACK use the fast send block?");
1521 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1522 	    SYSCTL_CHILDREN(rack_features),
1523 	    OID_AUTO, "rfo", CTLFLAG_RW,
1524 	    &rack_use_rfo, 1,
1525 	    "Should RACK use rack_fast_output()?");
1526 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1527 	    SYSCTL_CHILDREN(rack_features),
1528 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1529 	    &rack_use_rsm_rfo, 1,
1530 	    "Should RACK use rack_fast_rsm_output()?");
1531 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1532 	    SYSCTL_CHILDREN(rack_features),
1533 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1534 	    &rack_enable_mqueue_for_nonpaced, 0,
1535 	    "Should RACK use mbuf queuing for non-paced connections");
1536 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1537 	    SYSCTL_CHILDREN(rack_features),
1538 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1539 	    &rack_do_hystart, 0,
1540 	    "Should RACK enable HyStart++ on connections?");
1541 	/* Misc rack controls */
1542 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1543 	    SYSCTL_CHILDREN(rack_sysctl_root),
1544 	    OID_AUTO,
1545 	    "misc",
1546 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1547 	    "Misc related controls");
1548 #ifdef TCP_ACCOUNTING
1549 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1550 	    SYSCTL_CHILDREN(rack_misc),
1551 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1552 	    &rack_tcp_accounting, 0,
1553 	    "Should we turn on TCP accounting for all rack sessions?");
1554 #endif
1555 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1556 	    SYSCTL_CHILDREN(rack_misc),
1557 	    OID_AUTO, "dnd", CTLFLAG_RW,
1558 	    &rack_dnd_default, 0,
1559 	    "Do not disturb default for rack_rrr = 3");
1560 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1561 	    SYSCTL_CHILDREN(rack_misc),
1562 	    OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1563 	    &sad_seg_size_per, 800,
1564 	    "Percentage of segment size needed in a sack 800 = 80.0?");
1565 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1566 	    SYSCTL_CHILDREN(rack_misc),
1567 	    OID_AUTO, "rxt_controls", CTLFLAG_RW,
1568 	    &rack_rxt_controls, 0,
1569 	    "Retransmit sending size controls (valid  values 0, 1, 2 default=1)?");
1570 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1571 	    SYSCTL_CHILDREN(rack_misc),
1572 	    OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1573 	    &rack_hibeta_setting, 0,
1574 	    "Do we ue a high beta (80 instead of 50)?");
1575 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1576 	    SYSCTL_CHILDREN(rack_misc),
1577 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1578 	    &rack_apply_rtt_with_reduced_conf, 0,
1579 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1580 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1581 	    SYSCTL_CHILDREN(rack_misc),
1582 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1583 	    &rack_dsack_std_based, 3,
1584 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1585 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1586 	    SYSCTL_CHILDREN(rack_misc),
1587 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1588 	    &rack_prr_addbackmax, 2,
1589 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1590 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1591 	    SYSCTL_CHILDREN(rack_misc),
1592 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1593 	    &rack_stats_gets_ms_rtt, 1,
1594 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1595 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1596 	    SYSCTL_CHILDREN(rack_misc),
1597 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1598 	    &rack_client_low_buf, 0,
1599 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1600 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1601 	    SYSCTL_CHILDREN(rack_misc),
1602 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1603 	    &rack_def_profile, 0,
1604 	    "Should RACK use a default profile (0=no, num == profile num)?");
1605 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1606 	    SYSCTL_CHILDREN(rack_misc),
1607 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1608 	    &rack_enable_shared_cwnd, 1,
1609 	    "Should RACK try to use the shared cwnd on connections where allowed");
1610 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1611 	    SYSCTL_CHILDREN(rack_misc),
1612 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1613 	    &rack_limits_scwnd, 1,
1614 	    "Should RACK place low end time limits on the shared cwnd feature");
1615 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1616 	    SYSCTL_CHILDREN(rack_misc),
1617 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1618 	    &rack_disable_prr, 0,
1619 	    "Should RACK not use prr and only pace (must have pacing on)");
1620 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1621 	    SYSCTL_CHILDREN(rack_misc),
1622 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1623 	    &rack_verbose_logging, 0,
1624 	    "Should RACK black box logging be verbose");
1625 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1626 	    SYSCTL_CHILDREN(rack_misc),
1627 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1628 	    &rack_ignore_data_after_close, 1,
1629 	    "Do we hold off sending a RST until all pending data is ack'd");
1630 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1631 	    SYSCTL_CHILDREN(rack_misc),
1632 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1633 	    &rack_sack_not_required, 1,
1634 	    "Do we allow rack to run on connections not supporting SACK");
1635 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1636 	    SYSCTL_CHILDREN(rack_misc),
1637 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1638 	    &rack_send_a_lot_in_prr, 1,
1639 	    "Send a lot in prr");
1640 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1641 	    SYSCTL_CHILDREN(rack_misc),
1642 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1643 	    &rack_autosndbuf_inc, 20,
1644 	    "What percentage should rack scale up its snd buffer by?");
1645 
1646 
1647 	/* Sack Attacker detection stuff */
1648 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1649 	    SYSCTL_CHILDREN(rack_attack),
1650 	    OID_AUTO, "merge_out", CTLFLAG_RW,
1651 	    &rack_merge_out_sacks_on_attack, 0,
1652 	    "Do we merge the sendmap when we decide we are being attacked?");
1653 
1654 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1655 	    SYSCTL_CHILDREN(rack_attack),
1656 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1657 	    &rack_highest_sack_thresh_seen, 0,
1658 	    "Highest sack to ack ratio seen");
1659 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1660 	    SYSCTL_CHILDREN(rack_attack),
1661 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1662 	    &rack_highest_move_thresh_seen, 0,
1663 	    "Highest move to non-move ratio seen");
1664 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1665 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1666 	    SYSCTL_CHILDREN(rack_attack),
1667 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1668 	    &rack_ack_total,
1669 	    "Total number of Ack's");
1670 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1671 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1672 	    SYSCTL_CHILDREN(rack_attack),
1673 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1674 	    &rack_express_sack,
1675 	    "Total expresss number of Sack's");
1676 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1677 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1678 	    SYSCTL_CHILDREN(rack_attack),
1679 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1680 	    &rack_sack_total,
1681 	    "Total number of SACKs");
1682 	rack_move_none = counter_u64_alloc(M_WAITOK);
1683 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1684 	    SYSCTL_CHILDREN(rack_attack),
1685 	    OID_AUTO, "move_none", CTLFLAG_RD,
1686 	    &rack_move_none,
1687 	    "Total number of SACK index reuse of positions under threshold");
1688 	rack_move_some = counter_u64_alloc(M_WAITOK);
1689 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1690 	    SYSCTL_CHILDREN(rack_attack),
1691 	    OID_AUTO, "move_some", CTLFLAG_RD,
1692 	    &rack_move_some,
1693 	    "Total number of SACK index reuse of positions over threshold");
1694 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1695 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1696 	    SYSCTL_CHILDREN(rack_attack),
1697 	    OID_AUTO, "attacks", CTLFLAG_RD,
1698 	    &rack_sack_attacks_detected,
1699 	    "Total number of SACK attackers that had sack disabled");
1700 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1701 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1702 	    SYSCTL_CHILDREN(rack_attack),
1703 	    OID_AUTO, "reversed", CTLFLAG_RD,
1704 	    &rack_sack_attacks_reversed,
1705 	    "Total number of SACK attackers that were later determined false positive");
1706 	rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1707 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1708 	    SYSCTL_CHILDREN(rack_attack),
1709 	    OID_AUTO, "suspect", CTLFLAG_RD,
1710 	    &rack_sack_attacks_suspect,
1711 	    "Total number of SACKs that triggered early detection");
1712 
1713 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1714 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1715 	    SYSCTL_CHILDREN(rack_attack),
1716 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1717 	    &rack_sack_used_next_merge,
1718 	    "Total number of times we used the next merge");
1719 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1720 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1721 	    SYSCTL_CHILDREN(rack_attack),
1722 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1723 	    &rack_sack_used_prev_merge,
1724 	    "Total number of times we used the prev merge");
1725 	/* Counters */
1726 	rack_total_bytes = counter_u64_alloc(M_WAITOK);
1727 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1728 	    SYSCTL_CHILDREN(rack_counters),
1729 	    OID_AUTO, "totalbytes", CTLFLAG_RD,
1730 	    &rack_total_bytes,
1731 	    "Total number of bytes sent");
1732 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1733 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1734 	    SYSCTL_CHILDREN(rack_counters),
1735 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1736 	    &rack_fto_send, "Total number of rack_fast_output sends");
1737 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1738 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1739 	    SYSCTL_CHILDREN(rack_counters),
1740 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1741 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1742 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1743 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744 	    SYSCTL_CHILDREN(rack_counters),
1745 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1746 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1747 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1748 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1749 	    SYSCTL_CHILDREN(rack_counters),
1750 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1751 	    &rack_non_fto_send, "Total number of rack_output first sends");
1752 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1753 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1754 	    SYSCTL_CHILDREN(rack_counters),
1755 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1756 	    &rack_extended_rfo, "Total number of times we extended rfo");
1757 
1758 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1759 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1760 	    SYSCTL_CHILDREN(rack_counters),
1761 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1762 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1763 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1764 
1765 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1766 	    SYSCTL_CHILDREN(rack_counters),
1767 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1768 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1769 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1770 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1771 	    SYSCTL_CHILDREN(rack_counters),
1772 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1773 	    &rack_tlp_tot,
1774 	    "Total number of tail loss probe expirations");
1775 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1776 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1777 	    SYSCTL_CHILDREN(rack_counters),
1778 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1779 	    &rack_tlp_newdata,
1780 	    "Total number of tail loss probe sending new data");
1781 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1782 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1783 	    SYSCTL_CHILDREN(rack_counters),
1784 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1785 	    &rack_tlp_retran,
1786 	    "Total number of tail loss probe sending retransmitted data");
1787 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1788 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1789 	    SYSCTL_CHILDREN(rack_counters),
1790 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1791 	    &rack_tlp_retran_bytes,
1792 	    "Total bytes of tail loss probe sending retransmitted data");
1793 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1794 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1795 	    SYSCTL_CHILDREN(rack_counters),
1796 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1797 	    &rack_to_tot,
1798 	    "Total number of times the rack to expired");
1799 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1800 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1801 	    SYSCTL_CHILDREN(rack_counters),
1802 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1803 	    &rack_saw_enobuf,
1804 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1805 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1806 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1807 	    SYSCTL_CHILDREN(rack_counters),
1808 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1809 	    &rack_saw_enobuf_hw,
1810 	    "Total number of times a send returned enobuf for hdwr paced connections");
1811 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1812 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1813 	    SYSCTL_CHILDREN(rack_counters),
1814 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1815 	    &rack_saw_enetunreach,
1816 	    "Total number of times a send received a enetunreachable");
1817 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1818 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1819 	    SYSCTL_CHILDREN(rack_counters),
1820 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1821 	    &rack_hot_alloc,
1822 	    "Total allocations from the top of our list");
1823 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1824 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1825 	    SYSCTL_CHILDREN(rack_counters),
1826 	    OID_AUTO, "allocs", CTLFLAG_RD,
1827 	    &rack_to_alloc,
1828 	    "Total allocations of tracking structures");
1829 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1830 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1831 	    SYSCTL_CHILDREN(rack_counters),
1832 	    OID_AUTO, "allochard", CTLFLAG_RD,
1833 	    &rack_to_alloc_hard,
1834 	    "Total allocations done with sleeping the hard way");
1835 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1836 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1837 	    SYSCTL_CHILDREN(rack_counters),
1838 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1839 	    &rack_to_alloc_emerg,
1840 	    "Total allocations done from emergency cache");
1841 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1842 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1843 	    SYSCTL_CHILDREN(rack_counters),
1844 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1845 	    &rack_to_alloc_limited,
1846 	    "Total allocations dropped due to limit");
1847 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1848 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1849 	    SYSCTL_CHILDREN(rack_counters),
1850 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1851 	    &rack_alloc_limited_conns,
1852 	    "Connections with allocations dropped due to limit");
1853 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1854 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1855 	    SYSCTL_CHILDREN(rack_counters),
1856 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1857 	    &rack_split_limited,
1858 	    "Split allocations dropped due to limit");
1859 	rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1860 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1861 	    SYSCTL_CHILDREN(rack_counters),
1862 	    OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1863 	    &rack_rxt_clamps_cwnd,
1864 	    "Number of times that excessive rxt clamped the cwnd down");
1865 	rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1866 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1867 	    SYSCTL_CHILDREN(rack_counters),
1868 	    OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1869 	    &rack_rxt_clamps_cwnd_uniq,
1870 	    "Number of connections that have had excessive rxt clamped the cwnd down");
1871 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1872 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1873 	    SYSCTL_CHILDREN(rack_counters),
1874 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1875 	    &rack_persists_sends,
1876 	    "Number of times we sent a persist probe");
1877 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1878 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1879 	    SYSCTL_CHILDREN(rack_counters),
1880 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1881 	    &rack_persists_acks,
1882 	    "Number of times a persist probe was acked");
1883 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1884 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1885 	    SYSCTL_CHILDREN(rack_counters),
1886 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1887 	    &rack_persists_loss,
1888 	    "Number of times we detected a lost persist probe (no ack)");
1889 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1890 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1891 	    SYSCTL_CHILDREN(rack_counters),
1892 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1893 	    &rack_persists_lost_ends,
1894 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1895 #ifdef INVARIANTS
1896 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1897 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1898 	    SYSCTL_CHILDREN(rack_counters),
1899 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1900 	    &rack_adjust_map_bw,
1901 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1902 #endif
1903 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1904 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1905 	    SYSCTL_CHILDREN(rack_counters),
1906 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1907 	    &rack_multi_single_eq,
1908 	    "Number of compressed acks total represented");
1909 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1910 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1911 	    SYSCTL_CHILDREN(rack_counters),
1912 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1913 	    &rack_proc_non_comp_ack,
1914 	    "Number of non compresseds acks that we processed");
1915 
1916 
1917 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1918 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1919 	    SYSCTL_CHILDREN(rack_counters),
1920 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1921 	    &rack_sack_proc_all,
1922 	    "Total times we had to walk whole list for sack processing");
1923 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1924 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1925 	    SYSCTL_CHILDREN(rack_counters),
1926 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1927 	    &rack_sack_proc_restart,
1928 	    "Total times we had to walk whole list due to a restart");
1929 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1930 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1931 	    SYSCTL_CHILDREN(rack_counters),
1932 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1933 	    &rack_sack_proc_short,
1934 	    "Total times we took shortcut for sack processing");
1935 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1936 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1937 	    SYSCTL_CHILDREN(rack_attack),
1938 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1939 	    &rack_sack_skipped_acked,
1940 	    "Total number of times we skipped previously sacked");
1941 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1942 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1943 	    SYSCTL_CHILDREN(rack_attack),
1944 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1945 	    &rack_sack_splits,
1946 	    "Total number of times we did the old fashion tree split");
1947 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1948 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1949 	    SYSCTL_CHILDREN(rack_counters),
1950 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1951 	    &rack_input_idle_reduces,
1952 	    "Total number of idle reductions on input");
1953 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1954 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1955 	    SYSCTL_CHILDREN(rack_counters),
1956 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1957 	    &rack_collapsed_win_seen,
1958 	    "Total number of collapsed window events seen (where our window shrinks)");
1959 
1960 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1961 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1962 	    SYSCTL_CHILDREN(rack_counters),
1963 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1964 	    &rack_collapsed_win,
1965 	    "Total number of collapsed window events where we mark packets");
1966 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1967 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1968 	    SYSCTL_CHILDREN(rack_counters),
1969 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1970 	    &rack_collapsed_win_rxt,
1971 	    "Total number of packets that were retransmitted");
1972 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1973 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1974 	    SYSCTL_CHILDREN(rack_counters),
1975 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1976 	    &rack_collapsed_win_rxt_bytes,
1977 	    "Total number of bytes that were retransmitted");
1978 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1979 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1980 	    SYSCTL_CHILDREN(rack_counters),
1981 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1982 	    &rack_try_scwnd,
1983 	    "Total number of scwnd attempts");
1984 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1985 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1986 	    OID_AUTO, "outsize", CTLFLAG_RD,
1987 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1988 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1989 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1990 	    OID_AUTO, "opts", CTLFLAG_RD,
1991 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1992 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1993 	    SYSCTL_CHILDREN(rack_sysctl_root),
1994 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1995 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1996 }
1997 
1998 static uint32_t
1999 rc_init_window(struct tcp_rack *rack)
2000 {
2001 	return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
2002 
2003 }
2004 
2005 static uint64_t
2006 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
2007 {
2008 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
2009 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
2010 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2011 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
2012 	else
2013 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
2014 }
2015 
2016 static void
2017 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
2018 	uint64_t data, uint8_t mod, uint16_t aux,
2019 	struct tcp_sendfile_track *cur, int line)
2020 {
2021 #ifdef TCP_REQUEST_TRK
2022 	int do_log = 0;
2023 
2024 	/*
2025 	 * The rate cap one is noisy and only should come out when normal BB logging
2026 	 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
2027 	 * once per chunk and make up the BBpoint that can be turned on by the client.
2028 	 */
2029 	if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2030 		/*
2031 		 * The very noisy two need to only come out when
2032 		 * we have verbose logging on.
2033 		 */
2034 		if (rack_verbose_logging != 0)
2035 			do_log = tcp_bblogging_on(rack->rc_tp);
2036 		else
2037 			do_log = 0;
2038 	} else if (mod != HYBRID_LOG_BW_MEASURE) {
2039 		/*
2040 		 * All other less noisy logs here except the measure which
2041 		 * also needs to come out on the point and the log.
2042 		 */
2043 		do_log = tcp_bblogging_on(rack->rc_tp);
2044 	} else {
2045 		do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2046 	}
2047 
2048 	if (do_log) {
2049 		union tcp_log_stackspecific log;
2050 		struct timeval tv;
2051 		uint64_t lt_bw;
2052 
2053 		/* Convert our ms to a microsecond */
2054 		memset(&log, 0, sizeof(log));
2055 
2056 		log.u_bbr.cwnd_gain = line;
2057 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2058 		log.u_bbr.rttProp = tim;
2059 		log.u_bbr.bw_inuse = cbw;
2060 		log.u_bbr.delRate = rack_get_gp_est(rack);
2061 		lt_bw = rack_get_lt_bw(rack);
2062 		log.u_bbr.flex1 = seq;
2063 		log.u_bbr.pacing_gain = aux;
2064 		/* lt_bw = < flex3 | flex2 > */
2065 		log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2066 		log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2067 		/* Record the last obtained us rtt in inflight */
2068 		if (cur == NULL) {
2069 			/* Make sure we are looking at the right log if an overide comes in */
2070 			cur = rack->r_ctl.rc_last_sft;
2071 		}
2072 		if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2073 			log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2074 		else {
2075 			/* Use the last known rtt i.e. the rack-rtt */
2076 			log.u_bbr.inflight = rack->rc_rack_rtt;
2077 		}
2078 		if (cur != NULL) {
2079 			uint64_t off;
2080 
2081 			log.u_bbr.cur_del_rate = cur->deadline;
2082 			if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2083 				/* start = < lost | pkt_epoch > */
2084 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2085 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2086 				log.u_bbr.flex6 = cur->start_seq;
2087 				log.u_bbr.pkts_out = cur->end_seq;
2088 			} else {
2089 				/* start = < lost | pkt_epoch > */
2090 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2091 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2092 				/* end = < pkts_out | flex6 > */
2093 				log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2094 				log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2095 			}
2096 			/* first_send = <lt_epoch | epoch> */
2097 			log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2098 			log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2099 			/* localtime = <delivered | applimited>*/
2100 			log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2101 			log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2102 #ifdef TCP_REQUEST_TRK
2103 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2104 			log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2105 #endif
2106 			log.u_bbr.inhpts = 1;
2107 			log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2108 			log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2109 			log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2110 		} else {
2111 			log.u_bbr.flex7 = 0xffff;
2112 			log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2113 		}
2114 		/*
2115 		 * Compose bbr_state to be a bit wise 0000ADHF
2116 		 * where A is the always_pace flag
2117 		 * where D is the dgp_on flag
2118 		 * where H is the hybrid_mode on flag
2119 		 * where F is the use_fixed_rate flag.
2120 		 */
2121 		log.u_bbr.bbr_state = rack->rc_always_pace;
2122 		log.u_bbr.bbr_state <<= 1;
2123 		log.u_bbr.bbr_state |= rack->dgp_on;
2124 		log.u_bbr.bbr_state <<= 1;
2125 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2126 		log.u_bbr.bbr_state <<= 1;
2127 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2128 		log.u_bbr.flex8 = mod;
2129 		tcp_log_event(rack->rc_tp, NULL,
2130 		    &rack->rc_inp->inp_socket->so_rcv,
2131 		    &rack->rc_inp->inp_socket->so_snd,
2132 		    TCP_HYBRID_PACING_LOG, 0,
2133 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2134 
2135 	}
2136 #endif
2137 }
2138 
2139 #ifdef TCP_REQUEST_TRK
2140 static void
2141 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2142 {
2143 	if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2144 		union tcp_log_stackspecific log;
2145 		struct timeval tv;
2146 		uint64_t off;
2147 
2148 		/* Convert our ms to a microsecond */
2149 		memset(&log, 0, sizeof(log));
2150 
2151 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2152 		log.u_bbr.delRate = cur->sent_at_fs;
2153 
2154 		if ((cur->flags & TCP_TRK_TRACK_FLG_LSND) == 0) {
2155 			/*
2156 			 * We did not get a new Rules Applied to set so
2157 			 * no overlapping send occured, this means the
2158 			 * current byte counts are correct.
2159 			 */
2160 			log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2161 			log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2162 		} else {
2163 			/*
2164 			 * Overlapping send case, we switched to a new
2165 			 * send and did a rules applied.
2166 			 */
2167 			log.u_bbr.cur_del_rate = cur->sent_at_ls;
2168 			log.u_bbr.rttProp = cur->rxt_at_ls;
2169 		}
2170 		log.u_bbr.bw_inuse = cur->rxt_at_fs;
2171 		log.u_bbr.cwnd_gain = line;
2172 		off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2173 		log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2174 		/* start = < flex1 | flex2 > */
2175 		log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2176 		log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2177 		/* end = < flex3 | flex4 > */
2178 		log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2179 		log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2180 
2181 		/* localtime = <delivered | applimited>*/
2182 		log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2183 		log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2184 		/* client timestamp = <lt_epoch | epoch>*/
2185 		log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2186 		log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2187 		/* now set all the flags in */
2188 		log.u_bbr.pkts_out = cur->hybrid_flags;
2189 		log.u_bbr.lost = cur->playout_ms;
2190 		log.u_bbr.flex6 = cur->flags;
2191 		/*
2192 		 * Last send time  = <flex5 | pkt_epoch>  note we do not distinguish cases
2193 		 * where a false retransmit occurred so first_send  <-> lastsend may
2194 		 * include longer time then it actually took if we have a false rxt.
2195 		 */
2196 		log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2197 		log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2198 		/*
2199 		 * Compose bbr_state to be a bit wise 0000ADHF
2200 		 * where A is the always_pace flag
2201 		 * where D is the dgp_on flag
2202 		 * where H is the hybrid_mode on flag
2203 		 * where F is the use_fixed_rate flag.
2204 		 */
2205 		log.u_bbr.bbr_state = rack->rc_always_pace;
2206 		log.u_bbr.bbr_state <<= 1;
2207 		log.u_bbr.bbr_state |= rack->dgp_on;
2208 		log.u_bbr.bbr_state <<= 1;
2209 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2210 		log.u_bbr.bbr_state <<= 1;
2211 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2212 
2213 		log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2214 		tcp_log_event(rack->rc_tp, NULL,
2215 		    &rack->rc_inp->inp_socket->so_rcv,
2216 		    &rack->rc_inp->inp_socket->so_snd,
2217 		    TCP_HYBRID_PACING_LOG, 0,
2218 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2219 	}
2220 }
2221 #endif
2222 
2223 static inline uint64_t
2224 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2225 {
2226 	uint64_t ret_bw, ether;
2227 	uint64_t u_segsiz;
2228 
2229 	ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2230 	if (rack->r_is_v6){
2231 #ifdef INET6
2232 		ether += sizeof(struct ip6_hdr);
2233 #endif
2234 		ether += 14;	/* eheader size 6+6+2 */
2235 	} else {
2236 #ifdef INET
2237 		ether += sizeof(struct ip);
2238 #endif
2239 		ether += 14;	/* eheader size 6+6+2 */
2240 	}
2241 	u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2242 	ret_bw = bw;
2243 	ret_bw *= ether;
2244 	ret_bw /= u_segsiz;
2245 	return (ret_bw);
2246 }
2247 
2248 static void
2249 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2250 {
2251 #ifdef TCP_REQUEST_TRK
2252 	struct timeval tv;
2253 	uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2254 #endif
2255 
2256 	if (rack->r_ctl.bw_rate_cap == 0)
2257 		return;
2258 #ifdef TCP_REQUEST_TRK
2259 	if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2260 	    (rack->r_ctl.rc_last_sft != NULL)) {
2261 		/*
2262 		 * We have a dynamic cap. The original target
2263 		 * is in bw_rate_cap, but we need to look at
2264 		 * how long it is until we hit the deadline.
2265 		 */
2266 		struct tcp_sendfile_track *ent;
2267 
2268       		ent = rack->r_ctl.rc_last_sft;
2269 		microuptime(&tv);
2270 		timenow = tcp_tv_to_lusectick(&tv);
2271 		if (timenow >= ent->deadline) {
2272 			/* No time left we do DGP only */
2273 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2274 					   0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2275 			rack->r_ctl.bw_rate_cap = 0;
2276 			return;
2277 		}
2278 		/* We have the time */
2279 		timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2280 		if (timeleft < HPTS_MSEC_IN_SEC) {
2281 			/* If there is less than a ms left just use DGPs rate */
2282 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2283 					   0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2284 			rack->r_ctl.bw_rate_cap = 0;
2285 			return;
2286 		}
2287 		/*
2288 		 * Now lets find the amount of data left to send.
2289 		 *
2290 		 * Now ideally we want to use the end_seq to figure out how much more
2291 		 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2292 		 */
2293 		if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2294 			if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2295 				lenleft = ent->end_seq - rack->rc_tp->snd_una;
2296 			else {
2297 				/* TSNH, we should catch it at the send */
2298 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2299 						   0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2300 				rack->r_ctl.bw_rate_cap = 0;
2301 				return;
2302 			}
2303 		} else {
2304 			/*
2305 			 * The hard way, figure out how much is gone and then
2306 			 * take that away from the total the client asked for
2307 			 * (thats off by tls overhead if this is tls).
2308 			 */
2309 			if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2310 				lengone = rack->rc_tp->snd_una - ent->start_seq;
2311 			else
2312 				lengone = 0;
2313 			if (lengone < (ent->end - ent->start))
2314 				lenleft = (ent->end - ent->start) - lengone;
2315 			else {
2316 				/* TSNH, we should catch it at the send */
2317 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2318 						   0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2319 				rack->r_ctl.bw_rate_cap = 0;
2320 				return;
2321 			}
2322 		}
2323 		if (lenleft == 0) {
2324 			/* We have it all sent */
2325 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2326 					   0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2327 			if (rack->r_ctl.bw_rate_cap)
2328 				goto normal_ratecap;
2329 			else
2330 				return;
2331 		}
2332 		calcbw = lenleft * HPTS_USEC_IN_SEC;
2333 		calcbw /= timeleft;
2334 		/* Now we must compensate for IP/TCP overhead */
2335 		calcbw = rack_compensate_for_linerate(rack, calcbw);
2336 		/* Update the bit rate cap */
2337 		rack->r_ctl.bw_rate_cap = calcbw;
2338 		if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2339 		    (rack_hybrid_allow_set_maxseg == 1) &&
2340 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2341 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2342 			uint32_t orig_max;
2343 
2344 			orig_max = rack->r_ctl.rc_pace_max_segs;
2345 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2346 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2347 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2348 		}
2349 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2350 				   calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2351 		if ((calcbw > 0) && (*bw > calcbw)) {
2352 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2353 					   *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2354 			*capped = 1;
2355 			*bw = calcbw;
2356 		}
2357 		return;
2358 	}
2359 normal_ratecap:
2360 #endif
2361 	if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2362 #ifdef TCP_REQUEST_TRK
2363 		if (rack->rc_hybrid_mode &&
2364 		    rack->rc_catch_up &&
2365 		    (rack->r_ctl.rc_last_sft != NULL) &&
2366 		    (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2367 		    (rack_hybrid_allow_set_maxseg == 1) &&
2368 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2369 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2370 			uint32_t orig_max;
2371 
2372 			orig_max = rack->r_ctl.rc_pace_max_segs;
2373 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2374 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2375 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2376 		}
2377 #endif
2378 		*capped = 1;
2379 		*bw = rack->r_ctl.bw_rate_cap;
2380 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2381 				   *bw, 0, 0,
2382 				   HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2383 	}
2384 }
2385 
2386 static uint64_t
2387 rack_get_gp_est(struct tcp_rack *rack)
2388 {
2389 	uint64_t bw, lt_bw, ret_bw;
2390 
2391 	if (rack->rc_gp_filled == 0) {
2392 		/*
2393 		 * We have yet no b/w measurement,
2394 		 * if we have a user set initial bw
2395 		 * return it. If we don't have that and
2396 		 * we have an srtt, use the tcp IW (10) to
2397 		 * calculate a fictional b/w over the SRTT
2398 		 * which is more or less a guess. Note
2399 		 * we don't use our IW from rack on purpose
2400 		 * so if we have like IW=30, we are not
2401 		 * calculating a "huge" b/w.
2402 		 */
2403 		uint64_t srtt;
2404 
2405 		if (rack->dis_lt_bw == 1)
2406 			lt_bw = 0;
2407 		else
2408 			lt_bw = rack_get_lt_bw(rack);
2409 		if (lt_bw) {
2410 			/*
2411 			 * No goodput bw but a long-term b/w does exist
2412 			 * lets use that.
2413 			 */
2414 			ret_bw = lt_bw;
2415 			goto compensate;
2416 		}
2417 		if (rack->r_ctl.init_rate)
2418 			return (rack->r_ctl.init_rate);
2419 
2420 		/* Ok lets come up with the IW guess, if we have a srtt */
2421 		if (rack->rc_tp->t_srtt == 0) {
2422 			/*
2423 			 * Go with old pacing method
2424 			 * i.e. burst mitigation only.
2425 			 */
2426 			return (0);
2427 		}
2428 		/* Ok lets get the initial TCP win (not racks) */
2429 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2430 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2431 		bw *= (uint64_t)USECS_IN_SECOND;
2432 		bw /= srtt;
2433 		ret_bw = bw;
2434 		goto compensate;
2435 
2436 	}
2437 	if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2438 		/* Averaging is done, we can return the value */
2439 		bw = rack->r_ctl.gp_bw;
2440 	} else {
2441 		/* Still doing initial average must calculate */
2442 		bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2443 	}
2444 	if (rack->dis_lt_bw) {
2445 		/* We are not using lt-bw */
2446 		ret_bw = bw;
2447 		goto compensate;
2448 	}
2449 	lt_bw = rack_get_lt_bw(rack);
2450 	if (lt_bw == 0) {
2451 		/* If we don't have one then equate it to the gp_bw */
2452 		lt_bw = rack->r_ctl.gp_bw;
2453 	}
2454 	if (rack->use_lesser_lt_bw) {
2455 		if (lt_bw < bw)
2456 			ret_bw = lt_bw;
2457 		else
2458 			ret_bw = bw;
2459 	} else {
2460 		if (lt_bw > bw)
2461 			ret_bw = lt_bw;
2462 		else
2463 			ret_bw = bw;
2464 	}
2465 	/*
2466 	 * Now lets compensate based on the TCP/IP overhead. Our
2467 	 * Goodput estimate does not include this so we must pace out
2468 	 * a bit faster since our pacing calculations do. The pacing
2469 	 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2470 	 * we are using to do this, so we do that here in the opposite
2471 	 * direction as well. This means that if we are tunneled and the
2472 	 * segsiz is say 1200 bytes we will get quite a boost, but its
2473 	 * compensated for in the pacing time the opposite way.
2474 	 */
2475 compensate:
2476 	ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2477 	return(ret_bw);
2478 }
2479 
2480 
2481 static uint64_t
2482 rack_get_bw(struct tcp_rack *rack)
2483 {
2484 	uint64_t bw;
2485 
2486 	if (rack->use_fixed_rate) {
2487 		/* Return the fixed pacing rate */
2488 		return (rack_get_fixed_pacing_bw(rack));
2489 	}
2490 	bw = rack_get_gp_est(rack);
2491 	return (bw);
2492 }
2493 
2494 static uint16_t
2495 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2496 {
2497 	if (rack->use_fixed_rate) {
2498 		return (100);
2499 	} else if (rack->in_probe_rtt && (rsm == NULL))
2500 		return (rack->r_ctl.rack_per_of_gp_probertt);
2501 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2502 		  rack->r_ctl.rack_per_of_gp_rec)) {
2503 		if (rsm) {
2504 			/* a retransmission always use the recovery rate */
2505 			return (rack->r_ctl.rack_per_of_gp_rec);
2506 		} else if (rack->rack_rec_nonrxt_use_cr) {
2507 			/* Directed to use the configured rate */
2508 			goto configured_rate;
2509 		} else if (rack->rack_no_prr &&
2510 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2511 			/* No PRR, lets just use the b/w estimate only */
2512 			return (100);
2513 		} else {
2514 			/*
2515 			 * Here we may have a non-retransmit but we
2516 			 * have no overrides, so just use the recovery
2517 			 * rate (prr is in effect).
2518 			 */
2519 			return (rack->r_ctl.rack_per_of_gp_rec);
2520 		}
2521 	}
2522 configured_rate:
2523 	/* For the configured rate we look at our cwnd vs the ssthresh */
2524 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2525 		return (rack->r_ctl.rack_per_of_gp_ss);
2526 	else
2527 		return (rack->r_ctl.rack_per_of_gp_ca);
2528 }
2529 
2530 static void
2531 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2532 {
2533 	/*
2534 	 * Types of logs (mod value)
2535 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2536 	 * 2 = a dsack round begins, persist is reset to 16.
2537 	 * 3 = a dsack round ends
2538 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2539 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2540 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2541 	 */
2542 	if (tcp_bblogging_on(rack->rc_tp)) {
2543 		union tcp_log_stackspecific log;
2544 		struct timeval tv;
2545 
2546 		memset(&log, 0, sizeof(log));
2547 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2548 		log.u_bbr.flex1 <<= 1;
2549 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2550 		log.u_bbr.flex1 <<= 1;
2551 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2552 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2553 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2554 		log.u_bbr.flex4 = flex4;
2555 		log.u_bbr.flex5 = flex5;
2556 		log.u_bbr.flex6 = flex6;
2557 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2558 		log.u_bbr.flex8 = mod;
2559 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2560 		log.u_bbr.epoch = rack->r_ctl.current_round;
2561 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2562 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2563 		    &rack->rc_inp->inp_socket->so_rcv,
2564 		    &rack->rc_inp->inp_socket->so_snd,
2565 		    RACK_DSACK_HANDLING, 0,
2566 		    0, &log, false, &tv);
2567 	}
2568 }
2569 
2570 static void
2571 rack_log_hdwr_pacing(struct tcp_rack *rack,
2572 		     uint64_t rate, uint64_t hw_rate, int line,
2573 		     int error, uint16_t mod)
2574 {
2575 	if (tcp_bblogging_on(rack->rc_tp)) {
2576 		union tcp_log_stackspecific log;
2577 		struct timeval tv;
2578 		const struct ifnet *ifp;
2579 		uint64_t ifp64;
2580 
2581 		memset(&log, 0, sizeof(log));
2582 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2583 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2584 		if (rack->r_ctl.crte) {
2585 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2586 		} else if (rack->rc_inp->inp_route.ro_nh &&
2587 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2588 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2589 		} else
2590 			ifp = NULL;
2591 		if (ifp) {
2592 			ifp64 = (uintptr_t)ifp;
2593 			log.u_bbr.flex3 = ((ifp64  >> 32) & 0x00000000ffffffff);
2594 			log.u_bbr.flex4 = (ifp64 & 0x00000000ffffffff);
2595 		}
2596 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2597 		log.u_bbr.bw_inuse = rate;
2598 		log.u_bbr.flex5 = line;
2599 		log.u_bbr.flex6 = error;
2600 		log.u_bbr.flex7 = mod;
2601 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2602 		log.u_bbr.flex8 = rack->use_fixed_rate;
2603 		log.u_bbr.flex8 <<= 1;
2604 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2605 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2606 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2607 		if (rack->r_ctl.crte)
2608 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2609 		else
2610 			log.u_bbr.cur_del_rate = 0;
2611 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2612 		log.u_bbr.epoch = rack->r_ctl.current_round;
2613 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2614 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2615 		    &rack->rc_inp->inp_socket->so_rcv,
2616 		    &rack->rc_inp->inp_socket->so_snd,
2617 		    BBR_LOG_HDWR_PACE, 0,
2618 		    0, &log, false, &tv);
2619 	}
2620 }
2621 
2622 static uint64_t
2623 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2624 {
2625 	/*
2626 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2627 	 */
2628 	uint64_t bw_est, high_rate;
2629 	uint64_t gain;
2630 
2631 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2632 	bw_est = bw * gain;
2633 	bw_est /= (uint64_t)100;
2634 	/* Never fall below the minimum (def 64kbps) */
2635 	if (bw_est < RACK_MIN_BW)
2636 		bw_est = RACK_MIN_BW;
2637 	if (rack->r_rack_hw_rate_caps) {
2638 		/* Rate caps are in place */
2639 		if (rack->r_ctl.crte != NULL) {
2640 			/* We have a hdwr rate already */
2641 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2642 			if (bw_est >= high_rate) {
2643 				/* We are capping bw at the highest rate table entry */
2644 				if (rack_hw_rate_cap_per &&
2645 				    (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2646 					rack->r_rack_hw_rate_caps = 0;
2647 					goto done;
2648 				}
2649 				rack_log_hdwr_pacing(rack,
2650 						     bw_est, high_rate, __LINE__,
2651 						     0, 3);
2652 				bw_est = high_rate;
2653 				if (capped)
2654 					*capped = 1;
2655 			}
2656 		} else if ((rack->rack_hdrw_pacing == 0) &&
2657 			   (rack->rack_hdw_pace_ena) &&
2658 			   (rack->rack_attempt_hdwr_pace == 0) &&
2659 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2660 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2661 			/*
2662 			 * Special case, we have not yet attempted hardware
2663 			 * pacing, and yet we may, when we do, find out if we are
2664 			 * above the highest rate. We need to know the maxbw for the interface
2665 			 * in question (if it supports ratelimiting). We get back
2666 			 * a 0, if the interface is not found in the RL lists.
2667 			 */
2668 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2669 			if (high_rate) {
2670 				/* Yep, we have a rate is it above this rate? */
2671 				if (bw_est > high_rate) {
2672 					bw_est = high_rate;
2673 					if (capped)
2674 						*capped = 1;
2675 				}
2676 			}
2677 		}
2678 	}
2679 done:
2680 	return (bw_est);
2681 }
2682 
2683 static void
2684 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2685 {
2686 	if (tcp_bblogging_on(rack->rc_tp)) {
2687 		union tcp_log_stackspecific log;
2688 		struct timeval tv;
2689 
2690 		if ((mod != 1) && (rack_verbose_logging == 0))  {
2691 			/*
2692 			 * We get 3 values currently for mod
2693 			 * 1 - We are retransmitting and this tells the reason.
2694 			 * 2 - We are clearing a dup-ack count.
2695 			 * 3 - We are incrementing a dup-ack count.
2696 			 *
2697 			 * The clear/increment are only logged
2698 			 * if you have BBverbose on.
2699 			 */
2700 			return;
2701 		}
2702 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2703 		log.u_bbr.flex1 = tsused;
2704 		log.u_bbr.flex2 = thresh;
2705 		log.u_bbr.flex3 = rsm->r_flags;
2706 		log.u_bbr.flex4 = rsm->r_dupack;
2707 		log.u_bbr.flex5 = rsm->r_start;
2708 		log.u_bbr.flex6 = rsm->r_end;
2709 		log.u_bbr.flex8 = mod;
2710 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2711 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2712 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2713 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2714 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2715 		log.u_bbr.pacing_gain = rack->r_must_retran;
2716 		log.u_bbr.epoch = rack->r_ctl.current_round;
2717 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2718 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2719 		    &rack->rc_inp->inp_socket->so_rcv,
2720 		    &rack->rc_inp->inp_socket->so_snd,
2721 		    BBR_LOG_SETTINGS_CHG, 0,
2722 		    0, &log, false, &tv);
2723 	}
2724 }
2725 
2726 static void
2727 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2728 {
2729 	if (tcp_bblogging_on(rack->rc_tp)) {
2730 		union tcp_log_stackspecific log;
2731 		struct timeval tv;
2732 
2733 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2734 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2735 		log.u_bbr.flex2 = to;
2736 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2737 		log.u_bbr.flex4 = slot;
2738 		log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2739 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2740 		log.u_bbr.flex7 = rack->rc_in_persist;
2741 		log.u_bbr.flex8 = which;
2742 		if (rack->rack_no_prr)
2743 			log.u_bbr.pkts_out = 0;
2744 		else
2745 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2746 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2747 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2748 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2749 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2750 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2751 		log.u_bbr.pacing_gain = rack->r_must_retran;
2752 		log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2753 		log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2754 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2755 		log.u_bbr.lost = rack_rto_min;
2756 		log.u_bbr.epoch = rack->r_ctl.roundends;
2757 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2758 		log.u_bbr.bw_inuse <<= 32;
2759 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2760 		log.u_bbr.applimited = rack->rc_tp->t_flags2;
2761 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2762 		    &rack->rc_inp->inp_socket->so_rcv,
2763 		    &rack->rc_inp->inp_socket->so_snd,
2764 		    BBR_LOG_TIMERSTAR, 0,
2765 		    0, &log, false, &tv);
2766 	}
2767 }
2768 
2769 static void
2770 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2771 {
2772 	if (tcp_bblogging_on(rack->rc_tp)) {
2773 		union tcp_log_stackspecific log;
2774 		struct timeval tv;
2775 
2776 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2777 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2778 		log.u_bbr.flex8 = to_num;
2779 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2780 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2781 		if (rsm == NULL)
2782 			log.u_bbr.flex3 = 0;
2783 		else
2784 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2785 		if (rack->rack_no_prr)
2786 			log.u_bbr.flex5 = 0;
2787 		else
2788 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2789 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2790 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2791 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2792 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2793 		log.u_bbr.pacing_gain = rack->r_must_retran;
2794 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2795 		log.u_bbr.bw_inuse <<= 32;
2796 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2797 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2798 		    &rack->rc_inp->inp_socket->so_rcv,
2799 		    &rack->rc_inp->inp_socket->so_snd,
2800 		    BBR_LOG_RTO, 0,
2801 		    0, &log, false, &tv);
2802 	}
2803 }
2804 
2805 static void
2806 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2807 		 struct rack_sendmap *prev,
2808 		 struct rack_sendmap *rsm,
2809 		 struct rack_sendmap *next,
2810 		 int flag, uint32_t th_ack, int line)
2811 {
2812 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2813 		union tcp_log_stackspecific log;
2814 		struct timeval tv;
2815 
2816 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2817 		log.u_bbr.flex8 = flag;
2818 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2819 		log.u_bbr.cur_del_rate = (uintptr_t)prev;
2820 		log.u_bbr.delRate = (uintptr_t)rsm;
2821 		log.u_bbr.rttProp = (uintptr_t)next;
2822 		log.u_bbr.flex7 = 0;
2823 		if (prev) {
2824 			log.u_bbr.flex1 = prev->r_start;
2825 			log.u_bbr.flex2 = prev->r_end;
2826 			log.u_bbr.flex7 |= 0x4;
2827 		}
2828 		if (rsm) {
2829 			log.u_bbr.flex3 = rsm->r_start;
2830 			log.u_bbr.flex4 = rsm->r_end;
2831 			log.u_bbr.flex7 |= 0x2;
2832 		}
2833 		if (next) {
2834 			log.u_bbr.flex5 = next->r_start;
2835 			log.u_bbr.flex6 = next->r_end;
2836 			log.u_bbr.flex7 |= 0x1;
2837 		}
2838 		log.u_bbr.applimited = line;
2839 		log.u_bbr.pkts_out = th_ack;
2840 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2841 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2842 		if (rack->rack_no_prr)
2843 			log.u_bbr.lost = 0;
2844 		else
2845 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2846 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2847 		log.u_bbr.bw_inuse <<= 32;
2848 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2849 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2850 		    &rack->rc_inp->inp_socket->so_rcv,
2851 		    &rack->rc_inp->inp_socket->so_snd,
2852 		    TCP_LOG_MAPCHG, 0,
2853 		    0, &log, false, &tv);
2854 	}
2855 }
2856 
2857 static void
2858 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2859 		 struct rack_sendmap *rsm, int conf)
2860 {
2861 	if (tcp_bblogging_on(tp)) {
2862 		union tcp_log_stackspecific log;
2863 		struct timeval tv;
2864 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2865 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2866 		log.u_bbr.flex1 = t;
2867 		log.u_bbr.flex2 = len;
2868 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2869 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2870 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2871 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2872 		log.u_bbr.flex7 = conf;
2873 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2874 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2875 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2876 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2877 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2878 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2879 		if (rsm) {
2880 			log.u_bbr.pkt_epoch = rsm->r_start;
2881 			log.u_bbr.lost = rsm->r_end;
2882 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2883 			/* We loose any upper of the 24 bits */
2884 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2885 		} else {
2886 			/* Its a SYN */
2887 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2888 			log.u_bbr.lost = 0;
2889 			log.u_bbr.cwnd_gain = 0;
2890 			log.u_bbr.pacing_gain = 0;
2891 		}
2892 		/* Write out general bits of interest rrs here */
2893 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2894 		log.u_bbr.use_lt_bw <<= 1;
2895 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2896 		log.u_bbr.use_lt_bw <<= 1;
2897 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2898 		log.u_bbr.use_lt_bw <<= 1;
2899 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2900 		log.u_bbr.use_lt_bw <<= 1;
2901 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2902 		log.u_bbr.use_lt_bw <<= 1;
2903 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2904 		log.u_bbr.use_lt_bw <<= 1;
2905 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2906 		log.u_bbr.use_lt_bw <<= 1;
2907 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2908 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2909 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2910 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2911 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2912 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2913 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2914 		log.u_bbr.bw_inuse <<= 32;
2915 		if (rsm)
2916 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2917 		TCP_LOG_EVENTP(tp, NULL,
2918 		    &rack->rc_inp->inp_socket->so_rcv,
2919 		    &rack->rc_inp->inp_socket->so_snd,
2920 		    BBR_LOG_BBRRTT, 0,
2921 		    0, &log, false, &tv);
2922 
2923 
2924 	}
2925 }
2926 
2927 static void
2928 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2929 {
2930 	/*
2931 	 * Log the rtt sample we are
2932 	 * applying to the srtt algorithm in
2933 	 * useconds.
2934 	 */
2935 	if (tcp_bblogging_on(rack->rc_tp)) {
2936 		union tcp_log_stackspecific log;
2937 		struct timeval tv;
2938 
2939 		/* Convert our ms to a microsecond */
2940 		memset(&log, 0, sizeof(log));
2941 		log.u_bbr.flex1 = rtt;
2942 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2943 		log.u_bbr.flex7 = 1;
2944 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2945 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2946 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2947 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2948 		log.u_bbr.pacing_gain = rack->r_must_retran;
2949 		/*
2950 		 * We capture in delRate the upper 32 bits as
2951 		 * the confidence level we had declared, and the
2952 		 * lower 32 bits as the actual RTT using the arrival
2953 		 * timestamp.
2954 		 */
2955 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2956 		log.u_bbr.delRate <<= 32;
2957 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2958 		/* Lets capture all the things that make up t_rtxcur */
2959 		log.u_bbr.applimited = rack_rto_min;
2960 		log.u_bbr.epoch = rack_rto_max;
2961 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2962 		log.u_bbr.lost = rack_rto_min;
2963 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2964 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2965 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2966 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2967 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2968 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2969 		    &rack->rc_inp->inp_socket->so_rcv,
2970 		    &rack->rc_inp->inp_socket->so_snd,
2971 		    TCP_LOG_RTT, 0,
2972 		    0, &log, false, &tv);
2973 	}
2974 }
2975 
2976 static void
2977 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2978 {
2979 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2980 		union tcp_log_stackspecific log;
2981 		struct timeval tv;
2982 
2983 		/* Convert our ms to a microsecond */
2984 		memset(&log, 0, sizeof(log));
2985 		log.u_bbr.flex1 = rtt;
2986 		log.u_bbr.flex2 = send_time;
2987 		log.u_bbr.flex3 = ack_time;
2988 		log.u_bbr.flex4 = where;
2989 		log.u_bbr.flex7 = 2;
2990 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2991 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2992 		log.u_bbr.bw_inuse <<= 32;
2993 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2994 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2995 		    &rack->rc_inp->inp_socket->so_rcv,
2996 		    &rack->rc_inp->inp_socket->so_snd,
2997 		    TCP_LOG_RTT, 0,
2998 		    0, &log, false, &tv);
2999 	}
3000 }
3001 
3002 
3003 static void
3004 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
3005 {
3006 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3007 		union tcp_log_stackspecific log;
3008 		struct timeval tv;
3009 
3010 		/* Convert our ms to a microsecond */
3011 		memset(&log, 0, sizeof(log));
3012 		log.u_bbr.flex1 = idx;
3013 		log.u_bbr.flex2 = rack_ts_to_msec(tsv);
3014 		log.u_bbr.flex3 = tsecho;
3015 		log.u_bbr.flex7 = 3;
3016 		log.u_bbr.rttProp = tsv;
3017 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3018 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3019 		log.u_bbr.bw_inuse <<= 32;
3020 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3021 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3022 		    &rack->rc_inp->inp_socket->so_rcv,
3023 		    &rack->rc_inp->inp_socket->so_snd,
3024 		    TCP_LOG_RTT, 0,
3025 		    0, &log, false, &tv);
3026 	}
3027 }
3028 
3029 
3030 static inline void
3031 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
3032 {
3033 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3034 		union tcp_log_stackspecific log;
3035 		struct timeval tv;
3036 
3037 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3038 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3039 		log.u_bbr.flex1 = line;
3040 		log.u_bbr.flex2 = tick;
3041 		log.u_bbr.flex3 = tp->t_maxunacktime;
3042 		log.u_bbr.flex4 = tp->t_acktime;
3043 		log.u_bbr.flex8 = event;
3044 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3045 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3046 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3047 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3048 		log.u_bbr.pacing_gain = rack->r_must_retran;
3049 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3050 		log.u_bbr.bw_inuse <<= 32;
3051 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3052 		TCP_LOG_EVENTP(tp, NULL,
3053 		    &rack->rc_inp->inp_socket->so_rcv,
3054 		    &rack->rc_inp->inp_socket->so_snd,
3055 		    BBR_LOG_PROGRESS, 0,
3056 		    0, &log, false, &tv);
3057 	}
3058 }
3059 
3060 static void
3061 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
3062 {
3063 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3064 		union tcp_log_stackspecific log;
3065 
3066 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3067 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3068 		log.u_bbr.flex1 = slot;
3069 		if (rack->rack_no_prr)
3070 			log.u_bbr.flex2 = 0;
3071 		else
3072 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3073 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3074 		log.u_bbr.flex6 = line;
3075 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3076 		log.u_bbr.flex8 = rack->rc_in_persist;
3077 		log.u_bbr.timeStamp = cts;
3078 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3079 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3080 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3081 		log.u_bbr.pacing_gain = rack->r_must_retran;
3082 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3083 		    &rack->rc_inp->inp_socket->so_rcv,
3084 		    &rack->rc_inp->inp_socket->so_snd,
3085 		    BBR_LOG_BBRSND, 0,
3086 		    0, &log, false, tv);
3087 	}
3088 }
3089 
3090 static void
3091 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
3092 {
3093 	if (tcp_bblogging_on(rack->rc_tp)) {
3094 		union tcp_log_stackspecific log;
3095 		struct timeval tv;
3096 
3097 		memset(&log, 0, sizeof(log));
3098 		log.u_bbr.flex1 = did_out;
3099 		log.u_bbr.flex2 = nxt_pkt;
3100 		log.u_bbr.flex3 = way_out;
3101 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3102 		if (rack->rack_no_prr)
3103 			log.u_bbr.flex5 = 0;
3104 		else
3105 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3106 		log.u_bbr.flex6 = nsegs;
3107 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3108 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
3109 		log.u_bbr.flex7 <<= 1;
3110 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
3111 		log.u_bbr.flex7 <<= 1;
3112 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
3113 		log.u_bbr.flex8 = rack->rc_in_persist;
3114 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3115 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3116 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3117 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3118 		log.u_bbr.use_lt_bw <<= 1;
3119 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3120 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3121 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3122 		log.u_bbr.pacing_gain = rack->r_must_retran;
3123 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3124 		log.u_bbr.bw_inuse <<= 32;
3125 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3126 		log.u_bbr.epoch = rack->rc_inp->inp_socket->so_snd.sb_hiwat;
3127 		log.u_bbr.lt_epoch = rack->rc_inp->inp_socket->so_rcv.sb_hiwat;
3128 		log.u_bbr.lost = rack->rc_tp->t_srtt;
3129 		log.u_bbr.pkt_epoch = rack->rc_tp->rfbuf_cnt;
3130 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3131 		    &rack->rc_inp->inp_socket->so_rcv,
3132 		    &rack->rc_inp->inp_socket->so_snd,
3133 		    BBR_LOG_DOSEG_DONE, 0,
3134 		    0, &log, false, &tv);
3135 	}
3136 }
3137 
3138 static void
3139 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
3140 {
3141 	if (tcp_bblogging_on(rack->rc_tp)) {
3142 		union tcp_log_stackspecific log;
3143 		struct timeval tv;
3144 
3145 		memset(&log, 0, sizeof(log));
3146 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3147 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3148 		log.u_bbr.flex4 = arg1;
3149 		log.u_bbr.flex5 = arg2;
3150 		log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3151 		log.u_bbr.flex6 = arg3;
3152 		log.u_bbr.flex8 = frm;
3153 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3154 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3155 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3156 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3157 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3158 		log.u_bbr.pacing_gain = rack->r_must_retran;
3159 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3160 		    &tptosocket(tp)->so_snd,
3161 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3162 	}
3163 }
3164 
3165 static void
3166 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3167 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3168 {
3169 	if (tcp_bblogging_on(rack->rc_tp)) {
3170 		union tcp_log_stackspecific log;
3171 		struct timeval tv;
3172 
3173 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3174 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3175 		log.u_bbr.flex1 = slot;
3176 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3177 		log.u_bbr.flex4 = reason;
3178 		if (rack->rack_no_prr)
3179 			log.u_bbr.flex5 = 0;
3180 		else
3181 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3182 		log.u_bbr.flex7 = hpts_calling;
3183 		log.u_bbr.flex8 = rack->rc_in_persist;
3184 		log.u_bbr.lt_epoch = cwnd_to_use;
3185 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3186 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
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.cwnd_gain = rack->rc_has_collapsed;
3191 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3192 		log.u_bbr.bw_inuse <<= 32;
3193 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3194 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3195 		    &rack->rc_inp->inp_socket->so_rcv,
3196 		    &rack->rc_inp->inp_socket->so_snd,
3197 		    BBR_LOG_JUSTRET, 0,
3198 		    tlen, &log, false, &tv);
3199 	}
3200 }
3201 
3202 static void
3203 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3204 		   struct timeval *tv, uint32_t flags_on_entry)
3205 {
3206 	if (tcp_bblogging_on(rack->rc_tp)) {
3207 		union tcp_log_stackspecific log;
3208 
3209 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3210 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3211 		log.u_bbr.flex1 = line;
3212 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3213 		log.u_bbr.flex3 = flags_on_entry;
3214 		log.u_bbr.flex4 = us_cts;
3215 		if (rack->rack_no_prr)
3216 			log.u_bbr.flex5 = 0;
3217 		else
3218 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3219 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3220 		log.u_bbr.flex7 = hpts_removed;
3221 		log.u_bbr.flex8 = 1;
3222 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3223 		log.u_bbr.timeStamp = us_cts;
3224 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3225 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3226 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3227 		log.u_bbr.pacing_gain = rack->r_must_retran;
3228 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3229 		log.u_bbr.bw_inuse <<= 32;
3230 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3231 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3232 		    &rack->rc_inp->inp_socket->so_rcv,
3233 		    &rack->rc_inp->inp_socket->so_snd,
3234 		    BBR_LOG_TIMERCANC, 0,
3235 		    0, &log, false, tv);
3236 	}
3237 }
3238 
3239 static void
3240 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3241 			  uint32_t flex1, uint32_t flex2,
3242 			  uint32_t flex3, uint32_t flex4,
3243 			  uint32_t flex5, uint32_t flex6,
3244 			  uint16_t flex7, uint8_t mod)
3245 {
3246 	if (tcp_bblogging_on(rack->rc_tp)) {
3247 		union tcp_log_stackspecific log;
3248 		struct timeval tv;
3249 
3250 		if (mod == 1) {
3251 			/* No you can't use 1, its for the real to cancel */
3252 			return;
3253 		}
3254 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3255 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3256 		log.u_bbr.flex1 = flex1;
3257 		log.u_bbr.flex2 = flex2;
3258 		log.u_bbr.flex3 = flex3;
3259 		log.u_bbr.flex4 = flex4;
3260 		log.u_bbr.flex5 = flex5;
3261 		log.u_bbr.flex6 = flex6;
3262 		log.u_bbr.flex7 = flex7;
3263 		log.u_bbr.flex8 = mod;
3264 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3265 		    &rack->rc_inp->inp_socket->so_rcv,
3266 		    &rack->rc_inp->inp_socket->so_snd,
3267 		    BBR_LOG_TIMERCANC, 0,
3268 		    0, &log, false, &tv);
3269 	}
3270 }
3271 
3272 static void
3273 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3274 {
3275 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3276 		union tcp_log_stackspecific log;
3277 		struct timeval tv;
3278 
3279 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3280 		log.u_bbr.flex1 = timers;
3281 		log.u_bbr.flex2 = ret;
3282 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3283 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3284 		log.u_bbr.flex5 = cts;
3285 		if (rack->rack_no_prr)
3286 			log.u_bbr.flex6 = 0;
3287 		else
3288 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3289 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3290 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3291 		log.u_bbr.pacing_gain = rack->r_must_retran;
3292 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3293 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3294 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3295 		    &rack->rc_inp->inp_socket->so_rcv,
3296 		    &rack->rc_inp->inp_socket->so_snd,
3297 		    BBR_LOG_TO_PROCESS, 0,
3298 		    0, &log, false, &tv);
3299 	}
3300 }
3301 
3302 static void
3303 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3304 {
3305 	if (tcp_bblogging_on(rack->rc_tp)) {
3306 		union tcp_log_stackspecific log;
3307 		struct timeval tv;
3308 
3309 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3310 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3311 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3312 		if (rack->rack_no_prr)
3313 			log.u_bbr.flex3 = 0;
3314 		else
3315 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3316 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3317 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3318 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3319 		log.u_bbr.flex7 = line;
3320 		log.u_bbr.flex8 = frm;
3321 		log.u_bbr.pkts_out = orig_cwnd;
3322 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3323 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3324 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3325 		log.u_bbr.use_lt_bw <<= 1;
3326 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3327 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3328 		    &rack->rc_inp->inp_socket->so_rcv,
3329 		    &rack->rc_inp->inp_socket->so_snd,
3330 		    BBR_LOG_BBRUPD, 0,
3331 		    0, &log, false, &tv);
3332 	}
3333 }
3334 
3335 static void
3336 rack_counter_destroy(void)
3337 {
3338 	counter_u64_free(rack_total_bytes);
3339 	counter_u64_free(rack_fto_send);
3340 	counter_u64_free(rack_fto_rsm_send);
3341 	counter_u64_free(rack_nfto_resend);
3342 	counter_u64_free(rack_hw_pace_init_fail);
3343 	counter_u64_free(rack_hw_pace_lost);
3344 	counter_u64_free(rack_non_fto_send);
3345 	counter_u64_free(rack_extended_rfo);
3346 	counter_u64_free(rack_ack_total);
3347 	counter_u64_free(rack_express_sack);
3348 	counter_u64_free(rack_sack_total);
3349 	counter_u64_free(rack_move_none);
3350 	counter_u64_free(rack_move_some);
3351 	counter_u64_free(rack_sack_attacks_detected);
3352 	counter_u64_free(rack_sack_attacks_reversed);
3353 	counter_u64_free(rack_sack_attacks_suspect);
3354 	counter_u64_free(rack_sack_used_next_merge);
3355 	counter_u64_free(rack_sack_used_prev_merge);
3356 	counter_u64_free(rack_tlp_tot);
3357 	counter_u64_free(rack_tlp_newdata);
3358 	counter_u64_free(rack_tlp_retran);
3359 	counter_u64_free(rack_tlp_retran_bytes);
3360 	counter_u64_free(rack_to_tot);
3361 	counter_u64_free(rack_saw_enobuf);
3362 	counter_u64_free(rack_saw_enobuf_hw);
3363 	counter_u64_free(rack_saw_enetunreach);
3364 	counter_u64_free(rack_hot_alloc);
3365 	counter_u64_free(rack_to_alloc);
3366 	counter_u64_free(rack_to_alloc_hard);
3367 	counter_u64_free(rack_to_alloc_emerg);
3368 	counter_u64_free(rack_to_alloc_limited);
3369 	counter_u64_free(rack_alloc_limited_conns);
3370 	counter_u64_free(rack_split_limited);
3371 	counter_u64_free(rack_multi_single_eq);
3372 	counter_u64_free(rack_rxt_clamps_cwnd);
3373 	counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3374 	counter_u64_free(rack_proc_non_comp_ack);
3375 	counter_u64_free(rack_sack_proc_all);
3376 	counter_u64_free(rack_sack_proc_restart);
3377 	counter_u64_free(rack_sack_proc_short);
3378 	counter_u64_free(rack_sack_skipped_acked);
3379 	counter_u64_free(rack_sack_splits);
3380 	counter_u64_free(rack_input_idle_reduces);
3381 	counter_u64_free(rack_collapsed_win);
3382 	counter_u64_free(rack_collapsed_win_rxt);
3383 	counter_u64_free(rack_collapsed_win_rxt_bytes);
3384 	counter_u64_free(rack_collapsed_win_seen);
3385 	counter_u64_free(rack_try_scwnd);
3386 	counter_u64_free(rack_persists_sends);
3387 	counter_u64_free(rack_persists_acks);
3388 	counter_u64_free(rack_persists_loss);
3389 	counter_u64_free(rack_persists_lost_ends);
3390 #ifdef INVARIANTS
3391 	counter_u64_free(rack_adjust_map_bw);
3392 #endif
3393 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3394 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3395 }
3396 
3397 static struct rack_sendmap *
3398 rack_alloc(struct tcp_rack *rack)
3399 {
3400 	struct rack_sendmap *rsm;
3401 
3402 	/*
3403 	 * First get the top of the list it in
3404 	 * theory is the "hottest" rsm we have,
3405 	 * possibly just freed by ack processing.
3406 	 */
3407 	if (rack->rc_free_cnt > rack_free_cache) {
3408 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3409 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3410 		counter_u64_add(rack_hot_alloc, 1);
3411 		rack->rc_free_cnt--;
3412 		return (rsm);
3413 	}
3414 	/*
3415 	 * Once we get under our free cache we probably
3416 	 * no longer have a "hot" one available. Lets
3417 	 * get one from UMA.
3418 	 */
3419 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
3420 	if (rsm) {
3421 		rack->r_ctl.rc_num_maps_alloced++;
3422 		counter_u64_add(rack_to_alloc, 1);
3423 		return (rsm);
3424 	}
3425 	/*
3426 	 * Dig in to our aux rsm's (the last two) since
3427 	 * UMA failed to get us one.
3428 	 */
3429 	if (rack->rc_free_cnt) {
3430 		counter_u64_add(rack_to_alloc_emerg, 1);
3431 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3432 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3433 		rack->rc_free_cnt--;
3434 		return (rsm);
3435 	}
3436 	return (NULL);
3437 }
3438 
3439 static struct rack_sendmap *
3440 rack_alloc_full_limit(struct tcp_rack *rack)
3441 {
3442 	if ((V_tcp_map_entries_limit > 0) &&
3443 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3444 		counter_u64_add(rack_to_alloc_limited, 1);
3445 		if (!rack->alloc_limit_reported) {
3446 			rack->alloc_limit_reported = 1;
3447 			counter_u64_add(rack_alloc_limited_conns, 1);
3448 		}
3449 		return (NULL);
3450 	}
3451 	return (rack_alloc(rack));
3452 }
3453 
3454 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3455 static struct rack_sendmap *
3456 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3457 {
3458 	struct rack_sendmap *rsm;
3459 
3460 	if (limit_type) {
3461 		/* currently there is only one limit type */
3462 		if (rack->r_ctl.rc_split_limit > 0 &&
3463 		    rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3464 			counter_u64_add(rack_split_limited, 1);
3465 			if (!rack->alloc_limit_reported) {
3466 				rack->alloc_limit_reported = 1;
3467 				counter_u64_add(rack_alloc_limited_conns, 1);
3468 			}
3469 			return (NULL);
3470 		}
3471 	}
3472 
3473 	/* allocate and mark in the limit type, if set */
3474 	rsm = rack_alloc(rack);
3475 	if (rsm != NULL && limit_type) {
3476 		rsm->r_limit_type = limit_type;
3477 		rack->r_ctl.rc_num_split_allocs++;
3478 	}
3479 	return (rsm);
3480 }
3481 
3482 static void
3483 rack_free_trim(struct tcp_rack *rack)
3484 {
3485 	struct rack_sendmap *rsm;
3486 
3487 	/*
3488 	 * Free up all the tail entries until
3489 	 * we get our list down to the limit.
3490 	 */
3491 	while (rack->rc_free_cnt > rack_free_cache) {
3492 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3493 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3494 		rack->rc_free_cnt--;
3495 		rack->r_ctl.rc_num_maps_alloced--;
3496 		uma_zfree(rack_zone, rsm);
3497 	}
3498 }
3499 
3500 static void
3501 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3502 {
3503 	if (rsm->r_flags & RACK_APP_LIMITED) {
3504 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
3505 			rack->r_ctl.rc_app_limited_cnt--;
3506 		}
3507 	}
3508 	if (rsm->r_limit_type) {
3509 		/* currently there is only one limit type */
3510 		rack->r_ctl.rc_num_split_allocs--;
3511 	}
3512 	if (rsm == rack->r_ctl.rc_first_appl) {
3513 		rack->r_ctl.cleared_app_ack_seq = rsm->r_start + (rsm->r_end - rsm->r_start);
3514 		rack->r_ctl.cleared_app_ack = 1;
3515 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3516 			rack->r_ctl.rc_first_appl = NULL;
3517 		else
3518 			rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3519 	}
3520 	if (rsm == rack->r_ctl.rc_resend)
3521 		rack->r_ctl.rc_resend = NULL;
3522 	if (rsm == rack->r_ctl.rc_end_appl)
3523 		rack->r_ctl.rc_end_appl = NULL;
3524 	if (rack->r_ctl.rc_tlpsend == rsm)
3525 		rack->r_ctl.rc_tlpsend = NULL;
3526 	if (rack->r_ctl.rc_sacklast == rsm)
3527 		rack->r_ctl.rc_sacklast = NULL;
3528 	memset(rsm, 0, sizeof(struct rack_sendmap));
3529 	/* Make sure we are not going to overrun our count limit of 0xff */
3530 	if ((rack->rc_free_cnt + 1) > RACK_FREE_CNT_MAX) {
3531 		rack_free_trim(rack);
3532 	}
3533 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3534 	rack->rc_free_cnt++;
3535 }
3536 
3537 static uint32_t
3538 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3539 {
3540 	uint64_t srtt, bw, len, tim;
3541 	uint32_t segsiz, def_len, minl;
3542 
3543 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3544 	def_len = rack_def_data_window * segsiz;
3545 	if (rack->rc_gp_filled == 0) {
3546 		/*
3547 		 * We have no measurement (IW is in flight?) so
3548 		 * we can only guess using our data_window sysctl
3549 		 * value (usually 20MSS).
3550 		 */
3551 		return (def_len);
3552 	}
3553 	/*
3554 	 * Now we have a number of factors to consider.
3555 	 *
3556 	 * 1) We have a desired BDP which is usually
3557 	 *    at least 2.
3558 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3559 	 *    but we allow it too to be more.
3560 	 * 3) We want to make sure a measurement last N useconds (if
3561 	 *    we have set rack_min_measure_usec.
3562 	 *
3563 	 * We handle the first concern here by trying to create a data
3564 	 * window of max(rack_def_data_window, DesiredBDP). The
3565 	 * second concern we handle in not letting the measurement
3566 	 * window end normally until at least the required SRTT's
3567 	 * have gone by which is done further below in
3568 	 * rack_enough_for_measurement(). Finally the third concern
3569 	 * we also handle here by calculating how long that time
3570 	 * would take at the current BW and then return the
3571 	 * max of our first calculation and that length. Note
3572 	 * that if rack_min_measure_usec is 0, we don't deal
3573 	 * with concern 3. Also for both Concern 1 and 3 an
3574 	 * application limited period could end the measurement
3575 	 * earlier.
3576 	 *
3577 	 * So lets calculate the BDP with the "known" b/w using
3578 	 * the SRTT has our rtt and then multiply it by the
3579 	 * goal.
3580 	 */
3581 	bw = rack_get_bw(rack);
3582 	srtt = (uint64_t)tp->t_srtt;
3583 	len = bw * srtt;
3584 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3585 	len *= max(1, rack_goal_bdp);
3586 	/* Now we need to round up to the nearest MSS */
3587 	len = roundup(len, segsiz);
3588 	if (rack_min_measure_usec) {
3589 		/* Now calculate our min length for this b/w */
3590 		tim = rack_min_measure_usec;
3591 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3592 		if (minl == 0)
3593 			minl = 1;
3594 		minl = roundup(minl, segsiz);
3595 		if (len < minl)
3596 			len = minl;
3597 	}
3598 	/*
3599 	 * Now if we have a very small window we want
3600 	 * to attempt to get the window that is
3601 	 * as small as possible. This happens on
3602 	 * low b/w connections and we don't want to
3603 	 * span huge numbers of rtt's between measurements.
3604 	 *
3605 	 * We basically include 2 over our "MIN window" so
3606 	 * that the measurement can be shortened (possibly) by
3607 	 * an ack'ed packet.
3608 	 */
3609 	if (len < def_len)
3610 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3611 	else
3612 		return (max((uint32_t)len, def_len));
3613 
3614 }
3615 
3616 static int
3617 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3618 {
3619 	uint32_t tim, srtts, segsiz;
3620 
3621 	/*
3622 	 * Has enough time passed for the GP measurement to be valid?
3623 	 */
3624 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3625 		/* Not enough bytes yet */
3626 		return (0);
3627 	}
3628 	if ((tp->snd_max == tp->snd_una) ||
3629 	    (th_ack == tp->snd_max)){
3630 		/*
3631 		 * All is acked quality of all acked is
3632 		 * usually low or medium, but we in theory could split
3633 		 * all acked into two cases, where you got
3634 		 * a signifigant amount of your window and
3635 		 * where you did not. For now we leave it
3636 		 * but it is something to contemplate in the
3637 		 * future. The danger here is that delayed ack
3638 		 * is effecting the last byte (which is a 50:50 chance).
3639 		 */
3640 		*quality = RACK_QUALITY_ALLACKED;
3641 		return (1);
3642 	}
3643 	if (SEQ_GEQ(th_ack,  tp->gput_ack)) {
3644 		/*
3645 		 * We obtained our entire window of data we wanted
3646 		 * no matter if we are in recovery or not then
3647 		 * its ok since expanding the window does not
3648 		 * make things fuzzy (or at least not as much).
3649 		 */
3650 		*quality = RACK_QUALITY_HIGH;
3651 		return (1);
3652 	}
3653 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3654 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3655 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3656 		/* Not enough bytes yet */
3657 		return (0);
3658 	}
3659 	if (rack->r_ctl.rc_first_appl &&
3660 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3661 		/*
3662 		 * We are up to the app limited send point
3663 		 * we have to measure irrespective of the time..
3664 		 */
3665 		*quality = RACK_QUALITY_APPLIMITED;
3666 		return (1);
3667 	}
3668 	/* Now what about time? */
3669 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3670 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3671 	if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3672 		/*
3673 		 * We do not allow a measurement if we are in recovery
3674 		 * that would shrink the goodput window we wanted.
3675 		 * This is to prevent cloudyness of when the last send
3676 		 * was actually made.
3677 		 */
3678 		*quality = RACK_QUALITY_HIGH;
3679 		return (1);
3680 	}
3681 	/* Nope not even a full SRTT has passed */
3682 	return (0);
3683 }
3684 
3685 static void
3686 rack_log_timely(struct tcp_rack *rack,
3687 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3688 		uint64_t up_bnd, int line, uint8_t method)
3689 {
3690 	if (tcp_bblogging_on(rack->rc_tp)) {
3691 		union tcp_log_stackspecific log;
3692 		struct timeval tv;
3693 
3694 		memset(&log, 0, sizeof(log));
3695 		log.u_bbr.flex1 = logged;
3696 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3697 		log.u_bbr.flex2 <<= 4;
3698 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3699 		log.u_bbr.flex2 <<= 4;
3700 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3701 		log.u_bbr.flex2 <<= 4;
3702 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3703 		log.u_bbr.flex3 = rack->rc_gp_incr;
3704 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3705 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3706 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3707 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3708 		log.u_bbr.flex8 = method;
3709 		log.u_bbr.cur_del_rate = cur_bw;
3710 		log.u_bbr.delRate = low_bnd;
3711 		log.u_bbr.bw_inuse = up_bnd;
3712 		log.u_bbr.rttProp = rack_get_bw(rack);
3713 		log.u_bbr.pkt_epoch = line;
3714 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3715 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3716 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3717 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3718 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3719 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3720 		log.u_bbr.cwnd_gain <<= 1;
3721 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3722 		log.u_bbr.cwnd_gain <<= 1;
3723 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3724 		log.u_bbr.cwnd_gain <<= 1;
3725 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3726 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3727 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3728 		    &rack->rc_inp->inp_socket->so_rcv,
3729 		    &rack->rc_inp->inp_socket->so_snd,
3730 		    TCP_TIMELY_WORK, 0,
3731 		    0, &log, false, &tv);
3732 	}
3733 }
3734 
3735 static int
3736 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3737 {
3738 	/*
3739 	 * Before we increase we need to know if
3740 	 * the estimate just made was less than
3741 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3742 	 *
3743 	 * If we already are pacing at a fast enough
3744 	 * rate to push us faster there is no sense of
3745 	 * increasing.
3746 	 *
3747 	 * We first caculate our actual pacing rate (ss or ca multiplier
3748 	 * times our cur_bw).
3749 	 *
3750 	 * Then we take the last measured rate and multipy by our
3751 	 * maximum pacing overage to give us a max allowable rate.
3752 	 *
3753 	 * If our act_rate is smaller than our max_allowable rate
3754 	 * then we should increase. Else we should hold steady.
3755 	 *
3756 	 */
3757 	uint64_t act_rate, max_allow_rate;
3758 
3759 	if (rack_timely_no_stopping)
3760 		return (1);
3761 
3762 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3763 		/*
3764 		 * Initial startup case or
3765 		 * everything is acked case.
3766 		 */
3767 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3768 				__LINE__, 9);
3769 		return (1);
3770 	}
3771 	if (mult <= 100) {
3772 		/*
3773 		 * We can always pace at or slightly above our rate.
3774 		 */
3775 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3776 				__LINE__, 9);
3777 		return (1);
3778 	}
3779 	act_rate = cur_bw * (uint64_t)mult;
3780 	act_rate /= 100;
3781 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3782 	max_allow_rate /= 100;
3783 	if (act_rate < max_allow_rate) {
3784 		/*
3785 		 * Here the rate we are actually pacing at
3786 		 * is smaller than 10% above our last measurement.
3787 		 * This means we are pacing below what we would
3788 		 * like to try to achieve (plus some wiggle room).
3789 		 */
3790 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3791 				__LINE__, 9);
3792 		return (1);
3793 	} else {
3794 		/*
3795 		 * Here we are already pacing at least rack_max_per_above(10%)
3796 		 * what we are getting back. This indicates most likely
3797 		 * that we are being limited (cwnd/rwnd/app) and can't
3798 		 * get any more b/w. There is no sense of trying to
3799 		 * raise up the pacing rate its not speeding us up
3800 		 * and we already are pacing faster than we are getting.
3801 		 */
3802 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3803 				__LINE__, 8);
3804 		return (0);
3805 	}
3806 }
3807 
3808 static void
3809 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3810 {
3811 	/*
3812 	 * When we drag bottom, we want to assure
3813 	 * that no multiplier is below 1.0, if so
3814 	 * we want to restore it to at least that.
3815 	 */
3816 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3817 		/* This is unlikely we usually do not touch recovery */
3818 		rack->r_ctl.rack_per_of_gp_rec = 100;
3819 	}
3820 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3821 		rack->r_ctl.rack_per_of_gp_ca = 100;
3822 	}
3823 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3824 		rack->r_ctl.rack_per_of_gp_ss = 100;
3825 	}
3826 }
3827 
3828 static void
3829 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3830 {
3831 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3832 		rack->r_ctl.rack_per_of_gp_ca = 100;
3833 	}
3834 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3835 		rack->r_ctl.rack_per_of_gp_ss = 100;
3836 	}
3837 }
3838 
3839 static void
3840 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3841 {
3842 	int32_t  calc, logged, plus;
3843 
3844 	logged = 0;
3845 
3846 	if (rack->rc_skip_timely)
3847 		return;
3848 	if (override) {
3849 		/*
3850 		 * override is passed when we are
3851 		 * loosing b/w and making one last
3852 		 * gasp at trying to not loose out
3853 		 * to a new-reno flow.
3854 		 */
3855 		goto extra_boost;
3856 	}
3857 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3858 	if (rack->rc_gp_incr &&
3859 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3860 		/*
3861 		 * Reset and get 5 strokes more before the boost. Note
3862 		 * that the count is 0 based so we have to add one.
3863 		 */
3864 extra_boost:
3865 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3866 		rack->rc_gp_timely_inc_cnt = 0;
3867 	} else
3868 		plus = (uint32_t)rack_gp_increase_per;
3869 	/* Must be at least 1% increase for true timely increases */
3870 	if ((plus < 1) &&
3871 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3872 		plus = 1;
3873 	if (rack->rc_gp_saw_rec &&
3874 	    (rack->rc_gp_no_rec_chg == 0) &&
3875 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3876 				  rack->r_ctl.rack_per_of_gp_rec)) {
3877 		/* We have been in recovery ding it too */
3878 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3879 		if (calc > 0xffff)
3880 			calc = 0xffff;
3881 		logged |= 1;
3882 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3883 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3884 		    (rack->rc_dragged_bottom == 0) &&
3885 		    (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3886 			rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3887 	}
3888 	if (rack->rc_gp_saw_ca &&
3889 	    (rack->rc_gp_saw_ss == 0) &&
3890 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3891 				  rack->r_ctl.rack_per_of_gp_ca)) {
3892 		/* In CA */
3893 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3894 		if (calc > 0xffff)
3895 			calc = 0xffff;
3896 		logged |= 2;
3897 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3898 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3899 		    (rack->rc_dragged_bottom == 0) &&
3900 		    (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3901 			rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3902 	}
3903 	if (rack->rc_gp_saw_ss &&
3904 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3905 				  rack->r_ctl.rack_per_of_gp_ss)) {
3906 		/* In SS */
3907 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3908 		if (calc > 0xffff)
3909 			calc = 0xffff;
3910 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3911 		if (rack->r_ctl.rack_per_upper_bound_ss &&
3912 		    (rack->rc_dragged_bottom == 0) &&
3913 		    (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3914 			rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3915 		logged |= 4;
3916 	}
3917 	if (logged &&
3918 	    (rack->rc_gp_incr == 0)){
3919 		/* Go into increment mode */
3920 		rack->rc_gp_incr = 1;
3921 		rack->rc_gp_timely_inc_cnt = 0;
3922 	}
3923 	if (rack->rc_gp_incr &&
3924 	    logged &&
3925 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3926 		rack->rc_gp_timely_inc_cnt++;
3927 	}
3928 	rack_log_timely(rack,  logged, plus, 0, 0,
3929 			__LINE__, 1);
3930 }
3931 
3932 static uint32_t
3933 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3934 {
3935 	/*-
3936 	 * norm_grad = rtt_diff / minrtt;
3937 	 * new_per = curper * (1 - B * norm_grad)
3938 	 *
3939 	 * B = rack_gp_decrease_per (default 80%)
3940 	 * rtt_dif = input var current rtt-diff
3941 	 * curper = input var current percentage
3942 	 * minrtt = from rack filter
3943 	 *
3944 	 * In order to do the floating point calculations above we
3945 	 * do an integer conversion. The code looks confusing so let me
3946 	 * translate it into something that use more variables and
3947 	 * is clearer for us humans :)
3948 	 *
3949 	 * uint64_t norm_grad, inverse, reduce_by, final_result;
3950 	 * uint32_t perf;
3951 	 *
3952 	 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
3953 	 *             (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
3954 	 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
3955 	 * inverse /= 1000000;
3956 	 * reduce_by = (1000000 - inverse);
3957 	 * final_result = (cur_per * reduce_by) / 1000000;
3958 	 * perf = (uint32_t)final_result;
3959 	 */
3960 	uint64_t perf;
3961 
3962 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3963 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3964 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3965 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3966 		     (uint64_t)1000000)) /
3967 		(uint64_t)1000000);
3968 	if (perf > curper) {
3969 		/* TSNH */
3970 		perf = curper - 1;
3971 	}
3972 	return ((uint32_t)perf);
3973 }
3974 
3975 static uint32_t
3976 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3977 {
3978 	/*
3979 	 *                                   highrttthresh
3980 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3981 	 *                                     gp_srtt
3982 	 *
3983 	 * B = rack_gp_decrease_per (default .8 i.e. 80)
3984 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3985 	 */
3986 	uint64_t perf;
3987 	uint32_t highrttthresh;
3988 
3989 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3990 
3991 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3992 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3993 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3994 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3995 	if (tcp_bblogging_on(rack->rc_tp)) {
3996 		uint64_t log1;
3997 
3998 		log1 = rtt;
3999 		log1 <<= 32;
4000 		log1 |= highrttthresh;
4001 		rack_log_timely(rack,
4002 				rack_gp_decrease_per,
4003 				(uint64_t)curper,
4004 				log1,
4005 				perf,
4006 				__LINE__,
4007 				15);
4008 	}
4009 	return (perf);
4010 }
4011 
4012 static void
4013 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
4014 {
4015 	uint64_t logvar, logvar2, logvar3;
4016 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
4017 
4018 	if (rack->rc_skip_timely)
4019 		return;
4020 	if (rack->rc_gp_incr) {
4021 		/* Turn off increment counting */
4022 		rack->rc_gp_incr = 0;
4023 		rack->rc_gp_timely_inc_cnt = 0;
4024 	}
4025 	ss_red = ca_red = rec_red = 0;
4026 	logged = 0;
4027 	/* Calculate the reduction value */
4028 	if (rtt_diff < 0) {
4029 		rtt_diff *= -1;
4030 	}
4031 	/* Must be at least 1% reduction */
4032 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
4033 		/* We have been in recovery ding it too */
4034 		if (timely_says == 2) {
4035 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
4036 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4037 			if (alt < new_per)
4038 				val = alt;
4039 			else
4040 				val = new_per;
4041 		} else
4042 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4043 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
4044 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4045 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4046 		} else {
4047 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4048 			rec_red = 0;
4049 		}
4050 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4051 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4052 		logged |= 1;
4053 	}
4054 	if (rack->rc_gp_saw_ss) {
4055 		/* Sent in SS */
4056 		if (timely_says == 2) {
4057 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4058 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4059 			if (alt < new_per)
4060 				val = alt;
4061 			else
4062 				val = new_per;
4063 		} else
4064 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4065 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4066 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4067 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4068 		} else {
4069 			ss_red = new_per;
4070 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4071 			logvar = new_per;
4072 			logvar <<= 32;
4073 			logvar |= alt;
4074 			logvar2 = (uint32_t)rtt;
4075 			logvar2 <<= 32;
4076 			logvar2 |= (uint32_t)rtt_diff;
4077 			logvar3 = rack_gp_rtt_maxmul;
4078 			logvar3 <<= 32;
4079 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4080 			rack_log_timely(rack, timely_says,
4081 					logvar2, logvar3,
4082 					logvar, __LINE__, 10);
4083 		}
4084 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4085 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4086 		logged |= 4;
4087 	} else if (rack->rc_gp_saw_ca) {
4088 		/* Sent in CA */
4089 		if (timely_says == 2) {
4090 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4091 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4092 			if (alt < new_per)
4093 				val = alt;
4094 			else
4095 				val = new_per;
4096 		} else
4097 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4098 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
4099 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4100 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4101 		} else {
4102 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4103 			ca_red = 0;
4104 			logvar = new_per;
4105 			logvar <<= 32;
4106 			logvar |= alt;
4107 			logvar2 = (uint32_t)rtt;
4108 			logvar2 <<= 32;
4109 			logvar2 |= (uint32_t)rtt_diff;
4110 			logvar3 = rack_gp_rtt_maxmul;
4111 			logvar3 <<= 32;
4112 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4113 			rack_log_timely(rack, timely_says,
4114 					logvar2, logvar3,
4115 					logvar, __LINE__, 10);
4116 		}
4117 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4118 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4119 		logged |= 2;
4120 	}
4121 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
4122 		rack->rc_gp_timely_dec_cnt++;
4123 		if (rack_timely_dec_clear &&
4124 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4125 			rack->rc_gp_timely_dec_cnt = 0;
4126 	}
4127 	logvar = ss_red;
4128 	logvar <<= 32;
4129 	logvar |= ca_red;
4130 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
4131 			__LINE__, 2);
4132 }
4133 
4134 static void
4135 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4136 		     uint32_t rtt, uint32_t line, uint8_t reas)
4137 {
4138 	if (tcp_bblogging_on(rack->rc_tp)) {
4139 		union tcp_log_stackspecific log;
4140 		struct timeval tv;
4141 
4142 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4143 		log.u_bbr.flex1 = line;
4144 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4145 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4146 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4147 		log.u_bbr.flex5 = rtt;
4148 		log.u_bbr.flex6 = rack->rc_highly_buffered;
4149 		log.u_bbr.flex6 <<= 1;
4150 		log.u_bbr.flex6 |= rack->forced_ack;
4151 		log.u_bbr.flex6 <<= 1;
4152 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4153 		log.u_bbr.flex6 <<= 1;
4154 		log.u_bbr.flex6 |= rack->in_probe_rtt;
4155 		log.u_bbr.flex6 <<= 1;
4156 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4157 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4158 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4159 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4160 		log.u_bbr.flex8 = reas;
4161 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4162 		log.u_bbr.delRate = rack_get_bw(rack);
4163 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4164 		log.u_bbr.cur_del_rate <<= 32;
4165 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4166 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4167 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4168 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4169 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4170 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4171 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4172 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4173 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4174 		log.u_bbr.rttProp = us_cts;
4175 		log.u_bbr.rttProp <<= 32;
4176 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4177 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4178 		    &rack->rc_inp->inp_socket->so_rcv,
4179 		    &rack->rc_inp->inp_socket->so_snd,
4180 		    BBR_LOG_RTT_SHRINKS, 0,
4181 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4182 	}
4183 }
4184 
4185 static void
4186 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4187 {
4188 	uint64_t bwdp;
4189 
4190 	bwdp = rack_get_bw(rack);
4191 	bwdp *= (uint64_t)rtt;
4192 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4193 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4194 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4195 		/*
4196 		 * A window protocol must be able to have 4 packets
4197 		 * outstanding as the floor in order to function
4198 		 * (especially considering delayed ack :D).
4199 		 */
4200 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4201 	}
4202 }
4203 
4204 static void
4205 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4206 {
4207 	/**
4208 	 * ProbeRTT is a bit different in rack_pacing than in
4209 	 * BBR. It is like BBR in that it uses the lowering of
4210 	 * the RTT as a signal that we saw something new and
4211 	 * counts from there for how long between. But it is
4212 	 * different in that its quite simple. It does not
4213 	 * play with the cwnd and wait until we get down
4214 	 * to N segments outstanding and hold that for
4215 	 * 200ms. Instead it just sets the pacing reduction
4216 	 * rate to a set percentage (70 by default) and hold
4217 	 * that for a number of recent GP Srtt's.
4218 	 */
4219 	uint32_t segsiz;
4220 
4221 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4222 	if (rack->rc_gp_dyn_mul == 0)
4223 		return;
4224 
4225 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4226 		/* We are idle */
4227 		return;
4228 	}
4229 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4230 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4231 		/*
4232 		 * Stop the goodput now, the idea here is
4233 		 * that future measurements with in_probe_rtt
4234 		 * won't register if they are not greater so
4235 		 * we want to get what info (if any) is available
4236 		 * now.
4237 		 */
4238 		rack_do_goodput_measurement(rack->rc_tp, rack,
4239 					    rack->rc_tp->snd_una, __LINE__,
4240 					    RACK_QUALITY_PROBERTT);
4241 	}
4242 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4243 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4244 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4245 		     rack->r_ctl.rc_pace_min_segs);
4246 	rack->in_probe_rtt = 1;
4247 	rack->measure_saw_probe_rtt = 1;
4248 	rack->r_ctl.rc_time_probertt_starts = 0;
4249 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4250 	if (rack_probertt_use_min_rtt_entry)
4251 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4252 	else
4253 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4254 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4255 			     __LINE__, RACK_RTTS_ENTERPROBE);
4256 }
4257 
4258 static void
4259 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4260 {
4261 	struct rack_sendmap *rsm;
4262 	uint32_t segsiz;
4263 
4264 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4265 		     rack->r_ctl.rc_pace_min_segs);
4266 	rack->in_probe_rtt = 0;
4267 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4268 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4269 		/*
4270 		 * Stop the goodput now, the idea here is
4271 		 * that future measurements with in_probe_rtt
4272 		 * won't register if they are not greater so
4273 		 * we want to get what info (if any) is available
4274 		 * now.
4275 		 */
4276 		rack_do_goodput_measurement(rack->rc_tp, rack,
4277 					    rack->rc_tp->snd_una, __LINE__,
4278 					    RACK_QUALITY_PROBERTT);
4279 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4280 		/*
4281 		 * We don't have enough data to make a measurement.
4282 		 * So lets just stop and start here after exiting
4283 		 * probe-rtt. We probably are not interested in
4284 		 * the results anyway.
4285 		 */
4286 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4287 	}
4288 	/*
4289 	 * Measurements through the current snd_max are going
4290 	 * to be limited by the slower pacing rate.
4291 	 *
4292 	 * We need to mark these as app-limited so we
4293 	 * don't collapse the b/w.
4294 	 */
4295 	rsm = tqhash_max(rack->r_ctl.tqh);
4296 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4297 		if (rack->r_ctl.rc_app_limited_cnt == 0)
4298 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4299 		else {
4300 			/*
4301 			 * Go out to the end app limited and mark
4302 			 * this new one as next and move the end_appl up
4303 			 * to this guy.
4304 			 */
4305 			if (rack->r_ctl.rc_end_appl)
4306 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4307 			rack->r_ctl.rc_end_appl = rsm;
4308 		}
4309 		rsm->r_flags |= RACK_APP_LIMITED;
4310 		rack->r_ctl.rc_app_limited_cnt++;
4311 	}
4312 	/*
4313 	 * Now, we need to examine our pacing rate multipliers.
4314 	 * If its under 100%, we need to kick it back up to
4315 	 * 100%. We also don't let it be over our "max" above
4316 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4317 	 * Note setting clamp_atexit_prtt to 0 has the effect
4318 	 * of setting CA/SS to 100% always at exit (which is
4319 	 * the default behavior).
4320 	 */
4321 	if (rack_probertt_clear_is) {
4322 		rack->rc_gp_incr = 0;
4323 		rack->rc_gp_bwred = 0;
4324 		rack->rc_gp_timely_inc_cnt = 0;
4325 		rack->rc_gp_timely_dec_cnt = 0;
4326 	}
4327 	/* Do we do any clamping at exit? */
4328 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4329 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4330 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4331 	}
4332 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4333 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4334 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4335 	}
4336 	/*
4337 	 * Lets set rtt_diff to 0, so that we will get a "boost"
4338 	 * after exiting.
4339 	 */
4340 	rack->r_ctl.rc_rtt_diff = 0;
4341 
4342 	/* Clear all flags so we start fresh */
4343 	rack->rc_tp->t_bytes_acked = 0;
4344 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4345 	/*
4346 	 * If configured to, set the cwnd and ssthresh to
4347 	 * our targets.
4348 	 */
4349 	if (rack_probe_rtt_sets_cwnd) {
4350 		uint64_t ebdp;
4351 		uint32_t setto;
4352 
4353 		/* Set ssthresh so we get into CA once we hit our target */
4354 		if (rack_probertt_use_min_rtt_exit == 1) {
4355 			/* Set to min rtt */
4356 			rack_set_prtt_target(rack, segsiz,
4357 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4358 		} else if (rack_probertt_use_min_rtt_exit == 2) {
4359 			/* Set to current gp rtt */
4360 			rack_set_prtt_target(rack, segsiz,
4361 					     rack->r_ctl.rc_gp_srtt);
4362 		} else if (rack_probertt_use_min_rtt_exit == 3) {
4363 			/* Set to entry gp rtt */
4364 			rack_set_prtt_target(rack, segsiz,
4365 					     rack->r_ctl.rc_entry_gp_rtt);
4366 		} else {
4367 			uint64_t sum;
4368 			uint32_t setval;
4369 
4370 			sum = rack->r_ctl.rc_entry_gp_rtt;
4371 			sum *= 10;
4372 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4373 			if (sum >= 20) {
4374 				/*
4375 				 * A highly buffered path needs
4376 				 * cwnd space for timely to work.
4377 				 * Lets set things up as if
4378 				 * we are heading back here again.
4379 				 */
4380 				setval = rack->r_ctl.rc_entry_gp_rtt;
4381 			} else if (sum >= 15) {
4382 				/*
4383 				 * Lets take the smaller of the
4384 				 * two since we are just somewhat
4385 				 * buffered.
4386 				 */
4387 				setval = rack->r_ctl.rc_gp_srtt;
4388 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
4389 					setval = rack->r_ctl.rc_entry_gp_rtt;
4390 			} else {
4391 				/*
4392 				 * Here we are not highly buffered
4393 				 * and should pick the min we can to
4394 				 * keep from causing loss.
4395 				 */
4396 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4397 			}
4398 			rack_set_prtt_target(rack, segsiz,
4399 					     setval);
4400 		}
4401 		if (rack_probe_rtt_sets_cwnd > 1) {
4402 			/* There is a percentage here to boost */
4403 			ebdp = rack->r_ctl.rc_target_probertt_flight;
4404 			ebdp *= rack_probe_rtt_sets_cwnd;
4405 			ebdp /= 100;
4406 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4407 		} else
4408 			setto = rack->r_ctl.rc_target_probertt_flight;
4409 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4410 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4411 			/* Enforce a min */
4412 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4413 		}
4414 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
4415 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4416 	}
4417 	rack_log_rtt_shrinks(rack,  us_cts,
4418 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4419 			     __LINE__, RACK_RTTS_EXITPROBE);
4420 	/* Clear times last so log has all the info */
4421 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4422 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4423 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4424 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
4425 }
4426 
4427 static void
4428 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4429 {
4430 	/* Check in on probe-rtt */
4431 
4432 	if (rack->rc_gp_filled == 0) {
4433 		/* We do not do p-rtt unless we have gp measurements */
4434 		return;
4435 	}
4436 	if (rack->in_probe_rtt) {
4437 		uint64_t no_overflow;
4438 		uint32_t endtime, must_stay;
4439 
4440 		if (rack->r_ctl.rc_went_idle_time &&
4441 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4442 			/*
4443 			 * We went idle during prtt, just exit now.
4444 			 */
4445 			rack_exit_probertt(rack, us_cts);
4446 		} else if (rack_probe_rtt_safety_val &&
4447 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4448 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4449 			/*
4450 			 * Probe RTT safety value triggered!
4451 			 */
4452 			rack_log_rtt_shrinks(rack,  us_cts,
4453 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4454 					     __LINE__, RACK_RTTS_SAFETY);
4455 			rack_exit_probertt(rack, us_cts);
4456 		}
4457 		/* Calculate the max we will wait */
4458 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4459 		if (rack->rc_highly_buffered)
4460 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4461 		/* Calculate the min we must wait */
4462 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4463 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4464 		    TSTMP_LT(us_cts, endtime)) {
4465 			uint32_t calc;
4466 			/* Do we lower more? */
4467 no_exit:
4468 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4469 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4470 			else
4471 				calc = 0;
4472 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4473 			if (calc) {
4474 				/* Maybe */
4475 				calc *= rack_per_of_gp_probertt_reduce;
4476 				if (calc > rack_per_of_gp_probertt)
4477 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4478 				else
4479 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4480 				/* Limit it too */
4481 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4482 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4483 			}
4484 			/* We must reach target or the time set */
4485 			return;
4486 		}
4487 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
4488 			if ((TSTMP_LT(us_cts, must_stay) &&
4489 			     rack->rc_highly_buffered) ||
4490 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4491 			      rack->r_ctl.rc_target_probertt_flight)) {
4492 				/* We are not past the must_stay time */
4493 				goto no_exit;
4494 			}
4495 			rack_log_rtt_shrinks(rack,  us_cts,
4496 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4497 					     __LINE__, RACK_RTTS_REACHTARGET);
4498 			rack->r_ctl.rc_time_probertt_starts = us_cts;
4499 			if (rack->r_ctl.rc_time_probertt_starts == 0)
4500 				rack->r_ctl.rc_time_probertt_starts = 1;
4501 			/* Restore back to our rate we want to pace at in prtt */
4502 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4503 		}
4504 		/*
4505 		 * Setup our end time, some number of gp_srtts plus 200ms.
4506 		 */
4507 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4508 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4509 		if (rack_probertt_gpsrtt_cnt_div)
4510 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4511 		else
4512 			endtime = 0;
4513 		endtime += rack_min_probertt_hold;
4514 		endtime += rack->r_ctl.rc_time_probertt_starts;
4515 		if (TSTMP_GEQ(us_cts,  endtime)) {
4516 			/* yes, exit probertt */
4517 			rack_exit_probertt(rack, us_cts);
4518 		}
4519 
4520 	} else if ((rack->rc_skip_timely == 0) &&
4521 		   (TSTMP_GT(us_cts, rack->r_ctl.rc_lower_rtt_us_cts)) &&
4522 		   ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt)) {
4523 		/* Go into probertt, its been too long since we went lower */
4524 		rack_enter_probertt(rack, us_cts);
4525 	}
4526 }
4527 
4528 static void
4529 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4530 		       uint32_t rtt, int32_t rtt_diff)
4531 {
4532 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4533 	uint32_t losses;
4534 
4535 	if ((rack->rc_gp_dyn_mul == 0) ||
4536 	    (rack->use_fixed_rate) ||
4537 	    (rack->in_probe_rtt) ||
4538 	    (rack->rc_always_pace == 0)) {
4539 		/* No dynamic GP multiplier in play */
4540 		return;
4541 	}
4542 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4543 	cur_bw = rack_get_bw(rack);
4544 	/* Calculate our up and down range */
4545 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4546 	up_bnd /= 100;
4547 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4548 
4549 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4550 	subfr /= 100;
4551 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4552 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4553 		/*
4554 		 * This is the case where our RTT is above
4555 		 * the max target and we have been configured
4556 		 * to just do timely no bonus up stuff in that case.
4557 		 *
4558 		 * There are two configurations, set to 1, and we
4559 		 * just do timely if we are over our max. If its
4560 		 * set above 1 then we slam the multipliers down
4561 		 * to 100 and then decrement per timely.
4562 		 */
4563 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4564 				__LINE__, 3);
4565 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4566 			rack_validate_multipliers_at_or_below_100(rack);
4567 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4568 	} else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4569 		/*
4570 		 * We are decreasing this is a bit complicated this
4571 		 * means we are loosing ground. This could be
4572 		 * because another flow entered and we are competing
4573 		 * for b/w with it. This will push the RTT up which
4574 		 * makes timely unusable unless we want to get shoved
4575 		 * into a corner and just be backed off (the age
4576 		 * old problem with delay based CC).
4577 		 *
4578 		 * On the other hand if it was a route change we
4579 		 * would like to stay somewhat contained and not
4580 		 * blow out the buffers.
4581 		 */
4582 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4583 				__LINE__, 3);
4584 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4585 		if (rack->rc_gp_bwred == 0) {
4586 			/* Go into reduction counting */
4587 			rack->rc_gp_bwred = 1;
4588 			rack->rc_gp_timely_dec_cnt = 0;
4589 		}
4590 		if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4591 			/*
4592 			 * Push another time with a faster pacing
4593 			 * to try to gain back (we include override to
4594 			 * get a full raise factor).
4595 			 */
4596 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4597 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4598 			    (timely_says == 0) ||
4599 			    (rack_down_raise_thresh == 0)) {
4600 				/*
4601 				 * Do an override up in b/w if we were
4602 				 * below the threshold or if the threshold
4603 				 * is zero we always do the raise.
4604 				 */
4605 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4606 			} else {
4607 				/* Log it stays the same */
4608 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4609 						__LINE__, 11);
4610 			}
4611 			rack->rc_gp_timely_dec_cnt++;
4612 			/* We are not incrementing really no-count */
4613 			rack->rc_gp_incr = 0;
4614 			rack->rc_gp_timely_inc_cnt = 0;
4615 		} else {
4616 			/*
4617 			 * Lets just use the RTT
4618 			 * information and give up
4619 			 * pushing.
4620 			 */
4621 			goto use_timely;
4622 		}
4623 	} else if ((timely_says != 2) &&
4624 		    !losses &&
4625 		    (last_bw_est > up_bnd)) {
4626 		/*
4627 		 * We are increasing b/w lets keep going, updating
4628 		 * our b/w and ignoring any timely input, unless
4629 		 * of course we are at our max raise (if there is one).
4630 		 */
4631 
4632 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4633 				__LINE__, 3);
4634 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4635 		if (rack->rc_gp_saw_ss &&
4636 		    rack->r_ctl.rack_per_upper_bound_ss &&
4637 		     (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4638 			    /*
4639 			     * In cases where we can't go higher
4640 			     * we should just use timely.
4641 			     */
4642 			    goto use_timely;
4643 		}
4644 		if (rack->rc_gp_saw_ca &&
4645 		    rack->r_ctl.rack_per_upper_bound_ca &&
4646 		    (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4647 			    /*
4648 			     * In cases where we can't go higher
4649 			     * we should just use timely.
4650 			     */
4651 			    goto use_timely;
4652 		}
4653 		rack->rc_gp_bwred = 0;
4654 		rack->rc_gp_timely_dec_cnt = 0;
4655 		/* You get a set number of pushes if timely is trying to reduce */
4656 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4657 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4658 		} else {
4659 			/* Log it stays the same */
4660 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4661 			    __LINE__, 12);
4662 		}
4663 		return;
4664 	} else {
4665 		/*
4666 		 * We are staying between the lower and upper range bounds
4667 		 * so use timely to decide.
4668 		 */
4669 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4670 				__LINE__, 3);
4671 use_timely:
4672 		if (timely_says) {
4673 			rack->rc_gp_incr = 0;
4674 			rack->rc_gp_timely_inc_cnt = 0;
4675 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4676 			    !losses &&
4677 			    (last_bw_est < low_bnd)) {
4678 				/* We are loosing ground */
4679 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4680 				rack->rc_gp_timely_dec_cnt++;
4681 				/* We are not incrementing really no-count */
4682 				rack->rc_gp_incr = 0;
4683 				rack->rc_gp_timely_inc_cnt = 0;
4684 			} else
4685 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4686 		} else {
4687 			rack->rc_gp_bwred = 0;
4688 			rack->rc_gp_timely_dec_cnt = 0;
4689 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4690 		}
4691 	}
4692 }
4693 
4694 static int32_t
4695 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4696 {
4697 	int32_t timely_says;
4698 	uint64_t log_mult, log_rtt_a_diff;
4699 
4700 	log_rtt_a_diff = rtt;
4701 	log_rtt_a_diff <<= 32;
4702 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4703 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4704 		    rack_gp_rtt_maxmul)) {
4705 		/* Reduce the b/w multiplier */
4706 		timely_says = 2;
4707 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4708 		log_mult <<= 32;
4709 		log_mult |= prev_rtt;
4710 		rack_log_timely(rack,  timely_says, log_mult,
4711 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4712 				log_rtt_a_diff, __LINE__, 4);
4713 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4714 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4715 			    max(rack_gp_rtt_mindiv , 1)))) {
4716 		/* Increase the b/w multiplier */
4717 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4718 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4719 			 max(rack_gp_rtt_mindiv , 1));
4720 		log_mult <<= 32;
4721 		log_mult |= prev_rtt;
4722 		timely_says = 0;
4723 		rack_log_timely(rack,  timely_says, log_mult ,
4724 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4725 				log_rtt_a_diff, __LINE__, 5);
4726 	} else {
4727 		/*
4728 		 * Use a gradient to find it the timely gradient
4729 		 * is:
4730 		 * grad = rc_rtt_diff / min_rtt;
4731 		 *
4732 		 * anything below or equal to 0 will be
4733 		 * a increase indication. Anything above
4734 		 * zero is a decrease. Note we take care
4735 		 * of the actual gradient calculation
4736 		 * in the reduction (its not needed for
4737 		 * increase).
4738 		 */
4739 		log_mult = prev_rtt;
4740 		if (rtt_diff <= 0) {
4741 			/*
4742 			 * Rttdiff is less than zero, increase the
4743 			 * b/w multiplier (its 0 or negative)
4744 			 */
4745 			timely_says = 0;
4746 			rack_log_timely(rack,  timely_says, log_mult,
4747 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4748 		} else {
4749 			/* Reduce the b/w multiplier */
4750 			timely_says = 1;
4751 			rack_log_timely(rack,  timely_says, log_mult,
4752 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4753 		}
4754 	}
4755 	return (timely_says);
4756 }
4757 
4758 static __inline int
4759 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4760 {
4761 	if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4762 	    SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4763 		/**
4764 		 * This covers the case that the
4765 		 * resent is completely inside
4766 		 * the gp range or up to it.
4767 		 *      |----------------|
4768 		 *      |-----| <or>
4769 		 *            |----|
4770 		 *            <or>   |---|
4771 		 */
4772 		return (1);
4773 	} else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4774 		   SEQ_GT(rsm->r_end, tp->gput_seq)){
4775 		/**
4776 		 * This covers the case of
4777 		 *      |--------------|
4778 		 *  |-------->|
4779 		 */
4780 		return (1);
4781 	} else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4782 		   SEQ_LT(rsm->r_start, tp->gput_ack) &&
4783 		   SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4784 
4785 		/**
4786 		 * This covers the case of
4787 		 *      |--------------|
4788 		 *              |-------->|
4789 		 */
4790 		return (1);
4791 	}
4792 	return (0);
4793 }
4794 
4795 static __inline void
4796 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4797 {
4798 
4799 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
4800 		return;
4801 	/*
4802 	 * We have a Goodput measurement in progress. Mark
4803 	 * the send if its within the window. If its not
4804 	 * in the window make sure it does not have the mark.
4805 	 */
4806 	if (rack_in_gp_window(tp, rsm))
4807 		rsm->r_flags |= RACK_IN_GP_WIN;
4808 	else
4809 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4810 }
4811 
4812 static __inline void
4813 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4814 {
4815 	/* A GP measurement is ending, clear all marks on the send map*/
4816 	struct rack_sendmap *rsm = NULL;
4817 
4818 	rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4819 	if (rsm == NULL) {
4820 		rsm = tqhash_min(rack->r_ctl.tqh);
4821 	}
4822 	/* Nothing left? */
4823 	while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4824 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4825 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4826 	}
4827 }
4828 
4829 
4830 static __inline void
4831 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4832 {
4833 	struct rack_sendmap *rsm = NULL;
4834 
4835 	if (tp->snd_una == tp->snd_max) {
4836 		/* Nothing outstanding yet, nothing to do here */
4837 		return;
4838 	}
4839 	if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4840 		/*
4841 		 * We are measuring ahead of some outstanding
4842 		 * data. We need to walk through up until we get
4843 		 * to gp_seq marking so that no rsm is set incorrectly
4844 		 * with RACK_IN_GP_WIN.
4845 		 */
4846 		rsm = tqhash_min(rack->r_ctl.tqh);
4847 		while (rsm != NULL) {
4848 			rack_mark_in_gp_win(tp, rsm);
4849 			if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4850 				break;
4851 			rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4852 		}
4853 	}
4854 	if (rsm == NULL) {
4855 		/*
4856 		 * Need to find the GP seq, if rsm is
4857 		 * set we stopped as we hit it.
4858 		 */
4859 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4860 		if (rsm == NULL)
4861 			return;
4862 		rack_mark_in_gp_win(tp, rsm);
4863 	}
4864 	/*
4865 	 * Now we may need to mark already sent rsm, ahead of
4866 	 * gput_seq in the window since they may have been sent
4867 	 * *before* we started our measurment. The rsm, if non-null
4868 	 * has been marked (note if rsm would have been NULL we would have
4869 	 * returned in the previous block). So we go to the next, and continue
4870 	 * until we run out of entries or we exceed the gp_ack value.
4871 	 */
4872 	rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4873 	while (rsm) {
4874 		rack_mark_in_gp_win(tp, rsm);
4875 		if (SEQ_GT(rsm->r_end, tp->gput_ack))
4876 			break;
4877 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4878 	}
4879 }
4880 
4881 static void
4882 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)
4883 {
4884 	if (tcp_bblogging_on(rack->rc_tp)) {
4885 		union tcp_log_stackspecific log;
4886 		struct timeval tv;
4887 
4888 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4889 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4890 		log.u_bbr.flex1 = add_part;
4891 		log.u_bbr.flex2 = sub_part;
4892 		log.u_bbr.flex3 = rack_wma_divisor;
4893 		log.u_bbr.flex4 = srtt;
4894 		log.u_bbr.flex7 = (uint16_t)line;
4895 		log.u_bbr.flex8 = meth;
4896 		log.u_bbr.delRate = rack->r_ctl.gp_bw;
4897 		log.u_bbr.cur_del_rate = meas_bw;
4898 		log.u_bbr.rttProp = utim;
4899 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4900 		    &rack->rc_inp->inp_socket->so_rcv,
4901 		    &rack->rc_inp->inp_socket->so_snd,
4902 		    BBR_LOG_THRESH_CALC, 0,
4903 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4904 	}
4905 }
4906 
4907 static void
4908 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4909 			    tcp_seq th_ack, int line, uint8_t quality)
4910 {
4911 	uint64_t tim, bytes_ps, stim, utim;
4912 	uint32_t segsiz, bytes, reqbytes, us_cts;
4913 	int32_t gput, new_rtt_diff, timely_says;
4914 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4915 	int did_add = 0;
4916 
4917 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4918 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4919 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4920 		tim = us_cts - tp->gput_ts;
4921 	else
4922 		tim = 0;
4923 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4924 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4925 	else
4926 		stim = 0;
4927 	/*
4928 	 * Use the larger of the send time or ack time. This prevents us
4929 	 * from being influenced by ack artifacts to come up with too
4930 	 * high of measurement. Note that since we are spanning over many more
4931 	 * bytes in most of our measurements hopefully that is less likely to
4932 	 * occur.
4933 	 */
4934 	if (tim > stim)
4935 		utim = max(tim, 1);
4936 	else
4937 		utim = max(stim, 1);
4938 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4939 	rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
4940 	if ((tim == 0) && (stim == 0)) {
4941 		/*
4942 		 * Invalid measurement time, maybe
4943 		 * all on one ack/one send?
4944 		 */
4945 		bytes = 0;
4946 		bytes_ps = 0;
4947 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4948 					   0, 0, 0, 10, __LINE__, NULL, quality);
4949 		goto skip_measurement;
4950 	}
4951 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4952 		/* We never made a us_rtt measurement? */
4953 		bytes = 0;
4954 		bytes_ps = 0;
4955 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4956 					   0, 0, 0, 10, __LINE__, NULL, quality);
4957 		goto skip_measurement;
4958 	}
4959 	/*
4960 	 * Calculate the maximum possible b/w this connection
4961 	 * could have. We base our calculation on the lowest
4962 	 * rtt we have seen during the measurement and the
4963 	 * largest rwnd the client has given us in that time. This
4964 	 * forms a BDP that is the maximum that we could ever
4965 	 * get to the client. Anything larger is not valid.
4966 	 *
4967 	 * I originally had code here that rejected measurements
4968 	 * where the time was less than 1/2 the latest us_rtt.
4969 	 * But after thinking on that I realized its wrong since
4970 	 * say you had a 150Mbps or even 1Gbps link, and you
4971 	 * were a long way away.. example I am in Europe (100ms rtt)
4972 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4973 	 * bytes my time would be 1.2ms, and yet my rtt would say
4974 	 * the measurement was invalid the time was < 50ms. The
4975 	 * same thing is true for 150Mb (8ms of time).
4976 	 *
4977 	 * A better way I realized is to look at what the maximum
4978 	 * the connection could possibly do. This is gated on
4979 	 * the lowest RTT we have seen and the highest rwnd.
4980 	 * We should in theory never exceed that, if we are
4981 	 * then something on the path is storing up packets
4982 	 * and then feeding them all at once to our endpoint
4983 	 * messing up our measurement.
4984 	 */
4985 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4986 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4987 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4988 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4989 		/* No measurement can be made */
4990 		bytes = 0;
4991 		bytes_ps = 0;
4992 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4993 					   0, 0, 0, 10, __LINE__, NULL, quality);
4994 		goto skip_measurement;
4995 	} else
4996 		bytes = (th_ack - tp->gput_seq);
4997 	bytes_ps = (uint64_t)bytes;
4998 	/*
4999 	 * Don't measure a b/w for pacing unless we have gotten at least
5000 	 * an initial windows worth of data in this measurement interval.
5001 	 *
5002 	 * Small numbers of bytes get badly influenced by delayed ack and
5003 	 * other artifacts. Note we take the initial window or our
5004 	 * defined minimum GP (defaulting to 10 which hopefully is the
5005 	 * IW).
5006 	 */
5007 	if (rack->rc_gp_filled == 0) {
5008 		/*
5009 		 * The initial estimate is special. We
5010 		 * have blasted out an IW worth of packets
5011 		 * without a real valid ack ts results. We
5012 		 * then setup the app_limited_needs_set flag,
5013 		 * this should get the first ack in (probably 2
5014 		 * MSS worth) to be recorded as the timestamp.
5015 		 * We thus allow a smaller number of bytes i.e.
5016 		 * IW - 2MSS.
5017 		 */
5018 		reqbytes -= (2 * segsiz);
5019 		/* Also lets fill previous for our first measurement to be neutral */
5020 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5021 	}
5022 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
5023 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5024 					   rack->r_ctl.rc_app_limited_cnt,
5025 					   0, 0, 10, __LINE__, NULL, quality);
5026 		goto skip_measurement;
5027 	}
5028 	/*
5029 	 * We now need to calculate the Timely like status so
5030 	 * we can update (possibly) the b/w multipliers.
5031 	 */
5032 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
5033 	if (rack->rc_gp_filled == 0) {
5034 		/* No previous reading */
5035 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
5036 	} else {
5037 		if (rack->measure_saw_probe_rtt == 0) {
5038 			/*
5039 			 * We don't want a probertt to be counted
5040 			 * since it will be negative incorrectly. We
5041 			 * expect to be reducing the RTT when we
5042 			 * pace at a slower rate.
5043 			 */
5044 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
5045 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
5046 		}
5047 	}
5048 	timely_says = rack_make_timely_judgement(rack,
5049 	    rack->r_ctl.rc_gp_srtt,
5050 	    rack->r_ctl.rc_rtt_diff,
5051 	    rack->r_ctl.rc_prev_gp_srtt
5052 	);
5053 	bytes_ps *= HPTS_USEC_IN_SEC;
5054 	bytes_ps /= utim;
5055 	if (bytes_ps > rack->r_ctl.last_max_bw) {
5056 		/*
5057 		 * Something is on path playing
5058 		 * since this b/w is not possible based
5059 		 * on our BDP (highest rwnd and lowest rtt
5060 		 * we saw in the measurement window).
5061 		 *
5062 		 * Another option here would be to
5063 		 * instead skip the measurement.
5064 		 */
5065 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
5066 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
5067 					   11, __LINE__, NULL, quality);
5068 		bytes_ps = rack->r_ctl.last_max_bw;
5069 	}
5070 	/* We store gp for b/w in bytes per second */
5071 	if (rack->rc_gp_filled == 0) {
5072 		/* Initial measurement */
5073 		if (bytes_ps) {
5074 			rack->r_ctl.gp_bw = bytes_ps;
5075 			rack->rc_gp_filled = 1;
5076 			rack->r_ctl.num_measurements = 1;
5077 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5078 		} else {
5079 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5080 						   rack->r_ctl.rc_app_limited_cnt,
5081 						   0, 0, 10, __LINE__, NULL, quality);
5082 		}
5083 		if (tcp_in_hpts(rack->rc_tp) &&
5084 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5085 			/*
5086 			 * Ok we can't trust the pacer in this case
5087 			 * where we transition from un-paced to paced.
5088 			 * Or for that matter when the burst mitigation
5089 			 * was making a wild guess and got it wrong.
5090 			 * Stop the pacer and clear up all the aggregate
5091 			 * delays etc.
5092 			 */
5093 			tcp_hpts_remove(rack->rc_tp);
5094 			rack->r_ctl.rc_hpts_flags = 0;
5095 			rack->r_ctl.rc_last_output_to = 0;
5096 		}
5097 		did_add = 2;
5098 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5099 		/* Still a small number run an average */
5100 		rack->r_ctl.gp_bw += bytes_ps;
5101 		addpart = rack->r_ctl.num_measurements;
5102 		rack->r_ctl.num_measurements++;
5103 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5104 			/* We have collected enough to move forward */
5105 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5106 		}
5107 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5108 		did_add = 3;
5109 	} else {
5110 		/*
5111 		 * We want to take 1/wma of the goodput and add in to 7/8th
5112 		 * of the old value weighted by the srtt. So if your measurement
5113 		 * period is say 2 SRTT's long you would get 1/4 as the
5114 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
5115 		 *
5116 		 * But we must be careful not to take too much i.e. if the
5117 		 * srtt is say 20ms and the measurement is taken over
5118 		 * 400ms our weight would be 400/20 i.e. 20. On the
5119 		 * other hand if we get a measurement over 1ms with a
5120 		 * 10ms rtt we only want to take a much smaller portion.
5121 		 */
5122 		uint8_t meth;
5123 
5124 		if (rack->r_ctl.num_measurements < 0xff) {
5125 			rack->r_ctl.num_measurements++;
5126 		}
5127 		srtt = (uint64_t)tp->t_srtt;
5128 		if (srtt == 0) {
5129 			/*
5130 			 * Strange why did t_srtt go back to zero?
5131 			 */
5132 			if (rack->r_ctl.rc_rack_min_rtt)
5133 				srtt = rack->r_ctl.rc_rack_min_rtt;
5134 			else
5135 				srtt = HPTS_USEC_IN_MSEC;
5136 		}
5137 		/*
5138 		 * XXXrrs: Note for reviewers, in playing with
5139 		 * dynamic pacing I discovered this GP calculation
5140 		 * as done originally leads to some undesired results.
5141 		 * Basically you can get longer measurements contributing
5142 		 * too much to the WMA. Thus I changed it if you are doing
5143 		 * dynamic adjustments to only do the aportioned adjustment
5144 		 * if we have a very small (time wise) measurement. Longer
5145 		 * measurements just get there weight (defaulting to 1/8)
5146 		 * add to the WMA. We may want to think about changing
5147 		 * this to always do that for both sides i.e. dynamic
5148 		 * and non-dynamic... but considering lots of folks
5149 		 * were playing with this I did not want to change the
5150 		 * calculation per.se. without your thoughts.. Lawerence?
5151 		 * Peter??
5152 		 */
5153 		if (rack->rc_gp_dyn_mul == 0) {
5154 			subpart = rack->r_ctl.gp_bw * utim;
5155 			subpart /= (srtt * 8);
5156 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
5157 				/*
5158 				 * The b/w update takes no more
5159 				 * away then 1/2 our running total
5160 				 * so factor it in.
5161 				 */
5162 				addpart = bytes_ps * utim;
5163 				addpart /= (srtt * 8);
5164 				meth = 1;
5165 			} else {
5166 				/*
5167 				 * Don't allow a single measurement
5168 				 * to account for more than 1/2 of the
5169 				 * WMA. This could happen on a retransmission
5170 				 * where utim becomes huge compared to
5171 				 * srtt (multiple retransmissions when using
5172 				 * the sending rate which factors in all the
5173 				 * transmissions from the first one).
5174 				 */
5175 				subpart = rack->r_ctl.gp_bw / 2;
5176 				addpart = bytes_ps / 2;
5177 				meth = 2;
5178 			}
5179 			rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5180 			resid_bw = rack->r_ctl.gp_bw - subpart;
5181 			rack->r_ctl.gp_bw = resid_bw + addpart;
5182 			did_add = 1;
5183 		} else {
5184 			if ((utim / srtt) <= 1) {
5185 				/*
5186 				 * The b/w update was over a small period
5187 				 * of time. The idea here is to prevent a small
5188 				 * measurement time period from counting
5189 				 * too much. So we scale it based on the
5190 				 * time so it attributes less than 1/rack_wma_divisor
5191 				 * of its measurement.
5192 				 */
5193 				subpart = rack->r_ctl.gp_bw * utim;
5194 				subpart /= (srtt * rack_wma_divisor);
5195 				addpart = bytes_ps * utim;
5196 				addpart /= (srtt * rack_wma_divisor);
5197 				meth = 3;
5198 			} else {
5199 				/*
5200 				 * The scaled measurement was long
5201 				 * enough so lets just add in the
5202 				 * portion of the measurement i.e. 1/rack_wma_divisor
5203 				 */
5204 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5205 				addpart = bytes_ps / rack_wma_divisor;
5206 				meth = 4;
5207 			}
5208 			if ((rack->measure_saw_probe_rtt == 0) ||
5209 		            (bytes_ps > rack->r_ctl.gp_bw)) {
5210 				/*
5211 				 * For probe-rtt we only add it in
5212 				 * if its larger, all others we just
5213 				 * add in.
5214 				 */
5215 				did_add = 1;
5216 				rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5217 				resid_bw = rack->r_ctl.gp_bw - subpart;
5218 				rack->r_ctl.gp_bw = resid_bw + addpart;
5219 			}
5220 		}
5221 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5222 	}
5223 	/*
5224 	 * We only watch the growth of the GP during the initial startup
5225 	 * or first-slowstart that ensues. If we ever needed to watch
5226 	 * growth of gp outside of that period all we need to do is
5227 	 * remove the first clause of this if (rc_initial_ss_comp).
5228 	 */
5229 	if ((rack->rc_initial_ss_comp == 0) &&
5230 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG)) {
5231 		uint64_t gp_est;
5232 
5233 		gp_est = bytes_ps;
5234 		if (tcp_bblogging_on(rack->rc_tp)) {
5235 			union tcp_log_stackspecific log;
5236 			struct timeval tv;
5237 
5238 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5239 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5240 			log.u_bbr.flex1 = rack->r_ctl.current_round;
5241 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
5242 			log.u_bbr.delRate = gp_est;
5243 			log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5244 			log.u_bbr.flex8 = 41;
5245 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5246 					    0, &log, false, NULL, __func__, __LINE__,&tv);
5247 		}
5248 		if ((rack->r_ctl.num_measurements == RACK_REQ_AVG) ||
5249 		    (rack->r_ctl.last_gpest == 0)) {
5250 			/*
5251 			 * The round we get our measurement averaging going
5252 			 * is the base round so it always is the source point
5253 			 * for when we had our first increment. From there on
5254 			 * we only record the round that had a rise.
5255 			 */
5256 			rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5257 			rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5258 		} else if (gp_est >= rack->r_ctl.last_gpest) {
5259 			/*
5260 			 * Test to see if its gone up enough
5261 			 * to set the round count up to now. Note
5262 			 * that on the seeding of the 4th measurement we
5263 			 */
5264 			gp_est *= 1000;
5265 			gp_est /= rack->r_ctl.last_gpest;
5266 			if ((uint32_t)gp_est > rack->r_ctl.gp_gain_req) {
5267 				/*
5268 				 * We went up enough to record the round.
5269 				 */
5270 				if (tcp_bblogging_on(rack->rc_tp)) {
5271 					union tcp_log_stackspecific log;
5272 					struct timeval tv;
5273 
5274 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5275 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5276 					log.u_bbr.flex1 = rack->r_ctl.current_round;
5277 					log.u_bbr.flex2 = (uint32_t)gp_est;
5278 					log.u_bbr.flex3 = rack->r_ctl.gp_gain_req;
5279 					log.u_bbr.delRate = gp_est;
5280 					log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5281 					log.u_bbr.flex8 = 42;
5282 					(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5283 							    0, &log, false, NULL, __func__, __LINE__,&tv);
5284 				}
5285 				rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5286 				if (rack->r_ctl.use_gp_not_last == 1)
5287 					rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5288 				else
5289 					rack->r_ctl.last_gpest = bytes_ps;
5290 			}
5291 		}
5292 	}
5293 	if ((rack->gp_ready == 0) &&
5294 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5295 		/* We have enough measurements now */
5296 		rack->gp_ready = 1;
5297 		if (rack->dgp_on ||
5298 		    rack->rack_hibeta)
5299 			rack_set_cc_pacing(rack);
5300 		if (rack->defer_options)
5301 			rack_apply_deferred_options(rack);
5302 	}
5303 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5304 				   rack_get_bw(rack), 22, did_add, NULL, quality);
5305 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
5306 
5307 	if ((rack->measure_saw_probe_rtt == 0) &&
5308 	    rack->rc_gp_rtt_set) {
5309 		if (rack->rc_skip_timely == 0) {
5310 			rack_update_multiplier(rack, timely_says, bytes_ps,
5311 					       rack->r_ctl.rc_gp_srtt,
5312 					       rack->r_ctl.rc_rtt_diff);
5313 		}
5314 	}
5315 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5316 				   rack_get_bw(rack), 3, line, NULL, quality);
5317 	rack_log_pacing_delay_calc(rack,
5318 				   bytes, /* flex2 */
5319 				   tim, /* flex1 */
5320 				   bytes_ps, /* bw_inuse */
5321 				   rack->r_ctl.gp_bw, /* delRate */
5322 				   rack_get_lt_bw(rack), /* rttProp */
5323 				   20, line, NULL, 0);
5324 	/* reset the gp srtt and setup the new prev */
5325 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5326 	/* Record the lost count for the next measurement */
5327 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5328 skip_measurement:
5329 	/*
5330 	 * We restart our diffs based on the gpsrtt in the
5331 	 * measurement window.
5332 	 */
5333 	rack->rc_gp_rtt_set = 0;
5334 	rack->rc_gp_saw_rec = 0;
5335 	rack->rc_gp_saw_ca = 0;
5336 	rack->rc_gp_saw_ss = 0;
5337 	rack->rc_dragged_bottom = 0;
5338 	if (quality == RACK_QUALITY_HIGH) {
5339 		/*
5340 		 * Gput in the stats world is in kbps where bytes_ps is
5341 		 * bytes per second so we do ((x * 8)/ 1000).
5342 		 */
5343 		gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5344 #ifdef STATS
5345 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5346 					 gput);
5347 		/*
5348 		 * XXXLAS: This is a temporary hack, and should be
5349 		 * chained off VOI_TCP_GPUT when stats(9) grows an
5350 		 * API to deal with chained VOIs.
5351 		 */
5352 		if (tp->t_stats_gput_prev > 0)
5353 			stats_voi_update_abs_s32(tp->t_stats,
5354 						 VOI_TCP_GPUT_ND,
5355 						 ((gput - tp->t_stats_gput_prev) * 100) /
5356 						 tp->t_stats_gput_prev);
5357 #endif
5358 		tp->t_stats_gput_prev = gput;
5359 	}
5360 	tp->t_flags &= ~TF_GPUTINPROG;
5361 	/*
5362 	 * Now are we app limited now and there is space from where we
5363 	 * were to where we want to go?
5364 	 *
5365 	 * We don't do the other case i.e. non-applimited here since
5366 	 * the next send will trigger us picking up the missing data.
5367 	 */
5368 	if (rack->r_ctl.rc_first_appl &&
5369 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
5370 	    rack->r_ctl.rc_app_limited_cnt &&
5371 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5372 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5373 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5374 		/*
5375 		 * Yep there is enough outstanding to make a measurement here.
5376 		 */
5377 		struct rack_sendmap *rsm;
5378 
5379 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5380 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5381 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5382 		rack->app_limited_needs_set = 0;
5383 		tp->gput_seq = th_ack;
5384 		if (rack->in_probe_rtt)
5385 			rack->measure_saw_probe_rtt = 1;
5386 		else if ((rack->measure_saw_probe_rtt) &&
5387 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5388 			rack->measure_saw_probe_rtt = 0;
5389 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5390 			/* There is a full window to gain info from */
5391 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5392 		} else {
5393 			/* We can only measure up to the applimited point */
5394 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5395 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5396 				/*
5397 				 * We don't have enough to make a measurement.
5398 				 */
5399 				tp->t_flags &= ~TF_GPUTINPROG;
5400 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5401 							   0, 0, 0, 6, __LINE__, NULL, quality);
5402 				return;
5403 			}
5404 		}
5405 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
5406 			/*
5407 			 * We will get no more data into the SB
5408 			 * this means we need to have the data available
5409 			 * before we start a measurement.
5410 			 */
5411 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5412 				/* Nope not enough data. */
5413 				return;
5414 			}
5415 		}
5416 		tp->t_flags |= TF_GPUTINPROG;
5417 		/*
5418 		 * Now we need to find the timestamp of the send at tp->gput_seq
5419 		 * for the send based measurement.
5420 		 */
5421 		rack->r_ctl.rc_gp_cumack_ts = 0;
5422 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5423 		if (rsm) {
5424 			/* Ok send-based limit is set */
5425 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5426 				/*
5427 				 * Move back to include the earlier part
5428 				 * so our ack time lines up right (this may
5429 				 * make an overlapping measurement but thats
5430 				 * ok).
5431 				 */
5432 				tp->gput_seq = rsm->r_start;
5433 			}
5434 			if (rsm->r_flags & RACK_ACKED) {
5435 				struct rack_sendmap *nrsm;
5436 
5437 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5438 				tp->gput_seq = rsm->r_end;
5439 				nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5440 				if (nrsm)
5441 					rsm = nrsm;
5442 				else {
5443 					rack->app_limited_needs_set = 1;
5444 				}
5445 			} else
5446 				rack->app_limited_needs_set = 1;
5447 			/* We always go from the first send */
5448 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5449 		} else {
5450 			/*
5451 			 * If we don't find the rsm due to some
5452 			 * send-limit set the current time, which
5453 			 * basically disables the send-limit.
5454 			 */
5455 			struct timeval tv;
5456 
5457 			microuptime(&tv);
5458 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5459 		}
5460 		rack_tend_gp_marks(tp, rack);
5461 		rack_log_pacing_delay_calc(rack,
5462 					   tp->gput_seq,
5463 					   tp->gput_ack,
5464 					   (uintptr_t)rsm,
5465 					   tp->gput_ts,
5466 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5467 					   9,
5468 					   __LINE__, rsm, quality);
5469 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5470 	} else {
5471 		/*
5472 		 * To make sure proper timestamp merging occurs, we need to clear
5473 		 * all GP marks if we don't start a measurement.
5474 		 */
5475 		rack_clear_gp_marks(tp, rack);
5476 	}
5477 }
5478 
5479 /*
5480  * CC wrapper hook functions
5481  */
5482 static void
5483 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5484     uint16_t type, int32_t post_recovery)
5485 {
5486 	uint32_t prior_cwnd, acked;
5487 	struct tcp_log_buffer *lgb = NULL;
5488 	uint8_t labc_to_use, quality;
5489 
5490 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5491 	tp->t_ccv.nsegs = nsegs;
5492 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5493 	if ((post_recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5494 		uint32_t max;
5495 
5496 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5497 		if (tp->t_ccv.bytes_this_ack > max) {
5498 			tp->t_ccv.bytes_this_ack = max;
5499 		}
5500 	}
5501 #ifdef STATS
5502 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5503 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5504 #endif
5505 	if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5506 		/*
5507 		 * We will ack all the data, time to end any
5508 		 * lt_bw_up we have running until something
5509 		 * new is sent. Note we need to use the actual
5510 		 * ack_rcv_time which with pacing may be different.
5511 		 */
5512 		uint64_t tmark;
5513 
5514 		rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5515 		rack->r_ctl.lt_seq = tp->snd_max;
5516 		tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
5517 		if (tmark >= rack->r_ctl.lt_timemark) {
5518 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
5519 		}
5520 		rack->r_ctl.lt_timemark = tmark;
5521 		rack->lt_bw_up = 0;
5522 	}
5523 	quality = RACK_QUALITY_NONE;
5524 	if ((tp->t_flags & TF_GPUTINPROG) &&
5525 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5526 		/* Measure the Goodput */
5527 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5528 	}
5529 	/* Which way our we limited, if not cwnd limited no advance in CA */
5530 	if (tp->snd_cwnd <= tp->snd_wnd)
5531 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
5532 	else
5533 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5534 	if (tp->snd_cwnd > tp->snd_ssthresh) {
5535 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5536 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5537 		/* For the setting of a window past use the actual scwnd we are using */
5538 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5539 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5540 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5541 		}
5542 	} else {
5543 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5544 		tp->t_bytes_acked = 0;
5545 	}
5546 	prior_cwnd = tp->snd_cwnd;
5547 	if ((post_recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5548 	    (rack_client_low_buf && rack->client_bufferlvl &&
5549 	    (rack->client_bufferlvl < rack_client_low_buf)))
5550 		labc_to_use = rack->rc_labc;
5551 	else
5552 		labc_to_use = rack_max_abc_post_recovery;
5553 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5554 		union tcp_log_stackspecific log;
5555 		struct timeval tv;
5556 
5557 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5558 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5559 		log.u_bbr.flex1 = th_ack;
5560 		log.u_bbr.flex2 = tp->t_ccv.flags;
5561 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5562 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5563 		log.u_bbr.flex5 = labc_to_use;
5564 		log.u_bbr.flex6 = prior_cwnd;
5565 		log.u_bbr.flex7 = V_tcp_do_newsack;
5566 		log.u_bbr.flex8 = 1;
5567 		lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5568 				     0, &log, false, NULL, __func__, __LINE__,&tv);
5569 	}
5570 	if (CC_ALGO(tp)->ack_received != NULL) {
5571 		/* XXXLAS: Find a way to live without this */
5572 		tp->t_ccv.curack = th_ack;
5573 		tp->t_ccv.labc = labc_to_use;
5574 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5575 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5576 	}
5577 	if (lgb) {
5578 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5579 	}
5580 	if (rack->r_must_retran) {
5581 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5582 			/*
5583 			 * We now are beyond the rxt point so lets disable
5584 			 * the flag.
5585 			 */
5586 			rack->r_ctl.rc_out_at_rto = 0;
5587 			rack->r_must_retran = 0;
5588 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5589 			/*
5590 			 * Only decrement the rc_out_at_rto if the cwnd advances
5591 			 * at least a whole segment. Otherwise next time the peer
5592 			 * acks, we won't be able to send this generaly happens
5593 			 * when we are in Congestion Avoidance.
5594 			 */
5595 			if (acked <= rack->r_ctl.rc_out_at_rto){
5596 				rack->r_ctl.rc_out_at_rto -= acked;
5597 			} else {
5598 				rack->r_ctl.rc_out_at_rto = 0;
5599 			}
5600 		}
5601 	}
5602 #ifdef STATS
5603 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5604 #endif
5605 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5606 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5607 	}
5608 	if ((rack->rc_initial_ss_comp == 0) &&
5609 	    (tp->snd_cwnd >= tp->snd_ssthresh)) {
5610 		/*
5611 		 * The cwnd has grown beyond ssthresh we have
5612 		 * entered ca and completed our first Slowstart.
5613 		 */
5614 		rack->rc_initial_ss_comp = 1;
5615 	}
5616 }
5617 
5618 static void
5619 tcp_rack_partialack(struct tcpcb *tp)
5620 {
5621 	struct tcp_rack *rack;
5622 
5623 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5624 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5625 	/*
5626 	 * If we are doing PRR and have enough
5627 	 * room to send <or> we are pacing and prr
5628 	 * is disabled we will want to see if we
5629 	 * can send data (by setting r_wanted_output to
5630 	 * true).
5631 	 */
5632 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5633 	    rack->rack_no_prr)
5634 		rack->r_wanted_output = 1;
5635 }
5636 
5637 static void
5638 rack_exit_recovery(struct tcpcb *tp, struct tcp_rack *rack, int how)
5639 {
5640 	/*
5641 	 * Now exit recovery.
5642 	 */
5643 	EXIT_RECOVERY(tp->t_flags);
5644 }
5645 
5646 static void
5647 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
5648 {
5649 	struct tcp_rack *rack;
5650 	uint32_t orig_cwnd;
5651 
5652 	orig_cwnd = tp->snd_cwnd;
5653 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5654 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5655 	/* only alert CC if we alerted when we entered */
5656 	if (CC_ALGO(tp)->post_recovery != NULL) {
5657 		tp->t_ccv.curack = th_ack;
5658 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
5659 		if (tp->snd_cwnd < tp->snd_ssthresh) {
5660 			/*
5661 			 * Rack has burst control and pacing
5662 			 * so lets not set this any lower than
5663 			 * snd_ssthresh per RFC-6582 (option 2).
5664 			 */
5665 			tp->snd_cwnd = tp->snd_ssthresh;
5666 		}
5667 	}
5668 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5669 		union tcp_log_stackspecific log;
5670 		struct timeval tv;
5671 
5672 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5673 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5674 		log.u_bbr.flex1 = th_ack;
5675 		log.u_bbr.flex2 = tp->t_ccv.flags;
5676 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5677 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5678 		log.u_bbr.flex5 = V_tcp_abc_l_var;
5679 		log.u_bbr.flex6 = orig_cwnd;
5680 		log.u_bbr.flex7 = V_tcp_do_newsack;
5681 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
5682 		log.u_bbr.flex8 = 2;
5683 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5684 			       0, &log, false, NULL, __func__, __LINE__, &tv);
5685 	}
5686 	if ((rack->rack_no_prr == 0) &&
5687 	    (rack->no_prr_addback == 0) &&
5688 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
5689 		/*
5690 		 * Suck the next prr cnt back into cwnd, but
5691 		 * only do that if we are not application limited.
5692 		 */
5693 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
5694 			/*
5695 			 * We are allowed to add back to the cwnd the amount we did
5696 			 * not get out if:
5697 			 * a) no_prr_addback is off.
5698 			 * b) we are not app limited
5699 			 * c) we are doing prr
5700 			 * <and>
5701 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
5702 			 */
5703 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
5704 					    rack->r_ctl.rc_prr_sndcnt);
5705 		}
5706 		rack->r_ctl.rc_prr_sndcnt = 0;
5707 		rack_log_to_prr(rack, 1, 0, __LINE__);
5708 	}
5709 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
5710 	tp->snd_recover = tp->snd_una;
5711 	if (rack->r_ctl.dsack_persist) {
5712 		rack->r_ctl.dsack_persist--;
5713 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
5714 			rack->r_ctl.num_dsack = 0;
5715 		}
5716 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
5717 	}
5718 	if (rack->rto_from_rec == 1) {
5719 		rack->rto_from_rec = 0;
5720 		if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
5721 			tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
5722 	}
5723 	rack_exit_recovery(tp, rack, 1);
5724 }
5725 
5726 static void
5727 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
5728 {
5729 	struct tcp_rack *rack;
5730 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
5731 
5732 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5733 #ifdef STATS
5734 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
5735 #endif
5736 	if (IN_RECOVERY(tp->t_flags) == 0) {
5737 		in_rec_at_entry = 0;
5738 		ssthresh_enter = tp->snd_ssthresh;
5739 		cwnd_enter = tp->snd_cwnd;
5740 	} else
5741 		in_rec_at_entry = 1;
5742 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5743 	switch (type) {
5744 	case CC_NDUPACK:
5745 		tp->t_flags &= ~TF_WASFRECOVERY;
5746 		tp->t_flags &= ~TF_WASCRECOVERY;
5747 		if (!IN_FASTRECOVERY(tp->t_flags)) {
5748 			/* Check if this is the end of the initial Start-up i.e. initial slow-start */
5749 			if (rack->rc_initial_ss_comp == 0) {
5750 				/* Yep it is the end of the initial slowstart */
5751 				rack->rc_initial_ss_comp = 1;
5752 			}
5753 			rack->r_ctl.rc_prr_delivered = 0;
5754 			rack->r_ctl.rc_prr_out = 0;
5755 			rack->r_fast_output = 0;
5756 			if (rack->rack_no_prr == 0) {
5757 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5758 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
5759 			}
5760 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
5761 			tp->snd_recover = tp->snd_max;
5762 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5763 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5764 		}
5765 		break;
5766 	case CC_ECN:
5767 		if (!IN_CONGRECOVERY(tp->t_flags) ||
5768 		    /*
5769 		     * Allow ECN reaction on ACK to CWR, if
5770 		     * that data segment was also CE marked.
5771 		     */
5772 		    SEQ_GEQ(ack, tp->snd_recover)) {
5773 			EXIT_CONGRECOVERY(tp->t_flags);
5774 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5775 			rack->r_fast_output = 0;
5776 			tp->snd_recover = tp->snd_max + 1;
5777 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5778 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5779 		}
5780 		break;
5781 	case CC_RTO:
5782 		tp->t_dupacks = 0;
5783 		tp->t_bytes_acked = 0;
5784 		rack->r_fast_output = 0;
5785 		if (IN_RECOVERY(tp->t_flags))
5786 			rack_exit_recovery(tp, rack, 2);
5787 		orig_cwnd = tp->snd_cwnd;
5788 		rack_log_to_prr(rack, 16, orig_cwnd, line);
5789 		if (CC_ALGO(tp)->cong_signal == NULL) {
5790 			/* TSNH */
5791 			tp->snd_ssthresh = max(2,
5792 			    min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5793 			    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5794 			tp->snd_cwnd = ctf_fixed_maxseg(tp);
5795 		}
5796 		if (tp->t_flags2 & TF2_ECN_PERMIT)
5797 			tp->t_flags2 |= TF2_ECN_SND_CWR;
5798 		break;
5799 	case CC_RTO_ERR:
5800 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5801 		/* RTO was unnecessary, so reset everything. */
5802 		tp->snd_cwnd = tp->snd_cwnd_prev;
5803 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
5804 		tp->snd_recover = tp->snd_recover_prev;
5805 		if (tp->t_flags & TF_WASFRECOVERY) {
5806 			ENTER_FASTRECOVERY(tp->t_flags);
5807 			tp->t_flags &= ~TF_WASFRECOVERY;
5808 		}
5809 		if (tp->t_flags & TF_WASCRECOVERY) {
5810 			ENTER_CONGRECOVERY(tp->t_flags);
5811 			tp->t_flags &= ~TF_WASCRECOVERY;
5812 		}
5813 		tp->snd_nxt = tp->snd_max;
5814 		tp->t_badrxtwin = 0;
5815 		break;
5816 	}
5817 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
5818 	    (type != CC_RTO)){
5819 		tp->t_ccv.curack = ack;
5820 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
5821 	}
5822 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5823 		rack_log_to_prr(rack, 15, cwnd_enter, line);
5824 		rack->r_ctl.dsack_byte_cnt = 0;
5825 		rack->r_ctl.retran_during_recovery = 0;
5826 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5827 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5828 		rack->r_ent_rec_ns = 1;
5829 	}
5830 }
5831 
5832 static inline void
5833 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5834 {
5835 	uint32_t i_cwnd;
5836 
5837 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5838 
5839 	if (CC_ALGO(tp)->after_idle != NULL)
5840 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
5841 
5842 	if (tp->snd_cwnd == 1)
5843 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
5844 	else
5845 		i_cwnd = rc_init_window(rack);
5846 
5847 	/*
5848 	 * Being idle is no different than the initial window. If the cc
5849 	 * clamps it down below the initial window raise it to the initial
5850 	 * window.
5851 	 */
5852 	if (tp->snd_cwnd < i_cwnd) {
5853 		tp->snd_cwnd = i_cwnd;
5854 	}
5855 }
5856 
5857 /*
5858  * Indicate whether this ack should be delayed.  We can delay the ack if
5859  * following conditions are met:
5860  *	- There is no delayed ack timer in progress.
5861  *	- Our last ack wasn't a 0-sized window. We never want to delay
5862  *	  the ack that opens up a 0-sized window.
5863  *	- LRO wasn't used for this segment. We make sure by checking that the
5864  *	  segment size is not larger than the MSS.
5865  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
5866  *	  connection.
5867  */
5868 #define DELAY_ACK(tp, tlen)			 \
5869 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
5870 	((tp->t_flags & TF_DELACK) == 0) &&	 \
5871 	(tlen <= tp->t_maxseg) &&		 \
5872 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
5873 
5874 static struct rack_sendmap *
5875 rack_find_lowest_rsm(struct tcp_rack *rack)
5876 {
5877 	struct rack_sendmap *rsm;
5878 
5879 	/*
5880 	 * Walk the time-order transmitted list looking for an rsm that is
5881 	 * not acked. This will be the one that was sent the longest time
5882 	 * ago that is still outstanding.
5883 	 */
5884 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
5885 		if (rsm->r_flags & RACK_ACKED) {
5886 			continue;
5887 		}
5888 		goto finish;
5889 	}
5890 finish:
5891 	return (rsm);
5892 }
5893 
5894 static struct rack_sendmap *
5895 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5896 {
5897 	struct rack_sendmap *prsm;
5898 
5899 	/*
5900 	 * Walk the sequence order list backward until we hit and arrive at
5901 	 * the highest seq not acked. In theory when this is called it
5902 	 * should be the last segment (which it was not).
5903 	 */
5904 	prsm = rsm;
5905 
5906 	TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
5907 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5908 			continue;
5909 		}
5910 		return (prsm);
5911 	}
5912 	return (NULL);
5913 }
5914 
5915 static uint32_t
5916 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts, int line, int log_allowed)
5917 {
5918 	int32_t lro;
5919 	uint32_t thresh;
5920 
5921 	/*
5922 	 * lro is the flag we use to determine if we have seen reordering.
5923 	 * If it gets set we have seen reordering. The reorder logic either
5924 	 * works in one of two ways:
5925 	 *
5926 	 * If reorder-fade is configured, then we track the last time we saw
5927 	 * re-ordering occur. If we reach the point where enough time as
5928 	 * passed we no longer consider reordering has occuring.
5929 	 *
5930 	 * Or if reorder-face is 0, then once we see reordering we consider
5931 	 * the connection to alway be subject to reordering and just set lro
5932 	 * to 1.
5933 	 *
5934 	 * In the end if lro is non-zero we add the extra time for
5935 	 * reordering in.
5936 	 */
5937 	if (srtt == 0)
5938 		srtt = 1;
5939 	if (rack->r_ctl.rc_reorder_ts) {
5940 		if (rack->r_ctl.rc_reorder_fade) {
5941 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5942 				lro = cts - rack->r_ctl.rc_reorder_ts;
5943 				if (lro == 0) {
5944 					/*
5945 					 * No time as passed since the last
5946 					 * reorder, mark it as reordering.
5947 					 */
5948 					lro = 1;
5949 				}
5950 			} else {
5951 				/* Negative time? */
5952 				lro = 0;
5953 			}
5954 			if (lro > rack->r_ctl.rc_reorder_fade) {
5955 				/* Turn off reordering seen too */
5956 				rack->r_ctl.rc_reorder_ts = 0;
5957 				lro = 0;
5958 			}
5959 		} else {
5960 			/* Reodering does not fade */
5961 			lro = 1;
5962 		}
5963 	} else {
5964 		lro = 0;
5965 	}
5966 	if (rack->rc_rack_tmr_std_based == 0) {
5967 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
5968 	} else {
5969 		/* Standards based pkt-delay is 1/4 srtt */
5970 		thresh = srtt +  (srtt >> 2);
5971 	}
5972 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5973 		/* It must be set, if not you get 1/4 rtt */
5974 		if (rack->r_ctl.rc_reorder_shift)
5975 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5976 		else
5977 			thresh += (srtt >> 2);
5978 	}
5979 	if (rack->rc_rack_use_dsack &&
5980 	    lro &&
5981 	    (rack->r_ctl.num_dsack > 0)) {
5982 		/*
5983 		 * We only increase the reordering window if we
5984 		 * have seen reordering <and> we have a DSACK count.
5985 		 */
5986 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5987 		if (log_allowed)
5988 			rack_log_dsack_event(rack, 4, line, srtt, thresh);
5989 	}
5990 	/* SRTT * 2 is the ceiling */
5991 	if (thresh > (srtt * 2)) {
5992 		thresh = srtt * 2;
5993 	}
5994 	/* And we don't want it above the RTO max either */
5995 	if (thresh > rack_rto_max) {
5996 		thresh = rack_rto_max;
5997 	}
5998 	if (log_allowed)
5999 		rack_log_dsack_event(rack, 6, line,  srtt, thresh);
6000 	return (thresh);
6001 }
6002 
6003 static uint32_t
6004 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
6005 		     struct rack_sendmap *rsm, uint32_t srtt)
6006 {
6007 	struct rack_sendmap *prsm;
6008 	uint32_t thresh, len;
6009 	int segsiz;
6010 
6011 	if (srtt == 0)
6012 		srtt = 1;
6013 	if (rack->r_ctl.rc_tlp_threshold)
6014 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
6015 	else
6016 		thresh = (srtt * 2);
6017 
6018 	/* Get the previous sent packet, if any */
6019 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6020 	len = rsm->r_end - rsm->r_start;
6021 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
6022 		/* Exactly like the ID */
6023 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6024 			uint32_t alt_thresh;
6025 			/*
6026 			 * Compensate for delayed-ack with the d-ack time.
6027 			 */
6028 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6029 			if (alt_thresh > thresh)
6030 				thresh = alt_thresh;
6031 		}
6032 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6033 		/* 2.1 behavior */
6034 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6035 		if (prsm && (len <= segsiz)) {
6036 			/*
6037 			 * Two packets outstanding, thresh should be (2*srtt) +
6038 			 * possible inter-packet delay (if any).
6039 			 */
6040 			uint32_t inter_gap = 0;
6041 			int idx, nidx;
6042 
6043 			idx = rsm->r_rtr_cnt - 1;
6044 			nidx = prsm->r_rtr_cnt - 1;
6045 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6046 				/* Yes it was sent later (or at the same time) */
6047 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6048 			}
6049 			thresh += inter_gap;
6050 		} else if (len <= segsiz) {
6051 			/*
6052 			 * Possibly compensate for delayed-ack.
6053 			 */
6054 			uint32_t alt_thresh;
6055 
6056 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6057 			if (alt_thresh > thresh)
6058 				thresh = alt_thresh;
6059 		}
6060 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6061 		/* 2.2 behavior */
6062 		if (len <= segsiz) {
6063 			uint32_t alt_thresh;
6064 			/*
6065 			 * Compensate for delayed-ack with the d-ack time.
6066 			 */
6067 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6068 			if (alt_thresh > thresh)
6069 				thresh = alt_thresh;
6070 		}
6071 	}
6072 	/* Not above an RTO */
6073 	if (thresh > tp->t_rxtcur) {
6074 		thresh = tp->t_rxtcur;
6075 	}
6076 	/* Not above a RTO max */
6077 	if (thresh > rack_rto_max) {
6078 		thresh = rack_rto_max;
6079 	}
6080 	/* Apply user supplied min TLP */
6081 	if (thresh < rack_tlp_min) {
6082 		thresh = rack_tlp_min;
6083 	}
6084 	return (thresh);
6085 }
6086 
6087 static uint32_t
6088 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6089 {
6090 	/*
6091 	 * We want the rack_rtt which is the
6092 	 * last rtt we measured. However if that
6093 	 * does not exist we fallback to the srtt (which
6094 	 * we probably will never do) and then as a last
6095 	 * resort we use RACK_INITIAL_RTO if no srtt is
6096 	 * yet set.
6097 	 */
6098 	if (rack->rc_rack_rtt)
6099 		return (rack->rc_rack_rtt);
6100 	else if (tp->t_srtt == 0)
6101 		return (RACK_INITIAL_RTO);
6102 	return (tp->t_srtt);
6103 }
6104 
6105 static struct rack_sendmap *
6106 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6107 {
6108 	/*
6109 	 * Check to see that we don't need to fall into recovery. We will
6110 	 * need to do so if our oldest transmit is past the time we should
6111 	 * have had an ack.
6112 	 */
6113 	struct tcp_rack *rack;
6114 	struct rack_sendmap *rsm;
6115 	int32_t idx;
6116 	uint32_t srtt, thresh;
6117 
6118 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6119 	if (tqhash_empty(rack->r_ctl.tqh)) {
6120 		return (NULL);
6121 	}
6122 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6123 	if (rsm == NULL)
6124 		return (NULL);
6125 
6126 
6127 	if (rsm->r_flags & RACK_ACKED) {
6128 		rsm = rack_find_lowest_rsm(rack);
6129 		if (rsm == NULL)
6130 			return (NULL);
6131 	}
6132 	idx = rsm->r_rtr_cnt - 1;
6133 	srtt = rack_grab_rtt(tp, rack);
6134 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
6135 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6136 		return (NULL);
6137 	}
6138 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6139 		return (NULL);
6140 	}
6141 	/* Ok if we reach here we are over-due and this guy can be sent */
6142 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6143 	return (rsm);
6144 }
6145 
6146 static uint32_t
6147 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6148 {
6149 	int32_t t;
6150 	int32_t tt;
6151 	uint32_t ret_val;
6152 
6153 	t = (tp->t_srtt + (tp->t_rttvar << 2));
6154 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6155  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6156 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6157 	ret_val = (uint32_t)tt;
6158 	return (ret_val);
6159 }
6160 
6161 static uint32_t
6162 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6163 {
6164 	/*
6165 	 * Start the FR timer, we do this based on getting the first one in
6166 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6167 	 * events we need to stop the running timer (if its running) before
6168 	 * starting the new one.
6169 	 */
6170 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6171 	uint32_t srtt_cur;
6172 	int32_t idx;
6173 	int32_t is_tlp_timer = 0;
6174 	struct rack_sendmap *rsm;
6175 
6176 	if (rack->t_timers_stopped) {
6177 		/* All timers have been stopped none are to run */
6178 		return (0);
6179 	}
6180 	if (rack->rc_in_persist) {
6181 		/* We can't start any timer in persists */
6182 		return (rack_get_persists_timer_val(tp, rack));
6183 	}
6184 	rack->rc_on_min_to = 0;
6185 	if ((tp->t_state < TCPS_ESTABLISHED) ||
6186 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6187 		goto activate_rxt;
6188 	}
6189 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6190 	if ((rsm == NULL) || sup_rack) {
6191 		/* Nothing on the send map or no rack */
6192 activate_rxt:
6193 		time_since_sent = 0;
6194 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6195 		if (rsm) {
6196 			/*
6197 			 * Should we discount the RTX timer any?
6198 			 *
6199 			 * We want to discount it the smallest amount.
6200 			 * If a timer (Rack/TLP or RXT) has gone off more
6201 			 * recently thats the discount we want to use (now - timer time).
6202 			 * If the retransmit of the oldest packet was more recent then
6203 			 * we want to use that (now - oldest-packet-last_transmit_time).
6204 			 *
6205 			 */
6206 			idx = rsm->r_rtr_cnt - 1;
6207 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6208 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6209 			else
6210 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6211 			if (TSTMP_GT(cts, tstmp_touse))
6212 			    time_since_sent = cts - tstmp_touse;
6213 		}
6214 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6215 		    sbavail(&tptosocket(tp)->so_snd)) {
6216 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6217 			to = tp->t_rxtcur;
6218 			if (to > time_since_sent)
6219 				to -= time_since_sent;
6220 			else
6221 				to = rack->r_ctl.rc_min_to;
6222 			if (to == 0)
6223 				to = 1;
6224 			/* Special case for KEEPINIT */
6225 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6226 			    (TP_KEEPINIT(tp) != 0) &&
6227 			    rsm) {
6228 				/*
6229 				 * We have to put a ceiling on the rxt timer
6230 				 * of the keep-init timeout.
6231 				 */
6232 				uint32_t max_time, red;
6233 
6234 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6235 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6236 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6237 					if (red < max_time)
6238 						max_time -= red;
6239 					else
6240 						max_time = 1;
6241 				}
6242 				/* Reduce timeout to the keep value if needed */
6243 				if (max_time < to)
6244 					to = max_time;
6245 			}
6246 			return (to);
6247 		}
6248 		return (0);
6249 	}
6250 	if (rsm->r_flags & RACK_ACKED) {
6251 		rsm = rack_find_lowest_rsm(rack);
6252 		if (rsm == NULL) {
6253 			/* No lowest? */
6254 			goto activate_rxt;
6255 		}
6256 	}
6257 	/* Convert from ms to usecs */
6258 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
6259 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6260 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6261 		if ((tp->t_flags & TF_SENTFIN) &&
6262 		    ((tp->snd_max - tp->snd_una) == 1) &&
6263 		    (rsm->r_flags & RACK_HAS_FIN)) {
6264 			/*
6265 			 * We don't start a rack timer if all we have is a
6266 			 * FIN outstanding.
6267 			 */
6268 			goto activate_rxt;
6269 		}
6270 		if ((rack->use_rack_rr == 0) &&
6271 		    (IN_FASTRECOVERY(tp->t_flags)) &&
6272 		    (rack->rack_no_prr == 0) &&
6273 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
6274 			/*
6275 			 * We are not cheating, in recovery  and
6276 			 * not enough ack's to yet get our next
6277 			 * retransmission out.
6278 			 *
6279 			 * Note that classified attackers do not
6280 			 * get to use the rack-cheat.
6281 			 */
6282 			goto activate_tlp;
6283 		}
6284 		srtt = rack_grab_rtt(tp, rack);
6285 		thresh = rack_calc_thresh_rack(rack, srtt, cts, __LINE__, 1);
6286 		idx = rsm->r_rtr_cnt - 1;
6287 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6288 		if (SEQ_GEQ(exp, cts)) {
6289 			to = exp - cts;
6290 			if (to < rack->r_ctl.rc_min_to) {
6291 				to = rack->r_ctl.rc_min_to;
6292 				if (rack->r_rr_config == 3)
6293 					rack->rc_on_min_to = 1;
6294 			}
6295 		} else {
6296 			to = rack->r_ctl.rc_min_to;
6297 			if (rack->r_rr_config == 3)
6298 				rack->rc_on_min_to = 1;
6299 		}
6300 	} else {
6301 		/* Ok we need to do a TLP not RACK */
6302 activate_tlp:
6303 		if ((rack->rc_tlp_in_progress != 0) &&
6304 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6305 			/*
6306 			 * The previous send was a TLP and we have sent
6307 			 * N TLP's without sending new data.
6308 			 */
6309 			goto activate_rxt;
6310 		}
6311 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6312 		if (rsm == NULL) {
6313 			/* We found no rsm to TLP with. */
6314 			goto activate_rxt;
6315 		}
6316 		if (rsm->r_flags & RACK_HAS_FIN) {
6317 			/* If its a FIN we dont do TLP */
6318 			rsm = NULL;
6319 			goto activate_rxt;
6320 		}
6321 		idx = rsm->r_rtr_cnt - 1;
6322 		time_since_sent = 0;
6323 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6324 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6325 		else
6326 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6327 		if (TSTMP_GT(cts, tstmp_touse))
6328 		    time_since_sent = cts - tstmp_touse;
6329 		is_tlp_timer = 1;
6330 		if (tp->t_srtt) {
6331 			if ((rack->rc_srtt_measure_made == 0) &&
6332 			    (tp->t_srtt == 1)) {
6333 				/*
6334 				 * If another stack as run and set srtt to 1,
6335 				 * then the srtt was 0, so lets use the initial.
6336 				 */
6337 				srtt = RACK_INITIAL_RTO;
6338 			} else {
6339 				srtt_cur = tp->t_srtt;
6340 				srtt = srtt_cur;
6341 			}
6342 		} else
6343 			srtt = RACK_INITIAL_RTO;
6344 		/*
6345 		 * If the SRTT is not keeping up and the
6346 		 * rack RTT has spiked we want to use
6347 		 * the last RTT not the smoothed one.
6348 		 */
6349 		if (rack_tlp_use_greater &&
6350 		    tp->t_srtt &&
6351 		    (srtt < rack_grab_rtt(tp, rack))) {
6352 			srtt = rack_grab_rtt(tp, rack);
6353 		}
6354 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6355 		if (thresh > time_since_sent) {
6356 			to = thresh - time_since_sent;
6357 		} else {
6358 			to = rack->r_ctl.rc_min_to;
6359 			rack_log_alt_to_to_cancel(rack,
6360 						  thresh,		/* flex1 */
6361 						  time_since_sent,	/* flex2 */
6362 						  tstmp_touse,		/* flex3 */
6363 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6364 						  (uint32_t)rsm->r_tim_lastsent[idx],
6365 						  srtt,
6366 						  idx, 99);
6367 		}
6368 		if (to < rack_tlp_min) {
6369 			to = rack_tlp_min;
6370 		}
6371 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
6372 			/*
6373 			 * If the TLP time works out to larger than the max
6374 			 * RTO lets not do TLP.. just RTO.
6375 			 */
6376 			goto activate_rxt;
6377 		}
6378 	}
6379 	if (is_tlp_timer == 0) {
6380 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6381 	} else {
6382 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6383 	}
6384 	if (to == 0)
6385 		to = 1;
6386 	return (to);
6387 }
6388 
6389 static void
6390 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6391 {
6392 	if (rack->rc_in_persist == 0) {
6393 		if (tp->t_flags & TF_GPUTINPROG) {
6394 			/*
6395 			 * Stop the goodput now, the calling of the
6396 			 * measurement function clears the flag.
6397 			 */
6398 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6399 						    RACK_QUALITY_PERSIST);
6400 		}
6401 #ifdef NETFLIX_SHARED_CWND
6402 		if (rack->r_ctl.rc_scw) {
6403 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6404 			rack->rack_scwnd_is_idle = 1;
6405 		}
6406 #endif
6407 		rack->r_ctl.rc_went_idle_time = cts;
6408 		if (rack->r_ctl.rc_went_idle_time == 0)
6409 			rack->r_ctl.rc_went_idle_time = 1;
6410 		if (rack->lt_bw_up) {
6411 			/* Suspend our LT BW measurement */
6412 			uint64_t tmark;
6413 
6414 			rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6415 			rack->r_ctl.lt_seq = snd_una;
6416 			tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
6417 			if (tmark >= rack->r_ctl.lt_timemark) {
6418 				rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6419 			}
6420 			rack->r_ctl.lt_timemark = tmark;
6421 			rack->lt_bw_up = 0;
6422 			rack->r_persist_lt_bw_off = 1;
6423 		}
6424 		rack_timer_cancel(tp, rack, cts, __LINE__);
6425 		rack->r_ctl.persist_lost_ends = 0;
6426 		rack->probe_not_answered = 0;
6427 		rack->forced_ack = 0;
6428 		tp->t_rxtshift = 0;
6429 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6430 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6431 		rack->rc_in_persist = 1;
6432 	}
6433 }
6434 
6435 static void
6436 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6437 {
6438 	if (tcp_in_hpts(rack->rc_tp)) {
6439 		tcp_hpts_remove(rack->rc_tp);
6440 		rack->r_ctl.rc_hpts_flags = 0;
6441 	}
6442 #ifdef NETFLIX_SHARED_CWND
6443 	if (rack->r_ctl.rc_scw) {
6444 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6445 		rack->rack_scwnd_is_idle = 0;
6446 	}
6447 #endif
6448 	if (rack->rc_gp_dyn_mul &&
6449 	    (rack->use_fixed_rate == 0) &&
6450 	    (rack->rc_always_pace)) {
6451 		/*
6452 		 * Do we count this as if a probe-rtt just
6453 		 * finished?
6454 		 */
6455 		uint32_t time_idle, idle_min;
6456 
6457 		time_idle = cts - rack->r_ctl.rc_went_idle_time;
6458 		idle_min = rack_min_probertt_hold;
6459 		if (rack_probertt_gpsrtt_cnt_div) {
6460 			uint64_t extra;
6461 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
6462 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
6463 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
6464 			idle_min += (uint32_t)extra;
6465 		}
6466 		if (time_idle >= idle_min) {
6467 			/* Yes, we count it as a probe-rtt. */
6468 			uint32_t us_cts;
6469 
6470 			us_cts = tcp_get_usecs(NULL);
6471 			if (rack->in_probe_rtt == 0) {
6472 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6473 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
6474 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
6475 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
6476 			} else {
6477 				rack_exit_probertt(rack, us_cts);
6478 			}
6479 		}
6480 	}
6481 	if (rack->r_persist_lt_bw_off) {
6482 		/* Continue where we left off */
6483 		rack->r_ctl.lt_timemark = tcp_get_u64_usecs(NULL);
6484 		rack->lt_bw_up = 1;
6485 		rack->r_persist_lt_bw_off = 0;
6486 	}
6487 	rack->rc_in_persist = 0;
6488 	rack->r_ctl.rc_went_idle_time = 0;
6489 	tp->t_rxtshift = 0;
6490 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6491 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6492 	rack->r_ctl.rc_agg_delayed = 0;
6493 	rack->r_early = 0;
6494 	rack->r_late = 0;
6495 	rack->r_ctl.rc_agg_early = 0;
6496 }
6497 
6498 static void
6499 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
6500 		   struct hpts_diag *diag, struct timeval *tv)
6501 {
6502 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6503 		union tcp_log_stackspecific log;
6504 
6505 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6506 		log.u_bbr.flex1 = diag->p_nxt_slot;
6507 		log.u_bbr.flex2 = diag->p_cur_slot;
6508 		log.u_bbr.flex3 = diag->slot_req;
6509 		log.u_bbr.flex4 = diag->inp_hptsslot;
6510 		log.u_bbr.flex5 = diag->slot_remaining;
6511 		log.u_bbr.flex6 = diag->need_new_to;
6512 		log.u_bbr.flex7 = diag->p_hpts_active;
6513 		log.u_bbr.flex8 = diag->p_on_min_sleep;
6514 		/* Hijack other fields as needed */
6515 		log.u_bbr.epoch = diag->have_slept;
6516 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
6517 		log.u_bbr.pkts_out = diag->co_ret;
6518 		log.u_bbr.applimited = diag->hpts_sleep_time;
6519 		log.u_bbr.delivered = diag->p_prev_slot;
6520 		log.u_bbr.inflight = diag->p_runningslot;
6521 		log.u_bbr.bw_inuse = diag->wheel_slot;
6522 		log.u_bbr.rttProp = diag->wheel_cts;
6523 		log.u_bbr.timeStamp = cts;
6524 		log.u_bbr.delRate = diag->maxslots;
6525 		log.u_bbr.cur_del_rate = diag->p_curtick;
6526 		log.u_bbr.cur_del_rate <<= 32;
6527 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
6528 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6529 		    &rack->rc_inp->inp_socket->so_rcv,
6530 		    &rack->rc_inp->inp_socket->so_snd,
6531 		    BBR_LOG_HPTSDIAG, 0,
6532 		    0, &log, false, tv);
6533 	}
6534 
6535 }
6536 
6537 static void
6538 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
6539 {
6540 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6541 		union tcp_log_stackspecific log;
6542 		struct timeval tv;
6543 
6544 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6545 		log.u_bbr.flex1 = sb->sb_flags;
6546 		log.u_bbr.flex2 = len;
6547 		log.u_bbr.flex3 = sb->sb_state;
6548 		log.u_bbr.flex8 = type;
6549 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6550 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6551 		    &rack->rc_inp->inp_socket->so_rcv,
6552 		    &rack->rc_inp->inp_socket->so_snd,
6553 		    TCP_LOG_SB_WAKE, 0,
6554 		    len, &log, false, &tv);
6555 	}
6556 }
6557 
6558 static void
6559 rack_start_hpts_timer (struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
6560       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
6561 {
6562 	struct hpts_diag diag;
6563 	struct inpcb *inp = tptoinpcb(tp);
6564 	struct timeval tv;
6565 	uint32_t delayed_ack = 0;
6566 	uint32_t hpts_timeout;
6567 	uint32_t entry_slot = slot;
6568 	uint8_t stopped;
6569 	uint32_t left = 0;
6570 	uint32_t us_cts;
6571 
6572 	if ((tp->t_state == TCPS_CLOSED) ||
6573 	    (tp->t_state == TCPS_LISTEN)) {
6574 		return;
6575 	}
6576 	if (tcp_in_hpts(tp)) {
6577 		/* Already on the pacer */
6578 		return;
6579 	}
6580 	stopped = rack->rc_tmr_stopped;
6581 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
6582 		left = rack->r_ctl.rc_timer_exp - cts;
6583 	}
6584 	rack->r_ctl.rc_timer_exp = 0;
6585 	rack->r_ctl.rc_hpts_flags = 0;
6586 	us_cts = tcp_get_usecs(&tv);
6587 	/* Now early/late accounting */
6588 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
6589 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
6590 		/*
6591 		 * We have a early carry over set,
6592 		 * we can always add more time so we
6593 		 * can always make this compensation.
6594 		 *
6595 		 * Note if ack's are allowed to wake us do not
6596 		 * penalize the next timer for being awoke
6597 		 * by an ack aka the rc_agg_early (non-paced mode).
6598 		 */
6599 		slot += rack->r_ctl.rc_agg_early;
6600 		rack->r_early = 0;
6601 		rack->r_ctl.rc_agg_early = 0;
6602 	}
6603 	if ((rack->r_late) &&
6604 	    ((rack->r_use_hpts_min == 0) || (rack->dgp_on == 0))) {
6605 		/*
6606 		 * This is harder, we can
6607 		 * compensate some but it
6608 		 * really depends on what
6609 		 * the current pacing time is.
6610 		 */
6611 		if (rack->r_ctl.rc_agg_delayed >= slot) {
6612 			/*
6613 			 * We can't compensate for it all.
6614 			 * And we have to have some time
6615 			 * on the clock. We always have a min
6616 			 * 10 slots (10 x 10 i.e. 100 usecs).
6617 			 */
6618 			if (slot <= HPTS_TICKS_PER_SLOT) {
6619 				/* We gain delay */
6620 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
6621 				slot = HPTS_TICKS_PER_SLOT;
6622 			} else {
6623 				/* We take off some */
6624 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
6625 				slot = HPTS_TICKS_PER_SLOT;
6626 			}
6627 		} else {
6628 			slot -= rack->r_ctl.rc_agg_delayed;
6629 			rack->r_ctl.rc_agg_delayed = 0;
6630 			/* Make sure we have 100 useconds at minimum */
6631 			if (slot < HPTS_TICKS_PER_SLOT) {
6632 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
6633 				slot = HPTS_TICKS_PER_SLOT;
6634 			}
6635 			if (rack->r_ctl.rc_agg_delayed == 0)
6636 				rack->r_late = 0;
6637 		}
6638 	} else if (rack->r_late) {
6639 		/* r_use_hpts_min is on and so is DGP */
6640 		uint32_t max_red;
6641 
6642 		max_red = (slot * rack->r_ctl.max_reduction) / 100;
6643 		if (max_red >= rack->r_ctl.rc_agg_delayed) {
6644 			slot -= rack->r_ctl.rc_agg_delayed;
6645 			rack->r_ctl.rc_agg_delayed = 0;
6646 		} else {
6647 			slot -= max_red;
6648 			rack->r_ctl.rc_agg_delayed -= max_red;
6649 		}
6650 	}
6651 	if ((rack->r_use_hpts_min == 1) &&
6652 	    (slot > 0) &&
6653 	    (rack->dgp_on == 1)) {
6654 		/*
6655 		 * We are enforcing a min pacing timer
6656 		 * based on our hpts min timeout.
6657 		 */
6658 		uint32_t min;
6659 
6660 		min = get_hpts_min_sleep_time();
6661 		if (min > slot) {
6662 			slot = min;
6663 		}
6664 	}
6665 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
6666 	if (tp->t_flags & TF_DELACK) {
6667 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
6668 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
6669 	}
6670 	if (delayed_ack && ((hpts_timeout == 0) ||
6671 			    (delayed_ack < hpts_timeout)))
6672 		hpts_timeout = delayed_ack;
6673 	else
6674 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6675 	/*
6676 	 * If no timers are going to run and we will fall off the hptsi
6677 	 * wheel, we resort to a keep-alive timer if its configured.
6678 	 */
6679 	if ((hpts_timeout == 0) &&
6680 	    (slot == 0)) {
6681 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6682 		    (tp->t_state <= TCPS_CLOSING)) {
6683 			/*
6684 			 * Ok we have no timer (persists, rack, tlp, rxt  or
6685 			 * del-ack), we don't have segments being paced. So
6686 			 * all that is left is the keepalive timer.
6687 			 */
6688 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6689 				/* Get the established keep-alive time */
6690 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
6691 			} else {
6692 				/*
6693 				 * Get the initial setup keep-alive time,
6694 				 * note that this is probably not going to
6695 				 * happen, since rack will be running a rxt timer
6696 				 * if a SYN of some sort is outstanding. It is
6697 				 * actually handled in rack_timeout_rxt().
6698 				 */
6699 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
6700 			}
6701 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
6702 			if (rack->in_probe_rtt) {
6703 				/*
6704 				 * We want to instead not wake up a long time from
6705 				 * now but to wake up about the time we would
6706 				 * exit probe-rtt and initiate a keep-alive ack.
6707 				 * This will get us out of probe-rtt and update
6708 				 * our min-rtt.
6709 				 */
6710 				hpts_timeout = rack_min_probertt_hold;
6711 			}
6712 		}
6713 	}
6714 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
6715 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
6716 		/*
6717 		 * RACK, TLP, persists and RXT timers all are restartable
6718 		 * based on actions input .. i.e we received a packet (ack
6719 		 * or sack) and that changes things (rw, or snd_una etc).
6720 		 * Thus we can restart them with a new value. For
6721 		 * keep-alive, delayed_ack we keep track of what was left
6722 		 * and restart the timer with a smaller value.
6723 		 */
6724 		if (left < hpts_timeout)
6725 			hpts_timeout = left;
6726 	}
6727 	if (hpts_timeout) {
6728 		/*
6729 		 * Hack alert for now we can't time-out over 2,147,483
6730 		 * seconds (a bit more than 596 hours), which is probably ok
6731 		 * :).
6732 		 */
6733 		if (hpts_timeout > 0x7ffffffe)
6734 			hpts_timeout = 0x7ffffffe;
6735 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
6736 	}
6737 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
6738 	if ((rack->gp_ready == 0) &&
6739 	    (rack->use_fixed_rate == 0) &&
6740 	    (hpts_timeout < slot) &&
6741 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
6742 		/*
6743 		 * We have no good estimate yet for the
6744 		 * old clunky burst mitigation or the
6745 		 * real pacing. And the tlp or rxt is smaller
6746 		 * than the pacing calculation. Lets not
6747 		 * pace that long since we know the calculation
6748 		 * so far is not accurate.
6749 		 */
6750 		slot = hpts_timeout;
6751 	}
6752 	/**
6753 	 * Turn off all the flags for queuing by default. The
6754 	 * flags have important meanings to what happens when
6755 	 * LRO interacts with the transport. Most likely (by default now)
6756 	 * mbuf_queueing and ack compression are on. So the transport
6757 	 * has a couple of flags that control what happens (if those
6758 	 * are not on then these flags won't have any effect since it
6759 	 * won't go through the queuing LRO path).
6760 	 *
6761 	 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
6762 	 *                        pacing output, so don't disturb. But
6763 	 *                        it also means LRO can wake me if there
6764 	 *                        is a SACK arrival.
6765 	 *
6766 	 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
6767 	 *                       with the above flag (QUEUE_READY) and
6768 	 *                       when present it says don't even wake me
6769 	 *                       if a SACK arrives.
6770 	 *
6771 	 * The idea behind these flags is that if we are pacing we
6772 	 * set the MBUF_QUEUE_READY and only get woken up if
6773 	 * a SACK arrives (which could change things) or if
6774 	 * our pacing timer expires. If, however, we have a rack
6775 	 * timer running, then we don't even want a sack to wake
6776 	 * us since the rack timer has to expire before we can send.
6777 	 *
6778 	 * Other cases should usually have none of the flags set
6779 	 * so LRO can call into us.
6780 	 */
6781 	tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
6782 	if (slot) {
6783 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
6784 		rack->r_ctl.rc_last_output_to = us_cts + slot;
6785 		/*
6786 		 * A pacing timer (slot) is being set, in
6787 		 * such a case we cannot send (we are blocked by
6788 		 * the timer). So lets tell LRO that it should not
6789 		 * wake us unless there is a SACK. Note this only
6790 		 * will be effective if mbuf queueing is on or
6791 		 * compressed acks are being processed.
6792 		 */
6793 		tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
6794 		/*
6795 		 * But wait if we have a Rack timer running
6796 		 * even a SACK should not disturb us (with
6797 		 * the exception of r_rr_config 3).
6798 		 */
6799 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
6800 		    (IN_RECOVERY(tp->t_flags))) {
6801 			if (rack->r_rr_config != 3)
6802 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6803 			else if (rack->rc_pace_dnd) {
6804 				/*
6805 				 * When DND is on, we only let a sack
6806 				 * interrupt us if we are not in recovery.
6807 				 *
6808 				 * If DND is off, then we never hit here
6809 				 * and let all sacks wake us up.
6810 				 *
6811 				 */
6812 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6813 			}
6814 		}
6815 		if (rack->rc_ack_can_sendout_data) {
6816 			/*
6817 			 * Ahh but wait, this is that special case
6818 			 * where the pacing timer can be disturbed
6819 			 * backout the changes (used for non-paced
6820 			 * burst limiting).
6821 			 */
6822 			tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
6823 			    TF2_MBUF_QUEUE_READY);
6824 		}
6825 		if ((rack->use_rack_rr) &&
6826 		    (rack->r_rr_config < 2) &&
6827 		    ((hpts_timeout) && (hpts_timeout < slot))) {
6828 			/*
6829 			 * Arrange for the hpts to kick back in after the
6830 			 * t-o if the t-o does not cause a send.
6831 			 */
6832 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6833 						   __LINE__, &diag);
6834 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6835 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6836 		} else {
6837 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
6838 						   __LINE__, &diag);
6839 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6840 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
6841 		}
6842 	} else if (hpts_timeout) {
6843 		/*
6844 		 * With respect to t_flags2(?) here, lets let any new acks wake
6845 		 * us up here. Since we are not pacing (no pacing timer), output
6846 		 * can happen so we should let it. If its a Rack timer, then any inbound
6847 		 * packet probably won't change the sending (we will be blocked)
6848 		 * but it may change the prr stats so letting it in (the set defaults
6849 		 * at the start of this block) are good enough.
6850 		 */
6851 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6852 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6853 					   __LINE__, &diag);
6854 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6855 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6856 	} else {
6857 		/* No timer starting */
6858 #ifdef INVARIANTS
6859 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
6860 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
6861 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
6862 		}
6863 #endif
6864 	}
6865 	rack->rc_tmr_stopped = 0;
6866 	if (slot)
6867 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6868 }
6869 
6870 static void
6871 rack_mark_lost(struct tcpcb *tp,
6872     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
6873 {
6874 	struct rack_sendmap *nrsm;
6875 	uint32_t thresh,  exp;
6876 
6877 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
6878 	nrsm = rsm;
6879 	TAILQ_FOREACH_FROM(nrsm, &rack->r_ctl.rc_tmap, r_tnext) {
6880 		if ((nrsm->r_flags & RACK_SACK_PASSED) == 0) {
6881 			/* Got up to all that were marked sack-passed */
6882 			break;
6883 		}
6884 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
6885 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
6886 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
6887 				/* We now consider it lost */
6888 				nrsm->r_flags |= RACK_WAS_LOST;
6889 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
6890 			} else {
6891 				/* Past here it won't be lost so stop */
6892 				break;
6893 			}
6894 		}
6895 	}
6896 }
6897 
6898 /*
6899  * RACK Timer, here we simply do logging and house keeping.
6900  * the normal rack_output() function will call the
6901  * appropriate thing to check if we need to do a RACK retransmit.
6902  * We return 1, saying don't proceed with rack_output only
6903  * when all timers have been stopped (destroyed PCB?).
6904  */
6905 static int
6906 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6907 {
6908 	/*
6909 	 * This timer simply provides an internal trigger to send out data.
6910 	 * The check_recovery_mode call will see if there are needed
6911 	 * retransmissions, if so we will enter fast-recovery. The output
6912 	 * call may or may not do the same thing depending on sysctl
6913 	 * settings.
6914 	 */
6915 	struct rack_sendmap *rsm;
6916 
6917 	counter_u64_add(rack_to_tot, 1);
6918 	if (rack->r_state && (rack->r_state != tp->t_state))
6919 		rack_set_state(tp, rack);
6920 	rack->rc_on_min_to = 0;
6921 	rsm = rack_check_recovery_mode(tp, cts);
6922 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
6923 	if (rsm) {
6924 		/* We need to stroke any lost that are now declared as lost */
6925 		rack_mark_lost(tp, rack, rsm, cts);
6926 		rack->r_ctl.rc_resend = rsm;
6927 		rack->r_timer_override = 1;
6928 		if (rack->use_rack_rr) {
6929 			/*
6930 			 * Don't accumulate extra pacing delay
6931 			 * we are allowing the rack timer to
6932 			 * over-ride pacing i.e. rrr takes precedence
6933 			 * if the pacing interval is longer than the rrr
6934 			 * time (in other words we get the min pacing
6935 			 * time versus rrr pacing time).
6936 			 */
6937 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6938 		}
6939 	}
6940 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6941 	if (rsm == NULL) {
6942 		/* restart a timer and return 1 */
6943 		rack_start_hpts_timer(rack, tp, cts,
6944 				      0, 0, 0);
6945 		return (1);
6946 	}
6947 	return (0);
6948 }
6949 
6950 
6951 
6952 static void
6953 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6954 {
6955 
6956 	if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
6957 		/*
6958 		 * The trailing space changed, mbufs can grow
6959 		 * at the tail but they can't shrink from
6960 		 * it, KASSERT that. Adjust the orig_m_len to
6961 		 * compensate for this change.
6962 		 */
6963 		KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
6964 			("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
6965 			 rsm->m,
6966 			 rsm,
6967 			 (intmax_t)M_TRAILINGROOM(rsm->m),
6968 			 rsm->orig_t_space,
6969 			 rsm->orig_m_len,
6970 			 rsm->m->m_len));
6971 		rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
6972 		rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
6973 	}
6974 	if (rsm->m->m_len < rsm->orig_m_len) {
6975 		/*
6976 		 * Mbuf shrank, trimmed off the top by an ack, our
6977 		 * offset changes.
6978 		 */
6979 		KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
6980 			("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
6981 			 rsm->m, rsm->m->m_len,
6982 			 rsm, rsm->orig_m_len,
6983 			 rsm->soff));
6984 		if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
6985 			rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6986 		else
6987 			rsm->soff = 0;
6988 		rsm->orig_m_len = rsm->m->m_len;
6989 #ifdef INVARIANTS
6990 	} else if (rsm->m->m_len > rsm->orig_m_len) {
6991 		panic("rsm:%p m:%p m_len grew outside of t_space compensation",
6992 		      rsm, rsm->m);
6993 #endif
6994 	}
6995 }
6996 
6997 static void
6998 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6999 {
7000 	struct mbuf *m;
7001 	uint32_t soff;
7002 
7003 	if (src_rsm->m &&
7004 	    ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
7005 	     (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
7006 		/* Fix up the orig_m_len and possibly the mbuf offset */
7007 		rack_adjust_orig_mlen(src_rsm);
7008 	}
7009 	m = src_rsm->m;
7010 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
7011 	while (soff >= m->m_len) {
7012 		/* Move out past this mbuf */
7013 		soff -= m->m_len;
7014 		m = m->m_next;
7015 		KASSERT((m != NULL),
7016 			("rsm:%p nrsm:%p hit at soff:%u null m",
7017 			 src_rsm, rsm, soff));
7018 		if (m == NULL) {
7019 			/* This should *not* happen which is why there is a kassert */
7020 			src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7021 					       (src_rsm->r_start - rack->rc_tp->snd_una),
7022 					       &src_rsm->soff);
7023 			src_rsm->orig_m_len = src_rsm->m->m_len;
7024 			src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7025 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7026 					   (rsm->r_start - rack->rc_tp->snd_una),
7027 					   &rsm->soff);
7028 			rsm->orig_m_len = rsm->m->m_len;
7029 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7030 			return;
7031 		}
7032 	}
7033 	rsm->m = m;
7034 	rsm->soff = soff;
7035 	rsm->orig_m_len = m->m_len;
7036 	rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7037 }
7038 
7039 static __inline void
7040 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7041 	       struct rack_sendmap *rsm, uint32_t start)
7042 {
7043 	int idx;
7044 
7045 	nrsm->r_start = start;
7046 	nrsm->r_end = rsm->r_end;
7047 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7048 	nrsm->r_act_rxt_cnt = rsm->r_act_rxt_cnt;
7049 	nrsm->r_flags = rsm->r_flags;
7050 	nrsm->r_dupack = rsm->r_dupack;
7051 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7052 	nrsm->r_rtr_bytes = 0;
7053 	nrsm->r_fas = rsm->r_fas;
7054 	nrsm->r_bas = rsm->r_bas;
7055 	tqhash_update_end(rack->r_ctl.tqh, rsm, nrsm->r_start);
7056 	nrsm->r_just_ret = rsm->r_just_ret;
7057 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7058 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7059 	}
7060 	/* Now if we have SYN flag we keep it on the left edge */
7061 	if (nrsm->r_flags & RACK_HAS_SYN)
7062 		nrsm->r_flags &= ~RACK_HAS_SYN;
7063 	/* Now if we have a FIN flag we keep it on the right edge */
7064 	if (rsm->r_flags & RACK_HAS_FIN)
7065 		rsm->r_flags &= ~RACK_HAS_FIN;
7066 	/* Push bit must go to the right edge as well */
7067 	if (rsm->r_flags & RACK_HAD_PUSH)
7068 		rsm->r_flags &= ~RACK_HAD_PUSH;
7069 	/* Clone over the state of the hw_tls flag */
7070 	nrsm->r_hw_tls = rsm->r_hw_tls;
7071 	/*
7072 	 * Now we need to find nrsm's new location in the mbuf chain
7073 	 * we basically calculate a new offset, which is soff +
7074 	 * how much is left in original rsm. Then we walk out the mbuf
7075 	 * chain to find the righ position, it may be the same mbuf
7076 	 * or maybe not.
7077 	 */
7078 	KASSERT(((rsm->m != NULL) ||
7079 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7080 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7081 	if (rsm->m)
7082 		rack_setup_offset_for_rsm(rack, rsm, nrsm);
7083 }
7084 
7085 static struct rack_sendmap *
7086 rack_merge_rsm(struct tcp_rack *rack,
7087 	       struct rack_sendmap *l_rsm,
7088 	       struct rack_sendmap *r_rsm)
7089 {
7090 	/*
7091 	 * We are merging two ack'd RSM's,
7092 	 * the l_rsm is on the left (lower seq
7093 	 * values) and the r_rsm is on the right
7094 	 * (higher seq value). The simplest way
7095 	 * to merge these is to move the right
7096 	 * one into the left. I don't think there
7097 	 * is any reason we need to try to find
7098 	 * the oldest (or last oldest retransmitted).
7099 	 */
7100 	rack_log_map_chg(rack->rc_tp, rack, NULL,
7101 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7102 	tqhash_update_end(rack->r_ctl.tqh, l_rsm, r_rsm->r_end);
7103 	if (l_rsm->r_dupack < r_rsm->r_dupack)
7104 		l_rsm->r_dupack = r_rsm->r_dupack;
7105 	if (r_rsm->r_rtr_bytes)
7106 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7107 	if (r_rsm->r_in_tmap) {
7108 		/* This really should not happen */
7109 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7110 		r_rsm->r_in_tmap = 0;
7111 	}
7112 
7113 	/* Now the flags */
7114 	if (r_rsm->r_flags & RACK_HAS_FIN)
7115 		l_rsm->r_flags |= RACK_HAS_FIN;
7116 	if (r_rsm->r_flags & RACK_TLP)
7117 		l_rsm->r_flags |= RACK_TLP;
7118 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7119 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7120 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
7121 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7122 		/*
7123 		 * If both are app-limited then let the
7124 		 * free lower the count. If right is app
7125 		 * limited and left is not, transfer.
7126 		 */
7127 		l_rsm->r_flags |= RACK_APP_LIMITED;
7128 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
7129 		if (r_rsm == rack->r_ctl.rc_first_appl)
7130 			rack->r_ctl.rc_first_appl = l_rsm;
7131 	}
7132 	tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7133 	/*
7134 	 * We keep the largest value, which is the newest
7135 	 * send. We do this in case a segment that is
7136 	 * joined together and not part of a GP estimate
7137 	 * later gets expanded into the GP estimate.
7138 	 *
7139 	 * We prohibit the merging of unlike kinds i.e.
7140 	 * all pieces that are in the GP estimate can be
7141 	 * merged and all pieces that are not in a GP estimate
7142 	 * can be merged, but not disimilar pieces. Combine
7143 	 * this with taking the highest here and we should
7144 	 * be ok unless of course the client reneges. Then
7145 	 * all bets are off.
7146 	 */
7147 	if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7148 	   r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7149 		l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7150 	}
7151 	/*
7152 	 * When merging two RSM's we also need to consider the ack time and keep
7153 	 * newest. If the ack gets merged into a measurement then that is the
7154 	 * one we will want to be using.
7155 	 */
7156 	if(l_rsm->r_ack_arrival	 < r_rsm->r_ack_arrival)
7157 		l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7158 
7159 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7160 		/* Transfer the split limit to the map we free */
7161 		r_rsm->r_limit_type = l_rsm->r_limit_type;
7162 		l_rsm->r_limit_type = 0;
7163 	}
7164 	rack_free(rack, r_rsm);
7165 	l_rsm->r_flags |= RACK_MERGED;
7166 	return (l_rsm);
7167 }
7168 
7169 /*
7170  * TLP Timer, here we simply setup what segment we want to
7171  * have the TLP expire on, the normal rack_output() will then
7172  * send it out.
7173  *
7174  * We return 1, saying don't proceed with rack_output only
7175  * when all timers have been stopped (destroyed PCB?).
7176  */
7177 static int
7178 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7179 {
7180 	/*
7181 	 * Tail Loss Probe.
7182 	 */
7183 	struct rack_sendmap *rsm = NULL;
7184 	int insret __diagused;
7185 	struct socket *so = tptosocket(tp);
7186 	uint32_t amm;
7187 	uint32_t out, avail;
7188 	int collapsed_win = 0;
7189 
7190 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7191 		/* Its not time yet */
7192 		return (0);
7193 	}
7194 	if (ctf_progress_timeout_check(tp, true)) {
7195 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7196 		return (-ETIMEDOUT);	/* tcp_drop() */
7197 	}
7198 	/*
7199 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7200 	 * need to figure out how to force a full MSS segment out.
7201 	 */
7202 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7203 	rack->r_ctl.retran_during_recovery = 0;
7204 	rack->r_might_revert = 0;
7205 	rack->r_ctl.dsack_byte_cnt = 0;
7206 	counter_u64_add(rack_tlp_tot, 1);
7207 	if (rack->r_state && (rack->r_state != tp->t_state))
7208 		rack_set_state(tp, rack);
7209 	avail = sbavail(&so->so_snd);
7210 	out = tp->snd_max - tp->snd_una;
7211 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7212 		/* special case, we need a retransmission */
7213 		collapsed_win = 1;
7214 		goto need_retran;
7215 	}
7216 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7217 		rack->r_ctl.dsack_persist--;
7218 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7219 			rack->r_ctl.num_dsack = 0;
7220 		}
7221 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7222 	}
7223 	if ((tp->t_flags & TF_GPUTINPROG) &&
7224 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7225 		/*
7226 		 * If this is the second in a row
7227 		 * TLP and we are doing a measurement
7228 		 * its time to abandon the measurement.
7229 		 * Something is likely broken on
7230 		 * the clients network and measuring a
7231 		 * broken network does us no good.
7232 		 */
7233 		tp->t_flags &= ~TF_GPUTINPROG;
7234 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7235 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7236 					   tp->gput_seq,
7237 					   0, 0, 18, __LINE__, NULL, 0);
7238 	}
7239 	/*
7240 	 * Check our send oldest always settings, and if
7241 	 * there is an oldest to send jump to the need_retran.
7242 	 */
7243 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7244 		goto need_retran;
7245 
7246 	if (avail > out) {
7247 		/* New data is available */
7248 		amm = avail - out;
7249 		if (amm > ctf_fixed_maxseg(tp)) {
7250 			amm = ctf_fixed_maxseg(tp);
7251 			if ((amm + out) > tp->snd_wnd) {
7252 				/* We are rwnd limited */
7253 				goto need_retran;
7254 			}
7255 		} else if (amm < ctf_fixed_maxseg(tp)) {
7256 			/* not enough to fill a MTU */
7257 			goto need_retran;
7258 		}
7259 		if (IN_FASTRECOVERY(tp->t_flags)) {
7260 			/* Unlikely */
7261 			if (rack->rack_no_prr == 0) {
7262 				if (out + amm <= tp->snd_wnd) {
7263 					rack->r_ctl.rc_prr_sndcnt = amm;
7264 					rack->r_ctl.rc_tlp_new_data = amm;
7265 					rack_log_to_prr(rack, 4, 0, __LINE__);
7266 				}
7267 			} else
7268 				goto need_retran;
7269 		} else {
7270 			/* Set the send-new override */
7271 			if (out + amm <= tp->snd_wnd)
7272 				rack->r_ctl.rc_tlp_new_data = amm;
7273 			else
7274 				goto need_retran;
7275 		}
7276 		rack->r_ctl.rc_tlpsend = NULL;
7277 		counter_u64_add(rack_tlp_newdata, 1);
7278 		goto send;
7279 	}
7280 need_retran:
7281 	/*
7282 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
7283 	 * optionally the first un-acked segment.
7284 	 */
7285 	if (collapsed_win == 0) {
7286 		if (rack_always_send_oldest)
7287 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7288 		else {
7289 			rsm = tqhash_max(rack->r_ctl.tqh);
7290 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7291 				rsm = rack_find_high_nonack(rack, rsm);
7292 			}
7293 		}
7294 		if (rsm == NULL) {
7295 #ifdef TCP_BLACKBOX
7296 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7297 #endif
7298 			goto out;
7299 		}
7300 	} else {
7301 		/*
7302 		 * We had a collapsed window, lets find
7303 		 * the point before the collapse.
7304 		 */
7305 		if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7306 			rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7307 		else {
7308 			rsm = tqhash_min(rack->r_ctl.tqh);
7309 		}
7310 		if (rsm == NULL) {
7311 			/* Huh */
7312 			goto out;
7313 		}
7314 	}
7315 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7316 		/*
7317 		 * We need to split this the last segment in two.
7318 		 */
7319 		struct rack_sendmap *nrsm;
7320 
7321 		nrsm = rack_alloc_full_limit(rack);
7322 		if (nrsm == NULL) {
7323 			/*
7324 			 * No memory to split, we will just exit and punt
7325 			 * off to the RXT timer.
7326 			 */
7327 			goto out;
7328 		}
7329 		rack_clone_rsm(rack, nrsm, rsm,
7330 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
7331 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7332 #ifndef INVARIANTS
7333 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7334 #else
7335 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7336 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
7337 			      nrsm, insret, rack, rsm);
7338 		}
7339 #endif
7340 		if (rsm->r_in_tmap) {
7341 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7342 			nrsm->r_in_tmap = 1;
7343 		}
7344 		rsm = nrsm;
7345 	}
7346 	rack->r_ctl.rc_tlpsend = rsm;
7347 send:
7348 	/* Make sure output path knows we are doing a TLP */
7349 	*doing_tlp = 1;
7350 	rack->r_timer_override = 1;
7351 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7352 	return (0);
7353 out:
7354 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7355 	return (0);
7356 }
7357 
7358 /*
7359  * Delayed ack Timer, here we simply need to setup the
7360  * ACK_NOW flag and remove the DELACK flag. From there
7361  * the output routine will send the ack out.
7362  *
7363  * We only return 1, saying don't proceed, if all timers
7364  * are stopped (destroyed PCB?).
7365  */
7366 static int
7367 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7368 {
7369 
7370 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7371 	tp->t_flags &= ~TF_DELACK;
7372 	tp->t_flags |= TF_ACKNOW;
7373 	KMOD_TCPSTAT_INC(tcps_delack);
7374 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7375 	return (0);
7376 }
7377 
7378 static inline int
7379 rack_send_ack_challange(struct tcp_rack *rack)
7380 {
7381 	struct tcptemp *t_template;
7382 
7383 	t_template = tcpip_maketemplate(rack->rc_inp);
7384 	if (t_template) {
7385 		if (rack->forced_ack == 0) {
7386 			rack->forced_ack = 1;
7387 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7388 		} else {
7389 			rack->probe_not_answered = 1;
7390 		}
7391 		tcp_respond(rack->rc_tp, t_template->tt_ipgen,
7392 			    &t_template->tt_t, (struct mbuf *)NULL,
7393 			    rack->rc_tp->rcv_nxt, rack->rc_tp->snd_una - 1, 0);
7394 		free(t_template, M_TEMP);
7395 		/* This does send an ack so kill any D-ack timer */
7396 		if (rack->rc_tp->t_flags & TF_DELACK)
7397 			rack->rc_tp->t_flags &= ~TF_DELACK;
7398 		return(1);
7399 	} else
7400 		return (0);
7401 
7402 }
7403 
7404 /*
7405  * Persists timer, here we simply send the
7406  * same thing as a keepalive will.
7407  * the one byte send.
7408  *
7409  * We only return 1, saying don't proceed, if all timers
7410  * are stopped (destroyed PCB?).
7411  */
7412 static int
7413 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7414 {
7415 	int32_t retval = 1;
7416 
7417 	if (rack->rc_in_persist == 0)
7418 		return (0);
7419 	if (ctf_progress_timeout_check(tp, false)) {
7420 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7421 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7422 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7423 		return (-ETIMEDOUT);	/* tcp_drop() */
7424 	}
7425 	/*
7426 	 * Persistence timer into zero window. Force a byte to be output, if
7427 	 * possible.
7428 	 */
7429 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
7430 	/*
7431 	 * Hack: if the peer is dead/unreachable, we do not time out if the
7432 	 * window is closed.  After a full backoff, drop the connection if
7433 	 * the idle time (no responses to probes) reaches the maximum
7434 	 * backoff that we would use if retransmitting.
7435 	 */
7436 	if (tp->t_rxtshift >= V_tcp_retries &&
7437 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7438 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7439 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7440 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7441 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7442 		retval = -ETIMEDOUT;	/* tcp_drop() */
7443 		goto out;
7444 	}
7445 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7446 	    tp->snd_una == tp->snd_max)
7447 		rack_exit_persist(tp, rack, cts);
7448 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7449 	/*
7450 	 * If the user has closed the socket then drop a persisting
7451 	 * connection after a much reduced timeout.
7452 	 */
7453 	if (tp->t_state > TCPS_CLOSE_WAIT &&
7454 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7455 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7456 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7457 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7458 		retval = -ETIMEDOUT;	/* tcp_drop() */
7459 		goto out;
7460 	}
7461 	if (rack_send_ack_challange(rack)) {
7462 		/* only set it if we were answered */
7463 		if (rack->probe_not_answered) {
7464 			counter_u64_add(rack_persists_loss, 1);
7465 			rack->r_ctl.persist_lost_ends++;
7466 		}
7467 		counter_u64_add(rack_persists_sends, 1);
7468 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7469 	}
7470 	if (tp->t_rxtshift < V_tcp_retries)
7471 		tp->t_rxtshift++;
7472 out:
7473 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7474 	rack_start_hpts_timer(rack, tp, cts,
7475 			      0, 0, 0);
7476 	return (retval);
7477 }
7478 
7479 /*
7480  * If a keepalive goes off, we had no other timers
7481  * happening. We always return 1 here since this
7482  * routine either drops the connection or sends
7483  * out a segment with respond.
7484  */
7485 static int
7486 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7487 {
7488 	struct inpcb *inp = tptoinpcb(tp);
7489 
7490 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7491 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7492 	/*
7493 	 * Keep-alive timer went off; send something or drop connection if
7494 	 * idle for too long.
7495 	 */
7496 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
7497 	if (tp->t_state < TCPS_ESTABLISHED)
7498 		goto dropit;
7499 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7500 	    tp->t_state <= TCPS_CLOSING) {
7501 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7502 			goto dropit;
7503 		/*
7504 		 * Send a packet designed to force a response if the peer is
7505 		 * up and reachable: either an ACK if the connection is
7506 		 * still alive, or an RST if the peer has closed the
7507 		 * connection due to timeout or reboot. Using sequence
7508 		 * number tp->snd_una-1 causes the transmitted zero-length
7509 		 * segment to lie outside the receive window; by the
7510 		 * protocol spec, this requires the correspondent TCP to
7511 		 * respond.
7512 		 */
7513 		KMOD_TCPSTAT_INC(tcps_keepprobe);
7514 		rack_send_ack_challange(rack);
7515 	}
7516 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7517 	return (1);
7518 dropit:
7519 	KMOD_TCPSTAT_INC(tcps_keepdrops);
7520 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7521 	return (-ETIMEDOUT);	/* tcp_drop() */
7522 }
7523 
7524 /*
7525  * Retransmit helper function, clear up all the ack
7526  * flags and take care of important book keeping.
7527  */
7528 static void
7529 rack_remxt_tmr(struct tcpcb *tp)
7530 {
7531 	/*
7532 	 * The retransmit timer went off, all sack'd blocks must be
7533 	 * un-acked.
7534 	 */
7535 	struct rack_sendmap *rsm, *trsm = NULL;
7536 	struct tcp_rack *rack;
7537 
7538 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7539 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7540 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7541 	rack->r_timer_override = 1;
7542 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7543 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7544 	rack->r_late = 0;
7545 	rack->r_early = 0;
7546 	rack->r_ctl.rc_agg_delayed = 0;
7547 	rack->r_ctl.rc_agg_early = 0;
7548 	if (rack->r_state && (rack->r_state != tp->t_state))
7549 		rack_set_state(tp, rack);
7550 	if (tp->t_rxtshift <= rack_rxt_scoreboard_clear_thresh) {
7551 		/*
7552 		 * We do not clear the scoreboard until we have had
7553 		 * more than rack_rxt_scoreboard_clear_thresh time-outs.
7554 		 */
7555 		rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7556 		if (rack->r_ctl.rc_resend != NULL)
7557 			rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7558 
7559 		return;
7560 	}
7561 	/*
7562 	 * Ideally we would like to be able to
7563 	 * mark SACK-PASS on anything not acked here.
7564 	 *
7565 	 * However, if we do that we would burst out
7566 	 * all that data 1ms apart. This would be unwise,
7567 	 * so for now we will just let the normal rxt timer
7568 	 * and tlp timer take care of it.
7569 	 *
7570 	 * Also we really need to stick them back in sequence
7571 	 * order. This way we send in the proper order and any
7572 	 * sacks that come floating in will "re-ack" the data.
7573 	 * To do this we zap the tmap with an INIT and then
7574 	 * walk through and place every rsm in the tail queue
7575 	 * hash table back in its seq ordered place.
7576 	 */
7577 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
7578 
7579 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
7580 		rsm->r_dupack = 0;
7581 		if (rack_verbose_logging)
7582 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7583 		/* We must re-add it back to the tlist */
7584 		if (trsm == NULL) {
7585 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7586 		} else {
7587 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7588 		}
7589 		rsm->r_in_tmap = 1;
7590 		trsm = rsm;
7591 		if (rsm->r_flags & RACK_ACKED)
7592 			rsm->r_flags |= RACK_WAS_ACKED;
7593 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED | RACK_WAS_LOST);
7594 		rsm->r_flags |= RACK_MUST_RXT;
7595 	}
7596 	/* zero the lost since it's all gone */
7597 	rack->r_ctl.rc_considered_lost = 0;
7598 	/* Clear the count (we just un-acked them) */
7599 	rack->r_ctl.rc_sacked = 0;
7600 	rack->r_ctl.rc_sacklast = NULL;
7601 	/* Clear the tlp rtx mark */
7602 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7603 	if (rack->r_ctl.rc_resend != NULL)
7604 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7605 	rack->r_ctl.rc_prr_sndcnt = 0;
7606 	rack_log_to_prr(rack, 6, 0, __LINE__);
7607 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7608 	if (rack->r_ctl.rc_resend != NULL)
7609 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7610 	if (((tp->t_flags & TF_SACK_PERMIT) == 0) &&
7611 	    ((tp->t_flags & TF_SENTFIN) == 0)) {
7612 		/*
7613 		 * For non-sack customers new data
7614 		 * needs to go out as retransmits until
7615 		 * we retransmit up to snd_max.
7616 		 */
7617 		rack->r_must_retran = 1;
7618 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7619 							    rack->r_ctl.rc_sacked);
7620 	}
7621 }
7622 
7623 static void
7624 rack_convert_rtts(struct tcpcb *tp)
7625 {
7626 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7627 	tp->t_rxtcur = RACK_REXMTVAL(tp);
7628 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7629 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7630 	}
7631 	if (tp->t_rxtcur > rack_rto_max) {
7632 		tp->t_rxtcur = rack_rto_max;
7633 	}
7634 }
7635 
7636 static void
7637 rack_cc_conn_init(struct tcpcb *tp)
7638 {
7639 	struct tcp_rack *rack;
7640 	uint32_t srtt;
7641 
7642 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7643 	srtt = tp->t_srtt;
7644 	cc_conn_init(tp);
7645 	/*
7646 	 * Now convert to rack's internal format,
7647 	 * if required.
7648 	 */
7649 	if ((srtt == 0) && (tp->t_srtt != 0))
7650 		rack_convert_rtts(tp);
7651 	/*
7652 	 * We want a chance to stay in slowstart as
7653 	 * we create a connection. TCP spec says that
7654 	 * initially ssthresh is infinite. For our
7655 	 * purposes that is the snd_wnd.
7656 	 */
7657 	if (tp->snd_ssthresh < tp->snd_wnd) {
7658 		tp->snd_ssthresh = tp->snd_wnd;
7659 	}
7660 	/*
7661 	 * We also want to assure a IW worth of
7662 	 * data can get inflight.
7663 	 */
7664 	if (rc_init_window(rack) < tp->snd_cwnd)
7665 		tp->snd_cwnd = rc_init_window(rack);
7666 }
7667 
7668 /*
7669  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7670  * we will setup to retransmit the lowest seq number outstanding.
7671  */
7672 static int
7673 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7674 {
7675 	struct inpcb *inp = tptoinpcb(tp);
7676 	int32_t rexmt;
7677 	int32_t retval = 0;
7678 	bool isipv6;
7679 
7680 	if ((tp->t_flags & TF_GPUTINPROG) &&
7681 	    (tp->t_rxtshift)) {
7682 		/*
7683 		 * We have had a second timeout
7684 		 * measurements on successive rxt's are not profitable.
7685 		 * It is unlikely to be of any use (the network is
7686 		 * broken or the client went away).
7687 		 */
7688 		tp->t_flags &= ~TF_GPUTINPROG;
7689 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7690 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7691 					   tp->gput_seq,
7692 					   0, 0, 18, __LINE__, NULL, 0);
7693 	}
7694 	if (ctf_progress_timeout_check(tp, false)) {
7695 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7696 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7697 		return (-ETIMEDOUT);	/* tcp_drop() */
7698 	}
7699 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7700 	rack->r_ctl.retran_during_recovery = 0;
7701 	rack->rc_ack_required = 1;
7702 	rack->r_ctl.dsack_byte_cnt = 0;
7703 	if (IN_RECOVERY(tp->t_flags) &&
7704 	    (rack->rto_from_rec == 0)) {
7705 		/*
7706 		 * Mark that we had a rto while in recovery
7707 		 * and save the ssthresh so if we go back
7708 		 * into recovery we will have a chance
7709 		 * to slowstart back to the level.
7710 		 */
7711 		rack->rto_from_rec = 1;
7712 		rack->r_ctl.rto_ssthresh = tp->snd_ssthresh;
7713 	}
7714 	if (IN_FASTRECOVERY(tp->t_flags))
7715 		tp->t_flags |= TF_WASFRECOVERY;
7716 	else
7717 		tp->t_flags &= ~TF_WASFRECOVERY;
7718 	if (IN_CONGRECOVERY(tp->t_flags))
7719 		tp->t_flags |= TF_WASCRECOVERY;
7720 	else
7721 		tp->t_flags &= ~TF_WASCRECOVERY;
7722 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7723 	    (tp->snd_una == tp->snd_max)) {
7724 		/* Nothing outstanding .. nothing to do */
7725 		return (0);
7726 	}
7727 	if (rack->r_ctl.dsack_persist) {
7728 		rack->r_ctl.dsack_persist--;
7729 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7730 			rack->r_ctl.num_dsack = 0;
7731 		}
7732 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7733 	}
7734 	/*
7735 	 * Rack can only run one timer  at a time, so we cannot
7736 	 * run a KEEPINIT (gating SYN sending) and a retransmit
7737 	 * timer for the SYN. So if we are in a front state and
7738 	 * have a KEEPINIT timer we need to check the first transmit
7739 	 * against now to see if we have exceeded the KEEPINIT time
7740 	 * (if one is set).
7741 	 */
7742 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7743 	    (TP_KEEPINIT(tp) != 0)) {
7744 		struct rack_sendmap *rsm;
7745 
7746 		rsm = tqhash_min(rack->r_ctl.tqh);
7747 		if (rsm) {
7748 			/* Ok we have something outstanding to test keepinit with */
7749 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7750 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7751 				/* We have exceeded the KEEPINIT time */
7752 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7753 				goto drop_it;
7754 			}
7755 		}
7756 	}
7757 	/*
7758 	 * Retransmission timer went off.  Message has not been acked within
7759 	 * retransmit interval.  Back off to a longer retransmit interval
7760 	 * and retransmit one segment.
7761 	 */
7762 	if ((rack->r_ctl.rc_resend == NULL) ||
7763 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7764 		/*
7765 		 * If the rwnd collapsed on
7766 		 * the one we are retransmitting
7767 		 * it does not count against the
7768 		 * rxt count.
7769 		 */
7770 		tp->t_rxtshift++;
7771 	}
7772 	rack_remxt_tmr(tp);
7773 	if (tp->t_rxtshift > V_tcp_retries) {
7774 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7775 drop_it:
7776 		tp->t_rxtshift = V_tcp_retries;
7777 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7778 		/* XXXGL: previously t_softerror was casted to uint16_t */
7779 		MPASS(tp->t_softerror >= 0);
7780 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7781 		goto out;	/* tcp_drop() */
7782 	}
7783 	if (tp->t_state == TCPS_SYN_SENT) {
7784 		/*
7785 		 * If the SYN was retransmitted, indicate CWND to be limited
7786 		 * to 1 segment in cc_conn_init().
7787 		 */
7788 		tp->snd_cwnd = 1;
7789 	} else if (tp->t_rxtshift == 1) {
7790 		/*
7791 		 * first retransmit; record ssthresh and cwnd so they can be
7792 		 * recovered if this turns out to be a "bad" retransmit. A
7793 		 * retransmit is considered "bad" if an ACK for this segment
7794 		 * is received within RTT/2 interval; the assumption here is
7795 		 * that the ACK was already in flight.  See "On Estimating
7796 		 * End-to-End Network Path Properties" by Allman and Paxson
7797 		 * for more details.
7798 		 */
7799 		tp->snd_cwnd_prev = tp->snd_cwnd;
7800 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
7801 		tp->snd_recover_prev = tp->snd_recover;
7802 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7803 		tp->t_flags |= TF_PREVVALID;
7804 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7805 		tp->t_flags &= ~TF_PREVVALID;
7806 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7807 	if ((tp->t_state == TCPS_SYN_SENT) ||
7808 	    (tp->t_state == TCPS_SYN_RECEIVED))
7809 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7810 	else
7811 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7812 
7813 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7814 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7815 	/*
7816 	 * We enter the path for PLMTUD if connection is established or, if
7817 	 * connection is FIN_WAIT_1 status, reason for the last is that if
7818 	 * amount of data we send is very small, we could send it in couple
7819 	 * of packets and process straight to FIN. In that case we won't
7820 	 * catch ESTABLISHED state.
7821 	 */
7822 #ifdef INET6
7823 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7824 #else
7825 	isipv6 = false;
7826 #endif
7827 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
7828 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7829 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7830 	    ((tp->t_state == TCPS_ESTABLISHED) ||
7831 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
7832 		/*
7833 		 * Idea here is that at each stage of mtu probe (usually,
7834 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
7835 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7836 		 * should take care of that.
7837 		 */
7838 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7839 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7840 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7841 		    tp->t_rxtshift % 2 == 0)) {
7842 			/*
7843 			 * Enter Path MTU Black-hole Detection mechanism: -
7844 			 * Disable Path MTU Discovery (IP "DF" bit). -
7845 			 * Reduce MTU to lower value than what we negotiated
7846 			 * with peer.
7847 			 */
7848 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7849 				/* Record that we may have found a black hole. */
7850 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7851 				/* Keep track of previous MSS. */
7852 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7853 			}
7854 
7855 			/*
7856 			 * Reduce the MSS to blackhole value or to the
7857 			 * default in an attempt to retransmit.
7858 			 */
7859 #ifdef INET6
7860 			if (isipv6 &&
7861 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7862 				/* Use the sysctl tuneable blackhole MSS. */
7863 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7864 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7865 			} else if (isipv6) {
7866 				/* Use the default MSS. */
7867 				tp->t_maxseg = V_tcp_v6mssdflt;
7868 				/*
7869 				 * Disable Path MTU Discovery when we switch
7870 				 * to minmss.
7871 				 */
7872 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7873 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7874 			}
7875 #endif
7876 #if defined(INET6) && defined(INET)
7877 			else
7878 #endif
7879 #ifdef INET
7880 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7881 				/* Use the sysctl tuneable blackhole MSS. */
7882 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
7883 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7884 			} else {
7885 				/* Use the default MSS. */
7886 				tp->t_maxseg = V_tcp_mssdflt;
7887 				/*
7888 				 * Disable Path MTU Discovery when we switch
7889 				 * to minmss.
7890 				 */
7891 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7892 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7893 			}
7894 #endif
7895 		} else {
7896 			/*
7897 			 * If further retransmissions are still unsuccessful
7898 			 * with a lowered MTU, maybe this isn't a blackhole
7899 			 * and we restore the previous MSS and blackhole
7900 			 * detection flags. The limit '6' is determined by
7901 			 * giving each probe stage (1448, 1188, 524) 2
7902 			 * chances to recover.
7903 			 */
7904 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
7905 			    (tp->t_rxtshift >= 6)) {
7906 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
7907 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
7908 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
7909 				if (tp->t_maxseg < V_tcp_mssdflt) {
7910 					/*
7911 					 * The MSS is so small we should not
7912 					 * process incoming SACK's since we are
7913 					 * subject to attack in such a case.
7914 					 */
7915 					tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
7916 				} else {
7917 					tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
7918 				}
7919 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
7920 			}
7921 		}
7922 	}
7923 	/*
7924 	 * Disable RFC1323 and SACK if we haven't got any response to
7925 	 * our third SYN to work-around some broken terminal servers
7926 	 * (most of which have hopefully been retired) that have bad VJ
7927 	 * header compression code which trashes TCP segments containing
7928 	 * unknown-to-them TCP options.
7929 	 */
7930 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
7931 	    (tp->t_rxtshift == 3))
7932 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
7933 	/*
7934 	 * If we backed off this far, our srtt estimate is probably bogus.
7935 	 * Clobber it so we'll take the next rtt measurement as our srtt;
7936 	 * move the current srtt into rttvar to keep the current retransmit
7937 	 * times until then.
7938 	 */
7939 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
7940 #ifdef INET6
7941 		if ((inp->inp_vflag & INP_IPV6) != 0)
7942 			in6_losing(inp);
7943 		else
7944 #endif
7945 			in_losing(inp);
7946 		tp->t_rttvar += tp->t_srtt;
7947 		tp->t_srtt = 0;
7948 	}
7949 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
7950 	tp->snd_recover = tp->snd_max;
7951 	tp->t_flags |= TF_ACKNOW;
7952 	tp->t_rtttime = 0;
7953 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
7954 out:
7955 	return (retval);
7956 }
7957 
7958 static int
7959 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
7960 {
7961 	int32_t ret = 0;
7962 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
7963 
7964 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7965 	    (tp->t_flags & TF_GPUTINPROG)) {
7966 		/*
7967 		 * We have a goodput in progress
7968 		 * and we have entered a late state.
7969 		 * Do we have enough data in the sb
7970 		 * to handle the GPUT request?
7971 		 */
7972 		uint32_t bytes;
7973 
7974 		bytes = tp->gput_ack - tp->gput_seq;
7975 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
7976 			bytes += tp->gput_seq - tp->snd_una;
7977 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
7978 			/*
7979 			 * There are not enough bytes in the socket
7980 			 * buffer that have been sent to cover this
7981 			 * measurement. Cancel it.
7982 			 */
7983 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7984 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
7985 						   tp->gput_seq,
7986 						   0, 0, 18, __LINE__, NULL, 0);
7987 			tp->t_flags &= ~TF_GPUTINPROG;
7988 		}
7989 	}
7990 	if (timers == 0) {
7991 		return (0);
7992 	}
7993 	if (tp->t_state == TCPS_LISTEN) {
7994 		/* no timers on listen sockets */
7995 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7996 			return (0);
7997 		return (1);
7998 	}
7999 	if ((timers & PACE_TMR_RACK) &&
8000 	    rack->rc_on_min_to) {
8001 		/*
8002 		 * For the rack timer when we
8003 		 * are on a min-timeout (which means rrr_conf = 3)
8004 		 * we don't want to check the timer. It may
8005 		 * be going off for a pace and thats ok we
8006 		 * want to send the retransmit (if its ready).
8007 		 *
8008 		 * If its on a normal rack timer (non-min) then
8009 		 * we will check if its expired.
8010 		 */
8011 		goto skip_time_check;
8012 	}
8013 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
8014 		uint32_t left;
8015 
8016 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
8017 			ret = -1;
8018 			rack_log_to_processing(rack, cts, ret, 0);
8019 			return (0);
8020 		}
8021 		if (hpts_calling == 0) {
8022 			/*
8023 			 * A user send or queued mbuf (sack) has called us? We
8024 			 * return 0 and let the pacing guards
8025 			 * deal with it if they should or
8026 			 * should not cause a send.
8027 			 */
8028 			ret = -2;
8029 			rack_log_to_processing(rack, cts, ret, 0);
8030 			return (0);
8031 		}
8032 		/*
8033 		 * Ok our timer went off early and we are not paced false
8034 		 * alarm, go back to sleep. We make sure we don't have
8035 		 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8036 		 * flag in place.
8037 		 */
8038 		rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8039 		ret = -3;
8040 		left = rack->r_ctl.rc_timer_exp - cts;
8041 		tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8042 		rack_log_to_processing(rack, cts, ret, left);
8043 		return (1);
8044 	}
8045 skip_time_check:
8046 	rack->rc_tmr_stopped = 0;
8047 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8048 	if (timers & PACE_TMR_DELACK) {
8049 		ret = rack_timeout_delack(tp, rack, cts);
8050 	} else if (timers & PACE_TMR_RACK) {
8051 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8052 		rack->r_fast_output = 0;
8053 		ret = rack_timeout_rack(tp, rack, cts);
8054 	} else if (timers & PACE_TMR_TLP) {
8055 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8056 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8057 	} else if (timers & PACE_TMR_RXT) {
8058 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8059 		rack->r_fast_output = 0;
8060 		ret = rack_timeout_rxt(tp, rack, cts);
8061 	} else if (timers & PACE_TMR_PERSIT) {
8062 		ret = rack_timeout_persist(tp, rack, cts);
8063 	} else if (timers & PACE_TMR_KEEP) {
8064 		ret = rack_timeout_keepalive(tp, rack, cts);
8065 	}
8066 	rack_log_to_processing(rack, cts, ret, timers);
8067 	return (ret);
8068 }
8069 
8070 static void
8071 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8072 {
8073 	struct timeval tv;
8074 	uint32_t us_cts, flags_on_entry;
8075 	uint8_t hpts_removed = 0;
8076 
8077 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
8078 	us_cts = tcp_get_usecs(&tv);
8079 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8080 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8081 	     ((tp->snd_max - tp->snd_una) == 0))) {
8082 		tcp_hpts_remove(rack->rc_tp);
8083 		hpts_removed = 1;
8084 		/* If we were not delayed cancel out the flag. */
8085 		if ((tp->snd_max - tp->snd_una) == 0)
8086 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8087 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8088 	}
8089 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8090 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8091 		if (tcp_in_hpts(rack->rc_tp) &&
8092 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8093 			/*
8094 			 * Canceling timer's when we have no output being
8095 			 * paced. We also must remove ourselves from the
8096 			 * hpts.
8097 			 */
8098 			tcp_hpts_remove(rack->rc_tp);
8099 			hpts_removed = 1;
8100 		}
8101 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8102 	}
8103 	if (hpts_removed == 0)
8104 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8105 }
8106 
8107 static int
8108 rack_stopall(struct tcpcb *tp)
8109 {
8110 	struct tcp_rack *rack;
8111 
8112 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8113 	rack->t_timers_stopped = 1;
8114 
8115 	tcp_hpts_remove(tp);
8116 
8117 	return (0);
8118 }
8119 
8120 static void
8121 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8122 {
8123 	/*
8124 	 * Assure no timers are running.
8125 	 */
8126 	if (tcp_timer_active(tp, TT_PERSIST)) {
8127 		/* We enter in persists, set the flag appropriately */
8128 		rack->rc_in_persist = 1;
8129 	}
8130 	if (tcp_in_hpts(rack->rc_tp)) {
8131 		tcp_hpts_remove(rack->rc_tp);
8132 	}
8133 }
8134 
8135 static void
8136 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8137     struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz)
8138 {
8139 	int32_t idx;
8140 
8141 	rsm->r_rtr_cnt++;
8142 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8143 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8144 		rsm->r_flags |= RACK_OVERMAX;
8145 	}
8146 	rsm->r_act_rxt_cnt++;
8147 	/* Peg the count/index */
8148 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8149 	rsm->r_dupack = 0;
8150 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8151 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8152 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8153 	}
8154 	if (rsm->r_flags & RACK_WAS_LOST) {
8155 		/*
8156 		 * We retransmitted it putting it back in flight
8157 		 * remove the lost desgination and reduce the
8158 		 * bytes considered lost.
8159 		 */
8160 		rsm->r_flags  &= ~RACK_WAS_LOST;
8161 		KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
8162 			("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
8163 		if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
8164 			rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
8165 		else
8166 			rack->r_ctl.rc_considered_lost = 0;
8167 	}
8168 	idx = rsm->r_rtr_cnt - 1;
8169 	rsm->r_tim_lastsent[idx] = ts;
8170 	/*
8171 	 * Here we don't add in the len of send, since its already
8172 	 * in snduna <->snd_max.
8173 	 */
8174 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
8175 				     rack->r_ctl.rc_sacked);
8176 	if (rsm->r_flags & RACK_ACKED) {
8177 		/* Problably MTU discovery messing with us */
8178 		rsm->r_flags &= ~RACK_ACKED;
8179 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8180 	}
8181 	if (rsm->r_in_tmap) {
8182 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8183 		rsm->r_in_tmap = 0;
8184 	}
8185 	/* Lets make sure it really is in or not the GP window */
8186 	rack_mark_in_gp_win(tp, rsm);
8187 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8188 	rsm->r_in_tmap = 1;
8189 	rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8190 	/* Take off the must retransmit flag, if its on */
8191 	if (rsm->r_flags & RACK_MUST_RXT) {
8192 		if (rack->r_must_retran)
8193 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8194 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8195 			/*
8196 			 * We have retransmitted all we need. Clear
8197 			 * any must retransmit flags.
8198 			 */
8199 			rack->r_must_retran = 0;
8200 			rack->r_ctl.rc_out_at_rto = 0;
8201 		}
8202 		rsm->r_flags &= ~RACK_MUST_RXT;
8203 	}
8204 	/* Remove any collapsed flag */
8205 	rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8206 	if (rsm->r_flags & RACK_SACK_PASSED) {
8207 		/* We have retransmitted due to the SACK pass */
8208 		rsm->r_flags &= ~RACK_SACK_PASSED;
8209 		rsm->r_flags |= RACK_WAS_SACKPASS;
8210 	}
8211 }
8212 
8213 static uint32_t
8214 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8215     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint32_t add_flag, int segsiz)
8216 {
8217 	/*
8218 	 * We (re-)transmitted starting at rsm->r_start for some length
8219 	 * (possibly less than r_end.
8220 	 */
8221 	struct rack_sendmap *nrsm;
8222 	int insret __diagused;
8223 	uint32_t c_end;
8224 	int32_t len;
8225 
8226 	len = *lenp;
8227 	c_end = rsm->r_start + len;
8228 	if (SEQ_GEQ(c_end, rsm->r_end)) {
8229 		/*
8230 		 * We retransmitted the whole piece or more than the whole
8231 		 * slopping into the next rsm.
8232 		 */
8233 		rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8234 		if (c_end == rsm->r_end) {
8235 			*lenp = 0;
8236 			return (0);
8237 		} else {
8238 			int32_t act_len;
8239 
8240 			/* Hangs over the end return whats left */
8241 			act_len = rsm->r_end - rsm->r_start;
8242 			*lenp = (len - act_len);
8243 			return (rsm->r_end);
8244 		}
8245 		/* We don't get out of this block. */
8246 	}
8247 	/*
8248 	 * Here we retransmitted less than the whole thing which means we
8249 	 * have to split this into what was transmitted and what was not.
8250 	 */
8251 	nrsm = rack_alloc_full_limit(rack);
8252 	if (nrsm == NULL) {
8253 		/*
8254 		 * We can't get memory, so lets not proceed.
8255 		 */
8256 		*lenp = 0;
8257 		return (0);
8258 	}
8259 	/*
8260 	 * So here we are going to take the original rsm and make it what we
8261 	 * retransmitted. nrsm will be the tail portion we did not
8262 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8263 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8264 	 * 1, 6 and the new piece will be 6, 11.
8265 	 */
8266 	rack_clone_rsm(rack, nrsm, rsm, c_end);
8267 	nrsm->r_dupack = 0;
8268 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8269 #ifndef INVARIANTS
8270 	(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8271 #else
8272 	if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8273 		panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8274 		      nrsm, insret, rack, rsm);
8275 	}
8276 #endif
8277 	if (rsm->r_in_tmap) {
8278 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8279 		nrsm->r_in_tmap = 1;
8280 	}
8281 	rsm->r_flags &= (~RACK_HAS_FIN);
8282 	rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8283 	/* Log a split of rsm into rsm and nrsm */
8284 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8285 	*lenp = 0;
8286 	return (0);
8287 }
8288 
8289 static void
8290 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8291 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8292 		struct rack_sendmap *hintrsm, uint32_t add_flag, struct mbuf *s_mb,
8293 		uint32_t s_moff, int hw_tls, int segsiz)
8294 {
8295 	struct tcp_rack *rack;
8296 	struct rack_sendmap *rsm, *nrsm;
8297 	int insret __diagused;
8298 
8299 	register uint32_t snd_max, snd_una;
8300 
8301 	/*
8302 	 * Add to the RACK log of packets in flight or retransmitted. If
8303 	 * there is a TS option we will use the TS echoed, if not we will
8304 	 * grab a TS.
8305 	 *
8306 	 * Retransmissions will increment the count and move the ts to its
8307 	 * proper place. Note that if options do not include TS's then we
8308 	 * won't be able to effectively use the ACK for an RTT on a retran.
8309 	 *
8310 	 * Notes about r_start and r_end. Lets consider a send starting at
8311 	 * sequence 1 for 10 bytes. In such an example the r_start would be
8312 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8313 	 * This means that r_end is actually the first sequence for the next
8314 	 * slot (11).
8315 	 *
8316 	 */
8317 	/*
8318 	 * If err is set what do we do XXXrrs? should we not add the thing?
8319 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
8320 	 * i.e. proceed with add ** do this for now.
8321 	 */
8322 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8323 	if (err)
8324 		/*
8325 		 * We don't log errors -- we could but snd_max does not
8326 		 * advance in this case either.
8327 		 */
8328 		return;
8329 
8330 	if (th_flags & TH_RST) {
8331 		/*
8332 		 * We don't log resets and we return immediately from
8333 		 * sending
8334 		 */
8335 		return;
8336 	}
8337 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8338 	snd_una = tp->snd_una;
8339 	snd_max = tp->snd_max;
8340 	if (th_flags & (TH_SYN | TH_FIN)) {
8341 		/*
8342 		 * The call to rack_log_output is made before bumping
8343 		 * snd_max. This means we can record one extra byte on a SYN
8344 		 * or FIN if seq_out is adding more on and a FIN is present
8345 		 * (and we are not resending).
8346 		 */
8347 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8348 			len++;
8349 		if (th_flags & TH_FIN)
8350 			len++;
8351 	}
8352 	if (SEQ_LEQ((seq_out + len), snd_una)) {
8353 		/* Are sending an old segment to induce an ack (keep-alive)? */
8354 		return;
8355 	}
8356 	if (SEQ_LT(seq_out, snd_una)) {
8357 		/* huh? should we panic? */
8358 		uint32_t end;
8359 
8360 		end = seq_out + len;
8361 		seq_out = snd_una;
8362 		if (SEQ_GEQ(end, seq_out))
8363 			len = end - seq_out;
8364 		else
8365 			len = 0;
8366 	}
8367 	if (len == 0) {
8368 		/* We don't log zero window probes */
8369 		return;
8370 	}
8371 	if (IN_FASTRECOVERY(tp->t_flags)) {
8372 		rack->r_ctl.rc_prr_out += len;
8373 	}
8374 	/* First question is it a retransmission or new? */
8375 	if (seq_out == snd_max) {
8376 		/* Its new */
8377 		rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
8378 again:
8379 		rsm = rack_alloc(rack);
8380 		if (rsm == NULL) {
8381 			/*
8382 			 * Hmm out of memory and the tcb got destroyed while
8383 			 * we tried to wait.
8384 			 */
8385 			return;
8386 		}
8387 		if (th_flags & TH_FIN) {
8388 			rsm->r_flags = RACK_HAS_FIN|add_flag;
8389 		} else {
8390 			rsm->r_flags = add_flag;
8391 		}
8392 		if (hw_tls)
8393 			rsm->r_hw_tls = 1;
8394 		rsm->r_tim_lastsent[0] = cts;
8395 		rsm->r_rtr_cnt = 1;
8396  		rsm->r_act_rxt_cnt = 0;
8397 		rsm->r_rtr_bytes = 0;
8398 		if (th_flags & TH_SYN) {
8399 			/* The data space is one beyond snd_una */
8400 			rsm->r_flags |= RACK_HAS_SYN;
8401 		}
8402 		rsm->r_start = seq_out;
8403 		rsm->r_end = rsm->r_start + len;
8404 		rack_mark_in_gp_win(tp, rsm);
8405 		rsm->r_dupack = 0;
8406 		/*
8407 		 * save off the mbuf location that
8408 		 * sndmbuf_noadv returned (which is
8409 		 * where we started copying from)..
8410 		 */
8411 		rsm->m = s_mb;
8412 		rsm->soff = s_moff;
8413 		/*
8414 		 * Here we do add in the len of send, since its not yet
8415 		 * reflected in in snduna <->snd_max
8416 		 */
8417 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8418 					      rack->r_ctl.rc_sacked) +
8419 			      (rsm->r_end - rsm->r_start));
8420 		if ((rack->rc_initial_ss_comp == 0) &&
8421 		    (rack->r_ctl.ss_hi_fs < rsm->r_fas)) {
8422 			   rack->r_ctl.ss_hi_fs = rsm->r_fas;
8423 		}
8424 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8425 		if (rsm->m) {
8426 			if (rsm->m->m_len <= rsm->soff) {
8427 				/*
8428 				 * XXXrrs Question, will this happen?
8429 				 *
8430 				 * If sbsndptr is set at the correct place
8431 				 * then s_moff should always be somewhere
8432 				 * within rsm->m. But if the sbsndptr was
8433 				 * off then that won't be true. If it occurs
8434 				 * we need to walkout to the correct location.
8435 				 */
8436 				struct mbuf *lm;
8437 
8438 				lm = rsm->m;
8439 				while (lm->m_len <= rsm->soff) {
8440 					rsm->soff -= lm->m_len;
8441 					lm = lm->m_next;
8442 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8443 							     __func__, rack, s_moff, s_mb, rsm->soff));
8444 				}
8445 				rsm->m = lm;
8446 			}
8447 			rsm->orig_m_len = rsm->m->m_len;
8448 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8449 		} else {
8450 			rsm->orig_m_len = 0;
8451 			rsm->orig_t_space = 0;
8452 		}
8453 		rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8454 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8455 		/* Log a new rsm */
8456 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8457 #ifndef INVARIANTS
8458 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
8459 #else
8460 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8461 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8462 			      nrsm, insret, rack, rsm);
8463 		}
8464 #endif
8465 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8466 		rsm->r_in_tmap = 1;
8467 		if (rsm->r_flags & RACK_IS_PCM) {
8468 			rack->r_ctl.pcm_i.send_time = cts;
8469 			rack->r_ctl.pcm_i.eseq = rsm->r_end;
8470 			/* First time through we set the start too */
8471 			if (rack->pcm_in_progress == 0)
8472 				rack->r_ctl.pcm_i.sseq = rsm->r_start;
8473 		}
8474 		/*
8475 		 * Special case detection, is there just a single
8476 		 * packet outstanding when we are not in recovery?
8477 		 *
8478 		 * If this is true mark it so.
8479 		 */
8480 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8481 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8482 			struct rack_sendmap *prsm;
8483 
8484 			prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8485 			if (prsm)
8486 				prsm->r_one_out_nr = 1;
8487 		}
8488 		return;
8489 	}
8490 	/*
8491 	 * If we reach here its a retransmission and we need to find it.
8492 	 */
8493 more:
8494 	if (hintrsm && (hintrsm->r_start == seq_out)) {
8495 		rsm = hintrsm;
8496 		hintrsm = NULL;
8497 	} else {
8498 		/* No hints sorry */
8499 		rsm = NULL;
8500 	}
8501 	if ((rsm) && (rsm->r_start == seq_out)) {
8502 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8503 		if (len == 0) {
8504 			return;
8505 		} else {
8506 			goto more;
8507 		}
8508 	}
8509 	/* Ok it was not the last pointer go through it the hard way. */
8510 refind:
8511 	rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8512 	if (rsm) {
8513 		if (rsm->r_start == seq_out) {
8514 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8515 			if (len == 0) {
8516 				return;
8517 			} else {
8518 				goto refind;
8519 			}
8520 		}
8521 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8522 			/* Transmitted within this piece */
8523 			/*
8524 			 * Ok we must split off the front and then let the
8525 			 * update do the rest
8526 			 */
8527 			nrsm = rack_alloc_full_limit(rack);
8528 			if (nrsm == NULL) {
8529 				rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8530 				return;
8531 			}
8532 			/*
8533 			 * copy rsm to nrsm and then trim the front of rsm
8534 			 * to not include this part.
8535 			 */
8536 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
8537 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8538 #ifndef INVARIANTS
8539 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8540 #else
8541 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8542 				panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8543 				      nrsm, insret, rack, rsm);
8544 			}
8545 #endif
8546 			if (rsm->r_in_tmap) {
8547 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8548 				nrsm->r_in_tmap = 1;
8549 			}
8550 			rsm->r_flags &= (~RACK_HAS_FIN);
8551 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8552 			if (len == 0) {
8553 				return;
8554 			} else if (len > 0)
8555 				goto refind;
8556 		}
8557 	}
8558 	/*
8559 	 * Hmm not found in map did they retransmit both old and on into the
8560 	 * new?
8561 	 */
8562 	if (seq_out == tp->snd_max) {
8563 		goto again;
8564 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
8565 #ifdef INVARIANTS
8566 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8567 		       seq_out, len, tp->snd_una, tp->snd_max);
8568 		printf("Starting Dump of all rack entries\n");
8569 		TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8570 			printf("rsm:%p start:%u end:%u\n",
8571 			       rsm, rsm->r_start, rsm->r_end);
8572 		}
8573 		printf("Dump complete\n");
8574 		panic("seq_out not found rack:%p tp:%p",
8575 		      rack, tp);
8576 #endif
8577 	} else {
8578 #ifdef INVARIANTS
8579 		/*
8580 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
8581 		 * flag)
8582 		 */
8583 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8584 		      seq_out, len, tp->snd_max, tp);
8585 #endif
8586 	}
8587 }
8588 
8589 /*
8590  * Record one of the RTT updates from an ack into
8591  * our sample structure.
8592  */
8593 
8594 static void
8595 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8596 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8597 {
8598 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8599 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8600 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8601 	}
8602 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8603 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8604 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8605 	}
8606 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8607 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8608 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
8609 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8610 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8611 	}
8612 	if ((confidence == 1) &&
8613 	    ((rsm == NULL) ||
8614 	     (rsm->r_just_ret) ||
8615 	     (rsm->r_one_out_nr &&
8616 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8617 		/*
8618 		 * If the rsm had a just return
8619 		 * hit it then we can't trust the
8620 		 * rtt measurement for buffer deterimination
8621 		 * Note that a confidence of 2, indicates
8622 		 * SACK'd which overrides the r_just_ret or
8623 		 * the r_one_out_nr. If it was a CUM-ACK and
8624 		 * we had only two outstanding, but get an
8625 		 * ack for only 1. Then that also lowers our
8626 		 * confidence.
8627 		 */
8628 		confidence = 0;
8629 	}
8630 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8631 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8632 		if (rack->r_ctl.rack_rs.confidence == 0) {
8633 			/*
8634 			 * We take anything with no current confidence
8635 			 * saved.
8636 			 */
8637 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8638 			rack->r_ctl.rack_rs.confidence = confidence;
8639 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8640 		} else if (confidence != 0) {
8641 			/*
8642 			 * Once we have a confident number,
8643 			 * we can update it with a smaller
8644 			 * value since this confident number
8645 			 * may include the DSACK time until
8646 			 * the next segment (the second one) arrived.
8647 			 */
8648 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8649 			rack->r_ctl.rack_rs.confidence = confidence;
8650 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8651 		}
8652 	}
8653 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8654 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8655 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8656 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
8657 }
8658 
8659 /*
8660  * Collect new round-trip time estimate
8661  * and update averages and current timeout.
8662  */
8663 static void
8664 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8665 {
8666 	int32_t delta;
8667 	int32_t rtt;
8668 
8669 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8670 		/* No valid sample */
8671 		return;
8672 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8673 		/* We are to use the lowest RTT seen in a single ack */
8674 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8675 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8676 		/* We are to use the highest RTT seen in a single ack */
8677 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8678 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8679 		/* We are to use the average RTT seen in a single ack */
8680 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8681 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8682 	} else {
8683 #ifdef INVARIANTS
8684 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8685 #endif
8686 		return;
8687 	}
8688 	if (rtt == 0)
8689 		rtt = 1;
8690 	if (rack->rc_gp_rtt_set == 0) {
8691 		/*
8692 		 * With no RTT we have to accept
8693 		 * even one we are not confident of.
8694 		 */
8695 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8696 		rack->rc_gp_rtt_set = 1;
8697 	} else if (rack->r_ctl.rack_rs.confidence) {
8698 		/* update the running gp srtt */
8699 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8700 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8701 	}
8702 	if (rack->r_ctl.rack_rs.confidence) {
8703 		/*
8704 		 * record the low and high for highly buffered path computation,
8705 		 * we only do this if we are confident (not a retransmission).
8706 		 */
8707 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8708 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8709 		}
8710 		if (rack->rc_highly_buffered == 0) {
8711 			/*
8712 			 * Currently once we declare a path has
8713 			 * highly buffered there is no going
8714 			 * back, which may be a problem...
8715 			 */
8716 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8717 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8718 						     rack->r_ctl.rc_highest_us_rtt,
8719 						     rack->r_ctl.rc_lowest_us_rtt,
8720 						     RACK_RTTS_SEEHBP);
8721 				rack->rc_highly_buffered = 1;
8722 			}
8723 		}
8724 	}
8725 	if ((rack->r_ctl.rack_rs.confidence) ||
8726 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8727 		/*
8728 		 * If we are highly confident of it <or> it was
8729 		 * never retransmitted we accept it as the last us_rtt.
8730 		 */
8731 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8732 		/* The lowest rtt can be set if its was not retransmited */
8733 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8734 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8735 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
8736 				rack->r_ctl.rc_lowest_us_rtt = 1;
8737 		}
8738 	}
8739 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8740 	if (tp->t_srtt != 0) {
8741 		/*
8742 		 * We keep a simple srtt in microseconds, like our rtt
8743 		 * measurement. We don't need to do any tricks with shifting
8744 		 * etc. Instead we just add in 1/8th of the new measurement
8745 		 * and subtract out 1/8 of the old srtt. We do the same with
8746 		 * the variance after finding the absolute value of the
8747 		 * difference between this sample and the current srtt.
8748 		 */
8749 		delta = tp->t_srtt - rtt;
8750 		/* Take off 1/8th of the current sRTT */
8751 		tp->t_srtt -= (tp->t_srtt >> 3);
8752 		/* Add in 1/8th of the new RTT just measured */
8753 		tp->t_srtt += (rtt >> 3);
8754 		if (tp->t_srtt <= 0)
8755 			tp->t_srtt = 1;
8756 		/* Now lets make the absolute value of the variance */
8757 		if (delta < 0)
8758 			delta = -delta;
8759 		/* Subtract out 1/8th */
8760 		tp->t_rttvar -= (tp->t_rttvar >> 3);
8761 		/* Add in 1/8th of the new variance we just saw */
8762 		tp->t_rttvar += (delta >> 3);
8763 		if (tp->t_rttvar <= 0)
8764 			tp->t_rttvar = 1;
8765 	} else {
8766 		/*
8767 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
8768 		 * variance to half the rtt (so our first retransmit happens
8769 		 * at 3*rtt).
8770 		 */
8771 		tp->t_srtt = rtt;
8772 		tp->t_rttvar = rtt >> 1;
8773 	}
8774 	rack->rc_srtt_measure_made = 1;
8775 	KMOD_TCPSTAT_INC(tcps_rttupdated);
8776 	if (tp->t_rttupdated < UCHAR_MAX)
8777 		tp->t_rttupdated++;
8778 #ifdef STATS
8779 	if (rack_stats_gets_ms_rtt == 0) {
8780 		/* Send in the microsecond rtt used for rxt timeout purposes */
8781 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8782 	} else if (rack_stats_gets_ms_rtt == 1) {
8783 		/* Send in the millisecond rtt used for rxt timeout purposes */
8784 		int32_t ms_rtt;
8785 
8786 		/* Round up */
8787 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8788 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8789 	} else if (rack_stats_gets_ms_rtt == 2) {
8790 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
8791 		int32_t ms_rtt;
8792 
8793 		/* Round up */
8794 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8795 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8796 	}  else {
8797 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
8798 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8799 	}
8800 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8801 #endif
8802 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
8803 	/*
8804 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
8805 	 * way we do the smoothing, srtt and rttvar will each average +1/2
8806 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
8807 	 * tick of rounding and 1 extra tick because of +-1/2 tick
8808 	 * uncertainty in the firing of the timer.  The bias will give us
8809 	 * exactly the 1.5 tick we need.  But, because the bias is
8810 	 * statistical, we have to test that we don't drop below the minimum
8811 	 * feasible timer (which is 2 ticks).
8812 	 */
8813 	tp->t_rxtshift = 0;
8814 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8815 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8816 	rack_log_rtt_sample(rack, rtt);
8817 	tp->t_softerror = 0;
8818 }
8819 
8820 
8821 static void
8822 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8823 {
8824 	/*
8825 	 * Apply to filter the inbound us-rtt at us_cts.
8826 	 */
8827 	uint32_t old_rtt;
8828 
8829 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8830 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8831 			       us_rtt, us_cts);
8832 	if (old_rtt > us_rtt) {
8833 		/* We just hit a new lower rtt time */
8834 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
8835 				     __LINE__, RACK_RTTS_NEWRTT);
8836 		/*
8837 		 * Only count it if its lower than what we saw within our
8838 		 * calculated range.
8839 		 */
8840 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8841 			if (rack_probertt_lower_within &&
8842 			    rack->rc_gp_dyn_mul &&
8843 			    (rack->use_fixed_rate == 0) &&
8844 			    (rack->rc_always_pace)) {
8845 				/*
8846 				 * We are seeing a new lower rtt very close
8847 				 * to the time that we would have entered probe-rtt.
8848 				 * This is probably due to the fact that a peer flow
8849 				 * has entered probe-rtt. Lets go in now too.
8850 				 */
8851 				uint32_t val;
8852 
8853 				val = rack_probertt_lower_within * rack_time_between_probertt;
8854 				val /= 100;
8855 				if ((rack->in_probe_rtt == 0)  &&
8856 				    (rack->rc_skip_timely == 0) &&
8857 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
8858 					rack_enter_probertt(rack, us_cts);
8859 				}
8860 			}
8861 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8862 		}
8863 	}
8864 }
8865 
8866 static int
8867 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8868     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8869 {
8870 	uint32_t us_rtt;
8871 	int32_t i, all;
8872 	uint32_t t, len_acked;
8873 
8874 	if ((rsm->r_flags & RACK_ACKED) ||
8875 	    (rsm->r_flags & RACK_WAS_ACKED))
8876 		/* Already done */
8877 		return (0);
8878 	if (rsm->r_no_rtt_allowed) {
8879 		/* Not allowed */
8880 		return (0);
8881 	}
8882 	if (ack_type == CUM_ACKED) {
8883 		if (SEQ_GT(th_ack, rsm->r_end)) {
8884 			len_acked = rsm->r_end - rsm->r_start;
8885 			all = 1;
8886 		} else {
8887 			len_acked = th_ack - rsm->r_start;
8888 			all = 0;
8889 		}
8890 	} else {
8891 		len_acked = rsm->r_end - rsm->r_start;
8892 		all = 0;
8893 	}
8894 	if (rsm->r_rtr_cnt == 1) {
8895 
8896 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8897 		if ((int)t <= 0)
8898 			t = 1;
8899 		if (!tp->t_rttlow || tp->t_rttlow > t)
8900 			tp->t_rttlow = t;
8901 		if (!rack->r_ctl.rc_rack_min_rtt ||
8902 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8903 			rack->r_ctl.rc_rack_min_rtt = t;
8904 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
8905 				rack->r_ctl.rc_rack_min_rtt = 1;
8906 			}
8907 		}
8908 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8909 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8910 		else
8911 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8912 		if (us_rtt == 0)
8913 			us_rtt = 1;
8914 		if (CC_ALGO(tp)->rttsample != NULL) {
8915 			/* Kick the RTT to the CC */
8916 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8917 		}
8918 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
8919 		if (ack_type == SACKED) {
8920 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
8921 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
8922 		} else {
8923 			/*
8924 			 * We need to setup what our confidence
8925 			 * is in this ack.
8926 			 *
8927 			 * If the rsm was app limited and it is
8928 			 * less than a mss in length (the end
8929 			 * of the send) then we have a gap. If we
8930 			 * were app limited but say we were sending
8931 			 * multiple MSS's then we are more confident
8932 			 * int it.
8933 			 *
8934 			 * When we are not app-limited then we see if
8935 			 * the rsm is being included in the current
8936 			 * measurement, we tell this by the app_limited_needs_set
8937 			 * flag.
8938 			 *
8939 			 * Note that being cwnd blocked is not applimited
8940 			 * as well as the pacing delay between packets which
8941 			 * are sending only 1 or 2 MSS's also will show up
8942 			 * in the RTT. We probably need to examine this algorithm
8943 			 * a bit more and enhance it to account for the delay
8944 			 * between rsm's. We could do that by saving off the
8945 			 * pacing delay of each rsm (in an rsm) and then
8946 			 * factoring that in somehow though for now I am
8947 			 * not sure how :)
8948 			 */
8949 			int calc_conf = 0;
8950 
8951 			if (rsm->r_flags & RACK_APP_LIMITED) {
8952 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
8953 					calc_conf = 0;
8954 				else
8955 					calc_conf = 1;
8956 			} else if (rack->app_limited_needs_set == 0) {
8957 				calc_conf = 1;
8958 			} else {
8959 				calc_conf = 0;
8960 			}
8961 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
8962 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
8963 					    calc_conf, rsm, rsm->r_rtr_cnt);
8964 		}
8965 		if ((rsm->r_flags & RACK_TLP) &&
8966 		    (!IN_FASTRECOVERY(tp->t_flags))) {
8967 			/* Segment was a TLP and our retrans matched */
8968 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
8969 				rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
8970 			}
8971 		}
8972 		if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
8973 		    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
8974 			    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
8975 			/* New more recent rack_tmit_time */
8976 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8977 			if (rack->r_ctl.rc_rack_tmit_time == 0)
8978 				rack->r_ctl.rc_rack_tmit_time = 1;
8979 			rack->rc_rack_rtt = t;
8980 		}
8981 		return (1);
8982 	}
8983 	/*
8984 	 * We clear the soft/rxtshift since we got an ack.
8985 	 * There is no assurance we will call the commit() function
8986 	 * so we need to clear these to avoid incorrect handling.
8987 	 */
8988 	tp->t_rxtshift = 0;
8989 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8990 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
8991 	tp->t_softerror = 0;
8992 	if (to && (to->to_flags & TOF_TS) &&
8993 	    (ack_type == CUM_ACKED) &&
8994 	    (to->to_tsecr) &&
8995 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
8996 		/*
8997 		 * Now which timestamp does it match? In this block the ACK
8998 		 * must be coming from a previous transmission.
8999 		 */
9000 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
9001 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
9002 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9003 				if ((int)t <= 0)
9004 					t = 1;
9005 				if (CC_ALGO(tp)->rttsample != NULL) {
9006 					/*
9007 					 * Kick the RTT to the CC, here
9008 					 * we lie a bit in that we know the
9009 					 * retransmission is correct even though
9010 					 * we retransmitted. This is because
9011 					 * we match the timestamps.
9012 					 */
9013 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
9014 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
9015 					else
9016 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
9017 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9018 				}
9019 				if ((i + 1) < rsm->r_rtr_cnt) {
9020 					/*
9021 					 * The peer ack'd from our previous
9022 					 * transmission. We have a spurious
9023 					 * retransmission and thus we dont
9024 					 * want to update our rack_rtt.
9025 					 *
9026 					 * Hmm should there be a CC revert here?
9027 					 *
9028 					 */
9029 					return (0);
9030 				}
9031 				if (!tp->t_rttlow || tp->t_rttlow > t)
9032 					tp->t_rttlow = t;
9033 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9034 					rack->r_ctl.rc_rack_min_rtt = t;
9035 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
9036 						rack->r_ctl.rc_rack_min_rtt = 1;
9037 					}
9038 				}
9039 				if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9040 				    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9041 					    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9042 					/* New more recent rack_tmit_time */
9043 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9044 					if (rack->r_ctl.rc_rack_tmit_time == 0)
9045 						rack->r_ctl.rc_rack_tmit_time = 1;
9046 					rack->rc_rack_rtt = t;
9047 				}
9048 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9049 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9050 						    rsm->r_rtr_cnt);
9051 				return (1);
9052 			}
9053 		}
9054 		/* If we are logging log out the sendmap */
9055 		if (tcp_bblogging_on(rack->rc_tp)) {
9056 			for (i = 0; i < rsm->r_rtr_cnt; i++) {
9057 				rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9058 			}
9059 		}
9060 		goto ts_not_found;
9061 	} else {
9062 		/*
9063 		 * Ok its a SACK block that we retransmitted. or a windows
9064 		 * machine without timestamps. We can tell nothing from the
9065 		 * time-stamp since its not there or the time the peer last
9066 		 * received a segment that moved forward its cum-ack point.
9067 		 */
9068 ts_not_found:
9069 		i = rsm->r_rtr_cnt - 1;
9070 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9071 		if ((int)t <= 0)
9072 			t = 1;
9073 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9074 			/*
9075 			 * We retransmitted and the ack came back in less
9076 			 * than the smallest rtt we have observed. We most
9077 			 * likely did an improper retransmit as outlined in
9078 			 * 6.2 Step 2 point 2 in the rack-draft so we
9079 			 * don't want to update our rack_rtt. We in
9080 			 * theory (in future) might want to think about reverting our
9081 			 * cwnd state but we won't for now.
9082 			 */
9083 			return (0);
9084 		} else if (rack->r_ctl.rc_rack_min_rtt) {
9085 			/*
9086 			 * We retransmitted it and the retransmit did the
9087 			 * job.
9088 			 */
9089 			if (!rack->r_ctl.rc_rack_min_rtt ||
9090 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9091 				rack->r_ctl.rc_rack_min_rtt = t;
9092 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
9093 					rack->r_ctl.rc_rack_min_rtt = 1;
9094 				}
9095 			}
9096 			if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9097 			    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9098 				    (uint32_t)rsm->r_tim_lastsent[i]))) {
9099 				/* New more recent rack_tmit_time */
9100 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9101 				if (rack->r_ctl.rc_rack_tmit_time == 0)
9102 					rack->r_ctl.rc_rack_tmit_time = 1;
9103 				rack->rc_rack_rtt = t;
9104 			}
9105 			return (1);
9106 		}
9107 	}
9108 	return (0);
9109 }
9110 
9111 /*
9112  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9113  */
9114 static void
9115 rack_log_sack_passed(struct tcpcb *tp,
9116     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
9117 {
9118 	struct rack_sendmap *nrsm;
9119 	uint32_t thresh;
9120 
9121 	/* Get our rxt threshold for lost consideration */
9122 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
9123 	/* Now start looking at rsm's */
9124 	nrsm = rsm;
9125 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9126 	    rack_head, r_tnext) {
9127 		if (nrsm == rsm) {
9128 			/* Skip original segment he is acked */
9129 			continue;
9130 		}
9131 		if (nrsm->r_flags & RACK_ACKED) {
9132 			/*
9133 			 * Skip ack'd segments, though we
9134 			 * should not see these, since tmap
9135 			 * should not have ack'd segments.
9136 			 */
9137 			continue;
9138 		}
9139 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9140 			/*
9141 			 * If the peer dropped the rwnd on
9142 			 * these then we don't worry about them.
9143 			 */
9144 			continue;
9145 		}
9146 		/* Check lost state */
9147 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
9148 			uint32_t exp;
9149 
9150 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
9151 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
9152 				/* We consider it lost */
9153 				nrsm->r_flags |= RACK_WAS_LOST;
9154 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
9155 			}
9156 		}
9157 		if (nrsm->r_flags & RACK_SACK_PASSED) {
9158 			/*
9159 			 * We found one that is already marked
9160 			 * passed, we have been here before and
9161 			 * so all others below this are marked.
9162 			 */
9163 			break;
9164 		}
9165 		nrsm->r_flags |= RACK_SACK_PASSED;
9166 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9167 	}
9168 }
9169 
9170 static void
9171 rack_need_set_test(struct tcpcb *tp,
9172 		   struct tcp_rack *rack,
9173 		   struct rack_sendmap *rsm,
9174 		   tcp_seq th_ack,
9175 		   int line,
9176 		   int use_which)
9177 {
9178 	struct rack_sendmap *s_rsm;
9179 
9180 	if ((tp->t_flags & TF_GPUTINPROG) &&
9181 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9182 		/*
9183 		 * We were app limited, and this ack
9184 		 * butts up or goes beyond the point where we want
9185 		 * to start our next measurement. We need
9186 		 * to record the new gput_ts as here and
9187 		 * possibly update the start sequence.
9188 		 */
9189 		uint32_t seq, ts;
9190 
9191 		if (rsm->r_rtr_cnt > 1) {
9192 			/*
9193 			 * This is a retransmit, can we
9194 			 * really make any assessment at this
9195 			 * point?  We are not really sure of
9196 			 * the timestamp, is it this or the
9197 			 * previous transmission?
9198 			 *
9199 			 * Lets wait for something better that
9200 			 * is not retransmitted.
9201 			 */
9202 			return;
9203 		}
9204 		seq = tp->gput_seq;
9205 		ts = tp->gput_ts;
9206 		rack->app_limited_needs_set = 0;
9207 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9208 		/* Do we start at a new end? */
9209 		if ((use_which == RACK_USE_BEG) &&
9210 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9211 			/*
9212 			 * When we get an ACK that just eats
9213 			 * up some of the rsm, we set RACK_USE_BEG
9214 			 * since whats at r_start (i.e. th_ack)
9215 			 * is left unacked and thats where the
9216 			 * measurement now starts.
9217 			 */
9218 			tp->gput_seq = rsm->r_start;
9219 		}
9220 		if ((use_which == RACK_USE_END) &&
9221 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9222 			/*
9223 			 * We use the end when the cumack
9224 			 * is moving forward and completely
9225 			 * deleting the rsm passed so basically
9226 			 * r_end holds th_ack.
9227 			 *
9228 			 * For SACK's we also want to use the end
9229 			 * since this piece just got sacked and
9230 			 * we want to target anything after that
9231 			 * in our measurement.
9232 			 */
9233 			tp->gput_seq = rsm->r_end;
9234 		}
9235 		if (use_which == RACK_USE_END_OR_THACK) {
9236 			/*
9237 			 * special case for ack moving forward,
9238 			 * not a sack, we need to move all the
9239 			 * way up to where this ack cum-ack moves
9240 			 * to.
9241 			 */
9242 			if (SEQ_GT(th_ack, rsm->r_end))
9243 				tp->gput_seq = th_ack;
9244 			else
9245 				tp->gput_seq = rsm->r_end;
9246 		}
9247 		if (SEQ_LT(tp->gput_seq, tp->snd_max))
9248 			s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9249 		else
9250 			s_rsm = NULL;
9251 		/*
9252 		 * Pick up the correct send time if we can the rsm passed in
9253 		 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9254 		 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9255 		 * find a different seq i.e. the next send up.
9256 		 *
9257 		 * If that has not been sent, s_rsm will be NULL and we must
9258 		 * arrange it so this function will get called again by setting
9259 		 * app_limited_needs_set.
9260 		 */
9261 		if (s_rsm)
9262 			rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9263 		else {
9264 			/* If we hit here we have to have *not* sent tp->gput_seq */
9265 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9266 			/* Set it up so we will go through here again */
9267 			rack->app_limited_needs_set = 1;
9268 		}
9269 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9270 			/*
9271 			 * We moved beyond this guy's range, re-calculate
9272 			 * the new end point.
9273 			 */
9274 			if (rack->rc_gp_filled == 0) {
9275 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9276 			} else {
9277 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9278 			}
9279 		}
9280 		/*
9281 		 * We are moving the goal post, we may be able to clear the
9282 		 * measure_saw_probe_rtt flag.
9283 		 */
9284 		if ((rack->in_probe_rtt == 0) &&
9285 		    (rack->measure_saw_probe_rtt) &&
9286 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9287 			rack->measure_saw_probe_rtt = 0;
9288 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9289 					   seq, tp->gput_seq,
9290 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9291 					    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9292 					   5, line, NULL, 0);
9293 		if (rack->rc_gp_filled &&
9294 		    ((tp->gput_ack - tp->gput_seq) <
9295 		     max(rc_init_window(rack), (MIN_GP_WIN *
9296 						ctf_fixed_maxseg(tp))))) {
9297 			uint32_t ideal_amount;
9298 
9299 			ideal_amount = rack_get_measure_window(tp, rack);
9300 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9301 				/*
9302 				 * There is no sense of continuing this measurement
9303 				 * because its too small to gain us anything we
9304 				 * trust. Skip it and that way we can start a new
9305 				 * measurement quicker.
9306 				 */
9307 				tp->t_flags &= ~TF_GPUTINPROG;
9308 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9309 							   0, 0,
9310 							   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9311 							    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9312 							   6, __LINE__, NULL, 0);
9313 			} else {
9314 				/*
9315 				 * Reset the window further out.
9316 				 */
9317 				tp->gput_ack = tp->gput_seq + ideal_amount;
9318 			}
9319 		}
9320 		rack_tend_gp_marks(tp, rack);
9321 		rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9322 	}
9323 }
9324 
9325 static inline int
9326 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9327 {
9328 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9329 		/* Behind our TLP definition or right at */
9330 		return (0);
9331 	}
9332 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9333 		/* The start is beyond or right at our end of TLP definition */
9334 		return (0);
9335 	}
9336 	/* It has to be a sub-part of the original TLP recorded */
9337 	return (1);
9338 }
9339 
9340 static uint32_t
9341 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9342 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9343 		   uint32_t segsiz)
9344 {
9345 	uint32_t start, end, changed = 0;
9346 	struct rack_sendmap stack_map;
9347 	struct rack_sendmap *rsm, *nrsm, *prev, *next;
9348 	int insret __diagused;
9349 	int32_t used_ref = 1;
9350 	int can_use_hookery = 0;
9351 
9352 	start = sack->start;
9353 	end = sack->end;
9354 	rsm = *prsm;
9355 
9356 do_rest_ofb:
9357 	if ((rsm == NULL) ||
9358 	    (SEQ_LT(end, rsm->r_start)) ||
9359 	    (SEQ_GEQ(start, rsm->r_end)) ||
9360 	    (SEQ_LT(start, rsm->r_start))) {
9361 		/*
9362 		 * We are not in the right spot,
9363 		 * find the correct spot in the tree.
9364 		 */
9365 		used_ref = 0;
9366 		rsm = tqhash_find(rack->r_ctl.tqh, start);
9367 	}
9368 	if (rsm == NULL) {
9369 		/* TSNH */
9370 		goto out;
9371 	}
9372 	/* Ok we have an ACK for some piece of this rsm */
9373 	if (rsm->r_start != start) {
9374 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9375 			/*
9376 			 * Before any splitting or hookery is
9377 			 * done is it a TLP of interest i.e. rxt?
9378 			 */
9379 			if ((rsm->r_flags & RACK_TLP) &&
9380 			    (rsm->r_rtr_cnt > 1)) {
9381 				/*
9382 				 * We are splitting a rxt TLP, check
9383 				 * if we need to save off the start/end
9384 				 */
9385 				if (rack->rc_last_tlp_acked_set &&
9386 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9387 					/*
9388 					 * We already turned this on since we are inside
9389 					 * the previous one was a partially sack now we
9390 					 * are getting another one (maybe all of it).
9391 					 *
9392 					 */
9393 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9394 					/*
9395 					 * Lets make sure we have all of it though.
9396 					 */
9397 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9398 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9399 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9400 								     rack->r_ctl.last_tlp_acked_end);
9401 					}
9402 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9403 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9404 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9405 								     rack->r_ctl.last_tlp_acked_end);
9406 					}
9407 				} else {
9408 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9409 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9410 					rack->rc_last_tlp_past_cumack = 0;
9411 					rack->rc_last_tlp_acked_set = 1;
9412 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9413 				}
9414 			}
9415 			/**
9416 			 * Need to split this in two pieces the before and after,
9417 			 * the before remains in the map, the after must be
9418 			 * added. In other words we have:
9419 			 * rsm        |--------------|
9420 			 * sackblk        |------->
9421 			 * rsm will become
9422 			 *     rsm    |---|
9423 			 * and nrsm will be  the sacked piece
9424 			 *     nrsm       |----------|
9425 			 *
9426 			 * But before we start down that path lets
9427 			 * see if the sack spans over on top of
9428 			 * the next guy and it is already sacked.
9429 			 *
9430 			 */
9431 			/*
9432 			 * Hookery can only be used if the two entries
9433 			 * are in the same bucket and neither one of
9434 			 * them staddle the bucket line.
9435 			 */
9436 			next = tqhash_next(rack->r_ctl.tqh, rsm);
9437 			if (next &&
9438 			    (rsm->bindex == next->bindex) &&
9439 			    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9440 			    ((next->r_flags & RACK_STRADDLE) == 0) &&
9441 			    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9442 			    ((next->r_flags & RACK_IS_PCM) == 0) &&
9443 			    (rsm->r_flags & RACK_IN_GP_WIN) &&
9444 			    (next->r_flags & RACK_IN_GP_WIN))
9445 				can_use_hookery = 1;
9446 			else
9447 				can_use_hookery = 0;
9448 			if (next && can_use_hookery &&
9449 			    (next->r_flags & RACK_ACKED) &&
9450 			    SEQ_GEQ(end, next->r_start)) {
9451 				/**
9452 				 * So the next one is already acked, and
9453 				 * we can thus by hookery use our stack_map
9454 				 * to reflect the piece being sacked and
9455 				 * then adjust the two tree entries moving
9456 				 * the start and ends around. So we start like:
9457 				 *  rsm     |------------|             (not-acked)
9458 				 *  next                 |-----------| (acked)
9459 				 *  sackblk        |-------->
9460 				 *  We want to end like so:
9461 				 *  rsm     |------|                   (not-acked)
9462 				 *  next           |-----------------| (acked)
9463 				 *  nrsm           |-----|
9464 				 * Where nrsm is a temporary stack piece we
9465 				 * use to update all the gizmos.
9466 				 */
9467 				/* Copy up our fudge block */
9468 				nrsm = &stack_map;
9469 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9470 				/* Now adjust our tree blocks */
9471 				tqhash_update_end(rack->r_ctl.tqh, rsm, start);
9472 				next->r_start = start;
9473  				rsm->r_flags |= RACK_SHUFFLED;
9474 				next->r_flags |= RACK_SHUFFLED;
9475 				/* Now we must adjust back where next->m is */
9476 				rack_setup_offset_for_rsm(rack, rsm, next);
9477 				/*
9478 				 * Which timestamp do we keep? It is rather
9479 				 * important in GP measurements to have the
9480 				 * accurate end of the send window.
9481 				 *
9482 				 * We keep the largest value, which is the newest
9483 				 * send. We do this in case a segment that is
9484 				 * joined together and not part of a GP estimate
9485 				 * later gets expanded into the GP estimate.
9486 				 *
9487 				 * We prohibit the merging of unlike kinds i.e.
9488 				 * all pieces that are in the GP estimate can be
9489 				 * merged and all pieces that are not in a GP estimate
9490 				 * can be merged, but not disimilar pieces. Combine
9491 				 * this with taking the highest here and we should
9492 				 * be ok unless of course the client reneges. Then
9493 				 * all bets are off.
9494 				 */
9495 				if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9496 				    nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9497 					next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9498 				/*
9499 				 * And we must keep the newest ack arrival time.
9500 				 */
9501 				if (next->r_ack_arrival <
9502 				    rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9503 					next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9504 
9505 
9506 				/* We don't need to adjust rsm, it did not change */
9507 				/* Clear out the dup ack count of the remainder */
9508 				rsm->r_dupack = 0;
9509 				rsm->r_just_ret = 0;
9510 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9511 				/* Now lets make sure our fudge block is right */
9512 				nrsm->r_start = start;
9513 				/* Now lets update all the stats and such */
9514 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9515 				if (rack->app_limited_needs_set)
9516 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9517 				changed += (nrsm->r_end - nrsm->r_start);
9518 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9519 				if (rsm->r_flags & RACK_WAS_LOST) {
9520 					int my_chg;
9521 
9522 					my_chg = (nrsm->r_end - nrsm->r_start);
9523 					KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9524 						("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9525 					if (my_chg <= rack->r_ctl.rc_considered_lost)
9526 						rack->r_ctl.rc_considered_lost -= my_chg;
9527 					else
9528 						rack->r_ctl.rc_considered_lost = 0;
9529 				}
9530 				if (nrsm->r_flags & RACK_SACK_PASSED) {
9531 					rack->r_ctl.rc_reorder_ts = cts;
9532 					if (rack->r_ctl.rc_reorder_ts == 0)
9533 						rack->r_ctl.rc_reorder_ts = 1;
9534 				}
9535 				/*
9536 				 * Now we want to go up from rsm (the
9537 				 * one left un-acked) to the next one
9538 				 * in the tmap. We do this so when
9539 				 * we walk backwards we include marking
9540 				 * sack-passed on rsm (The one passed in
9541 				 * is skipped since it is generally called
9542 				 * on something sacked before removing it
9543 				 * from the tmap).
9544 				 */
9545 				if (rsm->r_in_tmap) {
9546 					nrsm = TAILQ_NEXT(rsm, r_tnext);
9547 					/*
9548 					 * Now that we have the next
9549 					 * one walk backwards from there.
9550 					 */
9551 					if (nrsm && nrsm->r_in_tmap)
9552 						rack_log_sack_passed(tp, rack, nrsm, cts);
9553 				}
9554 				/* Now are we done? */
9555 				if (SEQ_LT(end, next->r_end) ||
9556 				    (end == next->r_end)) {
9557 					/* Done with block */
9558 					goto out;
9559 				}
9560 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9561 				counter_u64_add(rack_sack_used_next_merge, 1);
9562 				/* Postion for the next block */
9563 				start = next->r_end;
9564 				rsm = tqhash_next(rack->r_ctl.tqh, next);
9565 				if (rsm == NULL)
9566 					goto out;
9567 			} else {
9568 				/**
9569 				 * We can't use any hookery here, so we
9570 				 * need to split the map. We enter like
9571 				 * so:
9572 				 *  rsm      |--------|
9573 				 *  sackblk       |----->
9574 				 * We will add the new block nrsm and
9575 				 * that will be the new portion, and then
9576 				 * fall through after reseting rsm. So we
9577 				 * split and look like this:
9578 				 *  rsm      |----|
9579 				 *  sackblk       |----->
9580 				 *  nrsm          |---|
9581 				 * We then fall through reseting
9582 				 * rsm to nrsm, so the next block
9583 				 * picks it up.
9584 				 */
9585 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9586 				if (nrsm == NULL) {
9587 					/*
9588 					 * failed XXXrrs what can we do but loose the sack
9589 					 * info?
9590 					 */
9591 					goto out;
9592 				}
9593 				counter_u64_add(rack_sack_splits, 1);
9594 				rack_clone_rsm(rack, nrsm, rsm, start);
9595 				rsm->r_just_ret = 0;
9596 #ifndef INVARIANTS
9597 				(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9598 #else
9599 				if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9600 					panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9601 					      nrsm, insret, rack, rsm);
9602 				}
9603 #endif
9604 				if (rsm->r_in_tmap) {
9605 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9606 					nrsm->r_in_tmap = 1;
9607 				}
9608 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9609 				rsm->r_flags &= (~RACK_HAS_FIN);
9610 				/* Position us to point to the new nrsm that starts the sack blk */
9611 				rsm = nrsm;
9612 			}
9613 		} else {
9614 			/* Already sacked this piece */
9615 			counter_u64_add(rack_sack_skipped_acked, 1);
9616 			if (end == rsm->r_end) {
9617 				/* Done with block */
9618 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9619 				goto out;
9620 			} else if (SEQ_LT(end, rsm->r_end)) {
9621 				/* A partial sack to a already sacked block */
9622 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9623 				goto out;
9624 			} else {
9625 				/*
9626 				 * The end goes beyond this guy
9627 				 * reposition the start to the
9628 				 * next block.
9629 				 */
9630 				start = rsm->r_end;
9631 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9632 				if (rsm == NULL)
9633 					goto out;
9634 			}
9635 		}
9636 	}
9637 	if (SEQ_GEQ(end, rsm->r_end)) {
9638 		/**
9639 		 * The end of this block is either beyond this guy or right
9640 		 * at this guy. I.e.:
9641 		 *  rsm ---                 |-----|
9642 		 *  end                     |-----|
9643 		 *  <or>
9644 		 *  end                     |---------|
9645 		 */
9646 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9647 			/*
9648 			 * Is it a TLP of interest?
9649 			 */
9650 			if ((rsm->r_flags & RACK_TLP) &&
9651 			    (rsm->r_rtr_cnt > 1)) {
9652 				/*
9653 				 * We are splitting a rxt TLP, check
9654 				 * if we need to save off the start/end
9655 				 */
9656 				if (rack->rc_last_tlp_acked_set &&
9657 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9658 					/*
9659 					 * We already turned this on since we are inside
9660 					 * the previous one was a partially sack now we
9661 					 * are getting another one (maybe all of it).
9662 					 */
9663 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9664 					/*
9665 					 * Lets make sure we have all of it though.
9666 					 */
9667 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9668 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9669 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9670 								     rack->r_ctl.last_tlp_acked_end);
9671 					}
9672 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9673 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9674 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9675 								     rack->r_ctl.last_tlp_acked_end);
9676 					}
9677 				} else {
9678 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9679 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9680 					rack->rc_last_tlp_past_cumack = 0;
9681 					rack->rc_last_tlp_acked_set = 1;
9682 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9683 				}
9684 			}
9685 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9686 			changed += (rsm->r_end - rsm->r_start);
9687 			/* You get a count for acking a whole segment or more */
9688 			if (rsm->r_flags & RACK_WAS_LOST) {
9689 				int my_chg;
9690 
9691 				my_chg = (rsm->r_end - rsm->r_start);
9692 				rsm->r_flags &= ~RACK_WAS_LOST;
9693 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9694 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9695 				if (my_chg <= rack->r_ctl.rc_considered_lost)
9696 					rack->r_ctl.rc_considered_lost -= my_chg;
9697 				else
9698 					rack->r_ctl.rc_considered_lost = 0;
9699 			}
9700 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9701 			if (rsm->r_in_tmap) /* should be true */
9702 				rack_log_sack_passed(tp, rack, rsm, cts);
9703 			/* Is Reordering occuring? */
9704 			if (rsm->r_flags & RACK_SACK_PASSED) {
9705 				rsm->r_flags &= ~RACK_SACK_PASSED;
9706 				rack->r_ctl.rc_reorder_ts = cts;
9707 				if (rack->r_ctl.rc_reorder_ts == 0)
9708 					rack->r_ctl.rc_reorder_ts = 1;
9709 			}
9710 			if (rack->app_limited_needs_set)
9711 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9712 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9713 			rsm->r_flags |= RACK_ACKED;
9714 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
9715 			if (rsm->r_in_tmap) {
9716 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9717 				rsm->r_in_tmap = 0;
9718 			}
9719 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9720 		} else {
9721 			counter_u64_add(rack_sack_skipped_acked, 1);
9722 		}
9723 		if (end == rsm->r_end) {
9724 			/* This block only - done, setup for next */
9725 			goto out;
9726 		}
9727 		/*
9728 		 * There is more not coverend by this rsm move on
9729 		 * to the next block in the tail queue hash table.
9730 		 */
9731 		nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9732 		start = rsm->r_end;
9733 		rsm = nrsm;
9734 		if (rsm == NULL)
9735 			goto out;
9736 		goto do_rest_ofb;
9737 	}
9738 	/**
9739 	 * The end of this sack block is smaller than
9740 	 * our rsm i.e.:
9741 	 *  rsm ---                 |-----|
9742 	 *  end                     |--|
9743 	 */
9744 	if ((rsm->r_flags & RACK_ACKED) == 0) {
9745 		/*
9746 		 * Is it a TLP of interest?
9747 		 */
9748 		if ((rsm->r_flags & RACK_TLP) &&
9749 		    (rsm->r_rtr_cnt > 1)) {
9750 			/*
9751 			 * We are splitting a rxt TLP, check
9752 			 * if we need to save off the start/end
9753 			 */
9754 			if (rack->rc_last_tlp_acked_set &&
9755 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9756 				/*
9757 				 * We already turned this on since we are inside
9758 				 * the previous one was a partially sack now we
9759 				 * are getting another one (maybe all of it).
9760 				 */
9761 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9762 				/*
9763 				 * Lets make sure we have all of it though.
9764 				 */
9765 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9766 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9767 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9768 							     rack->r_ctl.last_tlp_acked_end);
9769 				}
9770 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9771 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9772 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9773 							     rack->r_ctl.last_tlp_acked_end);
9774 				}
9775 			} else {
9776 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9777 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9778 				rack->rc_last_tlp_past_cumack = 0;
9779 				rack->rc_last_tlp_acked_set = 1;
9780 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9781 			}
9782 		}
9783 		/*
9784 		 * Hookery can only be used if the two entries
9785 		 * are in the same bucket and neither one of
9786 		 * them staddle the bucket line.
9787 		 */
9788 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9789 		if (prev &&
9790 		    (rsm->bindex == prev->bindex) &&
9791 		    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9792 		    ((prev->r_flags & RACK_STRADDLE) == 0) &&
9793 		    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9794 		    ((prev->r_flags & RACK_IS_PCM) == 0) &&
9795 		    (rsm->r_flags & RACK_IN_GP_WIN) &&
9796 		    (prev->r_flags & RACK_IN_GP_WIN))
9797 			can_use_hookery = 1;
9798 		else
9799 			can_use_hookery = 0;
9800 		if (prev && can_use_hookery &&
9801 		    (prev->r_flags & RACK_ACKED)) {
9802 			/**
9803 			 * Goal, we want the right remainder of rsm to shrink
9804 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
9805 			 * We want to expand prev to go all the way
9806 			 * to prev->r_end <- end.
9807 			 * so in the tree we have before:
9808 			 *   prev     |--------|         (acked)
9809 			 *   rsm               |-------| (non-acked)
9810 			 *   sackblk           |-|
9811 			 * We churn it so we end up with
9812 			 *   prev     |----------|       (acked)
9813 			 *   rsm                 |-----| (non-acked)
9814 			 *   nrsm              |-| (temporary)
9815 			 *
9816 			 * Note if either prev/rsm is a TLP we don't
9817 			 * do this.
9818 			 */
9819 			nrsm = &stack_map;
9820 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9821 			tqhash_update_end(rack->r_ctl.tqh, prev, end);
9822 			rsm->r_start = end;
9823 			rsm->r_flags |= RACK_SHUFFLED;
9824 			prev->r_flags |= RACK_SHUFFLED;
9825 			/* Now adjust nrsm (stack copy) to be
9826 			 * the one that is the small
9827 			 * piece that was "sacked".
9828 			 */
9829 			nrsm->r_end = end;
9830 			rsm->r_dupack = 0;
9831 			/*
9832 			 * Which timestamp do we keep? It is rather
9833 			 * important in GP measurements to have the
9834 			 * accurate end of the send window.
9835 			 *
9836 			 * We keep the largest value, which is the newest
9837 			 * send. We do this in case a segment that is
9838 			 * joined together and not part of a GP estimate
9839 			 * later gets expanded into the GP estimate.
9840 			 *
9841 			 * We prohibit the merging of unlike kinds i.e.
9842 			 * all pieces that are in the GP estimate can be
9843 			 * merged and all pieces that are not in a GP estimate
9844 			 * can be merged, but not disimilar pieces. Combine
9845 			 * this with taking the highest here and we should
9846 			 * be ok unless of course the client reneges. Then
9847 			 * all bets are off.
9848 			 */
9849 			if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
9850 			   nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
9851 				prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9852 			}
9853 			/*
9854 			 * And we must keep the newest ack arrival time.
9855 			 */
9856 
9857 			if(prev->r_ack_arrival <
9858 			   rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9859 				prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9860 
9861 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9862 			/*
9863 			 * Now that the rsm has had its start moved forward
9864 			 * lets go ahead and get its new place in the world.
9865 			 */
9866 			rack_setup_offset_for_rsm(rack, prev, rsm);
9867 			/*
9868 			 * Now nrsm is our new little piece
9869 			 * that is acked (which was merged
9870 			 * to prev). Update the rtt and changed
9871 			 * based on that. Also check for reordering.
9872 			 */
9873 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9874 			if (rack->app_limited_needs_set)
9875 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9876 			changed += (nrsm->r_end - nrsm->r_start);
9877 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9878 			if (rsm->r_flags & RACK_WAS_LOST) {
9879 				int my_chg;
9880 
9881 				my_chg = (nrsm->r_end - nrsm->r_start);
9882 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9883 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9884 				if (my_chg <= rack->r_ctl.rc_considered_lost)
9885 					rack->r_ctl.rc_considered_lost -= my_chg;
9886 				else
9887 					rack->r_ctl.rc_considered_lost = 0;
9888 			}
9889 			if (nrsm->r_flags & RACK_SACK_PASSED) {
9890 				rack->r_ctl.rc_reorder_ts = cts;
9891 				if (rack->r_ctl.rc_reorder_ts == 0)
9892 					rack->r_ctl.rc_reorder_ts = 1;
9893 			}
9894 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
9895 			rsm = prev;
9896 			counter_u64_add(rack_sack_used_prev_merge, 1);
9897 		} else {
9898 			/**
9899 			 * This is the case where our previous
9900 			 * block is not acked either, so we must
9901 			 * split the block in two.
9902 			 */
9903 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9904 			if (nrsm == NULL) {
9905 				/* failed rrs what can we do but loose the sack info? */
9906 				goto out;
9907 			}
9908 			if ((rsm->r_flags & RACK_TLP) &&
9909 			    (rsm->r_rtr_cnt > 1)) {
9910 				/*
9911 				 * We are splitting a rxt TLP, check
9912 				 * if we need to save off the start/end
9913 				 */
9914 				if (rack->rc_last_tlp_acked_set &&
9915 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9916 					/*
9917 					 * We already turned this on since this block is inside
9918 					 * the previous one was a partially sack now we
9919 					 * are getting another one (maybe all of it).
9920 					 */
9921 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9922 					/*
9923 					 * Lets make sure we have all of it though.
9924 					 */
9925 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9926 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9927 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9928 								     rack->r_ctl.last_tlp_acked_end);
9929 					}
9930 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9931 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9932 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9933 								     rack->r_ctl.last_tlp_acked_end);
9934 					}
9935 				} else {
9936 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9937 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9938 					rack->rc_last_tlp_acked_set = 1;
9939 					rack->rc_last_tlp_past_cumack = 0;
9940 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9941 				}
9942 			}
9943 			/**
9944 			 * In this case nrsm becomes
9945 			 * nrsm->r_start = end;
9946 			 * nrsm->r_end = rsm->r_end;
9947 			 * which is un-acked.
9948 			 * <and>
9949 			 * rsm->r_end = nrsm->r_start;
9950 			 * i.e. the remaining un-acked
9951 			 * piece is left on the left
9952 			 * hand side.
9953 			 *
9954 			 * So we start like this
9955 			 * rsm      |----------| (not acked)
9956 			 * sackblk  |---|
9957 			 * build it so we have
9958 			 * rsm      |---|         (acked)
9959 			 * nrsm         |------|  (not acked)
9960 			 */
9961 			counter_u64_add(rack_sack_splits, 1);
9962 			rack_clone_rsm(rack, nrsm, rsm, end);
9963 			rsm->r_flags &= (~RACK_HAS_FIN);
9964 			rsm->r_just_ret = 0;
9965 #ifndef INVARIANTS
9966 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9967 #else
9968 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9969 				panic("Insert in tailq_hash of %p fails ret:% rack:%p rsm:%p",
9970 				      nrsm, insret, rack, rsm);
9971 			}
9972 #endif
9973 			if (rsm->r_in_tmap) {
9974 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9975 				nrsm->r_in_tmap = 1;
9976 			}
9977 			nrsm->r_dupack = 0;
9978 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
9979 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9980 			changed += (rsm->r_end - rsm->r_start);
9981 			if (rsm->r_flags & RACK_WAS_LOST) {
9982 				int my_chg;
9983 
9984 				my_chg = (rsm->r_end - rsm->r_start);
9985 				rsm->r_flags &= ~RACK_WAS_LOST;
9986 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9987 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9988 				if (my_chg <= rack->r_ctl.rc_considered_lost)
9989 					rack->r_ctl.rc_considered_lost -= my_chg;
9990 				else
9991 					rack->r_ctl.rc_considered_lost = 0;
9992 			}
9993 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9994 
9995 			if (rsm->r_in_tmap) /* should be true */
9996 				rack_log_sack_passed(tp, rack, rsm, cts);
9997 			/* Is Reordering occuring? */
9998 			if (rsm->r_flags & RACK_SACK_PASSED) {
9999 				rsm->r_flags &= ~RACK_SACK_PASSED;
10000 				rack->r_ctl.rc_reorder_ts = cts;
10001 				if (rack->r_ctl.rc_reorder_ts == 0)
10002 					rack->r_ctl.rc_reorder_ts = 1;
10003 			}
10004 			if (rack->app_limited_needs_set)
10005 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10006 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10007 			rsm->r_flags |= RACK_ACKED;
10008 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
10009 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
10010 			if (rsm->r_in_tmap) {
10011 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10012 				rsm->r_in_tmap = 0;
10013 			}
10014 		}
10015 	} else if (start != end){
10016 		/*
10017 		 * The block was already acked.
10018 		 */
10019 		counter_u64_add(rack_sack_skipped_acked, 1);
10020 	}
10021 out:
10022 	if (rsm &&
10023 	    ((rsm->r_flags & RACK_TLP) == 0) &&
10024 	    (rsm->r_flags & RACK_ACKED)) {
10025 		/*
10026 		 * Now can we merge where we worked
10027 		 * with either the previous or
10028 		 * next block?
10029 		 */
10030 		next = tqhash_next(rack->r_ctl.tqh, rsm);
10031 		while (next) {
10032 			if (next->r_flags & RACK_TLP)
10033 				break;
10034 			/* Only allow merges between ones in or out of GP window */
10035 			if ((next->r_flags & RACK_IN_GP_WIN) &&
10036 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10037 				break;
10038 			}
10039 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10040 			    ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10041 				break;
10042 			}
10043 			if (rsm->bindex != next->bindex)
10044 				break;
10045 			if (rsm->r_flags & RACK_STRADDLE)
10046 				break;
10047 			if (rsm->r_flags & RACK_IS_PCM)
10048 				break;
10049 			if (next->r_flags & RACK_STRADDLE)
10050 				break;
10051 			if (next->r_flags & RACK_IS_PCM)
10052 				break;
10053 			if (next->r_flags & RACK_ACKED) {
10054 				/* yep this and next can be merged */
10055 				rsm = rack_merge_rsm(rack, rsm, next);
10056 				next = tqhash_next(rack->r_ctl.tqh, rsm);
10057 			} else
10058 				break;
10059 		}
10060 		/* Now what about the previous? */
10061 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10062 		while (prev) {
10063 			if (prev->r_flags & RACK_TLP)
10064 				break;
10065 			/* Only allow merges between ones in or out of GP window */
10066 			if ((prev->r_flags & RACK_IN_GP_WIN) &&
10067 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10068 				break;
10069 			}
10070 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10071 			    ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10072 				break;
10073 			}
10074 			if (rsm->bindex != prev->bindex)
10075 				break;
10076 			if (rsm->r_flags & RACK_STRADDLE)
10077 				break;
10078 			if (rsm->r_flags & RACK_IS_PCM)
10079 				break;
10080 			if (prev->r_flags & RACK_STRADDLE)
10081 				break;
10082 			if (prev->r_flags & RACK_IS_PCM)
10083 				break;
10084 			if (prev->r_flags & RACK_ACKED) {
10085 				/* yep the previous and this can be merged */
10086 				rsm = rack_merge_rsm(rack, prev, rsm);
10087 				prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10088 			} else
10089 				break;
10090 		}
10091 	}
10092 	if (used_ref == 0) {
10093 		counter_u64_add(rack_sack_proc_all, 1);
10094 	} else {
10095 		counter_u64_add(rack_sack_proc_short, 1);
10096 	}
10097 	/* Save off the next one for quick reference. */
10098 	nrsm = tqhash_find(rack->r_ctl.tqh, end);
10099 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
10100 	return (changed);
10101 }
10102 
10103 static void inline
10104 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10105 {
10106 	struct rack_sendmap *tmap;
10107 
10108 	tmap = NULL;
10109 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
10110 		/* Its no longer sacked, mark it so */
10111 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10112 #ifdef INVARIANTS
10113 		if (rsm->r_in_tmap) {
10114 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
10115 			      rack, rsm, rsm->r_flags);
10116 		}
10117 #endif
10118 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10119 		/* Rebuild it into our tmap */
10120 		if (tmap == NULL) {
10121 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10122 			tmap = rsm;
10123 		} else {
10124 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10125 			tmap = rsm;
10126 		}
10127 		tmap->r_in_tmap = 1;
10128 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10129 	}
10130 	/*
10131 	 * Now lets possibly clear the sack filter so we start
10132 	 * recognizing sacks that cover this area.
10133 	 */
10134 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10135 
10136 }
10137 
10138 
10139 static void inline
10140 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10141 {
10142 	/*
10143 	 * We look at advancing the end send time for our GP
10144 	 * measurement tracking only as the cumulative acknowledgment
10145 	 * moves forward. You might wonder about this, why not
10146 	 * at every transmission or retransmission within the
10147 	 * GP window update the rc_gp_cumack_ts? Well its rather
10148 	 * nuanced but basically the GP window *may* expand (as
10149 	 * it does below) or worse and harder to track it may shrink.
10150 	 *
10151 	 * This last makes it impossible to track at the time of
10152 	 * the send, since you may set forward your rc_gp_cumack_ts
10153 	 * when you send, because that send *is* in your currently
10154 	 * "guessed" window, but then it shrinks. Now which was
10155 	 * the send time of the last bytes in the window, by the
10156 	 * time you ask that question that part of the sendmap
10157 	 * is freed. So you don't know and you will have too
10158 	 * long of send window. Instead by updating the time
10159 	 * marker only when the cumack advances this assures us
10160 	 * that we will have only the sends in the window of our
10161 	 * GP measurement.
10162 	 *
10163 	 * Another complication from this is the
10164 	 * merging of sendmap entries. During SACK processing this
10165 	 * can happen to conserve the sendmap size. That breaks
10166 	 * everything down in tracking the send window of the GP
10167 	 * estimate. So to prevent that and keep it working with
10168 	 * a tiny bit more limited merging, we only allow like
10169 	 * types to be merged. I.e. if two sends are in the GP window
10170 	 * then its ok to merge them together. If two sends are not
10171 	 * in the GP window its ok to merge them together too. Though
10172 	 * one send in and one send out cannot be merged. We combine
10173 	 * this with never allowing the shrinking of the GP window when
10174 	 * we are in recovery so that we can properly calculate the
10175 	 * sending times.
10176 	 *
10177 	 * This all of course seems complicated, because it is.. :)
10178 	 *
10179 	 * The cum-ack is being advanced upon the sendmap.
10180 	 * If we are not doing a GP estimate don't
10181 	 * proceed.
10182 	 */
10183 	uint64_t ts;
10184 
10185 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
10186 		return;
10187 	/*
10188 	 * If this sendmap entry is going
10189 	 * beyond the measurement window we had picked,
10190 	 * expand the measurement window by that much.
10191 	 */
10192 	if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10193 		tp->gput_ack = rsm->r_end;
10194 	}
10195 	/*
10196 	 * If we have not setup a ack, then we
10197 	 * have no idea if the newly acked pieces
10198 	 * will be "in our seq measurement range". If
10199 	 * it is when we clear the app_limited_needs_set
10200 	 * flag the timestamp will be updated.
10201 	 */
10202 	if (rack->app_limited_needs_set)
10203 		return;
10204 	/*
10205 	 * Finally, we grab out the latest timestamp
10206 	 * that this packet was sent and then see
10207 	 * if:
10208 	 *  a) The packet touches are newly defined GP range.
10209 	 *  b) The time is greater than (newer) than the
10210 	 *     one we currently have. If so we update
10211 	 *     our sending end time window.
10212 	 *
10213 	 * Note we *do not* do this at send time. The reason
10214 	 * is that if you do you *may* pick up a newer timestamp
10215 	 * for a range you are not going to measure. We project
10216 	 * out how far and then sometimes modify that to be
10217 	 * smaller. If that occurs then you will have a send
10218 	 * that does not belong to the range included.
10219 	 */
10220 	if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10221 	    rack->r_ctl.rc_gp_cumack_ts)
10222 		return;
10223 	if (rack_in_gp_window(tp, rsm)) {
10224 		rack->r_ctl.rc_gp_cumack_ts = ts;
10225 		rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10226 			       __LINE__, from, rsm);
10227 	}
10228 }
10229 
10230 static void
10231 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)
10232 {
10233 	struct rack_sendmap *rsm;
10234 	/*
10235 	 * The ACK point is advancing to th_ack, we must drop off
10236 	 * the packets in the rack log and calculate any eligble
10237 	 * RTT's.
10238 	 */
10239 
10240 	if (sack_filter_blks_used(&rack->r_ctl.rack_sf)) {
10241 		/*
10242 		 * If we have some sack blocks in the filter
10243 		 * lets prune them out by calling sfb with no blocks.
10244 		 */
10245 		sack_filter_blks(tp, &rack->r_ctl.rack_sf, NULL, 0, th_ack);
10246 	}
10247 	if (SEQ_GT(th_ack, tp->snd_una)) {
10248 		/* Clear any app ack remembered settings */
10249 		rack->r_ctl.cleared_app_ack = 0;
10250 	}
10251 	rack->r_wanted_output = 1;
10252 	if (SEQ_GT(th_ack, tp->snd_una))
10253 		rack->r_ctl.last_cumack_advance = acktime;
10254 
10255 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
10256 	if ((rack->rc_last_tlp_acked_set == 1)&&
10257 	    (rack->rc_last_tlp_past_cumack == 1) &&
10258 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10259 		/*
10260 		 * We have reached the point where our last rack
10261 		 * tlp retransmit sequence is ahead of the cum-ack.
10262 		 * This can only happen when the cum-ack moves all
10263 		 * the way around (its been a full 2^^31+1 bytes
10264 		 * or more since we sent a retransmitted TLP). Lets
10265 		 * turn off the valid flag since its not really valid.
10266 		 *
10267 		 * Note since sack's also turn on this event we have
10268 		 * a complication, we have to wait to age it out until
10269 		 * the cum-ack is by the TLP before checking which is
10270 		 * what the next else clause does.
10271 		 */
10272 		rack_log_dsack_event(rack, 9, __LINE__,
10273 				     rack->r_ctl.last_tlp_acked_start,
10274 				     rack->r_ctl.last_tlp_acked_end);
10275 		rack->rc_last_tlp_acked_set = 0;
10276 		rack->rc_last_tlp_past_cumack = 0;
10277 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
10278 		   (rack->rc_last_tlp_past_cumack == 0) &&
10279 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10280 		/*
10281 		 * It is safe to start aging TLP's out.
10282 		 */
10283 		rack->rc_last_tlp_past_cumack = 1;
10284 	}
10285 	/* We do the same for the tlp send seq as well */
10286 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10287 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
10288 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
10289 		rack_log_dsack_event(rack, 9, __LINE__,
10290 				     rack->r_ctl.last_sent_tlp_seq,
10291 				     (rack->r_ctl.last_sent_tlp_seq +
10292 				      rack->r_ctl.last_sent_tlp_len));
10293 		rack->rc_last_sent_tlp_seq_valid = 0;
10294 		rack->rc_last_sent_tlp_past_cumack = 0;
10295 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10296 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
10297 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10298 		/*
10299 		 * It is safe to start aging TLP's send.
10300 		 */
10301 		rack->rc_last_sent_tlp_past_cumack = 1;
10302 	}
10303 more:
10304 	rsm = tqhash_min(rack->r_ctl.tqh);
10305 	if (rsm == NULL) {
10306 		if ((th_ack - 1) == tp->iss) {
10307 			/*
10308 			 * For the SYN incoming case we will not
10309 			 * have called tcp_output for the sending of
10310 			 * the SYN, so there will be no map. All
10311 			 * other cases should probably be a panic.
10312 			 */
10313 			return;
10314 		}
10315 		if (tp->t_flags & TF_SENTFIN) {
10316 			/* if we sent a FIN we often will not have map */
10317 			return;
10318 		}
10319 #ifdef INVARIANTS
10320 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u\n",
10321 		      tp,
10322 		      tp->t_state, th_ack, rack,
10323 		      tp->snd_una, tp->snd_max);
10324 #endif
10325 		return;
10326 	}
10327 	if (SEQ_LT(th_ack, rsm->r_start)) {
10328 		/* Huh map is missing this */
10329 #ifdef INVARIANTS
10330 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10331 		       rsm->r_start,
10332 		       th_ack, tp->t_state, rack->r_state);
10333 #endif
10334 		return;
10335 	}
10336 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10337 
10338 	/* Now was it a retransmitted TLP? */
10339 	if ((rsm->r_flags & RACK_TLP) &&
10340 	    (rsm->r_rtr_cnt > 1)) {
10341 		/*
10342 		 * Yes, this rsm was a TLP and retransmitted, remember that
10343 		 * since if a DSACK comes back on this we don't want
10344 		 * to think of it as a reordered segment. This may
10345 		 * get updated again with possibly even other TLPs
10346 		 * in flight, but thats ok. Only when we don't send
10347 		 * a retransmitted TLP for 1/2 the sequences space
10348 		 * will it get turned off (above).
10349 		 */
10350 		if (rack->rc_last_tlp_acked_set &&
10351 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10352 			/*
10353 			 * We already turned this on since the end matches,
10354 			 * the previous one was a partially ack now we
10355 			 * are getting another one (maybe all of it).
10356 			 */
10357 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10358 			/*
10359 			 * Lets make sure we have all of it though.
10360 			 */
10361 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10362 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10363 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10364 						     rack->r_ctl.last_tlp_acked_end);
10365 			}
10366 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10367 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10368 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10369 						     rack->r_ctl.last_tlp_acked_end);
10370 			}
10371 		} else {
10372 			rack->rc_last_tlp_past_cumack = 1;
10373 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10374 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10375 			rack->rc_last_tlp_acked_set = 1;
10376 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10377 		}
10378 	}
10379 	/* Now do we consume the whole thing? */
10380 	rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
10381 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
10382 		/* Its all consumed. */
10383 		uint32_t left;
10384 		uint8_t newly_acked;
10385 
10386 		if (rsm->r_flags & RACK_WAS_LOST) {
10387 			/*
10388 			 * This can happen when we marked it as lost
10389 			 * and yet before retransmitting we get an ack
10390 			 * which can happen due to reordering.
10391 			 */
10392 			rsm->r_flags  &= ~RACK_WAS_LOST;
10393 			KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
10394 				("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10395 			if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
10396 				rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
10397 			else
10398 				rack->r_ctl.rc_considered_lost = 0;
10399 		}
10400 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10401 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10402 		rsm->r_rtr_bytes = 0;
10403 		/*
10404 		 * Record the time of highest cumack sent if its in our measurement
10405 		 * window and possibly bump out the end.
10406 		 */
10407 		rack_rsm_sender_update(rack, tp, rsm, 4);
10408 		tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10409 		if (rsm->r_in_tmap) {
10410 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10411 			rsm->r_in_tmap = 0;
10412 		}
10413 		newly_acked = 1;
10414 		if (rsm->r_flags & RACK_ACKED) {
10415 			/*
10416 			 * It was acked on the scoreboard -- remove
10417 			 * it from total
10418 			 */
10419 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10420 			newly_acked = 0;
10421 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
10422 			/*
10423 			 * There are segments ACKED on the
10424 			 * scoreboard further up. We are seeing
10425 			 * reordering.
10426 			 */
10427 			rsm->r_flags &= ~RACK_SACK_PASSED;
10428 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10429 			rsm->r_flags |= RACK_ACKED;
10430 			rack->r_ctl.rc_reorder_ts = cts;
10431 			if (rack->r_ctl.rc_reorder_ts == 0)
10432 				rack->r_ctl.rc_reorder_ts = 1;
10433 			if (rack->r_ent_rec_ns) {
10434 				/*
10435 				 * We have sent no more, and we saw an sack
10436 				 * then ack arrive.
10437 				 */
10438 				rack->r_might_revert = 1;
10439 			}
10440 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10441 		} else {
10442 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10443 		}
10444 		if ((rsm->r_flags & RACK_TO_REXT) &&
10445 		    (tp->t_flags & TF_RCVD_TSTMP) &&
10446 		    (to->to_flags & TOF_TS) &&
10447 		    (to->to_tsecr != 0) &&
10448 		    (tp->t_flags & TF_PREVVALID)) {
10449 			/*
10450 			 * We can use the timestamp to see
10451 			 * if this retransmission was from the
10452 			 * first transmit. If so we made a mistake.
10453 			 */
10454 			tp->t_flags &= ~TF_PREVVALID;
10455 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10456 				/* The first transmit is what this ack is for */
10457 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10458 			}
10459 		}
10460 		left = th_ack - rsm->r_end;
10461 		if (rack->app_limited_needs_set && newly_acked)
10462 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10463 		/* Free back to zone */
10464 		rack_free(rack, rsm);
10465 		if (left) {
10466 			goto more;
10467 		}
10468 		/* Check for reneging */
10469 		rsm = tqhash_min(rack->r_ctl.tqh);
10470 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10471 			/*
10472 			 * The peer has moved snd_una up to
10473 			 * the edge of this send, i.e. one
10474 			 * that it had previously acked. The only
10475 			 * way that can be true if the peer threw
10476 			 * away data (space issues) that it had
10477 			 * previously sacked (else it would have
10478 			 * given us snd_una up to (rsm->r_end).
10479 			 * We need to undo the acked markings here.
10480 			 *
10481 			 * Note we have to look to make sure th_ack is
10482 			 * our rsm->r_start in case we get an old ack
10483 			 * where th_ack is behind snd_una.
10484 			 */
10485 			rack_peer_reneges(rack, rsm, th_ack);
10486 		}
10487 		return;
10488 	}
10489 	if (rsm->r_flags & RACK_ACKED) {
10490 		/*
10491 		 * It was acked on the scoreboard -- remove it from
10492 		 * total for the part being cum-acked.
10493 		 */
10494 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10495 	} else {
10496 		rack_update_pcm_ack(rack, 1, rsm->r_start, th_ack);
10497 	}
10498 	/* And what about the lost flag? */
10499 	if (rsm->r_flags & RACK_WAS_LOST) {
10500 		/*
10501 		 * This can happen when we marked it as lost
10502 		 * and yet before retransmitting we get an ack
10503 		 * which can happen due to reordering. In this
10504 		 * case its only a partial ack of the send.
10505 		 */
10506 		KASSERT((rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start)),
10507 			("rsm:%p rack:%p rc_considered_lost goes negative th_ack:%u", rsm,  rack, th_ack));
10508 		if (rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start))
10509 			rack->r_ctl.rc_considered_lost -= th_ack - rsm->r_start;
10510 		else
10511 			rack->r_ctl.rc_considered_lost = 0;
10512 	}
10513 	/*
10514 	 * Clear the dup ack count for
10515 	 * the piece that remains.
10516 	 */
10517 	rsm->r_dupack = 0;
10518 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10519 	if (rsm->r_rtr_bytes) {
10520 		/*
10521 		 * It was retransmitted adjust the
10522 		 * sack holes for what was acked.
10523 		 */
10524 		int ack_am;
10525 
10526 		ack_am = (th_ack - rsm->r_start);
10527 		if (ack_am >= rsm->r_rtr_bytes) {
10528 			rack->r_ctl.rc_holes_rxt -= ack_am;
10529 			rsm->r_rtr_bytes -= ack_am;
10530 		}
10531 	}
10532 	/*
10533 	 * Update where the piece starts and record
10534 	 * the time of send of highest cumack sent if
10535 	 * its in our GP range.
10536 	 */
10537 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10538 	/* Now we need to move our offset forward too */
10539 	if (rsm->m &&
10540 	    ((rsm->orig_m_len != rsm->m->m_len) ||
10541 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10542 		/* Fix up the orig_m_len and possibly the mbuf offset */
10543 		rack_adjust_orig_mlen(rsm);
10544 	}
10545 	rsm->soff += (th_ack - rsm->r_start);
10546 	rack_rsm_sender_update(rack, tp, rsm, 5);
10547 	/* The trim will move th_ack into r_start for us */
10548 	tqhash_trim(rack->r_ctl.tqh, th_ack);
10549 	/* Now do we need to move the mbuf fwd too? */
10550 	{
10551 		struct mbuf *m;
10552 		uint32_t soff;
10553 
10554 		m = rsm->m;
10555 		soff = rsm->soff;
10556 		if (m) {
10557 			while (soff >= m->m_len) {
10558 				soff -= m->m_len;
10559 				KASSERT((m->m_next != NULL),
10560 					(" rsm:%p  off:%u soff:%u m:%p",
10561 					 rsm, rsm->soff, soff, m));
10562 				m = m->m_next;
10563 				if (m == NULL) {
10564 					/*
10565 					 * This is a fall-back that prevents a panic. In reality
10566 					 * we should be able to walk the mbuf's and find our place.
10567 					 * At this point snd_una has not been updated with the sbcut() yet
10568 					 * but tqhash_trim did update rsm->r_start so the offset calcuation
10569 					 * should work fine. This is undesirable since we will take cache
10570 					 * hits to access the socket buffer. And even more puzzling is that
10571 					 * it happens occasionally. It should not :(
10572 					 */
10573 					m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
10574 						      (rsm->r_start - tp->snd_una),
10575 						      &soff);
10576 					break;
10577 				}
10578 			}
10579 			/*
10580 			 * Now save in our updated values.
10581 			 */
10582 			rsm->m = m;
10583 			rsm->soff = soff;
10584 			rsm->orig_m_len = rsm->m->m_len;
10585 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10586 		}
10587 	}
10588 	if (rack->app_limited_needs_set &&
10589 	    SEQ_GEQ(th_ack, tp->gput_seq))
10590 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10591 }
10592 
10593 static void
10594 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10595 {
10596 	struct rack_sendmap *rsm;
10597 	int sack_pass_fnd = 0;
10598 
10599 	if (rack->r_might_revert) {
10600 		/*
10601 		 * Ok we have reordering, have not sent anything, we
10602 		 * might want to revert the congestion state if nothing
10603 		 * further has SACK_PASSED on it. Lets check.
10604 		 *
10605 		 * We also get here when we have DSACKs come in for
10606 		 * all the data that we FR'd. Note that a rxt or tlp
10607 		 * timer clears this from happening.
10608 		 */
10609 
10610 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10611 			if (rsm->r_flags & RACK_SACK_PASSED) {
10612 				sack_pass_fnd = 1;
10613 				break;
10614 			}
10615 		}
10616 		if (sack_pass_fnd == 0) {
10617 			/*
10618 			 * We went into recovery
10619 			 * incorrectly due to reordering!
10620 			 */
10621 			int orig_cwnd;
10622 
10623 			rack->r_ent_rec_ns = 0;
10624 			orig_cwnd = tp->snd_cwnd;
10625 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10626 			tp->snd_recover = tp->snd_una;
10627 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10628 			if (IN_RECOVERY(tp->t_flags)) {
10629 				rack_exit_recovery(tp, rack, 3);
10630 				if ((rack->rto_from_rec == 1) && (rack_ssthresh_rest_rto_rec != 0) ){
10631 					/*
10632 					 * We were in recovery, had an RTO
10633 					 * and then re-entered recovery (more sack's arrived)
10634 					 * and we have properly recorded the old ssthresh from
10635 					 * the first recovery. We want to be able to slow-start
10636 					 * back to this level. The ssthresh from the timeout
10637 					 * and then back into recovery will end up most likely
10638 					 * to be min(cwnd=1mss, 2mss). Which makes it basically
10639 					 * so we get no slow-start after our RTO.
10640 					 */
10641 					rack->rto_from_rec = 0;
10642 					if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
10643 						tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
10644 				}
10645 			}
10646 		}
10647 		rack->r_might_revert = 0;
10648 	}
10649 }
10650 
10651 
10652 static int
10653 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10654 {
10655 
10656 	uint32_t am, l_end;
10657 	int was_tlp = 0;
10658 
10659 	if (SEQ_GT(end, start))
10660 		am = end - start;
10661 	else
10662 		am = 0;
10663 	if ((rack->rc_last_tlp_acked_set ) &&
10664 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
10665 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
10666 		/*
10667 		 * The DSACK is because of a TLP which we don't
10668 		 * do anything with the reordering window over since
10669 		 * it was not reordering that caused the DSACK but
10670 		 * our previous retransmit TLP.
10671 		 */
10672 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
10673 		was_tlp = 1;
10674 		goto skip_dsack_round;
10675 	}
10676 	if (rack->rc_last_sent_tlp_seq_valid) {
10677 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
10678 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
10679 		    (SEQ_LEQ(end, l_end))) {
10680 			/*
10681 			 * This dsack is from the last sent TLP, ignore it
10682 			 * for reordering purposes.
10683 			 */
10684 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
10685 			was_tlp = 1;
10686 			goto skip_dsack_round;
10687 		}
10688 	}
10689 	if (rack->rc_dsack_round_seen == 0) {
10690 		rack->rc_dsack_round_seen = 1;
10691 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
10692 		rack->r_ctl.num_dsack++;
10693 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
10694 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
10695 	}
10696 skip_dsack_round:
10697 	/*
10698 	 * We keep track of how many DSACK blocks we get
10699 	 * after a recovery incident.
10700 	 */
10701 	rack->r_ctl.dsack_byte_cnt += am;
10702 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
10703 	    rack->r_ctl.retran_during_recovery &&
10704 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
10705 		/*
10706 		 * False recovery most likely culprit is reordering. If
10707 		 * nothing else is missing we need to revert.
10708 		 */
10709 		rack->r_might_revert = 1;
10710 		rack_handle_might_revert(rack->rc_tp, rack);
10711 		rack->r_might_revert = 0;
10712 		rack->r_ctl.retran_during_recovery = 0;
10713 		rack->r_ctl.dsack_byte_cnt = 0;
10714 	}
10715 	return (was_tlp);
10716 }
10717 
10718 static uint32_t
10719 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
10720 {
10721 	return (((tp->snd_max - snd_una) -
10722 		 (rack->r_ctl.rc_sacked + rack->r_ctl.rc_considered_lost)) + rack->r_ctl.rc_holes_rxt);
10723 }
10724 
10725 static int32_t
10726 rack_compute_pipe(struct tcpcb *tp)
10727 {
10728 	return ((int32_t)do_rack_compute_pipe(tp,
10729 					      (struct tcp_rack *)tp->t_fb_ptr,
10730 					      tp->snd_una));
10731 }
10732 
10733 static void
10734 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
10735 {
10736 	/* Deal with changed and PRR here (in recovery only) */
10737 	uint32_t pipe, snd_una;
10738 
10739 	rack->r_ctl.rc_prr_delivered += changed;
10740 
10741 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
10742 		/*
10743 		 * It is all outstanding, we are application limited
10744 		 * and thus we don't need more room to send anything.
10745 		 * Note we use tp->snd_una here and not th_ack because
10746 		 * the data as yet not been cut from the sb.
10747 		 */
10748 		rack->r_ctl.rc_prr_sndcnt = 0;
10749 		return;
10750 	}
10751 	/* Compute prr_sndcnt */
10752 	if (SEQ_GT(tp->snd_una, th_ack)) {
10753 		snd_una = tp->snd_una;
10754 	} else {
10755 		snd_una = th_ack;
10756 	}
10757 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
10758 	if (pipe > tp->snd_ssthresh) {
10759 		long sndcnt;
10760 
10761 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
10762 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
10763 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
10764 		else {
10765 			rack->r_ctl.rc_prr_sndcnt = 0;
10766 			rack_log_to_prr(rack, 9, 0, __LINE__);
10767 			sndcnt = 0;
10768 		}
10769 		sndcnt++;
10770 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
10771 			sndcnt -= rack->r_ctl.rc_prr_out;
10772 		else
10773 			sndcnt = 0;
10774 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
10775 		rack_log_to_prr(rack, 10, 0, __LINE__);
10776 	} else {
10777 		uint32_t limit;
10778 
10779 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
10780 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
10781 		else
10782 			limit = 0;
10783 		if (changed > limit)
10784 			limit = changed;
10785 		limit += ctf_fixed_maxseg(tp);
10786 		if (tp->snd_ssthresh > pipe) {
10787 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
10788 			rack_log_to_prr(rack, 11, 0, __LINE__);
10789 		} else {
10790 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
10791 			rack_log_to_prr(rack, 12, 0, __LINE__);
10792 		}
10793 	}
10794 }
10795 
10796 static void
10797 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
10798 	     int *dsack_seen, int *sacks_seen)
10799 {
10800 	uint32_t changed;
10801 	struct tcp_rack *rack;
10802 	struct rack_sendmap *rsm;
10803 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
10804 	register uint32_t th_ack;
10805 	int32_t i, j, k, num_sack_blks = 0;
10806 	uint32_t cts, acked, ack_point;
10807 	int loop_start = 0;
10808 	uint32_t tsused;
10809 	uint32_t segsiz;
10810 
10811 
10812 	INP_WLOCK_ASSERT(tptoinpcb(tp));
10813 	if (tcp_get_flags(th) & TH_RST) {
10814 		/* We don't log resets */
10815 		return;
10816 	}
10817 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10818 	cts = tcp_get_usecs(NULL);
10819 	rsm = tqhash_min(rack->r_ctl.tqh);
10820 	changed = 0;
10821 	th_ack = th->th_ack;
10822 	segsiz = ctf_fixed_maxseg(rack->rc_tp);
10823 	if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
10824 		/*
10825 		 * You only get credit for
10826 		 * MSS and greater (and you get extra
10827 		 * credit for larger cum-ack moves).
10828 		 */
10829 		int ac;
10830 
10831 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
10832 		counter_u64_add(rack_ack_total, ac);
10833 	}
10834 	if (SEQ_GT(th_ack, tp->snd_una)) {
10835 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
10836 		tp->t_acktime = ticks;
10837 	}
10838 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
10839 		changed = th_ack - rsm->r_start;
10840 	if (changed) {
10841 		rack_process_to_cumack(tp, rack, th_ack, cts, to,
10842 				       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
10843 	}
10844 	if ((to->to_flags & TOF_SACK) == 0) {
10845 		/* We are done nothing left and no sack. */
10846 		rack_handle_might_revert(tp, rack);
10847 		/*
10848 		 * For cases where we struck a dup-ack
10849 		 * with no SACK, add to the changes so
10850 		 * PRR will work right.
10851 		 */
10852 		if (dup_ack_struck && (changed == 0)) {
10853 			changed += ctf_fixed_maxseg(rack->rc_tp);
10854 		}
10855 		goto out;
10856 	}
10857 	/* Sack block processing */
10858 	if (SEQ_GT(th_ack, tp->snd_una))
10859 		ack_point = th_ack;
10860 	else
10861 		ack_point = tp->snd_una;
10862 	for (i = 0; i < to->to_nsacks; i++) {
10863 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
10864 		      &sack, sizeof(sack));
10865 		sack.start = ntohl(sack.start);
10866 		sack.end = ntohl(sack.end);
10867 		if (SEQ_GT(sack.end, sack.start) &&
10868 		    SEQ_GT(sack.start, ack_point) &&
10869 		    SEQ_LT(sack.start, tp->snd_max) &&
10870 		    SEQ_GT(sack.end, ack_point) &&
10871 		    SEQ_LEQ(sack.end, tp->snd_max)) {
10872 			sack_blocks[num_sack_blks] = sack;
10873 			num_sack_blks++;
10874 		} else if (SEQ_LEQ(sack.start, th_ack) &&
10875 			   SEQ_LEQ(sack.end, th_ack)) {
10876 			int was_tlp;
10877 
10878 			if (dsack_seen != NULL)
10879 				*dsack_seen = 1;
10880 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
10881 			/*
10882 			 * Its a D-SACK block.
10883 			 */
10884 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
10885 		}
10886 	}
10887 	if (rack->rc_dsack_round_seen) {
10888 		/* Is the dsack roound over? */
10889 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
10890 			/* Yes it is */
10891 			rack->rc_dsack_round_seen = 0;
10892 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
10893 		}
10894 	}
10895 	/*
10896 	 * Sort the SACK blocks so we can update the rack scoreboard with
10897 	 * just one pass.
10898 	 */
10899 	num_sack_blks = sack_filter_blks(tp, &rack->r_ctl.rack_sf, sack_blocks,
10900 					 num_sack_blks, th->th_ack);
10901 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
10902 	if (sacks_seen != NULL)
10903 		*sacks_seen = num_sack_blks;
10904 	if (num_sack_blks == 0) {
10905 		/* Nothing to sack, but we need to update counts */
10906 		goto out_with_totals;
10907 	}
10908 	/* Its a sack of some sort */
10909 	if (num_sack_blks < 2) {
10910 		/* Only one, we don't need to sort */
10911 		goto do_sack_work;
10912 	}
10913 	/* Sort the sacks */
10914 	for (i = 0; i < num_sack_blks; i++) {
10915 		for (j = i + 1; j < num_sack_blks; j++) {
10916 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
10917 				sack = sack_blocks[i];
10918 				sack_blocks[i] = sack_blocks[j];
10919 				sack_blocks[j] = sack;
10920 			}
10921 		}
10922 	}
10923 	/*
10924 	 * Now are any of the sack block ends the same (yes some
10925 	 * implementations send these)?
10926 	 */
10927 again:
10928 	if (num_sack_blks == 0)
10929 		goto out_with_totals;
10930 	if (num_sack_blks > 1) {
10931 		for (i = 0; i < num_sack_blks; i++) {
10932 			for (j = i + 1; j < num_sack_blks; j++) {
10933 				if (sack_blocks[i].end == sack_blocks[j].end) {
10934 					/*
10935 					 * Ok these two have the same end we
10936 					 * want the smallest end and then
10937 					 * throw away the larger and start
10938 					 * again.
10939 					 */
10940 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
10941 						/*
10942 						 * The second block covers
10943 						 * more area use that
10944 						 */
10945 						sack_blocks[i].start = sack_blocks[j].start;
10946 					}
10947 					/*
10948 					 * Now collapse out the dup-sack and
10949 					 * lower the count
10950 					 */
10951 					for (k = (j + 1); k < num_sack_blks; k++) {
10952 						sack_blocks[j].start = sack_blocks[k].start;
10953 						sack_blocks[j].end = sack_blocks[k].end;
10954 						j++;
10955 					}
10956 					num_sack_blks--;
10957 					goto again;
10958 				}
10959 			}
10960 		}
10961 	}
10962 do_sack_work:
10963 	/*
10964 	 * First lets look to see if
10965 	 * we have retransmitted and
10966 	 * can use the transmit next?
10967 	 */
10968 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
10969 	if (rsm &&
10970 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
10971 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
10972 		/*
10973 		 * We probably did the FR and the next
10974 		 * SACK in continues as we would expect.
10975 		 */
10976 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, segsiz);
10977 		if (acked) {
10978 			rack->r_wanted_output = 1;
10979 			changed += acked;
10980 		}
10981 		if (num_sack_blks == 1) {
10982 			/*
10983 			 * This is what we would expect from
10984 			 * a normal implementation to happen
10985 			 * after we have retransmitted the FR,
10986 			 * i.e the sack-filter pushes down
10987 			 * to 1 block and the next to be retransmitted
10988 			 * is the sequence in the sack block (has more
10989 			 * are acked). Count this as ACK'd data to boost
10990 			 * up the chances of recovering any false positives.
10991 			 */
10992 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
10993 			counter_u64_add(rack_express_sack, 1);
10994 			goto out_with_totals;
10995 		} else {
10996 			/*
10997 			 * Start the loop through the
10998 			 * rest of blocks, past the first block.
10999 			 */
11000 			loop_start = 1;
11001 		}
11002 	}
11003 	counter_u64_add(rack_sack_total, 1);
11004 	rsm = rack->r_ctl.rc_sacklast;
11005 	for (i = loop_start; i < num_sack_blks; i++) {
11006 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts,  segsiz);
11007 		if (acked) {
11008 			rack->r_wanted_output = 1;
11009 			changed += acked;
11010 		}
11011 	}
11012 out_with_totals:
11013 	if (num_sack_blks > 1) {
11014 		/*
11015 		 * You get an extra stroke if
11016 		 * you have more than one sack-blk, this
11017 		 * could be where we are skipping forward
11018 		 * and the sack-filter is still working, or
11019 		 * it could be an attacker constantly
11020 		 * moving us.
11021 		 */
11022 		counter_u64_add(rack_move_some, 1);
11023 	}
11024 out:
11025 	if (changed) {
11026 		/* Something changed cancel the rack timer */
11027 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11028 	}
11029 	tsused = tcp_get_usecs(NULL);
11030 	rsm = tcp_rack_output(tp, rack, tsused);
11031 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11032 	    rsm &&
11033 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11034 		/* Enter recovery */
11035 		entered_recovery = 1;
11036 		rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
11037 		/*
11038 		 * When we enter recovery we need to assure we send
11039 		 * one packet.
11040 		 */
11041 		if (rack->rack_no_prr == 0) {
11042 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11043 			rack_log_to_prr(rack, 8, 0, __LINE__);
11044 		}
11045 		rack->r_timer_override = 1;
11046 		rack->r_early = 0;
11047 		rack->r_ctl.rc_agg_early = 0;
11048 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
11049 		   rsm &&
11050 		   (rack->r_rr_config == 3)) {
11051 		/*
11052 		 * Assure we can output and we get no
11053 		 * remembered pace time except the retransmit.
11054 		 */
11055 		rack->r_timer_override = 1;
11056 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11057 		rack->r_ctl.rc_resend = rsm;
11058 	}
11059 	if (IN_FASTRECOVERY(tp->t_flags) &&
11060 	    (rack->rack_no_prr == 0) &&
11061 	    (entered_recovery == 0)) {
11062 		rack_update_prr(tp, rack, changed, th_ack);
11063 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11064 		     ((tcp_in_hpts(rack->rc_tp) == 0) &&
11065 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11066 			/*
11067 			 * If you are pacing output you don't want
11068 			 * to override.
11069 			 */
11070 			rack->r_early = 0;
11071 			rack->r_ctl.rc_agg_early = 0;
11072 			rack->r_timer_override = 1;
11073 		}
11074 	}
11075 }
11076 
11077 static void
11078 rack_strike_dupack(struct tcp_rack *rack, tcp_seq th_ack)
11079 {
11080 	struct rack_sendmap *rsm;
11081 
11082 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11083 	while (rsm) {
11084 		/*
11085 		 * We need to skip anything already set
11086 		 * to be retransmitted.
11087 		 */
11088 		if ((rsm->r_dupack >= DUP_ACK_THRESHOLD)  ||
11089 		    (rsm->r_flags & RACK_MUST_RXT)) {
11090 			rsm = TAILQ_NEXT(rsm, r_tnext);
11091 			continue;
11092 		}
11093 		break;
11094 	}
11095 	if (rsm && (rsm->r_dupack < 0xff)) {
11096 		rsm->r_dupack++;
11097 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11098 			struct timeval tv;
11099 			uint32_t cts;
11100 			/*
11101 			 * Here we see if we need to retransmit. For
11102 			 * a SACK type connection if enough time has passed
11103 			 * we will get a return of the rsm. For a non-sack
11104 			 * connection we will get the rsm returned if the
11105 			 * dupack value is 3 or more.
11106 			 */
11107 			cts = tcp_get_usecs(&tv);
11108 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11109 			if (rack->r_ctl.rc_resend != NULL) {
11110 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11111 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11112 							 th_ack,  __LINE__);
11113 				}
11114 				rack->r_wanted_output = 1;
11115 				rack->r_timer_override = 1;
11116 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11117 			}
11118 		} else {
11119 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11120 		}
11121 	}
11122 }
11123 
11124 static void
11125 rack_check_bottom_drag(struct tcpcb *tp,
11126 		       struct tcp_rack *rack,
11127 		       struct socket *so)
11128 {
11129 	/*
11130 	 * So what is dragging bottom?
11131 	 *
11132 	 * Dragging bottom means you were under pacing and had a
11133 	 * delay in processing inbound acks waiting on our pacing
11134 	 * timer to expire. While you were waiting all of the acknowledgments
11135 	 * for the packets you sent have arrived. This means we are pacing
11136 	 * way underneath the bottleneck to the point where our Goodput
11137 	 * measurements stop working, since they require more than one
11138 	 * ack (usually at least 8 packets worth with multiple acks so we can
11139 	 * gauge the inter-ack times). If that occurs we have a real problem
11140 	 * since we are stuck in a hole that we can't get out of without
11141 	 * something speeding us up.
11142 	 *
11143 	 * We also check to see if we are widdling down to just one segment
11144 	 * outstanding. If this occurs and we have room to send in our cwnd/rwnd
11145 	 * then we are adding the delayed ack interval into our measurments and
11146 	 * we need to speed up slightly.
11147 	 */
11148 	uint32_t segsiz, minseg;
11149 
11150 	segsiz = ctf_fixed_maxseg(tp);
11151 	minseg = segsiz;
11152 	if (tp->snd_max == tp->snd_una) {
11153 		/*
11154 		 * We are doing dynamic pacing and we are way
11155 		 * under. Basically everything got acked while
11156 		 * we were still waiting on the pacer to expire.
11157 		 *
11158 		 * This means we need to boost the b/w in
11159 		 * addition to any earlier boosting of
11160 		 * the multiplier.
11161 		 */
11162 		uint64_t lt_bw;
11163 
11164 		tcp_trace_point(rack->rc_tp, TCP_TP_PACED_BOTTOM);
11165 		lt_bw = rack_get_lt_bw(rack);
11166 		rack->rc_dragged_bottom = 1;
11167 		rack_validate_multipliers_at_or_above100(rack);
11168 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11169 		    (rack->dis_lt_bw == 0) &&
11170 		    (rack->use_lesser_lt_bw == 0) &&
11171 		    (lt_bw > 0)) {
11172 			/*
11173 			 * Lets use the long-term b/w we have
11174 			 * been getting as a base.
11175 			 */
11176 			if (rack->rc_gp_filled == 0) {
11177 				if (lt_bw > ONE_POINT_TWO_MEG) {
11178 					/*
11179 					 * If we have no measurement
11180 					 * don't let us set in more than
11181 					 * 1.2Mbps. If we are still too
11182 					 * low after pacing with this we
11183 					 * will hopefully have a max b/w
11184 					 * available to sanity check things.
11185 					 */
11186 					lt_bw = ONE_POINT_TWO_MEG;
11187 				}
11188 				rack->r_ctl.rc_rtt_diff = 0;
11189 				rack->r_ctl.gp_bw = lt_bw;
11190 				rack->rc_gp_filled = 1;
11191 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11192 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11193 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11194 			} else if (lt_bw > rack->r_ctl.gp_bw) {
11195 				rack->r_ctl.rc_rtt_diff = 0;
11196 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11197 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11198 				rack->r_ctl.gp_bw = lt_bw;
11199 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11200 			} else
11201 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
11202 			if ((rack->gp_ready == 0) &&
11203 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11204 				/* We have enough measurements now */
11205 				rack->gp_ready = 1;
11206 				if (rack->dgp_on ||
11207 				    rack->rack_hibeta)
11208 					rack_set_cc_pacing(rack);
11209 				if (rack->defer_options)
11210 					rack_apply_deferred_options(rack);
11211 			}
11212 		} else {
11213 			/*
11214 			 * zero rtt possibly?, settle for just an old increase.
11215 			 */
11216 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
11217 		}
11218 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11219 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11220 					       minseg)) &&
11221 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11222 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11223 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11224 		    (segsiz * rack_req_segs))) {
11225 		/*
11226 		 * We are doing dynamic GP pacing and
11227 		 * we have everything except 1MSS or less
11228 		 * bytes left out. We are still pacing away.
11229 		 * And there is data that could be sent, This
11230 		 * means we are inserting delayed ack time in
11231 		 * our measurements because we are pacing too slow.
11232 		 */
11233 		rack_validate_multipliers_at_or_above100(rack);
11234 		rack->rc_dragged_bottom = 1;
11235 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
11236 	}
11237 }
11238 
11239 #ifdef TCP_REQUEST_TRK
11240 static void
11241 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11242 		struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11243 {
11244 	int do_log;
11245 
11246 	do_log = tcp_bblogging_on(rack->rc_tp);
11247 	if (do_log == 0) {
11248 		if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11249 			return;
11250 		/* We only allow the three below with point logging on */
11251 		if ((mod != HYBRID_LOG_RULES_APP) &&
11252 		    (mod != HYBRID_LOG_RULES_SET) &&
11253 		    (mod != HYBRID_LOG_REQ_COMP))
11254 			return;
11255 
11256 	}
11257 	if (do_log) {
11258 		union tcp_log_stackspecific log;
11259 		struct timeval tv;
11260 
11261 		/* Convert our ms to a microsecond */
11262 		memset(&log, 0, sizeof(log));
11263 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11264 		log.u_bbr.flex1 = seq;
11265 		log.u_bbr.cwnd_gain = line;
11266 		if (cur != NULL) {
11267 			uint64_t off;
11268 
11269 			log.u_bbr.flex2 = cur->start_seq;
11270 			log.u_bbr.flex3 = cur->end_seq;
11271 			log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11272 			log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11273 			log.u_bbr.flex6 = cur->flags;
11274 			log.u_bbr.pkts_out = cur->hybrid_flags;
11275 			log.u_bbr.rttProp = cur->timestamp;
11276 			log.u_bbr.cur_del_rate = cur->cspr;
11277 			log.u_bbr.bw_inuse = cur->start;
11278 			log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11279 			log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11280 			log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11281 			log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11282 			log.u_bbr.inhpts = 1;
11283 #ifdef TCP_REQUEST_TRK
11284 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11285 			log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11286 #endif
11287 		} else {
11288 			log.u_bbr.flex2 = err;
11289 		}
11290 		/*
11291 		 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11292 		 */
11293 		log.u_bbr.flex7 = rack->rc_catch_up;
11294 		log.u_bbr.flex7 <<= 1;
11295 		log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11296 		log.u_bbr.flex7 <<= 1;
11297 		log.u_bbr.flex7 |= rack->dgp_on;
11298 		/*
11299 		 * Compose bbr_state to be a bit wise 0000ADHF
11300 		 * where A is the always_pace flag
11301 		 * where D is the dgp_on flag
11302 		 * where H is the hybrid_mode on flag
11303 		 * where F is the use_fixed_rate flag.
11304 		 */
11305 		log.u_bbr.bbr_state = rack->rc_always_pace;
11306 		log.u_bbr.bbr_state <<= 1;
11307 		log.u_bbr.bbr_state |= rack->dgp_on;
11308 		log.u_bbr.bbr_state <<= 1;
11309 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
11310 		log.u_bbr.bbr_state <<= 1;
11311 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
11312 		log.u_bbr.flex8 = mod;
11313 		log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11314 		log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11315 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11316 		log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11317 		log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11318 		log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11319 		tcp_log_event(rack->rc_tp, NULL,
11320 		    &rack->rc_inp->inp_socket->so_rcv,
11321 		    &rack->rc_inp->inp_socket->so_snd,
11322 		    TCP_HYBRID_PACING_LOG, 0,
11323 	            0, &log, false, NULL, __func__, __LINE__, &tv);
11324 	}
11325 }
11326 #endif
11327 
11328 #ifdef TCP_REQUEST_TRK
11329 static void
11330 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11331 {
11332 	struct tcp_sendfile_track *rc_cur, *orig_ent;
11333 	struct tcpcb *tp;
11334 	int err = 0;
11335 
11336 	orig_ent = rack->r_ctl.rc_last_sft;
11337 	rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
11338 	if (rc_cur == NULL) {
11339 		/* If not in the beginning what about the end piece */
11340 		if (rack->rc_hybrid_mode)
11341 			rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11342 		rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11343 	} else {
11344 		err = 12345;
11345 	}
11346 	/* If we find no parameters we are in straight DGP mode */
11347 	if(rc_cur == NULL) {
11348 		/* None found for this seq, just DGP for now */
11349 		if (rack->rc_hybrid_mode) {
11350 			rack->r_ctl.client_suggested_maxseg = 0;
11351 			rack->rc_catch_up = 0;
11352 			if (rack->cspr_is_fcc == 0)
11353 				rack->r_ctl.bw_rate_cap = 0;
11354 			else
11355 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11356 		}
11357 		if (rack->rc_hybrid_mode) {
11358 			rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11359 		}
11360 		if (rack->r_ctl.rc_last_sft) {
11361 			rack->r_ctl.rc_last_sft = NULL;
11362 		}
11363 		return;
11364 	}
11365 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
11366 		/* This entry was never setup for hybrid pacing on/off etc */
11367 		if (rack->rc_hybrid_mode) {
11368 			rack->r_ctl.client_suggested_maxseg = 0;
11369 			rack->rc_catch_up = 0;
11370 			rack->r_ctl.bw_rate_cap = 0;
11371 		}
11372 		if (rack->r_ctl.rc_last_sft) {
11373 			rack->r_ctl.rc_last_sft = NULL;
11374 		}
11375 		if ((rc_cur->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11376 			rc_cur->flags |= TCP_TRK_TRACK_FLG_FSND;
11377 			rc_cur->first_send = cts;
11378 			rc_cur->sent_at_fs = rack->rc_tp->t_sndbytes;
11379 			rc_cur->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11380 		}
11381 		return;
11382 	}
11383 	/*
11384 	 * Ok if we have a new entry *or* have never
11385 	 * set up an entry we need to proceed. If
11386 	 * we have already set it up this entry we
11387 	 * just continue along with what we already
11388 	 * setup.
11389 	 */
11390 	tp = rack->rc_tp;
11391 	if ((rack->r_ctl.rc_last_sft != NULL) &&
11392 	    (rack->r_ctl.rc_last_sft == rc_cur)) {
11393 		/* Its already in place */
11394 		if (rack->rc_hybrid_mode)
11395 			rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11396 		return;
11397 	}
11398 	if (rack->rc_hybrid_mode == 0) {
11399 		rack->r_ctl.rc_last_sft = rc_cur;
11400 		if (orig_ent) {
11401 			orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11402 			orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11403 			orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11404 		}
11405 		rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11406 		return;
11407 	}
11408 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11409 		/* Compensate for all the header overhead's */
11410 		if (rack->cspr_is_fcc == 0)
11411 			rack->r_ctl.bw_rate_cap	= rack_compensate_for_linerate(rack, rc_cur->cspr);
11412 		else
11413 			rack->r_ctl.fillcw_cap =  rack_compensate_for_linerate(rack, rc_cur->cspr);
11414 	} else {
11415 		if (rack->rc_hybrid_mode) {
11416 			if (rack->cspr_is_fcc == 0)
11417 				rack->r_ctl.bw_rate_cap = 0;
11418 			else
11419 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11420 		}
11421 	}
11422 	if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11423 		rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11424 	else
11425 		rack->r_ctl.client_suggested_maxseg = 0;
11426 	if (rc_cur->timestamp == rack->r_ctl.last_tm_mark) {
11427 		/*
11428 		 * It is the same timestamp as the previous one
11429 		 * add the hybrid flag that will indicate we use
11430 		 * sendtime not arrival time for catch-up mode.
11431 		 */
11432 		rc_cur->hybrid_flags |= TCP_HYBRID_PACING_SENDTIME;
11433 	}
11434 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11435 	    (rc_cur->cspr > 0)) {
11436 		uint64_t len;
11437 
11438 		rack->rc_catch_up = 1;
11439 		/*
11440 		 * Calculate the deadline time, first set the
11441 		 * time to when the request arrived.
11442 		 */
11443 		if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_SENDTIME) {
11444 			/*
11445 			 * For cases where its a duplicate tm (we received more
11446 			 * than one request for a tm) we want to use now, the point
11447 			 * where we are just sending the first bit of the request.
11448 			 */
11449 			rc_cur->deadline = cts;
11450 		} else {
11451 			/*
11452 			 * Here we have a different tm from the last request
11453 			 * so we want to use arrival time as our base.
11454 			 */
11455 			rc_cur->deadline = rc_cur->localtime;
11456 		}
11457 		/*
11458 		 * Next calculate the length and compensate for
11459 		 * TLS if need be.
11460 		 */
11461 		len = rc_cur->end - rc_cur->start;
11462 		if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11463 			/*
11464 			 * This session is doing TLS. Take a swag guess
11465 			 * at the overhead.
11466 			 */
11467 			len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11468 		}
11469 		/*
11470 		 * Now considering the size, and the cspr, what is the time that
11471 		 * would be required at the cspr rate. Here we use the raw
11472 		 * cspr value since the client only looks at the raw data. We
11473 		 * do use len which includes TLS overhead, but not the TCP/IP etc.
11474 		 * That will get made up for in the CU pacing rate set.
11475 		 */
11476 		len *= HPTS_USEC_IN_SEC;
11477 		len /= rc_cur->cspr;
11478 		rc_cur->deadline += len;
11479 	} else {
11480 		rack->rc_catch_up = 0;
11481 		rc_cur->deadline = 0;
11482 	}
11483 	if (rack->r_ctl.client_suggested_maxseg != 0) {
11484 		/*
11485 		 * We need to reset the max pace segs if we have a
11486 		 * client_suggested_maxseg.
11487 		 */
11488 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
11489 	}
11490 	if (orig_ent) {
11491 		orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11492 		orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11493 		orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11494 	}
11495 	rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11496 	/* Remember it for next time and for CU mode */
11497 	rack->r_ctl.rc_last_sft = rc_cur;
11498 	rack->r_ctl.last_tm_mark = rc_cur->timestamp;
11499 }
11500 #endif
11501 
11502 static void
11503 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11504 {
11505 #ifdef TCP_REQUEST_TRK
11506 	struct tcp_sendfile_track *ent;
11507 
11508 	ent = rack->r_ctl.rc_last_sft;
11509 	if ((ent == NULL) ||
11510 	    (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
11511 	    (SEQ_GEQ(seq, ent->end_seq))) {
11512 		/* Time to update the track. */
11513 		rack_set_dgp_hybrid_mode(rack, seq, len, cts);
11514 		ent = rack->r_ctl.rc_last_sft;
11515 	}
11516 	/* Out of all */
11517 	if (ent == NULL) {
11518 		return;
11519 	}
11520 	if (SEQ_LT(ent->end_seq, (seq + len))) {
11521 		/*
11522 		 * This is the case where our end_seq guess
11523 		 * was wrong. This is usually due to TLS having
11524 		 * more bytes then our guess. It could also be the
11525 		 * case that the client sent in two requests closely
11526 		 * and the SB is full of both so we are sending part
11527 		 * of each (end|beg). In such a case lets move this
11528 		 * guys end to match the end of this send. That
11529 		 * way it will complete when all of it is acked.
11530 		 */
11531 		ent->end_seq = (seq + len);
11532 		if (rack->rc_hybrid_mode)
11533 			rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
11534 	}
11535 	/* Now validate we have set the send time of this one */
11536 	if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11537 		ent->flags |= TCP_TRK_TRACK_FLG_FSND;
11538 		ent->first_send = cts;
11539 		ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11540 		ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11541 	}
11542 #endif
11543 }
11544 
11545 static void
11546 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11547 {
11548 	/*
11549 	 * The fast output path is enabled and we
11550 	 * have moved the cumack forward. Lets see if
11551 	 * we can expand forward the fast path length by
11552 	 * that amount. What we would ideally like to
11553 	 * do is increase the number of bytes in the
11554 	 * fast path block (left_to_send) by the
11555 	 * acked amount. However we have to gate that
11556 	 * by two factors:
11557 	 * 1) The amount outstanding and the rwnd of the peer
11558 	 *    (i.e. we don't want to exceed the rwnd of the peer).
11559 	 *    <and>
11560 	 * 2) The amount of data left in the socket buffer (i.e.
11561 	 *    we can't send beyond what is in the buffer).
11562 	 *
11563 	 * Note that this does not take into account any increase
11564 	 * in the cwnd. We will only extend the fast path by
11565 	 * what was acked.
11566 	 */
11567 	uint32_t new_total, gating_val;
11568 
11569 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11570 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11571 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11572 	if (new_total <= gating_val) {
11573 		/* We can increase left_to_send by the acked amount */
11574 		counter_u64_add(rack_extended_rfo, 1);
11575 		rack->r_ctl.fsb.left_to_send = new_total;
11576 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11577 			("rack:%p left_to_send:%u sbavail:%u out:%u",
11578 			 rack, rack->r_ctl.fsb.left_to_send,
11579 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
11580 			 (tp->snd_max - tp->snd_una)));
11581 
11582 	}
11583 }
11584 
11585 static void
11586 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11587 {
11588 	/*
11589 	 * Here any sendmap entry that points to the
11590 	 * beginning mbuf must be adjusted to the correct
11591 	 * offset. This must be called with:
11592 	 * 1) The socket buffer locked
11593 	 * 2) snd_una adjusted to its new position.
11594 	 *
11595 	 * Note that (2) implies rack_ack_received has also
11596 	 * been called and all the sbcut's have been done.
11597 	 *
11598 	 * We grab the first mbuf in the socket buffer and
11599 	 * then go through the front of the sendmap, recalculating
11600 	 * the stored offset for any sendmap entry that has
11601 	 * that mbuf. We must use the sb functions to do this
11602 	 * since its possible an add was done has well as
11603 	 * the subtraction we may have just completed. This should
11604 	 * not be a penalty though, since we just referenced the sb
11605 	 * to go in and trim off the mbufs that we freed (of course
11606 	 * there will be a penalty for the sendmap references though).
11607 	 *
11608 	 * Note also with INVARIANT on, we validate with a KASSERT
11609 	 * that the first sendmap entry has a soff of 0.
11610 	 *
11611 	 */
11612 	struct mbuf *m;
11613 	struct rack_sendmap *rsm;
11614 	tcp_seq snd_una;
11615 #ifdef INVARIANTS
11616 	int first_processed = 0;
11617 #endif
11618 
11619 	snd_una = rack->rc_tp->snd_una;
11620 	SOCKBUF_LOCK_ASSERT(sb);
11621 	m = sb->sb_mb;
11622 	rsm = tqhash_min(rack->r_ctl.tqh);
11623 	if ((rsm == NULL) || (m == NULL)) {
11624 		/* Nothing outstanding */
11625 		return;
11626 	}
11627 	/* The very first RSM's mbuf must point to the head mbuf in the sb */
11628 	KASSERT((rsm->m == m),
11629 		("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
11630 		 rack, sb, rsm));
11631 	while (rsm->m && (rsm->m == m)) {
11632 		/* one to adjust */
11633 #ifdef INVARIANTS
11634 		struct mbuf *tm;
11635 		uint32_t soff;
11636 
11637 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
11638 		if ((rsm->orig_m_len != m->m_len) ||
11639 		    (rsm->orig_t_space != M_TRAILINGROOM(m))){
11640 			rack_adjust_orig_mlen(rsm);
11641 		}
11642 		if (first_processed == 0) {
11643 			KASSERT((rsm->soff == 0),
11644 				("Rack:%p rsm:%p -- rsm at head but soff not zero",
11645 				 rack, rsm));
11646 			first_processed = 1;
11647 		}
11648 		if ((rsm->soff != soff) || (rsm->m != tm)) {
11649 			/*
11650 			 * This is not a fatal error, we anticipate it
11651 			 * might happen (the else code), so we count it here
11652 			 * so that under invariant we can see that it really
11653 			 * does happen.
11654 			 */
11655 			counter_u64_add(rack_adjust_map_bw, 1);
11656 		}
11657 		rsm->m = tm;
11658 		rsm->soff = soff;
11659 		if (tm) {
11660 			rsm->orig_m_len = rsm->m->m_len;
11661 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11662 		} else {
11663 			rsm->orig_m_len = 0;
11664 			rsm->orig_t_space = 0;
11665 		}
11666 #else
11667 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
11668 		if (rsm->m) {
11669 			rsm->orig_m_len = rsm->m->m_len;
11670 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11671 		} else {
11672 			rsm->orig_m_len = 0;
11673 			rsm->orig_t_space = 0;
11674 		}
11675 #endif
11676 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
11677 		if (rsm == NULL)
11678 			break;
11679 	}
11680 }
11681 
11682 #ifdef TCP_REQUEST_TRK
11683 static inline void
11684 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
11685 {
11686 	struct tcp_sendfile_track *ent;
11687 	int i;
11688 
11689 	if ((rack->rc_hybrid_mode == 0) &&
11690 	    (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
11691 		/*
11692 		 * Just do normal completions hybrid pacing is not on
11693 		 * and CLDL is off as well.
11694 		 */
11695 		tcp_req_check_for_comp(rack->rc_tp, th_ack);
11696 		return;
11697 	}
11698 	/*
11699 	 * Originally I was just going to find the th_ack associated
11700 	 * with an entry. But then I realized a large strech ack could
11701 	 * in theory ack two or more requests at once. So instead we
11702 	 * need to find all entries that are completed by th_ack not
11703 	 * just a single entry and do our logging.
11704 	 */
11705 	ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11706 	while (ent != NULL) {
11707 		/*
11708 		 * We may be doing hybrid pacing or CLDL and need more details possibly
11709 		 * so we do it manually instead of calling
11710 		 * tcp_req_check_for_comp()
11711 		 */
11712 		uint64_t laa, tim, data, cbw, ftim;
11713 
11714 		/* Ok this ack frees it */
11715 		rack_log_hybrid(rack, th_ack,
11716 				ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
11717 		rack_log_hybrid_sends(rack, ent, __LINE__);
11718 		/* calculate the time based on the ack arrival */
11719 		data = ent->end - ent->start;
11720 		laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
11721 		if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
11722 			if (ent->first_send > ent->localtime)
11723 				ftim = ent->first_send;
11724 			else
11725 				ftim = ent->localtime;
11726 		} else {
11727 			/* TSNH */
11728 			ftim = ent->localtime;
11729 		}
11730 		if (laa > ent->localtime)
11731 			tim = laa - ftim;
11732 		else
11733 			tim = 0;
11734 		cbw = data * HPTS_USEC_IN_SEC;
11735 		if (tim > 0)
11736 			cbw /= tim;
11737 		else
11738 			cbw = 0;
11739 		rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
11740 		/*
11741 		 * Check to see if we are freeing what we are pointing to send wise
11742 		 * if so be sure to NULL the pointer so we know we are no longer
11743 		 * set to anything.
11744 		 */
11745 		if (ent == rack->r_ctl.rc_last_sft) {
11746 			rack->r_ctl.rc_last_sft = NULL;
11747 			if (rack->rc_hybrid_mode) {
11748 				rack->rc_catch_up = 0;
11749 				if (rack->cspr_is_fcc == 0)
11750 					rack->r_ctl.bw_rate_cap = 0;
11751 				else
11752 					rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11753 				rack->r_ctl.client_suggested_maxseg = 0;
11754 			}
11755 		}
11756 		/* Generate the log that the tcp_netflix call would have */
11757 		tcp_req_log_req_info(rack->rc_tp, ent,
11758 				      i, TCP_TRK_REQ_LOG_FREED, 0, 0);
11759 		/* Free it and see if there is another one */
11760 		tcp_req_free_a_slot(rack->rc_tp, ent);
11761 		ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11762 	}
11763 }
11764 #endif
11765 
11766 
11767 /*
11768  * Return value of 1, we do not need to call rack_process_data().
11769  * return value of 0, rack_process_data can be called.
11770  * For ret_val if its 0 the TCP is locked, if its non-zero
11771  * its unlocked and probably unsafe to touch the TCB.
11772  */
11773 static int
11774 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11775     struct tcpcb *tp, struct tcpopt *to,
11776     uint32_t tiwin, int32_t tlen,
11777     int32_t * ofia, int32_t thflags, int32_t *ret_val, int32_t orig_tlen)
11778 {
11779 	int32_t ourfinisacked = 0;
11780 	int32_t nsegs, acked_amount;
11781 	int32_t acked;
11782 	struct mbuf *mfree;
11783 	struct tcp_rack *rack;
11784 	int32_t under_pacing = 0;
11785 	int32_t post_recovery = 0;
11786 	uint32_t p_cwnd;
11787 
11788 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11789 
11790 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11791 	if (SEQ_GEQ(tp->snd_una, tp->iss + (65535 << tp->snd_scale))) {
11792 		/* Checking SEG.ACK against ISS is definitely redundant. */
11793 		tp->t_flags2 |= TF2_NO_ISS_CHECK;
11794 	}
11795 	if (!V_tcp_insecure_ack) {
11796 		tcp_seq seq_min;
11797 		bool ghost_ack_check;
11798 
11799 		if (tp->t_flags2 & TF2_NO_ISS_CHECK) {
11800 			/* Check for too old ACKs (RFC 5961, Section 5.2). */
11801 			seq_min = tp->snd_una - tp->max_sndwnd;
11802 			ghost_ack_check = false;
11803 		} else {
11804 			if (SEQ_GT(tp->iss + 1, tp->snd_una - tp->max_sndwnd)) {
11805 				/* Checking for ghost ACKs is stricter. */
11806 				seq_min = tp->iss + 1;
11807 				ghost_ack_check = true;
11808 			} else {
11809 				/*
11810 				 * Checking for too old ACKs (RFC 5961,
11811 				 * Section 5.2) is stricter.
11812 				 */
11813 				seq_min = tp->snd_una - tp->max_sndwnd;
11814 				ghost_ack_check = false;
11815 			}
11816 		}
11817 		if (SEQ_LT(th->th_ack, seq_min)) {
11818 			if (ghost_ack_check)
11819 				TCPSTAT_INC(tcps_rcvghostack);
11820 			else
11821 				TCPSTAT_INC(tcps_rcvacktooold);
11822 			/* Send challenge ACK. */
11823 			ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11824 			rack->r_wanted_output = 1;
11825 			return (1);
11826 		}
11827 	}
11828 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
11829 		ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11830 		rack->r_wanted_output = 1;
11831 		return (1);
11832 	}
11833 	if (rack->gp_ready &&
11834 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11835 		under_pacing = 1;
11836 	}
11837 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
11838 		int in_rec, dup_ack_struck = 0;
11839 		int dsack_seen = 0, sacks_seen = 0;
11840 
11841 		in_rec = IN_FASTRECOVERY(tp->t_flags);
11842 		if (rack->rc_in_persist) {
11843 			tp->t_rxtshift = 0;
11844 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11845 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11846 		}
11847 
11848 		if ((th->th_ack == tp->snd_una) &&
11849 		    (tiwin == tp->snd_wnd) &&
11850 		    (orig_tlen == 0) &&
11851 		    ((to->to_flags & TOF_SACK) == 0)) {
11852 			rack_strike_dupack(rack, th->th_ack);
11853 			dup_ack_struck = 1;
11854 		}
11855 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
11856 			     dup_ack_struck, &dsack_seen, &sacks_seen);
11857 
11858 	}
11859 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
11860 		/*
11861 		 * Old ack, behind (or duplicate to) the last one rcv'd
11862 		 * Note: We mark reordering is occuring if its
11863 		 * less than and we have not closed our window.
11864 		 */
11865 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
11866 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
11867 			if (rack->r_ctl.rc_reorder_ts == 0)
11868 				rack->r_ctl.rc_reorder_ts = 1;
11869 		}
11870 		return (0);
11871 	}
11872 	/*
11873 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
11874 	 * something we sent.
11875 	 */
11876 	if (tp->t_flags & TF_NEEDSYN) {
11877 		/*
11878 		 * T/TCP: Connection was half-synchronized, and our SYN has
11879 		 * been ACK'd (so connection is now fully synchronized).  Go
11880 		 * to non-starred state, increment snd_una for ACK of SYN,
11881 		 * and check if we can do window scaling.
11882 		 */
11883 		tp->t_flags &= ~TF_NEEDSYN;
11884 		tp->snd_una++;
11885 		/* Do window scaling? */
11886 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11887 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11888 			tp->rcv_scale = tp->request_r_scale;
11889 			/* Send window already scaled. */
11890 		}
11891 	}
11892 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
11893 
11894 	acked = BYTES_THIS_ACK(tp, th);
11895 	if (acked) {
11896 		/*
11897 		 * Any time we move the cum-ack forward clear
11898 		 * keep-alive tied probe-not-answered. The
11899 		 * persists clears its own on entry.
11900 		 */
11901 		rack->probe_not_answered = 0;
11902 	}
11903 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11904 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11905 	/*
11906 	 * If we just performed our first retransmit, and the ACK arrives
11907 	 * within our recovery window, then it was a mistake to do the
11908 	 * retransmit in the first place.  Recover our original cwnd and
11909 	 * ssthresh, and proceed to transmit where we left off.
11910 	 */
11911 	if ((tp->t_flags & TF_PREVVALID) &&
11912 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11913 		tp->t_flags &= ~TF_PREVVALID;
11914 		if (tp->t_rxtshift == 1 &&
11915 		    (int)(ticks - tp->t_badrxtwin) < 0)
11916 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
11917 	}
11918 	if (acked) {
11919 		/* assure we are not backed off */
11920 		tp->t_rxtshift = 0;
11921 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11922 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11923 		rack->rc_tlp_in_progress = 0;
11924 		rack->r_ctl.rc_tlp_cnt_out = 0;
11925 		/*
11926 		 * If it is the RXT timer we want to
11927 		 * stop it, so we can restart a TLP.
11928 		 */
11929 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11930 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11931 #ifdef TCP_REQUEST_TRK
11932 		rack_req_check_for_comp(rack, th->th_ack);
11933 #endif
11934 	}
11935 	/*
11936 	 * If we have a timestamp reply, update smoothed round trip time. If
11937 	 * no timestamp is present but transmit timer is running and timed
11938 	 * sequence number was acked, update smoothed round trip time. Since
11939 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
11940 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
11941 	 * timer.
11942 	 *
11943 	 * Some boxes send broken timestamp replies during the SYN+ACK
11944 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
11945 	 * and blow up the retransmit timer.
11946 	 */
11947 	/*
11948 	 * If all outstanding data is acked, stop retransmit timer and
11949 	 * remember to restart (more output or persist). If there is more
11950 	 * data to be acked, restart retransmit timer, using current
11951 	 * (possibly backed-off) value.
11952 	 */
11953 	if (acked == 0) {
11954 		if (ofia)
11955 			*ofia = ourfinisacked;
11956 		return (0);
11957 	}
11958 	if (IN_RECOVERY(tp->t_flags)) {
11959 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
11960 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
11961 			tcp_rack_partialack(tp);
11962 		} else {
11963 			rack_post_recovery(tp, th->th_ack);
11964 			post_recovery = 1;
11965 			/*
11966 			 * Grab the segsiz, multiply by 2 and add the snd_cwnd
11967 			 * that is the max the CC should add if we are exiting
11968 			 * recovery and doing a late add.
11969 			 */
11970 			p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
11971 			p_cwnd <<= 1;
11972 			p_cwnd += tp->snd_cwnd;
11973 		}
11974 	} else if ((rack->rto_from_rec == 1) &&
11975 		   SEQ_GEQ(th->th_ack, tp->snd_recover)) {
11976 		/*
11977 		 * We were in recovery, hit a rxt timeout
11978 		 * and never re-entered recovery. The timeout(s)
11979 		 * made up all the lost data. In such a case
11980 		 * we need to clear the rto_from_rec flag.
11981 		 */
11982 		rack->rto_from_rec = 0;
11983 	}
11984 	/*
11985 	 * Let the congestion control algorithm update congestion control
11986 	 * related information. This typically means increasing the
11987 	 * congestion window.
11988 	 */
11989 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, post_recovery);
11990 	if (post_recovery &&
11991 	    (tp->snd_cwnd > p_cwnd)) {
11992 		/* Must be non-newreno (cubic) getting too ahead of itself */
11993 		tp->snd_cwnd = p_cwnd;
11994 	}
11995 	SOCKBUF_LOCK(&so->so_snd);
11996 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
11997 	tp->snd_wnd -= acked_amount;
11998 	mfree = sbcut_locked(&so->so_snd, acked_amount);
11999 	if ((sbused(&so->so_snd) == 0) &&
12000 	    (acked > acked_amount) &&
12001 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
12002 	    (tp->t_flags & TF_SENTFIN)) {
12003 		/*
12004 		 * We must be sure our fin
12005 		 * was sent and acked (we can be
12006 		 * in FIN_WAIT_1 without having
12007 		 * sent the fin).
12008 		 */
12009 		ourfinisacked = 1;
12010 	}
12011 	tp->snd_una = th->th_ack;
12012 	/* wakeups? */
12013 	if (acked_amount && sbavail(&so->so_snd))
12014 		rack_adjust_sendmap_head(rack, &so->so_snd);
12015 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12016 	/* NB: sowwakeup_locked() does an implicit unlock. */
12017 	sowwakeup_locked(so);
12018 	m_freem(mfree);
12019 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
12020 		tp->snd_recover = tp->snd_una;
12021 
12022 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12023 		tp->snd_nxt = tp->snd_max;
12024 	}
12025 	if (under_pacing &&
12026 	    (rack->use_fixed_rate == 0) &&
12027 	    (rack->in_probe_rtt == 0) &&
12028 	    rack->rc_gp_dyn_mul &&
12029 	    rack->rc_always_pace) {
12030 		/* Check if we are dragging bottom */
12031 		rack_check_bottom_drag(tp, rack, so);
12032 	}
12033 	if (tp->snd_una == tp->snd_max) {
12034 		/* Nothing left outstanding */
12035 		tp->t_flags &= ~TF_PREVVALID;
12036 		if (rack->r_ctl.rc_went_idle_time == 0)
12037 			rack->r_ctl.rc_went_idle_time = 1;
12038 		rack->r_ctl.retran_during_recovery = 0;
12039 		rack->r_ctl.dsack_byte_cnt = 0;
12040 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12041 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12042 			tp->t_acktime = 0;
12043 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12044 		rack->rc_suspicious = 0;
12045 		/* Set need output so persist might get set */
12046 		rack->r_wanted_output = 1;
12047 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12048 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12049 		    (sbavail(&so->so_snd) == 0) &&
12050 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12051 			/*
12052 			 * The socket was gone and the
12053 			 * peer sent data (now or in the past), time to
12054 			 * reset him.
12055 			 */
12056 			*ret_val = 1;
12057 			/* tcp_close will kill the inp pre-log the Reset */
12058 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12059 			tp = tcp_close(tp);
12060 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12061 			return (1);
12062 		}
12063 	}
12064 	if (ofia)
12065 		*ofia = ourfinisacked;
12066 	return (0);
12067 }
12068 
12069 
12070 static void
12071 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12072 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
12073 {
12074 	if (tcp_bblogging_on(rack->rc_tp)) {
12075 		union tcp_log_stackspecific log;
12076 		struct timeval tv;
12077 
12078 		memset(&log, 0, sizeof(log));
12079 		log.u_bbr.flex1 = cnt;
12080 		log.u_bbr.flex2 = split;
12081 		log.u_bbr.flex3 = out;
12082 		log.u_bbr.flex4 = line;
12083 		log.u_bbr.flex5 = rack->r_must_retran;
12084 		log.u_bbr.flex6 = flags;
12085 		log.u_bbr.flex7 = rack->rc_has_collapsed;
12086 		log.u_bbr.flex8 = dir;	/*
12087 					 * 1 is collapsed, 0 is uncollapsed,
12088 					 * 2 is log of a rsm being marked, 3 is a split.
12089 					 */
12090 		if (rsm == NULL)
12091 			log.u_bbr.rttProp = 0;
12092 		else
12093 			log.u_bbr.rttProp = (uintptr_t)rsm;
12094 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12095 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12096 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
12097 		    &rack->rc_inp->inp_socket->so_rcv,
12098 		    &rack->rc_inp->inp_socket->so_snd,
12099 		    TCP_RACK_LOG_COLLAPSE, 0,
12100 		    0, &log, false, &tv);
12101 	}
12102 }
12103 
12104 static void
12105 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12106 {
12107 	/*
12108 	 * Here all we do is mark the collapsed point and set the flag.
12109 	 * This may happen again and again, but there is no
12110 	 * sense splitting our map until we know where the
12111 	 * peer finally lands in the collapse.
12112 	 */
12113 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12114 	if ((rack->rc_has_collapsed == 0) ||
12115 	    (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12116 		counter_u64_add(rack_collapsed_win_seen, 1);
12117 	rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12118 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12119 	rack->rc_has_collapsed = 1;
12120 	rack->r_collapse_point_valid = 1;
12121 	rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12122 }
12123 
12124 static void
12125 rack_un_collapse_window(struct tcp_rack *rack, int line)
12126 {
12127 	struct rack_sendmap *nrsm, *rsm;
12128 	int cnt = 0, split = 0;
12129 	int insret __diagused;
12130 
12131 
12132 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12133 	rack->rc_has_collapsed = 0;
12134 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12135 	if (rsm == NULL) {
12136 		/* Nothing to do maybe the peer ack'ed it all */
12137 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12138 		return;
12139 	}
12140 	/* Now do we need to split this one? */
12141 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12142 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12143 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12144 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12145 		if (nrsm == NULL) {
12146 			/* We can't get a rsm, mark all? */
12147 			nrsm = rsm;
12148 			goto no_split;
12149 		}
12150 		/* Clone it */
12151 		split = 1;
12152 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12153 #ifndef INVARIANTS
12154 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12155 #else
12156 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12157 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
12158 			      nrsm, insret, rack, rsm);
12159 		}
12160 #endif
12161 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12162 				 rack->r_ctl.last_collapse_point, __LINE__);
12163 		if (rsm->r_in_tmap) {
12164 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12165 			nrsm->r_in_tmap = 1;
12166 		}
12167 		/*
12168 		 * Set in the new RSM as the
12169 		 * collapsed starting point
12170 		 */
12171 		rsm = nrsm;
12172 	}
12173 
12174 no_split:
12175 	TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
12176 		cnt++;
12177 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
12178 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12179 		cnt++;
12180 	}
12181 	if (cnt) {
12182 		counter_u64_add(rack_collapsed_win, 1);
12183 	}
12184 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12185 }
12186 
12187 static void
12188 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12189 			int32_t tlen, int32_t tfo_syn)
12190 {
12191 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
12192 		rack_timer_cancel(tp, rack,
12193 				  rack->r_ctl.rc_rcvtime, __LINE__);
12194 		tp->t_flags |= TF_DELACK;
12195 	} else {
12196 		rack->r_wanted_output = 1;
12197 		tp->t_flags |= TF_ACKNOW;
12198 	}
12199 }
12200 
12201 static void
12202 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12203 {
12204 	/*
12205 	 * If fast output is in progress, lets validate that
12206 	 * the new window did not shrink on us and make it
12207 	 * so fast output should end.
12208 	 */
12209 	if (rack->r_fast_output) {
12210 		uint32_t out;
12211 
12212 		/*
12213 		 * Calculate what we will send if left as is
12214 		 * and compare that to our send window.
12215 		 */
12216 		out = ctf_outstanding(tp);
12217 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12218 			/* ok we have an issue */
12219 			if (out >= tp->snd_wnd) {
12220 				/* Turn off fast output the window is met or collapsed */
12221 				rack->r_fast_output = 0;
12222 			} else {
12223 				/* we have some room left */
12224 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12225 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12226 					/* If not at least 1 full segment never mind */
12227 					rack->r_fast_output = 0;
12228 				}
12229 			}
12230 		}
12231 	}
12232 }
12233 
12234 /*
12235  * Return value of 1, the TCB is unlocked and most
12236  * likely gone, return value of 0, the TCP is still
12237  * locked.
12238  */
12239 static int
12240 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12241     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12242     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12243 {
12244 	/*
12245 	 * Update window information. Don't look at window if no ACK: TAC's
12246 	 * send garbage on first SYN.
12247 	 */
12248 	int32_t nsegs;
12249 	int32_t tfo_syn;
12250 	struct tcp_rack *rack;
12251 
12252 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12253 
12254 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12255 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12256 	if ((thflags & TH_ACK) &&
12257 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12258 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12259 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12260 		/* keep track of pure window updates */
12261 		if (tlen == 0 &&
12262 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12263 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12264 		tp->snd_wnd = tiwin;
12265 		rack_validate_fo_sendwin_up(tp, rack);
12266 		tp->snd_wl1 = th->th_seq;
12267 		tp->snd_wl2 = th->th_ack;
12268 		if (tp->snd_wnd > tp->max_sndwnd)
12269 			tp->max_sndwnd = tp->snd_wnd;
12270 		rack->r_wanted_output = 1;
12271 	} else if (thflags & TH_ACK) {
12272 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12273 			tp->snd_wnd = tiwin;
12274 			rack_validate_fo_sendwin_up(tp, rack);
12275 			tp->snd_wl1 = th->th_seq;
12276 			tp->snd_wl2 = th->th_ack;
12277 		}
12278 	}
12279 	if (tp->snd_wnd < ctf_outstanding(tp))
12280 		/* The peer collapsed the window */
12281 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12282 	else if (rack->rc_has_collapsed)
12283 		rack_un_collapse_window(rack, __LINE__);
12284 	if ((rack->r_collapse_point_valid) &&
12285 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12286 		rack->r_collapse_point_valid = 0;
12287 	/* Was persist timer active and now we have window space? */
12288 	if ((rack->rc_in_persist != 0) &&
12289 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12290 				rack->r_ctl.rc_pace_min_segs))) {
12291 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12292 		tp->snd_nxt = tp->snd_max;
12293 		/* Make sure we output to start the timer */
12294 		rack->r_wanted_output = 1;
12295 	}
12296 	/* Do we enter persists? */
12297 	if ((rack->rc_in_persist == 0) &&
12298 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12299 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12300 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12301 	    sbavail(&tptosocket(tp)->so_snd) &&
12302 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12303 		/*
12304 		 * Here the rwnd is less than
12305 		 * the pacing size, we are established,
12306 		 * nothing is outstanding, and there is
12307 		 * data to send. Enter persists.
12308 		 */
12309 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12310 	}
12311 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12312 		m_freem(m);
12313 		return (0);
12314 	}
12315 	/*
12316 	 * don't process the URG bit, ignore them drag
12317 	 * along the up.
12318 	 */
12319 	tp->rcv_up = tp->rcv_nxt;
12320 
12321 	/*
12322 	 * Process the segment text, merging it into the TCP sequencing
12323 	 * queue, and arranging for acknowledgment of receipt if necessary.
12324 	 * This process logically involves adjusting tp->rcv_wnd as data is
12325 	 * presented to the user (this happens in tcp_usrreq.c, case
12326 	 * PRU_RCVD).  If a FIN has already been received on this connection
12327 	 * then we just ignore the text.
12328 	 */
12329 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12330 	    (tp->t_flags & TF_FASTOPEN));
12331 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12332 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12333 		tcp_seq save_start = th->th_seq;
12334 		tcp_seq save_rnxt  = tp->rcv_nxt;
12335 		int     save_tlen  = tlen;
12336 
12337 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12338 		/*
12339 		 * Insert segment which includes th into TCP reassembly
12340 		 * queue with control block tp.  Set thflags to whether
12341 		 * reassembly now includes a segment with FIN.  This handles
12342 		 * the common case inline (segment is the next to be
12343 		 * received on an established connection, and the queue is
12344 		 * empty), avoiding linkage into and removal from the queue
12345 		 * and repetition of various conversions. Set DELACK for
12346 		 * segments received in order, but ack immediately when
12347 		 * segments are out of order (so fast retransmit can work).
12348 		 */
12349 		if (th->th_seq == tp->rcv_nxt &&
12350 		    SEGQ_EMPTY(tp) &&
12351 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
12352 		    tfo_syn)) {
12353 #ifdef NETFLIX_SB_LIMITS
12354 			u_int mcnt, appended;
12355 
12356 			if (so->so_rcv.sb_shlim) {
12357 				mcnt = m_memcnt(m);
12358 				appended = 0;
12359 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12360 				    CFO_NOSLEEP, NULL) == false) {
12361 					counter_u64_add(tcp_sb_shlim_fails, 1);
12362 					m_freem(m);
12363 					return (0);
12364 				}
12365 			}
12366 #endif
12367 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12368 			tp->rcv_nxt += tlen;
12369 			if (tlen &&
12370 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12371 			    (tp->t_fbyte_in == 0)) {
12372 				tp->t_fbyte_in = ticks;
12373 				if (tp->t_fbyte_in == 0)
12374 					tp->t_fbyte_in = 1;
12375 				if (tp->t_fbyte_out && tp->t_fbyte_in)
12376 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12377 			}
12378 			thflags = tcp_get_flags(th) & TH_FIN;
12379 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12380 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12381 			SOCKBUF_LOCK(&so->so_rcv);
12382 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12383 				m_freem(m);
12384 			} else {
12385 				int32_t newsize;
12386 
12387 				if (tlen > 0) {
12388 					newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12389 					if (newsize)
12390 						if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12391 							so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12392 				}
12393 #ifdef NETFLIX_SB_LIMITS
12394 				appended =
12395 #endif
12396 					sbappendstream_locked(&so->so_rcv, m, 0);
12397 			}
12398 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12399 			/* NB: sorwakeup_locked() does an implicit unlock. */
12400 			sorwakeup_locked(so);
12401 #ifdef NETFLIX_SB_LIMITS
12402 			if (so->so_rcv.sb_shlim && appended != mcnt)
12403 				counter_fo_release(so->so_rcv.sb_shlim,
12404 				    mcnt - appended);
12405 #endif
12406 		} else {
12407 			/*
12408 			 * XXX: Due to the header drop above "th" is
12409 			 * theoretically invalid by now.  Fortunately
12410 			 * m_adj() doesn't actually frees any mbufs when
12411 			 * trimming from the head.
12412 			 */
12413 			tcp_seq temp = save_start;
12414 
12415 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
12416 			tp->t_flags |= TF_ACKNOW;
12417 			if (tp->t_flags & TF_WAKESOR) {
12418 				tp->t_flags &= ~TF_WAKESOR;
12419 				/* NB: sorwakeup_locked() does an implicit unlock. */
12420 				sorwakeup_locked(so);
12421 			}
12422 		}
12423 		if ((tp->t_flags & TF_SACK_PERMIT) &&
12424 		    (save_tlen > 0) &&
12425 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
12426 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12427 				/*
12428 				 * DSACK actually handled in the fastpath
12429 				 * above.
12430 				 */
12431 				tcp_update_sack_list(tp, save_start,
12432 				    save_start + save_tlen);
12433 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12434 				if ((tp->rcv_numsacks >= 1) &&
12435 				    (tp->sackblks[0].end == save_start)) {
12436 					/*
12437 					 * Partial overlap, recorded at todrop
12438 					 * above.
12439 					 */
12440 					tcp_update_sack_list(tp,
12441 					    tp->sackblks[0].start,
12442 					    tp->sackblks[0].end);
12443 				} else {
12444 					tcp_update_dsack_list(tp, save_start,
12445 					    save_start + save_tlen);
12446 				}
12447 			} else if (tlen >= save_tlen) {
12448 				/* Update of sackblks. */
12449 				tcp_update_dsack_list(tp, save_start,
12450 				    save_start + save_tlen);
12451 			} else if (tlen > 0) {
12452 				tcp_update_dsack_list(tp, save_start,
12453 				    save_start + tlen);
12454 			}
12455 		}
12456 	} else {
12457 		m_freem(m);
12458 		thflags &= ~TH_FIN;
12459 	}
12460 
12461 	/*
12462 	 * If FIN is received ACK the FIN and let the user know that the
12463 	 * connection is closing.
12464 	 */
12465 	if (thflags & TH_FIN) {
12466 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12467 			/* The socket upcall is handled by socantrcvmore. */
12468 			socantrcvmore(so);
12469 			/*
12470 			 * If connection is half-synchronized (ie NEEDSYN
12471 			 * flag on) then delay ACK, so it may be piggybacked
12472 			 * when SYN is sent. Otherwise, since we received a
12473 			 * FIN then no more input can be expected, send ACK
12474 			 * now.
12475 			 */
12476 			if (tp->t_flags & TF_NEEDSYN) {
12477 				rack_timer_cancel(tp, rack,
12478 				    rack->r_ctl.rc_rcvtime, __LINE__);
12479 				tp->t_flags |= TF_DELACK;
12480 			} else {
12481 				tp->t_flags |= TF_ACKNOW;
12482 			}
12483 			tp->rcv_nxt++;
12484 		}
12485 		switch (tp->t_state) {
12486 			/*
12487 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
12488 			 * CLOSE_WAIT state.
12489 			 */
12490 		case TCPS_SYN_RECEIVED:
12491 			tp->t_starttime = ticks;
12492 			/* FALLTHROUGH */
12493 		case TCPS_ESTABLISHED:
12494 			rack_timer_cancel(tp, rack,
12495 			    rack->r_ctl.rc_rcvtime, __LINE__);
12496 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
12497 			break;
12498 
12499 			/*
12500 			 * If still in FIN_WAIT_1 STATE FIN has not been
12501 			 * acked so enter the CLOSING state.
12502 			 */
12503 		case TCPS_FIN_WAIT_1:
12504 			rack_timer_cancel(tp, rack,
12505 			    rack->r_ctl.rc_rcvtime, __LINE__);
12506 			tcp_state_change(tp, TCPS_CLOSING);
12507 			break;
12508 
12509 			/*
12510 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
12511 			 * starting the time-wait timer, turning off the
12512 			 * other standard timers.
12513 			 */
12514 		case TCPS_FIN_WAIT_2:
12515 			rack_timer_cancel(tp, rack,
12516 			    rack->r_ctl.rc_rcvtime, __LINE__);
12517 			tcp_twstart(tp);
12518 			return (1);
12519 		}
12520 	}
12521 	/*
12522 	 * Return any desired output.
12523 	 */
12524 	if ((tp->t_flags & TF_ACKNOW) ||
12525 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12526 		rack->r_wanted_output = 1;
12527 	}
12528 	return (0);
12529 }
12530 
12531 /*
12532  * Here nothing is really faster, its just that we
12533  * have broken out the fast-data path also just like
12534  * the fast-ack.
12535  */
12536 static int
12537 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12538     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12539     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12540 {
12541 	int32_t nsegs;
12542 	int32_t newsize = 0;	/* automatic sockbuf scaling */
12543 	struct tcp_rack *rack;
12544 #ifdef NETFLIX_SB_LIMITS
12545 	u_int mcnt, appended;
12546 #endif
12547 
12548 	/*
12549 	 * If last ACK falls within this segment's sequence numbers, record
12550 	 * the timestamp. NOTE that the test is modified according to the
12551 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12552 	 */
12553 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12554 		return (0);
12555 	}
12556 	if (tiwin && tiwin != tp->snd_wnd) {
12557 		return (0);
12558 	}
12559 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12560 		return (0);
12561 	}
12562 	if (__predict_false((to->to_flags & TOF_TS) &&
12563 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12564 		return (0);
12565 	}
12566 	if (__predict_false((th->th_ack != tp->snd_una))) {
12567 		return (0);
12568 	}
12569 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12570 		return (0);
12571 	}
12572 	if ((to->to_flags & TOF_TS) != 0 &&
12573 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12574 		tp->ts_recent_age = tcp_ts_getticks();
12575 		tp->ts_recent = to->to_tsval;
12576 	}
12577 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12578 	/*
12579 	 * This is a pure, in-sequence data packet with nothing on the
12580 	 * reassembly queue and we have enough buffer space to take it.
12581 	 */
12582 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12583 
12584 #ifdef NETFLIX_SB_LIMITS
12585 	if (so->so_rcv.sb_shlim) {
12586 		mcnt = m_memcnt(m);
12587 		appended = 0;
12588 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12589 		    CFO_NOSLEEP, NULL) == false) {
12590 			counter_u64_add(tcp_sb_shlim_fails, 1);
12591 			m_freem(m);
12592 			return (1);
12593 		}
12594 	}
12595 #endif
12596 	/* Clean receiver SACK report if present */
12597 	if (tp->rcv_numsacks)
12598 		tcp_clean_sackreport(tp);
12599 	KMOD_TCPSTAT_INC(tcps_preddat);
12600 	tp->rcv_nxt += tlen;
12601 	if (tlen &&
12602 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12603 	    (tp->t_fbyte_in == 0)) {
12604 		tp->t_fbyte_in = ticks;
12605 		if (tp->t_fbyte_in == 0)
12606 			tp->t_fbyte_in = 1;
12607 		if (tp->t_fbyte_out && tp->t_fbyte_in)
12608 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12609 	}
12610 	/*
12611 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12612 	 */
12613 	tp->snd_wl1 = th->th_seq;
12614 	/*
12615 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12616 	 */
12617 	tp->rcv_up = tp->rcv_nxt;
12618 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12619 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12620 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12621 
12622 	/* Add data to socket buffer. */
12623 	SOCKBUF_LOCK(&so->so_rcv);
12624 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12625 		m_freem(m);
12626 	} else {
12627 		/*
12628 		 * Set new socket buffer size. Give up when limit is
12629 		 * reached.
12630 		 */
12631 		if (newsize)
12632 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12633 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12634 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12635 #ifdef NETFLIX_SB_LIMITS
12636 		appended =
12637 #endif
12638 			sbappendstream_locked(&so->so_rcv, m, 0);
12639 		ctf_calc_rwin(so, tp);
12640 	}
12641 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12642 	/* NB: sorwakeup_locked() does an implicit unlock. */
12643 	sorwakeup_locked(so);
12644 #ifdef NETFLIX_SB_LIMITS
12645 	if (so->so_rcv.sb_shlim && mcnt != appended)
12646 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12647 #endif
12648 	rack_handle_delayed_ack(tp, rack, tlen, 0);
12649 	if (tp->snd_una == tp->snd_max)
12650 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12651 	return (1);
12652 }
12653 
12654 /*
12655  * This subfunction is used to try to highly optimize the
12656  * fast path. We again allow window updates that are
12657  * in sequence to remain in the fast-path. We also add
12658  * in the __predict's to attempt to help the compiler.
12659  * Note that if we return a 0, then we can *not* process
12660  * it and the caller should push the packet into the
12661  * slow-path.
12662  */
12663 static int
12664 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12665     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12666     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12667 {
12668 	int32_t acked;
12669 	int32_t nsegs;
12670 	int32_t under_pacing = 0;
12671 	struct tcp_rack *rack;
12672 
12673 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12674 		/* Old ack, behind (or duplicate to) the last one rcv'd */
12675 		return (0);
12676 	}
12677 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
12678 		/* Above what we have sent? */
12679 		return (0);
12680 	}
12681 	if (__predict_false(tiwin == 0)) {
12682 		/* zero window */
12683 		return (0);
12684 	}
12685 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
12686 		/* We need a SYN or a FIN, unlikely.. */
12687 		return (0);
12688 	}
12689 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
12690 		/* Timestamp is behind .. old ack with seq wrap? */
12691 		return (0);
12692 	}
12693 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
12694 		/* Still recovering */
12695 		return (0);
12696 	}
12697 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12698 	if (rack->r_ctl.rc_sacked) {
12699 		/* We have sack holes on our scoreboard */
12700 		return (0);
12701 	}
12702 	/* Ok if we reach here, we can process a fast-ack */
12703 	if (rack->gp_ready &&
12704 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12705 		under_pacing = 1;
12706 	}
12707 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12708 	rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
12709 	/* Did the window get updated? */
12710 	if (tiwin != tp->snd_wnd) {
12711 		tp->snd_wnd = tiwin;
12712 		rack_validate_fo_sendwin_up(tp, rack);
12713 		tp->snd_wl1 = th->th_seq;
12714 		if (tp->snd_wnd > tp->max_sndwnd)
12715 			tp->max_sndwnd = tp->snd_wnd;
12716 	}
12717 	/* Do we exit persists? */
12718 	if ((rack->rc_in_persist != 0) &&
12719 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12720 			       rack->r_ctl.rc_pace_min_segs))) {
12721 		rack_exit_persist(tp, rack, cts);
12722 	}
12723 	/* Do we enter persists? */
12724 	if ((rack->rc_in_persist == 0) &&
12725 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12726 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12727 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12728 	    sbavail(&tptosocket(tp)->so_snd) &&
12729 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12730 		/*
12731 		 * Here the rwnd is less than
12732 		 * the pacing size, we are established,
12733 		 * nothing is outstanding, and there is
12734 		 * data to send. Enter persists.
12735 		 */
12736 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
12737 	}
12738 	/*
12739 	 * If last ACK falls within this segment's sequence numbers, record
12740 	 * the timestamp. NOTE that the test is modified according to the
12741 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12742 	 */
12743 	if ((to->to_flags & TOF_TS) != 0 &&
12744 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12745 		tp->ts_recent_age = tcp_ts_getticks();
12746 		tp->ts_recent = to->to_tsval;
12747 	}
12748 	/*
12749 	 * This is a pure ack for outstanding data.
12750 	 */
12751 	KMOD_TCPSTAT_INC(tcps_predack);
12752 
12753 	/*
12754 	 * "bad retransmit" recovery.
12755 	 */
12756 	if ((tp->t_flags & TF_PREVVALID) &&
12757 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12758 		tp->t_flags &= ~TF_PREVVALID;
12759 		if (tp->t_rxtshift == 1 &&
12760 		    (int)(ticks - tp->t_badrxtwin) < 0)
12761 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12762 	}
12763 	/*
12764 	 * Recalculate the transmit timer / rtt.
12765 	 *
12766 	 * Some boxes send broken timestamp replies during the SYN+ACK
12767 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
12768 	 * and blow up the retransmit timer.
12769 	 */
12770 	acked = BYTES_THIS_ACK(tp, th);
12771 
12772 #ifdef TCP_HHOOK
12773 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
12774 	hhook_run_tcp_est_in(tp, th, to);
12775 #endif
12776 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12777 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12778 	if (acked) {
12779 		struct mbuf *mfree;
12780 
12781 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
12782 		SOCKBUF_LOCK(&so->so_snd);
12783 		mfree = sbcut_locked(&so->so_snd, acked);
12784 		tp->snd_una = th->th_ack;
12785 		/* Note we want to hold the sb lock through the sendmap adjust */
12786 		rack_adjust_sendmap_head(rack, &so->so_snd);
12787 		/* Wake up the socket if we have room to write more */
12788 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12789 		sowwakeup_locked(so);
12790 		m_freem(mfree);
12791 		tp->t_rxtshift = 0;
12792 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12793 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12794 		rack->rc_tlp_in_progress = 0;
12795 		rack->r_ctl.rc_tlp_cnt_out = 0;
12796 		/*
12797 		 * If it is the RXT timer we want to
12798 		 * stop it, so we can restart a TLP.
12799 		 */
12800 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12801 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12802 
12803 #ifdef TCP_REQUEST_TRK
12804 		rack_req_check_for_comp(rack, th->th_ack);
12805 #endif
12806 	}
12807 	/*
12808 	 * Let the congestion control algorithm update congestion control
12809 	 * related information. This typically means increasing the
12810 	 * congestion window.
12811 	 */
12812 	if (tp->snd_wnd < ctf_outstanding(tp)) {
12813 		/* The peer collapsed the window */
12814 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12815 	} else if (rack->rc_has_collapsed)
12816 		rack_un_collapse_window(rack, __LINE__);
12817 	if ((rack->r_collapse_point_valid) &&
12818 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
12819 		rack->r_collapse_point_valid = 0;
12820 	/*
12821 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
12822 	 */
12823 	tp->snd_wl2 = th->th_ack;
12824 	tp->t_dupacks = 0;
12825 	m_freem(m);
12826 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
12827 
12828 	/*
12829 	 * If all outstanding data are acked, stop retransmit timer,
12830 	 * otherwise restart timer using current (possibly backed-off)
12831 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
12832 	 * If data are ready to send, let tcp_output decide between more
12833 	 * output or persist.
12834 	 */
12835 	if (under_pacing &&
12836 	    (rack->use_fixed_rate == 0) &&
12837 	    (rack->in_probe_rtt == 0) &&
12838 	    rack->rc_gp_dyn_mul &&
12839 	    rack->rc_always_pace) {
12840 		/* Check if we are dragging bottom */
12841 		rack_check_bottom_drag(tp, rack, so);
12842 	}
12843 	if (tp->snd_una == tp->snd_max) {
12844 		tp->t_flags &= ~TF_PREVVALID;
12845 		rack->r_ctl.retran_during_recovery = 0;
12846 		rack->rc_suspicious = 0;
12847 		rack->r_ctl.dsack_byte_cnt = 0;
12848 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12849 		if (rack->r_ctl.rc_went_idle_time == 0)
12850 			rack->r_ctl.rc_went_idle_time = 1;
12851 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12852 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12853 			tp->t_acktime = 0;
12854 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12855 	}
12856 	if (acked && rack->r_fast_output)
12857 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
12858 	if (sbavail(&so->so_snd)) {
12859 		rack->r_wanted_output = 1;
12860 	}
12861 	return (1);
12862 }
12863 
12864 /*
12865  * Return value of 1, the TCB is unlocked and most
12866  * likely gone, return value of 0, the TCP is still
12867  * locked.
12868  */
12869 static int
12870 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
12871     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12872     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12873 {
12874 	int32_t ret_val = 0;
12875 	int32_t orig_tlen = tlen;
12876 	int32_t todrop;
12877 	int32_t ourfinisacked = 0;
12878 	struct tcp_rack *rack;
12879 
12880 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12881 
12882 	ctf_calc_rwin(so, tp);
12883 	/*
12884 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
12885 	 * SYN, drop the input. if seg contains a RST, then drop the
12886 	 * connection. if seg does not contain SYN, then drop it. Otherwise
12887 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
12888 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
12889 	 * contains an ECE and ECN support is enabled, the stream is ECN
12890 	 * capable. if SYN has been acked change to ESTABLISHED else
12891 	 * SYN_RCVD state arrange for segment to be acked (eventually)
12892 	 * continue processing rest of data/controls.
12893 	 */
12894 	if ((thflags & TH_ACK) &&
12895 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
12896 	    SEQ_GT(th->th_ack, tp->snd_max))) {
12897 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
12898 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12899 		return (1);
12900 	}
12901 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
12902 		TCP_PROBE5(connect__refused, NULL, tp,
12903 		    mtod(m, const char *), tp, th);
12904 		tp = tcp_drop(tp, ECONNREFUSED);
12905 		ctf_do_drop(m, tp);
12906 		return (1);
12907 	}
12908 	if (thflags & TH_RST) {
12909 		ctf_do_drop(m, tp);
12910 		return (1);
12911 	}
12912 	if (!(thflags & TH_SYN)) {
12913 		ctf_do_drop(m, tp);
12914 		return (1);
12915 	}
12916 	tp->irs = th->th_seq;
12917 	tcp_rcvseqinit(tp);
12918 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12919 	if (thflags & TH_ACK) {
12920 		int tfo_partial = 0;
12921 
12922 		KMOD_TCPSTAT_INC(tcps_connects);
12923 		soisconnected(so);
12924 #ifdef MAC
12925 		mac_socketpeer_set_from_mbuf(m, so);
12926 #endif
12927 		/* Do window scaling on this connection? */
12928 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12929 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12930 			tp->rcv_scale = tp->request_r_scale;
12931 		}
12932 		tp->rcv_adv += min(tp->rcv_wnd,
12933 		    TCP_MAXWIN << tp->rcv_scale);
12934 		/*
12935 		 * If not all the data that was sent in the TFO SYN
12936 		 * has been acked, resend the remainder right away.
12937 		 */
12938 		if ((tp->t_flags & TF_FASTOPEN) &&
12939 		    (tp->snd_una != tp->snd_max)) {
12940 			/* Was it a partial ack? */
12941 			if (SEQ_LT(th->th_ack, tp->snd_max))
12942 				tfo_partial = 1;
12943 		}
12944 		/*
12945 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
12946 		 * will be turned on later.
12947 		 */
12948 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
12949 			rack_timer_cancel(tp, rack,
12950 					  rack->r_ctl.rc_rcvtime, __LINE__);
12951 			tp->t_flags |= TF_DELACK;
12952 		} else {
12953 			rack->r_wanted_output = 1;
12954 			tp->t_flags |= TF_ACKNOW;
12955 		}
12956 
12957 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
12958 
12959 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
12960 			/*
12961 			 * We advance snd_una for the
12962 			 * fast open case. If th_ack is
12963 			 * acknowledging data beyond
12964 			 * snd_una we can't just call
12965 			 * ack-processing since the
12966 			 * data stream in our send-map
12967 			 * will start at snd_una + 1 (one
12968 			 * beyond the SYN). If its just
12969 			 * equal we don't need to do that
12970 			 * and there is no send_map.
12971 			 */
12972 			tp->snd_una++;
12973 			if (tfo_partial && (SEQ_GT(tp->snd_max, tp->snd_una))) {
12974 				/*
12975 				 * We sent a SYN with data, and thus have a
12976 				 * sendmap entry with a SYN set. Lets find it
12977 				 * and take off the send bit and the byte and
12978 				 * set it up to be what we send (send it next).
12979 				 */
12980 				struct rack_sendmap *rsm;
12981 
12982 				rsm = tqhash_min(rack->r_ctl.tqh);
12983 				if (rsm) {
12984 					if (rsm->r_flags & RACK_HAS_SYN) {
12985 						rsm->r_flags &= ~RACK_HAS_SYN;
12986 						rsm->r_start++;
12987 					}
12988 					rack->r_ctl.rc_resend = rsm;
12989 				}
12990 			}
12991 		}
12992 		/*
12993 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
12994 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
12995 		 */
12996 		tp->t_starttime = ticks;
12997 		if (tp->t_flags & TF_NEEDFIN) {
12998 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
12999 			tp->t_flags &= ~TF_NEEDFIN;
13000 			thflags &= ~TH_SYN;
13001 		} else {
13002 			tcp_state_change(tp, TCPS_ESTABLISHED);
13003 			TCP_PROBE5(connect__established, NULL, tp,
13004 			    mtod(m, const char *), tp, th);
13005 			rack_cc_conn_init(tp);
13006 		}
13007 	} else {
13008 		/*
13009 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
13010 		 * open.  If segment contains CC option and there is a
13011 		 * cached CC, apply TAO test. If it succeeds, connection is *
13012 		 * half-synchronized. Otherwise, do 3-way handshake:
13013 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
13014 		 * there was no CC option, clear cached CC value.
13015 		 */
13016 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
13017 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
13018 	}
13019 	/*
13020 	 * Advance th->th_seq to correspond to first data byte. If data,
13021 	 * trim to stay within window, dropping FIN if necessary.
13022 	 */
13023 	th->th_seq++;
13024 	if (tlen > tp->rcv_wnd) {
13025 		todrop = tlen - tp->rcv_wnd;
13026 		m_adj(m, -todrop);
13027 		tlen = tp->rcv_wnd;
13028 		thflags &= ~TH_FIN;
13029 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13030 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13031 	}
13032 	tp->snd_wl1 = th->th_seq - 1;
13033 	tp->rcv_up = th->th_seq;
13034 	/*
13035 	 * Client side of transaction: already sent SYN and data. If the
13036 	 * remote host used T/TCP to validate the SYN, our data will be
13037 	 * ACK'd; if so, enter normal data segment processing in the middle
13038 	 * of step 5, ack processing. Otherwise, goto step 6.
13039 	 */
13040 	if (thflags & TH_ACK) {
13041 		/* For syn-sent we need to possibly update the rtt */
13042 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13043 			uint32_t t, mcts;
13044 
13045 			mcts = tcp_ts_getticks();
13046 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13047 			if (!tp->t_rttlow || tp->t_rttlow > t)
13048 				tp->t_rttlow = t;
13049 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13050 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13051 			tcp_rack_xmit_timer_commit(rack, tp);
13052 		}
13053 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen))
13054 			return (ret_val);
13055 		/* We may have changed to FIN_WAIT_1 above */
13056 		if (tp->t_state == TCPS_FIN_WAIT_1) {
13057 			/*
13058 			 * In FIN_WAIT_1 STATE in addition to the processing
13059 			 * for the ESTABLISHED state if our FIN is now
13060 			 * acknowledged then enter FIN_WAIT_2.
13061 			 */
13062 			if (ourfinisacked) {
13063 				/*
13064 				 * If we can't receive any more data, then
13065 				 * closing user can proceed. Starting the
13066 				 * timer is contrary to the specification,
13067 				 * but if we don't get a FIN we'll hang
13068 				 * forever.
13069 				 *
13070 				 * XXXjl: we should release the tp also, and
13071 				 * use a compressed state.
13072 				 */
13073 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13074 					soisdisconnected(so);
13075 					tcp_timer_activate(tp, TT_2MSL,
13076 					    (tcp_fast_finwait2_recycle ?
13077 					    tcp_finwait2_timeout :
13078 					    TP_MAXIDLE(tp)));
13079 				}
13080 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13081 			}
13082 		}
13083 	}
13084 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13085 	   tiwin, thflags, nxt_pkt));
13086 }
13087 
13088 /*
13089  * Return value of 1, the TCB is unlocked and most
13090  * likely gone, return value of 0, the TCP is still
13091  * locked.
13092  */
13093 static int
13094 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13095     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13096     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13097 {
13098 	struct tcp_rack *rack;
13099 	int32_t orig_tlen = tlen;
13100 	int32_t ret_val = 0;
13101 	int32_t ourfinisacked = 0;
13102 
13103 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13104 	ctf_calc_rwin(so, tp);
13105 	if ((thflags & TH_RST) ||
13106 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13107 		return (ctf_process_rst(m, th, so, tp));
13108 	if ((thflags & TH_ACK) &&
13109 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13110 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13111 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13112 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13113 		return (1);
13114 	}
13115 	if (tp->t_flags & TF_FASTOPEN) {
13116 		/*
13117 		 * When a TFO connection is in SYN_RECEIVED, the
13118 		 * only valid packets are the initial SYN, a
13119 		 * retransmit/copy of the initial SYN (possibly with
13120 		 * a subset of the original data), a valid ACK, a
13121 		 * FIN, or a RST.
13122 		 */
13123 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13124 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13125 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13126 			return (1);
13127 		} else if (thflags & TH_SYN) {
13128 			/* non-initial SYN is ignored */
13129 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13130 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13131 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13132 				ctf_do_drop(m, NULL);
13133 				return (0);
13134 			}
13135 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13136 			ctf_do_drop(m, NULL);
13137 			return (0);
13138 		}
13139 	}
13140 
13141 	/*
13142 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13143 	 * it's less than ts_recent, drop it.
13144 	 */
13145 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13146 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13147 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13148 			return (ret_val);
13149 	}
13150 	/*
13151 	 * In the SYN-RECEIVED state, validate that the packet belongs to
13152 	 * this connection before trimming the data to fit the receive
13153 	 * window.  Check the sequence number versus IRS since we know the
13154 	 * sequence numbers haven't wrapped.  This is a partial fix for the
13155 	 * "LAND" DoS attack.
13156 	 */
13157 	if (SEQ_LT(th->th_seq, tp->irs)) {
13158 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13159 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13160 		return (1);
13161 	}
13162 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13163 		return (ret_val);
13164 	}
13165 	/*
13166 	 * If last ACK falls within this segment's sequence numbers, record
13167 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13168 	 * from the latest proposal of the tcplw@cray.com list (Braden
13169 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13170 	 * with our earlier PAWS tests, so this check should be solely
13171 	 * predicated on the sequence space of this segment. 3) That we
13172 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13173 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13174 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13175 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13176 	 * p.869. In such cases, we can still calculate the RTT correctly
13177 	 * when RCV.NXT == Last.ACK.Sent.
13178 	 */
13179 	if ((to->to_flags & TOF_TS) != 0 &&
13180 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13181 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13182 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13183 		tp->ts_recent_age = tcp_ts_getticks();
13184 		tp->ts_recent = to->to_tsval;
13185 	}
13186 	tp->snd_wnd = tiwin;
13187 	rack_validate_fo_sendwin_up(tp, rack);
13188 	/*
13189 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13190 	 * is on (half-synchronized state), then queue data for later
13191 	 * processing; else drop segment and return.
13192 	 */
13193 	if ((thflags & TH_ACK) == 0) {
13194 		if (tp->t_flags & TF_FASTOPEN) {
13195 			rack_cc_conn_init(tp);
13196 		}
13197 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13198 		    tiwin, thflags, nxt_pkt));
13199 	}
13200 	KMOD_TCPSTAT_INC(tcps_connects);
13201 	if (tp->t_flags & TF_SONOTCONN) {
13202 		tp->t_flags &= ~TF_SONOTCONN;
13203 		soisconnected(so);
13204 	}
13205 	/* Do window scaling? */
13206 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13207 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13208 		tp->rcv_scale = tp->request_r_scale;
13209 	}
13210 	/*
13211 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
13212 	 * FIN-WAIT-1
13213 	 */
13214 	tp->t_starttime = ticks;
13215 	if ((tp->t_flags & TF_FASTOPEN) && tp->t_tfo_pending) {
13216 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13217 		tp->t_tfo_pending = NULL;
13218 	}
13219 	if (tp->t_flags & TF_NEEDFIN) {
13220 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
13221 		tp->t_flags &= ~TF_NEEDFIN;
13222 	} else {
13223 		tcp_state_change(tp, TCPS_ESTABLISHED);
13224 		TCP_PROBE5(accept__established, NULL, tp,
13225 		    mtod(m, const char *), tp, th);
13226 		/*
13227 		 * TFO connections call cc_conn_init() during SYN
13228 		 * processing.  Calling it again here for such connections
13229 		 * is not harmless as it would undo the snd_cwnd reduction
13230 		 * that occurs when a TFO SYN|ACK is retransmitted.
13231 		 */
13232 		if (!(tp->t_flags & TF_FASTOPEN))
13233 			rack_cc_conn_init(tp);
13234 	}
13235 	/*
13236 	 * Account for the ACK of our SYN prior to
13237 	 * regular ACK processing below, except for
13238 	 * simultaneous SYN, which is handled later.
13239 	 */
13240 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13241 		tp->snd_una++;
13242 	/*
13243 	 * If segment contains data or ACK, will call tcp_reass() later; if
13244 	 * not, do so now to pass queued data to user.
13245 	 */
13246 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
13247 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13248 		    (struct mbuf *)0);
13249 		if (tp->t_flags & TF_WAKESOR) {
13250 			tp->t_flags &= ~TF_WAKESOR;
13251 			/* NB: sorwakeup_locked() does an implicit unlock. */
13252 			sorwakeup_locked(so);
13253 		}
13254 	}
13255 	tp->snd_wl1 = th->th_seq - 1;
13256 	/* For syn-recv we need to possibly update the rtt */
13257 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13258 		uint32_t t, mcts;
13259 
13260 		mcts = tcp_ts_getticks();
13261 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13262 		if (!tp->t_rttlow || tp->t_rttlow > t)
13263 			tp->t_rttlow = t;
13264 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13265 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13266 		tcp_rack_xmit_timer_commit(rack, tp);
13267 	}
13268 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13269 		return (ret_val);
13270 	}
13271 	if (tp->t_state == TCPS_FIN_WAIT_1) {
13272 		/* We could have went to FIN_WAIT_1 (or EST) above */
13273 		/*
13274 		 * In FIN_WAIT_1 STATE in addition to the processing for the
13275 		 * ESTABLISHED state if our FIN is now acknowledged then
13276 		 * enter FIN_WAIT_2.
13277 		 */
13278 		if (ourfinisacked) {
13279 			/*
13280 			 * If we can't receive any more data, then closing
13281 			 * user can proceed. Starting the timer is contrary
13282 			 * to the specification, but if we don't get a FIN
13283 			 * we'll hang forever.
13284 			 *
13285 			 * XXXjl: we should release the tp also, and use a
13286 			 * compressed state.
13287 			 */
13288 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13289 				soisdisconnected(so);
13290 				tcp_timer_activate(tp, TT_2MSL,
13291 				    (tcp_fast_finwait2_recycle ?
13292 				    tcp_finwait2_timeout :
13293 				    TP_MAXIDLE(tp)));
13294 			}
13295 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
13296 		}
13297 	}
13298 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13299 	    tiwin, thflags, nxt_pkt));
13300 }
13301 
13302 /*
13303  * Return value of 1, the TCB is unlocked and most
13304  * likely gone, return value of 0, the TCP is still
13305  * locked.
13306  */
13307 static int
13308 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13309     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13310     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13311 {
13312 	int32_t ret_val = 0;
13313 	int32_t orig_tlen = tlen;
13314 	struct tcp_rack *rack;
13315 
13316 	/*
13317 	 * Header prediction: check for the two common cases of a
13318 	 * uni-directional data xfer.  If the packet has no control flags,
13319 	 * is in-sequence, the window didn't change and we're not
13320 	 * retransmitting, it's a candidate.  If the length is zero and the
13321 	 * ack moved forward, we're the sender side of the xfer.  Just free
13322 	 * the data acked & wake any higher level process that was blocked
13323 	 * waiting for space.  If the length is non-zero and the ack didn't
13324 	 * move, we're the receiver side.  If we're getting packets in-order
13325 	 * (the reassembly queue is empty), add the data toc The socket
13326 	 * buffer and note that we need a delayed ack. Make sure that the
13327 	 * hidden state-flags are also off. Since we check for
13328 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13329 	 */
13330 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13331 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13332 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13333 	    __predict_true(SEGQ_EMPTY(tp)) &&
13334 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
13335 		if (tlen == 0) {
13336 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13337 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13338 				return (0);
13339 			}
13340 		} else {
13341 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13342 			    tiwin, nxt_pkt, iptos)) {
13343 				return (0);
13344 			}
13345 		}
13346 	}
13347 	ctf_calc_rwin(so, tp);
13348 
13349 	if ((thflags & TH_RST) ||
13350 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13351 		return (ctf_process_rst(m, th, so, tp));
13352 
13353 	/*
13354 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13355 	 * synchronized state.
13356 	 */
13357 	if (thflags & TH_SYN) {
13358 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13359 		return (ret_val);
13360 	}
13361 	/*
13362 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13363 	 * it's less than ts_recent, drop it.
13364 	 */
13365 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13366 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13367 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13368 			return (ret_val);
13369 	}
13370 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13371 		return (ret_val);
13372 	}
13373 	/*
13374 	 * If last ACK falls within this segment's sequence numbers, record
13375 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13376 	 * from the latest proposal of the tcplw@cray.com list (Braden
13377 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13378 	 * with our earlier PAWS tests, so this check should be solely
13379 	 * predicated on the sequence space of this segment. 3) That we
13380 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13381 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13382 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13383 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13384 	 * p.869. In such cases, we can still calculate the RTT correctly
13385 	 * when RCV.NXT == Last.ACK.Sent.
13386 	 */
13387 	if ((to->to_flags & TOF_TS) != 0 &&
13388 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13389 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13390 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13391 		tp->ts_recent_age = tcp_ts_getticks();
13392 		tp->ts_recent = to->to_tsval;
13393 	}
13394 	/*
13395 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13396 	 * is on (half-synchronized state), then queue data for later
13397 	 * processing; else drop segment and return.
13398 	 */
13399 	if ((thflags & TH_ACK) == 0) {
13400 		if (tp->t_flags & TF_NEEDSYN) {
13401 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13402 			    tiwin, thflags, nxt_pkt));
13403 
13404 		} else if (tp->t_flags & TF_ACKNOW) {
13405 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13406 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13407 			return (ret_val);
13408 		} else {
13409 			ctf_do_drop(m, NULL);
13410 			return (0);
13411 		}
13412 	}
13413 	/*
13414 	 * Ack processing.
13415 	 */
13416 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13417 		return (ret_val);
13418 	}
13419 	if (sbavail(&so->so_snd)) {
13420 		if (ctf_progress_timeout_check(tp, true)) {
13421 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13422 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13423 			return (1);
13424 		}
13425 	}
13426 	/* State changes only happen in rack_process_data() */
13427 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13428 	    tiwin, thflags, nxt_pkt));
13429 }
13430 
13431 /*
13432  * Return value of 1, the TCB is unlocked and most
13433  * likely gone, return value of 0, the TCP is still
13434  * locked.
13435  */
13436 static int
13437 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13438     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13439     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13440 {
13441 	int32_t ret_val = 0;
13442 	int32_t orig_tlen = tlen;
13443 
13444 	ctf_calc_rwin(so, tp);
13445 	if ((thflags & TH_RST) ||
13446 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13447 		return (ctf_process_rst(m, th, so, tp));
13448 	/*
13449 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13450 	 * synchronized state.
13451 	 */
13452 	if (thflags & TH_SYN) {
13453 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13454 		return (ret_val);
13455 	}
13456 	/*
13457 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13458 	 * it's less than ts_recent, drop it.
13459 	 */
13460 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13461 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13462 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13463 			return (ret_val);
13464 	}
13465 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13466 		return (ret_val);
13467 	}
13468 	/*
13469 	 * If last ACK falls within this segment's sequence numbers, record
13470 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13471 	 * from the latest proposal of the tcplw@cray.com list (Braden
13472 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13473 	 * with our earlier PAWS tests, so this check should be solely
13474 	 * predicated on the sequence space of this segment. 3) That we
13475 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13476 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13477 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13478 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13479 	 * p.869. In such cases, we can still calculate the RTT correctly
13480 	 * when RCV.NXT == Last.ACK.Sent.
13481 	 */
13482 	if ((to->to_flags & TOF_TS) != 0 &&
13483 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13484 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13485 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13486 		tp->ts_recent_age = tcp_ts_getticks();
13487 		tp->ts_recent = to->to_tsval;
13488 	}
13489 	/*
13490 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13491 	 * is on (half-synchronized state), then queue data for later
13492 	 * processing; else drop segment and return.
13493 	 */
13494 	if ((thflags & TH_ACK) == 0) {
13495 		if (tp->t_flags & TF_NEEDSYN) {
13496 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13497 			    tiwin, thflags, nxt_pkt));
13498 
13499 		} else if (tp->t_flags & TF_ACKNOW) {
13500 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13501 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13502 			return (ret_val);
13503 		} else {
13504 			ctf_do_drop(m, NULL);
13505 			return (0);
13506 		}
13507 	}
13508 	/*
13509 	 * Ack processing.
13510 	 */
13511 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13512 		return (ret_val);
13513 	}
13514 	if (sbavail(&so->so_snd)) {
13515 		if (ctf_progress_timeout_check(tp, true)) {
13516 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13517 						tp, tick, PROGRESS_DROP, __LINE__);
13518 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13519 			return (1);
13520 		}
13521 	}
13522 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13523 	    tiwin, thflags, nxt_pkt));
13524 }
13525 
13526 static int
13527 rack_check_data_after_close(struct mbuf *m,
13528     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13529 {
13530 	struct tcp_rack *rack;
13531 
13532 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13533 	if (rack->rc_allow_data_af_clo == 0) {
13534 	close_now:
13535 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13536 		/* tcp_close will kill the inp pre-log the Reset */
13537 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13538 		tp = tcp_close(tp);
13539 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13540 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
13541 		return (1);
13542 	}
13543 	if (sbavail(&so->so_snd) == 0)
13544 		goto close_now;
13545 	/* Ok we allow data that is ignored and a followup reset */
13546 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13547 	tp->rcv_nxt = th->th_seq + *tlen;
13548 	tp->t_flags2 |= TF2_DROP_AF_DATA;
13549 	rack->r_wanted_output = 1;
13550 	*tlen = 0;
13551 	return (0);
13552 }
13553 
13554 /*
13555  * Return value of 1, the TCB is unlocked and most
13556  * likely gone, return value of 0, the TCP is still
13557  * locked.
13558  */
13559 static int
13560 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13561     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13562     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13563 {
13564 	int32_t ret_val = 0;
13565 	int32_t orig_tlen = tlen;
13566 	int32_t ourfinisacked = 0;
13567 
13568 	ctf_calc_rwin(so, tp);
13569 
13570 	if ((thflags & TH_RST) ||
13571 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13572 		return (ctf_process_rst(m, th, so, tp));
13573 	/*
13574 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13575 	 * synchronized state.
13576 	 */
13577 	if (thflags & TH_SYN) {
13578 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13579 		return (ret_val);
13580 	}
13581 	/*
13582 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13583 	 * it's less than ts_recent, drop it.
13584 	 */
13585 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13586 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13587 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13588 			return (ret_val);
13589 	}
13590 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13591 		return (ret_val);
13592 	}
13593 	/*
13594 	 * If new data are received on a connection after the user processes
13595 	 * are gone, then RST the other end.
13596 	 */
13597 	if ((tp->t_flags & TF_CLOSED) && tlen &&
13598 	    rack_check_data_after_close(m, tp, &tlen, th, so))
13599 		return (1);
13600 	/*
13601 	 * If last ACK falls within this segment's sequence numbers, record
13602 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13603 	 * from the latest proposal of the tcplw@cray.com list (Braden
13604 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13605 	 * with our earlier PAWS tests, so this check should be solely
13606 	 * predicated on the sequence space of this segment. 3) That we
13607 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13608 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13609 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13610 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13611 	 * p.869. In such cases, we can still calculate the RTT correctly
13612 	 * when RCV.NXT == Last.ACK.Sent.
13613 	 */
13614 	if ((to->to_flags & TOF_TS) != 0 &&
13615 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13616 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13617 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13618 		tp->ts_recent_age = tcp_ts_getticks();
13619 		tp->ts_recent = to->to_tsval;
13620 	}
13621 	/*
13622 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13623 	 * is on (half-synchronized state), then queue data for later
13624 	 * processing; else drop segment and return.
13625 	 */
13626 	if ((thflags & TH_ACK) == 0) {
13627 		if (tp->t_flags & TF_NEEDSYN) {
13628 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13629 			    tiwin, thflags, nxt_pkt));
13630 		} else if (tp->t_flags & TF_ACKNOW) {
13631 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13632 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13633 			return (ret_val);
13634 		} else {
13635 			ctf_do_drop(m, NULL);
13636 			return (0);
13637 		}
13638 	}
13639 	/*
13640 	 * Ack processing.
13641 	 */
13642 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13643 		return (ret_val);
13644 	}
13645 	if (ourfinisacked) {
13646 		/*
13647 		 * If we can't receive any more data, then closing user can
13648 		 * proceed. Starting the timer is contrary to the
13649 		 * specification, but if we don't get a FIN we'll hang
13650 		 * forever.
13651 		 *
13652 		 * XXXjl: we should release the tp also, and use a
13653 		 * compressed state.
13654 		 */
13655 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13656 			soisdisconnected(so);
13657 			tcp_timer_activate(tp, TT_2MSL,
13658 			    (tcp_fast_finwait2_recycle ?
13659 			    tcp_finwait2_timeout :
13660 			    TP_MAXIDLE(tp)));
13661 		}
13662 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
13663 	}
13664 	if (sbavail(&so->so_snd)) {
13665 		if (ctf_progress_timeout_check(tp, true)) {
13666 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13667 						tp, tick, PROGRESS_DROP, __LINE__);
13668 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13669 			return (1);
13670 		}
13671 	}
13672 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13673 	    tiwin, thflags, nxt_pkt));
13674 }
13675 
13676 /*
13677  * Return value of 1, the TCB is unlocked and most
13678  * likely gone, return value of 0, the TCP is still
13679  * locked.
13680  */
13681 static int
13682 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
13683     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13684     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13685 {
13686 	int32_t ret_val = 0;
13687 	int32_t orig_tlen = tlen;
13688 	int32_t ourfinisacked = 0;
13689 
13690 	ctf_calc_rwin(so, tp);
13691 
13692 	if ((thflags & TH_RST) ||
13693 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13694 		return (ctf_process_rst(m, th, so, tp));
13695 	/*
13696 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13697 	 * synchronized state.
13698 	 */
13699 	if (thflags & TH_SYN) {
13700 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13701 		return (ret_val);
13702 	}
13703 	/*
13704 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13705 	 * it's less than ts_recent, drop it.
13706 	 */
13707 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13708 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13709 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13710 			return (ret_val);
13711 	}
13712 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13713 		return (ret_val);
13714 	}
13715 	/*
13716 	 * If last ACK falls within this segment's sequence numbers, record
13717 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13718 	 * from the latest proposal of the tcplw@cray.com list (Braden
13719 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13720 	 * with our earlier PAWS tests, so this check should be solely
13721 	 * predicated on the sequence space of this segment. 3) That we
13722 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13723 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13724 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13725 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13726 	 * p.869. In such cases, we can still calculate the RTT correctly
13727 	 * when RCV.NXT == Last.ACK.Sent.
13728 	 */
13729 	if ((to->to_flags & TOF_TS) != 0 &&
13730 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13731 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13732 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13733 		tp->ts_recent_age = tcp_ts_getticks();
13734 		tp->ts_recent = to->to_tsval;
13735 	}
13736 	/*
13737 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13738 	 * is on (half-synchronized state), then queue data for later
13739 	 * processing; else drop segment and return.
13740 	 */
13741 	if ((thflags & TH_ACK) == 0) {
13742 		if (tp->t_flags & TF_NEEDSYN) {
13743 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13744 			    tiwin, thflags, nxt_pkt));
13745 		} else if (tp->t_flags & TF_ACKNOW) {
13746 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13747 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13748 			return (ret_val);
13749 		} else {
13750 			ctf_do_drop(m, NULL);
13751 			return (0);
13752 		}
13753 	}
13754 	/*
13755 	 * Ack processing.
13756 	 */
13757 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13758 		return (ret_val);
13759 	}
13760 	if (ourfinisacked) {
13761 		tcp_twstart(tp);
13762 		m_freem(m);
13763 		return (1);
13764 	}
13765 	if (sbavail(&so->so_snd)) {
13766 		if (ctf_progress_timeout_check(tp, true)) {
13767 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13768 						tp, tick, PROGRESS_DROP, __LINE__);
13769 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13770 			return (1);
13771 		}
13772 	}
13773 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13774 	    tiwin, thflags, nxt_pkt));
13775 }
13776 
13777 /*
13778  * Return value of 1, the TCB is unlocked and most
13779  * likely gone, return value of 0, the TCP is still
13780  * locked.
13781  */
13782 static int
13783 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13784     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13785     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13786 {
13787 	int32_t ret_val = 0;
13788 	int32_t orig_tlen;
13789 	int32_t ourfinisacked = 0;
13790 
13791 	ctf_calc_rwin(so, tp);
13792 
13793 	if ((thflags & TH_RST) ||
13794 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13795 		return (ctf_process_rst(m, th, so, tp));
13796 	/*
13797 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13798 	 * synchronized state.
13799 	 */
13800 	if (thflags & TH_SYN) {
13801 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13802 		return (ret_val);
13803 	}
13804 	/*
13805 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13806 	 * it's less than ts_recent, drop it.
13807 	 */
13808 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13809 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13810 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13811 			return (ret_val);
13812 	}
13813 	orig_tlen = tlen;
13814 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13815 		return (ret_val);
13816 	}
13817 	/*
13818 	 * If last ACK falls within this segment's sequence numbers, record
13819 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13820 	 * from the latest proposal of the tcplw@cray.com list (Braden
13821 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13822 	 * with our earlier PAWS tests, so this check should be solely
13823 	 * predicated on the sequence space of this segment. 3) That we
13824 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13825 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13826 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13827 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13828 	 * p.869. In such cases, we can still calculate the RTT correctly
13829 	 * when RCV.NXT == Last.ACK.Sent.
13830 	 */
13831 	if ((to->to_flags & TOF_TS) != 0 &&
13832 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13833 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13834 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13835 		tp->ts_recent_age = tcp_ts_getticks();
13836 		tp->ts_recent = to->to_tsval;
13837 	}
13838 	/*
13839 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13840 	 * is on (half-synchronized state), then queue data for later
13841 	 * processing; else drop segment and return.
13842 	 */
13843 	if ((thflags & TH_ACK) == 0) {
13844 		if (tp->t_flags & TF_NEEDSYN) {
13845 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13846 			    tiwin, thflags, nxt_pkt));
13847 		} else if (tp->t_flags & TF_ACKNOW) {
13848 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13849 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13850 			return (ret_val);
13851 		} else {
13852 			ctf_do_drop(m, NULL);
13853 			return (0);
13854 		}
13855 	}
13856 	/*
13857 	 * case TCPS_LAST_ACK: Ack processing.
13858 	 */
13859 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13860 		return (ret_val);
13861 	}
13862 	if (ourfinisacked) {
13863 		tp = tcp_close(tp);
13864 		ctf_do_drop(m, tp);
13865 		return (1);
13866 	}
13867 	if (sbavail(&so->so_snd)) {
13868 		if (ctf_progress_timeout_check(tp, true)) {
13869 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13870 						tp, tick, PROGRESS_DROP, __LINE__);
13871 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13872 			return (1);
13873 		}
13874 	}
13875 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13876 	    tiwin, thflags, nxt_pkt));
13877 }
13878 
13879 /*
13880  * Return value of 1, the TCB is unlocked and most
13881  * likely gone, return value of 0, the TCP is still
13882  * locked.
13883  */
13884 static int
13885 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
13886     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13887     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13888 {
13889 	int32_t ret_val = 0;
13890 	int32_t orig_tlen = tlen;
13891 	int32_t ourfinisacked = 0;
13892 
13893 	ctf_calc_rwin(so, tp);
13894 
13895 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
13896 	if ((thflags & TH_RST) ||
13897 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13898 		return (ctf_process_rst(m, th, so, tp));
13899 	/*
13900 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13901 	 * synchronized state.
13902 	 */
13903 	if (thflags & TH_SYN) {
13904 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13905 		return (ret_val);
13906 	}
13907 	/*
13908 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13909 	 * it's less than ts_recent, drop it.
13910 	 */
13911 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13912 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13913 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13914 			return (ret_val);
13915 	}
13916 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13917 		return (ret_val);
13918 	}
13919 	/*
13920 	 * If new data are received on a connection after the user processes
13921 	 * are gone, then RST the other end.
13922 	 */
13923 	if ((tp->t_flags & TF_CLOSED) && tlen &&
13924 	    rack_check_data_after_close(m, tp, &tlen, th, so))
13925 		return (1);
13926 	/*
13927 	 * If last ACK falls within this segment's sequence numbers, record
13928 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13929 	 * from the latest proposal of the tcplw@cray.com list (Braden
13930 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13931 	 * with our earlier PAWS tests, so this check should be solely
13932 	 * predicated on the sequence space of this segment. 3) That we
13933 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13934 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13935 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13936 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13937 	 * p.869. In such cases, we can still calculate the RTT correctly
13938 	 * when RCV.NXT == Last.ACK.Sent.
13939 	 */
13940 	if ((to->to_flags & TOF_TS) != 0 &&
13941 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13942 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13943 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13944 		tp->ts_recent_age = tcp_ts_getticks();
13945 		tp->ts_recent = to->to_tsval;
13946 	}
13947 	/*
13948 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13949 	 * is on (half-synchronized state), then queue data for later
13950 	 * processing; else drop segment and return.
13951 	 */
13952 	if ((thflags & TH_ACK) == 0) {
13953 		if (tp->t_flags & TF_NEEDSYN) {
13954 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13955 			    tiwin, thflags, nxt_pkt));
13956 		} else if (tp->t_flags & TF_ACKNOW) {
13957 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13958 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13959 			return (ret_val);
13960 		} else {
13961 			ctf_do_drop(m, NULL);
13962 			return (0);
13963 		}
13964 	}
13965 	/*
13966 	 * Ack processing.
13967 	 */
13968 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13969 		return (ret_val);
13970 	}
13971 	if (sbavail(&so->so_snd)) {
13972 		if (ctf_progress_timeout_check(tp, true)) {
13973 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13974 						tp, tick, PROGRESS_DROP, __LINE__);
13975 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13976 			return (1);
13977 		}
13978 	}
13979 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13980 	    tiwin, thflags, nxt_pkt));
13981 }
13982 
13983 static void inline
13984 rack_clear_rate_sample(struct tcp_rack *rack)
13985 {
13986 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
13987 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
13988 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
13989 }
13990 
13991 static void
13992 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
13993 {
13994 	uint64_t bw_est, rate_wanted;
13995 	int chged = 0;
13996 	uint32_t user_max, orig_min, orig_max;
13997 
13998 #ifdef TCP_REQUEST_TRK
13999 	if (rack->rc_hybrid_mode &&
14000 	    (rack->r_ctl.rc_pace_max_segs != 0) &&
14001 	    (rack_hybrid_allow_set_maxseg == 1) &&
14002 	    (rack->r_ctl.rc_last_sft != NULL)) {
14003 		rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
14004 		return;
14005 	}
14006 #endif
14007 	orig_min = rack->r_ctl.rc_pace_min_segs;
14008 	orig_max = rack->r_ctl.rc_pace_max_segs;
14009 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
14010 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
14011 		chged = 1;
14012 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
14013 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
14014 		if (user_max != rack->r_ctl.rc_pace_max_segs)
14015 			chged = 1;
14016 	}
14017 	if (rack->rc_force_max_seg) {
14018 		rack->r_ctl.rc_pace_max_segs = user_max;
14019 	} else if (rack->use_fixed_rate) {
14020 		bw_est = rack_get_bw(rack);
14021 		if ((rack->r_ctl.crte == NULL) ||
14022 		    (bw_est != rack->r_ctl.crte->rate)) {
14023 			rack->r_ctl.rc_pace_max_segs = user_max;
14024 		} else {
14025 			/* We are pacing right at the hardware rate */
14026 			uint32_t segsiz, pace_one;
14027 
14028 			if (rack_pace_one_seg ||
14029 			    (rack->r_ctl.rc_user_set_min_segs == 1))
14030 				pace_one = 1;
14031 			else
14032 				pace_one = 0;
14033 			segsiz = min(ctf_fixed_maxseg(tp),
14034 				     rack->r_ctl.rc_pace_min_segs);
14035 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14036 				tp, bw_est, segsiz, pace_one,
14037 				rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14038 		}
14039 	} else if (rack->rc_always_pace) {
14040 		if (rack->r_ctl.gp_bw ||
14041 		    rack->r_ctl.init_rate) {
14042 			/* We have a rate of some sort set */
14043 			uint32_t  orig;
14044 
14045 			bw_est = rack_get_bw(rack);
14046 			orig = rack->r_ctl.rc_pace_max_segs;
14047 			if (fill_override)
14048 				rate_wanted = *fill_override;
14049 			else
14050 				rate_wanted = rack_get_gp_est(rack);
14051 			if (rate_wanted) {
14052 				/* We have something */
14053 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14054 										   rate_wanted,
14055 										   ctf_fixed_maxseg(rack->rc_tp));
14056 			} else
14057 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14058 			if (orig != rack->r_ctl.rc_pace_max_segs)
14059 				chged = 1;
14060 		} else if ((rack->r_ctl.gp_bw == 0) &&
14061 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
14062 			/*
14063 			 * If we have nothing limit us to bursting
14064 			 * out IW sized pieces.
14065 			 */
14066 			chged = 1;
14067 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14068 		}
14069 	}
14070 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14071 		chged = 1;
14072 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14073 	}
14074 	if (chged)
14075 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14076 }
14077 
14078 
14079 static void
14080 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14081 {
14082 #ifdef INET6
14083 	struct ip6_hdr *ip6 = NULL;
14084 #endif
14085 #ifdef INET
14086 	struct ip *ip = NULL;
14087 #endif
14088 	struct udphdr *udp = NULL;
14089 
14090 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
14091 #ifdef INET6
14092 	if (rack->r_is_v6) {
14093 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14094 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14095 		if (tp->t_port) {
14096 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14097 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14098 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14099 			udp->uh_dport = tp->t_port;
14100 			rack->r_ctl.fsb.udp = udp;
14101 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14102 		} else
14103 		{
14104 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14105 			rack->r_ctl.fsb.udp = NULL;
14106 		}
14107 		tcpip_fillheaders(rack->rc_inp,
14108 				  tp->t_port,
14109 				  ip6, rack->r_ctl.fsb.th);
14110 		rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14111 	} else
14112 #endif				/* INET6 */
14113 #ifdef INET
14114 	{
14115 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14116 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14117 		if (tp->t_port) {
14118 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14119 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14120 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14121 			udp->uh_dport = tp->t_port;
14122 			rack->r_ctl.fsb.udp = udp;
14123 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14124 		} else
14125 		{
14126 			rack->r_ctl.fsb.udp = NULL;
14127 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14128 		}
14129 		tcpip_fillheaders(rack->rc_inp,
14130 				  tp->t_port,
14131 				  ip, rack->r_ctl.fsb.th);
14132 		rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14133 	}
14134 #endif
14135 	rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14136 	    (long)TCP_MAXWIN << tp->rcv_scale);
14137 	rack->r_fsb_inited = 1;
14138 }
14139 
14140 static int
14141 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14142 {
14143 	/*
14144 	 * Allocate the larger of spaces V6 if available else just
14145 	 * V4 and include udphdr (overbook)
14146 	 */
14147 #ifdef INET6
14148 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14149 #else
14150 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14151 #endif
14152 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14153 					    M_TCPFSB, M_NOWAIT|M_ZERO);
14154 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14155 		return (ENOMEM);
14156 	}
14157 	rack->r_fsb_inited = 0;
14158 	return (0);
14159 }
14160 
14161 static void
14162 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14163 {
14164 	/*
14165 	 * Types of logs (mod value)
14166 	 * 20 - Initial round setup
14167 	 * 21 - Rack declares a new round.
14168 	 */
14169 	struct tcpcb *tp;
14170 
14171 	tp = rack->rc_tp;
14172 	if (tcp_bblogging_on(tp)) {
14173 		union tcp_log_stackspecific log;
14174 		struct timeval tv;
14175 
14176 		memset(&log, 0, sizeof(log));
14177 		log.u_bbr.flex1 = rack->r_ctl.current_round;
14178 		log.u_bbr.flex2 = rack->r_ctl.roundends;
14179 		log.u_bbr.flex3 = high_seq;
14180 		log.u_bbr.flex4 = tp->snd_max;
14181 		log.u_bbr.flex8 = mod;
14182 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14183 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14184 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14185 		TCP_LOG_EVENTP(tp, NULL,
14186 		    &tptosocket(tp)->so_rcv,
14187 		    &tptosocket(tp)->so_snd,
14188 		    TCP_HYSTART, 0,
14189 		    0, &log, false, &tv);
14190 	}
14191 }
14192 
14193 static void
14194 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14195 {
14196 	rack->rack_deferred_inited = 1;
14197 	rack->r_ctl.roundends = tp->snd_max;
14198 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14199 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14200 }
14201 
14202 static void
14203 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14204 {
14205 	/* Retransmit bit controls.
14206 	 *
14207 	 * The setting of these values control one of
14208 	 * three settings you can have and dictate
14209 	 * how rack does retransmissions. Note this
14210 	 * is in *any* mode i.e. pacing on or off DGP
14211 	 * fixed rate pacing, or just bursting rack.
14212 	 *
14213 	 * 1 - Use full sized retransmits i.e. limit
14214 	 *     the size to whatever the pace_max_segments
14215 	 *     size is.
14216 	 *
14217 	 * 2 - Use pacer min granularity as a guide to
14218 	 *     the size combined with the current calculated
14219 	 *     goodput b/w measurement. So for example if
14220 	 *     the goodput is measured at 20Mbps we would
14221 	 *     calculate 8125 (pacer minimum 250usec in
14222 	 *     that b/w) and then round it up to the next
14223 	 *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14224 	 *
14225 	 * 0 - The rack default 1 MSS (anything not 0/1/2
14226 	 *     fall here too if we are setting via rack_init()).
14227 	 *
14228 	 */
14229 	if (ctl == 1) {
14230 		rack->full_size_rxt = 1;
14231 		rack->shape_rxt_to_pacing_min  = 0;
14232 	} else if (ctl == 2) {
14233 		rack->full_size_rxt = 0;
14234 		rack->shape_rxt_to_pacing_min  = 1;
14235 	} else {
14236 		rack->full_size_rxt = 0;
14237 		rack->shape_rxt_to_pacing_min  = 0;
14238 	}
14239 }
14240 
14241 static void
14242 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14243 		  uint32_t flex1,
14244 		  uint32_t flex2,
14245 		  uint32_t flex3)
14246 {
14247 	if (tcp_bblogging_on(rack->rc_tp)) {
14248 		union tcp_log_stackspecific log;
14249 		struct timeval tv;
14250 
14251 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14252 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14253 		log.u_bbr.flex8 = mod;
14254 		log.u_bbr.flex1 = flex1;
14255 		log.u_bbr.flex2 = flex2;
14256 		log.u_bbr.flex3 = flex3;
14257 		tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14258 			       0, &log, false, NULL, __func__, __LINE__, &tv);
14259 	}
14260 }
14261 
14262 static int
14263 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14264 {
14265 	struct tcp_rack *rack;
14266 	struct rack_sendmap *rsm;
14267 	int i;
14268 
14269 
14270 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14271 	switch (reqr->req) {
14272 	case TCP_QUERY_SENDMAP:
14273 		if ((reqr->req_param == tp->snd_max) ||
14274 		    (tp->snd_max == tp->snd_una)){
14275 			/* Unlikely */
14276 			return (0);
14277 		}
14278 		rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14279 		if (rsm == NULL) {
14280 			/* Can't find that seq -- unlikely */
14281 			return (0);
14282 		}
14283 		reqr->sendmap_start = rsm->r_start;
14284 		reqr->sendmap_end = rsm->r_end;
14285 		reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14286 		reqr->sendmap_fas = rsm->r_fas;
14287 		if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14288 			reqr->sendmap_send_cnt = SNDMAP_NRTX;
14289 		for(i=0; i<reqr->sendmap_send_cnt; i++)
14290 			reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14291 		reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14292 		reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14293 		reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14294 		reqr->sendmap_dupacks = rsm->r_dupack;
14295 		rack_log_chg_info(tp, rack, 1,
14296 				  rsm->r_start,
14297 				  rsm->r_end,
14298 				  rsm->r_flags);
14299 		return(1);
14300 		break;
14301 	case TCP_QUERY_TIMERS_UP:
14302 		if (rack->r_ctl.rc_hpts_flags == 0) {
14303 			/* no timers up */
14304 			return (0);
14305 		}
14306 		reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14307 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14308 			reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14309 		}
14310 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14311 			reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14312 		}
14313 		rack_log_chg_info(tp, rack, 2,
14314 				  rack->r_ctl.rc_hpts_flags,
14315 				  rack->r_ctl.rc_last_output_to,
14316 				  rack->r_ctl.rc_timer_exp);
14317 		return (1);
14318 		break;
14319 	case TCP_QUERY_RACK_TIMES:
14320 		/* Reordering items */
14321 		reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14322 		reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14323 		/* Timerstamps and timers */
14324 		reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14325 		reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14326 		reqr->rack_rtt = rack->rc_rack_rtt;
14327 		reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14328 		reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14329 		/* PRR data */
14330 		reqr->rack_sacked = rack->r_ctl.rc_sacked;
14331 		reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14332 		reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14333 		reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14334 		reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14335 		reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14336 		/* TLP and persists info */
14337 		reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14338 		reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14339 		if (rack->rc_in_persist) {
14340 			reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14341 			reqr->rack_in_persist = 1;
14342 		} else {
14343 			reqr->rack_time_went_idle = 0;
14344 			reqr->rack_in_persist = 0;
14345 		}
14346 		if (rack->r_wanted_output)
14347 			reqr->rack_wanted_output = 1;
14348 		else
14349 			reqr->rack_wanted_output = 0;
14350 		return (1);
14351 		break;
14352 	default:
14353 		return (-EINVAL);
14354 	}
14355 }
14356 
14357 static void
14358 rack_switch_failed(struct tcpcb *tp)
14359 {
14360 	/*
14361 	 * This method gets called if a stack switch was
14362 	 * attempted and it failed. We are left
14363 	 * but our hpts timers were stopped and we
14364 	 * need to validate time units and t_flags2.
14365 	 */
14366 	struct tcp_rack *rack;
14367 	struct timeval tv;
14368 	uint32_t cts;
14369 	uint32_t toval;
14370 	struct hpts_diag diag;
14371 
14372 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14373 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14374 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14375 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14376 	else
14377 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14378 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14379 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
14380 	if (tp->t_in_hpts > IHPTS_NONE) {
14381 		/* Strange */
14382 		return;
14383 	}
14384 	cts = tcp_get_usecs(&tv);
14385 	if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14386 		if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14387 			toval = rack->r_ctl.rc_last_output_to - cts;
14388 		} else {
14389 			/* one slot please */
14390 			toval = HPTS_TICKS_PER_SLOT;
14391 		}
14392 	} else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14393 		if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14394 			toval = rack->r_ctl.rc_timer_exp - cts;
14395 		} else {
14396 			/* one slot please */
14397 			toval = HPTS_TICKS_PER_SLOT;
14398 		}
14399 	} else
14400 		toval = HPTS_TICKS_PER_SLOT;
14401 	(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14402 				   __LINE__, &diag);
14403 	rack_log_hpts_diag(rack, cts, &diag, &tv);
14404 }
14405 
14406 static int
14407 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14408 {
14409 	struct rack_sendmap *rsm, *ersm;
14410 	int insret __diagused;
14411 	/*
14412 	 * When initing outstanding, we must be quite careful
14413 	 * to not refer to tp->t_fb_ptr. This has the old rack
14414 	 * pointer in it, not the "new" one (when we are doing
14415 	 * a stack switch).
14416 	 */
14417 
14418 
14419 	if (tp->t_fb->tfb_chg_query == NULL) {
14420 		/* Create a send map for the current outstanding data */
14421 
14422 		rsm = rack_alloc(rack);
14423 		if (rsm == NULL) {
14424 			uma_zfree(rack_pcb_zone, ptr);
14425 			return (ENOMEM);
14426 		}
14427 		rsm->r_no_rtt_allowed = 1;
14428 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14429 		rsm->r_rtr_cnt = 1;
14430 		rsm->r_rtr_bytes = 0;
14431 		if (tp->t_flags & TF_SENTFIN)
14432 			rsm->r_flags |= RACK_HAS_FIN;
14433 		rsm->r_end = tp->snd_max;
14434 		if (tp->snd_una == tp->iss) {
14435 			/* The data space is one beyond snd_una */
14436 			rsm->r_flags |= RACK_HAS_SYN;
14437 			rsm->r_start = tp->iss;
14438 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14439 		} else
14440 			rsm->r_start = tp->snd_una;
14441 		rsm->r_dupack = 0;
14442 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14443 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14444 			if (rsm->m) {
14445 				rsm->orig_m_len = rsm->m->m_len;
14446 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14447 			} else {
14448 				rsm->orig_m_len = 0;
14449 				rsm->orig_t_space = 0;
14450 			}
14451 		} else {
14452 			/*
14453 			 * This can happen if we have a stand-alone FIN or
14454 			 *  SYN.
14455 			 */
14456 			rsm->m = NULL;
14457 			rsm->orig_m_len = 0;
14458 			rsm->orig_t_space = 0;
14459 			rsm->soff = 0;
14460 		}
14461 #ifdef INVARIANTS
14462 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14463 			panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14464 			      insret, rack, rsm);
14465 		}
14466 #else
14467 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14468 #endif
14469 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14470 		rsm->r_in_tmap = 1;
14471 	} else {
14472 		/* We have a query mechanism, lets use it */
14473 		struct tcp_query_resp qr;
14474 		int i;
14475 		tcp_seq at;
14476 
14477 		at = tp->snd_una;
14478 		while (at != tp->snd_max) {
14479 			memset(&qr, 0, sizeof(qr));
14480 			qr.req = TCP_QUERY_SENDMAP;
14481 			qr.req_param = at;
14482 			if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14483 				break;
14484 			/* Move forward */
14485 			at = qr.sendmap_end;
14486 			/* Now lets build the entry for this one */
14487 			rsm = rack_alloc(rack);
14488 			if (rsm == NULL) {
14489 				uma_zfree(rack_pcb_zone, ptr);
14490 				return (ENOMEM);
14491 			}
14492 			memset(rsm, 0, sizeof(struct rack_sendmap));
14493 			/* Now configure the rsm and insert it */
14494 			rsm->r_dupack = qr.sendmap_dupacks;
14495 			rsm->r_start = qr.sendmap_start;
14496 			rsm->r_end = qr.sendmap_end;
14497 			if (qr.sendmap_fas)
14498 				rsm->r_fas = qr.sendmap_end;
14499 			else
14500 				rsm->r_fas = rsm->r_start - tp->snd_una;
14501 			/*
14502 			 * We have carefully aligned the bits
14503 			 * so that all we have to do is copy over
14504 			 * the bits with the mask.
14505 			 */
14506 			rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14507 			rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14508 			rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14509 			rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14510 			for (i=0 ; i<rsm->r_rtr_cnt; i++)
14511 				rsm->r_tim_lastsent[i]	= qr.sendmap_time[i];
14512 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14513 					   (rsm->r_start - tp->snd_una), &rsm->soff);
14514 			if (rsm->m) {
14515 				rsm->orig_m_len = rsm->m->m_len;
14516 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14517 			} else {
14518 				rsm->orig_m_len = 0;
14519 				rsm->orig_t_space = 0;
14520 			}
14521 #ifdef INVARIANTS
14522 			if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14523 				panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14524 				      insret, rack, rsm);
14525 			}
14526 #else
14527 			(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14528 #endif
14529 			if ((rsm->r_flags & RACK_ACKED) == 0)  {
14530 				TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14531 					if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14532 					    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14533 						/*
14534 						 * If the existing ersm was sent at
14535 						 * a later time than the new one, then
14536 						 * the new one should appear ahead of this
14537 						 * ersm.
14538 						 */
14539 						rsm->r_in_tmap = 1;
14540 						TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14541 						break;
14542 					}
14543 				}
14544 				if (rsm->r_in_tmap == 0) {
14545 					/*
14546 					 * Not found so shove it on the tail.
14547 					 */
14548 					TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14549 					rsm->r_in_tmap = 1;
14550 				}
14551  			} else {
14552 				if ((rack->r_ctl.rc_sacklast == NULL) ||
14553 				    (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14554 					rack->r_ctl.rc_sacklast = rsm;
14555 				}
14556 			}
14557 			rack_log_chg_info(tp, rack, 3,
14558 					  rsm->r_start,
14559 					  rsm->r_end,
14560 					  rsm->r_flags);
14561 		}
14562 	}
14563 	return (0);
14564 }
14565 
14566 
14567 static int32_t
14568 rack_init(struct tcpcb *tp, void **ptr)
14569 {
14570 	struct inpcb *inp = tptoinpcb(tp);
14571 	struct tcp_rack *rack = NULL;
14572 	uint32_t iwin, snt, us_cts;
14573 	size_t sz;
14574 	int err, no_query;
14575 
14576 	tcp_hpts_init(tp);
14577 
14578 	/*
14579 	 * First are we the initial or are we a switched stack?
14580 	 * If we are initing via tcp_newtcppcb the ptr passed
14581 	 * will be tp->t_fb_ptr. If its a stack switch that
14582 	 * has a previous stack we can query it will be a local
14583 	 * var that will in the end be set into t_fb_ptr.
14584 	 */
14585 	if (ptr == &tp->t_fb_ptr)
14586 		no_query = 1;
14587 	else
14588 		no_query = 0;
14589 	*ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14590 	if (*ptr == NULL) {
14591 		/*
14592 		 * We need to allocate memory but cant. The INP and INP_INFO
14593 		 * locks and they are recursive (happens during setup. So a
14594 		 * scheme to drop the locks fails :(
14595 		 *
14596 		 */
14597 		return(ENOMEM);
14598 	}
14599 	memset(*ptr, 0, sizeof(struct tcp_rack));
14600 	rack = (struct tcp_rack *)*ptr;
14601 	rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
14602 	if (rack->r_ctl.tqh == NULL) {
14603 		uma_zfree(rack_pcb_zone, rack);
14604 		return(ENOMEM);
14605 	}
14606 	tqhash_init(rack->r_ctl.tqh);
14607 	TAILQ_INIT(&rack->r_ctl.rc_free);
14608 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
14609 	rack->rc_tp = tp;
14610 	rack->rc_inp = inp;
14611 	/* Set the flag */
14612 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
14613 	/* Probably not needed but lets be sure */
14614 	rack_clear_rate_sample(rack);
14615 	/*
14616 	 * Save off the default values, socket options will poke
14617 	 * at these if pacing is not on or we have not yet
14618 	 * reached where pacing is on (gp_ready/fixed enabled).
14619 	 * When they get set into the CC module (when gp_ready
14620 	 * is enabled or we enable fixed) then we will set these
14621 	 * values into the CC and place in here the old values
14622 	 * so we have a restoral. Then we will set the flag
14623 	 * rc_pacing_cc_set. That way whenever we turn off pacing
14624 	 * or switch off this stack, we will know to go restore
14625 	 * the saved values.
14626 	 *
14627 	 * We specifically put into the beta the ecn value for pacing.
14628 	 */
14629 	rack->rc_new_rnd_needed = 1;
14630 	rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
14631 	/* We want abe like behavior as well */
14632 
14633 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
14634 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
14635 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
14636 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
14637 	if (rack_fill_cw_state)
14638 		rack->rc_pace_to_cwnd = 1;
14639 	if (rack_pacing_min_seg)
14640 		rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
14641 	if (use_rack_rr)
14642 		rack->use_rack_rr = 1;
14643 	if (rack_dnd_default) {
14644 		rack->rc_pace_dnd = 1;
14645 	}
14646 	if (V_tcp_delack_enabled)
14647 		tp->t_delayed_ack = 1;
14648 	else
14649 		tp->t_delayed_ack = 0;
14650 #ifdef TCP_ACCOUNTING
14651 	if (rack_tcp_accounting) {
14652 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
14653 	}
14654 #endif
14655 	rack->r_ctl.pcm_i.cnt_alloc = RACK_DEFAULT_PCM_ARRAY;
14656 	sz = (sizeof(struct rack_pcm_stats) * rack->r_ctl.pcm_i.cnt_alloc);
14657 	rack->r_ctl.pcm_s = malloc(sz,M_TCPPCM, M_NOWAIT);
14658 	if (rack->r_ctl.pcm_s == NULL) {
14659 		rack->r_ctl.pcm_i.cnt_alloc = 0;
14660 	}
14661 #ifdef NETFLIX_STATS
14662 	rack->r_ctl.side_chan_dis_mask = tcp_sidechannel_disable_mask;
14663 #endif
14664 	rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
14665 	rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
14666 	if (rack_enable_shared_cwnd)
14667 		rack->rack_enable_scwnd = 1;
14668 	rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
14669 	rack->rc_user_set_max_segs = rack_hptsi_segments;
14670 	rack->r_ctl.max_reduction = rack_max_reduce;
14671 	rack->rc_force_max_seg = 0;
14672 	TAILQ_INIT(&rack->r_ctl.opt_list);
14673 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
14674 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
14675 	if (rack_hibeta_setting) {
14676 		rack->rack_hibeta = 1;
14677 		if ((rack_hibeta_setting >= 50) &&
14678 		    (rack_hibeta_setting <= 100)) {
14679 			rack->r_ctl.rc_saved_beta.beta = rack_hibeta_setting;
14680 			rack->r_ctl.saved_hibeta = rack_hibeta_setting;
14681 		}
14682 	} else {
14683 		rack->r_ctl.saved_hibeta = 50;
14684 	}
14685 	/*
14686 	 * We initialize to all ones so we never match 0
14687 	 * just in case the client sends in 0, it hopefully
14688 	 * will never have all 1's in ms :-)
14689 	 */
14690 	rack->r_ctl.last_tm_mark = 0xffffffffffffffff;
14691 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
14692 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
14693 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
14694 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
14695 	rack->r_ctl.rc_highest_us_rtt = 0;
14696 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
14697 	rack->pcm_enabled = rack_pcm_is_enabled;
14698 	if (rack_fillcw_bw_cap)
14699 		rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
14700 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
14701 	if (rack_use_cmp_acks)
14702 		rack->r_use_cmp_ack = 1;
14703 	if (rack_disable_prr)
14704 		rack->rack_no_prr = 1;
14705 	if (rack_gp_no_rec_chg)
14706 		rack->rc_gp_no_rec_chg = 1;
14707 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
14708 		rack->r_ctl.pacing_method |= RACK_REG_PACING;
14709 		rack->rc_always_pace = 1;
14710 		if (rack->rack_hibeta)
14711 			rack_set_cc_pacing(rack);
14712 	} else
14713 		rack->rc_always_pace = 0;
14714 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
14715 		rack->r_mbuf_queue = 1;
14716 	else
14717 		rack->r_mbuf_queue = 0;
14718 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
14719 	if (rack_limits_scwnd)
14720 		rack->r_limit_scw = 1;
14721 	else
14722 		rack->r_limit_scw = 0;
14723 	rack_init_retransmit_value(rack, rack_rxt_controls);
14724 	rack->rc_labc = V_tcp_abc_l_var;
14725 	if (rack_honors_hpts_min_to)
14726 		rack->r_use_hpts_min = 1;
14727 	if (tp->snd_una != 0) {
14728 		rack->rc_sendvars_notset = 0;
14729 		/*
14730 		 * Make sure any TCP timers are not running.
14731 		 */
14732 		tcp_timer_stop(tp);
14733 	} else {
14734 		/*
14735 		 * Server side, we are called from the
14736 		 * syn-cache. This means none of the
14737 		 * snd_una/max are set yet so we have
14738 		 * to defer this until the first send.
14739 		 */
14740 		rack->rc_sendvars_notset = 1;
14741 	}
14742 
14743 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
14744 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
14745 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
14746 	rack->r_ctl.rc_min_to = rack_min_to;
14747 	microuptime(&rack->r_ctl.act_rcv_time);
14748 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
14749 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
14750 	if (rack_hw_up_only)
14751 		rack->r_up_only = 1;
14752 	if (rack_do_dyn_mul) {
14753 		/* When dynamic adjustment is on CA needs to start at 100% */
14754 		rack->rc_gp_dyn_mul = 1;
14755 		if (rack_do_dyn_mul >= 100)
14756 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
14757 	} else
14758 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
14759 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
14760 	if (rack_timely_off) {
14761 		rack->rc_skip_timely = 1;
14762 	}
14763 	if (rack->rc_skip_timely) {
14764 		rack->r_ctl.rack_per_of_gp_rec = 90;
14765 		rack->r_ctl.rack_per_of_gp_ca = 100;
14766 		rack->r_ctl.rack_per_of_gp_ss = 250;
14767 	}
14768 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
14769 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
14770 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
14771 
14772 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
14773 				rack_probertt_filter_life);
14774 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14775 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
14776 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
14777 	rack->r_ctl.rc_went_idle_time = us_cts;
14778 	rack->r_ctl.rc_time_probertt_starts = 0;
14779 
14780 	rack->r_ctl.gp_rnd_thresh = rack_rnd_cnt_req & 0xff;
14781 	if (rack_rnd_cnt_req  & 0x10000)
14782 		rack->r_ctl.gate_to_fs = 1;
14783 	rack->r_ctl.gp_gain_req = rack_gp_gain_req;
14784 	if ((rack_rnd_cnt_req & 0x100) > 0) {
14785 
14786 	}
14787 	if (rack_dsack_std_based & 0x1) {
14788 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
14789 		rack->rc_rack_tmr_std_based = 1;
14790 	}
14791 	if (rack_dsack_std_based & 0x2) {
14792 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
14793 		rack->rc_rack_use_dsack = 1;
14794 	}
14795 	/* We require at least one measurement, even if the sysctl is 0 */
14796 	if (rack_req_measurements)
14797 		rack->r_ctl.req_measurements = rack_req_measurements;
14798 	else
14799 		rack->r_ctl.req_measurements = 1;
14800 	if (rack_enable_hw_pacing)
14801 		rack->rack_hdw_pace_ena = 1;
14802 	if (rack_hw_rate_caps)
14803 		rack->r_rack_hw_rate_caps = 1;
14804 	if (rack_non_rxt_use_cr)
14805 		rack->rack_rec_nonrxt_use_cr = 1;
14806 	/* Lets setup the fsb block */
14807 	err = rack_init_fsb(tp, rack);
14808 	if (err) {
14809 		uma_zfree(rack_pcb_zone, *ptr);
14810 		*ptr = NULL;
14811 		return (err);
14812 	}
14813 	if (rack_do_hystart) {
14814 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
14815 		if (rack_do_hystart > 1)
14816 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
14817 		if (rack_do_hystart > 2)
14818 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
14819 	}
14820 	/* Log what we will do with queries */
14821 	rack_log_chg_info(tp, rack, 7,
14822 			  no_query, 0, 0);
14823 	if (rack_def_profile)
14824 		rack_set_profile(rack, rack_def_profile);
14825 	/* Cancel the GP measurement in progress */
14826 	tp->t_flags &= ~TF_GPUTINPROG;
14827 	if ((tp->t_state != TCPS_CLOSED) &&
14828 	    (tp->t_state != TCPS_TIME_WAIT)) {
14829 		/*
14830 		 * We are already open, we may
14831 		 * need to adjust a few things.
14832 		 */
14833 		if (SEQ_GT(tp->snd_max, tp->iss))
14834 			snt = tp->snd_max - tp->iss;
14835 		else
14836 			snt = 0;
14837 		iwin = rc_init_window(rack);
14838 		if ((snt < iwin) &&
14839 		    (no_query == 1)) {
14840 			/* We are not past the initial window
14841 			 * on the first init (i.e. a stack switch
14842 			 * has not yet occured) so we need to make
14843 			 * sure cwnd and ssthresh is correct.
14844 			 */
14845 			if (tp->snd_cwnd < iwin)
14846 				tp->snd_cwnd = iwin;
14847 			/*
14848 			 * If we are within the initial window
14849 			 * we want ssthresh to be unlimited. Setting
14850 			 * it to the rwnd (which the default stack does
14851 			 * and older racks) is not really a good idea
14852 			 * since we want to be in SS and grow both the
14853 			 * cwnd and the rwnd (via dynamic rwnd growth). If
14854 			 * we set it to the rwnd then as the peer grows its
14855 			 * rwnd we will be stuck in CA and never hit SS.
14856 			 *
14857 			 * Its far better to raise it up high (this takes the
14858 			 * risk that there as been a loss already, probably
14859 			 * we should have an indicator in all stacks of loss
14860 			 * but we don't), but considering the normal use this
14861 			 * is a risk worth taking. The consequences of not
14862 			 * hitting SS are far worse than going one more time
14863 			 * into it early on (before we have sent even a IW).
14864 			 * It is highly unlikely that we will have had a loss
14865 			 * before getting the IW out.
14866 			 */
14867 			tp->snd_ssthresh = 0xffffffff;
14868 		}
14869 		/*
14870 		 * Any init based on sequence numbers
14871 		 * should be done in the deferred init path
14872 		 * since we can be CLOSED and not have them
14873 		 * inited when rack_init() is called. We
14874 		 * are not closed so lets call it.
14875 		 */
14876 		rack_deferred_init(tp, rack);
14877 	}
14878 	if ((tp->t_state != TCPS_CLOSED) &&
14879 	    (tp->t_state != TCPS_TIME_WAIT) &&
14880 	    (no_query == 0) &&
14881 	    (tp->snd_una != tp->snd_max))  {
14882 		err = rack_init_outstanding(tp, rack, us_cts, *ptr);
14883 		if (err) {
14884 			*ptr = NULL;
14885 			return(err);
14886 		}
14887 	}
14888 	rack_stop_all_timers(tp, rack);
14889 	/* Setup all the t_flags2 */
14890 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14891 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14892 	else
14893 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14894 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14895 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
14896 	/*
14897 	 * Timers in Rack are kept in microseconds so lets
14898 	 * convert any initial incoming variables
14899 	 * from ticks into usecs. Note that we
14900 	 * also change the values of t_srtt and t_rttvar, if
14901 	 * they are non-zero. They are kept with a 5
14902 	 * bit decimal so we have to carefully convert
14903 	 * these to get the full precision.
14904 	 */
14905 	rack_convert_rtts(tp);
14906 	rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
14907 	if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
14908 		/* We do not start any timers on DROPPED connections */
14909 		if (tp->t_fb->tfb_chg_query == NULL) {
14910 			rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14911 		} else {
14912 			struct tcp_query_resp qr;
14913 			int ret;
14914 
14915 			memset(&qr, 0, sizeof(qr));
14916 
14917 			/* Get the misc time stamps and such for rack */
14918 			qr.req = TCP_QUERY_RACK_TIMES;
14919 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14920 			if (ret == 1) {
14921 				rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
14922 				rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
14923 				rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
14924 				rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
14925 				rack->rc_rack_rtt = qr.rack_rtt;
14926 				rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
14927 				rack->r_ctl.rc_sacked = qr.rack_sacked;
14928 				rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
14929 				rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
14930 				rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
14931 				rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
14932 				rack->r_ctl.rc_prr_out = qr.rack_prr_out;
14933 				if (qr.rack_tlp_out) {
14934 					rack->rc_tlp_in_progress = 1;
14935 					rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
14936 				} else {
14937 					rack->rc_tlp_in_progress = 0;
14938 					rack->r_ctl.rc_tlp_cnt_out = 0;
14939 				}
14940 				if (qr.rack_srtt_measured)
14941 					rack->rc_srtt_measure_made = 1;
14942 				if (qr.rack_in_persist == 1) {
14943 					rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
14944 #ifdef NETFLIX_SHARED_CWND
14945 					if (rack->r_ctl.rc_scw) {
14946 						tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
14947 						rack->rack_scwnd_is_idle = 1;
14948 					}
14949 #endif
14950 					rack->r_ctl.persist_lost_ends = 0;
14951 					rack->probe_not_answered = 0;
14952 					rack->forced_ack = 0;
14953 					tp->t_rxtshift = 0;
14954 					rack->rc_in_persist = 1;
14955 					RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
14956 							   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
14957 				}
14958 				if (qr.rack_wanted_output)
14959 					rack->r_wanted_output = 1;
14960 				rack_log_chg_info(tp, rack, 6,
14961 						  qr.rack_min_rtt,
14962 						  qr.rack_rtt,
14963 						  qr.rack_reorder_ts);
14964 			}
14965 			/* Get the old stack timers */
14966 			qr.req_param = 0;
14967 			qr.req = TCP_QUERY_TIMERS_UP;
14968 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14969 			if (ret) {
14970 				/*
14971 				 * non-zero return means we have a timer('s)
14972 				 * to start. Zero means no timer (no keepalive
14973 				 * I suppose).
14974 				 */
14975 				uint32_t tov = 0;
14976 
14977 				rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
14978 				if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
14979 					rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
14980 					if (TSTMP_GT(qr.timer_pacing_to, us_cts))
14981 						tov = qr.timer_pacing_to - us_cts;
14982 					else
14983 						tov = HPTS_TICKS_PER_SLOT;
14984 				}
14985 				if (qr.timer_hpts_flags & PACE_TMR_MASK) {
14986 					rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
14987 					if (tov == 0) {
14988 						if (TSTMP_GT(qr.timer_timer_exp, us_cts))
14989 							tov = qr.timer_timer_exp - us_cts;
14990 						else
14991 							tov = HPTS_TICKS_PER_SLOT;
14992 					}
14993 				}
14994 				rack_log_chg_info(tp, rack, 4,
14995 						  rack->r_ctl.rc_hpts_flags,
14996 						  rack->r_ctl.rc_last_output_to,
14997 						  rack->r_ctl.rc_timer_exp);
14998 				if (tov) {
14999 					struct hpts_diag diag;
15000 
15001 					(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
15002 								   __LINE__, &diag);
15003 					rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
15004 				}
15005 			}
15006 		}
15007 		rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
15008 				     __LINE__, RACK_RTTS_INIT);
15009 	}
15010 	return (0);
15011 }
15012 
15013 static int
15014 rack_handoff_ok(struct tcpcb *tp)
15015 {
15016 	if ((tp->t_state == TCPS_CLOSED) ||
15017 	    (tp->t_state == TCPS_LISTEN)) {
15018 		/* Sure no problem though it may not stick */
15019 		return (0);
15020 	}
15021 	if ((tp->t_state == TCPS_SYN_SENT) ||
15022 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
15023 		/*
15024 		 * We really don't know if you support sack,
15025 		 * you have to get to ESTAB or beyond to tell.
15026 		 */
15027 		return (EAGAIN);
15028 	}
15029 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15030 		/*
15031 		 * Rack will only send a FIN after all data is acknowledged.
15032 		 * So in this case we have more data outstanding. We can't
15033 		 * switch stacks until either all data and only the FIN
15034 		 * is left (in which case rack_init() now knows how
15035 		 * to deal with that) <or> all is acknowledged and we
15036 		 * are only left with incoming data, though why you
15037 		 * would want to switch to rack after all data is acknowledged
15038 		 * I have no idea (rrs)!
15039 		 */
15040 		return (EAGAIN);
15041 	}
15042 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15043 		return (0);
15044 	}
15045 	/*
15046 	 * If we reach here we don't do SACK on this connection so we can
15047 	 * never do rack.
15048 	 */
15049 	return (EINVAL);
15050 }
15051 
15052 static void
15053 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15054 {
15055 
15056 	if (tp->t_fb_ptr) {
15057 		uint32_t cnt_free = 0;
15058 		struct tcp_rack *rack;
15059 		struct rack_sendmap *rsm;
15060 
15061 		tcp_handle_orphaned_packets(tp);
15062 		tp->t_flags &= ~TF_FORCEDATA;
15063 		rack = (struct tcp_rack *)tp->t_fb_ptr;
15064 		rack_log_pacing_delay_calc(rack,
15065 					   0,
15066 					   0,
15067 					   0,
15068 					   rack_get_gp_est(rack), /* delRate */
15069 					   rack_get_lt_bw(rack), /* rttProp */
15070 					   20, __LINE__, NULL, 0);
15071 #ifdef NETFLIX_SHARED_CWND
15072 		if (rack->r_ctl.rc_scw) {
15073 			uint32_t limit;
15074 
15075 			if (rack->r_limit_scw)
15076 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15077 			else
15078 				limit = 0;
15079 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15080 						  rack->r_ctl.rc_scw_index,
15081 						  limit);
15082 			rack->r_ctl.rc_scw = NULL;
15083 		}
15084 #endif
15085 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
15086 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15087 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15088 			rack->r_ctl.fsb.th = NULL;
15089 		}
15090 		if (rack->rc_always_pace == 1) {
15091 			rack_remove_pacing(rack);
15092 		}
15093 		/* Clean up any options if they were not applied */
15094 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15095 			struct deferred_opt_list *dol;
15096 
15097 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15098 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15099 			free(dol, M_TCPDO);
15100 		}
15101 		/* rack does not use force data but other stacks may clear it */
15102 		if (rack->r_ctl.crte != NULL) {
15103 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15104 			rack->rack_hdrw_pacing = 0;
15105 			rack->r_ctl.crte = NULL;
15106 		}
15107 #ifdef TCP_BLACKBOX
15108 		tcp_log_flowend(tp);
15109 #endif
15110 		/*
15111 		 * Lets take a different approach to purging just
15112 		 * get each one and free it like a cum-ack would and
15113 		 * not use a foreach loop.
15114 		 */
15115 		rsm = tqhash_min(rack->r_ctl.tqh);
15116 		while (rsm) {
15117 			tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15118 			rack->r_ctl.rc_num_maps_alloced--;
15119 			uma_zfree(rack_zone, rsm);
15120 			rsm = tqhash_min(rack->r_ctl.tqh);
15121 		}
15122 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15123 		while (rsm) {
15124 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15125 			rack->r_ctl.rc_num_maps_alloced--;
15126 			rack->rc_free_cnt--;
15127 			cnt_free++;
15128 			uma_zfree(rack_zone, rsm);
15129 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15130 		}
15131 		if (rack->r_ctl.pcm_s != NULL) {
15132 			free(rack->r_ctl.pcm_s, M_TCPPCM);
15133 			rack->r_ctl.pcm_s = NULL;
15134 			rack->r_ctl.pcm_i.cnt_alloc = 0;
15135 			rack->r_ctl.pcm_i.cnt = 0;
15136 		}
15137 		if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15138 		    (tcp_bblogging_on(tp))) {
15139 			union tcp_log_stackspecific log;
15140 			struct timeval tv;
15141 
15142 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15143 			log.u_bbr.flex8 = 10;
15144 			log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15145 			log.u_bbr.flex2 = rack->rc_free_cnt;
15146 			log.u_bbr.flex3 = cnt_free;
15147 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15148 			rsm = tqhash_min(rack->r_ctl.tqh);
15149 			log.u_bbr.delRate = (uintptr_t)rsm;
15150 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15151 			log.u_bbr.cur_del_rate = (uintptr_t)rsm;
15152 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15153 			log.u_bbr.pkt_epoch = __LINE__;
15154 			(void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15155 					     0, &log, false, NULL, NULL, 0, &tv);
15156 		}
15157 		KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15158 			("rack:%p num_aloc:%u after freeing all?",
15159 			 rack,
15160 			 rack->r_ctl.rc_num_maps_alloced));
15161 		rack->rc_free_cnt = 0;
15162 		free(rack->r_ctl.tqh, M_TCPFSB);
15163 		rack->r_ctl.tqh = NULL;
15164 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15165 		tp->t_fb_ptr = NULL;
15166 	}
15167 	/* Make sure snd_nxt is correctly set */
15168 	tp->snd_nxt = tp->snd_max;
15169 }
15170 
15171 static void
15172 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15173 {
15174 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15175 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15176 	}
15177 	switch (tp->t_state) {
15178 	case TCPS_SYN_SENT:
15179 		rack->r_state = TCPS_SYN_SENT;
15180 		rack->r_substate = rack_do_syn_sent;
15181 		break;
15182 	case TCPS_SYN_RECEIVED:
15183 		rack->r_state = TCPS_SYN_RECEIVED;
15184 		rack->r_substate = rack_do_syn_recv;
15185 		break;
15186 	case TCPS_ESTABLISHED:
15187 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15188 		rack->r_state = TCPS_ESTABLISHED;
15189 		rack->r_substate = rack_do_established;
15190 		break;
15191 	case TCPS_CLOSE_WAIT:
15192 		rack->r_state = TCPS_CLOSE_WAIT;
15193 		rack->r_substate = rack_do_close_wait;
15194 		break;
15195 	case TCPS_FIN_WAIT_1:
15196 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15197 		rack->r_state = TCPS_FIN_WAIT_1;
15198 		rack->r_substate = rack_do_fin_wait_1;
15199 		break;
15200 	case TCPS_CLOSING:
15201 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15202 		rack->r_state = TCPS_CLOSING;
15203 		rack->r_substate = rack_do_closing;
15204 		break;
15205 	case TCPS_LAST_ACK:
15206 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15207 		rack->r_state = TCPS_LAST_ACK;
15208 		rack->r_substate = rack_do_lastack;
15209 		break;
15210 	case TCPS_FIN_WAIT_2:
15211 		rack->r_state = TCPS_FIN_WAIT_2;
15212 		rack->r_substate = rack_do_fin_wait_2;
15213 		break;
15214 	case TCPS_LISTEN:
15215 	case TCPS_CLOSED:
15216 	case TCPS_TIME_WAIT:
15217 	default:
15218 		break;
15219 	};
15220 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15221 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15222 
15223 }
15224 
15225 static void
15226 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15227 {
15228 	/*
15229 	 * We received an ack, and then did not
15230 	 * call send or were bounced out due to the
15231 	 * hpts was running. Now a timer is up as well, is
15232 	 * it the right timer?
15233 	 */
15234 	struct rack_sendmap *rsm;
15235 	int tmr_up;
15236 
15237 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15238 	if (tcp_in_hpts(rack->rc_tp) == 0) {
15239 		/*
15240 		 * Ok we probably need some timer up, but no
15241 		 * matter what the mask we are not in hpts. We
15242 		 * may have received an old ack and thus did nothing.
15243 		 */
15244 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15245 		rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15246 		return;
15247 	}
15248 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15249 		return;
15250 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15251 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15252 	    (tmr_up == PACE_TMR_RXT)) {
15253 		/* Should be an RXT */
15254 		return;
15255 	}
15256 	if (rsm == NULL) {
15257 		/* Nothing outstanding? */
15258 		if (tp->t_flags & TF_DELACK) {
15259 			if (tmr_up == PACE_TMR_DELACK)
15260 				/* We are supposed to have delayed ack up and we do */
15261 				return;
15262 		} else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
15263 			/*
15264 			 * if we hit enobufs then we would expect the possibility
15265 			 * of nothing outstanding and the RXT up (and the hptsi timer).
15266 			 */
15267 			return;
15268 		} else if (((V_tcp_always_keepalive ||
15269 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15270 			    (tp->t_state <= TCPS_CLOSING)) &&
15271 			   (tmr_up == PACE_TMR_KEEP) &&
15272 			   (tp->snd_max == tp->snd_una)) {
15273 			/* We should have keep alive up and we do */
15274 			return;
15275 		}
15276 	}
15277 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15278 		   ((tmr_up == PACE_TMR_TLP) ||
15279 		    (tmr_up == PACE_TMR_RACK) ||
15280 		    (tmr_up == PACE_TMR_RXT))) {
15281 		/*
15282 		 * Either a Rack, TLP or RXT is fine if  we
15283 		 * have outstanding data.
15284 		 */
15285 		return;
15286 	} else if (tmr_up == PACE_TMR_DELACK) {
15287 		/*
15288 		 * If the delayed ack was going to go off
15289 		 * before the rtx/tlp/rack timer were going to
15290 		 * expire, then that would be the timer in control.
15291 		 * Note we don't check the time here trusting the
15292 		 * code is correct.
15293 		 */
15294 		return;
15295 	}
15296 	/*
15297 	 * Ok the timer originally started is not what we want now.
15298 	 * We will force the hpts to be stopped if any, and restart
15299 	 * with the slot set to what was in the saved slot.
15300 	 */
15301 	if (tcp_in_hpts(rack->rc_tp)) {
15302 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15303 			uint32_t us_cts;
15304 
15305 			us_cts = tcp_get_usecs(NULL);
15306 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15307 				rack->r_early = 1;
15308 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15309 			}
15310 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15311 		}
15312 		tcp_hpts_remove(rack->rc_tp);
15313 	}
15314 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15315 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15316 }
15317 
15318 
15319 static void
15320 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15321 {
15322 	if ((SEQ_LT(tp->snd_wl1, seq) ||
15323 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15324 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15325 		/* keep track of pure window updates */
15326 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15327 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15328 		tp->snd_wnd = tiwin;
15329 		rack_validate_fo_sendwin_up(tp, rack);
15330 		tp->snd_wl1 = seq;
15331 		tp->snd_wl2 = ack;
15332 		if (tp->snd_wnd > tp->max_sndwnd)
15333 			tp->max_sndwnd = tp->snd_wnd;
15334 	    rack->r_wanted_output = 1;
15335 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15336 		tp->snd_wnd = tiwin;
15337 		rack_validate_fo_sendwin_up(tp, rack);
15338 		tp->snd_wl1 = seq;
15339 		tp->snd_wl2 = ack;
15340 	} else {
15341 		/* Not a valid win update */
15342 		return;
15343 	}
15344 	if (tp->snd_wnd > tp->max_sndwnd)
15345 		tp->max_sndwnd = tp->snd_wnd;
15346 	/* Do we exit persists? */
15347 	if ((rack->rc_in_persist != 0) &&
15348 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15349 				rack->r_ctl.rc_pace_min_segs))) {
15350 		rack_exit_persist(tp, rack, cts);
15351 	}
15352 	/* Do we enter persists? */
15353 	if ((rack->rc_in_persist == 0) &&
15354 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15355 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
15356 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15357 	    sbavail(&tptosocket(tp)->so_snd) &&
15358 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15359 		/*
15360 		 * Here the rwnd is less than
15361 		 * the pacing size, we are established,
15362 		 * nothing is outstanding, and there is
15363 		 * data to send. Enter persists.
15364 		 */
15365 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15366 	}
15367 }
15368 
15369 static void
15370 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15371 {
15372 
15373 	if (tcp_bblogging_on(rack->rc_tp)) {
15374 		struct inpcb *inp = tptoinpcb(tp);
15375 		union tcp_log_stackspecific log;
15376 		struct timeval ltv;
15377 		char tcp_hdr_buf[60];
15378 		struct tcphdr *th;
15379 		struct timespec ts;
15380 		uint32_t orig_snd_una;
15381 		uint8_t xx = 0;
15382 
15383 #ifdef TCP_REQUEST_TRK
15384 		struct tcp_sendfile_track *tcp_req;
15385 
15386 		if (SEQ_GT(ae->ack, tp->snd_una)) {
15387 			tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
15388 		} else {
15389 			tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
15390 		}
15391 #endif
15392 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15393 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
15394 		if (rack->rack_no_prr == 0)
15395 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15396 		else
15397 			log.u_bbr.flex1 = 0;
15398 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15399 		log.u_bbr.use_lt_bw <<= 1;
15400 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
15401 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15402 		log.u_bbr.bbr_state = rack->rc_free_cnt;
15403 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15404 		log.u_bbr.pkts_out = tp->t_maxseg;
15405 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15406 		log.u_bbr.flex7 = 1;
15407 		log.u_bbr.lost = ae->flags;
15408 		log.u_bbr.cwnd_gain = ackval;
15409 		log.u_bbr.pacing_gain = 0x2;
15410 		if (ae->flags & TSTMP_HDWR) {
15411 			/* Record the hardware timestamp if present */
15412 			log.u_bbr.flex3 = M_TSTMP;
15413 			ts.tv_sec = ae->timestamp / 1000000000;
15414 			ts.tv_nsec = ae->timestamp % 1000000000;
15415 			ltv.tv_sec = ts.tv_sec;
15416 			ltv.tv_usec = ts.tv_nsec / 1000;
15417 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
15418 		} else if (ae->flags & TSTMP_LRO) {
15419 			/* Record the LRO the arrival timestamp */
15420 			log.u_bbr.flex3 = M_TSTMP_LRO;
15421 			ts.tv_sec = ae->timestamp / 1000000000;
15422 			ts.tv_nsec = ae->timestamp % 1000000000;
15423 			ltv.tv_sec = ts.tv_sec;
15424 			ltv.tv_usec = ts.tv_nsec / 1000;
15425 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
15426 		}
15427 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
15428 		/* Log the rcv time */
15429 		log.u_bbr.delRate = ae->timestamp;
15430 #ifdef TCP_REQUEST_TRK
15431 		log.u_bbr.applimited = tp->t_tcpreq_closed;
15432 		log.u_bbr.applimited <<= 8;
15433 		log.u_bbr.applimited |= tp->t_tcpreq_open;
15434 		log.u_bbr.applimited <<= 8;
15435 		log.u_bbr.applimited |= tp->t_tcpreq_req;
15436 		if (tcp_req) {
15437 			/* Copy out any client req info */
15438 			/* seconds */
15439 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
15440 			/* useconds */
15441 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
15442 			log.u_bbr.rttProp = tcp_req->timestamp;
15443 			log.u_bbr.cur_del_rate = tcp_req->start;
15444 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
15445 				log.u_bbr.flex8 |= 1;
15446 			} else {
15447 				log.u_bbr.flex8 |= 2;
15448 				log.u_bbr.bw_inuse = tcp_req->end;
15449 			}
15450 			log.u_bbr.flex6 = tcp_req->start_seq;
15451 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
15452 				log.u_bbr.flex8 |= 4;
15453 				log.u_bbr.epoch = tcp_req->end_seq;
15454 			}
15455 		}
15456 #endif
15457 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15458 		th = (struct tcphdr *)tcp_hdr_buf;
15459 		th->th_seq = ae->seq;
15460 		th->th_ack = ae->ack;
15461 		th->th_win = ae->win;
15462 		/* Now fill in the ports */
15463 		th->th_sport = inp->inp_fport;
15464 		th->th_dport = inp->inp_lport;
15465 		tcp_set_flags(th, ae->flags);
15466 		/* Now do we have a timestamp option? */
15467 		if (ae->flags & HAS_TSTMP) {
15468 			u_char *cp;
15469 			uint32_t val;
15470 
15471 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15472 			cp = (u_char *)(th + 1);
15473 			*cp = TCPOPT_NOP;
15474 			cp++;
15475 			*cp = TCPOPT_NOP;
15476 			cp++;
15477 			*cp = TCPOPT_TIMESTAMP;
15478 			cp++;
15479 			*cp = TCPOLEN_TIMESTAMP;
15480 			cp++;
15481 			val = htonl(ae->ts_value);
15482 			bcopy((char *)&val,
15483 			      (char *)cp, sizeof(uint32_t));
15484 			val = htonl(ae->ts_echo);
15485 			bcopy((char *)&val,
15486 			      (char *)(cp + 4), sizeof(uint32_t));
15487 		} else
15488 			th->th_off = (sizeof(struct tcphdr) >> 2);
15489 
15490 		/*
15491 		 * For sane logging we need to play a little trick.
15492 		 * If the ack were fully processed we would have moved
15493 		 * snd_una to high_seq, but since compressed acks are
15494 		 * processed in two phases, at this point (logging) snd_una
15495 		 * won't be advanced. So we would see multiple acks showing
15496 		 * the advancement. We can prevent that by "pretending" that
15497 		 * snd_una was advanced and then un-advancing it so that the
15498 		 * logging code has the right value for tlb_snd_una.
15499 		 */
15500 		if (tp->snd_una != high_seq) {
15501 			orig_snd_una = tp->snd_una;
15502 			tp->snd_una = high_seq;
15503 			xx = 1;
15504 		} else
15505 			xx = 0;
15506 		TCP_LOG_EVENTP(tp, th,
15507 			       &tptosocket(tp)->so_rcv,
15508 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15509 			       0, &log, true, &ltv);
15510 		if (xx) {
15511 			tp->snd_una = orig_snd_una;
15512 		}
15513 	}
15514 
15515 }
15516 
15517 static void
15518 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15519 {
15520 	uint32_t us_rtt;
15521 	/*
15522 	 * A persist or keep-alive was forced out, update our
15523 	 * min rtt time. Note now worry about lost responses.
15524 	 * When a subsequent keep-alive or persist times out
15525 	 * and forced_ack is still on, then the last probe
15526 	 * was not responded to. In such cases we have a
15527 	 * sysctl that controls the behavior. Either we apply
15528 	 * the rtt but with reduced confidence (0). Or we just
15529 	 * plain don't apply the rtt estimate. Having data flow
15530 	 * will clear the probe_not_answered flag i.e. cum-ack
15531 	 * move forward <or> exiting and reentering persists.
15532 	 */
15533 
15534 	rack->forced_ack = 0;
15535 	rack->rc_tp->t_rxtshift = 0;
15536 	if ((rack->rc_in_persist &&
15537 	     (tiwin == rack->rc_tp->snd_wnd)) ||
15538 	    (rack->rc_in_persist == 0)) {
15539 		/*
15540 		 * In persists only apply the RTT update if this is
15541 		 * a response to our window probe. And that
15542 		 * means the rwnd sent must match the current
15543 		 * snd_wnd. If it does not, then we got a
15544 		 * window update ack instead. For keepalive
15545 		 * we allow the answer no matter what the window.
15546 		 *
15547 		 * Note that if the probe_not_answered is set then
15548 		 * the forced_ack_ts is the oldest one i.e. the first
15549 		 * probe sent that might have been lost. This assures
15550 		 * us that if we do calculate an RTT it is longer not
15551 		 * some short thing.
15552 		 */
15553 		if (rack->rc_in_persist)
15554 			counter_u64_add(rack_persists_acks, 1);
15555 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15556 		if (us_rtt == 0)
15557 			us_rtt = 1;
15558 		if (rack->probe_not_answered == 0) {
15559 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15560 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15561 		} else {
15562 			/* We have a retransmitted probe here too */
15563 			if (rack_apply_rtt_with_reduced_conf) {
15564 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15565 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15566 			}
15567 		}
15568 	}
15569 }
15570 
15571 static void
15572 rack_new_round_starts(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15573 {
15574 	/*
15575 	 * The next send has occurred mark the end of the round
15576 	 * as when that data gets acknowledged. We can
15577 	 * also do common things we might need to do when
15578 	 * a round begins.
15579 	 */
15580 	rack->r_ctl.roundends = tp->snd_max;
15581 	rack->rc_new_rnd_needed = 0;
15582 	rack_log_hystart_event(rack, tp->snd_max, 4);
15583 }
15584 
15585 
15586 static void
15587 rack_log_pcm(struct tcp_rack *rack, uint8_t mod, uint32_t flex1, uint32_t flex2,
15588 	     uint32_t flex3)
15589 {
15590 	if (tcp_bblogging_on(rack->rc_tp)) {
15591 		union tcp_log_stackspecific log;
15592 		struct timeval tv;
15593 
15594 		(void)tcp_get_usecs(&tv);
15595 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15596 		log.u_bbr.timeStamp = tcp_tv_to_usectick(&tv);
15597 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15598 		log.u_bbr.flex8 = mod;
15599 		log.u_bbr.flex1 = flex1;
15600 		log.u_bbr.flex2 = flex2;
15601 		log.u_bbr.flex3 = flex3;
15602 		log.u_bbr.flex4 = rack_pcm_every_n_rounds;
15603 		log.u_bbr.flex5 = rack->r_ctl.pcm_idle_rounds;
15604 		log.u_bbr.bbr_substate = rack->pcm_needed;
15605 		log.u_bbr.bbr_substate <<= 1;
15606 		log.u_bbr.bbr_substate |= rack->pcm_in_progress;
15607 		log.u_bbr.bbr_substate <<= 1;
15608 		log.u_bbr.bbr_substate |= rack->pcm_enabled; /* bits are NIE for Needed, Inprogress, Enabled */
15609 		(void)tcp_log_event(rack->rc_tp, NULL, NULL, NULL, TCP_PCM_MEASURE, ERRNO_UNK,
15610 				    0, &log, false, NULL, NULL, 0, &tv);
15611 	}
15612 }
15613 
15614 static void
15615 rack_new_round_setup(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15616 {
15617 	/*
15618 	 * The round (current_round) has ended. We now
15619 	 * setup for the next round by incrementing the
15620 	 * round numnber and doing any round specific
15621 	 * things.
15622 	 */
15623 	rack_log_hystart_event(rack, high_seq, 21);
15624 	rack->r_ctl.current_round++;
15625 	/* New round (current_round) begins at next send */
15626 	rack->rc_new_rnd_needed = 1;
15627 	if ((rack->pcm_enabled == 1) &&
15628 	    (rack->pcm_needed == 0) &&
15629 	    (rack->pcm_in_progress == 0)) {
15630 		/*
15631 		 * If we have enabled PCM, then we need to
15632 		 * check if the round has adanced to the state
15633 		 * where one is required.
15634 		 */
15635 		int rnds;
15636 
15637 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
15638 		if ((rnds + rack->r_ctl.pcm_idle_rounds) >= rack_pcm_every_n_rounds) {
15639 			rack->pcm_needed = 1;
15640 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15641 		} else if (rack_verbose_logging) {
15642 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15643 		}
15644 	}
15645 	if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
15646 		/* We have hystart enabled send the round info in */
15647 		if (CC_ALGO(tp)->newround != NULL) {
15648 			CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
15649 		}
15650 	}
15651 	/*
15652 	 * For DGP an initial startup check. We want to validate
15653 	 * that we are not just pushing on slow-start and just
15654 	 * not gaining.. i.e. filling buffers without getting any
15655 	 * boost in b/w during the inital slow-start.
15656 	 */
15657 	if (rack->dgp_on &&
15658 	    (rack->rc_initial_ss_comp == 0) &&
15659 	    (tp->snd_cwnd < tp->snd_ssthresh) &&
15660 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG) &&
15661 	    (rack->r_ctl.gp_rnd_thresh > 0) &&
15662 	    ((rack->r_ctl.current_round - rack->r_ctl.last_rnd_of_gp_rise) >= rack->r_ctl.gp_rnd_thresh)) {
15663 
15664 		/*
15665 		 * We are in the initial SS and we have hd rack_rnd_cnt_req rounds(def:5) where
15666 		 * we have not gained the required amount in the gp_est (120.0% aka 1200). Lets
15667 		 * exit SS.
15668 		 *
15669 		 * Pick up the flight size now as we enter slowstart (not the
15670 		 * cwnd which may be inflated).
15671 		 */
15672 		rack->rc_initial_ss_comp = 1;
15673 
15674 		if (tcp_bblogging_on(rack->rc_tp)) {
15675 			union tcp_log_stackspecific log;
15676 			struct timeval tv;
15677 
15678 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15679 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15680 			log.u_bbr.flex1 = rack->r_ctl.current_round;
15681 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
15682 			log.u_bbr.flex3 = rack->r_ctl.gp_rnd_thresh;
15683 			log.u_bbr.flex4 = rack->r_ctl.gate_to_fs;
15684 			log.u_bbr.flex5 = rack->r_ctl.ss_hi_fs;
15685 			log.u_bbr.flex8 = 40;
15686 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
15687 					    0, &log, false, NULL, __func__, __LINE__,&tv);
15688 		}
15689 		if ((rack->r_ctl.gate_to_fs == 1) &&
15690 		     (tp->snd_cwnd > rack->r_ctl.ss_hi_fs)) {
15691 			tp->snd_cwnd = rack->r_ctl.ss_hi_fs;
15692 		}
15693 		tp->snd_ssthresh = tp->snd_cwnd - 1;
15694 		/* Turn off any fast output running */
15695 		rack->r_fast_output = 0;
15696 	}
15697 }
15698 
15699 static int
15700 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15701 {
15702 	/*
15703 	 * Handle a "special" compressed ack mbuf. Each incoming
15704 	 * ack has only four possible dispositions:
15705 	 *
15706 	 * A) It moves the cum-ack forward
15707 	 * B) It is behind the cum-ack.
15708 	 * C) It is a window-update ack.
15709 	 * D) It is a dup-ack.
15710 	 *
15711 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15712 	 * in the incoming mbuf. We also need to still pay attention
15713 	 * to nxt_pkt since there may be another packet after this
15714 	 * one.
15715 	 */
15716 #ifdef TCP_ACCOUNTING
15717 	uint64_t ts_val;
15718 	uint64_t rdstc;
15719 #endif
15720 	int segsiz;
15721 	struct timespec ts;
15722 	struct tcp_rack *rack;
15723 	struct tcp_ackent *ae;
15724 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15725 	int cnt, i, did_out, ourfinisacked = 0;
15726 	struct tcpopt to_holder, *to = NULL;
15727 #ifdef TCP_ACCOUNTING
15728 	int win_up_req = 0;
15729 #endif
15730 	int nsegs = 0;
15731 	int under_pacing = 0;
15732 	int post_recovery = 0;
15733 #ifdef TCP_ACCOUNTING
15734 	sched_pin();
15735 #endif
15736 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15737 	if (rack->gp_ready &&
15738 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15739 		under_pacing = 1;
15740 
15741 	if (rack->r_state != tp->t_state)
15742 		rack_set_state(tp, rack);
15743 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15744 	    (tp->t_flags & TF_GPUTINPROG)) {
15745 		/*
15746 		 * We have a goodput in progress
15747 		 * and we have entered a late state.
15748 		 * Do we have enough data in the sb
15749 		 * to handle the GPUT request?
15750 		 */
15751 		uint32_t bytes;
15752 
15753 		bytes = tp->gput_ack - tp->gput_seq;
15754 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
15755 			bytes += tp->gput_seq - tp->snd_una;
15756 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15757 			/*
15758 			 * There are not enough bytes in the socket
15759 			 * buffer that have been sent to cover this
15760 			 * measurement. Cancel it.
15761 			 */
15762 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15763 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
15764 						   tp->gput_seq,
15765 						   0, 0, 18, __LINE__, NULL, 0);
15766 			tp->t_flags &= ~TF_GPUTINPROG;
15767 		}
15768 	}
15769 	to = &to_holder;
15770 	to->to_flags = 0;
15771 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15772 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15773 	cnt = m->m_len / sizeof(struct tcp_ackent);
15774 	counter_u64_add(rack_multi_single_eq, cnt);
15775 	high_seq = tp->snd_una;
15776 	the_win = tp->snd_wnd;
15777 	win_seq = tp->snd_wl1;
15778 	win_upd_ack = tp->snd_wl2;
15779 	cts = tcp_tv_to_usectick(tv);
15780 	ms_cts = tcp_tv_to_mssectick(tv);
15781 	rack->r_ctl.rc_rcvtime = cts;
15782 	segsiz = ctf_fixed_maxseg(tp);
15783 	if ((rack->rc_gp_dyn_mul) &&
15784 	    (rack->use_fixed_rate == 0) &&
15785 	    (rack->rc_always_pace)) {
15786 		/* Check in on probertt */
15787 		rack_check_probe_rtt(rack, cts);
15788 	}
15789 	for (i = 0; i < cnt; i++) {
15790 #ifdef TCP_ACCOUNTING
15791 		ts_val = get_cyclecount();
15792 #endif
15793 		rack_clear_rate_sample(rack);
15794 		ae = ((mtod(m, struct tcp_ackent *)) + i);
15795 		if (ae->flags & TH_FIN)
15796 			rack_log_pacing_delay_calc(rack,
15797 						   0,
15798 						   0,
15799 						   0,
15800 						   rack_get_gp_est(rack), /* delRate */
15801 						   rack_get_lt_bw(rack), /* rttProp */
15802 						   20, __LINE__, NULL, 0);
15803 		/* Setup the window */
15804 		tiwin = ae->win << tp->snd_scale;
15805 		if (tiwin > rack->r_ctl.rc_high_rwnd)
15806 			rack->r_ctl.rc_high_rwnd = tiwin;
15807 		/* figure out the type of ack */
15808 		if (SEQ_LT(ae->ack, high_seq)) {
15809 			/* Case B*/
15810 			ae->ack_val_set = ACK_BEHIND;
15811 		} else if (SEQ_GT(ae->ack, high_seq)) {
15812 			/* Case A */
15813 			ae->ack_val_set = ACK_CUMACK;
15814 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
15815 			/* Case D */
15816 			ae->ack_val_set = ACK_DUPACK;
15817 		} else {
15818 			/* Case C */
15819 			ae->ack_val_set = ACK_RWND;
15820 		}
15821 		rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
15822 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
15823 		/* Validate timestamp */
15824 		if (ae->flags & HAS_TSTMP) {
15825 			/* Setup for a timestamp */
15826 			to->to_flags = TOF_TS;
15827 			ae->ts_echo -= tp->ts_offset;
15828 			to->to_tsecr = ae->ts_echo;
15829 			to->to_tsval = ae->ts_value;
15830 			/*
15831 			 * If echoed timestamp is later than the current time, fall back to
15832 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
15833 			 * were used when this connection was established.
15834 			 */
15835 			if (TSTMP_GT(ae->ts_echo, ms_cts))
15836 				to->to_tsecr = 0;
15837 			if (tp->ts_recent &&
15838 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
15839 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
15840 #ifdef TCP_ACCOUNTING
15841 					rdstc = get_cyclecount();
15842 					if (rdstc > ts_val) {
15843 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15844 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15845 						}
15846 					}
15847 #endif
15848 					continue;
15849 				}
15850 			}
15851 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
15852 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
15853 				tp->ts_recent_age = tcp_ts_getticks();
15854 				tp->ts_recent = ae->ts_value;
15855 			}
15856 		} else {
15857 			/* Setup for a no options */
15858 			to->to_flags = 0;
15859 		}
15860 		/* Update the rcv time and perform idle reduction possibly */
15861 		if  (tp->t_idle_reduce &&
15862 		     (tp->snd_max == tp->snd_una) &&
15863 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
15864 			counter_u64_add(rack_input_idle_reduces, 1);
15865 			rack_cc_after_idle(rack, tp);
15866 		}
15867 		tp->t_rcvtime = ticks;
15868 		/* Now what about ECN of a chain of pure ACKs? */
15869 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
15870 			tcp_packets_this_ack(tp, ae->ack),
15871 			ae->codepoint))
15872 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
15873 #ifdef TCP_ACCOUNTING
15874 		/* Count for the specific type of ack in */
15875 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15876 			tp->tcp_cnt_counters[ae->ack_val_set]++;
15877 		}
15878 #endif
15879 		/*
15880 		 * Note how we could move up these in the determination
15881 		 * above, but we don't so that way the timestamp checks (and ECN)
15882 		 * is done first before we do any processing on the ACK.
15883 		 * The non-compressed path through the code has this
15884 		 * weakness (noted by @jtl) that it actually does some
15885 		 * processing before verifying the timestamp information.
15886 		 * We don't take that path here which is why we set
15887 		 * the ack_val_set first, do the timestamp and ecn
15888 		 * processing, and then look at what we have setup.
15889 		 */
15890 		if (ae->ack_val_set == ACK_BEHIND) {
15891 			/*
15892 			 * Case B flag reordering, if window is not closed
15893 			 * or it could be a keep-alive or persists
15894 			 */
15895 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
15896 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15897 				if (rack->r_ctl.rc_reorder_ts == 0)
15898 					rack->r_ctl.rc_reorder_ts = 1;
15899 			}
15900 		} else if (ae->ack_val_set == ACK_DUPACK) {
15901 			/* Case D */
15902 			rack_strike_dupack(rack, ae->ack);
15903 		} else if (ae->ack_val_set == ACK_RWND) {
15904 			/* Case C */
15905 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15906 				ts.tv_sec = ae->timestamp / 1000000000;
15907 				ts.tv_nsec = ae->timestamp % 1000000000;
15908 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15909 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15910 			} else {
15911 				rack->r_ctl.act_rcv_time = *tv;
15912 			}
15913 			if (rack->forced_ack) {
15914 				rack_handle_probe_response(rack, tiwin,
15915 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
15916 			}
15917 #ifdef TCP_ACCOUNTING
15918 			win_up_req = 1;
15919 #endif
15920 			win_upd_ack = ae->ack;
15921 			win_seq = ae->seq;
15922 			the_win = tiwin;
15923 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15924 		} else {
15925 			/* Case A */
15926 			if (SEQ_GT(ae->ack, tp->snd_max)) {
15927 				/*
15928 				 * We just send an ack since the incoming
15929 				 * ack is beyond the largest seq we sent.
15930 				 */
15931 				if ((tp->t_flags & TF_ACKNOW) == 0) {
15932 					ctf_ack_war_checks(tp);
15933 					if (tp->t_flags && TF_ACKNOW)
15934 						rack->r_wanted_output = 1;
15935 				}
15936 			} else {
15937 				nsegs++;
15938 				/* If the window changed setup to update */
15939 				if (tiwin != tp->snd_wnd) {
15940 					win_upd_ack = ae->ack;
15941 					win_seq = ae->seq;
15942 					the_win = tiwin;
15943 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15944 				}
15945 #ifdef TCP_ACCOUNTING
15946 				/* Account for the acks */
15947 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15948 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
15949 				}
15950 #endif
15951 				high_seq = ae->ack;
15952 				/* Setup our act_rcv_time */
15953 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15954 					ts.tv_sec = ae->timestamp / 1000000000;
15955 					ts.tv_nsec = ae->timestamp % 1000000000;
15956 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15957 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15958 				} else {
15959 					rack->r_ctl.act_rcv_time = *tv;
15960 				}
15961 				rack_process_to_cumack(tp, rack, ae->ack, cts, to,
15962 						       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
15963 #ifdef TCP_REQUEST_TRK
15964 				rack_req_check_for_comp(rack, high_seq);
15965 #endif
15966 				if (rack->rc_dsack_round_seen) {
15967 					/* Is the dsack round over? */
15968 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
15969 						/* Yes it is */
15970 						rack->rc_dsack_round_seen = 0;
15971 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
15972 					}
15973 				}
15974 			}
15975 		}
15976 		/* And lets be sure to commit the rtt measurements for this ack */
15977 		tcp_rack_xmit_timer_commit(rack, tp);
15978 #ifdef TCP_ACCOUNTING
15979 		rdstc = get_cyclecount();
15980 		if (rdstc > ts_val) {
15981 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15982 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15983 				if (ae->ack_val_set == ACK_CUMACK)
15984 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
15985 			}
15986 		}
15987 #endif
15988 	}
15989 #ifdef TCP_ACCOUNTING
15990 	ts_val = get_cyclecount();
15991 #endif
15992 	/* Tend to any collapsed window */
15993 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
15994 		/* The peer collapsed the window */
15995 		rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
15996 	} else if (rack->rc_has_collapsed)
15997 		rack_un_collapse_window(rack, __LINE__);
15998 	if ((rack->r_collapse_point_valid) &&
15999 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
16000 		rack->r_collapse_point_valid = 0;
16001 	acked_amount = acked = (high_seq - tp->snd_una);
16002 	if (acked) {
16003 		/*
16004 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
16005 		 * causes issues when we are just going app limited. Lets
16006 		 * instead use SEQ_GT <or> where its equal but more data
16007 		 * is outstanding.
16008 		 *
16009 		 * Also make sure we are on the last ack of a series. We
16010 		 * have to have all the ack's processed in queue to know
16011 		 * if there is something left outstanding.
16012 		 *
16013 		 */
16014 		if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
16015 		    (rack->rc_new_rnd_needed == 0) &&
16016 		    (nxt_pkt == 0)) {
16017 			/*
16018 			 * We have crossed into a new round with
16019 			 * this th_ack value.
16020 			 */
16021 			rack_new_round_setup(tp, rack, high_seq);
16022 		}
16023 		/*
16024 		 * Clear the probe not answered flag
16025 		 * since cum-ack moved forward.
16026 		 */
16027 		rack->probe_not_answered = 0;
16028 		if (tp->t_flags & TF_NEEDSYN) {
16029 			/*
16030 			 * T/TCP: Connection was half-synchronized, and our SYN has
16031 			 * been ACK'd (so connection is now fully synchronized).  Go
16032 			 * to non-starred state, increment snd_una for ACK of SYN,
16033 			 * and check if we can do window scaling.
16034 			 */
16035 			tp->t_flags &= ~TF_NEEDSYN;
16036 			tp->snd_una++;
16037 			acked_amount = acked = (high_seq - tp->snd_una);
16038 		}
16039 		if (acked > sbavail(&so->so_snd))
16040 			acked_amount = sbavail(&so->so_snd);
16041 		if (IN_FASTRECOVERY(tp->t_flags) &&
16042 		    (rack->rack_no_prr == 0))
16043 			rack_update_prr(tp, rack, acked_amount, high_seq);
16044 		if (IN_RECOVERY(tp->t_flags)) {
16045 			if (SEQ_LT(high_seq, tp->snd_recover) &&
16046 			    (SEQ_LT(high_seq, tp->snd_max))) {
16047 				tcp_rack_partialack(tp);
16048 			} else {
16049 				rack_post_recovery(tp, high_seq);
16050 				post_recovery = 1;
16051 			}
16052 		}  else if ((rack->rto_from_rec == 1) &&
16053 			    SEQ_GEQ(high_seq, tp->snd_recover)) {
16054 			/*
16055 			 * We were in recovery, hit a rxt timeout
16056 			 * and never re-entered recovery. The timeout(s)
16057 			 * made up all the lost data. In such a case
16058 			 * we need to clear the rto_from_rec flag.
16059 			 */
16060 			rack->rto_from_rec = 0;
16061 		}
16062 		/* Handle the rack-log-ack part (sendmap) */
16063 		if ((sbused(&so->so_snd) == 0) &&
16064 		    (acked > acked_amount) &&
16065 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16066 		    (tp->t_flags & TF_SENTFIN)) {
16067 			/*
16068 			 * We must be sure our fin
16069 			 * was sent and acked (we can be
16070 			 * in FIN_WAIT_1 without having
16071 			 * sent the fin).
16072 			 */
16073 			ourfinisacked = 1;
16074 			/*
16075 			 * Lets make sure snd_una is updated
16076 			 * since most likely acked_amount = 0 (it
16077 			 * should be).
16078 			 */
16079 			tp->snd_una = high_seq;
16080 		}
16081 		/* Did we make a RTO error? */
16082 		if ((tp->t_flags & TF_PREVVALID) &&
16083 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16084 			tp->t_flags &= ~TF_PREVVALID;
16085 			if (tp->t_rxtshift == 1 &&
16086 			    (int)(ticks - tp->t_badrxtwin) < 0)
16087 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16088 		}
16089 		/* Handle the data in the socket buffer */
16090 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16091 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16092 		if (acked_amount > 0) {
16093 			uint32_t p_cwnd;
16094 			struct mbuf *mfree;
16095 
16096 			if (post_recovery) {
16097 				/*
16098 				 * Grab the segsiz, multiply by 2 and add the snd_cwnd
16099 				 * that is the max the CC should add if we are exiting
16100 				 * recovery and doing a late add.
16101 				 */
16102 				p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16103 				p_cwnd <<= 1;
16104 				p_cwnd += tp->snd_cwnd;
16105 			}
16106 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, post_recovery);
16107 			if (post_recovery && (tp->snd_cwnd > p_cwnd)) {
16108 				/* Must be non-newreno (cubic) getting too ahead of itself */
16109 				tp->snd_cwnd = p_cwnd;
16110 			}
16111 			SOCKBUF_LOCK(&so->so_snd);
16112 			mfree = sbcut_locked(&so->so_snd, acked_amount);
16113 			tp->snd_una = high_seq;
16114 			/* Note we want to hold the sb lock through the sendmap adjust */
16115 			rack_adjust_sendmap_head(rack, &so->so_snd);
16116 			/* Wake up the socket if we have room to write more */
16117 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16118 			sowwakeup_locked(so);
16119 			m_freem(mfree);
16120 		}
16121 		/* update progress */
16122 		tp->t_acktime = ticks;
16123 		rack_log_progress_event(rack, tp, tp->t_acktime,
16124 					PROGRESS_UPDATE, __LINE__);
16125 		/* Clear out shifts and such */
16126 		tp->t_rxtshift = 0;
16127 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16128 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16129 		rack->rc_tlp_in_progress = 0;
16130 		rack->r_ctl.rc_tlp_cnt_out = 0;
16131 		/* Send recover and snd_nxt must be dragged along */
16132 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
16133 			tp->snd_recover = tp->snd_una;
16134 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
16135 			tp->snd_nxt = tp->snd_max;
16136 		/*
16137 		 * If the RXT timer is running we want to
16138 		 * stop it, so we can restart a TLP (or new RXT).
16139 		 */
16140 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16141 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16142 		tp->snd_wl2 = high_seq;
16143 		tp->t_dupacks = 0;
16144 		if (under_pacing &&
16145 		    (rack->use_fixed_rate == 0) &&
16146 		    (rack->in_probe_rtt == 0) &&
16147 		    rack->rc_gp_dyn_mul &&
16148 		    rack->rc_always_pace) {
16149 			/* Check if we are dragging bottom */
16150 			rack_check_bottom_drag(tp, rack, so);
16151 		}
16152 		if (tp->snd_una == tp->snd_max) {
16153 			tp->t_flags &= ~TF_PREVVALID;
16154 			rack->r_ctl.retran_during_recovery = 0;
16155 			rack->rc_suspicious = 0;
16156 			rack->r_ctl.dsack_byte_cnt = 0;
16157 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16158 			if (rack->r_ctl.rc_went_idle_time == 0)
16159 				rack->r_ctl.rc_went_idle_time = 1;
16160 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16161 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
16162 				tp->t_acktime = 0;
16163 			/* Set so we might enter persists... */
16164 			rack->r_wanted_output = 1;
16165 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16166 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16167 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16168 			    (sbavail(&so->so_snd) == 0) &&
16169 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16170 				/*
16171 				 * The socket was gone and the
16172 				 * peer sent data (not now in the past), time to
16173 				 * reset him.
16174 				 */
16175 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16176 				/* tcp_close will kill the inp pre-log the Reset */
16177 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16178 #ifdef TCP_ACCOUNTING
16179 				rdstc = get_cyclecount();
16180 				if (rdstc > ts_val) {
16181 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16182 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16183 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16184 					}
16185 				}
16186 #endif
16187 				m_freem(m);
16188 				tp = tcp_close(tp);
16189 				if (tp == NULL) {
16190 #ifdef TCP_ACCOUNTING
16191 					sched_unpin();
16192 #endif
16193 					return (1);
16194 				}
16195 				/*
16196 				 * We would normally do drop-with-reset which would
16197 				 * send back a reset. We can't since we don't have
16198 				 * all the needed bits. Instead lets arrange for
16199 				 * a call to tcp_output(). That way since we
16200 				 * are in the closed state we will generate a reset.
16201 				 *
16202 				 * Note if tcp_accounting is on we don't unpin since
16203 				 * we do that after the goto label.
16204 				 */
16205 				goto send_out_a_rst;
16206 			}
16207 			if ((sbused(&so->so_snd) == 0) &&
16208 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16209 			    (tp->t_flags & TF_SENTFIN)) {
16210 				/*
16211 				 * If we can't receive any more data, then closing user can
16212 				 * proceed. Starting the timer is contrary to the
16213 				 * specification, but if we don't get a FIN we'll hang
16214 				 * forever.
16215 				 *
16216 				 */
16217 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16218 					soisdisconnected(so);
16219 					tcp_timer_activate(tp, TT_2MSL,
16220 							   (tcp_fast_finwait2_recycle ?
16221 							    tcp_finwait2_timeout :
16222 							    TP_MAXIDLE(tp)));
16223 				}
16224 				if (ourfinisacked == 0) {
16225 					/*
16226 					 * We don't change to fin-wait-2 if we have our fin acked
16227 					 * which means we are probably in TCPS_CLOSING.
16228 					 */
16229 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
16230 				}
16231 			}
16232 		}
16233 		/* Wake up the socket if we have room to write more */
16234 		if (sbavail(&so->so_snd)) {
16235 			rack->r_wanted_output = 1;
16236 			if (ctf_progress_timeout_check(tp, true)) {
16237 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16238 							tp, tick, PROGRESS_DROP, __LINE__);
16239 				/*
16240 				 * We cheat here and don't send a RST, we should send one
16241 				 * when the pacer drops the connection.
16242 				 */
16243 #ifdef TCP_ACCOUNTING
16244 				rdstc = get_cyclecount();
16245 				if (rdstc > ts_val) {
16246 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16247 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16248 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16249 					}
16250 				}
16251 				sched_unpin();
16252 #endif
16253 				(void)tcp_drop(tp, ETIMEDOUT);
16254 				m_freem(m);
16255 				return (1);
16256 			}
16257 		}
16258 		if (ourfinisacked) {
16259 			switch(tp->t_state) {
16260 			case TCPS_CLOSING:
16261 #ifdef TCP_ACCOUNTING
16262 				rdstc = get_cyclecount();
16263 				if (rdstc > ts_val) {
16264 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16265 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16266 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16267 					}
16268 				}
16269 				sched_unpin();
16270 #endif
16271 				tcp_twstart(tp);
16272 				m_freem(m);
16273 				return (1);
16274 				break;
16275 			case TCPS_LAST_ACK:
16276 #ifdef TCP_ACCOUNTING
16277 				rdstc = get_cyclecount();
16278 				if (rdstc > ts_val) {
16279 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16280 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16281 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16282 					}
16283 				}
16284 				sched_unpin();
16285 #endif
16286 				tp = tcp_close(tp);
16287 				ctf_do_drop(m, tp);
16288 				return (1);
16289 				break;
16290 			case TCPS_FIN_WAIT_1:
16291 #ifdef TCP_ACCOUNTING
16292 				rdstc = get_cyclecount();
16293 				if (rdstc > ts_val) {
16294 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16295 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16296 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16297 					}
16298 				}
16299 #endif
16300 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16301 					soisdisconnected(so);
16302 					tcp_timer_activate(tp, TT_2MSL,
16303 							   (tcp_fast_finwait2_recycle ?
16304 							    tcp_finwait2_timeout :
16305 							    TP_MAXIDLE(tp)));
16306 				}
16307 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
16308 				break;
16309 			default:
16310 				break;
16311 			}
16312 		}
16313 		if (rack->r_fast_output) {
16314 			/*
16315 			 * We re doing fast output.. can we expand that?
16316 			 */
16317 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16318 		}
16319 #ifdef TCP_ACCOUNTING
16320 		rdstc = get_cyclecount();
16321 		if (rdstc > ts_val) {
16322 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16323 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16324 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16325 			}
16326 		}
16327 
16328 	} else if (win_up_req) {
16329 		rdstc = get_cyclecount();
16330 		if (rdstc > ts_val) {
16331 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16332 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16333 			}
16334 		}
16335 #endif
16336 	}
16337 	/* Now is there a next packet, if so we are done */
16338 	m_freem(m);
16339 	did_out = 0;
16340 	if (nxt_pkt) {
16341 #ifdef TCP_ACCOUNTING
16342 		sched_unpin();
16343 #endif
16344 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16345 		return (0);
16346 	}
16347 	rack_handle_might_revert(tp, rack);
16348 	ctf_calc_rwin(so, tp);
16349 	if ((rack->r_wanted_output != 0) ||
16350 	    (rack->r_fast_output != 0) ||
16351 	    (tp->t_flags & TF_ACKNOW )) {
16352 	send_out_a_rst:
16353 		if (tcp_output(tp) < 0) {
16354 #ifdef TCP_ACCOUNTING
16355 			sched_unpin();
16356 #endif
16357 			return (1);
16358 		}
16359 		did_out = 1;
16360 	}
16361 	if (tp->t_flags2 & TF2_HPTS_CALLS)
16362 		tp->t_flags2 &= ~TF2_HPTS_CALLS;
16363 	rack_free_trim(rack);
16364 #ifdef TCP_ACCOUNTING
16365 	sched_unpin();
16366 #endif
16367 	rack_timer_audit(tp, rack, &so->so_snd);
16368 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16369 	return (0);
16370 }
16371 
16372 #define	TCP_LRO_TS_OPTION \
16373     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16374 	  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16375 
16376 static int
16377 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16378     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16379     struct timeval *tv)
16380 {
16381 	struct inpcb *inp = tptoinpcb(tp);
16382 	struct socket *so = tptosocket(tp);
16383 #ifdef TCP_ACCOUNTING
16384 	uint64_t ts_val;
16385 #endif
16386 	int32_t thflags, retval, did_out = 0;
16387 	int32_t way_out = 0;
16388 	/*
16389 	 * cts - is the current time from tv (caller gets ts) in microseconds.
16390 	 * ms_cts - is the current time from tv in milliseconds.
16391 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16392 	 */
16393 	uint32_t cts, us_cts, ms_cts;
16394 	uint32_t tiwin;
16395 	struct timespec ts;
16396 	struct tcpopt to;
16397 	struct tcp_rack *rack;
16398 	struct rack_sendmap *rsm;
16399 	int32_t prev_state = 0;
16400 	int no_output = 0;
16401 	int slot_remaining = 0;
16402 #ifdef TCP_ACCOUNTING
16403 	int ack_val_set = 0xf;
16404 #endif
16405 	int nsegs;
16406 
16407 	NET_EPOCH_ASSERT();
16408 	INP_WLOCK_ASSERT(inp);
16409 
16410 	/*
16411 	 * tv passed from common code is from either M_TSTMP_LRO or
16412 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16413 	 */
16414 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16415 	if (rack->rack_deferred_inited == 0) {
16416 		/*
16417 		 * If we are the connecting socket we will
16418 		 * hit rack_init() when no sequence numbers
16419 		 * are setup. This makes it so we must defer
16420 		 * some initialization. Call that now.
16421 		 */
16422 		rack_deferred_init(tp, rack);
16423 	}
16424 	/*
16425 	 * Check to see if we need to skip any output plans. This
16426 	 * can happen in the non-LRO path where we are pacing and
16427 	 * must process the ack coming in but need to defer sending
16428 	 * anything becase a pacing timer is running.
16429 	 */
16430 	us_cts = tcp_tv_to_usectick(tv);
16431 	if (m->m_flags & M_ACKCMP) {
16432 		/*
16433 		 * All compressed ack's are ack's by definition so
16434 		 * remove any ack required flag and then do the processing.
16435 		 */
16436 		rack->rc_ack_required = 0;
16437 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16438 	}
16439 	thflags = tcp_get_flags(th);
16440 	if ((rack->rc_always_pace == 1) &&
16441 	    (rack->rc_ack_can_sendout_data == 0) &&
16442 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16443 	    (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16444 		/*
16445 		 * Ok conditions are right for queuing the packets
16446 		 * but we do have to check the flags in the inp, it
16447 		 * could be, if a sack is present, we want to be awoken and
16448 		 * so should process the packets.
16449 		 */
16450 		slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16451 		if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
16452 			no_output = 1;
16453 		} else {
16454 			/*
16455 			 * If there is no options, or just a
16456 			 * timestamp option, we will want to queue
16457 			 * the packets. This is the same that LRO does
16458 			 * and will need to change with accurate ECN.
16459 			 */
16460 			uint32_t *ts_ptr;
16461 			int optlen;
16462 
16463 			optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16464 			ts_ptr = (uint32_t *)(th + 1);
16465 			if ((optlen == 0) ||
16466 			    ((optlen == TCPOLEN_TSTAMP_APPA) &&
16467 			     (*ts_ptr == TCP_LRO_TS_OPTION)))
16468 				no_output = 1;
16469 		}
16470 		if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16471 			/*
16472 			 * It is unrealistic to think we can pace in less than
16473 			 * the minimum granularity of the pacer (def:250usec). So
16474 			 * if we have less than that time remaining we should go
16475 			 * ahead and allow output to be "early". We will attempt to
16476 			 * make up for it in any pacing time we try to apply on
16477 			 * the outbound packet.
16478 			 */
16479 			no_output = 0;
16480 		}
16481 	}
16482 	/*
16483 	 * If there is a RST or FIN lets dump out the bw
16484 	 * with a FIN the connection may go on but we
16485 	 * may not.
16486 	 */
16487 	if ((thflags & TH_FIN) || (thflags & TH_RST))
16488 		rack_log_pacing_delay_calc(rack,
16489 					   rack->r_ctl.gp_bw,
16490 					   0,
16491 					   0,
16492 					   rack_get_gp_est(rack), /* delRate */
16493 					   rack_get_lt_bw(rack), /* rttProp */
16494 					   20, __LINE__, NULL, 0);
16495 	if (m->m_flags & M_ACKCMP) {
16496 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16497 	}
16498 	cts = tcp_tv_to_usectick(tv);
16499 	ms_cts =  tcp_tv_to_mssectick(tv);
16500 	nsegs = m->m_pkthdr.lro_nsegs;
16501 	counter_u64_add(rack_proc_non_comp_ack, 1);
16502 #ifdef TCP_ACCOUNTING
16503 	sched_pin();
16504 	if (thflags & TH_ACK)
16505 		ts_val = get_cyclecount();
16506 #endif
16507 	if ((m->m_flags & M_TSTMP) ||
16508 	    (m->m_flags & M_TSTMP_LRO)) {
16509 		mbuf_tstmp2timespec(m, &ts);
16510 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16511 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16512 	} else
16513 		rack->r_ctl.act_rcv_time = *tv;
16514 	kern_prefetch(rack, &prev_state);
16515 	prev_state = 0;
16516 	/*
16517 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
16518 	 * the scale is zero.
16519 	 */
16520 	tiwin = th->th_win << tp->snd_scale;
16521 #ifdef TCP_ACCOUNTING
16522 	if (thflags & TH_ACK) {
16523 		/*
16524 		 * We have a tradeoff here. We can either do what we are
16525 		 * doing i.e. pinning to this CPU and then doing the accounting
16526 		 * <or> we could do a critical enter, setup the rdtsc and cpu
16527 		 * as in below, and then validate we are on the same CPU on
16528 		 * exit. I have choosen to not do the critical enter since
16529 		 * that often will gain you a context switch, and instead lock
16530 		 * us (line above this if) to the same CPU with sched_pin(). This
16531 		 * means we may be context switched out for a higher priority
16532 		 * interupt but we won't be moved to another CPU.
16533 		 *
16534 		 * If this occurs (which it won't very often since we most likely
16535 		 * are running this code in interupt context and only a higher
16536 		 * priority will bump us ... clock?) we will falsely add in
16537 		 * to the time the interupt processing time plus the ack processing
16538 		 * time. This is ok since its a rare event.
16539 		 */
16540 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16541 						    ctf_fixed_maxseg(tp));
16542 	}
16543 #endif
16544 	/*
16545 	 * Parse options on any incoming segment.
16546 	 */
16547 	memset(&to, 0, sizeof(to));
16548 	tcp_dooptions(&to, (u_char *)(th + 1),
16549 	    (th->th_off << 2) - sizeof(struct tcphdr),
16550 	    (thflags & TH_SYN) ? TO_SYN : 0);
16551 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16552 	    __func__));
16553 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16554 	    __func__));
16555 	if (tp->t_flags2 & TF2_PROC_SACK_PROHIBIT) {
16556 		/*
16557 		 * We don't look at sack's from the
16558 		 * peer because the MSS is too small which
16559 		 * can subject us to an attack.
16560 		 */
16561 		to.to_flags &= ~TOF_SACK;
16562 	}
16563 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16564 	    (tp->t_flags & TF_GPUTINPROG)) {
16565 		/*
16566 		 * We have a goodput in progress
16567 		 * and we have entered a late state.
16568 		 * Do we have enough data in the sb
16569 		 * to handle the GPUT request?
16570 		 */
16571 		uint32_t bytes;
16572 
16573 		bytes = tp->gput_ack - tp->gput_seq;
16574 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
16575 			bytes += tp->gput_seq - tp->snd_una;
16576 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16577 			/*
16578 			 * There are not enough bytes in the socket
16579 			 * buffer that have been sent to cover this
16580 			 * measurement. Cancel it.
16581 			 */
16582 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16583 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
16584 						   tp->gput_seq,
16585 						   0, 0, 18, __LINE__, NULL, 0);
16586 			tp->t_flags &= ~TF_GPUTINPROG;
16587 		}
16588 	}
16589 	if (tcp_bblogging_on(rack->rc_tp)) {
16590 		union tcp_log_stackspecific log;
16591 		struct timeval ltv;
16592 #ifdef TCP_REQUEST_TRK
16593 		struct tcp_sendfile_track *tcp_req;
16594 
16595 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
16596 			tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
16597 		} else {
16598 			tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
16599 		}
16600 #endif
16601 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16602 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16603 		if (rack->rack_no_prr == 0)
16604 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16605 		else
16606 			log.u_bbr.flex1 = 0;
16607 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16608 		log.u_bbr.use_lt_bw <<= 1;
16609 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
16610 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16611 		log.u_bbr.bbr_state = rack->rc_free_cnt;
16612 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16613 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16614 		log.u_bbr.flex3 = m->m_flags;
16615 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16616 		log.u_bbr.lost = thflags;
16617 		log.u_bbr.pacing_gain = 0x1;
16618 #ifdef TCP_ACCOUNTING
16619 		log.u_bbr.cwnd_gain = ack_val_set;
16620 #endif
16621 		log.u_bbr.flex7 = 2;
16622 		if (m->m_flags & M_TSTMP) {
16623 			/* Record the hardware timestamp if present */
16624 			mbuf_tstmp2timespec(m, &ts);
16625 			ltv.tv_sec = ts.tv_sec;
16626 			ltv.tv_usec = ts.tv_nsec / 1000;
16627 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
16628 		} else if (m->m_flags & M_TSTMP_LRO) {
16629 			/* Record the LRO the arrival timestamp */
16630 			mbuf_tstmp2timespec(m, &ts);
16631 			ltv.tv_sec = ts.tv_sec;
16632 			ltv.tv_usec = ts.tv_nsec / 1000;
16633 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
16634 		}
16635 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16636 		/* Log the rcv time */
16637 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16638 #ifdef TCP_REQUEST_TRK
16639 		log.u_bbr.applimited = tp->t_tcpreq_closed;
16640 		log.u_bbr.applimited <<= 8;
16641 		log.u_bbr.applimited |= tp->t_tcpreq_open;
16642 		log.u_bbr.applimited <<= 8;
16643 		log.u_bbr.applimited |= tp->t_tcpreq_req;
16644 		if (tcp_req) {
16645 			/* Copy out any client req info */
16646 			/* seconds */
16647 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16648 			/* useconds */
16649 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16650 			log.u_bbr.rttProp = tcp_req->timestamp;
16651 			log.u_bbr.cur_del_rate = tcp_req->start;
16652 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16653 				log.u_bbr.flex8 |= 1;
16654 			} else {
16655 				log.u_bbr.flex8 |= 2;
16656 				log.u_bbr.bw_inuse = tcp_req->end;
16657 			}
16658 			log.u_bbr.flex6 = tcp_req->start_seq;
16659 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16660 				log.u_bbr.flex8 |= 4;
16661 				log.u_bbr.epoch = tcp_req->end_seq;
16662 			}
16663 		}
16664 #endif
16665 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16666 		    tlen, &log, true, &ltv);
16667 	}
16668 	/* Remove ack required flag if set, we have one  */
16669 	if (thflags & TH_ACK)
16670 		rack->rc_ack_required = 0;
16671 	rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16672 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16673 		way_out = 4;
16674 		retval = 0;
16675 		m_freem(m);
16676 		goto done_with_input;
16677 	}
16678 	/*
16679 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
16680 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16681 	 */
16682 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16683 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16684 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16685 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
16686 #ifdef TCP_ACCOUNTING
16687 		sched_unpin();
16688 #endif
16689 		return (1);
16690 	}
16691 	/*
16692 	 * If timestamps were negotiated during SYN/ACK and a
16693 	 * segment without a timestamp is received, silently drop
16694 	 * the segment, unless it is a RST segment or missing timestamps are
16695 	 * tolerated.
16696 	 * See section 3.2 of RFC 7323.
16697 	 */
16698 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16699 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16700 		way_out = 5;
16701 		retval = 0;
16702 		m_freem(m);
16703 		goto done_with_input;
16704 	}
16705 	/*
16706 	 * Segment received on connection. Reset idle time and keep-alive
16707 	 * timer. XXX: This should be done after segment validation to
16708 	 * ignore broken/spoofed segs.
16709 	 */
16710 	if  (tp->t_idle_reduce &&
16711 	     (tp->snd_max == tp->snd_una) &&
16712 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16713 		counter_u64_add(rack_input_idle_reduces, 1);
16714 		rack_cc_after_idle(rack, tp);
16715 	}
16716 	tp->t_rcvtime = ticks;
16717 #ifdef STATS
16718 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16719 #endif
16720 	if (tiwin > rack->r_ctl.rc_high_rwnd)
16721 		rack->r_ctl.rc_high_rwnd = tiwin;
16722 	/*
16723 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16724 	 * this to occur after we've validated the segment.
16725 	 */
16726 	if (tcp_ecn_input_segment(tp, thflags, tlen,
16727 	    tcp_packets_this_ack(tp, th->th_ack),
16728 	    iptos))
16729 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16730 
16731 	/*
16732 	 * If echoed timestamp is later than the current time, fall back to
16733 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16734 	 * were used when this connection was established.
16735 	 */
16736 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16737 		to.to_tsecr -= tp->ts_offset;
16738 		if (TSTMP_GT(to.to_tsecr, ms_cts))
16739 			to.to_tsecr = 0;
16740 	}
16741 	if ((rack->r_rcvpath_rtt_up == 1) &&
16742 	    (to.to_flags & TOF_TS) &&
16743 	    (TSTMP_GEQ(to.to_tsecr, rack->r_ctl.last_rcv_tstmp_for_rtt))) {
16744 		uint32_t rtt = 0;
16745 
16746 		/*
16747 		 * We are receiving only and thus not sending
16748 		 * data to do an RTT. We set a flag when we first
16749 		 * sent this TS to the peer. We now have it back
16750 		 * and have an RTT to share. We log it as a conf
16751 		 * 4, we are not so sure about it.. since we
16752 		 * may have lost an ack.
16753 		 */
16754 		if (TSTMP_GT(cts, rack->r_ctl.last_time_of_arm_rcv))
16755 		    rtt = (cts - rack->r_ctl.last_time_of_arm_rcv);
16756 		rack->r_rcvpath_rtt_up = 0;
16757 		/* Submit and commit the timer */
16758 		if (rtt > 0) {
16759 			tcp_rack_xmit_timer(rack, rtt, 0, rtt, 4, NULL, 1);
16760 			tcp_rack_xmit_timer_commit(rack, tp);
16761 		}
16762 	}
16763 	/*
16764 	 * If its the first time in we need to take care of options and
16765 	 * verify we can do SACK for rack!
16766 	 */
16767 	if (rack->r_state == 0) {
16768 		/* Should be init'd by rack_init() */
16769 		KASSERT(rack->rc_inp != NULL,
16770 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
16771 		if (rack->rc_inp == NULL) {
16772 			rack->rc_inp = inp;
16773 		}
16774 
16775 		/*
16776 		 * Process options only when we get SYN/ACK back. The SYN
16777 		 * case for incoming connections is handled in tcp_syncache.
16778 		 * According to RFC1323 the window field in a SYN (i.e., a
16779 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16780 		 * this is traditional behavior, may need to be cleaned up.
16781 		 */
16782 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16783 			/* Handle parallel SYN for ECN */
16784 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
16785 			if ((to.to_flags & TOF_SCALE) &&
16786 			    (tp->t_flags & TF_REQ_SCALE)) {
16787 				tp->t_flags |= TF_RCVD_SCALE;
16788 				tp->snd_scale = to.to_wscale;
16789 			} else
16790 				tp->t_flags &= ~TF_REQ_SCALE;
16791 			/*
16792 			 * Initial send window.  It will be updated with the
16793 			 * next incoming segment to the scaled value.
16794 			 */
16795 			tp->snd_wnd = th->th_win;
16796 			rack_validate_fo_sendwin_up(tp, rack);
16797 			if ((to.to_flags & TOF_TS) &&
16798 			    (tp->t_flags & TF_REQ_TSTMP)) {
16799 				tp->t_flags |= TF_RCVD_TSTMP;
16800 				tp->ts_recent = to.to_tsval;
16801 				tp->ts_recent_age = cts;
16802 			} else
16803 				tp->t_flags &= ~TF_REQ_TSTMP;
16804 			if (to.to_flags & TOF_MSS) {
16805 				tcp_mss(tp, to.to_mss);
16806 			}
16807 			if ((tp->t_flags & TF_SACK_PERMIT) &&
16808 			    (to.to_flags & TOF_SACKPERM) == 0)
16809 				tp->t_flags &= ~TF_SACK_PERMIT;
16810 			if (tp->t_flags & TF_FASTOPEN) {
16811 				if (to.to_flags & TOF_FASTOPEN) {
16812 					uint16_t mss;
16813 
16814 					if (to.to_flags & TOF_MSS)
16815 						mss = to.to_mss;
16816 					else
16817 						if ((inp->inp_vflag & INP_IPV6) != 0)
16818 							mss = TCP6_MSS;
16819 						else
16820 							mss = TCP_MSS;
16821 					tcp_fastopen_update_cache(tp, mss,
16822 					    to.to_tfo_len, to.to_tfo_cookie);
16823 				} else
16824 					tcp_fastopen_disable_path(tp);
16825 			}
16826 		}
16827 		/*
16828 		 * At this point we are at the initial call. Here we decide
16829 		 * if we are doing RACK or not. We do this by seeing if
16830 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
16831 		 * The code now does do dup-ack counting so if you don't
16832 		 * switch back you won't get rack & TLP, but you will still
16833 		 * get this stack.
16834 		 */
16835 
16836 		if ((rack_sack_not_required == 0) &&
16837 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
16838 			tcp_switch_back_to_default(tp);
16839 			(*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
16840 			    tlen, iptos);
16841 #ifdef TCP_ACCOUNTING
16842 			sched_unpin();
16843 #endif
16844 			return (1);
16845 		}
16846 		tcp_set_hpts(tp);
16847 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
16848 	}
16849 	if (thflags & TH_FIN)
16850 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
16851 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16852 	if ((rack->rc_gp_dyn_mul) &&
16853 	    (rack->use_fixed_rate == 0) &&
16854 	    (rack->rc_always_pace)) {
16855 		/* Check in on probertt */
16856 		rack_check_probe_rtt(rack, cts);
16857 	}
16858 	rack_clear_rate_sample(rack);
16859 	if ((rack->forced_ack) &&
16860 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
16861 		rack_handle_probe_response(rack, tiwin, us_cts);
16862 	}
16863 	/*
16864 	 * This is the one exception case where we set the rack state
16865 	 * always. All other times (timers etc) we must have a rack-state
16866 	 * set (so we assure we have done the checks above for SACK).
16867 	 */
16868 	rack->r_ctl.rc_rcvtime = cts;
16869 	if (rack->r_state != tp->t_state)
16870 		rack_set_state(tp, rack);
16871 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
16872 	    (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
16873 		kern_prefetch(rsm, &prev_state);
16874 	prev_state = rack->r_state;
16875 	if ((thflags & TH_RST) &&
16876 	    ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
16877 	      SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
16878 	     (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
16879 		/* The connection will be killed by a reset check the tracepoint */
16880 		tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
16881 	}
16882 	retval = (*rack->r_substate) (m, th, so,
16883 	    tp, &to, drop_hdrlen,
16884 	    tlen, tiwin, thflags, nxt_pkt, iptos);
16885 	if (retval == 0) {
16886 		/*
16887 		 * If retval is 1 the tcb is unlocked and most likely the tp
16888 		 * is gone.
16889 		 */
16890 		INP_WLOCK_ASSERT(inp);
16891 		if ((rack->rc_gp_dyn_mul) &&
16892 		    (rack->rc_always_pace) &&
16893 		    (rack->use_fixed_rate == 0) &&
16894 		    rack->in_probe_rtt &&
16895 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
16896 			/*
16897 			 * If we are going for target, lets recheck before
16898 			 * we output.
16899 			 */
16900 			rack_check_probe_rtt(rack, cts);
16901 		}
16902 		if (rack->set_pacing_done_a_iw == 0) {
16903 			/* How much has been acked? */
16904 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
16905 				/* We have enough to set in the pacing segment size */
16906 				rack->set_pacing_done_a_iw = 1;
16907 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
16908 			}
16909 		}
16910 		tcp_rack_xmit_timer_commit(rack, tp);
16911 #ifdef TCP_ACCOUNTING
16912 		/*
16913 		 * If we set the ack_val_se to what ack processing we are doing
16914 		 * we also want to track how many cycles we burned. Note
16915 		 * the bits after tcp_output we let be "free". This is because
16916 		 * we are also tracking the tcp_output times as well. Note the
16917 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
16918 		 * 0xf cannot be returned and is what we initialize it too to
16919 		 * indicate we are not doing the tabulations.
16920 		 */
16921 		if (ack_val_set != 0xf) {
16922 			uint64_t crtsc;
16923 
16924 			crtsc = get_cyclecount();
16925 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16926 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
16927 			}
16928 		}
16929 #endif
16930 		if ((nxt_pkt == 0) && (no_output == 0)) {
16931 			if ((rack->r_wanted_output != 0) ||
16932 			    (tp->t_flags & TF_ACKNOW) ||
16933 			    (rack->r_fast_output != 0)) {
16934 
16935 do_output_now:
16936 				if (tcp_output(tp) < 0) {
16937 #ifdef TCP_ACCOUNTING
16938 					sched_unpin();
16939 #endif
16940 					return (1);
16941 				}
16942 				did_out = 1;
16943 			}
16944 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16945 			rack_free_trim(rack);
16946 		} else if ((nxt_pkt == 0) && (tp->t_flags & TF_ACKNOW)) {
16947 			goto do_output_now;
16948 		} else if ((no_output == 1) &&
16949 			   (nxt_pkt == 0)  &&
16950 			   (tcp_in_hpts(rack->rc_tp) == 0)) {
16951 			/*
16952 			 * We are not in hpts and we had a pacing timer up. Use
16953 			 * the remaining time (slot_remaining) to restart the timer.
16954 			 */
16955 			KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
16956 			rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
16957 			rack_free_trim(rack);
16958 		}
16959 		/* Clear the flag, it may have been cleared by output but we may not have  */
16960 		if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
16961 			tp->t_flags2 &= ~TF2_HPTS_CALLS;
16962 		/*
16963 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
16964 		 * causes issues when we are just going app limited. Lets
16965 		 * instead use SEQ_GT <or> where its equal but more data
16966 		 * is outstanding.
16967 		 *
16968 		 * Also make sure we are on the last ack of a series. We
16969 		 * have to have all the ack's processed in queue to know
16970 		 * if there is something left outstanding.
16971 		 */
16972 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
16973 		    (rack->rc_new_rnd_needed == 0) &&
16974 		    (nxt_pkt == 0)) {
16975 			/*
16976 			 * We have crossed into a new round with
16977 			 * the new snd_unae.
16978 			 */
16979 			rack_new_round_setup(tp, rack, tp->snd_una);
16980 		}
16981 		if ((nxt_pkt == 0) &&
16982 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
16983 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
16984 		     (tp->t_flags & TF_DELACK) ||
16985 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
16986 		      (tp->t_state <= TCPS_CLOSING)))) {
16987 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
16988 			if ((tp->snd_max == tp->snd_una) &&
16989 			    ((tp->t_flags & TF_DELACK) == 0) &&
16990 			    (tcp_in_hpts(rack->rc_tp)) &&
16991 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
16992 				/* keep alive not needed if we are hptsi output yet */
16993 				;
16994 			} else {
16995 				int late = 0;
16996 				if (tcp_in_hpts(tp)) {
16997 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
16998 						us_cts = tcp_get_usecs(NULL);
16999 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
17000 							rack->r_early = 1;
17001 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
17002 						} else
17003 							late = 1;
17004 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
17005 					}
17006 					tcp_hpts_remove(tp);
17007 				}
17008 				if (late && (did_out == 0)) {
17009 					/*
17010 					 * We are late in the sending
17011 					 * and we did not call the output
17012 					 * (this probably should not happen).
17013 					 */
17014 					goto do_output_now;
17015 				}
17016 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
17017 			}
17018 			way_out = 1;
17019 		} else if (nxt_pkt == 0) {
17020 			/* Do we have the correct timer running? */
17021 			rack_timer_audit(tp, rack, &so->so_snd);
17022 			way_out = 2;
17023 		}
17024 	done_with_input:
17025 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
17026 		if (did_out)
17027 			rack->r_wanted_output = 0;
17028 	}
17029 
17030 #ifdef TCP_ACCOUNTING
17031 	sched_unpin();
17032 #endif
17033 	return (retval);
17034 }
17035 
17036 static void
17037 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17038     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17039 {
17040 	struct timeval tv;
17041 
17042 	/* First lets see if we have old packets */
17043 	if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17044 		if (ctf_do_queued_segments(tp, 1)) {
17045 			m_freem(m);
17046 			return;
17047 		}
17048 	}
17049 	if (m->m_flags & M_TSTMP_LRO) {
17050 		mbuf_tstmp2timeval(m, &tv);
17051 	} else {
17052 		/* Should not be should we kassert instead? */
17053 		tcp_get_usecs(&tv);
17054 	}
17055 	if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17056 	    &tv) == 0) {
17057 		INP_WUNLOCK(tptoinpcb(tp));
17058 	}
17059 }
17060 
17061 struct rack_sendmap *
17062 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17063 {
17064 	struct rack_sendmap *rsm = NULL;
17065 	int32_t idx;
17066 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
17067 
17068 	/* Return the next guy to be re-transmitted */
17069 	if (tqhash_empty(rack->r_ctl.tqh)) {
17070 		return (NULL);
17071 	}
17072 	if (tp->t_flags & TF_SENTFIN) {
17073 		/* retran the end FIN? */
17074 		return (NULL);
17075 	}
17076 	/* ok lets look at this one */
17077 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17078 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17079 		return (rsm);
17080 	}
17081 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17082 		goto check_it;
17083 	}
17084 	rsm = rack_find_lowest_rsm(rack);
17085 	if (rsm == NULL) {
17086 		return (NULL);
17087 	}
17088 check_it:
17089 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
17090 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17091 		/*
17092 		 * No sack so we automatically do the 3 strikes and
17093 		 * retransmit (no rack timer would be started).
17094 		 */
17095 		return (rsm);
17096 	}
17097 	if (rsm->r_flags & RACK_ACKED) {
17098 		return (NULL);
17099 	}
17100 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17101 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17102 		/* Its not yet ready */
17103 		return (NULL);
17104 	}
17105 	srtt = rack_grab_rtt(tp, rack);
17106 	idx = rsm->r_rtr_cnt - 1;
17107 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17108 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
17109 	if ((tsused == ts_low) ||
17110 	    (TSTMP_LT(tsused, ts_low))) {
17111 		/* No time since sending */
17112 		return (NULL);
17113 	}
17114 	if ((tsused - ts_low) < thresh) {
17115 		/* It has not been long enough yet */
17116 		return (NULL);
17117 	}
17118 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17119 	    ((rsm->r_flags & RACK_SACK_PASSED))) {
17120 		/*
17121 		 * We have passed the dup-ack threshold <or>
17122 		 * a SACK has indicated this is missing.
17123 		 * Note that if you are a declared attacker
17124 		 * it is only the dup-ack threshold that
17125 		 * will cause retransmits.
17126 		 */
17127 		/* log retransmit reason */
17128 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17129 		rack->r_fast_output = 0;
17130 		return (rsm);
17131 	}
17132 	return (NULL);
17133 }
17134 
17135 static void
17136 rack_log_pacing_delay_calc (struct tcp_rack *rack, uint32_t len, uint32_t slot,
17137 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17138 			   int line, struct rack_sendmap *rsm, uint8_t quality)
17139 {
17140 	if (tcp_bblogging_on(rack->rc_tp)) {
17141 		union tcp_log_stackspecific log;
17142 		struct timeval tv;
17143 
17144 		if (rack_verbose_logging == 0) {
17145 			/*
17146 			 * We are not verbose screen out all but
17147 			 * ones we always want.
17148 			 */
17149 			if ((method != 2) &&
17150 			    (method != 3) &&
17151 			    (method != 7) &&
17152 			    (method != 89) &&
17153 			    (method != 14) &&
17154 			    (method != 20)) {
17155 				return;
17156 			}
17157 		}
17158 		memset(&log, 0, sizeof(log));
17159 		log.u_bbr.flex1 = slot;
17160 		log.u_bbr.flex2 = len;
17161 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17162 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17163 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17164 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17165 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17166 		log.u_bbr.use_lt_bw <<= 1;
17167 		log.u_bbr.use_lt_bw |= rack->r_late;
17168 		log.u_bbr.use_lt_bw <<= 1;
17169 		log.u_bbr.use_lt_bw |= rack->r_early;
17170 		log.u_bbr.use_lt_bw <<= 1;
17171 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17172 		log.u_bbr.use_lt_bw <<= 1;
17173 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17174 		log.u_bbr.use_lt_bw <<= 1;
17175 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17176 		log.u_bbr.use_lt_bw <<= 1;
17177 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17178 		log.u_bbr.use_lt_bw <<= 1;
17179 		log.u_bbr.use_lt_bw |= rack->gp_ready;
17180 		log.u_bbr.pkt_epoch = line;
17181 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17182 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17183 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17184 		log.u_bbr.bw_inuse = bw_est;
17185 		log.u_bbr.delRate = bw;
17186 		if (rack->r_ctl.gp_bw == 0)
17187 			log.u_bbr.cur_del_rate = 0;
17188 		else
17189 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
17190 		log.u_bbr.rttProp = len_time;
17191 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17192 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17193 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17194 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17195 			/* We are in slow start */
17196 			log.u_bbr.flex7 = 1;
17197 		} else {
17198 			/* we are on congestion avoidance */
17199 			log.u_bbr.flex7 = 0;
17200 		}
17201 		log.u_bbr.flex8 = method;
17202 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17203 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17204 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17205 		log.u_bbr.cwnd_gain <<= 1;
17206 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17207 		log.u_bbr.cwnd_gain <<= 1;
17208 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17209 		log.u_bbr.bbr_substate = quality;
17210 		log.u_bbr.bbr_state = rack->dgp_on;
17211 		log.u_bbr.bbr_state <<= 1;
17212 		log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17213 		log.u_bbr.bbr_state <<= 2;
17214 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
17215 		    &rack->rc_inp->inp_socket->so_rcv,
17216 		    &rack->rc_inp->inp_socket->so_snd,
17217 		    BBR_LOG_HPTSI_CALC, 0,
17218 		    0, &log, false, &tv);
17219 	}
17220 }
17221 
17222 static uint32_t
17223 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17224 {
17225 	uint32_t new_tso, user_max, pace_one;
17226 
17227 	user_max = rack->rc_user_set_max_segs * mss;
17228 	if (rack->rc_force_max_seg) {
17229 		return (user_max);
17230 	}
17231 	if (rack->use_fixed_rate &&
17232 	    ((rack->r_ctl.crte == NULL) ||
17233 	     (bw != rack->r_ctl.crte->rate))) {
17234 		/* Use the user mss since we are not exactly matched */
17235 		return (user_max);
17236 	}
17237 	if (rack_pace_one_seg ||
17238 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17239 		pace_one = 1;
17240 	else
17241 		pace_one = 0;
17242 
17243 	new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17244 		     pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17245 	if (new_tso > user_max)
17246 		new_tso = user_max;
17247 	if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17248 		if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17249 			new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17250 	}
17251 	if (rack->r_ctl.rc_user_set_min_segs &&
17252 	    ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17253 	    new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17254 	return (new_tso);
17255 }
17256 
17257 static uint64_t
17258 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
17259 {
17260 	uint64_t reduced_win;
17261 	uint32_t gain;
17262 
17263 	if (window_input < rc_init_window(rack)) {
17264 		/*
17265 		 * The cwnd is collapsed to
17266 		 * nearly zero, maybe because of a time-out?
17267 		 * Lets drop back to the lt-bw.
17268 		 */
17269 		reduced_win = rack_get_lt_bw(rack);
17270 		/* Set the flag so the caller knows its a rate and not a reduced window */
17271 		*rate_set = 1;
17272 		gain = 100;
17273 	} else if  (IN_RECOVERY(rack->rc_tp->t_flags)) {
17274 		/*
17275 		 * If we are in recover our cwnd needs to be less for
17276 		 * our pacing consideration.
17277 		 */
17278 		if (rack->rack_hibeta == 0) {
17279 			reduced_win = window_input / 2;
17280 			gain = 50;
17281 		} else {
17282 			reduced_win = window_input * rack->r_ctl.saved_hibeta;
17283 			reduced_win /= 100;
17284 			gain = rack->r_ctl.saved_hibeta;
17285 		}
17286 	} else {
17287 		/*
17288 		 * Apply Timely factor to increase/decrease the
17289 		 * amount we are pacing at.
17290 		 */
17291 		gain = rack_get_output_gain(rack, NULL);
17292 		if (gain > rack_gain_p5_ub) {
17293 			gain = rack_gain_p5_ub;
17294 		}
17295 		reduced_win = window_input * gain;
17296 		reduced_win /= 100;
17297 	}
17298 	if (gain_b != NULL)
17299 		*gain_b = gain;
17300 	/*
17301 	 * What is being returned here is a trimmed down
17302 	 * window values in all cases where rate_set is left
17303 	 * at 0. In one case we actually return the rate (lt_bw).
17304 	 * the "reduced_win" is returned as a slimmed down cwnd that
17305 	 * is then calculated by the caller into a rate when rate_set
17306 	 * is 0.
17307 	 */
17308 	return (reduced_win);
17309 }
17310 
17311 static int32_t
17312 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)
17313 {
17314 	uint64_t lentim, fill_bw;
17315 
17316 	rack->r_via_fill_cw = 0;
17317 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17318 		return (slot);
17319 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17320 		return (slot);
17321 	if (rack->r_ctl.rc_last_us_rtt == 0)
17322 		return (slot);
17323 	if (rack->rc_pace_fill_if_rttin_range &&
17324 	    (rack->r_ctl.rc_last_us_rtt >=
17325 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17326 		/* The rtt is huge, N * smallest, lets not fill */
17327 		return (slot);
17328 	}
17329 	if (rack->r_ctl.fillcw_cap && *rate_wanted >= rack->r_ctl.fillcw_cap)
17330 		return (slot);
17331 	/*
17332 	 * first lets calculate the b/w based on the last us-rtt
17333 	 * and the the smallest send window.
17334 	 */
17335 	fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17336 	if (rack->rc_fillcw_apply_discount) {
17337 		uint32_t rate_set = 0;
17338 
17339 		fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
17340 		if (rate_set) {
17341 			goto at_lt_bw;
17342 		}
17343 	}
17344 	/* Take the rwnd if its smaller */
17345 	if (fill_bw > rack->rc_tp->snd_wnd)
17346 		fill_bw = rack->rc_tp->snd_wnd;
17347 	/* Now lets make it into a b/w */
17348 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17349 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17350 	/* Adjust to any cap */
17351 	if (rack->r_ctl.fillcw_cap && fill_bw >= rack->r_ctl.fillcw_cap)
17352 		fill_bw = rack->r_ctl.fillcw_cap;
17353 
17354 at_lt_bw:
17355 	if (rack_bw_multipler > 0) {
17356 		/*
17357 		 * We want to limit fill-cw to the some multiplier
17358 		 * of the max(lt_bw, gp_est). The normal default
17359 		 * is 0 for off, so a sysctl has enabled it.
17360 		 */
17361 		uint64_t lt_bw, gp, rate;
17362 
17363 		gp = rack_get_gp_est(rack);
17364 		lt_bw = rack_get_lt_bw(rack);
17365 		if (lt_bw > gp)
17366 			rate = lt_bw;
17367 		else
17368 			rate = gp;
17369 		rate *= rack_bw_multipler;
17370 		rate /= 100;
17371 		if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
17372 			union tcp_log_stackspecific log;
17373 			struct timeval tv;
17374 
17375 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17376 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17377 			log.u_bbr.flex1 = rack_bw_multipler;
17378 			log.u_bbr.flex2 = len;
17379 			log.u_bbr.cur_del_rate = gp;
17380 			log.u_bbr.delRate = lt_bw;
17381 			log.u_bbr.bw_inuse = rate;
17382 			log.u_bbr.rttProp = fill_bw;
17383 			log.u_bbr.flex8 = 44;
17384 			tcp_log_event(rack->rc_tp, NULL, NULL, NULL,
17385 				      BBR_LOG_CWND, 0,
17386 				      0, &log, false, NULL,
17387 				      __func__, __LINE__, &tv);
17388 		}
17389 		if (fill_bw > rate)
17390 			fill_bw = rate;
17391 	}
17392 	/* We are below the min b/w */
17393 	if (non_paced)
17394 		*rate_wanted = fill_bw;
17395 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17396 		return (slot);
17397 	rack->r_via_fill_cw = 1;
17398 	if (rack->r_rack_hw_rate_caps &&
17399 	    (rack->r_ctl.crte != NULL)) {
17400 		uint64_t high_rate;
17401 
17402 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17403 		if (fill_bw > high_rate) {
17404 			/* We are capping bw at the highest rate table entry */
17405 			if (*rate_wanted > high_rate) {
17406 				/* The original rate was also capped */
17407 				rack->r_via_fill_cw = 0;
17408 			}
17409 			rack_log_hdwr_pacing(rack,
17410 					     fill_bw, high_rate, __LINE__,
17411 					     0, 3);
17412 			fill_bw = high_rate;
17413 			if (capped)
17414 				*capped = 1;
17415 		}
17416 	} else if ((rack->r_ctl.crte == NULL) &&
17417 		   (rack->rack_hdrw_pacing == 0) &&
17418 		   (rack->rack_hdw_pace_ena) &&
17419 		   rack->r_rack_hw_rate_caps &&
17420 		   (rack->rack_attempt_hdwr_pace == 0) &&
17421 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
17422 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17423 		/*
17424 		 * Ok we may have a first attempt that is greater than our top rate
17425 		 * lets check.
17426 		 */
17427 		uint64_t high_rate;
17428 
17429 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17430 		if (high_rate) {
17431 			if (fill_bw > high_rate) {
17432 				fill_bw = high_rate;
17433 				if (capped)
17434 					*capped = 1;
17435 			}
17436 		}
17437 	}
17438 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17439 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17440 				   fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
17441 		fill_bw = rack->r_ctl.bw_rate_cap;
17442 	}
17443 	/*
17444 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
17445 	 * in an rtt (unless it was capped), what does that
17446 	 * time wise equate too?
17447 	 */
17448 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17449 	lentim /= fill_bw;
17450 	*rate_wanted = fill_bw;
17451 	if (non_paced || (lentim < slot)) {
17452 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17453 					   0, lentim, 12, __LINE__, NULL, 0);
17454 		return ((int32_t)lentim);
17455 	} else
17456 		return (slot);
17457 }
17458 
17459 static int32_t
17460 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz, int line)
17461 {
17462 	uint64_t srtt;
17463 	int32_t slot = 0;
17464 	int32_t minslot = 0;
17465 	int can_start_hw_pacing = 1;
17466 	int err;
17467 	int pace_one;
17468 
17469 	if (rack_pace_one_seg ||
17470 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17471 		pace_one = 1;
17472 	else
17473 		pace_one = 0;
17474 	if (rack->rc_always_pace == 0) {
17475 		/*
17476 		 * We use the most optimistic possible cwnd/srtt for
17477 		 * sending calculations. This will make our
17478 		 * calculation anticipate getting more through
17479 		 * quicker then possible. But thats ok we don't want
17480 		 * the peer to have a gap in data sending.
17481 		 */
17482 		uint64_t cwnd, tr_perms = 0;
17483 		int32_t reduce = 0;
17484 
17485 	old_method:
17486 		/*
17487 		 * We keep no precise pacing with the old method
17488 		 * instead we use the pacer to mitigate bursts.
17489 		 */
17490 		if (rack->r_ctl.rc_rack_min_rtt)
17491 			srtt = rack->r_ctl.rc_rack_min_rtt;
17492 		else
17493 			srtt = max(tp->t_srtt, 1);
17494 		if (rack->r_ctl.rc_rack_largest_cwnd)
17495 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17496 		else
17497 			cwnd = rack->r_ctl.cwnd_to_use;
17498 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
17499 		tr_perms = (cwnd * 1000) / srtt;
17500 		if (tr_perms == 0) {
17501 			tr_perms = ctf_fixed_maxseg(tp);
17502 		}
17503 		/*
17504 		 * Calculate how long this will take to drain, if
17505 		 * the calculation comes out to zero, thats ok we
17506 		 * will use send_a_lot to possibly spin around for
17507 		 * more increasing tot_len_this_send to the point
17508 		 * that its going to require a pace, or we hit the
17509 		 * cwnd. Which in that case we are just waiting for
17510 		 * a ACK.
17511 		 */
17512 		slot = len / tr_perms;
17513 		/* Now do we reduce the time so we don't run dry? */
17514 		if (slot && rack_slot_reduction) {
17515 			reduce = (slot / rack_slot_reduction);
17516 			if (reduce < slot) {
17517 				slot -= reduce;
17518 			} else
17519 				slot = 0;
17520 		}
17521 		slot *= HPTS_USEC_IN_MSEC;
17522 		if (rack->rc_pace_to_cwnd) {
17523 			uint64_t rate_wanted = 0;
17524 
17525 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17526 			rack->rc_ack_can_sendout_data = 1;
17527 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17528 		} else
17529 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17530 		/*******************************************************/
17531 		/* RRS: We insert non-paced call to stats here for len */
17532 		/*******************************************************/
17533 	} else {
17534 		uint64_t bw_est, res, lentim, rate_wanted;
17535 		uint32_t segs, oh;
17536 		int capped = 0;
17537 		int prev_fill;
17538 
17539 		if ((rack->r_rr_config == 1) && rsm) {
17540 			return (rack->r_ctl.rc_min_to);
17541 		}
17542 		if (rack->use_fixed_rate) {
17543 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17544 		} else if ((rack->r_ctl.init_rate == 0) &&
17545 			   (rack->r_ctl.gp_bw == 0)) {
17546 			/* no way to yet do an estimate */
17547 			bw_est = rate_wanted = 0;
17548 		} else if (rack->dgp_on) {
17549 			bw_est = rack_get_bw(rack);
17550 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17551 		} else {
17552 			uint32_t gain, rate_set = 0;
17553 
17554 			rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17555 			rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
17556 			if (rate_set == 0) {
17557 				if (rate_wanted > rack->rc_tp->snd_wnd)
17558 					rate_wanted = rack->rc_tp->snd_wnd;
17559 				/* Now lets make it into a b/w */
17560 				rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
17561 				rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17562 			}
17563 			bw_est = rate_wanted;
17564 			rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
17565 						   rack->r_ctl.cwnd_to_use,
17566 						   rate_wanted, bw_est,
17567 						   rack->r_ctl.rc_last_us_rtt,
17568 						   88, __LINE__, NULL, gain);
17569 		}
17570 		if ((bw_est == 0) || (rate_wanted == 0) ||
17571 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
17572 			/*
17573 			 * No way yet to make a b/w estimate or
17574 			 * our raise is set incorrectly.
17575 			 */
17576 			goto old_method;
17577 		}
17578 		rack_rate_cap_bw(rack, &rate_wanted, &capped);
17579 		/* We need to account for all the overheads */
17580 		segs = (len + segsiz - 1) / segsiz;
17581 		/*
17582 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
17583 		 * and how much data we put in each packet. Yes this
17584 		 * means we may be off if we are larger than 1500 bytes
17585 		 * or smaller. But this just makes us more conservative.
17586 		 */
17587 
17588 		oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17589 		if (rack->r_is_v6) {
17590 #ifdef INET6
17591 			oh += sizeof(struct ip6_hdr);
17592 #endif
17593 		} else {
17594 #ifdef INET
17595 			oh += sizeof(struct ip);
17596 #endif
17597 		}
17598 		/* We add a fixed 14 for the ethernet header */
17599 		oh += 14;
17600 		segs *= oh;
17601 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17602 		res = lentim / rate_wanted;
17603 		slot = (uint32_t)res;
17604 		if (rack_hw_rate_min &&
17605 		    (rate_wanted < rack_hw_rate_min)) {
17606 			can_start_hw_pacing = 0;
17607 			if (rack->r_ctl.crte) {
17608 				/*
17609 				 * Ok we need to release it, we
17610 				 * have fallen too low.
17611 				 */
17612 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17613 				rack->r_ctl.crte = NULL;
17614 				rack->rack_attempt_hdwr_pace = 0;
17615 				rack->rack_hdrw_pacing = 0;
17616 			}
17617 		}
17618 		if (rack->r_ctl.crte &&
17619 		    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17620 			/*
17621 			 * We want more than the hardware can give us,
17622 			 * don't start any hw pacing.
17623 			 */
17624 			can_start_hw_pacing = 0;
17625 			if (rack->r_rack_hw_rate_caps == 0) {
17626 				/*
17627 				 * Ok we need to release it, we
17628 				 * want more than the card can give us and
17629 				 * no rate cap is in place. Set it up so
17630 				 * when we want less we can retry.
17631 				 */
17632 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17633 				rack->r_ctl.crte = NULL;
17634 				rack->rack_attempt_hdwr_pace = 0;
17635 				rack->rack_hdrw_pacing = 0;
17636 			}
17637 		}
17638 		if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17639 			/*
17640 			 * We lost our rate somehow, this can happen
17641 			 * if the interface changed underneath us.
17642 			 */
17643 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17644 			rack->r_ctl.crte = NULL;
17645 			/* Lets re-allow attempting to setup pacing */
17646 			rack->rack_hdrw_pacing = 0;
17647 			rack->rack_attempt_hdwr_pace = 0;
17648 			rack_log_hdwr_pacing(rack,
17649 					     rate_wanted, bw_est, __LINE__,
17650 					     0, 6);
17651 		}
17652 		prev_fill = rack->r_via_fill_cw;
17653 		if ((rack->rc_pace_to_cwnd) &&
17654 		    (capped == 0) &&
17655 		    (rack->dgp_on == 1) &&
17656 		    (rack->use_fixed_rate == 0) &&
17657 		    (rack->in_probe_rtt == 0) &&
17658 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17659 			/*
17660 			 * We want to pace at our rate *or* faster to
17661 			 * fill the cwnd to the max if its not full.
17662 			 */
17663 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17664 			/* Re-check to make sure we are not exceeding our max b/w */
17665 			if ((rack->r_ctl.crte != NULL) &&
17666 			    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17667 				/*
17668 				 * We want more than the hardware can give us,
17669 				 * don't start any hw pacing.
17670 				 */
17671 				can_start_hw_pacing = 0;
17672 				if (rack->r_rack_hw_rate_caps == 0) {
17673 					/*
17674 					 * Ok we need to release it, we
17675 					 * want more than the card can give us and
17676 					 * no rate cap is in place. Set it up so
17677 					 * when we want less we can retry.
17678 					 */
17679 					tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17680 					rack->r_ctl.crte = NULL;
17681 					rack->rack_attempt_hdwr_pace = 0;
17682 					rack->rack_hdrw_pacing = 0;
17683 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17684 				}
17685 			}
17686 		}
17687 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17688 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17689 			if ((rack->rack_hdw_pace_ena) &&
17690 			    (can_start_hw_pacing > 0) &&
17691 			    (rack->rack_hdrw_pacing == 0) &&
17692 			    (rack->rack_attempt_hdwr_pace == 0)) {
17693 				/*
17694 				 * Lets attempt to turn on hardware pacing
17695 				 * if we can.
17696 				 */
17697 				rack->rack_attempt_hdwr_pace = 1;
17698 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17699 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
17700 								       rate_wanted,
17701 								       RS_PACING_GEQ,
17702 								       &err, &rack->r_ctl.crte_prev_rate);
17703 				if (rack->r_ctl.crte) {
17704 					rack->rack_hdrw_pacing = 1;
17705 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17706 													   pace_one, rack->r_ctl.crte,
17707 													   NULL, rack->r_ctl.pace_len_divisor);
17708 					rack_log_hdwr_pacing(rack,
17709 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17710 							     err, 0);
17711 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17712 				} else {
17713 					counter_u64_add(rack_hw_pace_init_fail, 1);
17714 				}
17715 			} else if (rack->rack_hdrw_pacing &&
17716 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17717 				/* Do we need to adjust our rate? */
17718 				const struct tcp_hwrate_limit_table *nrte;
17719 
17720 				if (rack->r_up_only &&
17721 				    (rate_wanted < rack->r_ctl.crte->rate)) {
17722 					/**
17723 					 * We have four possible states here
17724 					 * having to do with the previous time
17725 					 * and this time.
17726 					 *   previous  |  this-time
17727 					 * A)     0      |     0   -- fill_cw not in the picture
17728 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
17729 					 * C)     1      |     1   -- all rates from fill_cw
17730 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
17731 					 *
17732 					 * For case A, C and D we don't allow a drop. But for
17733 					 * case B where we now our on our steady rate we do
17734 					 * allow a drop.
17735 					 *
17736 					 */
17737 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17738 						goto done_w_hdwr;
17739 				}
17740 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
17741 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17742 					if (rack_hw_rate_to_low &&
17743 					    (bw_est < rack_hw_rate_to_low)) {
17744 						/*
17745 						 * The pacing rate is too low for hardware, but
17746 						 * do allow hardware pacing to be restarted.
17747 						 */
17748 						rack_log_hdwr_pacing(rack,
17749 								     bw_est, rack->r_ctl.crte->rate, __LINE__,
17750 								     0, 5);
17751 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17752 						rack->r_ctl.crte = NULL;
17753 						rack->rack_attempt_hdwr_pace = 0;
17754 						rack->rack_hdrw_pacing = 0;
17755 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17756 						goto done_w_hdwr;
17757 					}
17758 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17759 								   rack->rc_tp,
17760 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
17761 								   rate_wanted,
17762 								   RS_PACING_GEQ,
17763 								   &err, &rack->r_ctl.crte_prev_rate);
17764 					if (nrte == NULL) {
17765 						/*
17766 						 * Lost the rate, lets drop hardware pacing
17767 						 * period.
17768 						 */
17769 						rack->rack_hdrw_pacing = 0;
17770 						rack->r_ctl.crte = NULL;
17771 						rack_log_hdwr_pacing(rack,
17772 								     rate_wanted, 0, __LINE__,
17773 								     err, 1);
17774 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17775 						counter_u64_add(rack_hw_pace_lost, 1);
17776 					} else if (nrte != rack->r_ctl.crte) {
17777 						rack->r_ctl.crte = nrte;
17778 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17779 														   segsiz, pace_one, rack->r_ctl.crte,
17780 														   NULL, rack->r_ctl.pace_len_divisor);
17781 						rack_log_hdwr_pacing(rack,
17782 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17783 								     err, 2);
17784 						rack->r_ctl.last_hw_bw_req = rate_wanted;
17785 					}
17786 				} else {
17787 					/* We just need to adjust the segment size */
17788 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17789 					rack_log_hdwr_pacing(rack,
17790 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17791 							     0, 4);
17792 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17793 				}
17794 			}
17795 		}
17796 		if (minslot && (minslot > slot)) {
17797 			rack_log_pacing_delay_calc(rack, minslot, slot, rack->r_ctl.crte->rate, bw_est, lentim,
17798 						   98, __LINE__, NULL, 0);
17799 			slot = minslot;
17800 		}
17801 	done_w_hdwr:
17802 		if (rack_limit_time_with_srtt &&
17803 		    (rack->use_fixed_rate == 0) &&
17804 		    (rack->rack_hdrw_pacing == 0)) {
17805 			/*
17806 			 * Sanity check, we do not allow the pacing delay
17807 			 * to be longer than the SRTT of the path. If it is
17808 			 * a slow path, then adding a packet should increase
17809 			 * the RTT and compensate for this i.e. the srtt will
17810 			 * be greater so the allowed pacing time will be greater.
17811 			 *
17812 			 * Note this restriction is not for where a peak rate
17813 			 * is set, we are doing fixed pacing or hardware pacing.
17814 			 */
17815 			if (rack->rc_tp->t_srtt)
17816 				srtt = rack->rc_tp->t_srtt;
17817 			else
17818 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
17819 			if (srtt < (uint64_t)slot) {
17820 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
17821 				slot = srtt;
17822 			}
17823 		}
17824 		/*******************************************************************/
17825 		/* RRS: We insert paced call to stats here for len and rate_wanted */
17826 		/*******************************************************************/
17827 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
17828 	}
17829 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
17830 		/*
17831 		 * If this rate is seeing enobufs when it
17832 		 * goes to send then either the nic is out
17833 		 * of gas or we are mis-estimating the time
17834 		 * somehow and not letting the queue empty
17835 		 * completely. Lets add to the pacing time.
17836 		 */
17837 		int hw_boost_delay;
17838 
17839 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
17840 		if (hw_boost_delay > rack_enobuf_hw_max)
17841 			hw_boost_delay = rack_enobuf_hw_max;
17842 		else if (hw_boost_delay < rack_enobuf_hw_min)
17843 			hw_boost_delay = rack_enobuf_hw_min;
17844 		slot += hw_boost_delay;
17845 	}
17846 	return (slot);
17847 }
17848 
17849 static void
17850 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
17851     tcp_seq startseq, uint32_t sb_offset)
17852 {
17853 	struct rack_sendmap *my_rsm = NULL;
17854 
17855 	if (tp->t_state < TCPS_ESTABLISHED) {
17856 		/*
17857 		 * We don't start any measurements if we are
17858 		 * not at least established.
17859 		 */
17860 		return;
17861 	}
17862 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
17863 		/*
17864 		 * We will get no more data into the SB
17865 		 * this means we need to have the data available
17866 		 * before we start a measurement.
17867 		 */
17868 
17869 		if (sbavail(&tptosocket(tp)->so_snd) <
17870 		    max(rc_init_window(rack),
17871 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
17872 			/* Nope not enough data */
17873 			return;
17874 		}
17875 	}
17876 	tp->t_flags |= TF_GPUTINPROG;
17877 	rack->r_ctl.rc_gp_cumack_ts = 0;
17878 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
17879 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
17880 	tp->gput_seq = startseq;
17881 	rack->app_limited_needs_set = 0;
17882 	if (rack->in_probe_rtt)
17883 		rack->measure_saw_probe_rtt = 1;
17884 	else if ((rack->measure_saw_probe_rtt) &&
17885 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
17886 		rack->measure_saw_probe_rtt = 0;
17887 	if (rack->rc_gp_filled)
17888 		tp->gput_ts = rack->r_ctl.last_cumack_advance;
17889 	else {
17890 		/* Special case initial measurement */
17891 		struct timeval tv;
17892 
17893 		tp->gput_ts = tcp_get_usecs(&tv);
17894 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
17895 	}
17896 	/*
17897 	 * We take a guess out into the future,
17898 	 * if we have no measurement and no
17899 	 * initial rate, we measure the first
17900 	 * initial-windows worth of data to
17901 	 * speed up getting some GP measurement and
17902 	 * thus start pacing.
17903 	 */
17904 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
17905 		rack->app_limited_needs_set = 1;
17906 		tp->gput_ack = startseq + max(rc_init_window(rack),
17907 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
17908 		rack_log_pacing_delay_calc(rack,
17909 					   tp->gput_seq,
17910 					   tp->gput_ack,
17911 					   0,
17912 					   tp->gput_ts,
17913 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17914 					   9,
17915 					   __LINE__, NULL, 0);
17916 		rack_tend_gp_marks(tp, rack);
17917 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
17918 		return;
17919 	}
17920 	if (sb_offset) {
17921 		/*
17922 		 * We are out somewhere in the sb
17923 		 * can we use the already outstanding data?
17924 		 */
17925 
17926 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
17927 			/*
17928 			 * Yes first one is good and in this case
17929 			 * the tp->gput_ts is correctly set based on
17930 			 * the last ack that arrived (no need to
17931 			 * set things up when an ack comes in).
17932 			 */
17933 			my_rsm = tqhash_min(rack->r_ctl.tqh);
17934 			if ((my_rsm == NULL) ||
17935 			    (my_rsm->r_rtr_cnt != 1)) {
17936 				/* retransmission? */
17937 				goto use_latest;
17938 			}
17939 		} else {
17940 			if (rack->r_ctl.rc_first_appl == NULL) {
17941 				/*
17942 				 * If rc_first_appl is NULL
17943 				 * then the cnt should be 0.
17944 				 * This is probably an error, maybe
17945 				 * a KASSERT would be approprate.
17946 				 */
17947 				goto use_latest;
17948 			}
17949 			/*
17950 			 * If we have a marker pointer to the last one that is
17951 			 * app limited we can use that, but we need to set
17952 			 * things up so that when it gets ack'ed we record
17953 			 * the ack time (if its not already acked).
17954 			 */
17955 			rack->app_limited_needs_set = 1;
17956 			/*
17957 			 * We want to get to the rsm that is either
17958 			 * next with space i.e. over 1 MSS or the one
17959 			 * after that (after the app-limited).
17960 			 */
17961 			my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
17962 			if (my_rsm) {
17963 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
17964 					/* Have to use the next one */
17965 					my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17966 				else {
17967 					/* Use after the first MSS of it is acked */
17968 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
17969 					goto start_set;
17970 				}
17971 			}
17972 			if ((my_rsm == NULL) ||
17973 			    (my_rsm->r_rtr_cnt != 1)) {
17974 				/*
17975 				 * Either its a retransmit or
17976 				 * the last is the app-limited one.
17977 				 */
17978 				goto use_latest;
17979 			}
17980 		}
17981 		tp->gput_seq = my_rsm->r_start;
17982 start_set:
17983 		if (my_rsm->r_flags & RACK_ACKED) {
17984 			/*
17985 			 * This one has been acked use the arrival ack time
17986 			 */
17987 			struct rack_sendmap *nrsm;
17988 
17989 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
17990 			rack->app_limited_needs_set = 0;
17991 			/*
17992 			 * Ok in this path we need to use the r_end now
17993 			 * since this guy is the starting ack.
17994 			 */
17995 			tp->gput_seq = my_rsm->r_end;
17996 			/*
17997 			 * We also need to adjust up the sendtime
17998 			 * to the send of the next data after my_rsm.
17999 			 */
18000 			nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
18001 			if (nrsm != NULL)
18002 				my_rsm = nrsm;
18003 			else {
18004 				/*
18005 				 * The next as not been sent, thats the
18006 				 * case for using the latest.
18007 				 */
18008 				goto use_latest;
18009 			}
18010 		}
18011 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18012 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
18013 		rack->r_ctl.rc_gp_cumack_ts = 0;
18014 		if ((rack->r_ctl.cleared_app_ack == 1) &&
18015 		    (SEQ_GEQ(rack->r_ctl.cleared_app_ack, tp->gput_seq))) {
18016 			/*
18017 			 * We just cleared an application limited period
18018 			 * so the next seq out needs to skip the first
18019 			 * ack.
18020 			 */
18021 			rack->app_limited_needs_set = 1;
18022 			rack->r_ctl.cleared_app_ack = 0;
18023 		}
18024 		rack_log_pacing_delay_calc(rack,
18025 					   tp->gput_seq,
18026 					   tp->gput_ack,
18027 					   (uintptr_t)my_rsm,
18028 					   tp->gput_ts,
18029 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18030 					   9,
18031 					   __LINE__, my_rsm, 0);
18032 		/* Now lets make sure all are marked as they should be */
18033 		rack_tend_gp_marks(tp, rack);
18034 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18035 		return;
18036 	}
18037 
18038 use_latest:
18039 	/*
18040 	 * We don't know how long we may have been
18041 	 * idle or if this is the first-send. Lets
18042 	 * setup the flag so we will trim off
18043 	 * the first ack'd data so we get a true
18044 	 * measurement.
18045 	 */
18046 	rack->app_limited_needs_set = 1;
18047 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
18048 	rack->r_ctl.rc_gp_cumack_ts = 0;
18049 	/* Find this guy so we can pull the send time */
18050 	my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
18051 	if (my_rsm) {
18052 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18053 		if (my_rsm->r_flags & RACK_ACKED) {
18054 			/*
18055 			 * Unlikely since its probably what was
18056 			 * just transmitted (but I am paranoid).
18057 			 */
18058 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18059 			rack->app_limited_needs_set = 0;
18060 		}
18061 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18062 			/* This also is unlikely */
18063 			tp->gput_seq = my_rsm->r_start;
18064 		}
18065 	} else {
18066 		/*
18067 		 * TSNH unless we have some send-map limit,
18068 		 * and even at that it should not be hitting
18069 		 * that limit (we should have stopped sending).
18070 		 */
18071 		struct timeval tv;
18072 
18073 		microuptime(&tv);
18074 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18075 	}
18076 	rack_tend_gp_marks(tp, rack);
18077 	rack_log_pacing_delay_calc(rack,
18078 				   tp->gput_seq,
18079 				   tp->gput_ack,
18080 				   (uintptr_t)my_rsm,
18081 				   tp->gput_ts,
18082 				   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18083 				   9, __LINE__, NULL, 0);
18084 	rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18085 }
18086 
18087 static inline uint32_t
18088 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
18089     uint32_t avail, int32_t sb_offset)
18090 {
18091 	uint32_t len;
18092 	uint32_t sendwin;
18093 
18094 	if (tp->snd_wnd > cwnd_to_use)
18095 		sendwin = cwnd_to_use;
18096 	else
18097 		sendwin = tp->snd_wnd;
18098 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
18099 		/* We never want to go over our peers rcv-window */
18100 		len = 0;
18101 	} else {
18102 		uint32_t flight;
18103 
18104 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18105 		if (flight >= sendwin) {
18106 			/*
18107 			 * We have in flight what we are allowed by cwnd (if
18108 			 * it was rwnd blocking it would have hit above out
18109 			 * >= tp->snd_wnd).
18110 			 */
18111 			return (0);
18112 		}
18113 		len = sendwin - flight;
18114 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18115 			/* We would send too much (beyond the rwnd) */
18116 			len = tp->snd_wnd - ctf_outstanding(tp);
18117 		}
18118 		if ((len + sb_offset) > avail) {
18119 			/*
18120 			 * We don't have that much in the SB, how much is
18121 			 * there?
18122 			 */
18123 			len = avail - sb_offset;
18124 		}
18125 	}
18126 	return (len);
18127 }
18128 
18129 static void
18130 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18131 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18132 	     int rsm_is_null, int optlen, int line, uint16_t mode)
18133 {
18134 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18135 		union tcp_log_stackspecific log;
18136 		struct timeval tv;
18137 
18138 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18139 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18140 		log.u_bbr.flex1 = error;
18141 		log.u_bbr.flex2 = flags;
18142 		log.u_bbr.flex3 = rsm_is_null;
18143 		log.u_bbr.flex4 = ipoptlen;
18144 		log.u_bbr.flex5 = tp->rcv_numsacks;
18145 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18146 		log.u_bbr.flex7 = optlen;
18147 		log.u_bbr.flex8 = rack->r_fsb_inited;
18148 		log.u_bbr.applimited = rack->r_fast_output;
18149 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18150 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18151 		log.u_bbr.cwnd_gain = mode;
18152 		log.u_bbr.pkts_out = orig_len;
18153 		log.u_bbr.lt_epoch = len;
18154 		log.u_bbr.delivered = line;
18155 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18156 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18157 		tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18158 			       len, &log, false, NULL, __func__, __LINE__, &tv);
18159 	}
18160 }
18161 
18162 
18163 static struct mbuf *
18164 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18165 		   struct rack_fast_send_blk *fsb,
18166 		   int32_t seglimit, int32_t segsize, int hw_tls)
18167 {
18168 #ifdef KERN_TLS
18169 	struct ktls_session *tls, *ntls;
18170 #ifdef INVARIANTS
18171 	struct mbuf *start;
18172 #endif
18173 #endif
18174 	struct mbuf *m, *n, **np, *smb;
18175 	struct mbuf *top;
18176 	int32_t off, soff;
18177 	int32_t len = *plen;
18178 	int32_t fragsize;
18179 	int32_t len_cp = 0;
18180 	uint32_t mlen, frags;
18181 
18182 	soff = off = the_off;
18183 	smb = m = the_m;
18184 	np = &top;
18185 	top = NULL;
18186 #ifdef KERN_TLS
18187 	if (hw_tls && (m->m_flags & M_EXTPG))
18188 		tls = m->m_epg_tls;
18189 	else
18190 		tls = NULL;
18191 #ifdef INVARIANTS
18192 	start = m;
18193 #endif
18194 #endif
18195 	while (len > 0) {
18196 		if (m == NULL) {
18197 			*plen = len_cp;
18198 			break;
18199 		}
18200 #ifdef KERN_TLS
18201 		if (hw_tls) {
18202 			if (m->m_flags & M_EXTPG)
18203 				ntls = m->m_epg_tls;
18204 			else
18205 				ntls = NULL;
18206 
18207 			/*
18208 			 * Avoid mixing TLS records with handshake
18209 			 * data or TLS records from different
18210 			 * sessions.
18211 			 */
18212 			if (tls != ntls) {
18213 				MPASS(m != start);
18214 				*plen = len_cp;
18215 				break;
18216 			}
18217 		}
18218 #endif
18219 		mlen = min(len, m->m_len - off);
18220 		if (seglimit) {
18221 			/*
18222 			 * For M_EXTPG mbufs, add 3 segments
18223 			 * + 1 in case we are crossing page boundaries
18224 			 * + 2 in case the TLS hdr/trailer are used
18225 			 * It is cheaper to just add the segments
18226 			 * than it is to take the cache miss to look
18227 			 * at the mbuf ext_pgs state in detail.
18228 			 */
18229 			if (m->m_flags & M_EXTPG) {
18230 				fragsize = min(segsize, PAGE_SIZE);
18231 				frags = 3;
18232 			} else {
18233 				fragsize = segsize;
18234 				frags = 0;
18235 			}
18236 
18237 			/* Break if we really can't fit anymore. */
18238 			if ((frags + 1) >= seglimit) {
18239 				*plen =	len_cp;
18240 				break;
18241 			}
18242 
18243 			/*
18244 			 * Reduce size if you can't copy the whole
18245 			 * mbuf. If we can't copy the whole mbuf, also
18246 			 * adjust len so the loop will end after this
18247 			 * mbuf.
18248 			 */
18249 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18250 				mlen = (seglimit - frags - 1) * fragsize;
18251 				len = mlen;
18252 				*plen = len_cp + len;
18253 			}
18254 			frags += howmany(mlen, fragsize);
18255 			if (frags == 0)
18256 				frags++;
18257 			seglimit -= frags;
18258 			KASSERT(seglimit > 0,
18259 			    ("%s: seglimit went too low", __func__));
18260 		}
18261 		n = m_get(M_NOWAIT, m->m_type);
18262 		*np = n;
18263 		if (n == NULL)
18264 			goto nospace;
18265 		n->m_len = mlen;
18266 		soff += mlen;
18267 		len_cp += n->m_len;
18268 		if (m->m_flags & (M_EXT | M_EXTPG)) {
18269 			n->m_data = m->m_data + off;
18270 			mb_dupcl(n, m);
18271 		} else {
18272 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18273 			    (u_int)n->m_len);
18274 		}
18275 		len -= n->m_len;
18276 		off = 0;
18277 		m = m->m_next;
18278 		np = &n->m_next;
18279 		if (len || (soff == smb->m_len)) {
18280 			/*
18281 			 * We have more so we move forward  or
18282 			 * we have consumed the entire mbuf and
18283 			 * len has fell to 0.
18284 			 */
18285 			soff = 0;
18286 			smb = m;
18287 		}
18288 
18289 	}
18290 	if (fsb != NULL) {
18291 		fsb->m = smb;
18292 		fsb->off = soff;
18293 		if (smb) {
18294 			/*
18295 			 * Save off the size of the mbuf. We do
18296 			 * this so that we can recognize when it
18297 			 * has been trimmed by sbcut() as acks
18298 			 * come in.
18299 			 */
18300 			fsb->o_m_len = smb->m_len;
18301 			fsb->o_t_len = M_TRAILINGROOM(smb);
18302 		} else {
18303 			/*
18304 			 * This is the case where the next mbuf went to NULL. This
18305 			 * means with this copy we have sent everything in the sb.
18306 			 * In theory we could clear the fast_output flag, but lets
18307 			 * not since its possible that we could get more added
18308 			 * and acks that call the extend function which would let
18309 			 * us send more.
18310 			 */
18311 			fsb->o_m_len = 0;
18312 			fsb->o_t_len = 0;
18313 		}
18314 	}
18315 	return (top);
18316 nospace:
18317 	if (top)
18318 		m_freem(top);
18319 	return (NULL);
18320 
18321 }
18322 
18323 /*
18324  * This is a copy of m_copym(), taking the TSO segment size/limit
18325  * constraints into account, and advancing the sndptr as it goes.
18326  */
18327 static struct mbuf *
18328 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18329 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18330 {
18331 	struct mbuf *m, *n;
18332 	int32_t soff;
18333 
18334 	m = rack->r_ctl.fsb.m;
18335 	if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18336 		/*
18337 		 * The trailing space changed, mbufs can grow
18338 		 * at the tail but they can't shrink from
18339 		 * it, KASSERT that. Adjust the orig_m_len to
18340 		 * compensate for this change.
18341 		 */
18342 		KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18343 			("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18344 			 m,
18345 			 rack,
18346 			 (intmax_t)M_TRAILINGROOM(m),
18347 			 rack->r_ctl.fsb.o_t_len,
18348 			 rack->r_ctl.fsb.o_m_len,
18349 			 m->m_len));
18350 		rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18351 		rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18352 	}
18353 	if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18354 		/*
18355 		 * Mbuf shrank, trimmed off the top by an ack, our
18356 		 * offset changes.
18357 		 */
18358 		KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18359 			("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18360 			 m, m->m_len,
18361 			 rack, rack->r_ctl.fsb.o_m_len,
18362 			 rack->r_ctl.fsb.off));
18363 
18364 		if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18365 			rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18366 		else
18367 			rack->r_ctl.fsb.off = 0;
18368 		rack->r_ctl.fsb.o_m_len = m->m_len;
18369 #ifdef INVARIANTS
18370 	} else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18371 		panic("rack:%p m:%p m_len grew outside of t_space compensation",
18372 		      rack, m);
18373 #endif
18374 	}
18375 	soff = rack->r_ctl.fsb.off;
18376 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18377 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18378 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18379 				 __FUNCTION__,
18380 				 rack, *plen, m, m->m_len));
18381 	/* Save off the right location before we copy and advance */
18382 	*s_soff = soff;
18383 	*s_mb = rack->r_ctl.fsb.m;
18384 	n = rack_fo_base_copym(m, soff, plen,
18385 			       &rack->r_ctl.fsb,
18386 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18387 	return (n);
18388 }
18389 
18390 /* Log the buffer level */
18391 static void
18392 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18393 		     int len, struct timeval *tv,
18394 		     uint32_t cts)
18395 {
18396 	uint32_t p_rate = 0, p_queue = 0, err = 0;
18397 	union tcp_log_stackspecific log;
18398 
18399 #ifdef RATELIMIT
18400 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18401 	err = in_pcbquery_txrtlmt(rack->rc_inp,	&p_rate);
18402 #endif
18403 	memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18404 	log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18405 	log.u_bbr.flex1 = p_rate;
18406 	log.u_bbr.flex2 = p_queue;
18407 	log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18408 	log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18409 	log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18410 	log.u_bbr.flex7 = 99;
18411 	log.u_bbr.flex8 = 0;
18412 	log.u_bbr.pkts_out = err;
18413 	log.u_bbr.delRate = rack->r_ctl.crte->rate;
18414 	log.u_bbr.timeStamp = cts;
18415 	log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18416 	tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18417 		       len, &log, false, NULL, __func__, __LINE__, tv);
18418 
18419 }
18420 
18421 static uint32_t
18422 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18423 		       struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18424 {
18425 	uint64_t lentime = 0;
18426 #ifdef RATELIMIT
18427 	uint32_t p_rate = 0, p_queue = 0, err;
18428 	union tcp_log_stackspecific log;
18429 	uint64_t bw;
18430 
18431 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18432 	/* Failed or queue is zero */
18433 	if (err || (p_queue == 0)) {
18434 		lentime = 0;
18435 		goto out;
18436 	}
18437 	err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18438 	if (err) {
18439 		lentime = 0;
18440 		goto out;
18441 	}
18442 	/*
18443 	 * If we reach here we have some bytes in
18444 	 * the queue. The number returned is a value
18445 	 * between 0 and 0xffff where ffff is full
18446 	 * and 0 is empty. So how best to make this into
18447 	 * something usable?
18448 	 *
18449 	 * The "safer" way is lets take the b/w gotten
18450 	 * from the query (which should be our b/w rate)
18451 	 * and pretend that a full send (our rc_pace_max_segs)
18452 	 * is outstanding. We factor it so its as if a full
18453 	 * number of our MSS segment is terms of full
18454 	 * ethernet segments are outstanding.
18455 	 */
18456 	bw = p_rate / 8;
18457 	if (bw) {
18458 		lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18459 		lentime *= ETHERNET_SEGMENT_SIZE;
18460 		lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18461 		lentime /= bw;
18462 	} else {
18463 		/* TSNH -- KASSERT? */
18464 		lentime = 0;
18465 	}
18466 out:
18467 	if (tcp_bblogging_on(tp)) {
18468 		memset(&log, 0, sizeof(log));
18469 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18470 		log.u_bbr.flex1 = p_rate;
18471 		log.u_bbr.flex2 = p_queue;
18472 		log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18473 		log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18474 		log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18475 		log.u_bbr.flex7 = 99;
18476 		log.u_bbr.flex8 = 0;
18477 		log.u_bbr.pkts_out = err;
18478 		log.u_bbr.delRate = rack->r_ctl.crte->rate;
18479 		log.u_bbr.cur_del_rate = lentime;
18480 		log.u_bbr.timeStamp = cts;
18481 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18482 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18483 			       len, &log, false, NULL, __func__, __LINE__,tv);
18484 	}
18485 #endif
18486 	return ((uint32_t)lentime);
18487 }
18488 
18489 static int
18490 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18491 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18492 {
18493 	/*
18494 	 * Enter the fast retransmit path. We are given that a sched_pin is
18495 	 * in place (if accounting is compliled in) and the cycle count taken
18496 	 * at the entry is in the ts_val. The concept her is that the rsm
18497 	 * now holds the mbuf offsets and such so we can directly transmit
18498 	 * without a lot of overhead, the len field is already set for
18499 	 * us to prohibit us from sending too much (usually its 1MSS).
18500 	 */
18501 	struct ip *ip = NULL;
18502 	struct udphdr *udp = NULL;
18503 	struct tcphdr *th = NULL;
18504 	struct mbuf *m = NULL;
18505 	struct inpcb *inp;
18506 	uint8_t *cpto;
18507 	struct tcp_log_buffer *lgb;
18508 #ifdef TCP_ACCOUNTING
18509 	uint64_t crtsc;
18510 	int cnt_thru = 1;
18511 #endif
18512 	struct tcpopt to;
18513 	u_char opt[TCP_MAXOLEN];
18514 	uint32_t hdrlen, optlen;
18515 	int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18516 	uint16_t flags;
18517 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
18518 	uint32_t if_hw_tsomaxsegsize;
18519 	int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18520 
18521 #ifdef INET6
18522 	struct ip6_hdr *ip6 = NULL;
18523 
18524 	if (rack->r_is_v6) {
18525 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18526 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18527 	} else
18528 #endif				/* INET6 */
18529 	{
18530 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18531 		hdrlen = sizeof(struct tcpiphdr);
18532 	}
18533 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18534 		goto failed;
18535 	}
18536 	if (doing_tlp) {
18537 		/* Its a TLP add the flag, it may already be there but be sure */
18538 		rsm->r_flags |= RACK_TLP;
18539 	} else {
18540 		/* If it was a TLP it is not not on this retransmit */
18541 		rsm->r_flags &= ~RACK_TLP;
18542 	}
18543 	startseq = rsm->r_start;
18544 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18545 	inp = rack->rc_inp;
18546 	to.to_flags = 0;
18547 	flags = tcp_outflags[tp->t_state];
18548 	if (flags & (TH_SYN|TH_RST)) {
18549 		goto failed;
18550 	}
18551 	if (rsm->r_flags & RACK_HAS_FIN) {
18552 		/* We can't send a FIN here */
18553 		goto failed;
18554 	}
18555 	if (flags & TH_FIN) {
18556 		/* We never send a FIN */
18557 		flags &= ~TH_FIN;
18558 	}
18559 	if (tp->t_flags & TF_RCVD_TSTMP) {
18560 		to.to_tsval = ms_cts + tp->ts_offset;
18561 		to.to_tsecr = tp->ts_recent;
18562 		to.to_flags = TOF_TS;
18563 	}
18564 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18565 	/* TCP-MD5 (RFC2385). */
18566 	if (tp->t_flags & TF_SIGNATURE)
18567 		to.to_flags |= TOF_SIGNATURE;
18568 #endif
18569 	optlen = tcp_addoptions(&to, opt);
18570 	hdrlen += optlen;
18571 	udp = rack->r_ctl.fsb.udp;
18572 	if (udp)
18573 		hdrlen += sizeof(struct udphdr);
18574 	if (rack->r_ctl.rc_pace_max_segs)
18575 		max_val = rack->r_ctl.rc_pace_max_segs;
18576 	else if (rack->rc_user_set_max_segs)
18577 		max_val = rack->rc_user_set_max_segs * segsiz;
18578 	else
18579 		max_val = len;
18580 	if ((tp->t_flags & TF_TSO) &&
18581 	    V_tcp_do_tso &&
18582 	    (len > segsiz) &&
18583 	    (tp->t_port == 0))
18584 		tso = 1;
18585 #ifdef INET6
18586 	if (MHLEN < hdrlen + max_linkhdr)
18587 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18588 	else
18589 #endif
18590 		m = m_gethdr(M_NOWAIT, MT_DATA);
18591 	if (m == NULL)
18592 		goto failed;
18593 	m->m_data += max_linkhdr;
18594 	m->m_len = hdrlen;
18595 	th = rack->r_ctl.fsb.th;
18596 	/* Establish the len to send */
18597 	if (len > max_val)
18598 		len = max_val;
18599 	if ((tso) && (len + optlen > segsiz)) {
18600 		uint32_t if_hw_tsomax;
18601 		int32_t max_len;
18602 
18603 		/* extract TSO information */
18604 		if_hw_tsomax = tp->t_tsomax;
18605 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18606 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18607 		/*
18608 		 * Check if we should limit by maximum payload
18609 		 * length:
18610 		 */
18611 		if (if_hw_tsomax != 0) {
18612 			/* compute maximum TSO length */
18613 			max_len = (if_hw_tsomax - hdrlen -
18614 				   max_linkhdr);
18615 			if (max_len <= 0) {
18616 				goto failed;
18617 			} else if (len > max_len) {
18618 				len = max_len;
18619 			}
18620 		}
18621 		if (len <= segsiz) {
18622 			/*
18623 			 * In case there are too many small fragments don't
18624 			 * use TSO:
18625 			 */
18626 			tso = 0;
18627 		}
18628 	} else {
18629 		tso = 0;
18630 	}
18631 	if ((tso == 0) && (len > segsiz))
18632 		len = segsiz;
18633 	(void)tcp_get_usecs(tv);
18634 	if ((len == 0) ||
18635 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
18636 		goto failed;
18637 	}
18638 	th->th_seq = htonl(rsm->r_start);
18639 	th->th_ack = htonl(tp->rcv_nxt);
18640 	/*
18641 	 * The PUSH bit should only be applied
18642 	 * if the full retransmission is made. If
18643 	 * we are sending less than this is the
18644 	 * left hand edge and should not have
18645 	 * the PUSH bit.
18646 	 */
18647 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
18648 	    (len == (rsm->r_end - rsm->r_start)))
18649 		flags |= TH_PUSH;
18650 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18651 	if (th->th_win == 0) {
18652 		tp->t_sndzerowin++;
18653 		tp->t_flags |= TF_RXWIN0SENT;
18654 	} else
18655 		tp->t_flags &= ~TF_RXWIN0SENT;
18656 	if (rsm->r_flags & RACK_TLP) {
18657 		/*
18658 		 * TLP should not count in retran count, but
18659 		 * in its own bin
18660 		 */
18661 		counter_u64_add(rack_tlp_retran, 1);
18662 		counter_u64_add(rack_tlp_retran_bytes, len);
18663 	} else {
18664 		tp->t_sndrexmitpack++;
18665 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18666 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18667 	}
18668 #ifdef STATS
18669 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18670 				 len);
18671 #endif
18672 	if (rsm->m == NULL)
18673 		goto failed;
18674 	if (rsm->m &&
18675 	    ((rsm->orig_m_len != rsm->m->m_len) ||
18676 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18677 		/* Fix up the orig_m_len and possibly the mbuf offset */
18678 		rack_adjust_orig_mlen(rsm);
18679 	}
18680 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18681 	if (len <= segsiz) {
18682 		/*
18683 		 * Must have ran out of mbufs for the copy
18684 		 * shorten it to no longer need tso. Lets
18685 		 * not put on sendalot since we are low on
18686 		 * mbufs.
18687 		 */
18688 		tso = 0;
18689 	}
18690 	if ((m->m_next == NULL) || (len <= 0)){
18691 		goto failed;
18692 	}
18693 	if (udp) {
18694 		if (rack->r_is_v6)
18695 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
18696 		else
18697 			ulen = hdrlen + len - sizeof(struct ip);
18698 		udp->uh_ulen = htons(ulen);
18699 	}
18700 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18701 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18702 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18703 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
18704 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18705 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18706 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18707 #ifdef INET6
18708 		if (rack->r_is_v6) {
18709 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18710 		    ip6->ip6_flow |= htonl(ect << 20);
18711 		}
18712 		else
18713 #endif
18714 		{
18715 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
18716 		    ip->ip_tos |= ect;
18717 		}
18718 	}
18719 	if (rack->r_ctl.crte != NULL) {
18720 		/* See if we can send via the hw queue */
18721 		slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18722 		/* If there is nothing in queue (no pacing time) we can send via the hw queue */
18723 		if (slot == 0)
18724 			ip_sendflag = 0;
18725 	}
18726 	tcp_set_flags(th, flags);
18727 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18728 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18729 	if (to.to_flags & TOF_SIGNATURE) {
18730 		/*
18731 		 * Calculate MD5 signature and put it into the place
18732 		 * determined before.
18733 		 * NOTE: since TCP options buffer doesn't point into
18734 		 * mbuf's data, calculate offset and use it.
18735 		 */
18736 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18737 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18738 			/*
18739 			 * Do not send segment if the calculation of MD5
18740 			 * digest has failed.
18741 			 */
18742 			goto failed;
18743 		}
18744 	}
18745 #endif
18746 #ifdef INET6
18747 	if (rack->r_is_v6) {
18748 		if (tp->t_port) {
18749 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18750 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18751 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18752 			th->th_sum = htons(0);
18753 			UDPSTAT_INC(udps_opackets);
18754 		} else {
18755 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18756 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18757 			th->th_sum = in6_cksum_pseudo(ip6,
18758 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18759 						      0);
18760 		}
18761 	}
18762 #endif
18763 #if defined(INET6) && defined(INET)
18764 	else
18765 #endif
18766 #ifdef INET
18767 	{
18768 		if (tp->t_port) {
18769 			m->m_pkthdr.csum_flags = CSUM_UDP;
18770 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18771 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18772 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18773 			th->th_sum = htons(0);
18774 			UDPSTAT_INC(udps_opackets);
18775 		} else {
18776 			m->m_pkthdr.csum_flags = CSUM_TCP;
18777 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18778 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18779 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18780 									IPPROTO_TCP + len + optlen));
18781 		}
18782 		/* IP version must be set here for ipv4/ipv6 checking later */
18783 		KASSERT(ip->ip_v == IPVERSION,
18784 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18785 	}
18786 #endif
18787 	if (tso) {
18788 		/*
18789 		 * Here we use segsiz since we have no added options besides
18790 		 * any standard timestamp options (no DSACKs or SACKS are sent
18791 		 * via either fast-path).
18792 		 */
18793 		KASSERT(len > segsiz,
18794 			("%s: len <= tso_segsz tp:%p", __func__, tp));
18795 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18796 		m->m_pkthdr.tso_segsz = segsiz;
18797 	}
18798 #ifdef INET6
18799 	if (rack->r_is_v6) {
18800 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18801 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18802 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18803 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18804 		else
18805 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18806 	}
18807 #endif
18808 #if defined(INET) && defined(INET6)
18809 	else
18810 #endif
18811 #ifdef INET
18812 	{
18813 		ip->ip_len = htons(m->m_pkthdr.len);
18814 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
18815 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18816 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18817 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18818 				ip->ip_off |= htons(IP_DF);
18819 			}
18820 		} else {
18821 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18822 		}
18823 	}
18824 #endif
18825 	if (doing_tlp == 0) {
18826 		/* Set we retransmitted */
18827 		rack->rc_gp_saw_rec = 1;
18828 	} else {
18829 		/* Its a TLP set ca or ss */
18830 		if (tp->snd_cwnd > tp->snd_ssthresh) {
18831 			/* Set we sent in CA */
18832 			rack->rc_gp_saw_ca = 1;
18833 		} else {
18834 			/* Set we sent in SS */
18835 			rack->rc_gp_saw_ss = 1;
18836 		}
18837 	}
18838 	/* Time to copy in our header */
18839 	cpto = mtod(m, uint8_t *);
18840 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18841 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18842 	if (optlen) {
18843 		bcopy(opt, th + 1, optlen);
18844 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18845 	} else {
18846 		th->th_off = sizeof(struct tcphdr) >> 2;
18847 	}
18848 	if (tcp_bblogging_on(rack->rc_tp)) {
18849 		union tcp_log_stackspecific log;
18850 
18851 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18852 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18853 			counter_u64_add(rack_collapsed_win_rxt, 1);
18854 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18855 		}
18856 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18857 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18858 		if (rack->rack_no_prr)
18859 			log.u_bbr.flex1 = 0;
18860 		else
18861 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18862 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18863 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18864 		log.u_bbr.flex4 = max_val;
18865 		/* Save off the early/late values */
18866 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18867 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18868 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18869 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
18870 		if (doing_tlp == 0)
18871 			log.u_bbr.flex8 = 1;
18872 		else
18873 			log.u_bbr.flex8 = 2;
18874 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18875 		log.u_bbr.flex7 = 55;
18876 		log.u_bbr.pkts_out = tp->t_maxseg;
18877 		log.u_bbr.timeStamp = cts;
18878 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18879 		if (rsm && (rsm->r_rtr_cnt > 0)) {
18880 			/*
18881 			 * When we have a retransmit we want to log the
18882 			 * burst at send and flight at send from before.
18883 			 */
18884 			log.u_bbr.flex5 = rsm->r_fas;
18885 			log.u_bbr.bbr_substate = rsm->r_bas;
18886 		} else {
18887 			/*
18888 			 * This is currently unlikely until we do the
18889 			 * packet pair probes but I will add it for completeness.
18890 			 */
18891 			log.u_bbr.flex5 = log.u_bbr.inflight;
18892 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
18893 		}
18894 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
18895 		log.u_bbr.delivered = 0;
18896 		log.u_bbr.rttProp = (uintptr_t)rsm;
18897 		log.u_bbr.delRate = rsm->r_flags;
18898 		log.u_bbr.delRate <<= 31;
18899 		log.u_bbr.delRate |= rack->r_must_retran;
18900 		log.u_bbr.delRate <<= 1;
18901 		log.u_bbr.delRate |= 1;
18902 		log.u_bbr.pkt_epoch = __LINE__;
18903 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
18904 				     len, &log, false, NULL, __func__, __LINE__, tv);
18905 	} else
18906 		lgb = NULL;
18907 	if ((rack->r_ctl.crte != NULL) &&
18908 	    tcp_bblogging_on(tp)) {
18909 		rack_log_queue_level(tp, rack, len, tv, cts);
18910 	}
18911 #ifdef INET6
18912 	if (rack->r_is_v6) {
18913 		error = ip6_output(m, inp->in6p_outputopts,
18914 				   &inp->inp_route6,
18915 				   ip_sendflag, NULL, NULL, inp);
18916 	}
18917 	else
18918 #endif
18919 #ifdef INET
18920 	{
18921 		error = ip_output(m, NULL,
18922 				  &inp->inp_route,
18923 				  ip_sendflag, 0, inp);
18924 	}
18925 #endif
18926 	m = NULL;
18927 	if (lgb) {
18928 		lgb->tlb_errno = error;
18929 		lgb = NULL;
18930 	}
18931 	/* Move snd_nxt to snd_max so we don't have false retransmissions */
18932 	tp->snd_nxt = tp->snd_max;
18933 	if (error) {
18934 		goto failed;
18935 	} else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
18936 		rack->rc_hw_nobuf = 0;
18937 		rack->r_ctl.rc_agg_delayed = 0;
18938 		rack->r_early = 0;
18939 		rack->r_late = 0;
18940 		rack->r_ctl.rc_agg_early = 0;
18941 	}
18942 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
18943 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
18944 	if (doing_tlp) {
18945 		rack->rc_tlp_in_progress = 1;
18946 		rack->r_ctl.rc_tlp_cnt_out++;
18947 	}
18948 	if (error == 0) {
18949 		counter_u64_add(rack_total_bytes, len);
18950 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
18951 		if (doing_tlp) {
18952 			rack->rc_last_sent_tlp_past_cumack = 0;
18953 			rack->rc_last_sent_tlp_seq_valid = 1;
18954 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18955 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18956 		}
18957 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18958 			rack->r_ctl.rc_prr_sndcnt -= len;
18959 		else
18960 			rack->r_ctl.rc_prr_sndcnt = 0;
18961 	}
18962 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18963 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
18964 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18965 		rack->r_ctl.retran_during_recovery += len;
18966 	{
18967 		int idx;
18968 
18969 		idx = (len / segsiz) + 3;
18970 		if (idx >= TCP_MSS_ACCT_ATIMER)
18971 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18972 		else
18973 			counter_u64_add(rack_out_size[idx], 1);
18974 	}
18975 	if (tp->t_rtttime == 0) {
18976 		tp->t_rtttime = ticks;
18977 		tp->t_rtseq = startseq;
18978 		KMOD_TCPSTAT_INC(tcps_segstimed);
18979 	}
18980 	counter_u64_add(rack_fto_rsm_send, 1);
18981 	if (error && (error == ENOBUFS)) {
18982 		if (rack->r_ctl.crte != NULL) {
18983 			tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
18984 			if (tcp_bblogging_on(rack->rc_tp))
18985 				rack_log_queue_level(tp, rack, len, tv, cts);
18986 		} else
18987 			tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
18988 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18989 		if (rack->rc_enobuf < 0x7f)
18990 			rack->rc_enobuf++;
18991 		if (slot < (10 * HPTS_USEC_IN_MSEC))
18992 			slot = 10 * HPTS_USEC_IN_MSEC;
18993 		if (rack->r_ctl.crte != NULL) {
18994 			counter_u64_add(rack_saw_enobuf_hw, 1);
18995 			tcp_rl_log_enobuf(rack->r_ctl.crte);
18996 		}
18997 		counter_u64_add(rack_saw_enobuf, 1);
18998 	} else {
18999 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz, __LINE__);
19000 	}
19001 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
19002 #ifdef TCP_ACCOUNTING
19003 	crtsc = get_cyclecount();
19004 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19005 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19006 	}
19007 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19008 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19009 	}
19010 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19011 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
19012 	}
19013 	sched_unpin();
19014 #endif
19015 	return (0);
19016 failed:
19017 	if (m)
19018 		m_free(m);
19019 	return (-1);
19020 }
19021 
19022 static void
19023 rack_sndbuf_autoscale(struct tcp_rack *rack)
19024 {
19025 	/*
19026 	 * Automatic sizing of send socket buffer.  Often the send buffer
19027 	 * size is not optimally adjusted to the actual network conditions
19028 	 * at hand (delay bandwidth product).  Setting the buffer size too
19029 	 * small limits throughput on links with high bandwidth and high
19030 	 * delay (eg. trans-continental/oceanic links).  Setting the
19031 	 * buffer size too big consumes too much real kernel memory,
19032 	 * especially with many connections on busy servers.
19033 	 *
19034 	 * The criteria to step up the send buffer one notch are:
19035 	 *  1. receive window of remote host is larger than send buffer
19036 	 *     (with a fudge factor of 5/4th);
19037 	 *  2. send buffer is filled to 7/8th with data (so we actually
19038 	 *     have data to make use of it);
19039 	 *  3. send buffer fill has not hit maximal automatic size;
19040 	 *  4. our send window (slow start and cogestion controlled) is
19041 	 *     larger than sent but unacknowledged data in send buffer.
19042 	 *
19043 	 * Note that the rack version moves things much faster since
19044 	 * we want to avoid hitting cache lines in the rack_fast_output()
19045 	 * path so this is called much less often and thus moves
19046 	 * the SB forward by a percentage.
19047 	 */
19048 	struct socket *so;
19049 	struct tcpcb *tp;
19050 	uint32_t sendwin, scaleup;
19051 
19052 	tp = rack->rc_tp;
19053 	so = rack->rc_inp->inp_socket;
19054 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
19055 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
19056 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
19057 		    sbused(&so->so_snd) >=
19058 		    (so->so_snd.sb_hiwat / 8 * 7) &&
19059 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
19060 		    sendwin >= (sbused(&so->so_snd) -
19061 		    (tp->snd_max - tp->snd_una))) {
19062 			if (rack_autosndbuf_inc)
19063 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
19064 			else
19065 				scaleup = V_tcp_autosndbuf_inc;
19066 			if (scaleup < V_tcp_autosndbuf_inc)
19067 				scaleup = V_tcp_autosndbuf_inc;
19068 			scaleup += so->so_snd.sb_hiwat;
19069 			if (scaleup > V_tcp_autosndbuf_max)
19070 				scaleup = V_tcp_autosndbuf_max;
19071 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
19072 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
19073 		}
19074 	}
19075 }
19076 
19077 static int
19078 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
19079 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
19080 {
19081 	/*
19082 	 * Enter to do fast output. We are given that the sched_pin is
19083 	 * in place (if accounting is compiled in) and the cycle count taken
19084 	 * at entry is in place in ts_val. The idea here is that
19085 	 * we know how many more bytes needs to be sent (presumably either
19086 	 * during pacing or to fill the cwnd and that was greater than
19087 	 * the max-burst). We have how much to send and all the info we
19088 	 * need to just send.
19089 	 */
19090 #ifdef INET
19091 	struct ip *ip = NULL;
19092 #endif
19093 	struct udphdr *udp = NULL;
19094 	struct tcphdr *th = NULL;
19095 	struct mbuf *m, *s_mb;
19096 	struct inpcb *inp;
19097 	uint8_t *cpto;
19098 	struct tcp_log_buffer *lgb;
19099 #ifdef TCP_ACCOUNTING
19100 	uint64_t crtsc;
19101 #endif
19102 	struct tcpopt to;
19103 	u_char opt[TCP_MAXOLEN];
19104 	uint32_t hdrlen, optlen;
19105 #ifdef TCP_ACCOUNTING
19106 	int cnt_thru = 1;
19107 #endif
19108 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19109 	uint16_t flags;
19110 	uint32_t s_soff;
19111 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
19112 	uint32_t if_hw_tsomaxsegsize;
19113 	uint32_t add_flag = RACK_SENT_FP;
19114 #ifdef INET6
19115 	struct ip6_hdr *ip6 = NULL;
19116 
19117 	if (rack->r_is_v6) {
19118 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19119 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19120 	} else
19121 #endif				/* INET6 */
19122 	{
19123 #ifdef INET
19124 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19125 		hdrlen = sizeof(struct tcpiphdr);
19126 #endif
19127 	}
19128 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19129 		m = NULL;
19130 		goto failed;
19131 	}
19132 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19133 	startseq = tp->snd_max;
19134 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19135 	inp = rack->rc_inp;
19136 	len = rack->r_ctl.fsb.left_to_send;
19137 	to.to_flags = 0;
19138 	flags = rack->r_ctl.fsb.tcp_flags;
19139 	if (tp->t_flags & TF_RCVD_TSTMP) {
19140 		to.to_tsval = ms_cts + tp->ts_offset;
19141 		to.to_tsecr = tp->ts_recent;
19142 		to.to_flags = TOF_TS;
19143 	}
19144 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19145 	/* TCP-MD5 (RFC2385). */
19146 	if (tp->t_flags & TF_SIGNATURE)
19147 		to.to_flags |= TOF_SIGNATURE;
19148 #endif
19149 	optlen = tcp_addoptions(&to, opt);
19150 	hdrlen += optlen;
19151 	udp = rack->r_ctl.fsb.udp;
19152 	if (udp)
19153 		hdrlen += sizeof(struct udphdr);
19154 	if (rack->r_ctl.rc_pace_max_segs)
19155 		max_val = rack->r_ctl.rc_pace_max_segs;
19156 	else if (rack->rc_user_set_max_segs)
19157 		max_val = rack->rc_user_set_max_segs * segsiz;
19158 	else
19159 		max_val = len;
19160 	if ((tp->t_flags & TF_TSO) &&
19161 	    V_tcp_do_tso &&
19162 	    (len > segsiz) &&
19163 	    (tp->t_port == 0))
19164 		tso = 1;
19165 again:
19166 #ifdef INET6
19167 	if (MHLEN < hdrlen + max_linkhdr)
19168 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19169 	else
19170 #endif
19171 		m = m_gethdr(M_NOWAIT, MT_DATA);
19172 	if (m == NULL)
19173 		goto failed;
19174 	m->m_data += max_linkhdr;
19175 	m->m_len = hdrlen;
19176 	th = rack->r_ctl.fsb.th;
19177 	/* Establish the len to send */
19178 	if (len > max_val)
19179 		len = max_val;
19180 	if ((tso) && (len + optlen > segsiz)) {
19181 		uint32_t if_hw_tsomax;
19182 		int32_t max_len;
19183 
19184 		/* extract TSO information */
19185 		if_hw_tsomax = tp->t_tsomax;
19186 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19187 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19188 		/*
19189 		 * Check if we should limit by maximum payload
19190 		 * length:
19191 		 */
19192 		if (if_hw_tsomax != 0) {
19193 			/* compute maximum TSO length */
19194 			max_len = (if_hw_tsomax - hdrlen -
19195 				   max_linkhdr);
19196 			if (max_len <= 0) {
19197 				goto failed;
19198 			} else if (len > max_len) {
19199 				len = max_len;
19200 			}
19201 		}
19202 		if (len <= segsiz) {
19203 			/*
19204 			 * In case there are too many small fragments don't
19205 			 * use TSO:
19206 			 */
19207 			tso = 0;
19208 		}
19209 	} else {
19210 		tso = 0;
19211 	}
19212 	if ((tso == 0) && (len > segsiz))
19213 		len = segsiz;
19214 	(void)tcp_get_usecs(tv);
19215 	if ((len == 0) ||
19216 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
19217 		goto failed;
19218 	}
19219 	sb_offset = tp->snd_max - tp->snd_una;
19220 	th->th_seq = htonl(tp->snd_max);
19221 	th->th_ack = htonl(tp->rcv_nxt);
19222 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19223 	if (th->th_win == 0) {
19224 		tp->t_sndzerowin++;
19225 		tp->t_flags |= TF_RXWIN0SENT;
19226 	} else
19227 		tp->t_flags &= ~TF_RXWIN0SENT;
19228 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
19229 	KMOD_TCPSTAT_INC(tcps_sndpack);
19230 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19231 #ifdef STATS
19232 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19233 				 len);
19234 #endif
19235 	if (rack->r_ctl.fsb.m == NULL)
19236 		goto failed;
19237 
19238 	/* s_mb and s_soff are saved for rack_log_output */
19239 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19240 				    &s_mb, &s_soff);
19241 	if (len <= segsiz) {
19242 		/*
19243 		 * Must have ran out of mbufs for the copy
19244 		 * shorten it to no longer need tso. Lets
19245 		 * not put on sendalot since we are low on
19246 		 * mbufs.
19247 		 */
19248 		tso = 0;
19249 	}
19250 	if (rack->r_ctl.fsb.rfo_apply_push &&
19251 	    (len == rack->r_ctl.fsb.left_to_send)) {
19252 		tcp_set_flags(th, flags | TH_PUSH);
19253 		add_flag |= RACK_HAD_PUSH;
19254 	}
19255 	if ((m->m_next == NULL) || (len <= 0)){
19256 		goto failed;
19257 	}
19258 	if (udp) {
19259 		if (rack->r_is_v6)
19260 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
19261 		else
19262 			ulen = hdrlen + len - sizeof(struct ip);
19263 		udp->uh_ulen = htons(ulen);
19264 	}
19265 	m->m_pkthdr.rcvif = (struct ifnet *)0;
19266 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
19267 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19268 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
19269 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19270 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
19271 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19272 #ifdef INET6
19273 		if (rack->r_is_v6) {
19274 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19275 			ip6->ip6_flow |= htonl(ect << 20);
19276 		}
19277 		else
19278 #endif
19279 		{
19280 #ifdef INET
19281 			ip->ip_tos &= ~IPTOS_ECN_MASK;
19282 			ip->ip_tos |= ect;
19283 #endif
19284 		}
19285 	}
19286 	tcp_set_flags(th, flags);
19287 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
19288 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19289 	if (to.to_flags & TOF_SIGNATURE) {
19290 		/*
19291 		 * Calculate MD5 signature and put it into the place
19292 		 * determined before.
19293 		 * NOTE: since TCP options buffer doesn't point into
19294 		 * mbuf's data, calculate offset and use it.
19295 		 */
19296 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
19297 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
19298 			/*
19299 			 * Do not send segment if the calculation of MD5
19300 			 * digest has failed.
19301 			 */
19302 			goto failed;
19303 		}
19304 	}
19305 #endif
19306 #ifdef INET6
19307 	if (rack->r_is_v6) {
19308 		if (tp->t_port) {
19309 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19310 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19311 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19312 			th->th_sum = htons(0);
19313 			UDPSTAT_INC(udps_opackets);
19314 		} else {
19315 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19316 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19317 			th->th_sum = in6_cksum_pseudo(ip6,
19318 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19319 						      0);
19320 		}
19321 	}
19322 #endif
19323 #if defined(INET6) && defined(INET)
19324 	else
19325 #endif
19326 #ifdef INET
19327 	{
19328 		if (tp->t_port) {
19329 			m->m_pkthdr.csum_flags = CSUM_UDP;
19330 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19331 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19332 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19333 			th->th_sum = htons(0);
19334 			UDPSTAT_INC(udps_opackets);
19335 		} else {
19336 			m->m_pkthdr.csum_flags = CSUM_TCP;
19337 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19338 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
19339 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19340 									IPPROTO_TCP + len + optlen));
19341 		}
19342 		/* IP version must be set here for ipv4/ipv6 checking later */
19343 		KASSERT(ip->ip_v == IPVERSION,
19344 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
19345 	}
19346 #endif
19347 	if (tso) {
19348 		/*
19349 		 * Here we use segsiz since we have no added options besides
19350 		 * any standard timestamp options (no DSACKs or SACKS are sent
19351 		 * via either fast-path).
19352 		 */
19353 		KASSERT(len > segsiz,
19354 			("%s: len <= tso_segsz tp:%p", __func__, tp));
19355 		m->m_pkthdr.csum_flags |= CSUM_TSO;
19356 		m->m_pkthdr.tso_segsz = segsiz;
19357 	}
19358 #ifdef INET6
19359 	if (rack->r_is_v6) {
19360 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19361 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19362 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19363 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19364 		else
19365 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19366 	}
19367 #endif
19368 #if defined(INET) && defined(INET6)
19369 	else
19370 #endif
19371 #ifdef INET
19372 	{
19373 		ip->ip_len = htons(m->m_pkthdr.len);
19374 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19375 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19376 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19377 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19378 				ip->ip_off |= htons(IP_DF);
19379 			}
19380 		} else {
19381 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19382 		}
19383 	}
19384 #endif
19385 	if (tp->snd_cwnd > tp->snd_ssthresh) {
19386 		/* Set we sent in CA */
19387 		rack->rc_gp_saw_ca = 1;
19388 	} else {
19389 		/* Set we sent in SS */
19390 		rack->rc_gp_saw_ss = 1;
19391 	}
19392 	/* Time to copy in our header */
19393 	cpto = mtod(m, uint8_t *);
19394 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19395 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19396 	if (optlen) {
19397 		bcopy(opt, th + 1, optlen);
19398 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19399 	} else {
19400 		th->th_off = sizeof(struct tcphdr) >> 2;
19401 	}
19402 	if ((rack->r_ctl.crte != NULL) &&
19403 	    tcp_bblogging_on(tp)) {
19404 		rack_log_queue_level(tp, rack, len, tv, cts);
19405 	}
19406 	if (tcp_bblogging_on(rack->rc_tp)) {
19407 		union tcp_log_stackspecific log;
19408 
19409 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19410 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19411 		if (rack->rack_no_prr)
19412 			log.u_bbr.flex1 = 0;
19413 		else
19414 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19415 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19416 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19417 		log.u_bbr.flex4 = max_val;
19418 		/* Save off the early/late values */
19419 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19420 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19421 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19422 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19423 		log.u_bbr.flex8 = 0;
19424 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19425 		log.u_bbr.flex7 = 44;
19426 		log.u_bbr.pkts_out = tp->t_maxseg;
19427 		log.u_bbr.timeStamp = cts;
19428 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19429 		log.u_bbr.flex5 = log.u_bbr.inflight;
19430 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19431 		log.u_bbr.delivered = 0;
19432 		log.u_bbr.rttProp = 0;
19433 		log.u_bbr.delRate = rack->r_must_retran;
19434 		log.u_bbr.delRate <<= 1;
19435 		log.u_bbr.pkt_epoch = __LINE__;
19436 		/* For fast output no retrans so just inflight and how many mss we send */
19437 		log.u_bbr.flex5 = log.u_bbr.inflight;
19438 		log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19439 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19440 				     len, &log, false, NULL, __func__, __LINE__, tv);
19441 	} else
19442 		lgb = NULL;
19443 #ifdef INET6
19444 	if (rack->r_is_v6) {
19445 		error = ip6_output(m, inp->in6p_outputopts,
19446 				   &inp->inp_route6,
19447 				   0, NULL, NULL, inp);
19448 	}
19449 #endif
19450 #if defined(INET) && defined(INET6)
19451 	else
19452 #endif
19453 #ifdef INET
19454 	{
19455 		error = ip_output(m, NULL,
19456 				  &inp->inp_route,
19457 				  0, 0, inp);
19458 	}
19459 #endif
19460 	if (lgb) {
19461 		lgb->tlb_errno = error;
19462 		lgb = NULL;
19463 	}
19464 	if (error) {
19465 		*send_err = error;
19466 		m = NULL;
19467 		goto failed;
19468 	} else if (rack->rc_hw_nobuf) {
19469 		rack->rc_hw_nobuf = 0;
19470 		rack->r_ctl.rc_agg_delayed = 0;
19471 		rack->r_early = 0;
19472 		rack->r_late = 0;
19473 		rack->r_ctl.rc_agg_early = 0;
19474 	}
19475 	if ((error == 0) && (rack->lt_bw_up == 0)) {
19476 		/* Unlikely */
19477 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
19478 		rack->r_ctl.lt_seq = tp->snd_una;
19479 		rack->lt_bw_up = 1;
19480 	} else if ((error == 0) &&
19481 		   (((tp->snd_max + len) - rack->r_ctl.lt_seq) > 0x7fffffff)) {
19482 		/*
19483 		 * Need to record what we have since we are
19484 		 * approaching seq wrap.
19485 		 */
19486 		struct timeval tv;
19487 		uint64_t tmark;
19488 
19489 		rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
19490 		rack->r_ctl.lt_seq = tp->snd_una;
19491 		tmark = tcp_get_u64_usecs(&tv);
19492 		if (tmark > rack->r_ctl.lt_timemark) {
19493 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
19494 			rack->r_ctl.lt_timemark = tmark;
19495 		}
19496 	}
19497 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19498 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19499 	m = NULL;
19500 	if (tp->snd_una == tp->snd_max) {
19501 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
19502 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19503 		tp->t_acktime = ticks;
19504 	}
19505 	counter_u64_add(rack_total_bytes, len);
19506 	tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19507 
19508 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19509 	tot_len += len;
19510 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
19511 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19512 	tp->snd_max += len;
19513 	tp->snd_nxt = tp->snd_max;
19514 	if (rack->rc_new_rnd_needed) {
19515 		rack_new_round_starts(tp, rack, tp->snd_max);
19516 	}
19517 	{
19518 		int idx;
19519 
19520 		idx = (len / segsiz) + 3;
19521 		if (idx >= TCP_MSS_ACCT_ATIMER)
19522 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19523 		else
19524 			counter_u64_add(rack_out_size[idx], 1);
19525 	}
19526 	if (len <= rack->r_ctl.fsb.left_to_send)
19527 		rack->r_ctl.fsb.left_to_send -= len;
19528 	else
19529 		rack->r_ctl.fsb.left_to_send = 0;
19530 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
19531 		rack->r_fast_output = 0;
19532 		rack->r_ctl.fsb.left_to_send = 0;
19533 		/* At the end of fast_output scale up the sb */
19534 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
19535 		rack_sndbuf_autoscale(rack);
19536 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
19537 	}
19538 	if (tp->t_rtttime == 0) {
19539 		tp->t_rtttime = ticks;
19540 		tp->t_rtseq = startseq;
19541 		KMOD_TCPSTAT_INC(tcps_segstimed);
19542 	}
19543 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19544 	    (max_val > len) &&
19545 	    (tso == 0)) {
19546 		max_val -= len;
19547 		len = segsiz;
19548 		th = rack->r_ctl.fsb.th;
19549 #ifdef TCP_ACCOUNTING
19550 		cnt_thru++;
19551 #endif
19552 		goto again;
19553 	}
19554 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19555 	counter_u64_add(rack_fto_send, 1);
19556 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz, __LINE__);
19557 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
19558 #ifdef TCP_ACCOUNTING
19559 	crtsc = get_cyclecount();
19560 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19561 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19562 	}
19563 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19564 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19565 	}
19566 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19567 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
19568 	}
19569 	sched_unpin();
19570 #endif
19571 	return (0);
19572 failed:
19573 	if (m)
19574 		m_free(m);
19575 	rack->r_fast_output = 0;
19576 	return (-1);
19577 }
19578 
19579 static inline void
19580 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19581 		       struct sockbuf *sb,
19582 		       int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19583 		       bool hw_tls,
19584 		       uint16_t flags)
19585 {
19586 	rack->r_fast_output = 1;
19587 	rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19588 	rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19589 	rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19590 	rack->r_ctl.fsb.tcp_flags = flags;
19591 	rack->r_ctl.fsb.left_to_send = orig_len - len;
19592 	if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19593 		/* Less than a full sized pace, lets not  */
19594 		rack->r_fast_output = 0;
19595 		return;
19596 	} else {
19597 		/* Round down to the nearest pace_max_seg */
19598 		rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19599 	}
19600 	if (hw_tls)
19601 		rack->r_ctl.fsb.hw_tls = 1;
19602 	else
19603 		rack->r_ctl.fsb.hw_tls = 0;
19604 	KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19605 		("rack:%p left_to_send:%u sbavail:%u out:%u",
19606 		 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19607 		 (tp->snd_max - tp->snd_una)));
19608 	if (rack->r_ctl.fsb.left_to_send < segsiz)
19609 		rack->r_fast_output = 0;
19610 	else {
19611 		if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19612 			rack->r_ctl.fsb.rfo_apply_push = 1;
19613 		else
19614 			rack->r_ctl.fsb.rfo_apply_push = 0;
19615 	}
19616 }
19617 
19618 static uint32_t
19619 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19620 {
19621 	uint64_t min_time;
19622 	uint32_t maxlen;
19623 
19624 	min_time = (uint64_t)get_hpts_min_sleep_time();
19625 	maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19626 	maxlen = roundup(maxlen, segsiz);
19627 	return (maxlen);
19628 }
19629 
19630 static struct rack_sendmap *
19631 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19632 {
19633 	struct rack_sendmap *rsm = NULL;
19634 	int thresh;
19635 
19636 restart:
19637 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19638 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19639 		/* Nothing, strange turn off validity  */
19640 		rack->r_collapse_point_valid = 0;
19641 		return (NULL);
19642 	}
19643 	/* Can we send it yet? */
19644 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19645 		/*
19646 		 * Receiver window has not grown enough for
19647 		 * the segment to be put on the wire.
19648 		 */
19649 		return (NULL);
19650 	}
19651 	if (rsm->r_flags & RACK_ACKED) {
19652 		/*
19653 		 * It has been sacked, lets move to the
19654 		 * next one if possible.
19655 		 */
19656 		rack->r_ctl.last_collapse_point = rsm->r_end;
19657 		/* Are we done? */
19658 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19659 			    rack->r_ctl.high_collapse_point)) {
19660 			rack->r_collapse_point_valid = 0;
19661 			return (NULL);
19662 		}
19663 		goto restart;
19664 	}
19665 	/* Now has it been long enough ? */
19666 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts, __LINE__, 1);
19667 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19668 		rack_log_collapse(rack, rsm->r_start,
19669 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19670 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
19671 		return (rsm);
19672 	}
19673 	/* Not enough time */
19674 	rack_log_collapse(rack, rsm->r_start,
19675 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19676 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
19677 	return (NULL);
19678 }
19679 
19680 static inline void
19681 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19682 {
19683 	if ((rack->full_size_rxt == 0) &&
19684 	    (rack->shape_rxt_to_pacing_min == 0) &&
19685 	    (*len >= segsiz)) {
19686 		*len = segsiz;
19687 	} else if (rack->shape_rxt_to_pacing_min &&
19688 		 rack->gp_ready) {
19689 		/* We use pacing min as shaping len req */
19690 		uint32_t maxlen;
19691 
19692 		maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19693 		if (*len > maxlen)
19694 			*len = maxlen;
19695 	} else {
19696 		/*
19697 		 * The else is full_size_rxt is on so send it all
19698 		 * note we do need to check this for exceeding
19699 		 * our max segment size due to the fact that
19700 		 * we do sometimes merge chunks together i.e.
19701 		 * we cannot just assume that we will never have
19702 		 * a chunk greater than pace_max_seg
19703 		 */
19704 		if (*len > pace_max_seg)
19705 			*len = pace_max_seg;
19706 	}
19707 }
19708 
19709 static int
19710 rack_output(struct tcpcb *tp)
19711 {
19712 	struct socket *so;
19713 	uint32_t recwin;
19714 	uint32_t sb_offset, s_moff = 0;
19715 	int32_t len, error = 0;
19716 	uint16_t flags;
19717 	struct mbuf *m, *s_mb = NULL;
19718 	struct mbuf *mb;
19719 	uint32_t if_hw_tsomaxsegcount = 0;
19720 	uint32_t if_hw_tsomaxsegsize;
19721 	int32_t segsiz, minseg;
19722 	long tot_len_this_send = 0;
19723 #ifdef INET
19724 	struct ip *ip = NULL;
19725 #endif
19726 	struct udphdr *udp = NULL;
19727 	struct tcp_rack *rack;
19728 	struct tcphdr *th;
19729 	uint8_t pass = 0;
19730 	uint8_t mark = 0;
19731 	uint8_t check_done = 0;
19732 	uint8_t wanted_cookie = 0;
19733 	u_char opt[TCP_MAXOLEN];
19734 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
19735 	uint32_t rack_seq;
19736 
19737 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19738 	unsigned ipsec_optlen = 0;
19739 
19740 #endif
19741 	int32_t idle, sendalot;
19742 	uint32_t tot_idle;
19743 	int32_t sub_from_prr = 0;
19744 	volatile int32_t sack_rxmit;
19745 	struct rack_sendmap *rsm = NULL;
19746 	int32_t tso, mtu;
19747 	struct tcpopt to;
19748 	int32_t slot = 0;
19749 	int32_t sup_rack = 0;
19750 	uint32_t cts, ms_cts, delayed, early;
19751 	uint32_t add_flag = RACK_SENT_SP;
19752 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19753 	uint8_t doing_tlp = 0;
19754 	uint32_t cwnd_to_use, pace_max_seg;
19755 	int32_t do_a_prefetch = 0;
19756 	int32_t prefetch_rsm = 0;
19757 	int32_t orig_len = 0;
19758 	struct timeval tv;
19759 	int32_t prefetch_so_done = 0;
19760 	struct tcp_log_buffer *lgb;
19761 	struct inpcb *inp = tptoinpcb(tp);
19762 	struct sockbuf *sb;
19763 	uint64_t ts_val = 0;
19764 #ifdef TCP_ACCOUNTING
19765 	uint64_t crtsc;
19766 #endif
19767 #ifdef INET6
19768 	struct ip6_hdr *ip6 = NULL;
19769 	int32_t isipv6;
19770 #endif
19771 	bool hpts_calling, hw_tls = false;
19772 
19773 	NET_EPOCH_ASSERT();
19774 	INP_WLOCK_ASSERT(inp);
19775 
19776 	/* setup and take the cache hits here */
19777 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19778 #ifdef TCP_ACCOUNTING
19779 	sched_pin();
19780 	ts_val = get_cyclecount();
19781 #endif
19782 	hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
19783 	tp->t_flags2 &= ~TF2_HPTS_CALLS;
19784 #ifdef TCP_OFFLOAD
19785 	if (tp->t_flags & TF_TOE) {
19786 #ifdef TCP_ACCOUNTING
19787 		sched_unpin();
19788 #endif
19789 		return (tcp_offload_output(tp));
19790 	}
19791 #endif
19792 	if (rack->rack_deferred_inited == 0) {
19793 		/*
19794 		 * If we are the connecting socket we will
19795 		 * hit rack_init() when no sequence numbers
19796 		 * are setup. This makes it so we must defer
19797 		 * some initialization. Call that now.
19798 		 */
19799 		rack_deferred_init(tp, rack);
19800 	}
19801 	/*
19802 	 * For TFO connections in SYN_RECEIVED, only allow the initial
19803 	 * SYN|ACK and those sent by the retransmit timer.
19804 	 */
19805 	if ((tp->t_flags & TF_FASTOPEN) &&
19806 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
19807 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
19808 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
19809 #ifdef TCP_ACCOUNTING
19810 		sched_unpin();
19811 #endif
19812 		return (0);
19813 	}
19814 #ifdef INET6
19815 	if (rack->r_state) {
19816 		/* Use the cache line loaded if possible */
19817 		isipv6 = rack->r_is_v6;
19818 	} else {
19819 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
19820 	}
19821 #endif
19822 	early = 0;
19823 	cts = tcp_get_usecs(&tv);
19824 	ms_cts = tcp_tv_to_mssectick(&tv);
19825 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
19826 	    tcp_in_hpts(rack->rc_tp)) {
19827 		/*
19828 		 * We are on the hpts for some timer but not hptsi output.
19829 		 * Remove from the hpts unconditionally.
19830 		 */
19831 		rack_timer_cancel(tp, rack, cts, __LINE__);
19832 	}
19833 	/* Are we pacing and late? */
19834 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19835 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
19836 		/* We are delayed */
19837 		delayed = cts - rack->r_ctl.rc_last_output_to;
19838 	} else {
19839 		delayed = 0;
19840 	}
19841 	/* Do the timers, which may override the pacer */
19842 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
19843 		int retval;
19844 
19845 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
19846 					     &doing_tlp);
19847 		if (retval != 0) {
19848 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
19849 #ifdef TCP_ACCOUNTING
19850 			sched_unpin();
19851 #endif
19852 			/*
19853 			 * If timers want tcp_drop(), then pass error out,
19854 			 * otherwise suppress it.
19855 			 */
19856 			return (retval < 0 ? retval : 0);
19857 		}
19858 	}
19859 	if (rack->rc_in_persist) {
19860 		if (tcp_in_hpts(rack->rc_tp) == 0) {
19861 			/* Timer is not running */
19862 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19863 		}
19864 #ifdef TCP_ACCOUNTING
19865 		sched_unpin();
19866 #endif
19867 		return (0);
19868 	}
19869 	if ((rack->rc_ack_required == 1) &&
19870 	    (rack->r_timer_override == 0)){
19871 		/* A timeout occurred and no ack has arrived */
19872 		if (tcp_in_hpts(rack->rc_tp) == 0) {
19873 			/* Timer is not running */
19874 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19875 		}
19876 #ifdef TCP_ACCOUNTING
19877 		sched_unpin();
19878 #endif
19879 		return (0);
19880 	}
19881 	if ((rack->r_timer_override) ||
19882 	    (rack->rc_ack_can_sendout_data) ||
19883 	    (delayed) ||
19884 	    (tp->t_state < TCPS_ESTABLISHED)) {
19885 		rack->rc_ack_can_sendout_data = 0;
19886 		if (tcp_in_hpts(rack->rc_tp))
19887 			tcp_hpts_remove(rack->rc_tp);
19888 	} else if (tcp_in_hpts(rack->rc_tp)) {
19889 		/*
19890 		 * On the hpts you can't pass even if ACKNOW is on, we will
19891 		 * when the hpts fires.
19892 		 */
19893 #ifdef TCP_ACCOUNTING
19894 		crtsc = get_cyclecount();
19895 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19896 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
19897 		}
19898 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19899 			tp->tcp_cnt_counters[SND_BLOCKED]++;
19900 		}
19901 		sched_unpin();
19902 #endif
19903 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
19904 		return (0);
19905 	}
19906 	/* Finish out both pacing early and late accounting */
19907 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19908 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
19909 		early = rack->r_ctl.rc_last_output_to - cts;
19910 	} else
19911 		early = 0;
19912 	if (delayed && (rack->rc_always_pace == 1)) {
19913 		rack->r_ctl.rc_agg_delayed += delayed;
19914 		rack->r_late = 1;
19915 	} else if (early && (rack->rc_always_pace == 1)) {
19916 		rack->r_ctl.rc_agg_early += early;
19917 		rack->r_early = 1;
19918 	} else if (rack->rc_always_pace == 0) {
19919 		/* Non-paced we are not late */
19920 		rack->r_ctl.rc_agg_delayed = rack->r_ctl.rc_agg_early = 0;
19921 		rack->r_early = rack->r_late = 0;
19922 	}
19923 	/* Now that early/late accounting is done turn off the flag */
19924 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
19925 	rack->r_wanted_output = 0;
19926 	rack->r_timer_override = 0;
19927 	if ((tp->t_state != rack->r_state) &&
19928 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
19929 		rack_set_state(tp, rack);
19930 	}
19931 	if ((rack->r_fast_output) &&
19932 	    (doing_tlp == 0) &&
19933 	    (tp->rcv_numsacks == 0)) {
19934 		int ret;
19935 
19936 		error = 0;
19937 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19938 		if (ret >= 0)
19939 			return(ret);
19940 		else if (error) {
19941 			inp = rack->rc_inp;
19942 			so = inp->inp_socket;
19943 			sb = &so->so_snd;
19944 			goto nomore;
19945 		}
19946 	}
19947 	inp = rack->rc_inp;
19948 	/*
19949 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
19950 	 * only allow the initial SYN or SYN|ACK and those sent
19951 	 * by the retransmit timer.
19952 	 */
19953 	if ((tp->t_flags & TF_FASTOPEN) &&
19954 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
19955 	     (tp->t_state == TCPS_SYN_SENT)) &&
19956 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
19957 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
19958 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19959 		so = inp->inp_socket;
19960 		sb = &so->so_snd;
19961 		goto just_return_nolock;
19962 	}
19963 	/*
19964 	 * Determine length of data that should be transmitted, and flags
19965 	 * that will be used. If there is some data or critical controls
19966 	 * (SYN, RST) to send, then transmit; otherwise, investigate
19967 	 * further.
19968 	 */
19969 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
19970 	if (tp->t_idle_reduce) {
19971 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
19972 			rack_cc_after_idle(rack, tp);
19973 	}
19974 	tp->t_flags &= ~TF_LASTIDLE;
19975 	if (idle) {
19976 		if (tp->t_flags & TF_MORETOCOME) {
19977 			tp->t_flags |= TF_LASTIDLE;
19978 			idle = 0;
19979 		}
19980 	}
19981 	if ((tp->snd_una == tp->snd_max) &&
19982 	    rack->r_ctl.rc_went_idle_time &&
19983 	    (cts > rack->r_ctl.rc_went_idle_time)) {
19984 		tot_idle = (cts - rack->r_ctl.rc_went_idle_time);
19985 		if (tot_idle > rack_min_probertt_hold) {
19986 			/* Count as a probe rtt */
19987 			if (rack->in_probe_rtt == 0) {
19988 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
19989 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
19990 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
19991 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
19992 			} else {
19993 				rack_exit_probertt(rack, cts);
19994 			}
19995 		}
19996 	}
19997 	if (rack_use_fsb &&
19998 	    (rack->r_ctl.fsb.tcp_ip_hdr) &&
19999 	    (rack->r_fsb_inited == 0) &&
20000 	    (rack->r_state != TCPS_CLOSED))
20001 		rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
20002 	if (rack->rc_sendvars_notset == 1) {
20003 		rack->rc_sendvars_notset = 0;
20004 		/*
20005 		 * Make sure any TCP timers (keep-alive) is not running.
20006 		 */
20007 		tcp_timer_stop(tp);
20008 	}
20009 	if ((rack->rack_no_prr == 1) &&
20010 	    (rack->rc_always_pace == 0)) {
20011 		/*
20012 		 * Sanity check before sending, if we have
20013 		 * no-pacing enabled and prr is turned off that
20014 		 * is a logistics error. Correct this by turnning
20015 		 * prr back on. A user *must* set some form of
20016 		 * pacing in order to turn PRR off. We do this
20017 		 * in the output path so that we can avoid socket
20018 		 * option ordering issues that would occur if we
20019 		 * tried to do it while setting rack_no_prr on.
20020 		 */
20021 		rack->rack_no_prr = 0;
20022 	}
20023 	if ((rack->pcm_enabled == 1) &&
20024 	    (rack->pcm_needed == 0) &&
20025 	    (tot_idle > 0)) {
20026 		/*
20027 		 * We have been idle some micro seconds. We need
20028 		 * to factor this in to see if a PCM is needed.
20029 		 */
20030 		uint32_t rtts_idle, rnds;
20031 
20032 		if (tp->t_srtt)
20033 			rtts_idle = tot_idle / tp->t_srtt;
20034 		else
20035 			rtts_idle = 0;
20036 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
20037 		rack->r_ctl.pcm_idle_rounds += rtts_idle;
20038 		if ((rnds + rack->r_ctl.pcm_idle_rounds)  >= rack_pcm_every_n_rounds) {
20039 			rack->pcm_needed = 1;
20040 			rack_log_pcm(rack, 8, rack->r_ctl.last_pcm_round, rtts_idle, rack->r_ctl.current_round );
20041 		}
20042 	}
20043 again:
20044 	sendalot = 0;
20045 	cts = tcp_get_usecs(&tv);
20046 	ms_cts = tcp_tv_to_mssectick(&tv);
20047 	tso = 0;
20048 	mtu = 0;
20049 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
20050 	minseg = segsiz;
20051 	if (rack->r_ctl.rc_pace_max_segs == 0)
20052 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
20053 	else
20054 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
20055 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
20056 	    (rack->r_ctl.pcm_max_seg == 0)) {
20057 		/*
20058 		 * We set in our first send so we know that the ctf_fixed_maxseg
20059 		 * has been fully set. If we do it in rack_init() we most likely
20060 		 * see 512 bytes so we end up at 5120, not desirable.
20061 		 */
20062 		rack->r_ctl.pcm_max_seg = rc_init_window(rack);
20063 		if (rack->r_ctl.pcm_max_seg < (ctf_fixed_maxseg(tp) * 10)) {
20064 			/*
20065 			 * Assure our initial PCM probe is at least 10 MSS.
20066 			 */
20067 			rack->r_ctl.pcm_max_seg = ctf_fixed_maxseg(tp) * 10;
20068 		}
20069 	}
20070 	if ((rack->r_ctl.pcm_max_seg != 0)  && (rack->pcm_needed == 1)) {
20071 		uint32_t rw_avail, cwa;
20072 
20073 		if (tp->snd_wnd > ctf_outstanding(tp))
20074 			rw_avail = tp->snd_wnd - ctf_outstanding(tp);
20075 		else
20076 			rw_avail = 0;
20077 		if (tp->snd_cwnd > ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked))
20078 			cwa = tp->snd_cwnd -ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20079 		else
20080 			cwa = 0;
20081 		if ((cwa >= rack->r_ctl.pcm_max_seg) &&
20082 		    (rw_avail > rack->r_ctl.pcm_max_seg)) {
20083 			/* Raise up the max seg for this trip through */
20084 			pace_max_seg = rack->r_ctl.pcm_max_seg;
20085 			/* Disable any fast output */
20086 			rack->r_fast_output = 0;
20087 		}
20088 		if (rack_verbose_logging) {
20089 			rack_log_pcm(rack, 4,
20090 				     cwa, rack->r_ctl.pcm_max_seg, rw_avail);
20091 		}
20092 	}
20093 	sb_offset = tp->snd_max - tp->snd_una;
20094 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20095 	flags = tcp_outflags[tp->t_state];
20096 	while (rack->rc_free_cnt < rack_free_cache) {
20097 		rsm = rack_alloc(rack);
20098 		if (rsm == NULL) {
20099 			if (hpts_calling)
20100 				/* Retry in a ms */
20101 				slot = (1 * HPTS_USEC_IN_MSEC);
20102 			so = inp->inp_socket;
20103 			sb = &so->so_snd;
20104 			goto just_return_nolock;
20105 		}
20106 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
20107 		rack->rc_free_cnt++;
20108 		rsm = NULL;
20109 	}
20110 	sack_rxmit = 0;
20111 	len = 0;
20112 	rsm = NULL;
20113 	if (flags & TH_RST) {
20114 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
20115 		so = inp->inp_socket;
20116 		sb = &so->so_snd;
20117 		goto send;
20118 	}
20119 	if (rack->r_ctl.rc_resend) {
20120 		/* Retransmit timer */
20121 		rsm = rack->r_ctl.rc_resend;
20122 		rack->r_ctl.rc_resend = NULL;
20123 		len = rsm->r_end - rsm->r_start;
20124 		sack_rxmit = 1;
20125 		sendalot = 0;
20126 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20127 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20128 			 __func__, __LINE__,
20129 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20130 		sb_offset = rsm->r_start - tp->snd_una;
20131 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20132 	} else if (rack->r_collapse_point_valid &&
20133 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
20134 		/*
20135 		 * If an RSM is returned then enough time has passed
20136 		 * for us to retransmit it. Move up the collapse point,
20137 		 * since this rsm has its chance to retransmit now.
20138 		 */
20139 		tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
20140 		rack->r_ctl.last_collapse_point = rsm->r_end;
20141 		/* Are we done? */
20142 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
20143 			    rack->r_ctl.high_collapse_point))
20144 			rack->r_collapse_point_valid = 0;
20145 		sack_rxmit = 1;
20146 		/* We are not doing a TLP */
20147 		doing_tlp = 0;
20148 		len = rsm->r_end - rsm->r_start;
20149 		sb_offset = rsm->r_start - tp->snd_una;
20150 		sendalot = 0;
20151 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20152 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
20153 		/* We have a retransmit that takes precedence */
20154 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
20155 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
20156 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
20157 			/* Enter recovery if not induced by a time-out */
20158 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
20159 		}
20160 #ifdef INVARIANTS
20161 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
20162 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
20163 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
20164 		}
20165 #endif
20166 		len = rsm->r_end - rsm->r_start;
20167 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20168 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20169 			 __func__, __LINE__,
20170 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20171 		sb_offset = rsm->r_start - tp->snd_una;
20172 		sendalot = 0;
20173 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20174 		if (len > 0) {
20175 			sack_rxmit = 1;
20176 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
20177 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
20178 					 min(len, segsiz));
20179 		}
20180 	} else if (rack->r_ctl.rc_tlpsend) {
20181 		/* Tail loss probe */
20182 		long cwin;
20183 		long tlen;
20184 
20185 		/*
20186 		 * Check if we can do a TLP with a RACK'd packet
20187 		 * this can happen if we are not doing the rack
20188 		 * cheat and we skipped to a TLP and it
20189 		 * went off.
20190 		 */
20191 		rsm = rack->r_ctl.rc_tlpsend;
20192 		/* We are doing a TLP make sure the flag is preent */
20193 		rsm->r_flags |= RACK_TLP;
20194 		rack->r_ctl.rc_tlpsend = NULL;
20195 		sack_rxmit = 1;
20196 		tlen = rsm->r_end - rsm->r_start;
20197 		if (tlen > segsiz)
20198 			tlen = segsiz;
20199 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20200 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20201 			 __func__, __LINE__,
20202 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20203 		sb_offset = rsm->r_start - tp->snd_una;
20204 		cwin = min(tp->snd_wnd, tlen);
20205 		len = cwin;
20206 	}
20207 	if (rack->r_must_retran &&
20208 	    (doing_tlp == 0) &&
20209 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20210 	    (rsm == NULL)) {
20211 		/*
20212 		 * There are two different ways that we
20213 		 * can get into this block:
20214 		 * a) This is a non-sack connection, we had a time-out
20215 		 *    and thus r_must_retran was set and everything
20216 		 *    left outstanding as been marked for retransmit.
20217 		 * b) The MTU of the path shrank, so that everything
20218 		 *    was marked to be retransmitted with the smaller
20219 		 *    mtu and r_must_retran was set.
20220 		 *
20221 		 * This means that we expect the sendmap (outstanding)
20222 		 * to all be marked must. We can use the tmap to
20223 		 * look at them.
20224 		 *
20225 		 */
20226 		int sendwin, flight;
20227 
20228 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20229 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20230 		if (flight >= sendwin) {
20231 			/*
20232 			 * We can't send yet.
20233 			 */
20234 			so = inp->inp_socket;
20235 			sb = &so->so_snd;
20236 			goto just_return_nolock;
20237 		}
20238 		/*
20239 		 * This is the case a/b mentioned above. All
20240 		 * outstanding/not-acked should be marked.
20241 		 * We can use the tmap to find them.
20242 		 */
20243 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20244 		if (rsm == NULL) {
20245 			/* TSNH */
20246 			rack->r_must_retran = 0;
20247 			rack->r_ctl.rc_out_at_rto = 0;
20248 			so = inp->inp_socket;
20249 			sb = &so->so_snd;
20250 			goto just_return_nolock;
20251 		}
20252 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20253 			/*
20254 			 * The first one does not have the flag, did we collapse
20255 			 * further up in our list?
20256 			 */
20257 			rack->r_must_retran = 0;
20258 			rack->r_ctl.rc_out_at_rto = 0;
20259 			rsm = NULL;
20260 			sack_rxmit = 0;
20261 		} else {
20262 			sack_rxmit = 1;
20263 			len = rsm->r_end - rsm->r_start;
20264 			sb_offset = rsm->r_start - tp->snd_una;
20265 			sendalot = 0;
20266 			if ((rack->full_size_rxt == 0) &&
20267 			    (rack->shape_rxt_to_pacing_min == 0) &&
20268 			    (len >= segsiz))
20269 				len = segsiz;
20270 			else if (rack->shape_rxt_to_pacing_min &&
20271 				 rack->gp_ready) {
20272 				/* We use pacing min as shaping len req */
20273 				uint32_t maxlen;
20274 
20275 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20276 				if (len > maxlen)
20277 					len = maxlen;
20278 			}
20279 			/*
20280 			 * Delay removing the flag RACK_MUST_RXT so
20281 			 * that the fastpath for retransmit will
20282 			 * work with this rsm.
20283 			 */
20284 		}
20285 	}
20286 	/*
20287 	 * Enforce a connection sendmap count limit if set
20288 	 * as long as we are not retransmiting.
20289 	 */
20290 	if ((rsm == NULL) &&
20291 	    (V_tcp_map_entries_limit > 0) &&
20292 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20293 		counter_u64_add(rack_to_alloc_limited, 1);
20294 		if (!rack->alloc_limit_reported) {
20295 			rack->alloc_limit_reported = 1;
20296 			counter_u64_add(rack_alloc_limited_conns, 1);
20297 		}
20298 		so = inp->inp_socket;
20299 		sb = &so->so_snd;
20300 		goto just_return_nolock;
20301 	}
20302 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20303 		/* we are retransmitting the fin */
20304 		len--;
20305 		if (len) {
20306 			/*
20307 			 * When retransmitting data do *not* include the
20308 			 * FIN. This could happen from a TLP probe.
20309 			 */
20310 			flags &= ~TH_FIN;
20311 		}
20312 	}
20313 	if (rsm && rack->r_fsb_inited &&
20314 	    rack_use_rsm_rfo &&
20315 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20316 		int ret;
20317 
20318 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20319 		if (ret == 0)
20320 			return (0);
20321 	}
20322 	so = inp->inp_socket;
20323 	sb = &so->so_snd;
20324 	if (do_a_prefetch == 0) {
20325 		kern_prefetch(sb, &do_a_prefetch);
20326 		do_a_prefetch = 1;
20327 	}
20328 #ifdef NETFLIX_SHARED_CWND
20329 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20330 	    rack->rack_enable_scwnd) {
20331 		/* We are doing cwnd sharing */
20332 		if (rack->gp_ready &&
20333 		    (rack->rack_attempted_scwnd == 0) &&
20334 		    (rack->r_ctl.rc_scw == NULL) &&
20335 		    tp->t_lib) {
20336 			/* The pcbid is in, lets make an attempt */
20337 			counter_u64_add(rack_try_scwnd, 1);
20338 			rack->rack_attempted_scwnd = 1;
20339 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20340 								   &rack->r_ctl.rc_scw_index,
20341 								   segsiz);
20342 		}
20343 		if (rack->r_ctl.rc_scw &&
20344 		    (rack->rack_scwnd_is_idle == 1) &&
20345 		    sbavail(&so->so_snd)) {
20346 			/* we are no longer out of data */
20347 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20348 			rack->rack_scwnd_is_idle = 0;
20349 		}
20350 		if (rack->r_ctl.rc_scw) {
20351 			/* First lets update and get the cwnd */
20352 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20353 										       rack->r_ctl.rc_scw_index,
20354 										       tp->snd_cwnd, tp->snd_wnd, segsiz);
20355 		}
20356 	}
20357 #endif
20358 	/*
20359 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
20360 	 * state flags.
20361 	 */
20362 	if (tp->t_flags & TF_NEEDFIN)
20363 		flags |= TH_FIN;
20364 	if (tp->t_flags & TF_NEEDSYN)
20365 		flags |= TH_SYN;
20366 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20367 		void *end_rsm;
20368 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20369 		if (end_rsm)
20370 			kern_prefetch(end_rsm, &prefetch_rsm);
20371 		prefetch_rsm = 1;
20372 	}
20373 	SOCKBUF_LOCK(sb);
20374 	if ((sack_rxmit == 0) &&
20375 	    (TCPS_HAVEESTABLISHED(tp->t_state) ||
20376 	    (tp->t_flags & TF_FASTOPEN))) {
20377 		/*
20378 		 * We are not retransmitting (sack_rxmit is 0) so we
20379 		 * are sending new data. This is always based on snd_max.
20380 		 * Now in theory snd_max may be equal to snd_una, if so
20381 		 * then nothing is outstanding and the offset would be 0.
20382 		 */
20383 		uint32_t avail;
20384 
20385 		avail = sbavail(sb);
20386 		if (SEQ_GT(tp->snd_max, tp->snd_una) && avail)
20387 			sb_offset = tp->snd_max - tp->snd_una;
20388 		else
20389 			sb_offset = 0;
20390 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20391 			if (rack->r_ctl.rc_tlp_new_data) {
20392 				/* TLP is forcing out new data */
20393 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20394 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20395 				}
20396 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20397 					if (tp->snd_wnd > sb_offset)
20398 						len = tp->snd_wnd - sb_offset;
20399 					else
20400 						len = 0;
20401 				} else {
20402 					len = rack->r_ctl.rc_tlp_new_data;
20403 				}
20404 				rack->r_ctl.rc_tlp_new_data = 0;
20405 			}  else {
20406 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20407 			}
20408 			if ((rack->r_ctl.crte == NULL) &&
20409 			    IN_FASTRECOVERY(tp->t_flags) &&
20410 			    (rack->full_size_rxt == 0) &&
20411 			    (rack->shape_rxt_to_pacing_min == 0) &&
20412 			    (len > segsiz)) {
20413 				/*
20414 				 * For prr=off, we need to send only 1 MSS
20415 				 * at a time. We do this because another sack could
20416 				 * be arriving that causes us to send retransmits and
20417 				 * we don't want to be on a long pace due to a larger send
20418 				 * that keeps us from sending out the retransmit.
20419 				 */
20420 				len = segsiz;
20421 			} else if (rack->shape_rxt_to_pacing_min &&
20422 				   rack->gp_ready) {
20423 				/* We use pacing min as shaping len req */
20424 				uint32_t maxlen;
20425 
20426 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20427 				if (len > maxlen)
20428 					len = maxlen;
20429 			}/* The else is full_size_rxt is on so send it all */
20430 		} else {
20431 			uint32_t outstanding;
20432 			/*
20433 			 * We are inside of a Fast recovery episode, this
20434 			 * is caused by a SACK or 3 dup acks. At this point
20435 			 * we have sent all the retransmissions and we rely
20436 			 * on PRR to dictate what we will send in the form of
20437 			 * new data.
20438 			 */
20439 
20440 			outstanding = tp->snd_max - tp->snd_una;
20441 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20442 				if (tp->snd_wnd > outstanding) {
20443 					len = tp->snd_wnd - outstanding;
20444 					/* Check to see if we have the data */
20445 					if ((sb_offset + len) > avail) {
20446 						/* It does not all fit */
20447 						if (avail > sb_offset)
20448 							len = avail - sb_offset;
20449 						else
20450 							len = 0;
20451 					}
20452 				} else {
20453 					len = 0;
20454 				}
20455 			} else if (avail > sb_offset) {
20456 				len = avail - sb_offset;
20457 			} else {
20458 				len = 0;
20459 			}
20460 			if (len > 0) {
20461 				if (len > rack->r_ctl.rc_prr_sndcnt) {
20462 					len = rack->r_ctl.rc_prr_sndcnt;
20463 				}
20464 				if (len > 0) {
20465 					sub_from_prr = 1;
20466 				}
20467 			}
20468 			if (len > segsiz) {
20469 				/*
20470 				 * We should never send more than a MSS when
20471 				 * retransmitting or sending new data in prr
20472 				 * mode unless the override flag is on. Most
20473 				 * likely the PRR algorithm is not going to
20474 				 * let us send a lot as well :-)
20475 				 */
20476 				if (rack->r_ctl.rc_prr_sendalot == 0) {
20477 					len = segsiz;
20478 				}
20479 			} else if (len < segsiz) {
20480 				/*
20481 				 * Do we send any? The idea here is if the
20482 				 * send empty's the socket buffer we want to
20483 				 * do it. However if not then lets just wait
20484 				 * for our prr_sndcnt to get bigger.
20485 				 */
20486 				long leftinsb;
20487 
20488 				leftinsb = sbavail(sb) - sb_offset;
20489 				if (leftinsb > len) {
20490 					/* This send does not empty the sb */
20491 					len = 0;
20492 				}
20493 			}
20494 		}
20495 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20496 		/*
20497 		 * If you have not established
20498 		 * and are not doing FAST OPEN
20499 		 * no data please.
20500 		 */
20501 		if ((sack_rxmit == 0) &&
20502 		    !(tp->t_flags & TF_FASTOPEN)) {
20503 			len = 0;
20504 			sb_offset = 0;
20505 		}
20506 	}
20507 	if (prefetch_so_done == 0) {
20508 		kern_prefetch(so, &prefetch_so_done);
20509 		prefetch_so_done = 1;
20510 	}
20511 	orig_len = len;
20512 	/*
20513 	 * Lop off SYN bit if it has already been sent.  However, if this is
20514 	 * SYN-SENT state and if segment contains data and if we don't know
20515 	 * that foreign host supports TAO, suppress sending segment.
20516 	 */
20517 	if ((flags & TH_SYN) &&
20518 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
20519 	    ((sack_rxmit == 0) &&
20520 	     (tp->t_rxtshift == 0))) {
20521 		/*
20522 		 * When sending additional segments following a TFO SYN|ACK,
20523 		 * do not include the SYN bit.
20524 		 */
20525 		if ((tp->t_flags & TF_FASTOPEN) &&
20526 		    (tp->t_state == TCPS_SYN_RECEIVED))
20527 			flags &= ~TH_SYN;
20528 	}
20529 	/*
20530 	 * Be careful not to send data and/or FIN on SYN segments. This
20531 	 * measure is needed to prevent interoperability problems with not
20532 	 * fully conformant TCP implementations.
20533 	 */
20534 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20535 		len = 0;
20536 		flags &= ~TH_FIN;
20537 	}
20538 	/*
20539 	 * On TFO sockets, ensure no data is sent in the following cases:
20540 	 *
20541 	 *  - When retransmitting SYN|ACK on a passively-created socket
20542 	 *
20543 	 *  - When retransmitting SYN on an actively created socket
20544 	 *
20545 	 *  - When sending a zero-length cookie (cookie request) on an
20546 	 *    actively created socket
20547 	 *
20548 	 *  - When the socket is in the CLOSED state (RST is being sent)
20549 	 */
20550 	if ((tp->t_flags & TF_FASTOPEN) &&
20551 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20552 	     ((tp->t_state == TCPS_SYN_SENT) &&
20553 	      (tp->t_tfo_client_cookie_len == 0)) ||
20554 	     (flags & TH_RST))) {
20555 		sack_rxmit = 0;
20556 		len = 0;
20557 	}
20558 	/* Without fast-open there should never be data sent on a SYN */
20559 	if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
20560 		len = 0;
20561 	}
20562 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20563 		/* We only send 1 MSS if we have a DSACK block */
20564 		add_flag |= RACK_SENT_W_DSACK;
20565 		len = segsiz;
20566 	}
20567 	if (len <= 0) {
20568 		/*
20569 		 * We have nothing to send, or the window shrank, or
20570 		 * is closed, do we need to go into persists?
20571 		 */
20572 		len = 0;
20573 		if ((tp->snd_wnd == 0) &&
20574 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20575 		    (tp->snd_una == tp->snd_max) &&
20576 		    (sb_offset < (int)sbavail(sb))) {
20577 			rack_enter_persist(tp, rack, cts, tp->snd_una);
20578 		}
20579 	} else if ((rsm == NULL) &&
20580 		   (doing_tlp == 0) &&
20581 		   (len < pace_max_seg)) {
20582 		/*
20583 		 * We are not sending a maximum sized segment for
20584 		 * some reason. Should we not send anything (think
20585 		 * sws or persists)?
20586 		 */
20587 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20588 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20589 		    (len < minseg) &&
20590 		    (len < (int)(sbavail(sb) - sb_offset))) {
20591 			/*
20592 			 * Here the rwnd is less than
20593 			 * the minimum pacing size, this is not a retransmit,
20594 			 * we are established and
20595 			 * the send is not the last in the socket buffer
20596 			 * we send nothing, and we may enter persists
20597 			 * if nothing is outstanding.
20598 			 */
20599 			len = 0;
20600 			if (tp->snd_max == tp->snd_una) {
20601 				/*
20602 				 * Nothing out we can
20603 				 * go into persists.
20604 				 */
20605 				rack_enter_persist(tp, rack, cts, tp->snd_una);
20606 			}
20607 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20608 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20609 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20610 			   (len < minseg)) {
20611 			/*
20612 			 * Here we are not retransmitting, and
20613 			 * the cwnd is not so small that we could
20614 			 * not send at least a min size (rxt timer
20615 			 * not having gone off), We have 2 segments or
20616 			 * more already in flight, its not the tail end
20617 			 * of the socket buffer  and the cwnd is blocking
20618 			 * us from sending out a minimum pacing segment size.
20619 			 * Lets not send anything.
20620 			 */
20621 			len = 0;
20622 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20623 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20624 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20625 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20626 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
20627 			/*
20628 			 * Here we have a send window but we have
20629 			 * filled it up and we can't send another pacing segment.
20630 			 * We also have in flight more than 2 segments
20631 			 * and we are not completing the sb i.e. we allow
20632 			 * the last bytes of the sb to go out even if
20633 			 * its not a full pacing segment.
20634 			 */
20635 			len = 0;
20636 		} else if ((rack->r_ctl.crte != NULL) &&
20637 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20638 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20639 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20640 			   (len < (int)(sbavail(sb) - sb_offset))) {
20641 			/*
20642 			 * Here we are doing hardware pacing, this is not a TLP,
20643 			 * we are not sending a pace max segment size, there is rwnd
20644 			 * room to send at least N pace_max_seg, the cwnd is greater
20645 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
20646 			 * more segments in flight and its not the tail of the socket buffer.
20647 			 *
20648 			 * We don't want to send instead we need to get more ack's in to
20649 			 * allow us to send a full pacing segment. Normally, if we are pacing
20650 			 * about the right speed, we should have finished our pacing
20651 			 * send as most of the acks have come back if we are at the
20652 			 * right rate. This is a bit fuzzy since return path delay
20653 			 * can delay the acks, which is why we want to make sure we
20654 			 * have cwnd space to have a bit more than a max pace segments in flight.
20655 			 *
20656 			 * If we have not gotten our acks back we are pacing at too high a
20657 			 * rate delaying will not hurt and will bring our GP estimate down by
20658 			 * injecting the delay. If we don't do this we will send
20659 			 * 2 MSS out in response to the acks being clocked in which
20660 			 * defeats the point of hw-pacing (i.e. to help us get
20661 			 * larger TSO's out).
20662 			 */
20663 			len = 0;
20664 		}
20665 
20666 	}
20667 	/* len will be >= 0 after this point. */
20668 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20669 	rack_sndbuf_autoscale(rack);
20670 	/*
20671 	 * Decide if we can use TCP Segmentation Offloading (if supported by
20672 	 * hardware).
20673 	 *
20674 	 * TSO may only be used if we are in a pure bulk sending state.  The
20675 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20676 	 * options prevent using TSO.  With TSO the TCP header is the same
20677 	 * (except for the sequence number) for all generated packets.  This
20678 	 * makes it impossible to transmit any options which vary per
20679 	 * generated segment or packet.
20680 	 *
20681 	 * IPv4 handling has a clear separation of ip options and ip header
20682 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20683 	 * the right thing below to provide length of just ip options and thus
20684 	 * checking for ipoptlen is enough to decide if ip options are present.
20685 	 */
20686 	ipoptlen = 0;
20687 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20688 	/*
20689 	 * Pre-calculate here as we save another lookup into the darknesses
20690 	 * of IPsec that way and can actually decide if TSO is ok.
20691 	 */
20692 #ifdef INET6
20693 	if (isipv6 && IPSEC_ENABLED(ipv6))
20694 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20695 #ifdef INET
20696 	else
20697 #endif
20698 #endif				/* INET6 */
20699 #ifdef INET
20700 		if (IPSEC_ENABLED(ipv4))
20701 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20702 #endif				/* INET */
20703 #endif
20704 
20705 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20706 	ipoptlen += ipsec_optlen;
20707 #endif
20708 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20709 	    (tp->t_port == 0) &&
20710 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
20711 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
20712 	    ipoptlen == 0)
20713 		tso = 1;
20714 	{
20715 		uint32_t outstanding __unused;
20716 
20717 		outstanding = tp->snd_max - tp->snd_una;
20718 		if (tp->t_flags & TF_SENTFIN) {
20719 			/*
20720 			 * If we sent a fin, snd_max is 1 higher than
20721 			 * snd_una
20722 			 */
20723 			outstanding--;
20724 		}
20725 		if (sack_rxmit) {
20726 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20727 				flags &= ~TH_FIN;
20728 		}
20729 	}
20730 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20731 		      (long)TCP_MAXWIN << tp->rcv_scale);
20732 
20733 	/*
20734 	 * Sender silly window avoidance.   We transmit under the following
20735 	 * conditions when len is non-zero:
20736 	 *
20737 	 * - We have a full segment (or more with TSO) - This is the last
20738 	 * buffer in a write()/send() and we are either idle or running
20739 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
20740 	 * then 1/2 the maximum send window's worth of data (receiver may be
20741 	 * limited the window size) - we need to retransmit
20742 	 */
20743 	if (len) {
20744 		if (len >= segsiz) {
20745 			goto send;
20746 		}
20747 		/*
20748 		 * NOTE! on localhost connections an 'ack' from the remote
20749 		 * end may occur synchronously with the output and cause us
20750 		 * to flush a buffer queued with moretocome.  XXX
20751 		 *
20752 		 */
20753 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
20754 		    (idle || (tp->t_flags & TF_NODELAY)) &&
20755 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20756 		    (tp->t_flags & TF_NOPUSH) == 0) {
20757 			pass = 2;
20758 			goto send;
20759 		}
20760 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
20761 			pass = 22;
20762 			goto send;
20763 		}
20764 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20765 			pass = 4;
20766 			goto send;
20767 		}
20768 		if (sack_rxmit) {
20769 			pass = 6;
20770 			goto send;
20771 		}
20772 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20773 		    (ctf_outstanding(tp) < (segsiz * 2))) {
20774 			/*
20775 			 * We have less than two MSS outstanding (delayed ack)
20776 			 * and our rwnd will not let us send a full sized
20777 			 * MSS. Lets go ahead and let this small segment
20778 			 * out because we want to try to have at least two
20779 			 * packets inflight to not be caught by delayed ack.
20780 			 */
20781 			pass = 12;
20782 			goto send;
20783 		}
20784 	}
20785 	/*
20786 	 * Sending of standalone window updates.
20787 	 *
20788 	 * Window updates are important when we close our window due to a
20789 	 * full socket buffer and are opening it again after the application
20790 	 * reads data from it.  Once the window has opened again and the
20791 	 * remote end starts to send again the ACK clock takes over and
20792 	 * provides the most current window information.
20793 	 *
20794 	 * We must avoid the silly window syndrome whereas every read from
20795 	 * the receive buffer, no matter how small, causes a window update
20796 	 * to be sent.  We also should avoid sending a flurry of window
20797 	 * updates when the socket buffer had queued a lot of data and the
20798 	 * application is doing small reads.
20799 	 *
20800 	 * Prevent a flurry of pointless window updates by only sending an
20801 	 * update when we can increase the advertized window by more than
20802 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
20803 	 * full or is very small be more aggressive and send an update
20804 	 * whenever we can increase by two mss sized segments. In all other
20805 	 * situations the ACK's to new incoming data will carry further
20806 	 * window increases.
20807 	 *
20808 	 * Don't send an independent window update if a delayed ACK is
20809 	 * pending (it will get piggy-backed on it) or the remote side
20810 	 * already has done a half-close and won't send more data.  Skip
20811 	 * this if the connection is in T/TCP half-open state.
20812 	 */
20813 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20814 	    !(tp->t_flags & TF_DELACK) &&
20815 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
20816 		/*
20817 		 * "adv" is the amount we could increase the window, taking
20818 		 * into account that we are limited by TCP_MAXWIN <<
20819 		 * tp->rcv_scale.
20820 		 */
20821 		int32_t adv;
20822 		int oldwin;
20823 
20824 		adv = recwin;
20825 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20826 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
20827 			if (adv > oldwin)
20828 				adv -= oldwin;
20829 			else {
20830 				/* We can't increase the window */
20831 				adv = 0;
20832 			}
20833 		} else
20834 			oldwin = 0;
20835 
20836 		/*
20837 		 * If the new window size ends up being the same as or less
20838 		 * than the old size when it is scaled, then don't force
20839 		 * a window update.
20840 		 */
20841 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20842 			goto dontupdate;
20843 
20844 		if (adv >= (int32_t)(2 * segsiz) &&
20845 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20846 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20847 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20848 			pass = 7;
20849 			goto send;
20850 		}
20851 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20852 			pass = 23;
20853 			goto send;
20854 		}
20855 	}
20856 dontupdate:
20857 
20858 	/*
20859 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
20860 	 * is also a catch-all for the retransmit timer timeout case.
20861 	 */
20862 	if (tp->t_flags & TF_ACKNOW) {
20863 		pass = 8;
20864 		goto send;
20865 	}
20866 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
20867 		pass = 9;
20868 		goto send;
20869 	}
20870 	/*
20871 	 * If our state indicates that FIN should be sent and we have not
20872 	 * yet done so, then we need to send.
20873 	 */
20874 	if ((flags & TH_FIN) &&
20875 	    (tp->snd_max == tp->snd_una)) {
20876 		pass = 11;
20877 		goto send;
20878 	}
20879 	/*
20880 	 * No reason to send a segment, just return.
20881 	 */
20882 just_return:
20883 	SOCKBUF_UNLOCK(sb);
20884 just_return_nolock:
20885 	{
20886 		int app_limited = CTF_JR_SENT_DATA;
20887 
20888 		if ((tp->t_flags & TF_FASTOPEN) == 0 &&
20889 		    (flags & TH_FIN) &&
20890 		    (len == 0) &&
20891 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
20892 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
20893 			/*
20894 			 * Ok less than or right at a MSS is
20895 			 * outstanding. The original FreeBSD stack would
20896 			 * have sent a FIN, which can speed things up for
20897 			 * a transactional application doing a MSG_WAITALL.
20898 			 * To speed things up since we do *not* send a FIN
20899 			 * if data is outstanding, we send a "challenge ack".
20900 			 * The idea behind that is instead of having to have
20901 			 * the peer wait for the delayed-ack timer to run off
20902 			 * we send an ack that makes the peer send us an ack.
20903 			 */
20904 			rack_send_ack_challange(rack);
20905 		}
20906 		if (tot_len_this_send > 0) {
20907 			rack->r_ctl.fsb.recwin = recwin;
20908 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz, __LINE__);
20909 			if ((error == 0) &&
20910 			    rack_use_rfo &&
20911 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
20912 			    (ipoptlen == 0) &&
20913 			    (tp->rcv_numsacks == 0) &&
20914 			    rack->r_fsb_inited &&
20915 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
20916 			    ((IN_RECOVERY(tp->t_flags)) == 0) &&
20917 			    (rack->r_must_retran == 0) &&
20918 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
20919 			    (len > 0) && (orig_len > 0) &&
20920 			    (orig_len > len) &&
20921 			    ((orig_len - len) >= segsiz) &&
20922 			    ((optlen == 0) ||
20923 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
20924 				/* We can send at least one more MSS using our fsb */
20925 				rack_setup_fast_output(tp, rack, sb, len, orig_len,
20926 						       segsiz, pace_max_seg, hw_tls, flags);
20927 			} else
20928 				rack->r_fast_output = 0;
20929 			rack_log_fsb(rack, tp, so, flags,
20930 				     ipoptlen, orig_len, len, 0,
20931 				     1, optlen, __LINE__, 1);
20932 			/* Assure when we leave that snd_nxt will point to top */
20933 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
20934 				tp->snd_nxt = tp->snd_max;
20935 		} else {
20936 			int end_window = 0;
20937 			uint32_t seq = tp->gput_ack;
20938 
20939 			rsm = tqhash_max(rack->r_ctl.tqh);
20940 			if (rsm) {
20941 				/*
20942 				 * Mark the last sent that we just-returned (hinting
20943 				 * that delayed ack may play a role in any rtt measurement).
20944 				 */
20945 				rsm->r_just_ret = 1;
20946 			}
20947 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
20948 			rack->r_ctl.rc_agg_delayed = 0;
20949 			rack->r_early = 0;
20950 			rack->r_late = 0;
20951 			rack->r_ctl.rc_agg_early = 0;
20952 			if ((ctf_outstanding(tp) +
20953 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
20954 				 minseg)) >= tp->snd_wnd) {
20955 				/* We are limited by the rwnd */
20956 				app_limited = CTF_JR_RWND_LIMITED;
20957 				if (IN_FASTRECOVERY(tp->t_flags))
20958 					rack->r_ctl.rc_prr_sndcnt = 0;
20959 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
20960 				/* We are limited by whats available -- app limited */
20961 				app_limited = CTF_JR_APP_LIMITED;
20962 				if (IN_FASTRECOVERY(tp->t_flags))
20963 					rack->r_ctl.rc_prr_sndcnt = 0;
20964 			} else if ((idle == 0) &&
20965 				   ((tp->t_flags & TF_NODELAY) == 0) &&
20966 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20967 				   (len < segsiz)) {
20968 				/*
20969 				 * No delay is not on and the
20970 				 * user is sending less than 1MSS. This
20971 				 * brings out SWS avoidance so we
20972 				 * don't send. Another app-limited case.
20973 				 */
20974 				app_limited = CTF_JR_APP_LIMITED;
20975 			} else if (tp->t_flags & TF_NOPUSH) {
20976 				/*
20977 				 * The user has requested no push of
20978 				 * the last segment and we are
20979 				 * at the last segment. Another app
20980 				 * limited case.
20981 				 */
20982 				app_limited = CTF_JR_APP_LIMITED;
20983 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
20984 				/* Its the cwnd */
20985 				app_limited = CTF_JR_CWND_LIMITED;
20986 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
20987 				   (rack->rack_no_prr == 0) &&
20988 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
20989 				app_limited = CTF_JR_PRR;
20990 			} else {
20991 				/* Now why here are we not sending? */
20992 #ifdef NOW
20993 #ifdef INVARIANTS
20994 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
20995 #endif
20996 #endif
20997 				app_limited = CTF_JR_ASSESSING;
20998 			}
20999 			/*
21000 			 * App limited in some fashion, for our pacing GP
21001 			 * measurements we don't want any gap (even cwnd).
21002 			 * Close  down the measurement window.
21003 			 */
21004 			if (rack_cwnd_block_ends_measure &&
21005 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
21006 			     (app_limited == CTF_JR_PRR))) {
21007 				/*
21008 				 * The reason we are not sending is
21009 				 * the cwnd (or prr). We have been configured
21010 				 * to end the measurement window in
21011 				 * this case.
21012 				 */
21013 				end_window = 1;
21014 			} else if (rack_rwnd_block_ends_measure &&
21015 				   (app_limited == CTF_JR_RWND_LIMITED)) {
21016 				/*
21017 				 * We are rwnd limited and have been
21018 				 * configured to end the measurement
21019 				 * window in this case.
21020 				 */
21021 				end_window = 1;
21022 			} else if (app_limited == CTF_JR_APP_LIMITED) {
21023 				/*
21024 				 * A true application limited period, we have
21025 				 * ran out of data.
21026 				 */
21027 				end_window = 1;
21028 			} else if (app_limited == CTF_JR_ASSESSING) {
21029 				/*
21030 				 * In the assessing case we hit the end of
21031 				 * the if/else and had no known reason
21032 				 * This will panic us under invariants..
21033 				 *
21034 				 * If we get this out in logs we need to
21035 				 * investagate which reason we missed.
21036 				 */
21037 				end_window = 1;
21038 			}
21039 			if (end_window) {
21040 				uint8_t log = 0;
21041 
21042 				/* Adjust the Gput measurement */
21043 				if ((tp->t_flags & TF_GPUTINPROG) &&
21044 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
21045 					tp->gput_ack = tp->snd_max;
21046 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
21047 						/*
21048 						 * There is not enough to measure.
21049 						 */
21050 						tp->t_flags &= ~TF_GPUTINPROG;
21051 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
21052 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
21053 									   tp->gput_seq,
21054 									   0, 0, 18, __LINE__, NULL, 0);
21055 					} else
21056 						log = 1;
21057 				}
21058 				/* Mark the last packet has app limited */
21059 				rsm = tqhash_max(rack->r_ctl.tqh);
21060 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
21061 					if (rack->r_ctl.rc_app_limited_cnt == 0)
21062 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
21063 					else {
21064 						/*
21065 						 * Go out to the end app limited and mark
21066 						 * this new one as next and move the end_appl up
21067 						 * to this guy.
21068 						 */
21069 						if (rack->r_ctl.rc_end_appl)
21070 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
21071 						rack->r_ctl.rc_end_appl = rsm;
21072 					}
21073 					rsm->r_flags |= RACK_APP_LIMITED;
21074 					rack->r_ctl.rc_app_limited_cnt++;
21075 				}
21076 				if (log)
21077 					rack_log_pacing_delay_calc(rack,
21078 								   rack->r_ctl.rc_app_limited_cnt, seq,
21079 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
21080 			}
21081 		}
21082 		/* Check if we need to go into persists or not */
21083 		if ((tp->snd_max == tp->snd_una) &&
21084 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
21085 		    sbavail(sb) &&
21086 		    (sbavail(sb) > tp->snd_wnd) &&
21087 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
21088 			/* Yes lets make sure to move to persist before timer-start */
21089 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
21090 		}
21091 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
21092 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
21093 	}
21094 #ifdef NETFLIX_SHARED_CWND
21095 	if ((sbavail(sb) == 0) &&
21096 	    rack->r_ctl.rc_scw) {
21097 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21098 		rack->rack_scwnd_is_idle = 1;
21099 	}
21100 #endif
21101 #ifdef TCP_ACCOUNTING
21102 	if (tot_len_this_send > 0) {
21103 		crtsc = get_cyclecount();
21104 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21105 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
21106 		}
21107 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21108 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
21109 		}
21110 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21111 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
21112 		}
21113 	} else {
21114 		crtsc = get_cyclecount();
21115 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21116 			tp->tcp_cnt_counters[SND_LIMITED]++;
21117 		}
21118 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21119 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
21120 		}
21121 	}
21122 	sched_unpin();
21123 #endif
21124 	return (0);
21125 
21126 send:
21127 	if ((rack->r_ctl.crte != NULL) &&
21128 	    (rsm == NULL) &&
21129 	    ((rack->rc_hw_nobuf == 1) ||
21130 	     (rack_hw_check_queue && (check_done == 0)))) {
21131 		/*
21132 		 * We only want to do this once with the hw_check_queue,
21133 		 * for the enobuf case we would only do it once if
21134 		 * we come around to again, the flag will be clear.
21135 		 */
21136 		check_done = 1;
21137 		slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
21138 		if (slot) {
21139 			rack->r_ctl.rc_agg_delayed = 0;
21140 			rack->r_ctl.rc_agg_early = 0;
21141 			rack->r_early = 0;
21142 			rack->r_late = 0;
21143 			SOCKBUF_UNLOCK(&so->so_snd);
21144 			goto skip_all_send;
21145 		}
21146 	}
21147 	if (rsm || sack_rxmit)
21148 		counter_u64_add(rack_nfto_resend, 1);
21149 	else
21150 		counter_u64_add(rack_non_fto_send, 1);
21151 	if ((flags & TH_FIN) &&
21152 	    sbavail(sb)) {
21153 		/*
21154 		 * We do not transmit a FIN
21155 		 * with data outstanding. We
21156 		 * need to make it so all data
21157 		 * is acked first.
21158 		 */
21159 		flags &= ~TH_FIN;
21160 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21161 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21162 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
21163 			/*
21164 			 * Ok less than or right at a MSS is
21165 			 * outstanding. The original FreeBSD stack would
21166 			 * have sent a FIN, which can speed things up for
21167 			 * a transactional application doing a MSG_WAITALL.
21168 			 * To speed things up since we do *not* send a FIN
21169 			 * if data is outstanding, we send a "challenge ack".
21170 			 * The idea behind that is instead of having to have
21171 			 * the peer wait for the delayed-ack timer to run off
21172 			 * we send an ack that makes the peer send us an ack.
21173 			 */
21174 			rack_send_ack_challange(rack);
21175 		}
21176 	}
21177 	/* Enforce stack imposed max seg size if we have one */
21178 	if (pace_max_seg &&
21179 	    (len > pace_max_seg)) {
21180 		mark = 1;
21181 		len = pace_max_seg;
21182 	}
21183 	if ((rsm == NULL) &&
21184 	    (rack->pcm_in_progress == 0) &&
21185 	    (rack->r_ctl.pcm_max_seg > 0) &&
21186 	    (len >= rack->r_ctl.pcm_max_seg)) {
21187 		/* It is large enough for a measurement */
21188 		add_flag |= RACK_IS_PCM;
21189 		rack_log_pcm(rack, 5, len, rack->r_ctl.pcm_max_seg,  add_flag);
21190 	} else if (rack_verbose_logging) {
21191 		rack_log_pcm(rack, 6, len, rack->r_ctl.pcm_max_seg,  add_flag);
21192 	}
21193 
21194 	SOCKBUF_LOCK_ASSERT(sb);
21195 	if (len > 0) {
21196 		if (len >= segsiz)
21197 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
21198 		else
21199 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
21200 	}
21201 	/*
21202 	 * Before ESTABLISHED, force sending of initial options unless TCP
21203 	 * set not to do any options. NOTE: we assume that the IP/TCP header
21204 	 * plus TCP options always fit in a single mbuf, leaving room for a
21205 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
21206 	 * + optlen <= MCLBYTES
21207 	 */
21208 	optlen = 0;
21209 #ifdef INET6
21210 	if (isipv6)
21211 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21212 	else
21213 #endif
21214 		hdrlen = sizeof(struct tcpiphdr);
21215 
21216 	/*
21217 	 * Ok what seq are we sending from. If we have
21218 	 * no rsm to use, then we look at various bits,
21219 	 * if we are putting out a SYN it will be ISS.
21220 	 * If we are retransmitting a FIN it will
21221 	 * be snd_max-1 else its snd_max.
21222 	 */
21223 	if (rsm == NULL) {
21224 		if (flags & TH_SYN)
21225 			rack_seq = tp->iss;
21226 		else if ((flags & TH_FIN) &&
21227 			 (tp->t_flags & TF_SENTFIN))
21228 			rack_seq = tp->snd_max - 1;
21229 		else
21230 			rack_seq = tp->snd_max;
21231 	} else {
21232 		rack_seq = rsm->r_start;
21233 	}
21234 	/*
21235 	 * Compute options for segment. We only have to care about SYN and
21236 	 * established connection segments.  Options for SYN-ACK segments
21237 	 * are handled in TCP syncache.
21238 	 */
21239 	to.to_flags = 0;
21240 	if ((tp->t_flags & TF_NOOPT) == 0) {
21241 		/* Maximum segment size. */
21242 		if (flags & TH_SYN) {
21243 			to.to_mss = tcp_mssopt(&inp->inp_inc);
21244 			if (tp->t_port)
21245 				to.to_mss -= V_tcp_udp_tunneling_overhead;
21246 			to.to_flags |= TOF_MSS;
21247 
21248 			/*
21249 			 * On SYN or SYN|ACK transmits on TFO connections,
21250 			 * only include the TFO option if it is not a
21251 			 * retransmit, as the presence of the TFO option may
21252 			 * have caused the original SYN or SYN|ACK to have
21253 			 * been dropped by a middlebox.
21254 			 */
21255 			if ((tp->t_flags & TF_FASTOPEN) &&
21256 			    (tp->t_rxtshift == 0)) {
21257 				if (tp->t_state == TCPS_SYN_RECEIVED) {
21258 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21259 					to.to_tfo_cookie =
21260 						(u_int8_t *)&tp->t_tfo_cookie.server;
21261 					to.to_flags |= TOF_FASTOPEN;
21262 					wanted_cookie = 1;
21263 				} else if (tp->t_state == TCPS_SYN_SENT) {
21264 					to.to_tfo_len =
21265 						tp->t_tfo_client_cookie_len;
21266 					to.to_tfo_cookie =
21267 						tp->t_tfo_cookie.client;
21268 					to.to_flags |= TOF_FASTOPEN;
21269 					wanted_cookie = 1;
21270 					/*
21271 					 * If we wind up having more data to
21272 					 * send with the SYN than can fit in
21273 					 * one segment, don't send any more
21274 					 * until the SYN|ACK comes back from
21275 					 * the other end.
21276 					 */
21277 					sendalot = 0;
21278 				}
21279 			}
21280 		}
21281 		/* Window scaling. */
21282 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21283 			to.to_wscale = tp->request_r_scale;
21284 			to.to_flags |= TOF_SCALE;
21285 		}
21286 		/* Timestamps. */
21287 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
21288 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21289 			uint32_t ts_to_use;
21290 
21291 			if ((rack->r_rcvpath_rtt_up == 1) &&
21292 			    (ms_cts == rack->r_ctl.last_rcv_tstmp_for_rtt)) {
21293 				/*
21294 				 * When we are doing a rcv_rtt probe all
21295 				 * other timestamps use the next msec. This
21296 				 * is safe since our previous ack is in the
21297 				 * air and we will just have a few more
21298 				 * on the next ms. This assures that only
21299 				 * the one ack has the ms_cts that was on
21300 				 * our ack-probe.
21301 				 */
21302 				ts_to_use = ms_cts + 1;
21303 			} else {
21304 				ts_to_use = ms_cts;
21305 			}
21306 			to.to_tsval = ts_to_use + tp->ts_offset;
21307 			to.to_tsecr = tp->ts_recent;
21308 			to.to_flags |= TOF_TS;
21309 			if ((len == 0) &&
21310 			    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
21311 			    ((ms_cts - rack->r_ctl.last_rcv_tstmp_for_rtt) > RCV_PATH_RTT_MS) &&
21312 			    (tp->snd_una == tp->snd_max) &&
21313 			    (flags & TH_ACK) &&
21314 			    (sbavail(sb) == 0) &&
21315 			    (rack->r_ctl.current_round != 0) &&
21316 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
21317 			    (rack->r_rcvpath_rtt_up == 0)) {
21318 				rack->r_ctl.last_rcv_tstmp_for_rtt = ms_cts;
21319 				rack->r_ctl.last_time_of_arm_rcv = cts;
21320 				rack->r_rcvpath_rtt_up = 1;
21321 				/* Subtract 1 from seq to force a response */
21322 				rack_seq--;
21323 			}
21324 		}
21325 		/* Set receive buffer autosizing timestamp. */
21326 		if (tp->rfbuf_ts == 0 &&
21327 		    (so->so_rcv.sb_flags & SB_AUTOSIZE)) {
21328 			tp->rfbuf_ts = ms_cts;
21329 		}
21330 		/* Selective ACK's. */
21331 		if (tp->t_flags & TF_SACK_PERMIT) {
21332 			if (flags & TH_SYN)
21333 				to.to_flags |= TOF_SACKPERM;
21334 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21335 				 tp->rcv_numsacks > 0) {
21336 				to.to_flags |= TOF_SACK;
21337 				to.to_nsacks = tp->rcv_numsacks;
21338 				to.to_sacks = (u_char *)tp->sackblks;
21339 			}
21340 		}
21341 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21342 		/* TCP-MD5 (RFC2385). */
21343 		if (tp->t_flags & TF_SIGNATURE)
21344 			to.to_flags |= TOF_SIGNATURE;
21345 #endif
21346 
21347 		/* Processing the options. */
21348 		hdrlen += optlen = tcp_addoptions(&to, opt);
21349 		/*
21350 		 * If we wanted a TFO option to be added, but it was unable
21351 		 * to fit, ensure no data is sent.
21352 		 */
21353 		if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
21354 		    !(to.to_flags & TOF_FASTOPEN))
21355 			len = 0;
21356 	}
21357 	if (tp->t_port) {
21358 		if (V_tcp_udp_tunneling_port == 0) {
21359 			/* The port was removed?? */
21360 			SOCKBUF_UNLOCK(&so->so_snd);
21361 #ifdef TCP_ACCOUNTING
21362 			crtsc = get_cyclecount();
21363 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21364 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21365 			}
21366 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21367 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21368 			}
21369 			sched_unpin();
21370 #endif
21371 			return (EHOSTUNREACH);
21372 		}
21373 		hdrlen += sizeof(struct udphdr);
21374 	}
21375 #ifdef INET6
21376 	if (isipv6)
21377 		ipoptlen = ip6_optlen(inp);
21378 	else
21379 #endif
21380 		if (inp->inp_options)
21381 			ipoptlen = inp->inp_options->m_len -
21382 				offsetof(struct ipoption, ipopt_list);
21383 		else
21384 			ipoptlen = 0;
21385 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21386 	ipoptlen += ipsec_optlen;
21387 #endif
21388 
21389 	/*
21390 	 * Adjust data length if insertion of options will bump the packet
21391 	 * length beyond the t_maxseg length. Clear the FIN bit because we
21392 	 * cut off the tail of the segment.
21393 	 */
21394 	if (len + optlen + ipoptlen > tp->t_maxseg) {
21395 		if (tso) {
21396 			uint32_t if_hw_tsomax;
21397 			uint32_t moff;
21398 			int32_t max_len;
21399 
21400 			/* extract TSO information */
21401 			if_hw_tsomax = tp->t_tsomax;
21402 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21403 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21404 			KASSERT(ipoptlen == 0,
21405 				("%s: TSO can't do IP options", __func__));
21406 
21407 			/*
21408 			 * Check if we should limit by maximum payload
21409 			 * length:
21410 			 */
21411 			if (if_hw_tsomax != 0) {
21412 				/* compute maximum TSO length */
21413 				max_len = (if_hw_tsomax - hdrlen -
21414 					   max_linkhdr);
21415 				if (max_len <= 0) {
21416 					len = 0;
21417 				} else if (len > max_len) {
21418 					sendalot = 1;
21419 					len = max_len;
21420 					mark = 2;
21421 				}
21422 			}
21423 			/*
21424 			 * Prevent the last segment from being fractional
21425 			 * unless the send sockbuf can be emptied:
21426 			 */
21427 			max_len = (tp->t_maxseg - optlen);
21428 			if ((sb_offset + len) < sbavail(sb)) {
21429 				moff = len % (u_int)max_len;
21430 				if (moff != 0) {
21431 					mark = 3;
21432 					len -= moff;
21433 				}
21434 			}
21435 			/*
21436 			 * In case there are too many small fragments don't
21437 			 * use TSO:
21438 			 */
21439 			if (len <= max_len) {
21440 				mark = 4;
21441 				tso = 0;
21442 			}
21443 			/*
21444 			 * Send the FIN in a separate segment after the bulk
21445 			 * sending is done. We don't trust the TSO
21446 			 * implementations to clear the FIN flag on all but
21447 			 * the last segment.
21448 			 */
21449 			if (tp->t_flags & TF_NEEDFIN) {
21450 				sendalot = 4;
21451 			}
21452 		} else {
21453 			mark = 5;
21454 			if (optlen + ipoptlen >= tp->t_maxseg) {
21455 				/*
21456 				 * Since we don't have enough space to put
21457 				 * the IP header chain and the TCP header in
21458 				 * one packet as required by RFC 7112, don't
21459 				 * send it. Also ensure that at least one
21460 				 * byte of the payload can be put into the
21461 				 * TCP segment.
21462 				 */
21463 				SOCKBUF_UNLOCK(&so->so_snd);
21464 				error = EMSGSIZE;
21465 				sack_rxmit = 0;
21466 				goto out;
21467 			}
21468 			len = tp->t_maxseg - optlen - ipoptlen;
21469 			sendalot = 5;
21470 		}
21471 	} else {
21472 		tso = 0;
21473 		mark = 6;
21474 	}
21475 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21476 		("%s: len > IP_MAXPACKET", __func__));
21477 #ifdef DIAGNOSTIC
21478 #ifdef INET6
21479 	if (max_linkhdr + hdrlen > MCLBYTES)
21480 #else
21481 		if (max_linkhdr + hdrlen > MHLEN)
21482 #endif
21483 			panic("tcphdr too big");
21484 #endif
21485 
21486 	/*
21487 	 * This KASSERT is here to catch edge cases at a well defined place.
21488 	 * Before, those had triggered (random) panic conditions further
21489 	 * down.
21490 	 */
21491 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21492 	if ((len == 0) &&
21493 	    (flags & TH_FIN) &&
21494 	    (sbused(sb))) {
21495 		/*
21496 		 * We have outstanding data, don't send a fin by itself!.
21497 		 *
21498 		 * Check to see if we need to send a challenge ack.
21499 		 */
21500 		if ((sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21501 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
21502 			/*
21503 			 * Ok less than or right at a MSS is
21504 			 * outstanding. The original FreeBSD stack would
21505 			 * have sent a FIN, which can speed things up for
21506 			 * a transactional application doing a MSG_WAITALL.
21507 			 * To speed things up since we do *not* send a FIN
21508 			 * if data is outstanding, we send a "challenge ack".
21509 			 * The idea behind that is instead of having to have
21510 			 * the peer wait for the delayed-ack timer to run off
21511 			 * we send an ack that makes the peer send us an ack.
21512 			 */
21513 			rack_send_ack_challange(rack);
21514 		}
21515 		goto just_return;
21516 	}
21517 	/*
21518 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
21519 	 * and initialize the header from the template for sends on this
21520 	 * connection.
21521 	 */
21522 	hw_tls = tp->t_nic_ktls_xmit != 0;
21523 	if (len) {
21524 		uint32_t max_val;
21525 		uint32_t moff;
21526 
21527 		if (pace_max_seg)
21528 			max_val = pace_max_seg;
21529 		else
21530 			max_val = len;
21531 		/*
21532 		 * We allow a limit on sending with hptsi.
21533 		 */
21534 		if (len > max_val) {
21535 			mark = 7;
21536 			len = max_val;
21537 		}
21538 #ifdef INET6
21539 		if (MHLEN < hdrlen + max_linkhdr)
21540 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21541 		else
21542 #endif
21543 			m = m_gethdr(M_NOWAIT, MT_DATA);
21544 
21545 		if (m == NULL) {
21546 			SOCKBUF_UNLOCK(sb);
21547 			error = ENOBUFS;
21548 			sack_rxmit = 0;
21549 			goto out;
21550 		}
21551 		m->m_data += max_linkhdr;
21552 		m->m_len = hdrlen;
21553 
21554 		/*
21555 		 * Start the m_copy functions from the closest mbuf to the
21556 		 * sb_offset in the socket buffer chain.
21557 		 */
21558 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
21559 		s_mb = mb;
21560 		s_moff = moff;
21561 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21562 			m_copydata(mb, moff, (int)len,
21563 				   mtod(m, caddr_t)+hdrlen);
21564 			/*
21565 			 * If we are not retransmitting advance the
21566 			 * sndptr to help remember the next place in
21567 			 * the sb.
21568 			 */
21569 			if (rsm == NULL)
21570 				sbsndptr_adv(sb, mb, len);
21571 			m->m_len += len;
21572 		} else {
21573 			struct sockbuf *msb;
21574 
21575 			/*
21576 			 * If we are not retransmitting pass in msb so
21577 			 * the socket buffer can be advanced. Otherwise
21578 			 * set it to NULL if its a retransmission since
21579 			 * we don't want to change the sb remembered
21580 			 * location.
21581 			 */
21582 			if (rsm == NULL)
21583 				msb = sb;
21584 			else
21585 				msb = NULL;
21586 			m->m_next = tcp_m_copym(
21587 				mb, moff, &len,
21588 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21589 				((rsm == NULL) ? hw_tls : 0)
21590 #ifdef NETFLIX_COPY_ARGS
21591 				, &s_mb, &s_moff
21592 #endif
21593 				);
21594 			if (len <= (tp->t_maxseg - optlen)) {
21595 				/*
21596 				 * Must have ran out of mbufs for the copy
21597 				 * shorten it to no longer need tso. Lets
21598 				 * not put on sendalot since we are low on
21599 				 * mbufs.
21600 				 */
21601 				tso = 0;
21602 			}
21603 			if (m->m_next == NULL) {
21604 				SOCKBUF_UNLOCK(sb);
21605 				(void)m_free(m);
21606 				error = ENOBUFS;
21607 				sack_rxmit = 0;
21608 				goto out;
21609 			}
21610 		}
21611 		if (sack_rxmit) {
21612 			if (rsm && (rsm->r_flags & RACK_TLP)) {
21613 				/*
21614 				 * TLP should not count in retran count, but
21615 				 * in its own bin
21616 				 */
21617 				counter_u64_add(rack_tlp_retran, 1);
21618 				counter_u64_add(rack_tlp_retran_bytes, len);
21619 			} else {
21620 				tp->t_sndrexmitpack++;
21621 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21622 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21623 			}
21624 #ifdef STATS
21625 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21626 						 len);
21627 #endif
21628 		} else {
21629 			KMOD_TCPSTAT_INC(tcps_sndpack);
21630 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21631 #ifdef STATS
21632 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21633 						 len);
21634 #endif
21635 		}
21636 		/*
21637 		 * If we're sending everything we've got, set PUSH. (This
21638 		 * will keep happy those implementations which only give
21639 		 * data to the user when a buffer fills or a PUSH comes in.)
21640 		 */
21641 		if (sb_offset + len == sbused(sb) &&
21642 		    sbused(sb) &&
21643 		    !(flags & TH_SYN)) {
21644 			flags |= TH_PUSH;
21645 			add_flag |= RACK_HAD_PUSH;
21646 		}
21647 
21648 		SOCKBUF_UNLOCK(sb);
21649 	} else {
21650 		SOCKBUF_UNLOCK(sb);
21651 		if (tp->t_flags & TF_ACKNOW)
21652 			KMOD_TCPSTAT_INC(tcps_sndacks);
21653 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
21654 			KMOD_TCPSTAT_INC(tcps_sndctrl);
21655 		else
21656 			KMOD_TCPSTAT_INC(tcps_sndwinup);
21657 
21658 		m = m_gethdr(M_NOWAIT, MT_DATA);
21659 		if (m == NULL) {
21660 			error = ENOBUFS;
21661 			sack_rxmit = 0;
21662 			goto out;
21663 		}
21664 #ifdef INET6
21665 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21666 		    MHLEN >= hdrlen) {
21667 			M_ALIGN(m, hdrlen);
21668 		} else
21669 #endif
21670 			m->m_data += max_linkhdr;
21671 		m->m_len = hdrlen;
21672 	}
21673 	SOCKBUF_UNLOCK_ASSERT(sb);
21674 	m->m_pkthdr.rcvif = (struct ifnet *)0;
21675 #ifdef MAC
21676 	mac_inpcb_create_mbuf(inp, m);
21677 #endif
21678 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
21679 #ifdef INET6
21680 		if (isipv6)
21681 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21682 		else
21683 #endif				/* INET6 */
21684 #ifdef INET
21685 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21686 #endif
21687 		th = rack->r_ctl.fsb.th;
21688 		udp = rack->r_ctl.fsb.udp;
21689 		if (udp) {
21690 #ifdef INET6
21691 			if (isipv6)
21692 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21693 			else
21694 #endif				/* INET6 */
21695 				ulen = hdrlen + len - sizeof(struct ip);
21696 			udp->uh_ulen = htons(ulen);
21697 		}
21698 	} else {
21699 #ifdef INET6
21700 		if (isipv6) {
21701 			ip6 = mtod(m, struct ip6_hdr *);
21702 			if (tp->t_port) {
21703 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21704 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21705 				udp->uh_dport = tp->t_port;
21706 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21707 				udp->uh_ulen = htons(ulen);
21708 				th = (struct tcphdr *)(udp + 1);
21709 			} else
21710 				th = (struct tcphdr *)(ip6 + 1);
21711 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
21712 		} else
21713 #endif				/* INET6 */
21714 		{
21715 #ifdef INET
21716 			ip = mtod(m, struct ip *);
21717 			if (tp->t_port) {
21718 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21719 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21720 				udp->uh_dport = tp->t_port;
21721 				ulen = hdrlen + len - sizeof(struct ip);
21722 				udp->uh_ulen = htons(ulen);
21723 				th = (struct tcphdr *)(udp + 1);
21724 			} else
21725 				th = (struct tcphdr *)(ip + 1);
21726 			tcpip_fillheaders(inp, tp->t_port, ip, th);
21727 #endif
21728 		}
21729 	}
21730 	/*
21731 	 * If we are starting a connection, send ECN setup SYN packet. If we
21732 	 * are on a retransmit, we may resend those bits a number of times
21733 	 * as per RFC 3168.
21734 	 */
21735 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21736 		flags |= tcp_ecn_output_syn_sent(tp);
21737 	}
21738 	/* Also handle parallel SYN for ECN */
21739 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
21740 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21741 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21742 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21743 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
21744 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21745 #ifdef INET6
21746 		if (isipv6) {
21747 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21748 			ip6->ip6_flow |= htonl(ect << 20);
21749 		}
21750 		else
21751 #endif
21752 		{
21753 #ifdef INET
21754 			ip->ip_tos &= ~IPTOS_ECN_MASK;
21755 			ip->ip_tos |= ect;
21756 #endif
21757 		}
21758 	}
21759 	th->th_seq = htonl(rack_seq);
21760 	th->th_ack = htonl(tp->rcv_nxt);
21761 	tcp_set_flags(th, flags);
21762 	/*
21763 	 * Calculate receive window.  Don't shrink window, but avoid silly
21764 	 * window syndrome.
21765 	 * If a RST segment is sent, advertise a window of zero.
21766 	 */
21767 	if (flags & TH_RST) {
21768 		recwin = 0;
21769 	} else {
21770 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21771 		    recwin < (long)segsiz) {
21772 			recwin = 0;
21773 		}
21774 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21775 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21776 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21777 	}
21778 
21779 	/*
21780 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21781 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
21782 	 * handled in syncache.
21783 	 */
21784 	if (flags & TH_SYN)
21785 		th->th_win = htons((u_short)
21786 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21787 	else {
21788 		/* Avoid shrinking window with window scaling. */
21789 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
21790 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21791 	}
21792 	/*
21793 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21794 	 * window.  This may cause the remote transmitter to stall.  This
21795 	 * flag tells soreceive() to disable delayed acknowledgements when
21796 	 * draining the buffer.  This can occur if the receiver is
21797 	 * attempting to read more data than can be buffered prior to
21798 	 * transmitting on the connection.
21799 	 */
21800 	if (th->th_win == 0) {
21801 		tp->t_sndzerowin++;
21802 		tp->t_flags |= TF_RXWIN0SENT;
21803 	} else
21804 		tp->t_flags &= ~TF_RXWIN0SENT;
21805 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
21806 	/* Now are we using fsb?, if so copy the template data to the mbuf */
21807 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21808 		uint8_t *cpto;
21809 
21810 		cpto = mtod(m, uint8_t *);
21811 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21812 		/*
21813 		 * We have just copied in:
21814 		 * IP/IP6
21815 		 * <optional udphdr>
21816 		 * tcphdr (no options)
21817 		 *
21818 		 * We need to grab the correct pointers into the mbuf
21819 		 * for both the tcp header, and possibly the udp header (if tunneling).
21820 		 * We do this by using the offset in the copy buffer and adding it
21821 		 * to the mbuf base pointer (cpto).
21822 		 */
21823 #ifdef INET6
21824 		if (isipv6)
21825 			ip6 = mtod(m, struct ip6_hdr *);
21826 		else
21827 #endif				/* INET6 */
21828 #ifdef INET
21829 			ip = mtod(m, struct ip *);
21830 #endif
21831 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21832 		/* If we have a udp header lets set it into the mbuf as well */
21833 		if (udp)
21834 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21835 	}
21836 	if (optlen) {
21837 		bcopy(opt, th + 1, optlen);
21838 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21839 	}
21840 	/*
21841 	 * Put TCP length in extended header, and then checksum extended
21842 	 * header and data.
21843 	 */
21844 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
21845 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21846 	if (to.to_flags & TOF_SIGNATURE) {
21847 		/*
21848 		 * Calculate MD5 signature and put it into the place
21849 		 * determined before.
21850 		 * NOTE: since TCP options buffer doesn't point into
21851 		 * mbuf's data, calculate offset and use it.
21852 		 */
21853 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21854 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21855 			/*
21856 			 * Do not send segment if the calculation of MD5
21857 			 * digest has failed.
21858 			 */
21859 			goto out;
21860 		}
21861 	}
21862 #endif
21863 #ifdef INET6
21864 	if (isipv6) {
21865 		/*
21866 		 * ip6_plen is not need to be filled now, and will be filled
21867 		 * in ip6_output.
21868 		 */
21869 		if (tp->t_port) {
21870 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21871 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21872 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21873 			th->th_sum = htons(0);
21874 			UDPSTAT_INC(udps_opackets);
21875 		} else {
21876 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21877 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21878 			th->th_sum = in6_cksum_pseudo(ip6,
21879 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21880 						      0);
21881 		}
21882 	}
21883 #endif
21884 #if defined(INET6) && defined(INET)
21885 	else
21886 #endif
21887 #ifdef INET
21888 	{
21889 		if (tp->t_port) {
21890 			m->m_pkthdr.csum_flags = CSUM_UDP;
21891 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21892 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21893 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21894 			th->th_sum = htons(0);
21895 			UDPSTAT_INC(udps_opackets);
21896 		} else {
21897 			m->m_pkthdr.csum_flags = CSUM_TCP;
21898 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21899 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
21900 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21901 									IPPROTO_TCP + len + optlen));
21902 		}
21903 		/* IP version must be set here for ipv4/ipv6 checking later */
21904 		KASSERT(ip->ip_v == IPVERSION,
21905 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
21906 	}
21907 #endif
21908 	/*
21909 	 * Enable TSO and specify the size of the segments. The TCP pseudo
21910 	 * header checksum is always provided. XXX: Fixme: This is currently
21911 	 * not the case for IPv6.
21912 	 */
21913 	if (tso) {
21914 		/*
21915 		 * Here we must use t_maxseg and the optlen since
21916 		 * the optlen may include SACK's (or DSACK).
21917 		 */
21918 		KASSERT(len > tp->t_maxseg - optlen,
21919 			("%s: len <= tso_segsz", __func__));
21920 		m->m_pkthdr.csum_flags |= CSUM_TSO;
21921 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21922 	}
21923 	KASSERT(len + hdrlen == m_length(m, NULL),
21924 		("%s: mbuf chain different than expected: %d + %u != %u",
21925 		 __func__, len, hdrlen, m_length(m, NULL)));
21926 
21927 #ifdef TCP_HHOOK
21928 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21929 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
21930 #endif
21931 	if ((rack->r_ctl.crte != NULL) &&
21932 	    (rack->rc_hw_nobuf == 0) &&
21933 	    tcp_bblogging_on(tp)) {
21934 		rack_log_queue_level(tp, rack, len, &tv, cts);
21935 	}
21936 	/* We're getting ready to send; log now. */
21937 	if (tcp_bblogging_on(rack->rc_tp)) {
21938 		union tcp_log_stackspecific log;
21939 
21940 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
21941 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
21942 		if (rack->rack_no_prr)
21943 			log.u_bbr.flex1 = 0;
21944 		else
21945 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21946 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
21947 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
21948 		log.u_bbr.flex4 = orig_len;
21949 		/* Save off the early/late values */
21950 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
21951 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
21952 		log.u_bbr.bw_inuse = rack_get_bw(rack);
21953 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
21954 		log.u_bbr.flex8 = 0;
21955 		if (rsm) {
21956 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
21957 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
21958 				counter_u64_add(rack_collapsed_win_rxt, 1);
21959 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
21960 			}
21961 			if (doing_tlp)
21962 				log.u_bbr.flex8 = 2;
21963 			else
21964 				log.u_bbr.flex8 = 1;
21965 		} else {
21966 			if (doing_tlp)
21967 				log.u_bbr.flex8 = 3;
21968 		}
21969 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
21970 		log.u_bbr.flex7 = mark;
21971 		log.u_bbr.flex7 <<= 8;
21972 		log.u_bbr.flex7 |= pass;
21973 		log.u_bbr.pkts_out = tp->t_maxseg;
21974 		log.u_bbr.timeStamp = cts;
21975 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
21976 		if (rsm && (rsm->r_rtr_cnt > 0)) {
21977 			/*
21978 			 * When we have a retransmit we want to log the
21979 			 * burst at send and flight at send from before.
21980 			 */
21981 			log.u_bbr.flex5 = rsm->r_fas;
21982 			log.u_bbr.bbr_substate = rsm->r_bas;
21983 		} else {
21984 			/*
21985 			 * New transmits we log in flex5 the inflight again as
21986 			 * well as the number of segments in our send in the
21987 			 * substate field.
21988 			 */
21989 			log.u_bbr.flex5 = log.u_bbr.inflight;
21990 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
21991 		}
21992 		log.u_bbr.lt_epoch = cwnd_to_use;
21993 		log.u_bbr.delivered = sendalot;
21994 		log.u_bbr.rttProp = (uintptr_t)rsm;
21995 		log.u_bbr.pkt_epoch = __LINE__;
21996 		if (rsm) {
21997 			log.u_bbr.delRate = rsm->r_flags;
21998 			log.u_bbr.delRate <<= 31;
21999 			log.u_bbr.delRate |= rack->r_must_retran;
22000 			log.u_bbr.delRate <<= 1;
22001 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22002 		} else {
22003 			log.u_bbr.delRate = rack->r_must_retran;
22004 			log.u_bbr.delRate <<= 1;
22005 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22006 		}
22007 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
22008 				    len, &log, false, NULL, __func__, __LINE__, &tv);
22009 	} else
22010 		lgb = NULL;
22011 
22012 	/*
22013 	 * Fill in IP length and desired time to live and send to IP level.
22014 	 * There should be a better way to handle ttl and tos; we could keep
22015 	 * them in the template, but need a way to checksum without them.
22016 	 */
22017 	/*
22018 	 * m->m_pkthdr.len should have been set before cksum calcuration,
22019 	 * because in6_cksum() need it.
22020 	 */
22021 #ifdef INET6
22022 	if (isipv6) {
22023 		/*
22024 		 * we separately set hoplimit for every segment, since the
22025 		 * user might want to change the value via setsockopt. Also,
22026 		 * desired default hop limit might be changed via Neighbor
22027 		 * Discovery.
22028 		 */
22029 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
22030 
22031 		/*
22032 		 * Set the packet size here for the benefit of DTrace
22033 		 * probes. ip6_output() will set it properly; it's supposed
22034 		 * to include the option header lengths as well.
22035 		 */
22036 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
22037 
22038 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
22039 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22040 		else
22041 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22042 
22043 		if (tp->t_state == TCPS_SYN_SENT)
22044 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
22045 
22046 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
22047 		/* TODO: IPv6 IP6TOS_ECT bit on */
22048 		error = ip6_output(m,
22049 				   inp->in6p_outputopts,
22050 				   &inp->inp_route6,
22051 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
22052 				   NULL, NULL, inp);
22053 
22054 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
22055 			mtu = inp->inp_route6.ro_nh->nh_mtu;
22056 	}
22057 #endif				/* INET6 */
22058 #if defined(INET) && defined(INET6)
22059 	else
22060 #endif
22061 #ifdef INET
22062 	{
22063 		ip->ip_len = htons(m->m_pkthdr.len);
22064 #ifdef INET6
22065 		if (inp->inp_vflag & INP_IPV6PROTO)
22066 			ip->ip_ttl = in6_selecthlim(inp, NULL);
22067 #endif				/* INET6 */
22068 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
22069 		/*
22070 		 * If we do path MTU discovery, then we set DF on every
22071 		 * packet. This might not be the best thing to do according
22072 		 * to RFC3390 Section 2. However the tcp hostcache migitates
22073 		 * the problem so it affects only the first tcp connection
22074 		 * with a host.
22075 		 *
22076 		 * NB: Don't set DF on small MTU/MSS to have a safe
22077 		 * fallback.
22078 		 */
22079 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
22080 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22081 			if (tp->t_port == 0 || len < V_tcp_minmss) {
22082 				ip->ip_off |= htons(IP_DF);
22083 			}
22084 		} else {
22085 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22086 		}
22087 
22088 		if (tp->t_state == TCPS_SYN_SENT)
22089 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
22090 
22091 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
22092 
22093 		error = ip_output(m,
22094 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22095 				  inp->inp_options,
22096 #else
22097 				  NULL,
22098 #endif
22099 				  &inp->inp_route,
22100 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
22101 				  inp);
22102 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
22103 			mtu = inp->inp_route.ro_nh->nh_mtu;
22104 	}
22105 #endif				/* INET */
22106 	if (lgb) {
22107 		lgb->tlb_errno = error;
22108 		lgb = NULL;
22109 	}
22110 
22111 out:
22112 	/*
22113 	 * In transmit state, time the transmission and arrange for the
22114 	 * retransmit.  In persist state, just set snd_max.
22115 	 */
22116 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
22117 			rack_to_usec_ts(&tv),
22118 			rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
22119 	if (error == 0) {
22120 		if (add_flag & RACK_IS_PCM) {
22121 			/* We just launched a PCM */
22122 			/* rrs here log */
22123 			rack->pcm_in_progress = 1;
22124 			rack->pcm_needed = 0;
22125 			rack_log_pcm(rack, 7, len, rack->r_ctl.pcm_max_seg,  add_flag);
22126 		}
22127 		if (rsm == NULL) {
22128 			if (rack->lt_bw_up == 0) {
22129 				rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
22130 				rack->r_ctl.lt_seq = tp->snd_una;
22131 				rack->lt_bw_up = 1;
22132 			} else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
22133 				/*
22134 				 * Need to record what we have since we are
22135 				 * approaching seq wrap.
22136 				 */
22137 				uint64_t tmark;
22138 
22139 				rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
22140 				rack->r_ctl.lt_seq = tp->snd_una;
22141 				tmark = tcp_get_u64_usecs(&tv);
22142 				if (tmark > rack->r_ctl.lt_timemark) {
22143 					rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
22144 					rack->r_ctl.lt_timemark = tmark;
22145 				}
22146 			}
22147 		}
22148 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
22149 		counter_u64_add(rack_total_bytes, len);
22150 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
22151 		if (rsm && doing_tlp) {
22152 			rack->rc_last_sent_tlp_past_cumack = 0;
22153 			rack->rc_last_sent_tlp_seq_valid = 1;
22154 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
22155 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
22156 		}
22157 		if (rack->rc_hw_nobuf) {
22158 			rack->rc_hw_nobuf = 0;
22159 			rack->r_ctl.rc_agg_delayed = 0;
22160 			rack->r_early = 0;
22161 			rack->r_late = 0;
22162 			rack->r_ctl.rc_agg_early = 0;
22163 		}
22164 		if (rsm && (doing_tlp == 0)) {
22165 			/* Set we retransmitted */
22166 			rack->rc_gp_saw_rec = 1;
22167 		} else {
22168 			if (cwnd_to_use > tp->snd_ssthresh) {
22169 				/* Set we sent in CA */
22170 				rack->rc_gp_saw_ca = 1;
22171 			} else {
22172 				/* Set we sent in SS */
22173 				rack->rc_gp_saw_ss = 1;
22174 			}
22175 		}
22176 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22177 		    (tp->t_flags & TF_SACK_PERMIT) &&
22178 		    tp->rcv_numsacks > 0)
22179 			tcp_clean_dsack_blocks(tp);
22180 		tot_len_this_send += len;
22181 		if (len == 0) {
22182 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
22183 		} else {
22184 			int idx;
22185 
22186 			idx = (len / segsiz) + 3;
22187 			if (idx >= TCP_MSS_ACCT_ATIMER)
22188 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
22189 			else
22190 				counter_u64_add(rack_out_size[idx], 1);
22191 		}
22192 	}
22193 	if ((rack->rack_no_prr == 0) &&
22194 	    sub_from_prr &&
22195 	    (error == 0)) {
22196 		if (rack->r_ctl.rc_prr_sndcnt >= len)
22197 			rack->r_ctl.rc_prr_sndcnt -= len;
22198 		else
22199 			rack->r_ctl.rc_prr_sndcnt = 0;
22200 	}
22201 	sub_from_prr = 0;
22202 	if (doing_tlp) {
22203 		/* Make sure the TLP is added */
22204 		add_flag |= RACK_TLP;
22205 	} else if (rsm) {
22206 		/* If its a resend without TLP then it must not have the flag */
22207 		rsm->r_flags &= ~RACK_TLP;
22208 	}
22209 
22210 
22211 	if ((error == 0) &&
22212 	    (len > 0) &&
22213 	    (tp->snd_una == tp->snd_max))
22214 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
22215 
22216 	{
22217 		/*
22218 		 * This block is not associated with the above error == 0 test.
22219 		 * It is used to advance snd_max if we have a new transmit.
22220 		 */
22221 		tcp_seq startseq = tp->snd_max;
22222 
22223 
22224 		if (rsm && (doing_tlp == 0))
22225 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
22226 		if (error)
22227 			/* We don't log or do anything with errors */
22228 			goto nomore;
22229 		if (doing_tlp == 0) {
22230 			if (rsm == NULL) {
22231 				/*
22232 				 * Not a retransmission of some
22233 				 * sort, new data is going out so
22234 				 * clear our TLP count and flag.
22235 				 */
22236 				rack->rc_tlp_in_progress = 0;
22237 				rack->r_ctl.rc_tlp_cnt_out = 0;
22238 			}
22239 		} else {
22240 			/*
22241 			 * We have just sent a TLP, mark that it is true
22242 			 * and make sure our in progress is set so we
22243 			 * continue to check the count.
22244 			 */
22245 			rack->rc_tlp_in_progress = 1;
22246 			rack->r_ctl.rc_tlp_cnt_out++;
22247 		}
22248 		/*
22249 		 * If we are retransmitting we are done, snd_max
22250 		 * does not get updated.
22251 		 */
22252 		if (sack_rxmit)
22253 			goto nomore;
22254 		if ((tp->snd_una == tp->snd_max) && (len > 0)) {
22255 			/*
22256 			 * Update the time we just added data since
22257 			 * nothing was outstanding.
22258 			 */
22259 			rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22260 			tp->t_acktime = ticks;
22261 		}
22262 		/*
22263 		 * Now for special SYN/FIN handling.
22264 		 */
22265 		if (flags & (TH_SYN | TH_FIN)) {
22266 			if ((flags & TH_SYN) &&
22267 			    ((tp->t_flags & TF_SENTSYN) == 0)) {
22268 				tp->snd_max++;
22269 				tp->t_flags |= TF_SENTSYN;
22270 			}
22271 			if ((flags & TH_FIN) &&
22272 			    ((tp->t_flags & TF_SENTFIN) == 0)) {
22273 				tp->snd_max++;
22274 				tp->t_flags |= TF_SENTFIN;
22275 			}
22276 		}
22277 		tp->snd_max += len;
22278 		if (rack->rc_new_rnd_needed) {
22279 			rack_new_round_starts(tp, rack, tp->snd_max);
22280 		}
22281 		/*
22282 		 * Time this transmission if not a retransmission and
22283 		 * not currently timing anything.
22284 		 * This is only relevant in case of switching back to
22285 		 * the base stack.
22286 		 */
22287 		if (tp->t_rtttime == 0) {
22288 			tp->t_rtttime = ticks;
22289 			tp->t_rtseq = startseq;
22290 			KMOD_TCPSTAT_INC(tcps_segstimed);
22291 		}
22292 		if (len &&
22293 		    ((tp->t_flags & TF_GPUTINPROG) == 0))
22294 			rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22295 		/*
22296 		 * If we are doing FO we need to update the mbuf position and subtract
22297 		 * this happens when the peer sends us duplicate information and
22298 		 * we thus want to send a DSACK.
22299 		 *
22300 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22301 		 * turned off? If not then we are going to echo multiple DSACK blocks
22302 		 * out (with the TSO), which we should not be doing.
22303 		 */
22304 		if (rack->r_fast_output && len) {
22305 			if (rack->r_ctl.fsb.left_to_send > len)
22306 				rack->r_ctl.fsb.left_to_send -= len;
22307 			else
22308 				rack->r_ctl.fsb.left_to_send = 0;
22309 			if (rack->r_ctl.fsb.left_to_send < segsiz)
22310 				rack->r_fast_output = 0;
22311 			if (rack->r_fast_output) {
22312 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22313 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22314 				rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22315 			}
22316 		}
22317 		if (rack_pcm_blast == 0) {
22318 			if ((orig_len > len) &&
22319 			    (add_flag & RACK_IS_PCM) &&
22320 			    (len < pace_max_seg) &&
22321 			    ((pace_max_seg - len) > segsiz)) {
22322 				/*
22323 				 * We are doing a PCM measurement and we did
22324 				 * not get enough data in the TSO to meet the
22325 				 * burst requirement.
22326 				 */
22327 				uint32_t n_len;
22328 
22329 				n_len = (orig_len - len);
22330 				orig_len -= len;
22331 				pace_max_seg -= len;
22332 				len = n_len;
22333 				sb_offset = tp->snd_max - tp->snd_una;
22334 				/* Re-lock for the next spin */
22335 				SOCKBUF_LOCK(sb);
22336 				goto send;
22337 			}
22338 		} else {
22339 			if ((orig_len > len) &&
22340 			    (add_flag & RACK_IS_PCM) &&
22341 			    ((orig_len - len) > segsiz)) {
22342 				/*
22343 				 * We are doing a PCM measurement and we did
22344 				 * not get enough data in the TSO to meet the
22345 				 * burst requirement.
22346 				 */
22347 				uint32_t n_len;
22348 
22349 				n_len = (orig_len - len);
22350 				orig_len -= len;
22351 				len = n_len;
22352 				sb_offset = tp->snd_max - tp->snd_una;
22353 				/* Re-lock for the next spin */
22354 				SOCKBUF_LOCK(sb);
22355 				goto send;
22356 			}
22357 		}
22358 	}
22359 nomore:
22360 	if (error) {
22361 		rack->r_ctl.rc_agg_delayed = 0;
22362 		rack->r_early = 0;
22363 		rack->r_late = 0;
22364 		rack->r_ctl.rc_agg_early = 0;
22365 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
22366 		/*
22367 		 * Failures do not advance the seq counter above. For the
22368 		 * case of ENOBUFS we will fall out and retry in 1ms with
22369 		 * the hpts. Everything else will just have to retransmit
22370 		 * with the timer.
22371 		 *
22372 		 * In any case, we do not want to loop around for another
22373 		 * send without a good reason.
22374 		 */
22375 		sendalot = 0;
22376 		switch (error) {
22377 		case EPERM:
22378 		case EACCES:
22379 			tp->t_softerror = error;
22380 #ifdef TCP_ACCOUNTING
22381 			crtsc = get_cyclecount();
22382 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22383 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22384 			}
22385 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22386 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22387 			}
22388 			sched_unpin();
22389 #endif
22390 			return (error);
22391 		case ENOBUFS:
22392 			/*
22393 			 * Pace us right away to retry in a some
22394 			 * time
22395 			 */
22396 			if (rack->r_ctl.crte != NULL) {
22397 				tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22398 				if (tcp_bblogging_on(rack->rc_tp))
22399 					rack_log_queue_level(tp, rack, len, &tv, cts);
22400 			} else
22401 				tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22402 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22403 			if (rack->rc_enobuf < 0x7f)
22404 				rack->rc_enobuf++;
22405 			if (slot < (10 * HPTS_USEC_IN_MSEC))
22406 				slot = 10 * HPTS_USEC_IN_MSEC;
22407 			if (rack->r_ctl.crte != NULL) {
22408 				counter_u64_add(rack_saw_enobuf_hw, 1);
22409 				tcp_rl_log_enobuf(rack->r_ctl.crte);
22410 			}
22411 			counter_u64_add(rack_saw_enobuf, 1);
22412 			goto enobufs;
22413 		case EMSGSIZE:
22414 			/*
22415 			 * For some reason the interface we used initially
22416 			 * to send segments changed to another or lowered
22417 			 * its MTU. If TSO was active we either got an
22418 			 * interface without TSO capabilits or TSO was
22419 			 * turned off. If we obtained mtu from ip_output()
22420 			 * then update it and try again.
22421 			 */
22422 			if (tso)
22423 				tp->t_flags &= ~TF_TSO;
22424 			if (mtu != 0) {
22425 				int saved_mtu;
22426 
22427 				saved_mtu = tp->t_maxseg;
22428 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
22429 				if (saved_mtu > tp->t_maxseg) {
22430 					goto again;
22431 				}
22432 			}
22433 			slot = 10 * HPTS_USEC_IN_MSEC;
22434 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22435 #ifdef TCP_ACCOUNTING
22436 			crtsc = get_cyclecount();
22437 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22438 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22439 			}
22440 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22441 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22442 			}
22443 			sched_unpin();
22444 #endif
22445 			return (error);
22446 		case ENETUNREACH:
22447 			counter_u64_add(rack_saw_enetunreach, 1);
22448 		case EHOSTDOWN:
22449 		case EHOSTUNREACH:
22450 		case ENETDOWN:
22451 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
22452 				tp->t_softerror = error;
22453 				error = 0;
22454 			}
22455 			/* FALLTHROUGH */
22456 		default:
22457 			slot = 10 * HPTS_USEC_IN_MSEC;
22458 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22459 #ifdef TCP_ACCOUNTING
22460 			crtsc = get_cyclecount();
22461 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22462 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22463 			}
22464 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22465 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22466 			}
22467 			sched_unpin();
22468 #endif
22469 			return (error);
22470 		}
22471 	} else {
22472 		rack->rc_enobuf = 0;
22473 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22474 			rack->r_ctl.retran_during_recovery += len;
22475 	}
22476 	KMOD_TCPSTAT_INC(tcps_sndtotal);
22477 
22478 	/*
22479 	 * Data sent (as far as we can tell). If this advertises a larger
22480 	 * window than any other segment, then remember the size of the
22481 	 * advertised window. Any pending ACK has now been sent.
22482 	 */
22483 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22484 		tp->rcv_adv = tp->rcv_nxt + recwin;
22485 
22486 	tp->last_ack_sent = tp->rcv_nxt;
22487 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22488 enobufs:
22489 	if (sendalot) {
22490 		/* Do we need to turn off sendalot? */
22491 		if (pace_max_seg &&
22492 		    (tot_len_this_send >= pace_max_seg)) {
22493 			/* We hit our max. */
22494 			sendalot = 0;
22495 		}
22496 	}
22497 	if ((error == 0) && (flags & TH_FIN))
22498 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22499 	if (flags & TH_RST) {
22500 		/*
22501 		 * We don't send again after sending a RST.
22502 		 */
22503 		slot = 0;
22504 		sendalot = 0;
22505 		if (error == 0)
22506 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22507 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22508 		/*
22509 		 * Get our pacing rate, if an error
22510 		 * occurred in sending (ENOBUF) we would
22511 		 * hit the else if with slot preset. Other
22512 		 * errors return.
22513 		 */
22514 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz, __LINE__);
22515 	}
22516 	/* We have sent clear the flag */
22517 	rack->r_ent_rec_ns = 0;
22518 	if (rack->r_must_retran) {
22519 		if (rsm) {
22520 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22521 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22522 				/*
22523 				 * We have retransmitted all.
22524 				 */
22525 				rack->r_must_retran = 0;
22526 				rack->r_ctl.rc_out_at_rto = 0;
22527 			}
22528 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22529 			/*
22530 			 * Sending new data will also kill
22531 			 * the loop.
22532 			 */
22533 			rack->r_must_retran = 0;
22534 			rack->r_ctl.rc_out_at_rto = 0;
22535 		}
22536 	}
22537 	rack->r_ctl.fsb.recwin = recwin;
22538 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22539 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22540 		/*
22541 		 * We hit an RTO and now have past snd_max at the RTO
22542 		 * clear all the WAS flags.
22543 		 */
22544 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22545 	}
22546 	if (slot) {
22547 		/* set the rack tcb into the slot N */
22548 		if ((error == 0) &&
22549 		    rack_use_rfo &&
22550 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22551 		    (rsm == NULL) &&
22552 		    (ipoptlen == 0) &&
22553 		    (tp->rcv_numsacks == 0) &&
22554 		    rack->r_fsb_inited &&
22555 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22556 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22557 		    (rack->r_must_retran == 0) &&
22558 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22559 		    (len > 0) && (orig_len > 0) &&
22560 		    (orig_len > len) &&
22561 		    ((orig_len - len) >= segsiz) &&
22562 		    ((optlen == 0) ||
22563 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22564 			/* We can send at least one more MSS using our fsb */
22565 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22566 					       segsiz, pace_max_seg, hw_tls, flags);
22567 		} else
22568 			rack->r_fast_output = 0;
22569 		rack_log_fsb(rack, tp, so, flags,
22570 			     ipoptlen, orig_len, len, error,
22571 			     (rsm == NULL), optlen, __LINE__, 2);
22572 	} else if (sendalot) {
22573 		int ret;
22574 
22575 		sack_rxmit = 0;
22576 		if ((error == 0) &&
22577 		    rack_use_rfo &&
22578 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22579 		    (rsm == NULL) &&
22580 		    (ipoptlen == 0) &&
22581 		    (tp->rcv_numsacks == 0) &&
22582 		    (rack->r_must_retran == 0) &&
22583 		    rack->r_fsb_inited &&
22584 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22585 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22586 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22587 		    (len > 0) && (orig_len > 0) &&
22588 		    (orig_len > len) &&
22589 		    ((orig_len - len) >= segsiz) &&
22590 		    ((optlen == 0) ||
22591 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22592 			/* we can use fast_output for more */
22593 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22594 					       segsiz, pace_max_seg, hw_tls, flags);
22595 			if (rack->r_fast_output) {
22596 				error = 0;
22597 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
22598 				if (ret >= 0)
22599 					return (ret);
22600 			        else if (error)
22601 					goto nomore;
22602 
22603 			}
22604 		}
22605 		goto again;
22606 	}
22607 skip_all_send:
22608 	/* Assure when we leave that snd_nxt will point to top */
22609 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22610 		tp->snd_nxt = tp->snd_max;
22611 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22612 #ifdef TCP_ACCOUNTING
22613 	crtsc = get_cyclecount() - ts_val;
22614 	if (tot_len_this_send) {
22615 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22616 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
22617 		}
22618 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22619 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22620 		}
22621 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22622 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22623 		}
22624 	} else {
22625 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22626 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
22627 		}
22628 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22629 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22630 		}
22631 	}
22632 	sched_unpin();
22633 #endif
22634 	if (error == ENOBUFS)
22635 		error = 0;
22636 	return (error);
22637 }
22638 
22639 static void
22640 rack_update_seg(struct tcp_rack *rack)
22641 {
22642 	uint32_t orig_val;
22643 
22644 	orig_val = rack->r_ctl.rc_pace_max_segs;
22645 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22646 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
22647 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22648 }
22649 
22650 static void
22651 rack_mtu_change(struct tcpcb *tp)
22652 {
22653 	/*
22654 	 * The MSS may have changed
22655 	 */
22656 	struct tcp_rack *rack;
22657 	struct rack_sendmap *rsm;
22658 
22659 	rack = (struct tcp_rack *)tp->t_fb_ptr;
22660 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22661 		/*
22662 		 * The MTU has changed we need to resend everything
22663 		 * since all we have sent is lost. We first fix
22664 		 * up the mtu though.
22665 		 */
22666 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
22667 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
22668 		rack_remxt_tmr(tp);
22669 		rack->r_fast_output = 0;
22670 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22671 						rack->r_ctl.rc_sacked);
22672 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22673 		rack->r_must_retran = 1;
22674 		/* Mark all inflight to needing to be rxt'd */
22675 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22676 			rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22677 		}
22678 	}
22679 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22680 	/* We don't use snd_nxt to retransmit */
22681 	tp->snd_nxt = tp->snd_max;
22682 }
22683 
22684 static int
22685 rack_set_dgp(struct tcp_rack *rack)
22686 {
22687 	if (rack->dgp_on == 1)
22688 		return(0);
22689 	if ((rack->use_fixed_rate == 1) &&
22690 	    (rack->rc_always_pace == 1)) {
22691 		/*
22692 		 * We are already pacing another
22693 		 * way.
22694 		 */
22695 		return (EBUSY);
22696 	}
22697 	if (rack->rc_always_pace == 1) {
22698 		rack_remove_pacing(rack);
22699 	}
22700 	if (tcp_incr_dgp_pacing_cnt() == 0)
22701 		return (ENOSPC);
22702 	rack->r_ctl.pacing_method |= RACK_DGP_PACING;
22703 	rack->rc_fillcw_apply_discount = 0;
22704 	rack->dgp_on = 1;
22705 	rack->rc_always_pace = 1;
22706 	rack->rc_pace_dnd = 1;
22707 	rack->use_fixed_rate = 0;
22708 	if (rack->gp_ready)
22709 		rack_set_cc_pacing(rack);
22710 	rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22711 	rack->rack_attempt_hdwr_pace = 0;
22712 	/* rxt settings */
22713 	rack->full_size_rxt = 1;
22714 	rack->shape_rxt_to_pacing_min  = 0;
22715 	/* cmpack=1 */
22716 	rack->r_use_cmp_ack = 1;
22717 	if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22718 	    rack->r_use_cmp_ack)
22719 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22720 	/* scwnd=1 */
22721 	rack->rack_enable_scwnd = 1;
22722 	/* dynamic=100 */
22723 	rack->rc_gp_dyn_mul = 1;
22724 	/* gp_inc_ca */
22725 	rack->r_ctl.rack_per_of_gp_ca = 100;
22726 	/* rrr_conf=3 */
22727 	rack->r_rr_config = 3;
22728 	/* npush=2 */
22729 	rack->r_ctl.rc_no_push_at_mrtt = 2;
22730 	/* fillcw=1 */
22731 	rack->rc_pace_to_cwnd = 1;
22732 	rack->rc_pace_fill_if_rttin_range = 0;
22733 	rack->rtt_limit_mul = 0;
22734 	/* noprr=1 */
22735 	rack->rack_no_prr = 1;
22736 	/* lscwnd=1 */
22737 	rack->r_limit_scw = 1;
22738 	/* gp_inc_rec */
22739 	rack->r_ctl.rack_per_of_gp_rec = 90;
22740 	return (0);
22741 }
22742 
22743 static int
22744 rack_set_profile(struct tcp_rack *rack, int prof)
22745 {
22746 	int err = EINVAL;
22747 	if (prof == 1) {
22748 		/*
22749 		 * Profile 1 is "standard" DGP. It ignores
22750 		 * client buffer level.
22751 		 */
22752 		err = rack_set_dgp(rack);
22753 		if (err)
22754 			return (err);
22755 	} else if (prof == 6) {
22756 		err = rack_set_dgp(rack);
22757 		if (err)
22758 			return (err);
22759 		/*
22760 		 * Profile 6 tweaks DGP so that it will apply to
22761 		 * fill-cw the same settings that profile5 does
22762 		 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
22763 		 */
22764 		rack->rc_fillcw_apply_discount = 1;
22765 	} else if (prof == 0) {
22766 		/* This changes things back to the default settings */
22767 		if (rack->rc_always_pace == 1) {
22768 			rack_remove_pacing(rack);
22769 		} else {
22770 			/* Make sure any stray flags are off */
22771 			rack->dgp_on = 0;
22772 			rack->rc_hybrid_mode = 0;
22773 			rack->use_fixed_rate = 0;
22774 		}
22775 		err = 0;
22776 		if (rack_fill_cw_state)
22777 			rack->rc_pace_to_cwnd = 1;
22778 		else
22779 			rack->rc_pace_to_cwnd = 0;
22780 
22781 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22782 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
22783 			rack->rc_always_pace = 1;
22784 			if (rack->rack_hibeta)
22785 				rack_set_cc_pacing(rack);
22786 		} else
22787 			rack->rc_always_pace = 0;
22788 		if (rack_dsack_std_based & 0x1) {
22789 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22790 			rack->rc_rack_tmr_std_based = 1;
22791 		}
22792 		if (rack_dsack_std_based & 0x2) {
22793 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
22794 			rack->rc_rack_use_dsack = 1;
22795 		}
22796 		if (rack_use_cmp_acks)
22797 			rack->r_use_cmp_ack = 1;
22798 		else
22799 			rack->r_use_cmp_ack = 0;
22800 		if (rack_disable_prr)
22801 			rack->rack_no_prr = 1;
22802 		else
22803 			rack->rack_no_prr = 0;
22804 		if (rack_gp_no_rec_chg)
22805 			rack->rc_gp_no_rec_chg = 1;
22806 		else
22807 			rack->rc_gp_no_rec_chg = 0;
22808 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22809 			rack->r_mbuf_queue = 1;
22810 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22811 				rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22812 			rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22813 		} else {
22814 			rack->r_mbuf_queue = 0;
22815 			rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22816 		}
22817 		if (rack_enable_shared_cwnd)
22818 			rack->rack_enable_scwnd = 1;
22819 		else
22820 			rack->rack_enable_scwnd = 0;
22821 		if (rack_do_dyn_mul) {
22822 			/* When dynamic adjustment is on CA needs to start at 100% */
22823 			rack->rc_gp_dyn_mul = 1;
22824 			if (rack_do_dyn_mul >= 100)
22825 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22826 		} else {
22827 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22828 			rack->rc_gp_dyn_mul = 0;
22829 		}
22830 		rack->r_rr_config = 0;
22831 		rack->r_ctl.rc_no_push_at_mrtt = 0;
22832 		rack->rc_pace_fill_if_rttin_range = 0;
22833 		rack->rtt_limit_mul = 0;
22834 
22835 		if (rack_enable_hw_pacing)
22836 			rack->rack_hdw_pace_ena = 1;
22837 		else
22838 			rack->rack_hdw_pace_ena = 0;
22839 		if (rack_disable_prr)
22840 			rack->rack_no_prr = 1;
22841 		else
22842 			rack->rack_no_prr = 0;
22843 		if (rack_limits_scwnd)
22844 			rack->r_limit_scw  = 1;
22845 		else
22846 			rack->r_limit_scw  = 0;
22847 		rack_init_retransmit_value(rack, rack_rxt_controls);
22848 		err = 0;
22849 	}
22850 	return (err);
22851 }
22852 
22853 static int
22854 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22855 {
22856 	struct deferred_opt_list *dol;
22857 
22858 	dol = malloc(sizeof(struct deferred_opt_list),
22859 		     M_TCPDO, M_NOWAIT|M_ZERO);
22860 	if (dol == NULL) {
22861 		/*
22862 		 * No space yikes -- fail out..
22863 		 */
22864 		return (0);
22865 	}
22866 	dol->optname = sopt_name;
22867 	dol->optval = loptval;
22868 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22869 	return (1);
22870 }
22871 
22872 static int
22873 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22874 {
22875 #ifdef TCP_REQUEST_TRK
22876 	struct tcp_sendfile_track *sft;
22877 	struct timeval tv;
22878 	tcp_seq seq;
22879 	int err;
22880 
22881 	microuptime(&tv);
22882 
22883 	/* Make sure no fixed rate is on */
22884 	rack->use_fixed_rate = 0;
22885 	rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22886 	rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22887 	rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22888 	/* Now allocate or find our entry that will have these settings */
22889 	sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
22890 	if (sft == NULL) {
22891 		rack->rc_tp->tcp_hybrid_error++;
22892 		/* no space, where would it have gone? */
22893 		seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22894 		rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22895 		return (ENOSPC);
22896 	}
22897 	/* mask our internal flags */
22898 	hybrid->hybrid_flags &= TCP_HYBRID_PACING_USER_MASK;
22899 	/* The seq will be snd_una + everything in the buffer */
22900 	seq = sft->start_seq;
22901 	if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22902 		/* Disabling hybrid pacing */
22903 		if (rack->rc_hybrid_mode) {
22904 			rack_set_profile(rack, 0);
22905 			rack->rc_tp->tcp_hybrid_stop++;
22906 		}
22907 		rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22908 		return (0);
22909 	}
22910 	if (rack->dgp_on == 0) {
22911 		/*
22912 		 * If we have not yet turned DGP on, do so
22913 		 * now setting pure DGP mode, no buffer level
22914 		 * response.
22915 		 */
22916 		if ((err = rack_set_profile(rack, 1)) != 0){
22917 			/* Failed to turn pacing on */
22918 			rack->rc_tp->tcp_hybrid_error++;
22919 			rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22920 			return (err);
22921 		}
22922 	}
22923 	/*
22924 	 * Now we must switch to hybrid mode as well which also
22925 	 * means moving to regular pacing.
22926 	 */
22927 	if (rack->rc_hybrid_mode == 0) {
22928 		/* First time */
22929 		if (tcp_can_enable_pacing()) {
22930 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
22931 			rack->rc_hybrid_mode = 1;
22932 		} else {
22933 			return (ENOSPC);
22934 		}
22935 		if (rack->r_ctl.pacing_method & RACK_DGP_PACING) {
22936 			/*
22937 			 * This should be true.
22938 			 */
22939 			tcp_dec_dgp_pacing_cnt();
22940 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
22941 		}
22942 	}
22943 	/* Now set in our flags */
22944 	sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
22945 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22946 		sft->cspr = hybrid->cspr;
22947 	else
22948 		sft->cspr = 0;
22949 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22950 		sft->hint_maxseg = hybrid->hint_maxseg;
22951 	else
22952 		sft->hint_maxseg = 0;
22953 	rack->rc_tp->tcp_hybrid_start++;
22954 	rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
22955 	return (0);
22956 #else
22957 	return (ENOTSUP);
22958 #endif
22959 }
22960 
22961 static int
22962 rack_stack_information(struct tcpcb *tp, struct stack_specific_info *si)
22963 {
22964 	/* We pulled a SSI info log out what was there */
22965 	si->bytes_transmitted = tp->t_sndbytes;
22966 	si->bytes_retransmitted = tp->t_snd_rxt_bytes;
22967 	return (0);
22968 }
22969 
22970 static int
22971 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
22972 		    uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
22973 
22974 {
22975 	struct epoch_tracker et;
22976 	struct sockopt sopt;
22977 	struct cc_newreno_opts opt;
22978 	uint64_t val;
22979 	int error = 0;
22980 	uint16_t ca, ss;
22981 
22982 	switch (sopt_name) {
22983 	case TCP_RACK_SET_RXT_OPTIONS:
22984 		if ((optval >= 0) && (optval <= 2)) {
22985 			rack_init_retransmit_value(rack, optval);
22986 		} else {
22987 			/*
22988 			 * You must send in 0, 1 or 2 all else is
22989 			 * invalid.
22990 			 */
22991 			error = EINVAL;
22992 		}
22993 		break;
22994 	case TCP_RACK_DSACK_OPT:
22995 		RACK_OPTS_INC(tcp_rack_dsack_opt);
22996 		if (optval & 0x1) {
22997 			rack->rc_rack_tmr_std_based = 1;
22998 		} else {
22999 			rack->rc_rack_tmr_std_based = 0;
23000 		}
23001 		if (optval & 0x2) {
23002 			rack->rc_rack_use_dsack = 1;
23003 		} else {
23004 			rack->rc_rack_use_dsack = 0;
23005 		}
23006 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
23007 		break;
23008 	case TCP_RACK_PACING_DIVISOR:
23009 		RACK_OPTS_INC(tcp_rack_pacing_divisor);
23010 		if (optval == 0) {
23011 			rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
23012 		} else {
23013 			if (optval < RL_MIN_DIVISOR)
23014 				rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
23015 			else
23016 				rack->r_ctl.pace_len_divisor = optval;
23017 		}
23018 		break;
23019 	case TCP_RACK_HI_BETA:
23020 		RACK_OPTS_INC(tcp_rack_hi_beta);
23021 		if (optval > 0) {
23022 			rack->rack_hibeta = 1;
23023 			if ((optval >= 50) &&
23024 			    (optval <= 100)) {
23025 				/*
23026 				 * User wants to set a custom beta.
23027 				 */
23028 				rack->r_ctl.saved_hibeta = optval;
23029 				if (rack->rc_pacing_cc_set)
23030 					rack_undo_cc_pacing(rack);
23031 				rack->r_ctl.rc_saved_beta.beta = optval;
23032 			}
23033 			if (rack->rc_pacing_cc_set == 0)
23034 				rack_set_cc_pacing(rack);
23035 		} else {
23036 			rack->rack_hibeta = 0;
23037 			if (rack->rc_pacing_cc_set)
23038 				rack_undo_cc_pacing(rack);
23039 		}
23040 		break;
23041 	case TCP_RACK_PACING_BETA:
23042 		error = EINVAL;
23043 		break;
23044 	case TCP_RACK_TIMER_SLOP:
23045 		RACK_OPTS_INC(tcp_rack_timer_slop);
23046 		rack->r_ctl.timer_slop = optval;
23047 		if (rack->rc_tp->t_srtt) {
23048 			/*
23049 			 * If we have an SRTT lets update t_rxtcur
23050 			 * to have the new slop.
23051 			 */
23052 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
23053 					   rack_rto_min, rack_rto_max,
23054 					   rack->r_ctl.timer_slop);
23055 		}
23056 		break;
23057 	case TCP_RACK_PACING_BETA_ECN:
23058 		RACK_OPTS_INC(tcp_rack_beta_ecn);
23059 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23060 			/* This only works for newreno. */
23061 			error = EINVAL;
23062 			break;
23063 		}
23064 		if (rack->rc_pacing_cc_set) {
23065 			/*
23066 			 * Set them into the real CC module
23067 			 * whats in the rack pcb is the old values
23068 			 * to be used on restoral/
23069 			 */
23070 			sopt.sopt_dir = SOPT_SET;
23071 			opt.name = CC_NEWRENO_BETA_ECN;
23072 			opt.val = optval;
23073 			if (CC_ALGO(tp)->ctl_output != NULL)
23074 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23075 			else
23076 				error = ENOENT;
23077 		} else {
23078 			/*
23079 			 * Not pacing yet so set it into our local
23080 			 * rack pcb storage.
23081 			 */
23082 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
23083 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
23084 		}
23085 		break;
23086 	case TCP_DEFER_OPTIONS:
23087 		RACK_OPTS_INC(tcp_defer_opt);
23088 		if (optval) {
23089 			if (rack->gp_ready) {
23090 				/* Too late */
23091 				error = EINVAL;
23092 				break;
23093 			}
23094 			rack->defer_options = 1;
23095 		} else
23096 			rack->defer_options = 0;
23097 		break;
23098 	case TCP_RACK_MEASURE_CNT:
23099 		RACK_OPTS_INC(tcp_rack_measure_cnt);
23100 		if (optval && (optval <= 0xff)) {
23101 			rack->r_ctl.req_measurements = optval;
23102 		} else
23103 			error = EINVAL;
23104 		break;
23105 	case TCP_REC_ABC_VAL:
23106 		RACK_OPTS_INC(tcp_rec_abc_val);
23107 		if (optval > 0)
23108 			rack->r_use_labc_for_rec = 1;
23109 		else
23110 			rack->r_use_labc_for_rec = 0;
23111 		break;
23112 	case TCP_RACK_ABC_VAL:
23113 		RACK_OPTS_INC(tcp_rack_abc_val);
23114 		if ((optval > 0) && (optval < 255))
23115 			rack->rc_labc = optval;
23116 		else
23117 			error = EINVAL;
23118 		break;
23119 	case TCP_HDWR_UP_ONLY:
23120 		RACK_OPTS_INC(tcp_pacing_up_only);
23121 		if (optval)
23122 			rack->r_up_only = 1;
23123 		else
23124 			rack->r_up_only = 0;
23125 		break;
23126 	case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
23127 		RACK_OPTS_INC(tcp_fillcw_rate_cap);
23128 		rack->r_ctl.fillcw_cap = loptval;
23129 		break;
23130 	case TCP_PACING_RATE_CAP:
23131 		RACK_OPTS_INC(tcp_pacing_rate_cap);
23132 		if ((rack->dgp_on == 1) &&
23133 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23134 			/*
23135 			 * If we are doing DGP we need to switch
23136 			 * to using the pacing limit.
23137 			 */
23138 			if (tcp_can_enable_pacing() == 0) {
23139 				error = ENOSPC;
23140 				break;
23141 			}
23142 			/*
23143 			 * Now change up the flags and counts to be correct.
23144 			 */
23145 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
23146 			tcp_dec_dgp_pacing_cnt();
23147 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23148 		}
23149 		rack->r_ctl.bw_rate_cap = loptval;
23150 		break;
23151 	case TCP_HYBRID_PACING:
23152 		if (hybrid == NULL) {
23153 			error = EINVAL;
23154 			break;
23155 		}
23156 		if (rack->r_ctl.side_chan_dis_mask & HYBRID_DIS_MASK) {
23157 			error = EPERM;
23158 			break;
23159 		}
23160 		error = process_hybrid_pacing(rack, hybrid);
23161 		break;
23162 	case TCP_SIDECHAN_DIS:			/*  URL:scodm */
23163 		if (optval)
23164 			rack->r_ctl.side_chan_dis_mask = optval;
23165 		else
23166 			rack->r_ctl.side_chan_dis_mask = 0;
23167 		break;
23168 	case TCP_RACK_PROFILE:
23169 		RACK_OPTS_INC(tcp_profile);
23170 		error = rack_set_profile(rack, optval);
23171 		break;
23172 	case TCP_USE_CMP_ACKS:
23173 		RACK_OPTS_INC(tcp_use_cmp_acks);
23174 		if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
23175 			/* You can't turn it off once its on! */
23176 			error = EINVAL;
23177 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
23178 			rack->r_use_cmp_ack = 1;
23179 			rack->r_mbuf_queue = 1;
23180 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23181 		}
23182 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
23183 			tp->t_flags2 |= TF2_MBUF_ACKCMP;
23184 		break;
23185 	case TCP_SHARED_CWND_TIME_LIMIT:
23186 		RACK_OPTS_INC(tcp_lscwnd);
23187 		if (optval)
23188 			rack->r_limit_scw = 1;
23189 		else
23190 			rack->r_limit_scw = 0;
23191 		break;
23192 	case TCP_RACK_DGP_IN_REC:
23193 		error = EINVAL;
23194 		break;
23195  	case TCP_RACK_PACE_TO_FILL:
23196 		RACK_OPTS_INC(tcp_fillcw);
23197 		if (optval == 0)
23198 			rack->rc_pace_to_cwnd = 0;
23199 		else {
23200 			rack->rc_pace_to_cwnd = 1;
23201 		}
23202 		if ((optval >= rack_gp_rtt_maxmul) &&
23203 		    rack_gp_rtt_maxmul &&
23204 		    (optval < 0xf)) {
23205 			rack->rc_pace_fill_if_rttin_range = 1;
23206 			rack->rtt_limit_mul = optval;
23207 		} else {
23208 			rack->rc_pace_fill_if_rttin_range = 0;
23209 			rack->rtt_limit_mul = 0;
23210 		}
23211 		break;
23212 	case TCP_RACK_NO_PUSH_AT_MAX:
23213 		RACK_OPTS_INC(tcp_npush);
23214 		if (optval == 0)
23215 			rack->r_ctl.rc_no_push_at_mrtt = 0;
23216 		else if (optval < 0xff)
23217 			rack->r_ctl.rc_no_push_at_mrtt = optval;
23218 		else
23219 			error = EINVAL;
23220 		break;
23221 	case TCP_SHARED_CWND_ENABLE:
23222 		RACK_OPTS_INC(tcp_rack_scwnd);
23223 		if (optval == 0)
23224 			rack->rack_enable_scwnd = 0;
23225 		else
23226 			rack->rack_enable_scwnd = 1;
23227 		break;
23228 	case TCP_RACK_MBUF_QUEUE:
23229 		/* Now do we use the LRO mbuf-queue feature */
23230 		RACK_OPTS_INC(tcp_rack_mbufq);
23231 		if (optval || rack->r_use_cmp_ack)
23232 			rack->r_mbuf_queue = 1;
23233 		else
23234 			rack->r_mbuf_queue = 0;
23235 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23236 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23237 		else
23238 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23239 		break;
23240 	case TCP_RACK_NONRXT_CFG_RATE:
23241 		RACK_OPTS_INC(tcp_rack_cfg_rate);
23242 		if (optval == 0)
23243 			rack->rack_rec_nonrxt_use_cr = 0;
23244 		else
23245 			rack->rack_rec_nonrxt_use_cr = 1;
23246 		break;
23247 	case TCP_NO_PRR:
23248 		RACK_OPTS_INC(tcp_rack_noprr);
23249 		if (optval == 0)
23250 			rack->rack_no_prr = 0;
23251 		else if (optval == 1)
23252 			rack->rack_no_prr = 1;
23253 		else if (optval == 2)
23254 			rack->no_prr_addback = 1;
23255 		else
23256 			error = EINVAL;
23257 		break;
23258 	case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
23259 		if (optval > 0)
23260 			rack->cspr_is_fcc = 1;
23261 		else
23262 			rack->cspr_is_fcc = 0;
23263 		break;
23264 	case TCP_TIMELY_DYN_ADJ:
23265 		RACK_OPTS_INC(tcp_timely_dyn);
23266 		if (optval == 0)
23267 			rack->rc_gp_dyn_mul = 0;
23268 		else {
23269 			rack->rc_gp_dyn_mul = 1;
23270 			if (optval >= 100) {
23271 				/*
23272 				 * If the user sets something 100 or more
23273 				 * its the gp_ca value.
23274 				 */
23275 				rack->r_ctl.rack_per_of_gp_ca  = optval;
23276 			}
23277 		}
23278 		break;
23279 	case TCP_RACK_DO_DETECTION:
23280 		error = EINVAL;
23281 		break;
23282 	case TCP_RACK_TLP_USE:
23283 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
23284 			error = EINVAL;
23285 			break;
23286 		}
23287 		RACK_OPTS_INC(tcp_tlp_use);
23288 		rack->rack_tlp_threshold_use = optval;
23289 		break;
23290 	case TCP_RACK_TLP_REDUCE:
23291 		/* RACK TLP cwnd reduction (bool) */
23292 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
23293 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
23294 		break;
23295 		/*  Pacing related ones */
23296 	case TCP_RACK_PACE_ALWAYS:
23297 		/*
23298 		 * zero is old rack method, 1 is new
23299 		 * method using a pacing rate.
23300 		 */
23301 		RACK_OPTS_INC(tcp_rack_pace_always);
23302 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23303 			error = EPERM;
23304 			break;
23305 		}
23306 		if (optval > 0) {
23307 			if (rack->rc_always_pace) {
23308 				error = EALREADY;
23309 				break;
23310 			} else if (tcp_can_enable_pacing()) {
23311 				rack->r_ctl.pacing_method |= RACK_REG_PACING;
23312 				rack->rc_always_pace = 1;
23313 				if (rack->rack_hibeta)
23314 					rack_set_cc_pacing(rack);
23315 			}
23316 			else {
23317 				error = ENOSPC;
23318 				break;
23319 			}
23320 		} else {
23321 			if (rack->rc_always_pace == 1) {
23322 				rack_remove_pacing(rack);
23323 			}
23324 		}
23325 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23326 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23327 		else
23328 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23329 		/* A rate may be set irate or other, if so set seg size */
23330 		rack_update_seg(rack);
23331 		break;
23332 	case TCP_BBR_RACK_INIT_RATE:
23333 		RACK_OPTS_INC(tcp_initial_rate);
23334 		val = optval;
23335 		/* Change from kbits per second to bytes per second */
23336 		val *= 1000;
23337 		val /= 8;
23338 		rack->r_ctl.init_rate = val;
23339 		if (rack->rc_always_pace)
23340 			rack_update_seg(rack);
23341 		break;
23342 	case TCP_BBR_IWINTSO:
23343 		error = EINVAL;
23344 		break;
23345 	case TCP_RACK_FORCE_MSEG:
23346 		RACK_OPTS_INC(tcp_rack_force_max_seg);
23347 		if (optval)
23348 			rack->rc_force_max_seg = 1;
23349 		else
23350 			rack->rc_force_max_seg = 0;
23351 		break;
23352 	case TCP_RACK_PACE_MIN_SEG:
23353 		RACK_OPTS_INC(tcp_rack_min_seg);
23354 		rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23355 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23356 		break;
23357 	case TCP_RACK_PACE_MAX_SEG:
23358 		/* Max segments size in a pace in bytes */
23359 		RACK_OPTS_INC(tcp_rack_max_seg);
23360 		if ((rack->dgp_on == 1) &&
23361 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23362 			/*
23363 			 * If we set a max-seg and are doing DGP then
23364 			 * we now fall under the pacing limits not the
23365 			 * DGP ones.
23366 			 */
23367 			if (tcp_can_enable_pacing() == 0) {
23368 				error = ENOSPC;
23369 				break;
23370 			}
23371 			/*
23372 			 * Now change up the flags and counts to be correct.
23373 			 */
23374 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
23375 			tcp_dec_dgp_pacing_cnt();
23376 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23377 		}
23378 		if (optval <= MAX_USER_SET_SEG)
23379 			rack->rc_user_set_max_segs = optval;
23380 		else
23381 			rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
23382 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23383 		break;
23384 	case TCP_RACK_PACE_RATE_REC:
23385 		/* Set the fixed pacing rate in Bytes per second ca */
23386 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23387 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23388 			error = EPERM;
23389 			break;
23390 		}
23391 		if (rack->dgp_on) {
23392 			/*
23393 			 * We are already pacing another
23394 			 * way.
23395 			 */
23396 			error = EBUSY;
23397 			break;
23398 		}
23399 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23400 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23401 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23402 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23403 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23404 		rack->use_fixed_rate = 1;
23405 		if (rack->rack_hibeta)
23406 			rack_set_cc_pacing(rack);
23407 		rack_log_pacing_delay_calc(rack,
23408 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23409 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23410 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23411 					   __LINE__, NULL,0);
23412 		break;
23413 
23414 	case TCP_RACK_PACE_RATE_SS:
23415 		/* Set the fixed pacing rate in Bytes per second ca */
23416 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23417 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23418 			error = EPERM;
23419 			break;
23420 		}
23421 		if (rack->dgp_on) {
23422 			/*
23423 			 * We are already pacing another
23424 			 * way.
23425 			 */
23426 			error = EBUSY;
23427 			break;
23428 		}
23429 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23430 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23431 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23432 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23433 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23434 		rack->use_fixed_rate = 1;
23435 		if (rack->rack_hibeta)
23436 			rack_set_cc_pacing(rack);
23437 		rack_log_pacing_delay_calc(rack,
23438 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23439 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23440 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23441 					   __LINE__, NULL, 0);
23442 		break;
23443 
23444 	case TCP_RACK_PACE_RATE_CA:
23445 		/* Set the fixed pacing rate in Bytes per second ca */
23446 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23447 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23448 			error = EPERM;
23449 			break;
23450 		}
23451 		if (rack->dgp_on) {
23452 			/*
23453 			 * We are already pacing another
23454 			 * way.
23455 			 */
23456 			error = EBUSY;
23457 			break;
23458 		}
23459 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23460 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23461 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23462 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23463 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23464 		rack->use_fixed_rate = 1;
23465 		if (rack->rack_hibeta)
23466 			rack_set_cc_pacing(rack);
23467 		rack_log_pacing_delay_calc(rack,
23468 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23469 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23470 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23471 					   __LINE__, NULL, 0);
23472 		break;
23473 	case TCP_RACK_GP_INCREASE_REC:
23474 		RACK_OPTS_INC(tcp_gp_inc_rec);
23475 		rack->r_ctl.rack_per_of_gp_rec = optval;
23476 		rack_log_pacing_delay_calc(rack,
23477 					   rack->r_ctl.rack_per_of_gp_ss,
23478 					   rack->r_ctl.rack_per_of_gp_ca,
23479 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23480 					   __LINE__, NULL, 0);
23481 		break;
23482 	case TCP_RACK_GP_INCREASE_CA:
23483 		RACK_OPTS_INC(tcp_gp_inc_ca);
23484 		ca = optval;
23485 		if (ca < 100) {
23486 			/*
23487 			 * We don't allow any reduction
23488 			 * over the GP b/w.
23489 			 */
23490 			error = EINVAL;
23491 			break;
23492 		}
23493 		rack->r_ctl.rack_per_of_gp_ca = ca;
23494 		rack_log_pacing_delay_calc(rack,
23495 					   rack->r_ctl.rack_per_of_gp_ss,
23496 					   rack->r_ctl.rack_per_of_gp_ca,
23497 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23498 					   __LINE__, NULL, 0);
23499 		break;
23500 	case TCP_RACK_GP_INCREASE_SS:
23501 		RACK_OPTS_INC(tcp_gp_inc_ss);
23502 		ss = optval;
23503 		if (ss < 100) {
23504 			/*
23505 			 * We don't allow any reduction
23506 			 * over the GP b/w.
23507 			 */
23508 			error = EINVAL;
23509 			break;
23510 		}
23511 		rack->r_ctl.rack_per_of_gp_ss = ss;
23512 		rack_log_pacing_delay_calc(rack,
23513 					   rack->r_ctl.rack_per_of_gp_ss,
23514 					   rack->r_ctl.rack_per_of_gp_ca,
23515 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23516 					   __LINE__, NULL, 0);
23517 		break;
23518 	case TCP_RACK_RR_CONF:
23519 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23520 		if (optval && optval <= 3)
23521 			rack->r_rr_config = optval;
23522 		else
23523 			rack->r_rr_config = 0;
23524 		break;
23525 	case TCP_PACING_DND:			/*  URL:dnd */
23526 		if (optval > 0)
23527 			rack->rc_pace_dnd = 1;
23528 		else
23529 			rack->rc_pace_dnd = 0;
23530 		break;
23531 	case TCP_HDWR_RATE_CAP:
23532 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
23533 		if (optval) {
23534 			if (rack->r_rack_hw_rate_caps == 0)
23535 				rack->r_rack_hw_rate_caps = 1;
23536 			else
23537 				error = EALREADY;
23538 		} else {
23539 			rack->r_rack_hw_rate_caps = 0;
23540 		}
23541 		break;
23542 	case TCP_DGP_UPPER_BOUNDS:
23543 	{
23544 		uint8_t val;
23545 		val = optval & 0x0000ff;
23546 		rack->r_ctl.rack_per_upper_bound_ca = val;
23547 		val = (optval >> 16) & 0x0000ff;
23548 		rack->r_ctl.rack_per_upper_bound_ss = val;
23549 		break;
23550 	}
23551 	case TCP_SS_EEXIT:			/*  URL:eexit */
23552 		if (optval > 0) {
23553 			rack->r_ctl.gp_rnd_thresh =  optval & 0x0ff;
23554 			if (optval & 0x10000) {
23555 				rack->r_ctl.gate_to_fs = 1;
23556 			} else {
23557 				rack->r_ctl.gate_to_fs = 0;
23558 			}
23559 			if (optval & 0x20000) {
23560 				rack->r_ctl.use_gp_not_last = 1;
23561 			} else {
23562 				rack->r_ctl.use_gp_not_last = 0;
23563 			}
23564 			if (optval & 0xfffc0000) {
23565 				uint32_t v;
23566 
23567 				v = (optval >> 18) & 0x00003fff;
23568 				if (v >= 1000)
23569 					rack->r_ctl.gp_gain_req = v;
23570 			}
23571 		} else {
23572 			/* We do not do ss early exit at all */
23573 			rack->rc_initial_ss_comp = 1;
23574 			rack->r_ctl.gp_rnd_thresh = 0;
23575 		}
23576 		break;
23577 	case TCP_RACK_SPLIT_LIMIT:
23578 		RACK_OPTS_INC(tcp_split_limit);
23579 		rack->r_ctl.rc_split_limit = optval;
23580 		break;
23581 	case TCP_BBR_HDWR_PACE:
23582 		RACK_OPTS_INC(tcp_hdwr_pacing);
23583 		if (optval){
23584 			if (rack->rack_hdrw_pacing == 0) {
23585 				rack->rack_hdw_pace_ena = 1;
23586 				rack->rack_attempt_hdwr_pace = 0;
23587 			} else
23588 				error = EALREADY;
23589 		} else {
23590 			rack->rack_hdw_pace_ena = 0;
23591 #ifdef RATELIMIT
23592 			if (rack->r_ctl.crte != NULL) {
23593 				rack->rack_hdrw_pacing = 0;
23594 				rack->rack_attempt_hdwr_pace = 0;
23595 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23596 				rack->r_ctl.crte = NULL;
23597 			}
23598 #endif
23599 		}
23600 		break;
23601 		/*  End Pacing related ones */
23602 	case TCP_RACK_PRR_SENDALOT:
23603 		/* Allow PRR to send more than one seg */
23604 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
23605 		rack->r_ctl.rc_prr_sendalot = optval;
23606 		break;
23607 	case TCP_RACK_MIN_TO:
23608 		/* Minimum time between rack t-o's in ms */
23609 		RACK_OPTS_INC(tcp_rack_min_to);
23610 		rack->r_ctl.rc_min_to = optval;
23611 		break;
23612 	case TCP_RACK_EARLY_SEG:
23613 		/* If early recovery max segments */
23614 		RACK_OPTS_INC(tcp_rack_early_seg);
23615 		rack->r_ctl.rc_early_recovery_segs = optval;
23616 		break;
23617 	case TCP_RACK_ENABLE_HYSTART:
23618 	{
23619 		if (optval) {
23620 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23621 			if (rack_do_hystart > RACK_HYSTART_ON)
23622 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23623 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23624 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23625 		} else {
23626 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23627 		}
23628 	}
23629 	break;
23630 	case TCP_RACK_REORD_THRESH:
23631 		/* RACK reorder threshold (shift amount) */
23632 		RACK_OPTS_INC(tcp_rack_reord_thresh);
23633 		if ((optval > 0) && (optval < 31))
23634 			rack->r_ctl.rc_reorder_shift = optval;
23635 		else
23636 			error = EINVAL;
23637 		break;
23638 	case TCP_RACK_REORD_FADE:
23639 		/* Does reordering fade after ms time */
23640 		RACK_OPTS_INC(tcp_rack_reord_fade);
23641 		rack->r_ctl.rc_reorder_fade = optval;
23642 		break;
23643 	case TCP_RACK_TLP_THRESH:
23644 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
23645 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
23646 		if (optval)
23647 			rack->r_ctl.rc_tlp_threshold = optval;
23648 		else
23649 			error = EINVAL;
23650 		break;
23651 	case TCP_BBR_USE_RACK_RR:
23652 		RACK_OPTS_INC(tcp_rack_rr);
23653 		if (optval)
23654 			rack->use_rack_rr = 1;
23655 		else
23656 			rack->use_rack_rr = 0;
23657 		break;
23658 	case TCP_RACK_PKT_DELAY:
23659 		/* RACK added ms i.e. rack-rtt + reord + N */
23660 		RACK_OPTS_INC(tcp_rack_pkt_delay);
23661 		rack->r_ctl.rc_pkt_delay = optval;
23662 		break;
23663 	case TCP_DELACK:
23664 		RACK_OPTS_INC(tcp_rack_delayed_ack);
23665 		if (optval == 0)
23666 			tp->t_delayed_ack = 0;
23667 		else
23668 			tp->t_delayed_ack = 1;
23669 		if (tp->t_flags & TF_DELACK) {
23670 			tp->t_flags &= ~TF_DELACK;
23671 			tp->t_flags |= TF_ACKNOW;
23672 			NET_EPOCH_ENTER(et);
23673 			rack_output(tp);
23674 			NET_EPOCH_EXIT(et);
23675 		}
23676 		break;
23677 
23678 	case TCP_BBR_RACK_RTT_USE:
23679 		RACK_OPTS_INC(tcp_rack_rtt_use);
23680 		if ((optval != USE_RTT_HIGH) &&
23681 		    (optval != USE_RTT_LOW) &&
23682 		    (optval != USE_RTT_AVG))
23683 			error = EINVAL;
23684 		else
23685 			rack->r_ctl.rc_rate_sample_method = optval;
23686 		break;
23687 	case TCP_HONOR_HPTS_MIN:
23688 		RACK_OPTS_INC(tcp_honor_hpts);
23689 		if (optval) {
23690 			rack->r_use_hpts_min = 1;
23691 			/*
23692 			 * Must be between 2 - 80% to be a reduction else
23693 			 * we keep the default (10%).
23694 			 */
23695 			if ((optval > 1) && (optval <= 80)) {
23696 				rack->r_ctl.max_reduction = optval;
23697 			}
23698 		} else
23699 			rack->r_use_hpts_min = 0;
23700 		break;
23701 	case TCP_REC_IS_DYN:			/*  URL:dynrec */
23702 		RACK_OPTS_INC(tcp_dyn_rec);
23703 		if (optval)
23704 			rack->rc_gp_no_rec_chg = 1;
23705 		else
23706 			rack->rc_gp_no_rec_chg = 0;
23707 		break;
23708 	case TCP_NO_TIMELY:
23709 		RACK_OPTS_INC(tcp_notimely);
23710 		if (optval) {
23711 			rack->rc_skip_timely = 1;
23712 			rack->r_ctl.rack_per_of_gp_rec = 90;
23713 			rack->r_ctl.rack_per_of_gp_ca = 100;
23714 			rack->r_ctl.rack_per_of_gp_ss = 250;
23715 		} else {
23716 			rack->rc_skip_timely = 0;
23717 		}
23718 		break;
23719 	case TCP_GP_USE_LTBW:
23720 		if (optval == 0) {
23721 			rack->use_lesser_lt_bw = 0;
23722 			rack->dis_lt_bw = 1;
23723 		} else if (optval == 1) {
23724 			rack->use_lesser_lt_bw = 1;
23725 			rack->dis_lt_bw = 0;
23726 		} else if (optval == 2) {
23727 			rack->use_lesser_lt_bw = 0;
23728 			rack->dis_lt_bw = 0;
23729 		}
23730 		break;
23731 	case TCP_DATA_AFTER_CLOSE:
23732 		RACK_OPTS_INC(tcp_data_after_close);
23733 		if (optval)
23734 			rack->rc_allow_data_af_clo = 1;
23735 		else
23736 			rack->rc_allow_data_af_clo = 0;
23737 		break;
23738 	default:
23739 		break;
23740 	}
23741 	tcp_log_socket_option(tp, sopt_name, optval, error);
23742 	return (error);
23743 }
23744 
23745 static void
23746 rack_inherit(struct tcpcb *tp, struct inpcb *parent)
23747 {
23748 	/*
23749 	 * A new connection has been created (tp) and
23750 	 * the parent is the inpcb given. We want to
23751 	 * apply a read-lock to the parent (we are already
23752 	 * holding a write lock on the tp) and copy anything
23753 	 * out of the rack specific data as long as its tfb is
23754 	 * the same as ours i.e. we are the same stack. Otherwise
23755 	 * we just return.
23756 	 */
23757 	struct tcpcb *par;
23758 	struct tcp_rack *dest, *src;
23759 	int cnt = 0;
23760 
23761 	par = intotcpcb(parent);
23762 	if (par->t_fb != tp->t_fb) {
23763 		/* Not the same stack */
23764 		tcp_log_socket_option(tp, 0, 0, 1);
23765 		return;
23766 	}
23767 	/* Ok if we reach here lets setup the two rack pointers */
23768 	dest = (struct tcp_rack *)tp->t_fb_ptr;
23769 	src = (struct tcp_rack *)par->t_fb_ptr;
23770 	if ((src == NULL) || (dest == NULL)) {
23771 		/* Huh? */
23772 		tcp_log_socket_option(tp, 0, 0, 2);
23773 		return;
23774 	}
23775 	/* Now copy out anything we wish to inherit i.e. things in socket-options */
23776 	/* TCP_RACK_PROFILE we can't know but we can set DGP if its on */
23777 	if ((src->dgp_on) && (dest->dgp_on == 0)) {
23778 		/* Profile 1 had to be set via sock opt */
23779 		rack_set_dgp(dest);
23780 		cnt++;
23781 	}
23782 	/* TCP_RACK_SET_RXT_OPTIONS */
23783 	if (dest->full_size_rxt != src->full_size_rxt) {
23784 		dest->full_size_rxt = src->full_size_rxt;
23785 		cnt++;
23786 	}
23787 	if (dest->shape_rxt_to_pacing_min  != src->shape_rxt_to_pacing_min) {
23788 		dest->shape_rxt_to_pacing_min = src->shape_rxt_to_pacing_min;
23789 		cnt++;
23790 	}
23791 	/* TCP_RACK_DSACK_OPT */
23792 	if (dest->rc_rack_tmr_std_based != src->rc_rack_tmr_std_based) {
23793 		dest->rc_rack_tmr_std_based = src->rc_rack_tmr_std_based;
23794 		cnt++;
23795 	}
23796 	if (dest->rc_rack_use_dsack != src->rc_rack_use_dsack) {
23797 		dest->rc_rack_use_dsack = src->rc_rack_use_dsack;
23798 		cnt++;
23799 	}
23800 	/* TCP_RACK_PACING_DIVISOR */
23801 	if (dest->r_ctl.pace_len_divisor != src->r_ctl.pace_len_divisor) {
23802 		dest->r_ctl.pace_len_divisor = src->r_ctl.pace_len_divisor;
23803 		cnt++;
23804 	}
23805 	/* TCP_RACK_HI_BETA */
23806 	if (src->rack_hibeta != dest->rack_hibeta) {
23807 		cnt++;
23808 		if (src->rack_hibeta) {
23809 			dest->r_ctl.rc_saved_beta.beta = src->r_ctl.rc_saved_beta.beta;
23810 			dest->rack_hibeta = 1;
23811 		} else {
23812 			dest->rack_hibeta = 0;
23813 		}
23814 	}
23815 	/* TCP_RACK_TIMER_SLOP */
23816 	if (dest->r_ctl.timer_slop != src->r_ctl.timer_slop) {
23817 		dest->r_ctl.timer_slop = src->r_ctl.timer_slop;
23818 		cnt++;
23819 	}
23820 	/* TCP_RACK_PACING_BETA_ECN */
23821 	if (dest->r_ctl.rc_saved_beta.beta_ecn != src->r_ctl.rc_saved_beta.beta_ecn) {
23822 		dest->r_ctl.rc_saved_beta.beta_ecn = src->r_ctl.rc_saved_beta.beta_ecn;
23823 		cnt++;
23824 	}
23825 	if (dest->r_ctl.rc_saved_beta.newreno_flags != src->r_ctl.rc_saved_beta.newreno_flags) {
23826 		dest->r_ctl.rc_saved_beta.newreno_flags = src->r_ctl.rc_saved_beta.newreno_flags;
23827 		cnt++;
23828 	}
23829 	/* We do not do TCP_DEFER_OPTIONS */
23830 	/* TCP_RACK_MEASURE_CNT */
23831 	if (dest->r_ctl.req_measurements != src->r_ctl.req_measurements) {
23832 		dest->r_ctl.req_measurements = src->r_ctl.req_measurements;
23833 		cnt++;
23834 	}
23835 	/* TCP_HDWR_UP_ONLY */
23836 	if (dest->r_up_only != src->r_up_only) {
23837 		dest->r_up_only = src->r_up_only;
23838 		cnt++;
23839 	}
23840 	/* TCP_FILLCW_RATE_CAP */
23841 	if (dest->r_ctl.fillcw_cap != src->r_ctl.fillcw_cap) {
23842 		dest->r_ctl.fillcw_cap = src->r_ctl.fillcw_cap;
23843 		cnt++;
23844 	}
23845 	/* TCP_PACING_RATE_CAP */
23846 	if (dest->r_ctl.bw_rate_cap != src->r_ctl.bw_rate_cap) {
23847 		dest->r_ctl.bw_rate_cap = src->r_ctl.bw_rate_cap;
23848 		cnt++;
23849 	}
23850 	/* A listener can't set TCP_HYBRID_PACING */
23851 	/* TCP_SIDECHAN_DIS */
23852 	if (dest->r_ctl.side_chan_dis_mask != src->r_ctl.side_chan_dis_mask) {
23853 		dest->r_ctl.side_chan_dis_mask = src->r_ctl.side_chan_dis_mask;
23854 		cnt++;
23855 	}
23856 	/* TCP_SHARED_CWND_TIME_LIMIT */
23857 	if (dest->r_limit_scw != src->r_limit_scw) {
23858 		dest->r_limit_scw = src->r_limit_scw;
23859 		cnt++;
23860 	}
23861 	/* TCP_RACK_PACE_TO_FILL */
23862 	if (dest->rc_pace_to_cwnd != src->rc_pace_to_cwnd) {
23863 		dest->rc_pace_to_cwnd = src->rc_pace_to_cwnd;
23864 		cnt++;
23865 	}
23866 	if (dest->rc_pace_fill_if_rttin_range != src->rc_pace_fill_if_rttin_range) {
23867 		dest->rc_pace_fill_if_rttin_range = src->rc_pace_fill_if_rttin_range;
23868 		cnt++;
23869 	}
23870 	if (dest->rtt_limit_mul != src->rtt_limit_mul) {
23871 		dest->rtt_limit_mul = src->rtt_limit_mul;
23872 		cnt++;
23873 	}
23874 	/* TCP_RACK_NO_PUSH_AT_MAX */
23875 	if (dest->r_ctl.rc_no_push_at_mrtt != src->r_ctl.rc_no_push_at_mrtt) {
23876 		dest->r_ctl.rc_no_push_at_mrtt = src->r_ctl.rc_no_push_at_mrtt;
23877 		cnt++;
23878 	}
23879 	/* TCP_SHARED_CWND_ENABLE */
23880 	if (dest->rack_enable_scwnd != src->rack_enable_scwnd) {
23881 		dest->rack_enable_scwnd = src->rack_enable_scwnd;
23882 		cnt++;
23883 	}
23884 	/* TCP_USE_CMP_ACKS */
23885 	if (dest->r_use_cmp_ack != src->r_use_cmp_ack) {
23886 		dest->r_use_cmp_ack = src->r_use_cmp_ack;
23887 		cnt++;
23888 	}
23889 
23890 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23891 		dest->r_mbuf_queue = src->r_mbuf_queue;
23892 		cnt++;
23893 	}
23894 	/* TCP_RACK_MBUF_QUEUE */
23895 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23896 		dest->r_mbuf_queue = src->r_mbuf_queue;
23897 		cnt++;
23898 	}
23899 	if  (dest->r_mbuf_queue || dest->rc_always_pace || dest->r_use_cmp_ack) {
23900 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23901 	} else {
23902 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23903 	}
23904 	if (dest->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) {
23905 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
23906 	}
23907 	/* TCP_RACK_NONRXT_CFG_RATE */
23908 	if (dest->rack_rec_nonrxt_use_cr != src->rack_rec_nonrxt_use_cr) {
23909 		dest->rack_rec_nonrxt_use_cr = src->rack_rec_nonrxt_use_cr;
23910 		cnt++;
23911 	}
23912 	/* TCP_NO_PRR */
23913 	if (dest->rack_no_prr != src->rack_no_prr) {
23914 		dest->rack_no_prr = src->rack_no_prr;
23915 		cnt++;
23916 	}
23917 	if (dest->no_prr_addback != src->no_prr_addback) {
23918 		dest->no_prr_addback = src->no_prr_addback;
23919 		cnt++;
23920 	}
23921 	/* RACK_CSPR_IS_FCC */
23922 	if (dest->cspr_is_fcc != src->cspr_is_fcc) {
23923 		dest->cspr_is_fcc = src->cspr_is_fcc;
23924 		cnt++;
23925 	}
23926 	/* TCP_TIMELY_DYN_ADJ */
23927 	if (dest->rc_gp_dyn_mul != src->rc_gp_dyn_mul) {
23928 		dest->rc_gp_dyn_mul = src->rc_gp_dyn_mul;
23929 		cnt++;
23930 	}
23931 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23932 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23933 		cnt++;
23934 	}
23935 	/* TCP_RACK_TLP_USE */
23936 	if (dest->rack_tlp_threshold_use != src->rack_tlp_threshold_use) {
23937 		dest->rack_tlp_threshold_use = src->rack_tlp_threshold_use;
23938 		cnt++;
23939 	}
23940 	/* we don't allow inheritence of TCP_RACK_PACE_ALWAYS */
23941 	/* TCP_BBR_RACK_INIT_RATE */
23942 	if (dest->r_ctl.init_rate != src->r_ctl.init_rate) {
23943 		dest->r_ctl.init_rate = src->r_ctl.init_rate;
23944 		cnt++;
23945 	}
23946 	/* TCP_RACK_FORCE_MSEG */
23947 	if (dest->rc_force_max_seg != src->rc_force_max_seg) {
23948 		dest->rc_force_max_seg = src->rc_force_max_seg;
23949 		cnt++;
23950 	}
23951 	/* TCP_RACK_PACE_MIN_SEG */
23952 	if (dest->r_ctl.rc_user_set_min_segs != src->r_ctl.rc_user_set_min_segs) {
23953 		dest->r_ctl.rc_user_set_min_segs = src->r_ctl.rc_user_set_min_segs;
23954 		cnt++;
23955 	}
23956 	/* we don't allow TCP_RACK_PACE_MAX_SEG */
23957 	/* TCP_RACK_PACE_RATE_REC, TCP_RACK_PACE_RATE_SS,  TCP_RACK_PACE_RATE_CA */
23958 	if (dest->r_ctl.rc_fixed_pacing_rate_ca != src->r_ctl.rc_fixed_pacing_rate_ca) {
23959 		dest->r_ctl.rc_fixed_pacing_rate_ca = src->r_ctl.rc_fixed_pacing_rate_ca;
23960 		cnt++;
23961 	}
23962 	if (dest->r_ctl.rc_fixed_pacing_rate_ss != src->r_ctl.rc_fixed_pacing_rate_ss) {
23963 		dest->r_ctl.rc_fixed_pacing_rate_ss = src->r_ctl.rc_fixed_pacing_rate_ss;
23964 		cnt++;
23965 	}
23966 	if (dest->r_ctl.rc_fixed_pacing_rate_rec != src->r_ctl.rc_fixed_pacing_rate_rec) {
23967 		dest->r_ctl.rc_fixed_pacing_rate_rec = src->r_ctl.rc_fixed_pacing_rate_rec;
23968 		cnt++;
23969 	}
23970 	/* TCP_RACK_GP_INCREASE_REC, TCP_RACK_GP_INCREASE_CA, TCP_RACK_GP_INCREASE_SS */
23971 	if (dest->r_ctl.rack_per_of_gp_rec != src->r_ctl.rack_per_of_gp_rec) {
23972 		dest->r_ctl.rack_per_of_gp_rec = src->r_ctl.rack_per_of_gp_rec;
23973 		cnt++;
23974 	}
23975 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23976 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23977 		cnt++;
23978 	}
23979 
23980 	if (dest->r_ctl.rack_per_of_gp_ss != src->r_ctl.rack_per_of_gp_ss) {
23981 		dest->r_ctl.rack_per_of_gp_ss = src->r_ctl.rack_per_of_gp_ss;
23982 		cnt++;
23983 	}
23984 	/* TCP_RACK_RR_CONF */
23985 	if (dest->r_rr_config != src->r_rr_config) {
23986 		dest->r_rr_config = src->r_rr_config;
23987 		cnt++;
23988 	}
23989 	/* TCP_PACING_DND */
23990 	if (dest->rc_pace_dnd != src->rc_pace_dnd) {
23991 		dest->rc_pace_dnd = src->rc_pace_dnd;
23992 		cnt++;
23993 	}
23994 	/* TCP_HDWR_RATE_CAP */
23995 	if (dest->r_rack_hw_rate_caps != src->r_rack_hw_rate_caps) {
23996 		dest->r_rack_hw_rate_caps = src->r_rack_hw_rate_caps;
23997 		cnt++;
23998 	}
23999 	/* TCP_DGP_UPPER_BOUNDS */
24000 	if (dest->r_ctl.rack_per_upper_bound_ca != src->r_ctl.rack_per_upper_bound_ca) {
24001 		dest->r_ctl.rack_per_upper_bound_ca = src->r_ctl.rack_per_upper_bound_ca;
24002 		cnt++;
24003 	}
24004 	if (dest->r_ctl.rack_per_upper_bound_ss != src->r_ctl.rack_per_upper_bound_ss) {
24005 		dest->r_ctl.rack_per_upper_bound_ss = src->r_ctl.rack_per_upper_bound_ss;
24006 		cnt++;
24007 	}
24008 	/* TCP_SS_EEXIT */
24009 	if (dest->r_ctl.gp_rnd_thresh != src->r_ctl.gp_rnd_thresh) {
24010 		dest->r_ctl.gp_rnd_thresh = src->r_ctl.gp_rnd_thresh;
24011 		cnt++;
24012 	}
24013 	if (dest->r_ctl.gate_to_fs != src->r_ctl.gate_to_fs) {
24014 		dest->r_ctl.gate_to_fs = src->r_ctl.gate_to_fs;
24015 		cnt++;
24016 	}
24017 	if (dest->r_ctl.use_gp_not_last != src->r_ctl.use_gp_not_last) {
24018 		dest->r_ctl.use_gp_not_last = src->r_ctl.use_gp_not_last;
24019 		cnt++;
24020 	}
24021 	if (dest->r_ctl.gp_gain_req != src->r_ctl.gp_gain_req) {
24022 		dest->r_ctl.gp_gain_req = src->r_ctl.gp_gain_req;
24023 		cnt++;
24024 	}
24025 	/* TCP_BBR_HDWR_PACE */
24026 	if (dest->rack_hdw_pace_ena != src->rack_hdw_pace_ena) {
24027 		dest->rack_hdw_pace_ena = src->rack_hdw_pace_ena;
24028 		cnt++;
24029 	}
24030 	if (dest->rack_attempt_hdwr_pace != src->rack_attempt_hdwr_pace) {
24031 		dest->rack_attempt_hdwr_pace = src->rack_attempt_hdwr_pace;
24032 		cnt++;
24033 	}
24034 	/* TCP_RACK_PRR_SENDALOT */
24035 	if (dest->r_ctl.rc_prr_sendalot != src->r_ctl.rc_prr_sendalot) {
24036 		dest->r_ctl.rc_prr_sendalot = src->r_ctl.rc_prr_sendalot;
24037 		cnt++;
24038 	}
24039 	/* TCP_RACK_MIN_TO */
24040 	if (dest->r_ctl.rc_min_to != src->r_ctl.rc_min_to) {
24041 		dest->r_ctl.rc_min_to = src->r_ctl.rc_min_to;
24042 		cnt++;
24043 	}
24044 	/* TCP_RACK_EARLY_SEG */
24045 	if (dest->r_ctl.rc_early_recovery_segs != src->r_ctl.rc_early_recovery_segs) {
24046 		dest->r_ctl.rc_early_recovery_segs = src->r_ctl.rc_early_recovery_segs;
24047 		cnt++;
24048 	}
24049 	/* TCP_RACK_ENABLE_HYSTART */
24050 	if (par->t_ccv.flags != tp->t_ccv.flags) {
24051 		cnt++;
24052 		if (par->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24053 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
24054 			if (rack_do_hystart > RACK_HYSTART_ON)
24055 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
24056 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
24057 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
24058 		} else {
24059 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
24060 		}
24061 	}
24062 	/* TCP_RACK_REORD_THRESH */
24063 	if (dest->r_ctl.rc_reorder_shift != src->r_ctl.rc_reorder_shift) {
24064 		dest->r_ctl.rc_reorder_shift = src->r_ctl.rc_reorder_shift;
24065 		cnt++;
24066 	}
24067 	/* TCP_RACK_REORD_FADE */
24068 	if (dest->r_ctl.rc_reorder_fade != src->r_ctl.rc_reorder_fade) {
24069 		dest->r_ctl.rc_reorder_fade = src->r_ctl.rc_reorder_fade;
24070 		cnt++;
24071 	}
24072 	/* TCP_RACK_TLP_THRESH */
24073 	if (dest->r_ctl.rc_tlp_threshold != src->r_ctl.rc_tlp_threshold) {
24074 		dest->r_ctl.rc_tlp_threshold = src->r_ctl.rc_tlp_threshold;
24075 		cnt++;
24076 	}
24077 	/* TCP_BBR_USE_RACK_RR */
24078 	if (dest->use_rack_rr != src->use_rack_rr) {
24079 		dest->use_rack_rr = src->use_rack_rr;
24080 		cnt++;
24081 	}
24082 	/* TCP_RACK_PKT_DELAY */
24083 	if (dest->r_ctl.rc_pkt_delay != src->r_ctl.rc_pkt_delay) {
24084 		dest->r_ctl.rc_pkt_delay = src->r_ctl.rc_pkt_delay;
24085 		cnt++;
24086 	}
24087 	/* TCP_DELACK will get copied via the main code if applicable */
24088 	/* TCP_BBR_RACK_RTT_USE */
24089 	if (dest->r_ctl.rc_rate_sample_method != src->r_ctl.rc_rate_sample_method) {
24090 		dest->r_ctl.rc_rate_sample_method = src->r_ctl.rc_rate_sample_method;
24091 		cnt++;
24092 	}
24093 	/* TCP_HONOR_HPTS_MIN */
24094 	if (dest->r_use_hpts_min != src->r_use_hpts_min) {
24095 		dest->r_use_hpts_min = src->r_use_hpts_min;
24096 		cnt++;
24097 	}
24098 	if (dest->r_ctl.max_reduction != src->r_ctl.max_reduction) {
24099 		dest->r_ctl.max_reduction = src->r_ctl.max_reduction;
24100 		cnt++;
24101 	}
24102 	/* TCP_REC_IS_DYN */
24103 	if (dest->rc_gp_no_rec_chg != src->rc_gp_no_rec_chg) {
24104 		dest->rc_gp_no_rec_chg = src->rc_gp_no_rec_chg;
24105 		cnt++;
24106 	}
24107 	if (dest->rc_skip_timely != src->rc_skip_timely) {
24108 		dest->rc_skip_timely = src->rc_skip_timely;
24109 		cnt++;
24110 	}
24111 	/* TCP_DATA_AFTER_CLOSE */
24112 	if (dest->rc_allow_data_af_clo != src->rc_allow_data_af_clo) {
24113 		dest->rc_allow_data_af_clo = src->rc_allow_data_af_clo;
24114 		cnt++;
24115 	}
24116 	/* TCP_GP_USE_LTBW */
24117 	if (src->use_lesser_lt_bw != dest->use_lesser_lt_bw) {
24118 		dest->use_lesser_lt_bw = src->use_lesser_lt_bw;
24119 		cnt++;
24120 	}
24121 	if (dest->dis_lt_bw != src->dis_lt_bw) {
24122 		dest->dis_lt_bw = src->dis_lt_bw;
24123 		cnt++;
24124 	}
24125 	tcp_log_socket_option(tp, 0, cnt, 0);
24126 }
24127 
24128 
24129 static void
24130 rack_apply_deferred_options(struct tcp_rack *rack)
24131 {
24132 	struct deferred_opt_list *dol, *sdol;
24133 	uint32_t s_optval;
24134 
24135 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
24136 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
24137 		/* Disadvantage of deferal is you loose the error return */
24138 		s_optval = (uint32_t)dol->optval;
24139 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
24140 		free(dol, M_TCPDO);
24141 	}
24142 }
24143 
24144 static void
24145 rack_hw_tls_change(struct tcpcb *tp, int chg)
24146 {
24147 	/* Update HW tls state */
24148 	struct tcp_rack *rack;
24149 
24150 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24151 	if (chg)
24152 		rack->r_ctl.fsb.hw_tls = 1;
24153 	else
24154 		rack->r_ctl.fsb.hw_tls = 0;
24155 }
24156 
24157 static int
24158 rack_pru_options(struct tcpcb *tp, int flags)
24159 {
24160 	if (flags & PRUS_OOB)
24161 		return (EOPNOTSUPP);
24162 	return (0);
24163 }
24164 
24165 static bool
24166 rack_wake_check(struct tcpcb *tp)
24167 {
24168 	struct tcp_rack *rack;
24169 	struct timeval tv;
24170 	uint32_t cts;
24171 
24172 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24173 	if (rack->r_ctl.rc_hpts_flags) {
24174 		cts = tcp_get_usecs(&tv);
24175 		if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
24176 			/*
24177 			 * Pacing timer is up, check if we are ready.
24178 			 */
24179 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
24180 				return (true);
24181 		} else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
24182 			/*
24183 			 * A timer is up, check if we are ready.
24184 			 */
24185 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
24186 				return (true);
24187 		}
24188 	}
24189 	return (false);
24190 }
24191 
24192 static struct tcp_function_block __tcp_rack = {
24193 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
24194 	.tfb_tcp_output = rack_output,
24195 	.tfb_do_queued_segments = ctf_do_queued_segments,
24196 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
24197 	.tfb_tcp_do_segment = rack_do_segment,
24198 	.tfb_tcp_ctloutput = rack_ctloutput,
24199 	.tfb_tcp_fb_init = rack_init,
24200 	.tfb_tcp_fb_fini = rack_fini,
24201 	.tfb_tcp_timer_stop_all = rack_stopall,
24202 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
24203 	.tfb_tcp_handoff_ok = rack_handoff_ok,
24204 	.tfb_tcp_mtu_chg = rack_mtu_change,
24205 	.tfb_pru_options = rack_pru_options,
24206 	.tfb_hwtls_change = rack_hw_tls_change,
24207 	.tfb_chg_query = rack_chg_query,
24208 	.tfb_switch_failed = rack_switch_failed,
24209 	.tfb_early_wake_check = rack_wake_check,
24210 	.tfb_compute_pipe = rack_compute_pipe,
24211 	.tfb_stack_info = rack_stack_information,
24212 	.tfb_inherit = rack_inherit,
24213 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP | TCP_FUNC_DEFAULT_OK,
24214 
24215 };
24216 
24217 /*
24218  * rack_ctloutput() must drop the inpcb lock before performing copyin on
24219  * socket option arguments.  When it re-acquires the lock after the copy, it
24220  * has to revalidate that the connection is still valid for the socket
24221  * option.
24222  */
24223 static int
24224 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24225 {
24226 	struct inpcb *inp = tptoinpcb(tp);
24227 #ifdef INET
24228 	struct ip *ip;
24229 #endif
24230 	struct tcp_rack *rack;
24231 	struct tcp_hybrid_req hybrid;
24232 	uint64_t loptval;
24233 	int32_t error = 0, optval;
24234 
24235 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24236 	if (rack == NULL) {
24237 		INP_WUNLOCK(inp);
24238 		return (EINVAL);
24239 	}
24240 #ifdef INET
24241 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
24242 #endif
24243 
24244 	switch (sopt->sopt_level) {
24245 #ifdef INET6
24246 	case IPPROTO_IPV6:
24247 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
24248 		switch (sopt->sopt_name) {
24249 		case IPV6_USE_MIN_MTU:
24250 			tcp6_use_min_mtu(tp);
24251 			break;
24252 		}
24253 		INP_WUNLOCK(inp);
24254 		return (0);
24255 #endif
24256 #ifdef INET
24257 	case IPPROTO_IP:
24258 		switch (sopt->sopt_name) {
24259 		case IP_TOS:
24260 			/*
24261 			 * The DSCP codepoint has changed, update the fsb.
24262 			 */
24263 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
24264 			break;
24265 		case IP_TTL:
24266 			/*
24267 			 * The TTL has changed, update the fsb.
24268 			 */
24269 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
24270 			break;
24271 		}
24272 		INP_WUNLOCK(inp);
24273 		return (0);
24274 #endif
24275 #ifdef SO_PEERPRIO
24276 	case SOL_SOCKET:
24277 		switch (sopt->sopt_name) {
24278 		case SO_PEERPRIO:			/*  SC-URL:bs */
24279 			/* Already read in and sanity checked in sosetopt(). */
24280 			if (inp->inp_socket) {
24281 				rack->client_bufferlvl = inp->inp_socket->so_peerprio;
24282 			}
24283 			break;
24284 		}
24285 		INP_WUNLOCK(inp);
24286 		return (0);
24287 #endif
24288 	case IPPROTO_TCP:
24289 		switch (sopt->sopt_name) {
24290 		case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
24291 		/*  Pacing related ones */
24292 		case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
24293 		case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
24294 		case TCP_RACK_PACE_MIN_SEG:		/*  URL:pace_min_seg */
24295 		case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
24296 		case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
24297 		case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
24298 		case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
24299 		case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
24300 		case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
24301 		case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
24302 		case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
24303 		case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
24304 		case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
24305 		case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
24306 		case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
24307 		case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
24308 		case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
24309 		case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
24310 		case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
24311 			/* End pacing related */
24312 		case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
24313 		case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
24314 		case TCP_RACK_MIN_TO:			/*  URL:min_to */
24315 		case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
24316 		case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
24317 		case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
24318 		case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
24319 		case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
24320 		case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
24321 		case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
24322 		case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
24323 		case TCP_NO_PRR:			/*  URL:noprr */
24324 		case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
24325 		case TCP_DATA_AFTER_CLOSE:		/*  no URL */
24326 		case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
24327 		case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
24328 		case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
24329 		case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
24330 		case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
24331 		case TCP_RACK_PROFILE:			/*  URL:profile */
24332 		case TCP_SIDECHAN_DIS:			/*  URL:scodm */
24333 		case TCP_HYBRID_PACING:			/*  URL:pacing=hybrid */
24334 		case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
24335 		case TCP_RACK_ABC_VAL:			/*  URL:labc */
24336 		case TCP_REC_ABC_VAL:			/*  URL:reclabc */
24337 		case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
24338 		case TCP_DEFER_OPTIONS:			/*  URL:defer */
24339 		case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
24340 		case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
24341 		case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
24342 		case TCP_RACK_SET_RXT_OPTIONS:		/*  URL:rxtsz */
24343 		case TCP_RACK_HI_BETA:			/*  URL:hibeta */
24344 		case TCP_RACK_SPLIT_LIMIT:		/*  URL:split */
24345 		case TCP_SS_EEXIT:			/*  URL:eexit */
24346 		case TCP_DGP_UPPER_BOUNDS:		/*  URL:upper */
24347 		case TCP_RACK_PACING_DIVISOR:		/*  URL:divisor */
24348 		case TCP_PACING_DND:			/*  URL:dnd */
24349 		case TCP_NO_TIMELY:			/*  URL:notimely */
24350 		case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
24351 		case TCP_HONOR_HPTS_MIN:		/*  URL:hptsmin */
24352 		case TCP_REC_IS_DYN:			/*  URL:dynrec */
24353 		case TCP_GP_USE_LTBW:			/*  URL:useltbw */
24354 			goto process_opt;
24355 			break;
24356 		default:
24357 			/* Filter off all unknown options to the base stack */
24358 			return (tcp_default_ctloutput(tp, sopt));
24359 			break;
24360 		}
24361 	default:
24362 		INP_WUNLOCK(inp);
24363 		return (0);
24364 	}
24365 process_opt:
24366 	INP_WUNLOCK(inp);
24367 	if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24368 	    (sopt->sopt_name == TCP_FILLCW_RATE_CAP)) {
24369 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
24370 		/*
24371 		 * We truncate it down to 32 bits for the socket-option trace this
24372 		 * means rates > 34Gbps won't show right, but thats probably ok.
24373 		 */
24374 		optval = (uint32_t)loptval;
24375 	} else if (sopt->sopt_name == TCP_HYBRID_PACING) {
24376 		error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
24377 	} else {
24378 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
24379 		/* Save it in 64 bit form too */
24380 		loptval = optval;
24381 	}
24382 	if (error)
24383 		return (error);
24384 	INP_WLOCK(inp);
24385 	if (tp->t_fb != &__tcp_rack) {
24386 		INP_WUNLOCK(inp);
24387 		return (ENOPROTOOPT);
24388 	}
24389 	if (rack->defer_options && (rack->gp_ready == 0) &&
24390 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
24391 	    (sopt->sopt_name != TCP_HYBRID_PACING) &&
24392 	    (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
24393 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
24394 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
24395 		/* Options are being deferred */
24396 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
24397 			INP_WUNLOCK(inp);
24398 			return (0);
24399 		} else {
24400 			/* No memory to defer, fail */
24401 			INP_WUNLOCK(inp);
24402 			return (ENOMEM);
24403 		}
24404 	}
24405 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
24406 	INP_WUNLOCK(inp);
24407 	return (error);
24408 }
24409 
24410 static void
24411 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
24412 {
24413 
24414 	INP_WLOCK_ASSERT(tptoinpcb(tp));
24415 	bzero(ti, sizeof(*ti));
24416 
24417 	ti->tcpi_state = tp->t_state;
24418 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
24419 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
24420 	if (tp->t_flags & TF_SACK_PERMIT)
24421 		ti->tcpi_options |= TCPI_OPT_SACK;
24422 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
24423 		ti->tcpi_options |= TCPI_OPT_WSCALE;
24424 		ti->tcpi_snd_wscale = tp->snd_scale;
24425 		ti->tcpi_rcv_wscale = tp->rcv_scale;
24426 	}
24427 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
24428 		ti->tcpi_options |= TCPI_OPT_ECN;
24429 	if (tp->t_flags & TF_FASTOPEN)
24430 		ti->tcpi_options |= TCPI_OPT_TFO;
24431 	/* still kept in ticks is t_rcvtime */
24432 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
24433 	/* Since we hold everything in precise useconds this is easy */
24434 	ti->tcpi_rtt = tp->t_srtt;
24435 	ti->tcpi_rttvar = tp->t_rttvar;
24436 	ti->tcpi_rto = tp->t_rxtcur;
24437 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
24438 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
24439 	/*
24440 	 * FreeBSD-specific extension fields for tcp_info.
24441 	 */
24442 	ti->tcpi_rcv_space = tp->rcv_wnd;
24443 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
24444 	ti->tcpi_snd_wnd = tp->snd_wnd;
24445 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
24446 	ti->tcpi_snd_nxt = tp->snd_nxt;
24447 	ti->tcpi_snd_mss = tp->t_maxseg;
24448 	ti->tcpi_rcv_mss = tp->t_maxseg;
24449 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
24450 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
24451 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
24452 	ti->tcpi_total_tlp = tp->t_sndtlppack;
24453 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
24454 	ti->tcpi_rttmin = tp->t_rttlow;
24455 #ifdef NETFLIX_STATS
24456 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
24457 #endif
24458 #ifdef TCP_OFFLOAD
24459 	if (tp->t_flags & TF_TOE) {
24460 		ti->tcpi_options |= TCPI_OPT_TOE;
24461 		tcp_offload_tcp_info(tp, ti);
24462 	}
24463 #endif
24464 }
24465 
24466 static int
24467 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24468 {
24469 	struct inpcb *inp = tptoinpcb(tp);
24470 	struct tcp_rack *rack;
24471 	int32_t error, optval;
24472 	uint64_t val, loptval;
24473 	struct	tcp_info ti;
24474 	/*
24475 	 * Because all our options are either boolean or an int, we can just
24476 	 * pull everything into optval and then unlock and copy. If we ever
24477 	 * add a option that is not a int, then this will have quite an
24478 	 * impact to this routine.
24479 	 */
24480 	error = 0;
24481 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24482 	if (rack == NULL) {
24483 		INP_WUNLOCK(inp);
24484 		return (EINVAL);
24485 	}
24486 	switch (sopt->sopt_name) {
24487 	case TCP_INFO:
24488 		/* First get the info filled */
24489 		rack_fill_info(tp, &ti);
24490 		/* Fix up the rtt related fields if needed */
24491 		INP_WUNLOCK(inp);
24492 		error = sooptcopyout(sopt, &ti, sizeof ti);
24493 		return (error);
24494 	/*
24495 	 * Beta is the congestion control value for NewReno that influences how
24496 	 * much of a backoff happens when loss is detected. It is normally set
24497 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
24498 	 * when you exit recovery.
24499 	 */
24500 	case TCP_RACK_PACING_BETA:
24501 		break;
24502 		/*
24503 		 * Beta_ecn is the congestion control value for NewReno that influences how
24504 		 * much of a backoff happens when a ECN mark is detected. It is normally set
24505 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
24506 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
24507 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
24508 		 */
24509 
24510 	case TCP_RACK_PACING_BETA_ECN:
24511 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24512 			error = EINVAL;
24513 		else if (rack->rc_pacing_cc_set == 0)
24514 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
24515 		else {
24516 			/*
24517 			 * Reach out into the CC data and report back what
24518 			 * I have previously set. Yeah it looks hackish but
24519 			 * we don't want to report the saved values.
24520 			 */
24521 			if (tp->t_ccv.cc_data)
24522 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
24523 			else
24524 				error = EINVAL;
24525 		}
24526 		break;
24527 	case TCP_RACK_DSACK_OPT:
24528 		optval = 0;
24529 		if (rack->rc_rack_tmr_std_based) {
24530 			optval |= 1;
24531 		}
24532 		if (rack->rc_rack_use_dsack) {
24533 			optval |= 2;
24534 		}
24535 		break;
24536 	case TCP_RACK_ENABLE_HYSTART:
24537 	{
24538 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24539 			optval = RACK_HYSTART_ON;
24540 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
24541 				optval = RACK_HYSTART_ON_W_SC;
24542 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
24543 				optval = RACK_HYSTART_ON_W_SC_C;
24544 		} else {
24545 			optval = RACK_HYSTART_OFF;
24546 		}
24547 	}
24548 	break;
24549 	case TCP_RACK_DGP_IN_REC:
24550 		error = EINVAL;
24551 		break;
24552 	case TCP_RACK_HI_BETA:
24553 		optval = rack->rack_hibeta;
24554 		break;
24555 	case TCP_DEFER_OPTIONS:
24556 		optval = rack->defer_options;
24557 		break;
24558 	case TCP_RACK_MEASURE_CNT:
24559 		optval = rack->r_ctl.req_measurements;
24560 		break;
24561 	case TCP_REC_ABC_VAL:
24562 		optval = rack->r_use_labc_for_rec;
24563 		break;
24564 	case TCP_RACK_ABC_VAL:
24565 		optval = rack->rc_labc;
24566 		break;
24567 	case TCP_HDWR_UP_ONLY:
24568 		optval= rack->r_up_only;
24569 		break;
24570 	case TCP_FILLCW_RATE_CAP:
24571 		loptval = rack->r_ctl.fillcw_cap;
24572 		break;
24573 	case TCP_PACING_RATE_CAP:
24574 		loptval = rack->r_ctl.bw_rate_cap;
24575 		break;
24576 	case TCP_RACK_PROFILE:
24577 		/* You cannot retrieve a profile, its write only */
24578 		error = EINVAL;
24579 		break;
24580 	case TCP_SIDECHAN_DIS:
24581 		optval = rack->r_ctl.side_chan_dis_mask;
24582 		break;
24583 	case TCP_HYBRID_PACING:
24584 		/* You cannot retrieve hybrid pacing information, its write only */
24585 		error = EINVAL;
24586 		break;
24587 	case TCP_USE_CMP_ACKS:
24588 		optval = rack->r_use_cmp_ack;
24589 		break;
24590 	case TCP_RACK_PACE_TO_FILL:
24591 		optval = rack->rc_pace_to_cwnd;
24592 		break;
24593 	case TCP_RACK_NO_PUSH_AT_MAX:
24594 		optval = rack->r_ctl.rc_no_push_at_mrtt;
24595 		break;
24596 	case TCP_SHARED_CWND_ENABLE:
24597 		optval = rack->rack_enable_scwnd;
24598 		break;
24599 	case TCP_RACK_NONRXT_CFG_RATE:
24600 		optval = rack->rack_rec_nonrxt_use_cr;
24601 		break;
24602 	case TCP_NO_PRR:
24603 		if (rack->rack_no_prr  == 1)
24604 			optval = 1;
24605 		else if (rack->no_prr_addback == 1)
24606 			optval = 2;
24607 		else
24608 			optval = 0;
24609 		break;
24610 	case TCP_GP_USE_LTBW:
24611 		if (rack->dis_lt_bw) {
24612 			/* It is not used */
24613 			optval = 0;
24614 		} else if (rack->use_lesser_lt_bw) {
24615 			/* we use min() */
24616 			optval = 1;
24617 		} else {
24618 			/* we use max() */
24619 			optval = 2;
24620 		}
24621 		break;
24622 	case TCP_RACK_DO_DETECTION:
24623 		error = EINVAL;
24624 		break;
24625 	case TCP_RACK_MBUF_QUEUE:
24626 		/* Now do we use the LRO mbuf-queue feature */
24627 		optval = rack->r_mbuf_queue;
24628 		break;
24629 	case RACK_CSPR_IS_FCC:
24630 		optval = rack->cspr_is_fcc;
24631 		break;
24632 	case TCP_TIMELY_DYN_ADJ:
24633 		optval = rack->rc_gp_dyn_mul;
24634 		break;
24635 	case TCP_BBR_IWINTSO:
24636 		error = EINVAL;
24637 		break;
24638 	case TCP_RACK_TLP_REDUCE:
24639 		/* RACK TLP cwnd reduction (bool) */
24640 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
24641 		break;
24642 	case TCP_BBR_RACK_INIT_RATE:
24643 		val = rack->r_ctl.init_rate;
24644 		/* convert to kbits per sec */
24645 		val *= 8;
24646 		val /= 1000;
24647 		optval = (uint32_t)val;
24648 		break;
24649 	case TCP_RACK_FORCE_MSEG:
24650 		optval = rack->rc_force_max_seg;
24651 		break;
24652 	case TCP_RACK_PACE_MIN_SEG:
24653 		optval = rack->r_ctl.rc_user_set_min_segs;
24654 		break;
24655 	case TCP_RACK_PACE_MAX_SEG:
24656 		/* Max segments in a pace */
24657 		optval = rack->rc_user_set_max_segs;
24658 		break;
24659 	case TCP_RACK_PACE_ALWAYS:
24660 		/* Use the always pace method */
24661 		optval = rack->rc_always_pace;
24662 		break;
24663 	case TCP_RACK_PRR_SENDALOT:
24664 		/* Allow PRR to send more than one seg */
24665 		optval = rack->r_ctl.rc_prr_sendalot;
24666 		break;
24667 	case TCP_RACK_MIN_TO:
24668 		/* Minimum time between rack t-o's in ms */
24669 		optval = rack->r_ctl.rc_min_to;
24670 		break;
24671 	case TCP_RACK_SPLIT_LIMIT:
24672 		optval = rack->r_ctl.rc_split_limit;
24673 		break;
24674 	case TCP_RACK_EARLY_SEG:
24675 		/* If early recovery max segments */
24676 		optval = rack->r_ctl.rc_early_recovery_segs;
24677 		break;
24678 	case TCP_RACK_REORD_THRESH:
24679 		/* RACK reorder threshold (shift amount) */
24680 		optval = rack->r_ctl.rc_reorder_shift;
24681 		break;
24682 	case TCP_SS_EEXIT:
24683 		if (rack->r_ctl.gp_rnd_thresh) {
24684 			uint32_t v;
24685 
24686 			v = rack->r_ctl.gp_gain_req;
24687 			v <<= 17;
24688 			optval = v | (rack->r_ctl.gp_rnd_thresh & 0xff);
24689 			if (rack->r_ctl.gate_to_fs == 1)
24690 				optval |= 0x10000;
24691 		} else
24692 			optval = 0;
24693 		break;
24694 	case TCP_RACK_REORD_FADE:
24695 		/* Does reordering fade after ms time */
24696 		optval = rack->r_ctl.rc_reorder_fade;
24697 		break;
24698 	case TCP_BBR_USE_RACK_RR:
24699 		/* Do we use the rack cheat for rxt */
24700 		optval = rack->use_rack_rr;
24701 		break;
24702 	case TCP_RACK_RR_CONF:
24703 		optval = rack->r_rr_config;
24704 		break;
24705 	case TCP_HDWR_RATE_CAP:
24706 		optval = rack->r_rack_hw_rate_caps;
24707 		break;
24708 	case TCP_BBR_HDWR_PACE:
24709 		optval = rack->rack_hdw_pace_ena;
24710 		break;
24711 	case TCP_RACK_TLP_THRESH:
24712 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
24713 		optval = rack->r_ctl.rc_tlp_threshold;
24714 		break;
24715 	case TCP_RACK_PKT_DELAY:
24716 		/* RACK added ms i.e. rack-rtt + reord + N */
24717 		optval = rack->r_ctl.rc_pkt_delay;
24718 		break;
24719 	case TCP_RACK_TLP_USE:
24720 		optval = rack->rack_tlp_threshold_use;
24721 		break;
24722 	case TCP_PACING_DND:
24723 		optval = rack->rc_pace_dnd;
24724 		break;
24725 	case TCP_RACK_PACE_RATE_CA:
24726 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
24727 		break;
24728 	case TCP_RACK_PACE_RATE_SS:
24729 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
24730 		break;
24731 	case TCP_RACK_PACE_RATE_REC:
24732 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
24733 		break;
24734 	case TCP_DGP_UPPER_BOUNDS:
24735 		optval = rack->r_ctl.rack_per_upper_bound_ss;
24736 		optval <<= 16;
24737 		optval |= rack->r_ctl.rack_per_upper_bound_ca;
24738 		break;
24739 	case TCP_RACK_GP_INCREASE_SS:
24740 		optval = rack->r_ctl.rack_per_of_gp_ca;
24741 		break;
24742 	case TCP_RACK_GP_INCREASE_CA:
24743 		optval = rack->r_ctl.rack_per_of_gp_ss;
24744 		break;
24745 	case TCP_RACK_PACING_DIVISOR:
24746 		optval = rack->r_ctl.pace_len_divisor;
24747 		break;
24748 	case TCP_BBR_RACK_RTT_USE:
24749 		optval = rack->r_ctl.rc_rate_sample_method;
24750 		break;
24751 	case TCP_DELACK:
24752 		optval = tp->t_delayed_ack;
24753 		break;
24754 	case TCP_DATA_AFTER_CLOSE:
24755 		optval = rack->rc_allow_data_af_clo;
24756 		break;
24757 	case TCP_SHARED_CWND_TIME_LIMIT:
24758 		optval = rack->r_limit_scw;
24759 		break;
24760 	case TCP_HONOR_HPTS_MIN:
24761 		if (rack->r_use_hpts_min)
24762 			optval = rack->r_ctl.max_reduction;
24763 		else
24764 			optval = 0;
24765 		break;
24766 	case TCP_REC_IS_DYN:
24767 		optval = rack->rc_gp_no_rec_chg;
24768 		break;
24769 	case TCP_NO_TIMELY:
24770 		optval = rack->rc_skip_timely;
24771 		break;
24772 	case TCP_RACK_TIMER_SLOP:
24773 		optval = rack->r_ctl.timer_slop;
24774 		break;
24775 	default:
24776 		return (tcp_default_ctloutput(tp, sopt));
24777 		break;
24778 	}
24779 	INP_WUNLOCK(inp);
24780 	if (error == 0) {
24781 		if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24782 		    (sopt->sopt_name == TCP_FILLCW_RATE_CAP))
24783 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
24784 		else
24785 			error = sooptcopyout(sopt, &optval, sizeof optval);
24786 	}
24787 	return (error);
24788 }
24789 
24790 static int
24791 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
24792 {
24793 	if (sopt->sopt_dir == SOPT_SET) {
24794 		return (rack_set_sockopt(tp, sopt));
24795 	} else if (sopt->sopt_dir == SOPT_GET) {
24796 		return (rack_get_sockopt(tp, sopt));
24797 	} else {
24798 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
24799 	}
24800 }
24801 
24802 static const char *rack_stack_names[] = {
24803 	__XSTRING(STACKNAME),
24804 #ifdef STACKALIAS
24805 	__XSTRING(STACKALIAS),
24806 #endif
24807 };
24808 
24809 static int
24810 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
24811 {
24812 	memset(mem, 0, size);
24813 	return (0);
24814 }
24815 
24816 static void
24817 rack_dtor(void *mem, int32_t size, void *arg)
24818 {
24819 
24820 }
24821 
24822 static bool rack_mod_inited = false;
24823 
24824 static int
24825 tcp_addrack(module_t mod, int32_t type, void *data)
24826 {
24827 	int32_t err = 0;
24828 	int num_stacks;
24829 
24830 	switch (type) {
24831 	case MOD_LOAD:
24832 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24833 		    sizeof(struct rack_sendmap),
24834 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24835 
24836 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24837 		    sizeof(struct tcp_rack),
24838 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24839 
24840 		sysctl_ctx_init(&rack_sysctl_ctx);
24841 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24842 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24843 		    OID_AUTO,
24844 #ifdef STACKALIAS
24845 		    __XSTRING(STACKALIAS),
24846 #else
24847 		    __XSTRING(STACKNAME),
24848 #endif
24849 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24850 		    "");
24851 		if (rack_sysctl_root == NULL) {
24852 			printf("Failed to add sysctl node\n");
24853 			err = EFAULT;
24854 			goto free_uma;
24855 		}
24856 		rack_init_sysctls();
24857 		num_stacks = nitems(rack_stack_names);
24858 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24859 		    rack_stack_names, &num_stacks);
24860 		if (err) {
24861 			printf("Failed to register %s stack name for "
24862 			    "%s module\n", rack_stack_names[num_stacks],
24863 			    __XSTRING(MODNAME));
24864 			sysctl_ctx_free(&rack_sysctl_ctx);
24865 free_uma:
24866 			uma_zdestroy(rack_zone);
24867 			uma_zdestroy(rack_pcb_zone);
24868 			rack_counter_destroy();
24869 			printf("Failed to register rack module -- err:%d\n", err);
24870 			return (err);
24871 		}
24872 		tcp_lro_reg_mbufq();
24873 		rack_mod_inited = true;
24874 		break;
24875 	case MOD_QUIESCE:
24876 		err = deregister_tcp_functions(&__tcp_rack, true, false);
24877 		break;
24878 	case MOD_UNLOAD:
24879 		err = deregister_tcp_functions(&__tcp_rack, false, true);
24880 		if (err == EBUSY)
24881 			break;
24882 		if (rack_mod_inited) {
24883 			uma_zdestroy(rack_zone);
24884 			uma_zdestroy(rack_pcb_zone);
24885 			sysctl_ctx_free(&rack_sysctl_ctx);
24886 			rack_counter_destroy();
24887 			rack_mod_inited = false;
24888 		}
24889 		tcp_lro_dereg_mbufq();
24890 		err = 0;
24891 		break;
24892 	default:
24893 		return (EOPNOTSUPP);
24894 	}
24895 	return (err);
24896 }
24897 
24898 static moduledata_t tcp_rack = {
24899 	.name = __XSTRING(MODNAME),
24900 	.evhand = tcp_addrack,
24901 	.priv = 0
24902 };
24903 
24904 MODULE_VERSION(MODNAME, 1);
24905 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24906 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24907 
24908 #endif /* #if !defined(INET) && !defined(INET6) */
24909