xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision f3087bef11543b42e0d69b708f367097a4118d24) 
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 	old.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
646 	/* Get the current values out */
647 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
648 	sopt.sopt_dir = SOPT_GET;
649 	opt.name = CC_NEWRENO_BETA;
650 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
651 	if (error)  {
652 		failed = 3;
653 		goto out;
654 	}
655 	old.beta = opt.val;
656 	opt.name = CC_NEWRENO_BETA_ECN;
657 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
658 	if (error)  {
659 		failed = 4;
660 		goto out;
661 	}
662 	old.beta_ecn = opt.val;
663 
664 	/* Now lets set in the values we have stored */
665 	sopt.sopt_dir = SOPT_SET;
666 	opt.name = CC_NEWRENO_BETA;
667 	opt.val = rack->r_ctl.rc_saved_beta.beta;
668 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
669 	if (error)  {
670 		failed = 5;
671 		goto out;
672 	}
673 	opt.name = CC_NEWRENO_BETA_ECN;
674 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
675 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
676 	if (error) {
677 		failed = 6;
678 		goto out;
679 	}
680 	/* Save off the values for restoral */
681 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
682 out:
683 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
684 		union tcp_log_stackspecific log;
685 		struct timeval tv;
686 		struct newreno *ptr;
687 
688 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
689 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
690 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
691 		log.u_bbr.flex1 = ptr->beta;
692 		log.u_bbr.flex2 = ptr->beta_ecn;
693 		log.u_bbr.flex3 = ptr->newreno_flags;
694 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
695 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
696 		log.u_bbr.flex6 = failed;
697 		log.u_bbr.flex7 = rack->gp_ready;
698 		log.u_bbr.flex7 <<= 1;
699 		log.u_bbr.flex7 |= rack->use_fixed_rate;
700 		log.u_bbr.flex7 <<= 1;
701 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
702 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
703 		log.u_bbr.flex8 = flex8;
704 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
705 			       0, &log, false, NULL, NULL, 0, &tv);
706 	}
707 }
708 
709 static void
710 rack_set_cc_pacing(struct tcp_rack *rack)
711 {
712 	if (rack->rc_pacing_cc_set)
713 		return;
714 	/*
715 	 * Use the swap utility placing in 3 for flex8 to id a
716 	 * set of a new set of values.
717 	 */
718 	rack->rc_pacing_cc_set = 1;
719 	rack_swap_beta_values(rack, 3);
720 }
721 
722 static void
723 rack_undo_cc_pacing(struct tcp_rack *rack)
724 {
725 	if (rack->rc_pacing_cc_set == 0)
726 		return;
727 	/*
728 	 * Use the swap utility placing in 4 for flex8 to id a
729 	 * restoral of the old values.
730 	 */
731 	rack->rc_pacing_cc_set = 0;
732 	rack_swap_beta_values(rack, 4);
733 }
734 
735 static void
736 rack_remove_pacing(struct tcp_rack *rack)
737 {
738 	if (rack->rc_pacing_cc_set)
739 		rack_undo_cc_pacing(rack);
740 	if (rack->r_ctl.pacing_method & RACK_REG_PACING)
741 		tcp_decrement_paced_conn();
742 	if (rack->r_ctl.pacing_method & RACK_DGP_PACING)
743 		tcp_dec_dgp_pacing_cnt();
744 	rack->rc_always_pace = 0;
745 	rack->r_ctl.pacing_method = RACK_PACING_NONE;
746 	rack->dgp_on = 0;
747 	rack->rc_hybrid_mode = 0;
748 	rack->use_fixed_rate = 0;
749 }
750 
751 static void
752 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
753 	       uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
754 {
755 	if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
756 		union tcp_log_stackspecific log;
757 		struct timeval tv;
758 
759 		memset(&log, 0, sizeof(log));
760 		log.u_bbr.flex1 = seq_end;
761 		log.u_bbr.flex2 = rack->rc_tp->gput_seq;
762 		log.u_bbr.flex3 = ack_end_t;
763 		log.u_bbr.flex4 = rack->rc_tp->gput_ts;
764 		log.u_bbr.flex5 = send_end_t;
765 		log.u_bbr.flex6 = rack->rc_tp->gput_ack;
766 		log.u_bbr.flex7 = mode;
767 		log.u_bbr.flex8 = 69;
768 		log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
769 		log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
770 		log.u_bbr.pkts_out = line;
771 		log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
772 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
773 		log.u_bbr.epoch = rack->r_ctl.current_round;
774 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
775 		if (rsm != NULL) {
776 			log.u_bbr.applimited = rsm->r_start;
777 			log.u_bbr.delivered = rsm->r_end;
778 			log.u_bbr.epoch = rsm->r_flags;
779 		}
780 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
781 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
782 		    &rack->rc_inp->inp_socket->so_rcv,
783 		    &rack->rc_inp->inp_socket->so_snd,
784 		    BBR_LOG_HPTSI_CALC, 0,
785 		    0, &log, false, &tv);
786 	}
787 }
788 
789 static int
790 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
791 {
792 	uint32_t stat;
793 	int32_t error;
794 
795 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
796 	if (error || req->newptr == NULL)
797 		return error;
798 
799 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
800 	if (error)
801 		return (error);
802 	if (stat == 1) {
803 #ifdef INVARIANTS
804 		printf("Clearing RACK counters\n");
805 #endif
806 		counter_u64_zero(rack_tlp_tot);
807 		counter_u64_zero(rack_tlp_newdata);
808 		counter_u64_zero(rack_tlp_retran);
809 		counter_u64_zero(rack_tlp_retran_bytes);
810 		counter_u64_zero(rack_to_tot);
811 		counter_u64_zero(rack_saw_enobuf);
812 		counter_u64_zero(rack_saw_enobuf_hw);
813 		counter_u64_zero(rack_saw_enetunreach);
814 		counter_u64_zero(rack_persists_sends);
815 		counter_u64_zero(rack_total_bytes);
816 		counter_u64_zero(rack_persists_acks);
817 		counter_u64_zero(rack_persists_loss);
818 		counter_u64_zero(rack_persists_lost_ends);
819 #ifdef INVARIANTS
820 		counter_u64_zero(rack_adjust_map_bw);
821 #endif
822 		counter_u64_zero(rack_to_alloc_hard);
823 		counter_u64_zero(rack_to_alloc_emerg);
824 		counter_u64_zero(rack_sack_proc_all);
825 		counter_u64_zero(rack_fto_send);
826 		counter_u64_zero(rack_fto_rsm_send);
827 		counter_u64_zero(rack_extended_rfo);
828 		counter_u64_zero(rack_hw_pace_init_fail);
829 		counter_u64_zero(rack_hw_pace_lost);
830 		counter_u64_zero(rack_non_fto_send);
831 		counter_u64_zero(rack_nfto_resend);
832 		counter_u64_zero(rack_sack_proc_short);
833 		counter_u64_zero(rack_sack_proc_restart);
834 		counter_u64_zero(rack_to_alloc);
835 		counter_u64_zero(rack_to_alloc_limited);
836 		counter_u64_zero(rack_alloc_limited_conns);
837 		counter_u64_zero(rack_split_limited);
838 		counter_u64_zero(rack_rxt_clamps_cwnd);
839 		counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
840 		counter_u64_zero(rack_multi_single_eq);
841 		counter_u64_zero(rack_proc_non_comp_ack);
842 		counter_u64_zero(rack_sack_attacks_detected);
843 		counter_u64_zero(rack_sack_attacks_reversed);
844 		counter_u64_zero(rack_sack_attacks_suspect);
845 		counter_u64_zero(rack_sack_used_next_merge);
846 		counter_u64_zero(rack_sack_used_prev_merge);
847 		counter_u64_zero(rack_sack_splits);
848 		counter_u64_zero(rack_sack_skipped_acked);
849 		counter_u64_zero(rack_ack_total);
850 		counter_u64_zero(rack_express_sack);
851 		counter_u64_zero(rack_sack_total);
852 		counter_u64_zero(rack_move_none);
853 		counter_u64_zero(rack_move_some);
854 		counter_u64_zero(rack_try_scwnd);
855 		counter_u64_zero(rack_collapsed_win);
856 		counter_u64_zero(rack_collapsed_win_rxt);
857 		counter_u64_zero(rack_collapsed_win_seen);
858 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
859 	} else if (stat == 2) {
860 #ifdef INVARIANTS
861 		printf("Clearing RACK option array\n");
862 #endif
863 		COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
864 	} else if (stat == 3) {
865 		printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
866 	} else if (stat == 4) {
867 #ifdef INVARIANTS
868 		printf("Clearing RACK out size array\n");
869 #endif
870 		COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
871 	}
872 	rack_clear_counter = 0;
873 	return (0);
874 }
875 
876 static void
877 rack_init_sysctls(void)
878 {
879 	struct sysctl_oid *rack_counters;
880 	struct sysctl_oid *rack_attack;
881 	struct sysctl_oid *rack_pacing;
882 	struct sysctl_oid *rack_timely;
883 	struct sysctl_oid *rack_timers;
884 	struct sysctl_oid *rack_tlp;
885 	struct sysctl_oid *rack_misc;
886 	struct sysctl_oid *rack_features;
887 	struct sysctl_oid *rack_measure;
888 	struct sysctl_oid *rack_probertt;
889 	struct sysctl_oid *rack_hw_pacing;
890 
891 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
892 	    SYSCTL_CHILDREN(rack_sysctl_root),
893 	    OID_AUTO,
894 	    "sack_attack",
895 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
896 	    "Rack Sack Attack Counters and Controls");
897 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
898 	    SYSCTL_CHILDREN(rack_sysctl_root),
899 	    OID_AUTO,
900 	    "stats",
901 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
902 	    "Rack Counters");
903 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
904 	    SYSCTL_CHILDREN(rack_sysctl_root),
905 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
906 	    &rack_rate_sample_method , USE_RTT_LOW,
907 	    "What method should we use for rate sampling 0=high, 1=low ");
908 	/* Probe rtt related controls */
909 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
910 	    SYSCTL_CHILDREN(rack_sysctl_root),
911 	    OID_AUTO,
912 	    "probertt",
913 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
914 	    "ProbeRTT related Controls");
915 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
916 	    SYSCTL_CHILDREN(rack_probertt),
917 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
918 	    &rack_atexit_prtt_hbp, 130,
919 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
920 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
921 	    SYSCTL_CHILDREN(rack_probertt),
922 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
923 	    &rack_atexit_prtt, 130,
924 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
925 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
926 	    SYSCTL_CHILDREN(rack_probertt),
927 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
928 	    &rack_per_of_gp_probertt, 60,
929 	    "What percentage of goodput do we pace at in probertt");
930 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
931 	    SYSCTL_CHILDREN(rack_probertt),
932 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
933 	    &rack_per_of_gp_probertt_reduce, 10,
934 	    "What percentage of goodput do we reduce every gp_srtt");
935 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
936 	    SYSCTL_CHILDREN(rack_probertt),
937 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
938 	    &rack_per_of_gp_lowthresh, 40,
939 	    "What percentage of goodput do we allow the multiplier to fall to");
940 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
941 	    SYSCTL_CHILDREN(rack_probertt),
942 	    OID_AUTO, "time_between", CTLFLAG_RW,
943 	    & rack_time_between_probertt, 96000000,
944 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
945 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
946 	    SYSCTL_CHILDREN(rack_probertt),
947 	    OID_AUTO, "safety", CTLFLAG_RW,
948 	    &rack_probe_rtt_safety_val, 2000000,
949 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
950 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
951 	    SYSCTL_CHILDREN(rack_probertt),
952 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
953 	    &rack_probe_rtt_sets_cwnd, 0,
954 	    "Do we set the cwnd too (if always_lower is on)");
955 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
956 	    SYSCTL_CHILDREN(rack_probertt),
957 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
958 	    &rack_max_drain_wait, 2,
959 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
960 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
961 	    SYSCTL_CHILDREN(rack_probertt),
962 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
963 	    &rack_must_drain, 1,
964 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
965 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
966 	    SYSCTL_CHILDREN(rack_probertt),
967 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
968 	    &rack_probertt_use_min_rtt_entry, 1,
969 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
970 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
971 	    SYSCTL_CHILDREN(rack_probertt),
972 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
973 	    &rack_probertt_use_min_rtt_exit, 0,
974 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
975 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
976 	    SYSCTL_CHILDREN(rack_probertt),
977 	    OID_AUTO, "length_div", CTLFLAG_RW,
978 	    &rack_probertt_gpsrtt_cnt_div, 0,
979 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
980 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
981 	    SYSCTL_CHILDREN(rack_probertt),
982 	    OID_AUTO, "length_mul", CTLFLAG_RW,
983 	    &rack_probertt_gpsrtt_cnt_mul, 0,
984 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
985 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
986 	    SYSCTL_CHILDREN(rack_probertt),
987 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
988 	    &rack_min_probertt_hold, 200000,
989 	    "What is the minimum time we hold probertt at target");
990 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
991 	    SYSCTL_CHILDREN(rack_probertt),
992 	    OID_AUTO, "filter_life", CTLFLAG_RW,
993 	    &rack_probertt_filter_life, 10000000,
994 	    "What is the time for the filters life in useconds");
995 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
996 	    SYSCTL_CHILDREN(rack_probertt),
997 	    OID_AUTO, "lower_within", CTLFLAG_RW,
998 	    &rack_probertt_lower_within, 10,
999 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
1000 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1001 	    SYSCTL_CHILDREN(rack_probertt),
1002 	    OID_AUTO, "must_move", CTLFLAG_RW,
1003 	    &rack_min_rtt_movement, 250,
1004 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
1005 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1006 	    SYSCTL_CHILDREN(rack_probertt),
1007 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
1008 	    &rack_probertt_clear_is, 1,
1009 	    "Do we clear I/S counts on exiting probe-rtt");
1010 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1011 	    SYSCTL_CHILDREN(rack_probertt),
1012 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
1013 	    &rack_max_drain_hbp, 1,
1014 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
1015 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1016 	    SYSCTL_CHILDREN(rack_probertt),
1017 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
1018 	    &rack_hbp_thresh, 3,
1019 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
1020 	/* Pacing related sysctls */
1021 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1022 	    SYSCTL_CHILDREN(rack_sysctl_root),
1023 	    OID_AUTO,
1024 	    "pacing",
1025 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1026 	    "Pacing related Controls");
1027 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1028 	    SYSCTL_CHILDREN(rack_pacing),
1029 	    OID_AUTO, "pcm_enabled", CTLFLAG_RW,
1030 	    &rack_pcm_is_enabled, 1,
1031 	    "Do we by default do PCM measurements?");
1032 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1033 	    SYSCTL_CHILDREN(rack_pacing),
1034 	    OID_AUTO, "pcm_rnds", CTLFLAG_RW,
1035 	    &rack_pcm_every_n_rounds, 100,
1036 	    "How many rounds before we need to do a PCM measurement");
1037 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1038 	    SYSCTL_CHILDREN(rack_pacing),
1039 	    OID_AUTO, "pcm_blast", CTLFLAG_RW,
1040 	    &rack_pcm_blast, 0,
1041 	    "Blast out the full cwnd/rwnd when doing a PCM measurement");
1042 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1043 	    SYSCTL_CHILDREN(rack_pacing),
1044 	    OID_AUTO, "rnd_gp_gain", CTLFLAG_RW,
1045 	    &rack_gp_gain_req, 1200,
1046 	    "How much do we have to increase the GP to record the round 1200 = 120.0");
1047 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1048 	    SYSCTL_CHILDREN(rack_pacing),
1049 	    OID_AUTO, "dgp_out_of_ss_at", CTLFLAG_RW,
1050 	    &rack_rnd_cnt_req, 0x10005,
1051 	    "How many rounds less than rnd_gp_gain will drop us out of SS");
1052 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1053 	    SYSCTL_CHILDREN(rack_pacing),
1054 	    OID_AUTO, "no_timely", CTLFLAG_RW,
1055 	    &rack_timely_off, 0,
1056 	    "Do we not use timely in DGP?");
1057 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1058 	    SYSCTL_CHILDREN(rack_pacing),
1059 	    OID_AUTO, "fullbufdisc", CTLFLAG_RW,
1060 	    &rack_full_buffer_discount, 10,
1061 	    "What percentage b/w reduction over the GP estimate for a full buffer (default=0 off)?");
1062 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1063 	    SYSCTL_CHILDREN(rack_pacing),
1064 	    OID_AUTO, "fillcw", CTLFLAG_RW,
1065 	    &rack_fill_cw_state, 0,
1066 	    "Enable fillcw on new connections (default=0 off)?");
1067 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1068 	    SYSCTL_CHILDREN(rack_pacing),
1069 	    OID_AUTO, "min_burst", CTLFLAG_RW,
1070 	    &rack_pacing_min_seg, 0,
1071 	    "What is the min burst size for pacing (0 disables)?");
1072 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1073 	    SYSCTL_CHILDREN(rack_pacing),
1074 	    OID_AUTO, "divisor", CTLFLAG_RW,
1075 	    &rack_default_pacing_divisor, 250,
1076 	    "What is the default divisor given to the rl code?");
1077 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1078 	    SYSCTL_CHILDREN(rack_pacing),
1079 	    OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1080 	    &rack_bw_multipler, 0,
1081 	    "What is the limit multiplier of the current gp_est that fillcw can increase the b/w too, 200 == 200% (0 = off)?");
1082 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1083 	    SYSCTL_CHILDREN(rack_pacing),
1084 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
1085 	    &rack_max_per_above, 30,
1086 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1087 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1088 	    SYSCTL_CHILDREN(rack_pacing),
1089 	    OID_AUTO, "allow1mss", CTLFLAG_RW,
1090 	    &rack_pace_one_seg, 0,
1091 	    "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1092 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1093 	    SYSCTL_CHILDREN(rack_pacing),
1094 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1095 	    &rack_limit_time_with_srtt, 0,
1096 	    "Do we limit pacing time based on srtt");
1097 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1098 	    SYSCTL_CHILDREN(rack_pacing),
1099 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1100 	    &rack_per_of_gp_ss, 250,
1101 	    "If non zero, what percentage of goodput to pace at in slow start");
1102 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1103 	    SYSCTL_CHILDREN(rack_pacing),
1104 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1105 	    &rack_per_of_gp_ca, 150,
1106 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1107 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1108 	    SYSCTL_CHILDREN(rack_pacing),
1109 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1110 	    &rack_per_of_gp_rec, 200,
1111 	    "If non zero, what percentage of goodput to pace at in recovery");
1112 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1113 	    SYSCTL_CHILDREN(rack_pacing),
1114 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1115 	    &rack_hptsi_segments, 40,
1116 	    "What size is the max for TSO segments in pacing and burst mitigation");
1117 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1118 	    SYSCTL_CHILDREN(rack_pacing),
1119 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1120 	    &rack_slot_reduction, 4,
1121 	    "When doing only burst mitigation what is the reduce divisor");
1122 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1123 	    SYSCTL_CHILDREN(rack_sysctl_root),
1124 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1125 	    &rack_pace_every_seg, 0,
1126 	    "If set we use pacing, if clear we use only the original burst mitigation");
1127 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1128 	    SYSCTL_CHILDREN(rack_pacing),
1129 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1130 	    &rack_bw_rate_cap, 0,
1131 	    "If set we apply this value to the absolute rate cap used by pacing");
1132 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1133 	    SYSCTL_CHILDREN(rack_pacing),
1134 	    OID_AUTO, "fillcw_cap", CTLFLAG_RW,
1135 	    &rack_fillcw_bw_cap, 3750000,
1136 	    "Do we have an absolute cap on the amount of b/w fillcw can specify (0 = no)?");
1137 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1138 	    SYSCTL_CHILDREN(rack_sysctl_root),
1139 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1140 	    &rack_req_measurements, 1,
1141 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1142 	/* Hardware pacing */
1143 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1144 	    SYSCTL_CHILDREN(rack_sysctl_root),
1145 	    OID_AUTO,
1146 	    "hdwr_pacing",
1147 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1148 	    "Pacing related Controls");
1149 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1150 	    SYSCTL_CHILDREN(rack_hw_pacing),
1151 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1152 	    &rack_hw_rwnd_factor, 2,
1153 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1154 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1155 	    SYSCTL_CHILDREN(rack_hw_pacing),
1156 	    OID_AUTO, "precheck", CTLFLAG_RW,
1157 	    &rack_hw_check_queue, 0,
1158 	    "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1159 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1160 	    SYSCTL_CHILDREN(rack_hw_pacing),
1161 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1162 	    &rack_enobuf_hw_boost_mult, 0,
1163 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1164 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1165 	    SYSCTL_CHILDREN(rack_hw_pacing),
1166 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1167 	    &rack_enobuf_hw_max, 2,
1168 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1169 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1170 	    SYSCTL_CHILDREN(rack_hw_pacing),
1171 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1172 	    &rack_enobuf_hw_min, 2,
1173 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1174 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1175 	    SYSCTL_CHILDREN(rack_hw_pacing),
1176 	    OID_AUTO, "enable", CTLFLAG_RW,
1177 	    &rack_enable_hw_pacing, 0,
1178 	    "Should RACK attempt to use hw pacing?");
1179 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1180 	    SYSCTL_CHILDREN(rack_hw_pacing),
1181 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1182 	    &rack_hw_rate_caps, 0,
1183 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1184 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1185 	    SYSCTL_CHILDREN(rack_hw_pacing),
1186 	    OID_AUTO, "uncap_per", CTLFLAG_RW,
1187 	    &rack_hw_rate_cap_per, 0,
1188 	    "If you go over b/w by this amount you will be uncapped (0 = never)");
1189 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1190 	    SYSCTL_CHILDREN(rack_hw_pacing),
1191 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1192 	    &rack_hw_rate_min, 0,
1193 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1194 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1195 	    SYSCTL_CHILDREN(rack_hw_pacing),
1196 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1197 	    &rack_hw_rate_to_low, 0,
1198 	    "If we fall below this rate, dis-engage hw pacing?");
1199 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1200 	    SYSCTL_CHILDREN(rack_hw_pacing),
1201 	    OID_AUTO, "up_only", CTLFLAG_RW,
1202 	    &rack_hw_up_only, 0,
1203 	    "Do we allow hw pacing to lower the rate selected?");
1204 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1205 	    SYSCTL_CHILDREN(rack_hw_pacing),
1206 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1207 	    &rack_hw_pace_extra_slots, 0,
1208 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1209 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1210 	    SYSCTL_CHILDREN(rack_sysctl_root),
1211 	    OID_AUTO,
1212 	    "timely",
1213 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1214 	    "Rack Timely RTT Controls");
1215 	/* Timely based GP dynmics */
1216 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1217 	    SYSCTL_CHILDREN(rack_timely),
1218 	    OID_AUTO, "upper", CTLFLAG_RW,
1219 	    &rack_gp_per_bw_mul_up, 2,
1220 	    "Rack timely upper range for equal b/w (in percentage)");
1221 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1222 	    SYSCTL_CHILDREN(rack_timely),
1223 	    OID_AUTO, "lower", CTLFLAG_RW,
1224 	    &rack_gp_per_bw_mul_down, 4,
1225 	    "Rack timely lower range for equal b/w (in percentage)");
1226 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1227 	    SYSCTL_CHILDREN(rack_timely),
1228 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1229 	    &rack_gp_rtt_maxmul, 3,
1230 	    "Rack timely multiplier of lowest rtt for rtt_max");
1231 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1232 	    SYSCTL_CHILDREN(rack_timely),
1233 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1234 	    &rack_gp_rtt_mindiv, 4,
1235 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1236 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1237 	    SYSCTL_CHILDREN(rack_timely),
1238 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1239 	    &rack_gp_rtt_minmul, 1,
1240 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1241 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1242 	    SYSCTL_CHILDREN(rack_timely),
1243 	    OID_AUTO, "decrease", CTLFLAG_RW,
1244 	    &rack_gp_decrease_per, 80,
1245 	    "Rack timely Beta value 80 = .8 (scaled by 100)");
1246 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1247 	    SYSCTL_CHILDREN(rack_timely),
1248 	    OID_AUTO, "increase", CTLFLAG_RW,
1249 	    &rack_gp_increase_per, 2,
1250 	    "Rack timely increase perentage of our GP multiplication factor");
1251 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1252 	    SYSCTL_CHILDREN(rack_timely),
1253 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1254 	    &rack_per_lower_bound, 50,
1255 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1256 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1257 	    SYSCTL_CHILDREN(rack_timely),
1258 	    OID_AUTO, "p5_upper", CTLFLAG_RW,
1259 	    &rack_gain_p5_ub, 250,
1260 	    "Profile 5 upper bound to timely gain");
1261 
1262 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1263 	    SYSCTL_CHILDREN(rack_timely),
1264 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1265 	    &rack_per_upper_bound_ss, 0,
1266 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1267 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1268 	    SYSCTL_CHILDREN(rack_timely),
1269 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1270 	    &rack_per_upper_bound_ca, 0,
1271 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1272 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1273 	    SYSCTL_CHILDREN(rack_timely),
1274 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1275 	    &rack_do_dyn_mul, 0,
1276 	    "Rack timely do we enable dynmaic timely goodput by default");
1277 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1278 	    SYSCTL_CHILDREN(rack_timely),
1279 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1280 	    &rack_gp_no_rec_chg, 1,
1281 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1282 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1283 	    SYSCTL_CHILDREN(rack_timely),
1284 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1285 	    &rack_timely_dec_clear, 6,
1286 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1287 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1288 	    SYSCTL_CHILDREN(rack_timely),
1289 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1290 	    &rack_timely_max_push_rise, 3,
1291 	    "Rack timely how many times do we push up with b/w increase");
1292 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1293 	    SYSCTL_CHILDREN(rack_timely),
1294 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1295 	    &rack_timely_max_push_drop, 3,
1296 	    "Rack timely how many times do we push back on b/w decent");
1297 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1298 	    SYSCTL_CHILDREN(rack_timely),
1299 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1300 	    &rack_timely_min_segs, 4,
1301 	    "Rack timely when setting the cwnd what is the min num segments");
1302 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1303 	    SYSCTL_CHILDREN(rack_timely),
1304 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1305 	    &rack_use_max_for_nobackoff, 0,
1306 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1307 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1308 	    SYSCTL_CHILDREN(rack_timely),
1309 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1310 	    &rack_timely_int_timely_only, 0,
1311 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1312 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1313 	    SYSCTL_CHILDREN(rack_timely),
1314 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1315 	    &rack_timely_no_stopping, 0,
1316 	    "Rack timely don't stop increase");
1317 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1318 	    SYSCTL_CHILDREN(rack_timely),
1319 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1320 	    &rack_down_raise_thresh, 100,
1321 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1322 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1323 	    SYSCTL_CHILDREN(rack_timely),
1324 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1325 	    &rack_req_segs, 1,
1326 	    "Bottom dragging if not these many segments outstanding and room");
1327 
1328 	/* TLP and Rack related parameters */
1329 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1330 	    SYSCTL_CHILDREN(rack_sysctl_root),
1331 	    OID_AUTO,
1332 	    "tlp",
1333 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1334 	    "TLP and Rack related Controls");
1335 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1336 	    SYSCTL_CHILDREN(rack_tlp),
1337 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1338 	    &use_rack_rr, 1,
1339 	    "Do we use Rack Rapid Recovery");
1340 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1341 	    SYSCTL_CHILDREN(rack_tlp),
1342 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1343 	    &rack_max_abc_post_recovery, 2,
1344 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1345 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1346 	    SYSCTL_CHILDREN(rack_tlp),
1347 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1348 	    &rack_non_rxt_use_cr, 0,
1349 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1350 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1351 	    SYSCTL_CHILDREN(rack_tlp),
1352 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1353 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1354 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1355 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1356 	    SYSCTL_CHILDREN(rack_tlp),
1357 	    OID_AUTO, "limit", CTLFLAG_RW,
1358 	    &rack_tlp_limit, 2,
1359 	    "How many TLP's can be sent without sending new data");
1360 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1361 	    SYSCTL_CHILDREN(rack_tlp),
1362 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1363 	    &rack_tlp_use_greater, 1,
1364 	    "Should we use the rack_rtt time if its greater than srtt");
1365 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1366 	    SYSCTL_CHILDREN(rack_tlp),
1367 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1368 	    &rack_tlp_min, 10000,
1369 	    "TLP minimum timeout per the specification (in microseconds)");
1370 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1371 	    SYSCTL_CHILDREN(rack_tlp),
1372 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1373 	    &rack_always_send_oldest, 0,
1374 	    "Should we always send the oldest TLP and RACK-TLP");
1375 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1376 	    SYSCTL_CHILDREN(rack_tlp),
1377 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1378 	    &rack_lower_cwnd_at_tlp, 0,
1379 	    "When a TLP completes a retran should we enter recovery");
1380 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1381 	    SYSCTL_CHILDREN(rack_tlp),
1382 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1383 	    &rack_reorder_thresh, 2,
1384 	    "What factor for rack will be added when seeing reordering (shift right)");
1385 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1386 	    SYSCTL_CHILDREN(rack_tlp),
1387 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1388 	    &rack_tlp_thresh, 1,
1389 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1390 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1391 	    SYSCTL_CHILDREN(rack_tlp),
1392 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1393 	    &rack_reorder_fade, 60000000,
1394 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1395 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1396 	    SYSCTL_CHILDREN(rack_tlp),
1397 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1398 	    &rack_pkt_delay, 1000,
1399 	    "Extra RACK time (in microseconds) besides reordering thresh");
1400 
1401 	/* Timer related controls */
1402 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1403 	    SYSCTL_CHILDREN(rack_sysctl_root),
1404 	    OID_AUTO,
1405 	    "timers",
1406 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1407 	    "Timer related controls");
1408 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1409 	    SYSCTL_CHILDREN(rack_timers),
1410 	    OID_AUTO, "reset_ssth_rec_rto", CTLFLAG_RW,
1411 	    &rack_ssthresh_rest_rto_rec, 0,
1412 	    "When doing recovery -> rto -> recovery do we reset SSthresh?");
1413 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1414 	    SYSCTL_CHILDREN(rack_timers),
1415 	    OID_AUTO, "scoreboard_thresh", CTLFLAG_RW,
1416 	    &rack_rxt_scoreboard_clear_thresh, 2,
1417 	    "How many RTO's are allowed before we clear the scoreboard");
1418 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1419 	    SYSCTL_CHILDREN(rack_timers),
1420 	    OID_AUTO, "honor_hpts_min", CTLFLAG_RW,
1421 	    &rack_honors_hpts_min_to, 1,
1422 	    "Do rack pacing timers honor hpts min timeout");
1423 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1424 	    SYSCTL_CHILDREN(rack_timers),
1425 	    OID_AUTO, "hpts_max_reduce", CTLFLAG_RW,
1426 	    &rack_max_reduce, 10,
1427 	    "Max percentage we will reduce slot by for pacing when we are behind");
1428 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1429 	    SYSCTL_CHILDREN(rack_timers),
1430 	    OID_AUTO, "persmin", CTLFLAG_RW,
1431 	    &rack_persist_min, 250000,
1432 	    "What is the minimum time in microseconds between persists");
1433 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1434 	    SYSCTL_CHILDREN(rack_timers),
1435 	    OID_AUTO, "persmax", CTLFLAG_RW,
1436 	    &rack_persist_max, 2000000,
1437 	    "What is the largest delay in microseconds between persists");
1438 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1439 	    SYSCTL_CHILDREN(rack_timers),
1440 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1441 	    &rack_delayed_ack_time, 40000,
1442 	    "Delayed ack time (40ms in microseconds)");
1443 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1444 	    SYSCTL_CHILDREN(rack_timers),
1445 	    OID_AUTO, "minrto", CTLFLAG_RW,
1446 	    &rack_rto_min, 30000,
1447 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1448 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1449 	    SYSCTL_CHILDREN(rack_timers),
1450 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1451 	    &rack_rto_max, 4000000,
1452 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1453 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1454 	    SYSCTL_CHILDREN(rack_timers),
1455 	    OID_AUTO, "minto", CTLFLAG_RW,
1456 	    &rack_min_to, 1000,
1457 	    "Minimum rack timeout in microseconds");
1458 	/* Measure controls */
1459 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1460 	    SYSCTL_CHILDREN(rack_sysctl_root),
1461 	    OID_AUTO,
1462 	    "measure",
1463 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1464 	    "Measure related controls");
1465 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1466 	    SYSCTL_CHILDREN(rack_measure),
1467 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1468 	    &rack_wma_divisor, 8,
1469 	    "When doing b/w calculation what is the  divisor for the WMA");
1470 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1471 	    SYSCTL_CHILDREN(rack_measure),
1472 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1473 	    &rack_cwnd_block_ends_measure, 0,
1474 	    "Does a cwnd just-return end the measurement window (app limited)");
1475 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1476 	    SYSCTL_CHILDREN(rack_measure),
1477 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1478 	    &rack_rwnd_block_ends_measure, 0,
1479 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1480 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1481 	    SYSCTL_CHILDREN(rack_measure),
1482 	    OID_AUTO, "min_target", CTLFLAG_RW,
1483 	    &rack_def_data_window, 20,
1484 	    "What is the minimum target window (in mss) for a GP measurements");
1485 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1486 	    SYSCTL_CHILDREN(rack_measure),
1487 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1488 	    &rack_goal_bdp, 2,
1489 	    "What is the goal BDP to measure");
1490 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1491 	    SYSCTL_CHILDREN(rack_measure),
1492 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1493 	    &rack_min_srtts, 1,
1494 	    "What is the goal BDP to measure");
1495 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1496 	    SYSCTL_CHILDREN(rack_measure),
1497 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1498 	    &rack_min_measure_usec, 0,
1499 	    "What is the Minimum time time for a measurement if 0, this is off");
1500 	/* Features */
1501 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1502 	    SYSCTL_CHILDREN(rack_sysctl_root),
1503 	    OID_AUTO,
1504 	    "features",
1505 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1506 	    "Feature controls");
1507 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1508 	    SYSCTL_CHILDREN(rack_features),
1509 	    OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1510 	    &rack_hybrid_allow_set_maxseg, 0,
1511 	    "Should hybrid pacing allow the setmss command");
1512 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1513 	    SYSCTL_CHILDREN(rack_features),
1514 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1515 	    &rack_use_cmp_acks, 1,
1516 	    "Should RACK have LRO send compressed acks");
1517 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1518 	    SYSCTL_CHILDREN(rack_features),
1519 	    OID_AUTO, "fsb", CTLFLAG_RW,
1520 	    &rack_use_fsb, 1,
1521 	    "Should RACK use the fast send block?");
1522 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1523 	    SYSCTL_CHILDREN(rack_features),
1524 	    OID_AUTO, "rfo", CTLFLAG_RW,
1525 	    &rack_use_rfo, 1,
1526 	    "Should RACK use rack_fast_output()?");
1527 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1528 	    SYSCTL_CHILDREN(rack_features),
1529 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1530 	    &rack_use_rsm_rfo, 1,
1531 	    "Should RACK use rack_fast_rsm_output()?");
1532 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1533 	    SYSCTL_CHILDREN(rack_features),
1534 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1535 	    &rack_enable_mqueue_for_nonpaced, 0,
1536 	    "Should RACK use mbuf queuing for non-paced connections");
1537 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1538 	    SYSCTL_CHILDREN(rack_features),
1539 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1540 	    &rack_do_hystart, 0,
1541 	    "Should RACK enable HyStart++ on connections?");
1542 	/* Misc rack controls */
1543 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1544 	    SYSCTL_CHILDREN(rack_sysctl_root),
1545 	    OID_AUTO,
1546 	    "misc",
1547 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1548 	    "Misc related controls");
1549 #ifdef TCP_ACCOUNTING
1550 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1551 	    SYSCTL_CHILDREN(rack_misc),
1552 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1553 	    &rack_tcp_accounting, 0,
1554 	    "Should we turn on TCP accounting for all rack sessions?");
1555 #endif
1556 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1557 	    SYSCTL_CHILDREN(rack_misc),
1558 	    OID_AUTO, "dnd", CTLFLAG_RW,
1559 	    &rack_dnd_default, 0,
1560 	    "Do not disturb default for rack_rrr = 3");
1561 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1562 	    SYSCTL_CHILDREN(rack_misc),
1563 	    OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1564 	    &sad_seg_size_per, 800,
1565 	    "Percentage of segment size needed in a sack 800 = 80.0?");
1566 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1567 	    SYSCTL_CHILDREN(rack_misc),
1568 	    OID_AUTO, "rxt_controls", CTLFLAG_RW,
1569 	    &rack_rxt_controls, 0,
1570 	    "Retransmit sending size controls (valid  values 0, 1, 2 default=1)?");
1571 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1572 	    SYSCTL_CHILDREN(rack_misc),
1573 	    OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1574 	    &rack_hibeta_setting, 0,
1575 	    "Do we ue a high beta (80 instead of 50)?");
1576 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1577 	    SYSCTL_CHILDREN(rack_misc),
1578 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1579 	    &rack_apply_rtt_with_reduced_conf, 0,
1580 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1581 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1582 	    SYSCTL_CHILDREN(rack_misc),
1583 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1584 	    &rack_dsack_std_based, 3,
1585 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1586 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1587 	    SYSCTL_CHILDREN(rack_misc),
1588 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1589 	    &rack_prr_addbackmax, 2,
1590 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1591 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1592 	    SYSCTL_CHILDREN(rack_misc),
1593 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1594 	    &rack_stats_gets_ms_rtt, 1,
1595 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1596 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1597 	    SYSCTL_CHILDREN(rack_misc),
1598 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1599 	    &rack_client_low_buf, 0,
1600 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1601 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1602 	    SYSCTL_CHILDREN(rack_misc),
1603 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1604 	    &rack_def_profile, 0,
1605 	    "Should RACK use a default profile (0=no, num == profile num)?");
1606 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1607 	    SYSCTL_CHILDREN(rack_misc),
1608 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1609 	    &rack_enable_shared_cwnd, 1,
1610 	    "Should RACK try to use the shared cwnd on connections where allowed");
1611 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1612 	    SYSCTL_CHILDREN(rack_misc),
1613 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1614 	    &rack_limits_scwnd, 1,
1615 	    "Should RACK place low end time limits on the shared cwnd feature");
1616 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1617 	    SYSCTL_CHILDREN(rack_misc),
1618 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1619 	    &rack_disable_prr, 0,
1620 	    "Should RACK not use prr and only pace (must have pacing on)");
1621 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1622 	    SYSCTL_CHILDREN(rack_misc),
1623 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1624 	    &rack_verbose_logging, 0,
1625 	    "Should RACK black box logging be verbose");
1626 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1627 	    SYSCTL_CHILDREN(rack_misc),
1628 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1629 	    &rack_ignore_data_after_close, 1,
1630 	    "Do we hold off sending a RST until all pending data is ack'd");
1631 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1632 	    SYSCTL_CHILDREN(rack_misc),
1633 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1634 	    &rack_sack_not_required, 1,
1635 	    "Do we allow rack to run on connections not supporting SACK");
1636 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1637 	    SYSCTL_CHILDREN(rack_misc),
1638 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1639 	    &rack_send_a_lot_in_prr, 1,
1640 	    "Send a lot in prr");
1641 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1642 	    SYSCTL_CHILDREN(rack_misc),
1643 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1644 	    &rack_autosndbuf_inc, 20,
1645 	    "What percentage should rack scale up its snd buffer by?");
1646 
1647 
1648 	/* Sack Attacker detection stuff */
1649 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1650 	    SYSCTL_CHILDREN(rack_attack),
1651 	    OID_AUTO, "merge_out", CTLFLAG_RW,
1652 	    &rack_merge_out_sacks_on_attack, 0,
1653 	    "Do we merge the sendmap when we decide we are being attacked?");
1654 
1655 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1656 	    SYSCTL_CHILDREN(rack_attack),
1657 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1658 	    &rack_highest_sack_thresh_seen, 0,
1659 	    "Highest sack to ack ratio seen");
1660 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1661 	    SYSCTL_CHILDREN(rack_attack),
1662 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1663 	    &rack_highest_move_thresh_seen, 0,
1664 	    "Highest move to non-move ratio seen");
1665 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1666 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1667 	    SYSCTL_CHILDREN(rack_attack),
1668 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1669 	    &rack_ack_total,
1670 	    "Total number of Ack's");
1671 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1672 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1673 	    SYSCTL_CHILDREN(rack_attack),
1674 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1675 	    &rack_express_sack,
1676 	    "Total expresss number of Sack's");
1677 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1678 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1679 	    SYSCTL_CHILDREN(rack_attack),
1680 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1681 	    &rack_sack_total,
1682 	    "Total number of SACKs");
1683 	rack_move_none = counter_u64_alloc(M_WAITOK);
1684 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1685 	    SYSCTL_CHILDREN(rack_attack),
1686 	    OID_AUTO, "move_none", CTLFLAG_RD,
1687 	    &rack_move_none,
1688 	    "Total number of SACK index reuse of positions under threshold");
1689 	rack_move_some = counter_u64_alloc(M_WAITOK);
1690 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1691 	    SYSCTL_CHILDREN(rack_attack),
1692 	    OID_AUTO, "move_some", CTLFLAG_RD,
1693 	    &rack_move_some,
1694 	    "Total number of SACK index reuse of positions over threshold");
1695 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1696 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1697 	    SYSCTL_CHILDREN(rack_attack),
1698 	    OID_AUTO, "attacks", CTLFLAG_RD,
1699 	    &rack_sack_attacks_detected,
1700 	    "Total number of SACK attackers that had sack disabled");
1701 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1702 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1703 	    SYSCTL_CHILDREN(rack_attack),
1704 	    OID_AUTO, "reversed", CTLFLAG_RD,
1705 	    &rack_sack_attacks_reversed,
1706 	    "Total number of SACK attackers that were later determined false positive");
1707 	rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1708 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1709 	    SYSCTL_CHILDREN(rack_attack),
1710 	    OID_AUTO, "suspect", CTLFLAG_RD,
1711 	    &rack_sack_attacks_suspect,
1712 	    "Total number of SACKs that triggered early detection");
1713 
1714 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1715 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1716 	    SYSCTL_CHILDREN(rack_attack),
1717 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1718 	    &rack_sack_used_next_merge,
1719 	    "Total number of times we used the next merge");
1720 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1721 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1722 	    SYSCTL_CHILDREN(rack_attack),
1723 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1724 	    &rack_sack_used_prev_merge,
1725 	    "Total number of times we used the prev merge");
1726 	/* Counters */
1727 	rack_total_bytes = counter_u64_alloc(M_WAITOK);
1728 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1729 	    SYSCTL_CHILDREN(rack_counters),
1730 	    OID_AUTO, "totalbytes", CTLFLAG_RD,
1731 	    &rack_total_bytes,
1732 	    "Total number of bytes sent");
1733 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1734 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1735 	    SYSCTL_CHILDREN(rack_counters),
1736 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1737 	    &rack_fto_send, "Total number of rack_fast_output sends");
1738 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1739 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1740 	    SYSCTL_CHILDREN(rack_counters),
1741 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1742 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1743 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1744 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1745 	    SYSCTL_CHILDREN(rack_counters),
1746 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1747 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1748 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1749 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750 	    SYSCTL_CHILDREN(rack_counters),
1751 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1752 	    &rack_non_fto_send, "Total number of rack_output first sends");
1753 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1754 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1755 	    SYSCTL_CHILDREN(rack_counters),
1756 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1757 	    &rack_extended_rfo, "Total number of times we extended rfo");
1758 
1759 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1760 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1761 	    SYSCTL_CHILDREN(rack_counters),
1762 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1763 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1764 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1765 
1766 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1767 	    SYSCTL_CHILDREN(rack_counters),
1768 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1769 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1770 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1771 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1772 	    SYSCTL_CHILDREN(rack_counters),
1773 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1774 	    &rack_tlp_tot,
1775 	    "Total number of tail loss probe expirations");
1776 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1777 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1778 	    SYSCTL_CHILDREN(rack_counters),
1779 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1780 	    &rack_tlp_newdata,
1781 	    "Total number of tail loss probe sending new data");
1782 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1783 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1784 	    SYSCTL_CHILDREN(rack_counters),
1785 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1786 	    &rack_tlp_retran,
1787 	    "Total number of tail loss probe sending retransmitted data");
1788 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1789 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1790 	    SYSCTL_CHILDREN(rack_counters),
1791 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1792 	    &rack_tlp_retran_bytes,
1793 	    "Total bytes of tail loss probe sending retransmitted data");
1794 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1795 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1796 	    SYSCTL_CHILDREN(rack_counters),
1797 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1798 	    &rack_to_tot,
1799 	    "Total number of times the rack to expired");
1800 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1801 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1802 	    SYSCTL_CHILDREN(rack_counters),
1803 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1804 	    &rack_saw_enobuf,
1805 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1806 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1807 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1808 	    SYSCTL_CHILDREN(rack_counters),
1809 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1810 	    &rack_saw_enobuf_hw,
1811 	    "Total number of times a send returned enobuf for hdwr paced connections");
1812 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1813 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1814 	    SYSCTL_CHILDREN(rack_counters),
1815 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1816 	    &rack_saw_enetunreach,
1817 	    "Total number of times a send received a enetunreachable");
1818 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1819 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1820 	    SYSCTL_CHILDREN(rack_counters),
1821 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1822 	    &rack_hot_alloc,
1823 	    "Total allocations from the top of our list");
1824 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1825 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1826 	    SYSCTL_CHILDREN(rack_counters),
1827 	    OID_AUTO, "allocs", CTLFLAG_RD,
1828 	    &rack_to_alloc,
1829 	    "Total allocations of tracking structures");
1830 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1831 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1832 	    SYSCTL_CHILDREN(rack_counters),
1833 	    OID_AUTO, "allochard", CTLFLAG_RD,
1834 	    &rack_to_alloc_hard,
1835 	    "Total allocations done with sleeping the hard way");
1836 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1837 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1838 	    SYSCTL_CHILDREN(rack_counters),
1839 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1840 	    &rack_to_alloc_emerg,
1841 	    "Total allocations done from emergency cache");
1842 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1843 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1844 	    SYSCTL_CHILDREN(rack_counters),
1845 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1846 	    &rack_to_alloc_limited,
1847 	    "Total allocations dropped due to limit");
1848 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1849 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1850 	    SYSCTL_CHILDREN(rack_counters),
1851 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1852 	    &rack_alloc_limited_conns,
1853 	    "Connections with allocations dropped due to limit");
1854 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1855 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1856 	    SYSCTL_CHILDREN(rack_counters),
1857 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1858 	    &rack_split_limited,
1859 	    "Split allocations dropped due to limit");
1860 	rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1861 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1862 	    SYSCTL_CHILDREN(rack_counters),
1863 	    OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1864 	    &rack_rxt_clamps_cwnd,
1865 	    "Number of times that excessive rxt clamped the cwnd down");
1866 	rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1867 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1868 	    SYSCTL_CHILDREN(rack_counters),
1869 	    OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1870 	    &rack_rxt_clamps_cwnd_uniq,
1871 	    "Number of connections that have had excessive rxt clamped the cwnd down");
1872 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1873 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1874 	    SYSCTL_CHILDREN(rack_counters),
1875 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1876 	    &rack_persists_sends,
1877 	    "Number of times we sent a persist probe");
1878 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1879 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1880 	    SYSCTL_CHILDREN(rack_counters),
1881 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1882 	    &rack_persists_acks,
1883 	    "Number of times a persist probe was acked");
1884 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1885 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1886 	    SYSCTL_CHILDREN(rack_counters),
1887 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1888 	    &rack_persists_loss,
1889 	    "Number of times we detected a lost persist probe (no ack)");
1890 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1891 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1892 	    SYSCTL_CHILDREN(rack_counters),
1893 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1894 	    &rack_persists_lost_ends,
1895 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1896 #ifdef INVARIANTS
1897 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1898 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1899 	    SYSCTL_CHILDREN(rack_counters),
1900 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1901 	    &rack_adjust_map_bw,
1902 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1903 #endif
1904 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1905 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1906 	    SYSCTL_CHILDREN(rack_counters),
1907 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1908 	    &rack_multi_single_eq,
1909 	    "Number of compressed acks total represented");
1910 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1911 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1912 	    SYSCTL_CHILDREN(rack_counters),
1913 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1914 	    &rack_proc_non_comp_ack,
1915 	    "Number of non compresseds acks that we processed");
1916 
1917 
1918 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1919 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1920 	    SYSCTL_CHILDREN(rack_counters),
1921 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1922 	    &rack_sack_proc_all,
1923 	    "Total times we had to walk whole list for sack processing");
1924 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1925 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1926 	    SYSCTL_CHILDREN(rack_counters),
1927 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1928 	    &rack_sack_proc_restart,
1929 	    "Total times we had to walk whole list due to a restart");
1930 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1931 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1932 	    SYSCTL_CHILDREN(rack_counters),
1933 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1934 	    &rack_sack_proc_short,
1935 	    "Total times we took shortcut for sack processing");
1936 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1937 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1938 	    SYSCTL_CHILDREN(rack_attack),
1939 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1940 	    &rack_sack_skipped_acked,
1941 	    "Total number of times we skipped previously sacked");
1942 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1943 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1944 	    SYSCTL_CHILDREN(rack_attack),
1945 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1946 	    &rack_sack_splits,
1947 	    "Total number of times we did the old fashion tree split");
1948 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1949 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1950 	    SYSCTL_CHILDREN(rack_counters),
1951 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1952 	    &rack_input_idle_reduces,
1953 	    "Total number of idle reductions on input");
1954 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1955 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1956 	    SYSCTL_CHILDREN(rack_counters),
1957 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1958 	    &rack_collapsed_win_seen,
1959 	    "Total number of collapsed window events seen (where our window shrinks)");
1960 
1961 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1962 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1963 	    SYSCTL_CHILDREN(rack_counters),
1964 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1965 	    &rack_collapsed_win,
1966 	    "Total number of collapsed window events where we mark packets");
1967 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1968 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1969 	    SYSCTL_CHILDREN(rack_counters),
1970 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1971 	    &rack_collapsed_win_rxt,
1972 	    "Total number of packets that were retransmitted");
1973 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1974 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1975 	    SYSCTL_CHILDREN(rack_counters),
1976 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1977 	    &rack_collapsed_win_rxt_bytes,
1978 	    "Total number of bytes that were retransmitted");
1979 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1980 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1981 	    SYSCTL_CHILDREN(rack_counters),
1982 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1983 	    &rack_try_scwnd,
1984 	    "Total number of scwnd attempts");
1985 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1986 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1987 	    OID_AUTO, "outsize", CTLFLAG_RD,
1988 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1989 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1990 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1991 	    OID_AUTO, "opts", CTLFLAG_RD,
1992 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1993 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1994 	    SYSCTL_CHILDREN(rack_sysctl_root),
1995 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1996 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1997 }
1998 
1999 static uint32_t
2000 rc_init_window(struct tcp_rack *rack)
2001 {
2002 	return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
2003 
2004 }
2005 
2006 static uint64_t
2007 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
2008 {
2009 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
2010 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
2011 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2012 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
2013 	else
2014 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
2015 }
2016 
2017 static void
2018 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
2019 	uint64_t data, uint8_t mod, uint16_t aux,
2020 	struct tcp_sendfile_track *cur, int line)
2021 {
2022 #ifdef TCP_REQUEST_TRK
2023 	int do_log = 0;
2024 
2025 	/*
2026 	 * The rate cap one is noisy and only should come out when normal BB logging
2027 	 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
2028 	 * once per chunk and make up the BBpoint that can be turned on by the client.
2029 	 */
2030 	if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2031 		/*
2032 		 * The very noisy two need to only come out when
2033 		 * we have verbose logging on.
2034 		 */
2035 		if (rack_verbose_logging != 0)
2036 			do_log = tcp_bblogging_on(rack->rc_tp);
2037 		else
2038 			do_log = 0;
2039 	} else if (mod != HYBRID_LOG_BW_MEASURE) {
2040 		/*
2041 		 * All other less noisy logs here except the measure which
2042 		 * also needs to come out on the point and the log.
2043 		 */
2044 		do_log = tcp_bblogging_on(rack->rc_tp);
2045 	} else {
2046 		do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2047 	}
2048 
2049 	if (do_log) {
2050 		union tcp_log_stackspecific log;
2051 		struct timeval tv;
2052 		uint64_t lt_bw;
2053 
2054 		/* Convert our ms to a microsecond */
2055 		memset(&log, 0, sizeof(log));
2056 
2057 		log.u_bbr.cwnd_gain = line;
2058 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2059 		log.u_bbr.rttProp = tim;
2060 		log.u_bbr.bw_inuse = cbw;
2061 		log.u_bbr.delRate = rack_get_gp_est(rack);
2062 		lt_bw = rack_get_lt_bw(rack);
2063 		log.u_bbr.flex1 = seq;
2064 		log.u_bbr.pacing_gain = aux;
2065 		/* lt_bw = < flex3 | flex2 > */
2066 		log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2067 		log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2068 		/* Record the last obtained us rtt in inflight */
2069 		if (cur == NULL) {
2070 			/* Make sure we are looking at the right log if an overide comes in */
2071 			cur = rack->r_ctl.rc_last_sft;
2072 		}
2073 		if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2074 			log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2075 		else {
2076 			/* Use the last known rtt i.e. the rack-rtt */
2077 			log.u_bbr.inflight = rack->rc_rack_rtt;
2078 		}
2079 		if (cur != NULL) {
2080 			uint64_t off;
2081 
2082 			log.u_bbr.cur_del_rate = cur->deadline;
2083 			if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2084 				/* start = < lost | pkt_epoch > */
2085 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2086 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2087 				log.u_bbr.flex6 = cur->start_seq;
2088 				log.u_bbr.pkts_out = cur->end_seq;
2089 			} else {
2090 				/* start = < lost | pkt_epoch > */
2091 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2092 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2093 				/* end = < pkts_out | flex6 > */
2094 				log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2095 				log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2096 			}
2097 			/* first_send = <lt_epoch | epoch> */
2098 			log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2099 			log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2100 			/* localtime = <delivered | applimited>*/
2101 			log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2102 			log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2103 #ifdef TCP_REQUEST_TRK
2104 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2105 			log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2106 #endif
2107 			log.u_bbr.inhpts = 1;
2108 			log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2109 			log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2110 			log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2111 		} else {
2112 			log.u_bbr.flex7 = 0xffff;
2113 			log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2114 		}
2115 		/*
2116 		 * Compose bbr_state to be a bit wise 0000ADHF
2117 		 * where A is the always_pace flag
2118 		 * where D is the dgp_on flag
2119 		 * where H is the hybrid_mode on flag
2120 		 * where F is the use_fixed_rate flag.
2121 		 */
2122 		log.u_bbr.bbr_state = rack->rc_always_pace;
2123 		log.u_bbr.bbr_state <<= 1;
2124 		log.u_bbr.bbr_state |= rack->dgp_on;
2125 		log.u_bbr.bbr_state <<= 1;
2126 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2127 		log.u_bbr.bbr_state <<= 1;
2128 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2129 		log.u_bbr.flex8 = mod;
2130 		tcp_log_event(rack->rc_tp, NULL,
2131 		    &rack->rc_inp->inp_socket->so_rcv,
2132 		    &rack->rc_inp->inp_socket->so_snd,
2133 		    TCP_HYBRID_PACING_LOG, 0,
2134 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2135 
2136 	}
2137 #endif
2138 }
2139 
2140 #ifdef TCP_REQUEST_TRK
2141 static void
2142 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2143 {
2144 	if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2145 		union tcp_log_stackspecific log;
2146 		struct timeval tv;
2147 		uint64_t off;
2148 
2149 		/* Convert our ms to a microsecond */
2150 		memset(&log, 0, sizeof(log));
2151 
2152 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2153 		log.u_bbr.delRate = cur->sent_at_fs;
2154 
2155 		if ((cur->flags & TCP_TRK_TRACK_FLG_LSND) == 0) {
2156 			/*
2157 			 * We did not get a new Rules Applied to set so
2158 			 * no overlapping send occured, this means the
2159 			 * current byte counts are correct.
2160 			 */
2161 			log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2162 			log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2163 		} else {
2164 			/*
2165 			 * Overlapping send case, we switched to a new
2166 			 * send and did a rules applied.
2167 			 */
2168 			log.u_bbr.cur_del_rate = cur->sent_at_ls;
2169 			log.u_bbr.rttProp = cur->rxt_at_ls;
2170 		}
2171 		log.u_bbr.bw_inuse = cur->rxt_at_fs;
2172 		log.u_bbr.cwnd_gain = line;
2173 		off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2174 		log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2175 		/* start = < flex1 | flex2 > */
2176 		log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2177 		log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2178 		/* end = < flex3 | flex4 > */
2179 		log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2180 		log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2181 
2182 		/* localtime = <delivered | applimited>*/
2183 		log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2184 		log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2185 		/* client timestamp = <lt_epoch | epoch>*/
2186 		log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2187 		log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2188 		/* now set all the flags in */
2189 		log.u_bbr.pkts_out = cur->hybrid_flags;
2190 		log.u_bbr.lost = cur->playout_ms;
2191 		log.u_bbr.flex6 = cur->flags;
2192 		/*
2193 		 * Last send time  = <flex5 | pkt_epoch>  note we do not distinguish cases
2194 		 * where a false retransmit occurred so first_send  <-> lastsend may
2195 		 * include longer time then it actually took if we have a false rxt.
2196 		 */
2197 		log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2198 		log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2199 		/*
2200 		 * Compose bbr_state to be a bit wise 0000ADHF
2201 		 * where A is the always_pace flag
2202 		 * where D is the dgp_on flag
2203 		 * where H is the hybrid_mode on flag
2204 		 * where F is the use_fixed_rate flag.
2205 		 */
2206 		log.u_bbr.bbr_state = rack->rc_always_pace;
2207 		log.u_bbr.bbr_state <<= 1;
2208 		log.u_bbr.bbr_state |= rack->dgp_on;
2209 		log.u_bbr.bbr_state <<= 1;
2210 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2211 		log.u_bbr.bbr_state <<= 1;
2212 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2213 
2214 		log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2215 		tcp_log_event(rack->rc_tp, NULL,
2216 		    &rack->rc_inp->inp_socket->so_rcv,
2217 		    &rack->rc_inp->inp_socket->so_snd,
2218 		    TCP_HYBRID_PACING_LOG, 0,
2219 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2220 	}
2221 }
2222 #endif
2223 
2224 static inline uint64_t
2225 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2226 {
2227 	uint64_t ret_bw, ether;
2228 	uint64_t u_segsiz;
2229 
2230 	ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2231 	if (rack->r_is_v6){
2232 #ifdef INET6
2233 		ether += sizeof(struct ip6_hdr);
2234 #endif
2235 		ether += 14;	/* eheader size 6+6+2 */
2236 	} else {
2237 #ifdef INET
2238 		ether += sizeof(struct ip);
2239 #endif
2240 		ether += 14;	/* eheader size 6+6+2 */
2241 	}
2242 	u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2243 	ret_bw = bw;
2244 	ret_bw *= ether;
2245 	ret_bw /= u_segsiz;
2246 	return (ret_bw);
2247 }
2248 
2249 static void
2250 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2251 {
2252 #ifdef TCP_REQUEST_TRK
2253 	struct timeval tv;
2254 	uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2255 #endif
2256 
2257 	if (rack->r_ctl.bw_rate_cap == 0)
2258 		return;
2259 #ifdef TCP_REQUEST_TRK
2260 	if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2261 	    (rack->r_ctl.rc_last_sft != NULL)) {
2262 		/*
2263 		 * We have a dynamic cap. The original target
2264 		 * is in bw_rate_cap, but we need to look at
2265 		 * how long it is until we hit the deadline.
2266 		 */
2267 		struct tcp_sendfile_track *ent;
2268 
2269       		ent = rack->r_ctl.rc_last_sft;
2270 		microuptime(&tv);
2271 		timenow = tcp_tv_to_lusectick(&tv);
2272 		if (timenow >= ent->deadline) {
2273 			/* No time left we do DGP only */
2274 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2275 					   0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2276 			rack->r_ctl.bw_rate_cap = 0;
2277 			return;
2278 		}
2279 		/* We have the time */
2280 		timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2281 		if (timeleft < HPTS_MSEC_IN_SEC) {
2282 			/* If there is less than a ms left just use DGPs rate */
2283 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2284 					   0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2285 			rack->r_ctl.bw_rate_cap = 0;
2286 			return;
2287 		}
2288 		/*
2289 		 * Now lets find the amount of data left to send.
2290 		 *
2291 		 * Now ideally we want to use the end_seq to figure out how much more
2292 		 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2293 		 */
2294 		if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2295 			if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2296 				lenleft = ent->end_seq - rack->rc_tp->snd_una;
2297 			else {
2298 				/* TSNH, we should catch it at the send */
2299 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2300 						   0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2301 				rack->r_ctl.bw_rate_cap = 0;
2302 				return;
2303 			}
2304 		} else {
2305 			/*
2306 			 * The hard way, figure out how much is gone and then
2307 			 * take that away from the total the client asked for
2308 			 * (thats off by tls overhead if this is tls).
2309 			 */
2310 			if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2311 				lengone = rack->rc_tp->snd_una - ent->start_seq;
2312 			else
2313 				lengone = 0;
2314 			if (lengone < (ent->end - ent->start))
2315 				lenleft = (ent->end - ent->start) - lengone;
2316 			else {
2317 				/* TSNH, we should catch it at the send */
2318 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2319 						   0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2320 				rack->r_ctl.bw_rate_cap = 0;
2321 				return;
2322 			}
2323 		}
2324 		if (lenleft == 0) {
2325 			/* We have it all sent */
2326 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2327 					   0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2328 			if (rack->r_ctl.bw_rate_cap)
2329 				goto normal_ratecap;
2330 			else
2331 				return;
2332 		}
2333 		calcbw = lenleft * HPTS_USEC_IN_SEC;
2334 		calcbw /= timeleft;
2335 		/* Now we must compensate for IP/TCP overhead */
2336 		calcbw = rack_compensate_for_linerate(rack, calcbw);
2337 		/* Update the bit rate cap */
2338 		rack->r_ctl.bw_rate_cap = calcbw;
2339 		if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2340 		    (rack_hybrid_allow_set_maxseg == 1) &&
2341 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2342 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2343 			uint32_t orig_max;
2344 
2345 			orig_max = rack->r_ctl.rc_pace_max_segs;
2346 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2347 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2348 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2349 		}
2350 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2351 				   calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2352 		if ((calcbw > 0) && (*bw > calcbw)) {
2353 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2354 					   *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2355 			*capped = 1;
2356 			*bw = calcbw;
2357 		}
2358 		return;
2359 	}
2360 normal_ratecap:
2361 #endif
2362 	if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2363 #ifdef TCP_REQUEST_TRK
2364 		if (rack->rc_hybrid_mode &&
2365 		    rack->rc_catch_up &&
2366 		    (rack->r_ctl.rc_last_sft != NULL) &&
2367 		    (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2368 		    (rack_hybrid_allow_set_maxseg == 1) &&
2369 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2370 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2371 			uint32_t orig_max;
2372 
2373 			orig_max = rack->r_ctl.rc_pace_max_segs;
2374 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2375 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2376 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2377 		}
2378 #endif
2379 		*capped = 1;
2380 		*bw = rack->r_ctl.bw_rate_cap;
2381 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2382 				   *bw, 0, 0,
2383 				   HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2384 	}
2385 }
2386 
2387 static uint64_t
2388 rack_get_gp_est(struct tcp_rack *rack)
2389 {
2390 	uint64_t bw, lt_bw, ret_bw;
2391 
2392 	if (rack->rc_gp_filled == 0) {
2393 		/*
2394 		 * We have yet no b/w measurement,
2395 		 * if we have a user set initial bw
2396 		 * return it. If we don't have that and
2397 		 * we have an srtt, use the tcp IW (10) to
2398 		 * calculate a fictional b/w over the SRTT
2399 		 * which is more or less a guess. Note
2400 		 * we don't use our IW from rack on purpose
2401 		 * so if we have like IW=30, we are not
2402 		 * calculating a "huge" b/w.
2403 		 */
2404 		uint64_t srtt;
2405 
2406 		if (rack->dis_lt_bw == 1)
2407 			lt_bw = 0;
2408 		else
2409 			lt_bw = rack_get_lt_bw(rack);
2410 		if (lt_bw) {
2411 			/*
2412 			 * No goodput bw but a long-term b/w does exist
2413 			 * lets use that.
2414 			 */
2415 			ret_bw = lt_bw;
2416 			goto compensate;
2417 		}
2418 		if (rack->r_ctl.init_rate)
2419 			return (rack->r_ctl.init_rate);
2420 
2421 		/* Ok lets come up with the IW guess, if we have a srtt */
2422 		if (rack->rc_tp->t_srtt == 0) {
2423 			/*
2424 			 * Go with old pacing method
2425 			 * i.e. burst mitigation only.
2426 			 */
2427 			return (0);
2428 		}
2429 		/* Ok lets get the initial TCP win (not racks) */
2430 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2431 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2432 		bw *= (uint64_t)USECS_IN_SECOND;
2433 		bw /= srtt;
2434 		ret_bw = bw;
2435 		goto compensate;
2436 
2437 	}
2438 	if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2439 		/* Averaging is done, we can return the value */
2440 		bw = rack->r_ctl.gp_bw;
2441 	} else {
2442 		/* Still doing initial average must calculate */
2443 		bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2444 	}
2445 	if (rack->dis_lt_bw) {
2446 		/* We are not using lt-bw */
2447 		ret_bw = bw;
2448 		goto compensate;
2449 	}
2450 	lt_bw = rack_get_lt_bw(rack);
2451 	if (lt_bw == 0) {
2452 		/* If we don't have one then equate it to the gp_bw */
2453 		lt_bw = rack->r_ctl.gp_bw;
2454 	}
2455 	if (rack->use_lesser_lt_bw) {
2456 		if (lt_bw < bw)
2457 			ret_bw = lt_bw;
2458 		else
2459 			ret_bw = bw;
2460 	} else {
2461 		if (lt_bw > bw)
2462 			ret_bw = lt_bw;
2463 		else
2464 			ret_bw = bw;
2465 	}
2466 	/*
2467 	 * Now lets compensate based on the TCP/IP overhead. Our
2468 	 * Goodput estimate does not include this so we must pace out
2469 	 * a bit faster since our pacing calculations do. The pacing
2470 	 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2471 	 * we are using to do this, so we do that here in the opposite
2472 	 * direction as well. This means that if we are tunneled and the
2473 	 * segsiz is say 1200 bytes we will get quite a boost, but its
2474 	 * compensated for in the pacing time the opposite way.
2475 	 */
2476 compensate:
2477 	ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2478 	return(ret_bw);
2479 }
2480 
2481 
2482 static uint64_t
2483 rack_get_bw(struct tcp_rack *rack)
2484 {
2485 	uint64_t bw;
2486 
2487 	if (rack->use_fixed_rate) {
2488 		/* Return the fixed pacing rate */
2489 		return (rack_get_fixed_pacing_bw(rack));
2490 	}
2491 	bw = rack_get_gp_est(rack);
2492 	return (bw);
2493 }
2494 
2495 static uint16_t
2496 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2497 {
2498 	if (rack->use_fixed_rate) {
2499 		return (100);
2500 	} else if (rack->in_probe_rtt && (rsm == NULL))
2501 		return (rack->r_ctl.rack_per_of_gp_probertt);
2502 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2503 		  rack->r_ctl.rack_per_of_gp_rec)) {
2504 		if (rsm) {
2505 			/* a retransmission always use the recovery rate */
2506 			return (rack->r_ctl.rack_per_of_gp_rec);
2507 		} else if (rack->rack_rec_nonrxt_use_cr) {
2508 			/* Directed to use the configured rate */
2509 			goto configured_rate;
2510 		} else if (rack->rack_no_prr &&
2511 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2512 			/* No PRR, lets just use the b/w estimate only */
2513 			return (100);
2514 		} else {
2515 			/*
2516 			 * Here we may have a non-retransmit but we
2517 			 * have no overrides, so just use the recovery
2518 			 * rate (prr is in effect).
2519 			 */
2520 			return (rack->r_ctl.rack_per_of_gp_rec);
2521 		}
2522 	}
2523 configured_rate:
2524 	/* For the configured rate we look at our cwnd vs the ssthresh */
2525 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2526 		return (rack->r_ctl.rack_per_of_gp_ss);
2527 	else
2528 		return (rack->r_ctl.rack_per_of_gp_ca);
2529 }
2530 
2531 static void
2532 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2533 {
2534 	/*
2535 	 * Types of logs (mod value)
2536 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2537 	 * 2 = a dsack round begins, persist is reset to 16.
2538 	 * 3 = a dsack round ends
2539 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2540 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2541 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2542 	 */
2543 	if (tcp_bblogging_on(rack->rc_tp)) {
2544 		union tcp_log_stackspecific log;
2545 		struct timeval tv;
2546 
2547 		memset(&log, 0, sizeof(log));
2548 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2549 		log.u_bbr.flex1 <<= 1;
2550 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2551 		log.u_bbr.flex1 <<= 1;
2552 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2553 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2554 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2555 		log.u_bbr.flex4 = flex4;
2556 		log.u_bbr.flex5 = flex5;
2557 		log.u_bbr.flex6 = flex6;
2558 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2559 		log.u_bbr.flex8 = mod;
2560 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2561 		log.u_bbr.epoch = rack->r_ctl.current_round;
2562 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2563 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2564 		    &rack->rc_inp->inp_socket->so_rcv,
2565 		    &rack->rc_inp->inp_socket->so_snd,
2566 		    RACK_DSACK_HANDLING, 0,
2567 		    0, &log, false, &tv);
2568 	}
2569 }
2570 
2571 static void
2572 rack_log_hdwr_pacing(struct tcp_rack *rack,
2573 		     uint64_t rate, uint64_t hw_rate, int line,
2574 		     int error, uint16_t mod)
2575 {
2576 	if (tcp_bblogging_on(rack->rc_tp)) {
2577 		union tcp_log_stackspecific log;
2578 		struct timeval tv;
2579 		const struct ifnet *ifp;
2580 		uint64_t ifp64;
2581 
2582 		memset(&log, 0, sizeof(log));
2583 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2584 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2585 		if (rack->r_ctl.crte) {
2586 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2587 		} else if (rack->rc_inp->inp_route.ro_nh &&
2588 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2589 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2590 		} else
2591 			ifp = NULL;
2592 		if (ifp) {
2593 			ifp64 = (uintptr_t)ifp;
2594 			log.u_bbr.flex3 = ((ifp64  >> 32) & 0x00000000ffffffff);
2595 			log.u_bbr.flex4 = (ifp64 & 0x00000000ffffffff);
2596 		}
2597 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2598 		log.u_bbr.bw_inuse = rate;
2599 		log.u_bbr.flex5 = line;
2600 		log.u_bbr.flex6 = error;
2601 		log.u_bbr.flex7 = mod;
2602 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2603 		log.u_bbr.flex8 = rack->use_fixed_rate;
2604 		log.u_bbr.flex8 <<= 1;
2605 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2606 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2607 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2608 		if (rack->r_ctl.crte)
2609 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2610 		else
2611 			log.u_bbr.cur_del_rate = 0;
2612 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2613 		log.u_bbr.epoch = rack->r_ctl.current_round;
2614 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2615 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2616 		    &rack->rc_inp->inp_socket->so_rcv,
2617 		    &rack->rc_inp->inp_socket->so_snd,
2618 		    BBR_LOG_HDWR_PACE, 0,
2619 		    0, &log, false, &tv);
2620 	}
2621 }
2622 
2623 static uint64_t
2624 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2625 {
2626 	/*
2627 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2628 	 */
2629 	uint64_t bw_est, high_rate;
2630 	uint64_t gain;
2631 
2632 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2633 	bw_est = bw * gain;
2634 	bw_est /= (uint64_t)100;
2635 	/* Never fall below the minimum (def 64kbps) */
2636 	if (bw_est < RACK_MIN_BW)
2637 		bw_est = RACK_MIN_BW;
2638 	if (rack->r_rack_hw_rate_caps) {
2639 		/* Rate caps are in place */
2640 		if (rack->r_ctl.crte != NULL) {
2641 			/* We have a hdwr rate already */
2642 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2643 			if (bw_est >= high_rate) {
2644 				/* We are capping bw at the highest rate table entry */
2645 				if (rack_hw_rate_cap_per &&
2646 				    (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2647 					rack->r_rack_hw_rate_caps = 0;
2648 					goto done;
2649 				}
2650 				rack_log_hdwr_pacing(rack,
2651 						     bw_est, high_rate, __LINE__,
2652 						     0, 3);
2653 				bw_est = high_rate;
2654 				if (capped)
2655 					*capped = 1;
2656 			}
2657 		} else if ((rack->rack_hdrw_pacing == 0) &&
2658 			   (rack->rack_hdw_pace_ena) &&
2659 			   (rack->rack_attempt_hdwr_pace == 0) &&
2660 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2661 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2662 			/*
2663 			 * Special case, we have not yet attempted hardware
2664 			 * pacing, and yet we may, when we do, find out if we are
2665 			 * above the highest rate. We need to know the maxbw for the interface
2666 			 * in question (if it supports ratelimiting). We get back
2667 			 * a 0, if the interface is not found in the RL lists.
2668 			 */
2669 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2670 			if (high_rate) {
2671 				/* Yep, we have a rate is it above this rate? */
2672 				if (bw_est > high_rate) {
2673 					bw_est = high_rate;
2674 					if (capped)
2675 						*capped = 1;
2676 				}
2677 			}
2678 		}
2679 	}
2680 done:
2681 	return (bw_est);
2682 }
2683 
2684 static void
2685 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2686 {
2687 	if (tcp_bblogging_on(rack->rc_tp)) {
2688 		union tcp_log_stackspecific log;
2689 		struct timeval tv;
2690 
2691 		if ((mod != 1) && (rack_verbose_logging == 0))  {
2692 			/*
2693 			 * We get 3 values currently for mod
2694 			 * 1 - We are retransmitting and this tells the reason.
2695 			 * 2 - We are clearing a dup-ack count.
2696 			 * 3 - We are incrementing a dup-ack count.
2697 			 *
2698 			 * The clear/increment are only logged
2699 			 * if you have BBverbose on.
2700 			 */
2701 			return;
2702 		}
2703 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2704 		log.u_bbr.flex1 = tsused;
2705 		log.u_bbr.flex2 = thresh;
2706 		log.u_bbr.flex3 = rsm->r_flags;
2707 		log.u_bbr.flex4 = rsm->r_dupack;
2708 		log.u_bbr.flex5 = rsm->r_start;
2709 		log.u_bbr.flex6 = rsm->r_end;
2710 		log.u_bbr.flex8 = mod;
2711 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2712 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2713 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2714 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2715 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2716 		log.u_bbr.pacing_gain = rack->r_must_retran;
2717 		log.u_bbr.epoch = rack->r_ctl.current_round;
2718 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2719 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2720 		    &rack->rc_inp->inp_socket->so_rcv,
2721 		    &rack->rc_inp->inp_socket->so_snd,
2722 		    BBR_LOG_SETTINGS_CHG, 0,
2723 		    0, &log, false, &tv);
2724 	}
2725 }
2726 
2727 static void
2728 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2729 {
2730 	if (tcp_bblogging_on(rack->rc_tp)) {
2731 		union tcp_log_stackspecific log;
2732 		struct timeval tv;
2733 
2734 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2735 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2736 		log.u_bbr.flex2 = to;
2737 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2738 		log.u_bbr.flex4 = slot;
2739 		log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2740 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2741 		log.u_bbr.flex7 = rack->rc_in_persist;
2742 		log.u_bbr.flex8 = which;
2743 		if (rack->rack_no_prr)
2744 			log.u_bbr.pkts_out = 0;
2745 		else
2746 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2747 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2748 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2749 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2750 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2751 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2752 		log.u_bbr.pacing_gain = rack->r_must_retran;
2753 		log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2754 		log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2755 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2756 		log.u_bbr.lost = rack_rto_min;
2757 		log.u_bbr.epoch = rack->r_ctl.roundends;
2758 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2759 		log.u_bbr.bw_inuse <<= 32;
2760 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2761 		log.u_bbr.applimited = rack->rc_tp->t_flags2;
2762 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2763 		    &rack->rc_inp->inp_socket->so_rcv,
2764 		    &rack->rc_inp->inp_socket->so_snd,
2765 		    BBR_LOG_TIMERSTAR, 0,
2766 		    0, &log, false, &tv);
2767 	}
2768 }
2769 
2770 static void
2771 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2772 {
2773 	if (tcp_bblogging_on(rack->rc_tp)) {
2774 		union tcp_log_stackspecific log;
2775 		struct timeval tv;
2776 
2777 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2778 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2779 		log.u_bbr.flex8 = to_num;
2780 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2781 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2782 		if (rsm == NULL)
2783 			log.u_bbr.flex3 = 0;
2784 		else
2785 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2786 		if (rack->rack_no_prr)
2787 			log.u_bbr.flex5 = 0;
2788 		else
2789 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2790 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2791 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2792 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2793 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2794 		log.u_bbr.pacing_gain = rack->r_must_retran;
2795 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2796 		log.u_bbr.bw_inuse <<= 32;
2797 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2798 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2799 		    &rack->rc_inp->inp_socket->so_rcv,
2800 		    &rack->rc_inp->inp_socket->so_snd,
2801 		    BBR_LOG_RTO, 0,
2802 		    0, &log, false, &tv);
2803 	}
2804 }
2805 
2806 static void
2807 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2808 		 struct rack_sendmap *prev,
2809 		 struct rack_sendmap *rsm,
2810 		 struct rack_sendmap *next,
2811 		 int flag, uint32_t th_ack, int line)
2812 {
2813 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2814 		union tcp_log_stackspecific log;
2815 		struct timeval tv;
2816 
2817 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2818 		log.u_bbr.flex8 = flag;
2819 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2820 		log.u_bbr.cur_del_rate = (uintptr_t)prev;
2821 		log.u_bbr.delRate = (uintptr_t)rsm;
2822 		log.u_bbr.rttProp = (uintptr_t)next;
2823 		log.u_bbr.flex7 = 0;
2824 		if (prev) {
2825 			log.u_bbr.flex1 = prev->r_start;
2826 			log.u_bbr.flex2 = prev->r_end;
2827 			log.u_bbr.flex7 |= 0x4;
2828 		}
2829 		if (rsm) {
2830 			log.u_bbr.flex3 = rsm->r_start;
2831 			log.u_bbr.flex4 = rsm->r_end;
2832 			log.u_bbr.flex7 |= 0x2;
2833 		}
2834 		if (next) {
2835 			log.u_bbr.flex5 = next->r_start;
2836 			log.u_bbr.flex6 = next->r_end;
2837 			log.u_bbr.flex7 |= 0x1;
2838 		}
2839 		log.u_bbr.applimited = line;
2840 		log.u_bbr.pkts_out = th_ack;
2841 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2842 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2843 		if (rack->rack_no_prr)
2844 			log.u_bbr.lost = 0;
2845 		else
2846 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2847 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2848 		log.u_bbr.bw_inuse <<= 32;
2849 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2850 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2851 		    &rack->rc_inp->inp_socket->so_rcv,
2852 		    &rack->rc_inp->inp_socket->so_snd,
2853 		    TCP_LOG_MAPCHG, 0,
2854 		    0, &log, false, &tv);
2855 	}
2856 }
2857 
2858 static void
2859 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2860 		 struct rack_sendmap *rsm, int conf)
2861 {
2862 	if (tcp_bblogging_on(tp)) {
2863 		union tcp_log_stackspecific log;
2864 		struct timeval tv;
2865 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2866 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2867 		log.u_bbr.flex1 = t;
2868 		log.u_bbr.flex2 = len;
2869 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2870 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2871 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2872 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2873 		log.u_bbr.flex7 = conf;
2874 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2875 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2876 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2877 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2878 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2879 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2880 		if (rsm) {
2881 			log.u_bbr.pkt_epoch = rsm->r_start;
2882 			log.u_bbr.lost = rsm->r_end;
2883 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2884 			/* We loose any upper of the 24 bits */
2885 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2886 		} else {
2887 			/* Its a SYN */
2888 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2889 			log.u_bbr.lost = 0;
2890 			log.u_bbr.cwnd_gain = 0;
2891 			log.u_bbr.pacing_gain = 0;
2892 		}
2893 		/* Write out general bits of interest rrs here */
2894 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2895 		log.u_bbr.use_lt_bw <<= 1;
2896 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2897 		log.u_bbr.use_lt_bw <<= 1;
2898 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2899 		log.u_bbr.use_lt_bw <<= 1;
2900 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2901 		log.u_bbr.use_lt_bw <<= 1;
2902 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2903 		log.u_bbr.use_lt_bw <<= 1;
2904 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2905 		log.u_bbr.use_lt_bw <<= 1;
2906 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2907 		log.u_bbr.use_lt_bw <<= 1;
2908 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2909 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2910 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2911 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2912 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2913 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2914 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2915 		log.u_bbr.bw_inuse <<= 32;
2916 		if (rsm)
2917 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2918 		TCP_LOG_EVENTP(tp, NULL,
2919 		    &rack->rc_inp->inp_socket->so_rcv,
2920 		    &rack->rc_inp->inp_socket->so_snd,
2921 		    BBR_LOG_BBRRTT, 0,
2922 		    0, &log, false, &tv);
2923 
2924 
2925 	}
2926 }
2927 
2928 static void
2929 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2930 {
2931 	/*
2932 	 * Log the rtt sample we are
2933 	 * applying to the srtt algorithm in
2934 	 * useconds.
2935 	 */
2936 	if (tcp_bblogging_on(rack->rc_tp)) {
2937 		union tcp_log_stackspecific log;
2938 		struct timeval tv;
2939 
2940 		/* Convert our ms to a microsecond */
2941 		memset(&log, 0, sizeof(log));
2942 		log.u_bbr.flex1 = rtt;
2943 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2944 		log.u_bbr.flex7 = 1;
2945 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2946 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2947 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2948 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2949 		log.u_bbr.pacing_gain = rack->r_must_retran;
2950 		/*
2951 		 * We capture in delRate the upper 32 bits as
2952 		 * the confidence level we had declared, and the
2953 		 * lower 32 bits as the actual RTT using the arrival
2954 		 * timestamp.
2955 		 */
2956 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2957 		log.u_bbr.delRate <<= 32;
2958 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2959 		/* Lets capture all the things that make up t_rtxcur */
2960 		log.u_bbr.applimited = rack_rto_min;
2961 		log.u_bbr.epoch = rack_rto_max;
2962 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2963 		log.u_bbr.lost = rack_rto_min;
2964 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2965 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2966 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2967 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2968 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2969 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2970 		    &rack->rc_inp->inp_socket->so_rcv,
2971 		    &rack->rc_inp->inp_socket->so_snd,
2972 		    TCP_LOG_RTT, 0,
2973 		    0, &log, false, &tv);
2974 	}
2975 }
2976 
2977 static void
2978 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2979 {
2980 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2981 		union tcp_log_stackspecific log;
2982 		struct timeval tv;
2983 
2984 		/* Convert our ms to a microsecond */
2985 		memset(&log, 0, sizeof(log));
2986 		log.u_bbr.flex1 = rtt;
2987 		log.u_bbr.flex2 = send_time;
2988 		log.u_bbr.flex3 = ack_time;
2989 		log.u_bbr.flex4 = where;
2990 		log.u_bbr.flex7 = 2;
2991 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2992 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2993 		log.u_bbr.bw_inuse <<= 32;
2994 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2995 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2996 		    &rack->rc_inp->inp_socket->so_rcv,
2997 		    &rack->rc_inp->inp_socket->so_snd,
2998 		    TCP_LOG_RTT, 0,
2999 		    0, &log, false, &tv);
3000 	}
3001 }
3002 
3003 
3004 static void
3005 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
3006 {
3007 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3008 		union tcp_log_stackspecific log;
3009 		struct timeval tv;
3010 
3011 		/* Convert our ms to a microsecond */
3012 		memset(&log, 0, sizeof(log));
3013 		log.u_bbr.flex1 = idx;
3014 		log.u_bbr.flex2 = rack_ts_to_msec(tsv);
3015 		log.u_bbr.flex3 = tsecho;
3016 		log.u_bbr.flex7 = 3;
3017 		log.u_bbr.rttProp = tsv;
3018 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3019 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3020 		log.u_bbr.bw_inuse <<= 32;
3021 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3022 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3023 		    &rack->rc_inp->inp_socket->so_rcv,
3024 		    &rack->rc_inp->inp_socket->so_snd,
3025 		    TCP_LOG_RTT, 0,
3026 		    0, &log, false, &tv);
3027 	}
3028 }
3029 
3030 
3031 static inline void
3032 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
3033 {
3034 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3035 		union tcp_log_stackspecific log;
3036 		struct timeval tv;
3037 
3038 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3039 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3040 		log.u_bbr.flex1 = line;
3041 		log.u_bbr.flex2 = tick;
3042 		log.u_bbr.flex3 = tp->t_maxunacktime;
3043 		log.u_bbr.flex4 = tp->t_acktime;
3044 		log.u_bbr.flex8 = event;
3045 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3046 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3047 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3048 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3049 		log.u_bbr.pacing_gain = rack->r_must_retran;
3050 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3051 		log.u_bbr.bw_inuse <<= 32;
3052 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3053 		TCP_LOG_EVENTP(tp, NULL,
3054 		    &rack->rc_inp->inp_socket->so_rcv,
3055 		    &rack->rc_inp->inp_socket->so_snd,
3056 		    BBR_LOG_PROGRESS, 0,
3057 		    0, &log, false, &tv);
3058 	}
3059 }
3060 
3061 static void
3062 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
3063 {
3064 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3065 		union tcp_log_stackspecific log;
3066 
3067 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3068 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3069 		log.u_bbr.flex1 = slot;
3070 		if (rack->rack_no_prr)
3071 			log.u_bbr.flex2 = 0;
3072 		else
3073 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3074 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3075 		log.u_bbr.flex6 = line;
3076 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3077 		log.u_bbr.flex8 = rack->rc_in_persist;
3078 		log.u_bbr.timeStamp = cts;
3079 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3080 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3081 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3082 		log.u_bbr.pacing_gain = rack->r_must_retran;
3083 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3084 		    &rack->rc_inp->inp_socket->so_rcv,
3085 		    &rack->rc_inp->inp_socket->so_snd,
3086 		    BBR_LOG_BBRSND, 0,
3087 		    0, &log, false, tv);
3088 	}
3089 }
3090 
3091 static void
3092 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
3093 {
3094 	if (tcp_bblogging_on(rack->rc_tp)) {
3095 		union tcp_log_stackspecific log;
3096 		struct timeval tv;
3097 
3098 		memset(&log, 0, sizeof(log));
3099 		log.u_bbr.flex1 = did_out;
3100 		log.u_bbr.flex2 = nxt_pkt;
3101 		log.u_bbr.flex3 = way_out;
3102 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3103 		if (rack->rack_no_prr)
3104 			log.u_bbr.flex5 = 0;
3105 		else
3106 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3107 		log.u_bbr.flex6 = nsegs;
3108 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3109 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
3110 		log.u_bbr.flex7 <<= 1;
3111 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
3112 		log.u_bbr.flex7 <<= 1;
3113 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
3114 		log.u_bbr.flex8 = rack->rc_in_persist;
3115 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3116 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3117 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3118 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3119 		log.u_bbr.use_lt_bw <<= 1;
3120 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3121 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3122 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3123 		log.u_bbr.pacing_gain = rack->r_must_retran;
3124 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3125 		log.u_bbr.bw_inuse <<= 32;
3126 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3127 		log.u_bbr.epoch = rack->rc_inp->inp_socket->so_snd.sb_hiwat;
3128 		log.u_bbr.lt_epoch = rack->rc_inp->inp_socket->so_rcv.sb_hiwat;
3129 		log.u_bbr.lost = rack->rc_tp->t_srtt;
3130 		log.u_bbr.pkt_epoch = rack->rc_tp->rfbuf_cnt;
3131 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3132 		    &rack->rc_inp->inp_socket->so_rcv,
3133 		    &rack->rc_inp->inp_socket->so_snd,
3134 		    BBR_LOG_DOSEG_DONE, 0,
3135 		    0, &log, false, &tv);
3136 	}
3137 }
3138 
3139 static void
3140 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
3141 {
3142 	if (tcp_bblogging_on(rack->rc_tp)) {
3143 		union tcp_log_stackspecific log;
3144 		struct timeval tv;
3145 
3146 		memset(&log, 0, sizeof(log));
3147 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3148 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3149 		log.u_bbr.flex4 = arg1;
3150 		log.u_bbr.flex5 = arg2;
3151 		log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3152 		log.u_bbr.flex6 = arg3;
3153 		log.u_bbr.flex8 = frm;
3154 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3155 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3156 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3157 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3158 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3159 		log.u_bbr.pacing_gain = rack->r_must_retran;
3160 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3161 		    &tptosocket(tp)->so_snd,
3162 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3163 	}
3164 }
3165 
3166 static void
3167 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3168 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3169 {
3170 	if (tcp_bblogging_on(rack->rc_tp)) {
3171 		union tcp_log_stackspecific log;
3172 		struct timeval tv;
3173 
3174 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3175 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3176 		log.u_bbr.flex1 = slot;
3177 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3178 		log.u_bbr.flex4 = reason;
3179 		if (rack->rack_no_prr)
3180 			log.u_bbr.flex5 = 0;
3181 		else
3182 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3183 		log.u_bbr.flex7 = hpts_calling;
3184 		log.u_bbr.flex8 = rack->rc_in_persist;
3185 		log.u_bbr.lt_epoch = cwnd_to_use;
3186 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3187 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3188 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3189 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3190 		log.u_bbr.pacing_gain = rack->r_must_retran;
3191 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3192 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3193 		log.u_bbr.bw_inuse <<= 32;
3194 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3195 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3196 		    &rack->rc_inp->inp_socket->so_rcv,
3197 		    &rack->rc_inp->inp_socket->so_snd,
3198 		    BBR_LOG_JUSTRET, 0,
3199 		    tlen, &log, false, &tv);
3200 	}
3201 }
3202 
3203 static void
3204 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3205 		   struct timeval *tv, uint32_t flags_on_entry)
3206 {
3207 	if (tcp_bblogging_on(rack->rc_tp)) {
3208 		union tcp_log_stackspecific log;
3209 
3210 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3211 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3212 		log.u_bbr.flex1 = line;
3213 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3214 		log.u_bbr.flex3 = flags_on_entry;
3215 		log.u_bbr.flex4 = us_cts;
3216 		if (rack->rack_no_prr)
3217 			log.u_bbr.flex5 = 0;
3218 		else
3219 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3220 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3221 		log.u_bbr.flex7 = hpts_removed;
3222 		log.u_bbr.flex8 = 1;
3223 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3224 		log.u_bbr.timeStamp = us_cts;
3225 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3226 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3227 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3228 		log.u_bbr.pacing_gain = rack->r_must_retran;
3229 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3230 		log.u_bbr.bw_inuse <<= 32;
3231 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3232 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3233 		    &rack->rc_inp->inp_socket->so_rcv,
3234 		    &rack->rc_inp->inp_socket->so_snd,
3235 		    BBR_LOG_TIMERCANC, 0,
3236 		    0, &log, false, tv);
3237 	}
3238 }
3239 
3240 static void
3241 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3242 			  uint32_t flex1, uint32_t flex2,
3243 			  uint32_t flex3, uint32_t flex4,
3244 			  uint32_t flex5, uint32_t flex6,
3245 			  uint16_t flex7, uint8_t mod)
3246 {
3247 	if (tcp_bblogging_on(rack->rc_tp)) {
3248 		union tcp_log_stackspecific log;
3249 		struct timeval tv;
3250 
3251 		if (mod == 1) {
3252 			/* No you can't use 1, its for the real to cancel */
3253 			return;
3254 		}
3255 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3256 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3257 		log.u_bbr.flex1 = flex1;
3258 		log.u_bbr.flex2 = flex2;
3259 		log.u_bbr.flex3 = flex3;
3260 		log.u_bbr.flex4 = flex4;
3261 		log.u_bbr.flex5 = flex5;
3262 		log.u_bbr.flex6 = flex6;
3263 		log.u_bbr.flex7 = flex7;
3264 		log.u_bbr.flex8 = mod;
3265 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3266 		    &rack->rc_inp->inp_socket->so_rcv,
3267 		    &rack->rc_inp->inp_socket->so_snd,
3268 		    BBR_LOG_TIMERCANC, 0,
3269 		    0, &log, false, &tv);
3270 	}
3271 }
3272 
3273 static void
3274 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3275 {
3276 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3277 		union tcp_log_stackspecific log;
3278 		struct timeval tv;
3279 
3280 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3281 		log.u_bbr.flex1 = timers;
3282 		log.u_bbr.flex2 = ret;
3283 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3284 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3285 		log.u_bbr.flex5 = cts;
3286 		if (rack->rack_no_prr)
3287 			log.u_bbr.flex6 = 0;
3288 		else
3289 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3290 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3291 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3292 		log.u_bbr.pacing_gain = rack->r_must_retran;
3293 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3294 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3295 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3296 		    &rack->rc_inp->inp_socket->so_rcv,
3297 		    &rack->rc_inp->inp_socket->so_snd,
3298 		    BBR_LOG_TO_PROCESS, 0,
3299 		    0, &log, false, &tv);
3300 	}
3301 }
3302 
3303 static void
3304 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3305 {
3306 	if (tcp_bblogging_on(rack->rc_tp)) {
3307 		union tcp_log_stackspecific log;
3308 		struct timeval tv;
3309 
3310 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3311 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3312 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3313 		if (rack->rack_no_prr)
3314 			log.u_bbr.flex3 = 0;
3315 		else
3316 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3317 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3318 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3319 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3320 		log.u_bbr.flex7 = line;
3321 		log.u_bbr.flex8 = frm;
3322 		log.u_bbr.pkts_out = orig_cwnd;
3323 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3324 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3325 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3326 		log.u_bbr.use_lt_bw <<= 1;
3327 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3328 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3329 		    &rack->rc_inp->inp_socket->so_rcv,
3330 		    &rack->rc_inp->inp_socket->so_snd,
3331 		    BBR_LOG_BBRUPD, 0,
3332 		    0, &log, false, &tv);
3333 	}
3334 }
3335 
3336 static void
3337 rack_counter_destroy(void)
3338 {
3339 	counter_u64_free(rack_total_bytes);
3340 	counter_u64_free(rack_fto_send);
3341 	counter_u64_free(rack_fto_rsm_send);
3342 	counter_u64_free(rack_nfto_resend);
3343 	counter_u64_free(rack_hw_pace_init_fail);
3344 	counter_u64_free(rack_hw_pace_lost);
3345 	counter_u64_free(rack_non_fto_send);
3346 	counter_u64_free(rack_extended_rfo);
3347 	counter_u64_free(rack_ack_total);
3348 	counter_u64_free(rack_express_sack);
3349 	counter_u64_free(rack_sack_total);
3350 	counter_u64_free(rack_move_none);
3351 	counter_u64_free(rack_move_some);
3352 	counter_u64_free(rack_sack_attacks_detected);
3353 	counter_u64_free(rack_sack_attacks_reversed);
3354 	counter_u64_free(rack_sack_attacks_suspect);
3355 	counter_u64_free(rack_sack_used_next_merge);
3356 	counter_u64_free(rack_sack_used_prev_merge);
3357 	counter_u64_free(rack_tlp_tot);
3358 	counter_u64_free(rack_tlp_newdata);
3359 	counter_u64_free(rack_tlp_retran);
3360 	counter_u64_free(rack_tlp_retran_bytes);
3361 	counter_u64_free(rack_to_tot);
3362 	counter_u64_free(rack_saw_enobuf);
3363 	counter_u64_free(rack_saw_enobuf_hw);
3364 	counter_u64_free(rack_saw_enetunreach);
3365 	counter_u64_free(rack_hot_alloc);
3366 	counter_u64_free(rack_to_alloc);
3367 	counter_u64_free(rack_to_alloc_hard);
3368 	counter_u64_free(rack_to_alloc_emerg);
3369 	counter_u64_free(rack_to_alloc_limited);
3370 	counter_u64_free(rack_alloc_limited_conns);
3371 	counter_u64_free(rack_split_limited);
3372 	counter_u64_free(rack_multi_single_eq);
3373 	counter_u64_free(rack_rxt_clamps_cwnd);
3374 	counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3375 	counter_u64_free(rack_proc_non_comp_ack);
3376 	counter_u64_free(rack_sack_proc_all);
3377 	counter_u64_free(rack_sack_proc_restart);
3378 	counter_u64_free(rack_sack_proc_short);
3379 	counter_u64_free(rack_sack_skipped_acked);
3380 	counter_u64_free(rack_sack_splits);
3381 	counter_u64_free(rack_input_idle_reduces);
3382 	counter_u64_free(rack_collapsed_win);
3383 	counter_u64_free(rack_collapsed_win_rxt);
3384 	counter_u64_free(rack_collapsed_win_rxt_bytes);
3385 	counter_u64_free(rack_collapsed_win_seen);
3386 	counter_u64_free(rack_try_scwnd);
3387 	counter_u64_free(rack_persists_sends);
3388 	counter_u64_free(rack_persists_acks);
3389 	counter_u64_free(rack_persists_loss);
3390 	counter_u64_free(rack_persists_lost_ends);
3391 #ifdef INVARIANTS
3392 	counter_u64_free(rack_adjust_map_bw);
3393 #endif
3394 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3395 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3396 }
3397 
3398 static struct rack_sendmap *
3399 rack_alloc(struct tcp_rack *rack)
3400 {
3401 	struct rack_sendmap *rsm;
3402 
3403 	/*
3404 	 * First get the top of the list it in
3405 	 * theory is the "hottest" rsm we have,
3406 	 * possibly just freed by ack processing.
3407 	 */
3408 	if (rack->rc_free_cnt > rack_free_cache) {
3409 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3410 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3411 		counter_u64_add(rack_hot_alloc, 1);
3412 		rack->rc_free_cnt--;
3413 		return (rsm);
3414 	}
3415 	/*
3416 	 * Once we get under our free cache we probably
3417 	 * no longer have a "hot" one available. Lets
3418 	 * get one from UMA.
3419 	 */
3420 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
3421 	if (rsm) {
3422 		rack->r_ctl.rc_num_maps_alloced++;
3423 		counter_u64_add(rack_to_alloc, 1);
3424 		return (rsm);
3425 	}
3426 	/*
3427 	 * Dig in to our aux rsm's (the last two) since
3428 	 * UMA failed to get us one.
3429 	 */
3430 	if (rack->rc_free_cnt) {
3431 		counter_u64_add(rack_to_alloc_emerg, 1);
3432 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3433 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3434 		rack->rc_free_cnt--;
3435 		return (rsm);
3436 	}
3437 	return (NULL);
3438 }
3439 
3440 static struct rack_sendmap *
3441 rack_alloc_full_limit(struct tcp_rack *rack)
3442 {
3443 	if ((V_tcp_map_entries_limit > 0) &&
3444 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3445 		counter_u64_add(rack_to_alloc_limited, 1);
3446 		if (!rack->alloc_limit_reported) {
3447 			rack->alloc_limit_reported = 1;
3448 			counter_u64_add(rack_alloc_limited_conns, 1);
3449 		}
3450 		return (NULL);
3451 	}
3452 	return (rack_alloc(rack));
3453 }
3454 
3455 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3456 static struct rack_sendmap *
3457 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3458 {
3459 	struct rack_sendmap *rsm;
3460 
3461 	if (limit_type) {
3462 		/* currently there is only one limit type */
3463 		if (rack->r_ctl.rc_split_limit > 0 &&
3464 		    rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3465 			counter_u64_add(rack_split_limited, 1);
3466 			if (!rack->alloc_limit_reported) {
3467 				rack->alloc_limit_reported = 1;
3468 				counter_u64_add(rack_alloc_limited_conns, 1);
3469 			}
3470 			return (NULL);
3471 		}
3472 	}
3473 
3474 	/* allocate and mark in the limit type, if set */
3475 	rsm = rack_alloc(rack);
3476 	if (rsm != NULL && limit_type) {
3477 		rsm->r_limit_type = limit_type;
3478 		rack->r_ctl.rc_num_split_allocs++;
3479 	}
3480 	return (rsm);
3481 }
3482 
3483 static void
3484 rack_free_trim(struct tcp_rack *rack)
3485 {
3486 	struct rack_sendmap *rsm;
3487 
3488 	/*
3489 	 * Free up all the tail entries until
3490 	 * we get our list down to the limit.
3491 	 */
3492 	while (rack->rc_free_cnt > rack_free_cache) {
3493 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3494 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3495 		rack->rc_free_cnt--;
3496 		rack->r_ctl.rc_num_maps_alloced--;
3497 		uma_zfree(rack_zone, rsm);
3498 	}
3499 }
3500 
3501 static void
3502 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3503 {
3504 	if (rsm->r_flags & RACK_APP_LIMITED) {
3505 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
3506 			rack->r_ctl.rc_app_limited_cnt--;
3507 		}
3508 	}
3509 	if (rsm->r_limit_type) {
3510 		/* currently there is only one limit type */
3511 		rack->r_ctl.rc_num_split_allocs--;
3512 	}
3513 	if (rsm == rack->r_ctl.rc_first_appl) {
3514 		rack->r_ctl.cleared_app_ack_seq = rsm->r_start + (rsm->r_end - rsm->r_start);
3515 		rack->r_ctl.cleared_app_ack = 1;
3516 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3517 			rack->r_ctl.rc_first_appl = NULL;
3518 		else
3519 			rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3520 	}
3521 	if (rsm == rack->r_ctl.rc_resend)
3522 		rack->r_ctl.rc_resend = NULL;
3523 	if (rsm == rack->r_ctl.rc_end_appl)
3524 		rack->r_ctl.rc_end_appl = NULL;
3525 	if (rack->r_ctl.rc_tlpsend == rsm)
3526 		rack->r_ctl.rc_tlpsend = NULL;
3527 	if (rack->r_ctl.rc_sacklast == rsm)
3528 		rack->r_ctl.rc_sacklast = NULL;
3529 	memset(rsm, 0, sizeof(struct rack_sendmap));
3530 	/* Make sure we are not going to overrun our count limit of 0xff */
3531 	if ((rack->rc_free_cnt + 1) > RACK_FREE_CNT_MAX) {
3532 		rack_free_trim(rack);
3533 	}
3534 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3535 	rack->rc_free_cnt++;
3536 }
3537 
3538 static uint32_t
3539 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3540 {
3541 	uint64_t srtt, bw, len, tim;
3542 	uint32_t segsiz, def_len, minl;
3543 
3544 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3545 	def_len = rack_def_data_window * segsiz;
3546 	if (rack->rc_gp_filled == 0) {
3547 		/*
3548 		 * We have no measurement (IW is in flight?) so
3549 		 * we can only guess using our data_window sysctl
3550 		 * value (usually 20MSS).
3551 		 */
3552 		return (def_len);
3553 	}
3554 	/*
3555 	 * Now we have a number of factors to consider.
3556 	 *
3557 	 * 1) We have a desired BDP which is usually
3558 	 *    at least 2.
3559 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3560 	 *    but we allow it too to be more.
3561 	 * 3) We want to make sure a measurement last N useconds (if
3562 	 *    we have set rack_min_measure_usec.
3563 	 *
3564 	 * We handle the first concern here by trying to create a data
3565 	 * window of max(rack_def_data_window, DesiredBDP). The
3566 	 * second concern we handle in not letting the measurement
3567 	 * window end normally until at least the required SRTT's
3568 	 * have gone by which is done further below in
3569 	 * rack_enough_for_measurement(). Finally the third concern
3570 	 * we also handle here by calculating how long that time
3571 	 * would take at the current BW and then return the
3572 	 * max of our first calculation and that length. Note
3573 	 * that if rack_min_measure_usec is 0, we don't deal
3574 	 * with concern 3. Also for both Concern 1 and 3 an
3575 	 * application limited period could end the measurement
3576 	 * earlier.
3577 	 *
3578 	 * So lets calculate the BDP with the "known" b/w using
3579 	 * the SRTT has our rtt and then multiply it by the
3580 	 * goal.
3581 	 */
3582 	bw = rack_get_bw(rack);
3583 	srtt = (uint64_t)tp->t_srtt;
3584 	len = bw * srtt;
3585 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3586 	len *= max(1, rack_goal_bdp);
3587 	/* Now we need to round up to the nearest MSS */
3588 	len = roundup(len, segsiz);
3589 	if (rack_min_measure_usec) {
3590 		/* Now calculate our min length for this b/w */
3591 		tim = rack_min_measure_usec;
3592 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3593 		if (minl == 0)
3594 			minl = 1;
3595 		minl = roundup(minl, segsiz);
3596 		if (len < minl)
3597 			len = minl;
3598 	}
3599 	/*
3600 	 * Now if we have a very small window we want
3601 	 * to attempt to get the window that is
3602 	 * as small as possible. This happens on
3603 	 * low b/w connections and we don't want to
3604 	 * span huge numbers of rtt's between measurements.
3605 	 *
3606 	 * We basically include 2 over our "MIN window" so
3607 	 * that the measurement can be shortened (possibly) by
3608 	 * an ack'ed packet.
3609 	 */
3610 	if (len < def_len)
3611 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3612 	else
3613 		return (max((uint32_t)len, def_len));
3614 
3615 }
3616 
3617 static int
3618 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3619 {
3620 	uint32_t tim, srtts, segsiz;
3621 
3622 	/*
3623 	 * Has enough time passed for the GP measurement to be valid?
3624 	 */
3625 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3626 		/* Not enough bytes yet */
3627 		return (0);
3628 	}
3629 	if ((tp->snd_max == tp->snd_una) ||
3630 	    (th_ack == tp->snd_max)){
3631 		/*
3632 		 * All is acked quality of all acked is
3633 		 * usually low or medium, but we in theory could split
3634 		 * all acked into two cases, where you got
3635 		 * a signifigant amount of your window and
3636 		 * where you did not. For now we leave it
3637 		 * but it is something to contemplate in the
3638 		 * future. The danger here is that delayed ack
3639 		 * is effecting the last byte (which is a 50:50 chance).
3640 		 */
3641 		*quality = RACK_QUALITY_ALLACKED;
3642 		return (1);
3643 	}
3644 	if (SEQ_GEQ(th_ack,  tp->gput_ack)) {
3645 		/*
3646 		 * We obtained our entire window of data we wanted
3647 		 * no matter if we are in recovery or not then
3648 		 * its ok since expanding the window does not
3649 		 * make things fuzzy (or at least not as much).
3650 		 */
3651 		*quality = RACK_QUALITY_HIGH;
3652 		return (1);
3653 	}
3654 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3655 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3656 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3657 		/* Not enough bytes yet */
3658 		return (0);
3659 	}
3660 	if (rack->r_ctl.rc_first_appl &&
3661 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3662 		/*
3663 		 * We are up to the app limited send point
3664 		 * we have to measure irrespective of the time..
3665 		 */
3666 		*quality = RACK_QUALITY_APPLIMITED;
3667 		return (1);
3668 	}
3669 	/* Now what about time? */
3670 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3671 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3672 	if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3673 		/*
3674 		 * We do not allow a measurement if we are in recovery
3675 		 * that would shrink the goodput window we wanted.
3676 		 * This is to prevent cloudyness of when the last send
3677 		 * was actually made.
3678 		 */
3679 		*quality = RACK_QUALITY_HIGH;
3680 		return (1);
3681 	}
3682 	/* Nope not even a full SRTT has passed */
3683 	return (0);
3684 }
3685 
3686 static void
3687 rack_log_timely(struct tcp_rack *rack,
3688 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3689 		uint64_t up_bnd, int line, uint8_t method)
3690 {
3691 	if (tcp_bblogging_on(rack->rc_tp)) {
3692 		union tcp_log_stackspecific log;
3693 		struct timeval tv;
3694 
3695 		memset(&log, 0, sizeof(log));
3696 		log.u_bbr.flex1 = logged;
3697 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3698 		log.u_bbr.flex2 <<= 4;
3699 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3700 		log.u_bbr.flex2 <<= 4;
3701 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3702 		log.u_bbr.flex2 <<= 4;
3703 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3704 		log.u_bbr.flex3 = rack->rc_gp_incr;
3705 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3706 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3707 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3708 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3709 		log.u_bbr.flex8 = method;
3710 		log.u_bbr.cur_del_rate = cur_bw;
3711 		log.u_bbr.delRate = low_bnd;
3712 		log.u_bbr.bw_inuse = up_bnd;
3713 		log.u_bbr.rttProp = rack_get_bw(rack);
3714 		log.u_bbr.pkt_epoch = line;
3715 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3716 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3717 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3718 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3719 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3720 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3721 		log.u_bbr.cwnd_gain <<= 1;
3722 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3723 		log.u_bbr.cwnd_gain <<= 1;
3724 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3725 		log.u_bbr.cwnd_gain <<= 1;
3726 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3727 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3728 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3729 		    &rack->rc_inp->inp_socket->so_rcv,
3730 		    &rack->rc_inp->inp_socket->so_snd,
3731 		    TCP_TIMELY_WORK, 0,
3732 		    0, &log, false, &tv);
3733 	}
3734 }
3735 
3736 static int
3737 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3738 {
3739 	/*
3740 	 * Before we increase we need to know if
3741 	 * the estimate just made was less than
3742 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3743 	 *
3744 	 * If we already are pacing at a fast enough
3745 	 * rate to push us faster there is no sense of
3746 	 * increasing.
3747 	 *
3748 	 * We first caculate our actual pacing rate (ss or ca multiplier
3749 	 * times our cur_bw).
3750 	 *
3751 	 * Then we take the last measured rate and multipy by our
3752 	 * maximum pacing overage to give us a max allowable rate.
3753 	 *
3754 	 * If our act_rate is smaller than our max_allowable rate
3755 	 * then we should increase. Else we should hold steady.
3756 	 *
3757 	 */
3758 	uint64_t act_rate, max_allow_rate;
3759 
3760 	if (rack_timely_no_stopping)
3761 		return (1);
3762 
3763 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3764 		/*
3765 		 * Initial startup case or
3766 		 * everything is acked case.
3767 		 */
3768 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3769 				__LINE__, 9);
3770 		return (1);
3771 	}
3772 	if (mult <= 100) {
3773 		/*
3774 		 * We can always pace at or slightly above our rate.
3775 		 */
3776 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3777 				__LINE__, 9);
3778 		return (1);
3779 	}
3780 	act_rate = cur_bw * (uint64_t)mult;
3781 	act_rate /= 100;
3782 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3783 	max_allow_rate /= 100;
3784 	if (act_rate < max_allow_rate) {
3785 		/*
3786 		 * Here the rate we are actually pacing at
3787 		 * is smaller than 10% above our last measurement.
3788 		 * This means we are pacing below what we would
3789 		 * like to try to achieve (plus some wiggle room).
3790 		 */
3791 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3792 				__LINE__, 9);
3793 		return (1);
3794 	} else {
3795 		/*
3796 		 * Here we are already pacing at least rack_max_per_above(10%)
3797 		 * what we are getting back. This indicates most likely
3798 		 * that we are being limited (cwnd/rwnd/app) and can't
3799 		 * get any more b/w. There is no sense of trying to
3800 		 * raise up the pacing rate its not speeding us up
3801 		 * and we already are pacing faster than we are getting.
3802 		 */
3803 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3804 				__LINE__, 8);
3805 		return (0);
3806 	}
3807 }
3808 
3809 static void
3810 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3811 {
3812 	/*
3813 	 * When we drag bottom, we want to assure
3814 	 * that no multiplier is below 1.0, if so
3815 	 * we want to restore it to at least that.
3816 	 */
3817 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3818 		/* This is unlikely we usually do not touch recovery */
3819 		rack->r_ctl.rack_per_of_gp_rec = 100;
3820 	}
3821 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3822 		rack->r_ctl.rack_per_of_gp_ca = 100;
3823 	}
3824 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3825 		rack->r_ctl.rack_per_of_gp_ss = 100;
3826 	}
3827 }
3828 
3829 static void
3830 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3831 {
3832 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3833 		rack->r_ctl.rack_per_of_gp_ca = 100;
3834 	}
3835 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3836 		rack->r_ctl.rack_per_of_gp_ss = 100;
3837 	}
3838 }
3839 
3840 static void
3841 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3842 {
3843 	int32_t  calc, logged, plus;
3844 
3845 	logged = 0;
3846 
3847 	if (rack->rc_skip_timely)
3848 		return;
3849 	if (override) {
3850 		/*
3851 		 * override is passed when we are
3852 		 * loosing b/w and making one last
3853 		 * gasp at trying to not loose out
3854 		 * to a new-reno flow.
3855 		 */
3856 		goto extra_boost;
3857 	}
3858 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3859 	if (rack->rc_gp_incr &&
3860 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3861 		/*
3862 		 * Reset and get 5 strokes more before the boost. Note
3863 		 * that the count is 0 based so we have to add one.
3864 		 */
3865 extra_boost:
3866 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3867 		rack->rc_gp_timely_inc_cnt = 0;
3868 	} else
3869 		plus = (uint32_t)rack_gp_increase_per;
3870 	/* Must be at least 1% increase for true timely increases */
3871 	if ((plus < 1) &&
3872 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3873 		plus = 1;
3874 	if (rack->rc_gp_saw_rec &&
3875 	    (rack->rc_gp_no_rec_chg == 0) &&
3876 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3877 				  rack->r_ctl.rack_per_of_gp_rec)) {
3878 		/* We have been in recovery ding it too */
3879 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3880 		if (calc > 0xffff)
3881 			calc = 0xffff;
3882 		logged |= 1;
3883 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3884 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3885 		    (rack->rc_dragged_bottom == 0) &&
3886 		    (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3887 			rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3888 	}
3889 	if (rack->rc_gp_saw_ca &&
3890 	    (rack->rc_gp_saw_ss == 0) &&
3891 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3892 				  rack->r_ctl.rack_per_of_gp_ca)) {
3893 		/* In CA */
3894 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3895 		if (calc > 0xffff)
3896 			calc = 0xffff;
3897 		logged |= 2;
3898 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3899 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3900 		    (rack->rc_dragged_bottom == 0) &&
3901 		    (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3902 			rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3903 	}
3904 	if (rack->rc_gp_saw_ss &&
3905 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3906 				  rack->r_ctl.rack_per_of_gp_ss)) {
3907 		/* In SS */
3908 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3909 		if (calc > 0xffff)
3910 			calc = 0xffff;
3911 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3912 		if (rack->r_ctl.rack_per_upper_bound_ss &&
3913 		    (rack->rc_dragged_bottom == 0) &&
3914 		    (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3915 			rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3916 		logged |= 4;
3917 	}
3918 	if (logged &&
3919 	    (rack->rc_gp_incr == 0)){
3920 		/* Go into increment mode */
3921 		rack->rc_gp_incr = 1;
3922 		rack->rc_gp_timely_inc_cnt = 0;
3923 	}
3924 	if (rack->rc_gp_incr &&
3925 	    logged &&
3926 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3927 		rack->rc_gp_timely_inc_cnt++;
3928 	}
3929 	rack_log_timely(rack,  logged, plus, 0, 0,
3930 			__LINE__, 1);
3931 }
3932 
3933 static uint32_t
3934 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3935 {
3936 	/*-
3937 	 * norm_grad = rtt_diff / minrtt;
3938 	 * new_per = curper * (1 - B * norm_grad)
3939 	 *
3940 	 * B = rack_gp_decrease_per (default 80%)
3941 	 * rtt_dif = input var current rtt-diff
3942 	 * curper = input var current percentage
3943 	 * minrtt = from rack filter
3944 	 *
3945 	 * In order to do the floating point calculations above we
3946 	 * do an integer conversion. The code looks confusing so let me
3947 	 * translate it into something that use more variables and
3948 	 * is clearer for us humans :)
3949 	 *
3950 	 * uint64_t norm_grad, inverse, reduce_by, final_result;
3951 	 * uint32_t perf;
3952 	 *
3953 	 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
3954 	 *             (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
3955 	 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
3956 	 * inverse /= 1000000;
3957 	 * reduce_by = (1000000 - inverse);
3958 	 * final_result = (cur_per * reduce_by) / 1000000;
3959 	 * perf = (uint32_t)final_result;
3960 	 */
3961 	uint64_t perf;
3962 
3963 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3964 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3965 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3966 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3967 		     (uint64_t)1000000)) /
3968 		(uint64_t)1000000);
3969 	if (perf > curper) {
3970 		/* TSNH */
3971 		perf = curper - 1;
3972 	}
3973 	return ((uint32_t)perf);
3974 }
3975 
3976 static uint32_t
3977 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3978 {
3979 	/*
3980 	 *                                   highrttthresh
3981 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3982 	 *                                     gp_srtt
3983 	 *
3984 	 * B = rack_gp_decrease_per (default .8 i.e. 80)
3985 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3986 	 */
3987 	uint64_t perf;
3988 	uint32_t highrttthresh;
3989 
3990 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3991 
3992 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3993 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3994 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3995 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3996 	if (tcp_bblogging_on(rack->rc_tp)) {
3997 		uint64_t log1;
3998 
3999 		log1 = rtt;
4000 		log1 <<= 32;
4001 		log1 |= highrttthresh;
4002 		rack_log_timely(rack,
4003 				rack_gp_decrease_per,
4004 				(uint64_t)curper,
4005 				log1,
4006 				perf,
4007 				__LINE__,
4008 				15);
4009 	}
4010 	return (perf);
4011 }
4012 
4013 static void
4014 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
4015 {
4016 	uint64_t logvar, logvar2, logvar3;
4017 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
4018 
4019 	if (rack->rc_skip_timely)
4020 		return;
4021 	if (rack->rc_gp_incr) {
4022 		/* Turn off increment counting */
4023 		rack->rc_gp_incr = 0;
4024 		rack->rc_gp_timely_inc_cnt = 0;
4025 	}
4026 	ss_red = ca_red = rec_red = 0;
4027 	logged = 0;
4028 	/* Calculate the reduction value */
4029 	if (rtt_diff < 0) {
4030 		rtt_diff *= -1;
4031 	}
4032 	/* Must be at least 1% reduction */
4033 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
4034 		/* We have been in recovery ding it too */
4035 		if (timely_says == 2) {
4036 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
4037 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4038 			if (alt < new_per)
4039 				val = alt;
4040 			else
4041 				val = new_per;
4042 		} else
4043 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4044 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
4045 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4046 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4047 		} else {
4048 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4049 			rec_red = 0;
4050 		}
4051 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4052 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4053 		logged |= 1;
4054 	}
4055 	if (rack->rc_gp_saw_ss) {
4056 		/* Sent in SS */
4057 		if (timely_says == 2) {
4058 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4059 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4060 			if (alt < new_per)
4061 				val = alt;
4062 			else
4063 				val = new_per;
4064 		} else
4065 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4066 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4067 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4068 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4069 		} else {
4070 			ss_red = new_per;
4071 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4072 			logvar = new_per;
4073 			logvar <<= 32;
4074 			logvar |= alt;
4075 			logvar2 = (uint32_t)rtt;
4076 			logvar2 <<= 32;
4077 			logvar2 |= (uint32_t)rtt_diff;
4078 			logvar3 = rack_gp_rtt_maxmul;
4079 			logvar3 <<= 32;
4080 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4081 			rack_log_timely(rack, timely_says,
4082 					logvar2, logvar3,
4083 					logvar, __LINE__, 10);
4084 		}
4085 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4086 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4087 		logged |= 4;
4088 	} else if (rack->rc_gp_saw_ca) {
4089 		/* Sent in CA */
4090 		if (timely_says == 2) {
4091 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4092 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4093 			if (alt < new_per)
4094 				val = alt;
4095 			else
4096 				val = new_per;
4097 		} else
4098 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4099 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
4100 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4101 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4102 		} else {
4103 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4104 			ca_red = 0;
4105 			logvar = new_per;
4106 			logvar <<= 32;
4107 			logvar |= alt;
4108 			logvar2 = (uint32_t)rtt;
4109 			logvar2 <<= 32;
4110 			logvar2 |= (uint32_t)rtt_diff;
4111 			logvar3 = rack_gp_rtt_maxmul;
4112 			logvar3 <<= 32;
4113 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4114 			rack_log_timely(rack, timely_says,
4115 					logvar2, logvar3,
4116 					logvar, __LINE__, 10);
4117 		}
4118 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4119 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4120 		logged |= 2;
4121 	}
4122 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
4123 		rack->rc_gp_timely_dec_cnt++;
4124 		if (rack_timely_dec_clear &&
4125 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4126 			rack->rc_gp_timely_dec_cnt = 0;
4127 	}
4128 	logvar = ss_red;
4129 	logvar <<= 32;
4130 	logvar |= ca_red;
4131 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
4132 			__LINE__, 2);
4133 }
4134 
4135 static void
4136 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4137 		     uint32_t rtt, uint32_t line, uint8_t reas)
4138 {
4139 	if (tcp_bblogging_on(rack->rc_tp)) {
4140 		union tcp_log_stackspecific log;
4141 		struct timeval tv;
4142 
4143 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4144 		log.u_bbr.flex1 = line;
4145 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4146 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4147 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4148 		log.u_bbr.flex5 = rtt;
4149 		log.u_bbr.flex6 = rack->rc_highly_buffered;
4150 		log.u_bbr.flex6 <<= 1;
4151 		log.u_bbr.flex6 |= rack->forced_ack;
4152 		log.u_bbr.flex6 <<= 1;
4153 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4154 		log.u_bbr.flex6 <<= 1;
4155 		log.u_bbr.flex6 |= rack->in_probe_rtt;
4156 		log.u_bbr.flex6 <<= 1;
4157 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4158 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4159 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4160 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4161 		log.u_bbr.flex8 = reas;
4162 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4163 		log.u_bbr.delRate = rack_get_bw(rack);
4164 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4165 		log.u_bbr.cur_del_rate <<= 32;
4166 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4167 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4168 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4169 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4170 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4171 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4172 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4173 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4174 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4175 		log.u_bbr.rttProp = us_cts;
4176 		log.u_bbr.rttProp <<= 32;
4177 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4178 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4179 		    &rack->rc_inp->inp_socket->so_rcv,
4180 		    &rack->rc_inp->inp_socket->so_snd,
4181 		    BBR_LOG_RTT_SHRINKS, 0,
4182 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4183 	}
4184 }
4185 
4186 static void
4187 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4188 {
4189 	uint64_t bwdp;
4190 
4191 	bwdp = rack_get_bw(rack);
4192 	bwdp *= (uint64_t)rtt;
4193 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4194 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4195 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4196 		/*
4197 		 * A window protocol must be able to have 4 packets
4198 		 * outstanding as the floor in order to function
4199 		 * (especially considering delayed ack :D).
4200 		 */
4201 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4202 	}
4203 }
4204 
4205 static void
4206 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4207 {
4208 	/**
4209 	 * ProbeRTT is a bit different in rack_pacing than in
4210 	 * BBR. It is like BBR in that it uses the lowering of
4211 	 * the RTT as a signal that we saw something new and
4212 	 * counts from there for how long between. But it is
4213 	 * different in that its quite simple. It does not
4214 	 * play with the cwnd and wait until we get down
4215 	 * to N segments outstanding and hold that for
4216 	 * 200ms. Instead it just sets the pacing reduction
4217 	 * rate to a set percentage (70 by default) and hold
4218 	 * that for a number of recent GP Srtt's.
4219 	 */
4220 	uint32_t segsiz;
4221 
4222 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4223 	if (rack->rc_gp_dyn_mul == 0)
4224 		return;
4225 
4226 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4227 		/* We are idle */
4228 		return;
4229 	}
4230 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4231 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4232 		/*
4233 		 * Stop the goodput now, the idea here is
4234 		 * that future measurements with in_probe_rtt
4235 		 * won't register if they are not greater so
4236 		 * we want to get what info (if any) is available
4237 		 * now.
4238 		 */
4239 		rack_do_goodput_measurement(rack->rc_tp, rack,
4240 					    rack->rc_tp->snd_una, __LINE__,
4241 					    RACK_QUALITY_PROBERTT);
4242 	}
4243 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4244 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4245 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4246 		     rack->r_ctl.rc_pace_min_segs);
4247 	rack->in_probe_rtt = 1;
4248 	rack->measure_saw_probe_rtt = 1;
4249 	rack->r_ctl.rc_time_probertt_starts = 0;
4250 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4251 	if (rack_probertt_use_min_rtt_entry)
4252 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4253 	else
4254 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4255 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4256 			     __LINE__, RACK_RTTS_ENTERPROBE);
4257 }
4258 
4259 static void
4260 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4261 {
4262 	struct rack_sendmap *rsm;
4263 	uint32_t segsiz;
4264 
4265 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4266 		     rack->r_ctl.rc_pace_min_segs);
4267 	rack->in_probe_rtt = 0;
4268 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4269 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4270 		/*
4271 		 * Stop the goodput now, the idea here is
4272 		 * that future measurements with in_probe_rtt
4273 		 * won't register if they are not greater so
4274 		 * we want to get what info (if any) is available
4275 		 * now.
4276 		 */
4277 		rack_do_goodput_measurement(rack->rc_tp, rack,
4278 					    rack->rc_tp->snd_una, __LINE__,
4279 					    RACK_QUALITY_PROBERTT);
4280 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4281 		/*
4282 		 * We don't have enough data to make a measurement.
4283 		 * So lets just stop and start here after exiting
4284 		 * probe-rtt. We probably are not interested in
4285 		 * the results anyway.
4286 		 */
4287 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4288 	}
4289 	/*
4290 	 * Measurements through the current snd_max are going
4291 	 * to be limited by the slower pacing rate.
4292 	 *
4293 	 * We need to mark these as app-limited so we
4294 	 * don't collapse the b/w.
4295 	 */
4296 	rsm = tqhash_max(rack->r_ctl.tqh);
4297 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4298 		if (rack->r_ctl.rc_app_limited_cnt == 0)
4299 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4300 		else {
4301 			/*
4302 			 * Go out to the end app limited and mark
4303 			 * this new one as next and move the end_appl up
4304 			 * to this guy.
4305 			 */
4306 			if (rack->r_ctl.rc_end_appl)
4307 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4308 			rack->r_ctl.rc_end_appl = rsm;
4309 		}
4310 		rsm->r_flags |= RACK_APP_LIMITED;
4311 		rack->r_ctl.rc_app_limited_cnt++;
4312 	}
4313 	/*
4314 	 * Now, we need to examine our pacing rate multipliers.
4315 	 * If its under 100%, we need to kick it back up to
4316 	 * 100%. We also don't let it be over our "max" above
4317 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4318 	 * Note setting clamp_atexit_prtt to 0 has the effect
4319 	 * of setting CA/SS to 100% always at exit (which is
4320 	 * the default behavior).
4321 	 */
4322 	if (rack_probertt_clear_is) {
4323 		rack->rc_gp_incr = 0;
4324 		rack->rc_gp_bwred = 0;
4325 		rack->rc_gp_timely_inc_cnt = 0;
4326 		rack->rc_gp_timely_dec_cnt = 0;
4327 	}
4328 	/* Do we do any clamping at exit? */
4329 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4330 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4331 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4332 	}
4333 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4334 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4335 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4336 	}
4337 	/*
4338 	 * Lets set rtt_diff to 0, so that we will get a "boost"
4339 	 * after exiting.
4340 	 */
4341 	rack->r_ctl.rc_rtt_diff = 0;
4342 
4343 	/* Clear all flags so we start fresh */
4344 	rack->rc_tp->t_bytes_acked = 0;
4345 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4346 	/*
4347 	 * If configured to, set the cwnd and ssthresh to
4348 	 * our targets.
4349 	 */
4350 	if (rack_probe_rtt_sets_cwnd) {
4351 		uint64_t ebdp;
4352 		uint32_t setto;
4353 
4354 		/* Set ssthresh so we get into CA once we hit our target */
4355 		if (rack_probertt_use_min_rtt_exit == 1) {
4356 			/* Set to min rtt */
4357 			rack_set_prtt_target(rack, segsiz,
4358 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4359 		} else if (rack_probertt_use_min_rtt_exit == 2) {
4360 			/* Set to current gp rtt */
4361 			rack_set_prtt_target(rack, segsiz,
4362 					     rack->r_ctl.rc_gp_srtt);
4363 		} else if (rack_probertt_use_min_rtt_exit == 3) {
4364 			/* Set to entry gp rtt */
4365 			rack_set_prtt_target(rack, segsiz,
4366 					     rack->r_ctl.rc_entry_gp_rtt);
4367 		} else {
4368 			uint64_t sum;
4369 			uint32_t setval;
4370 
4371 			sum = rack->r_ctl.rc_entry_gp_rtt;
4372 			sum *= 10;
4373 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4374 			if (sum >= 20) {
4375 				/*
4376 				 * A highly buffered path needs
4377 				 * cwnd space for timely to work.
4378 				 * Lets set things up as if
4379 				 * we are heading back here again.
4380 				 */
4381 				setval = rack->r_ctl.rc_entry_gp_rtt;
4382 			} else if (sum >= 15) {
4383 				/*
4384 				 * Lets take the smaller of the
4385 				 * two since we are just somewhat
4386 				 * buffered.
4387 				 */
4388 				setval = rack->r_ctl.rc_gp_srtt;
4389 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
4390 					setval = rack->r_ctl.rc_entry_gp_rtt;
4391 			} else {
4392 				/*
4393 				 * Here we are not highly buffered
4394 				 * and should pick the min we can to
4395 				 * keep from causing loss.
4396 				 */
4397 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4398 			}
4399 			rack_set_prtt_target(rack, segsiz,
4400 					     setval);
4401 		}
4402 		if (rack_probe_rtt_sets_cwnd > 1) {
4403 			/* There is a percentage here to boost */
4404 			ebdp = rack->r_ctl.rc_target_probertt_flight;
4405 			ebdp *= rack_probe_rtt_sets_cwnd;
4406 			ebdp /= 100;
4407 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4408 		} else
4409 			setto = rack->r_ctl.rc_target_probertt_flight;
4410 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4411 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4412 			/* Enforce a min */
4413 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4414 		}
4415 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
4416 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4417 	}
4418 	rack_log_rtt_shrinks(rack,  us_cts,
4419 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4420 			     __LINE__, RACK_RTTS_EXITPROBE);
4421 	/* Clear times last so log has all the info */
4422 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4423 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4424 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4425 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
4426 }
4427 
4428 static void
4429 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4430 {
4431 	/* Check in on probe-rtt */
4432 
4433 	if (rack->rc_gp_filled == 0) {
4434 		/* We do not do p-rtt unless we have gp measurements */
4435 		return;
4436 	}
4437 	if (rack->in_probe_rtt) {
4438 		uint64_t no_overflow;
4439 		uint32_t endtime, must_stay;
4440 
4441 		if (rack->r_ctl.rc_went_idle_time &&
4442 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4443 			/*
4444 			 * We went idle during prtt, just exit now.
4445 			 */
4446 			rack_exit_probertt(rack, us_cts);
4447 		} else if (rack_probe_rtt_safety_val &&
4448 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4449 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4450 			/*
4451 			 * Probe RTT safety value triggered!
4452 			 */
4453 			rack_log_rtt_shrinks(rack,  us_cts,
4454 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4455 					     __LINE__, RACK_RTTS_SAFETY);
4456 			rack_exit_probertt(rack, us_cts);
4457 		}
4458 		/* Calculate the max we will wait */
4459 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4460 		if (rack->rc_highly_buffered)
4461 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4462 		/* Calculate the min we must wait */
4463 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4464 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4465 		    TSTMP_LT(us_cts, endtime)) {
4466 			uint32_t calc;
4467 			/* Do we lower more? */
4468 no_exit:
4469 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4470 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4471 			else
4472 				calc = 0;
4473 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4474 			if (calc) {
4475 				/* Maybe */
4476 				calc *= rack_per_of_gp_probertt_reduce;
4477 				if (calc > rack_per_of_gp_probertt)
4478 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4479 				else
4480 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4481 				/* Limit it too */
4482 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4483 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4484 			}
4485 			/* We must reach target or the time set */
4486 			return;
4487 		}
4488 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
4489 			if ((TSTMP_LT(us_cts, must_stay) &&
4490 			     rack->rc_highly_buffered) ||
4491 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4492 			      rack->r_ctl.rc_target_probertt_flight)) {
4493 				/* We are not past the must_stay time */
4494 				goto no_exit;
4495 			}
4496 			rack_log_rtt_shrinks(rack,  us_cts,
4497 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4498 					     __LINE__, RACK_RTTS_REACHTARGET);
4499 			rack->r_ctl.rc_time_probertt_starts = us_cts;
4500 			if (rack->r_ctl.rc_time_probertt_starts == 0)
4501 				rack->r_ctl.rc_time_probertt_starts = 1;
4502 			/* Restore back to our rate we want to pace at in prtt */
4503 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4504 		}
4505 		/*
4506 		 * Setup our end time, some number of gp_srtts plus 200ms.
4507 		 */
4508 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4509 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4510 		if (rack_probertt_gpsrtt_cnt_div)
4511 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4512 		else
4513 			endtime = 0;
4514 		endtime += rack_min_probertt_hold;
4515 		endtime += rack->r_ctl.rc_time_probertt_starts;
4516 		if (TSTMP_GEQ(us_cts,  endtime)) {
4517 			/* yes, exit probertt */
4518 			rack_exit_probertt(rack, us_cts);
4519 		}
4520 
4521 	} else if ((rack->rc_skip_timely == 0) &&
4522 		   (TSTMP_GT(us_cts, rack->r_ctl.rc_lower_rtt_us_cts)) &&
4523 		   ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt)) {
4524 		/* Go into probertt, its been too long since we went lower */
4525 		rack_enter_probertt(rack, us_cts);
4526 	}
4527 }
4528 
4529 static void
4530 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4531 		       uint32_t rtt, int32_t rtt_diff)
4532 {
4533 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4534 	uint32_t losses;
4535 
4536 	if ((rack->rc_gp_dyn_mul == 0) ||
4537 	    (rack->use_fixed_rate) ||
4538 	    (rack->in_probe_rtt) ||
4539 	    (rack->rc_always_pace == 0)) {
4540 		/* No dynamic GP multiplier in play */
4541 		return;
4542 	}
4543 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4544 	cur_bw = rack_get_bw(rack);
4545 	/* Calculate our up and down range */
4546 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4547 	up_bnd /= 100;
4548 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4549 
4550 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4551 	subfr /= 100;
4552 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4553 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4554 		/*
4555 		 * This is the case where our RTT is above
4556 		 * the max target and we have been configured
4557 		 * to just do timely no bonus up stuff in that case.
4558 		 *
4559 		 * There are two configurations, set to 1, and we
4560 		 * just do timely if we are over our max. If its
4561 		 * set above 1 then we slam the multipliers down
4562 		 * to 100 and then decrement per timely.
4563 		 */
4564 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4565 				__LINE__, 3);
4566 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4567 			rack_validate_multipliers_at_or_below_100(rack);
4568 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4569 	} else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4570 		/*
4571 		 * We are decreasing this is a bit complicated this
4572 		 * means we are loosing ground. This could be
4573 		 * because another flow entered and we are competing
4574 		 * for b/w with it. This will push the RTT up which
4575 		 * makes timely unusable unless we want to get shoved
4576 		 * into a corner and just be backed off (the age
4577 		 * old problem with delay based CC).
4578 		 *
4579 		 * On the other hand if it was a route change we
4580 		 * would like to stay somewhat contained and not
4581 		 * blow out the buffers.
4582 		 */
4583 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4584 				__LINE__, 3);
4585 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4586 		if (rack->rc_gp_bwred == 0) {
4587 			/* Go into reduction counting */
4588 			rack->rc_gp_bwred = 1;
4589 			rack->rc_gp_timely_dec_cnt = 0;
4590 		}
4591 		if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4592 			/*
4593 			 * Push another time with a faster pacing
4594 			 * to try to gain back (we include override to
4595 			 * get a full raise factor).
4596 			 */
4597 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4598 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4599 			    (timely_says == 0) ||
4600 			    (rack_down_raise_thresh == 0)) {
4601 				/*
4602 				 * Do an override up in b/w if we were
4603 				 * below the threshold or if the threshold
4604 				 * is zero we always do the raise.
4605 				 */
4606 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4607 			} else {
4608 				/* Log it stays the same */
4609 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4610 						__LINE__, 11);
4611 			}
4612 			rack->rc_gp_timely_dec_cnt++;
4613 			/* We are not incrementing really no-count */
4614 			rack->rc_gp_incr = 0;
4615 			rack->rc_gp_timely_inc_cnt = 0;
4616 		} else {
4617 			/*
4618 			 * Lets just use the RTT
4619 			 * information and give up
4620 			 * pushing.
4621 			 */
4622 			goto use_timely;
4623 		}
4624 	} else if ((timely_says != 2) &&
4625 		    !losses &&
4626 		    (last_bw_est > up_bnd)) {
4627 		/*
4628 		 * We are increasing b/w lets keep going, updating
4629 		 * our b/w and ignoring any timely input, unless
4630 		 * of course we are at our max raise (if there is one).
4631 		 */
4632 
4633 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4634 				__LINE__, 3);
4635 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4636 		if (rack->rc_gp_saw_ss &&
4637 		    rack->r_ctl.rack_per_upper_bound_ss &&
4638 		     (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4639 			    /*
4640 			     * In cases where we can't go higher
4641 			     * we should just use timely.
4642 			     */
4643 			    goto use_timely;
4644 		}
4645 		if (rack->rc_gp_saw_ca &&
4646 		    rack->r_ctl.rack_per_upper_bound_ca &&
4647 		    (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4648 			    /*
4649 			     * In cases where we can't go higher
4650 			     * we should just use timely.
4651 			     */
4652 			    goto use_timely;
4653 		}
4654 		rack->rc_gp_bwred = 0;
4655 		rack->rc_gp_timely_dec_cnt = 0;
4656 		/* You get a set number of pushes if timely is trying to reduce */
4657 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4658 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4659 		} else {
4660 			/* Log it stays the same */
4661 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4662 			    __LINE__, 12);
4663 		}
4664 		return;
4665 	} else {
4666 		/*
4667 		 * We are staying between the lower and upper range bounds
4668 		 * so use timely to decide.
4669 		 */
4670 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4671 				__LINE__, 3);
4672 use_timely:
4673 		if (timely_says) {
4674 			rack->rc_gp_incr = 0;
4675 			rack->rc_gp_timely_inc_cnt = 0;
4676 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4677 			    !losses &&
4678 			    (last_bw_est < low_bnd)) {
4679 				/* We are loosing ground */
4680 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4681 				rack->rc_gp_timely_dec_cnt++;
4682 				/* We are not incrementing really no-count */
4683 				rack->rc_gp_incr = 0;
4684 				rack->rc_gp_timely_inc_cnt = 0;
4685 			} else
4686 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4687 		} else {
4688 			rack->rc_gp_bwred = 0;
4689 			rack->rc_gp_timely_dec_cnt = 0;
4690 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4691 		}
4692 	}
4693 }
4694 
4695 static int32_t
4696 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4697 {
4698 	int32_t timely_says;
4699 	uint64_t log_mult, log_rtt_a_diff;
4700 
4701 	log_rtt_a_diff = rtt;
4702 	log_rtt_a_diff <<= 32;
4703 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4704 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4705 		    rack_gp_rtt_maxmul)) {
4706 		/* Reduce the b/w multiplier */
4707 		timely_says = 2;
4708 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4709 		log_mult <<= 32;
4710 		log_mult |= prev_rtt;
4711 		rack_log_timely(rack,  timely_says, log_mult,
4712 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4713 				log_rtt_a_diff, __LINE__, 4);
4714 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4715 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4716 			    max(rack_gp_rtt_mindiv , 1)))) {
4717 		/* Increase the b/w multiplier */
4718 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4719 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4720 			 max(rack_gp_rtt_mindiv , 1));
4721 		log_mult <<= 32;
4722 		log_mult |= prev_rtt;
4723 		timely_says = 0;
4724 		rack_log_timely(rack,  timely_says, log_mult ,
4725 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4726 				log_rtt_a_diff, __LINE__, 5);
4727 	} else {
4728 		/*
4729 		 * Use a gradient to find it the timely gradient
4730 		 * is:
4731 		 * grad = rc_rtt_diff / min_rtt;
4732 		 *
4733 		 * anything below or equal to 0 will be
4734 		 * a increase indication. Anything above
4735 		 * zero is a decrease. Note we take care
4736 		 * of the actual gradient calculation
4737 		 * in the reduction (its not needed for
4738 		 * increase).
4739 		 */
4740 		log_mult = prev_rtt;
4741 		if (rtt_diff <= 0) {
4742 			/*
4743 			 * Rttdiff is less than zero, increase the
4744 			 * b/w multiplier (its 0 or negative)
4745 			 */
4746 			timely_says = 0;
4747 			rack_log_timely(rack,  timely_says, log_mult,
4748 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4749 		} else {
4750 			/* Reduce the b/w multiplier */
4751 			timely_says = 1;
4752 			rack_log_timely(rack,  timely_says, log_mult,
4753 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4754 		}
4755 	}
4756 	return (timely_says);
4757 }
4758 
4759 static __inline int
4760 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4761 {
4762 	if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4763 	    SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4764 		/**
4765 		 * This covers the case that the
4766 		 * resent is completely inside
4767 		 * the gp range or up to it.
4768 		 *      |----------------|
4769 		 *      |-----| <or>
4770 		 *            |----|
4771 		 *            <or>   |---|
4772 		 */
4773 		return (1);
4774 	} else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4775 		   SEQ_GT(rsm->r_end, tp->gput_seq)){
4776 		/**
4777 		 * This covers the case of
4778 		 *      |--------------|
4779 		 *  |-------->|
4780 		 */
4781 		return (1);
4782 	} else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4783 		   SEQ_LT(rsm->r_start, tp->gput_ack) &&
4784 		   SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4785 
4786 		/**
4787 		 * This covers the case of
4788 		 *      |--------------|
4789 		 *              |-------->|
4790 		 */
4791 		return (1);
4792 	}
4793 	return (0);
4794 }
4795 
4796 static __inline void
4797 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4798 {
4799 
4800 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
4801 		return;
4802 	/*
4803 	 * We have a Goodput measurement in progress. Mark
4804 	 * the send if its within the window. If its not
4805 	 * in the window make sure it does not have the mark.
4806 	 */
4807 	if (rack_in_gp_window(tp, rsm))
4808 		rsm->r_flags |= RACK_IN_GP_WIN;
4809 	else
4810 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4811 }
4812 
4813 static __inline void
4814 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4815 {
4816 	/* A GP measurement is ending, clear all marks on the send map*/
4817 	struct rack_sendmap *rsm = NULL;
4818 
4819 	rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4820 	if (rsm == NULL) {
4821 		rsm = tqhash_min(rack->r_ctl.tqh);
4822 	}
4823 	/* Nothing left? */
4824 	while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4825 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4826 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4827 	}
4828 }
4829 
4830 
4831 static __inline void
4832 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4833 {
4834 	struct rack_sendmap *rsm = NULL;
4835 
4836 	if (tp->snd_una == tp->snd_max) {
4837 		/* Nothing outstanding yet, nothing to do here */
4838 		return;
4839 	}
4840 	if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4841 		/*
4842 		 * We are measuring ahead of some outstanding
4843 		 * data. We need to walk through up until we get
4844 		 * to gp_seq marking so that no rsm is set incorrectly
4845 		 * with RACK_IN_GP_WIN.
4846 		 */
4847 		rsm = tqhash_min(rack->r_ctl.tqh);
4848 		while (rsm != NULL) {
4849 			rack_mark_in_gp_win(tp, rsm);
4850 			if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4851 				break;
4852 			rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4853 		}
4854 	}
4855 	if (rsm == NULL) {
4856 		/*
4857 		 * Need to find the GP seq, if rsm is
4858 		 * set we stopped as we hit it.
4859 		 */
4860 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4861 		if (rsm == NULL)
4862 			return;
4863 		rack_mark_in_gp_win(tp, rsm);
4864 	}
4865 	/*
4866 	 * Now we may need to mark already sent rsm, ahead of
4867 	 * gput_seq in the window since they may have been sent
4868 	 * *before* we started our measurment. The rsm, if non-null
4869 	 * has been marked (note if rsm would have been NULL we would have
4870 	 * returned in the previous block). So we go to the next, and continue
4871 	 * until we run out of entries or we exceed the gp_ack value.
4872 	 */
4873 	rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4874 	while (rsm) {
4875 		rack_mark_in_gp_win(tp, rsm);
4876 		if (SEQ_GT(rsm->r_end, tp->gput_ack))
4877 			break;
4878 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4879 	}
4880 }
4881 
4882 static void
4883 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)
4884 {
4885 	if (tcp_bblogging_on(rack->rc_tp)) {
4886 		union tcp_log_stackspecific log;
4887 		struct timeval tv;
4888 
4889 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4890 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4891 		log.u_bbr.flex1 = add_part;
4892 		log.u_bbr.flex2 = sub_part;
4893 		log.u_bbr.flex3 = rack_wma_divisor;
4894 		log.u_bbr.flex4 = srtt;
4895 		log.u_bbr.flex7 = (uint16_t)line;
4896 		log.u_bbr.flex8 = meth;
4897 		log.u_bbr.delRate = rack->r_ctl.gp_bw;
4898 		log.u_bbr.cur_del_rate = meas_bw;
4899 		log.u_bbr.rttProp = utim;
4900 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4901 		    &rack->rc_inp->inp_socket->so_rcv,
4902 		    &rack->rc_inp->inp_socket->so_snd,
4903 		    BBR_LOG_THRESH_CALC, 0,
4904 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4905 	}
4906 }
4907 
4908 static void
4909 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4910 			    tcp_seq th_ack, int line, uint8_t quality)
4911 {
4912 	uint64_t tim, bytes_ps, stim, utim;
4913 	uint32_t segsiz, bytes, reqbytes, us_cts;
4914 	int32_t gput, new_rtt_diff, timely_says;
4915 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4916 	int did_add = 0;
4917 
4918 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4919 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4920 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4921 		tim = us_cts - tp->gput_ts;
4922 	else
4923 		tim = 0;
4924 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4925 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4926 	else
4927 		stim = 0;
4928 	/*
4929 	 * Use the larger of the send time or ack time. This prevents us
4930 	 * from being influenced by ack artifacts to come up with too
4931 	 * high of measurement. Note that since we are spanning over many more
4932 	 * bytes in most of our measurements hopefully that is less likely to
4933 	 * occur.
4934 	 */
4935 	if (tim > stim)
4936 		utim = max(tim, 1);
4937 	else
4938 		utim = max(stim, 1);
4939 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4940 	rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
4941 	if ((tim == 0) && (stim == 0)) {
4942 		/*
4943 		 * Invalid measurement time, maybe
4944 		 * all on one ack/one send?
4945 		 */
4946 		bytes = 0;
4947 		bytes_ps = 0;
4948 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4949 					   0, 0, 0, 10, __LINE__, NULL, quality);
4950 		goto skip_measurement;
4951 	}
4952 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4953 		/* We never made a us_rtt measurement? */
4954 		bytes = 0;
4955 		bytes_ps = 0;
4956 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4957 					   0, 0, 0, 10, __LINE__, NULL, quality);
4958 		goto skip_measurement;
4959 	}
4960 	/*
4961 	 * Calculate the maximum possible b/w this connection
4962 	 * could have. We base our calculation on the lowest
4963 	 * rtt we have seen during the measurement and the
4964 	 * largest rwnd the client has given us in that time. This
4965 	 * forms a BDP that is the maximum that we could ever
4966 	 * get to the client. Anything larger is not valid.
4967 	 *
4968 	 * I originally had code here that rejected measurements
4969 	 * where the time was less than 1/2 the latest us_rtt.
4970 	 * But after thinking on that I realized its wrong since
4971 	 * say you had a 150Mbps or even 1Gbps link, and you
4972 	 * were a long way away.. example I am in Europe (100ms rtt)
4973 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4974 	 * bytes my time would be 1.2ms, and yet my rtt would say
4975 	 * the measurement was invalid the time was < 50ms. The
4976 	 * same thing is true for 150Mb (8ms of time).
4977 	 *
4978 	 * A better way I realized is to look at what the maximum
4979 	 * the connection could possibly do. This is gated on
4980 	 * the lowest RTT we have seen and the highest rwnd.
4981 	 * We should in theory never exceed that, if we are
4982 	 * then something on the path is storing up packets
4983 	 * and then feeding them all at once to our endpoint
4984 	 * messing up our measurement.
4985 	 */
4986 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4987 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4988 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4989 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4990 		/* No measurement can be made */
4991 		bytes = 0;
4992 		bytes_ps = 0;
4993 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4994 					   0, 0, 0, 10, __LINE__, NULL, quality);
4995 		goto skip_measurement;
4996 	} else
4997 		bytes = (th_ack - tp->gput_seq);
4998 	bytes_ps = (uint64_t)bytes;
4999 	/*
5000 	 * Don't measure a b/w for pacing unless we have gotten at least
5001 	 * an initial windows worth of data in this measurement interval.
5002 	 *
5003 	 * Small numbers of bytes get badly influenced by delayed ack and
5004 	 * other artifacts. Note we take the initial window or our
5005 	 * defined minimum GP (defaulting to 10 which hopefully is the
5006 	 * IW).
5007 	 */
5008 	if (rack->rc_gp_filled == 0) {
5009 		/*
5010 		 * The initial estimate is special. We
5011 		 * have blasted out an IW worth of packets
5012 		 * without a real valid ack ts results. We
5013 		 * then setup the app_limited_needs_set flag,
5014 		 * this should get the first ack in (probably 2
5015 		 * MSS worth) to be recorded as the timestamp.
5016 		 * We thus allow a smaller number of bytes i.e.
5017 		 * IW - 2MSS.
5018 		 */
5019 		reqbytes -= (2 * segsiz);
5020 		/* Also lets fill previous for our first measurement to be neutral */
5021 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5022 	}
5023 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
5024 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5025 					   rack->r_ctl.rc_app_limited_cnt,
5026 					   0, 0, 10, __LINE__, NULL, quality);
5027 		goto skip_measurement;
5028 	}
5029 	/*
5030 	 * We now need to calculate the Timely like status so
5031 	 * we can update (possibly) the b/w multipliers.
5032 	 */
5033 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
5034 	if (rack->rc_gp_filled == 0) {
5035 		/* No previous reading */
5036 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
5037 	} else {
5038 		if (rack->measure_saw_probe_rtt == 0) {
5039 			/*
5040 			 * We don't want a probertt to be counted
5041 			 * since it will be negative incorrectly. We
5042 			 * expect to be reducing the RTT when we
5043 			 * pace at a slower rate.
5044 			 */
5045 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
5046 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
5047 		}
5048 	}
5049 	timely_says = rack_make_timely_judgement(rack,
5050 	    rack->r_ctl.rc_gp_srtt,
5051 	    rack->r_ctl.rc_rtt_diff,
5052 	    rack->r_ctl.rc_prev_gp_srtt
5053 	);
5054 	bytes_ps *= HPTS_USEC_IN_SEC;
5055 	bytes_ps /= utim;
5056 	if (bytes_ps > rack->r_ctl.last_max_bw) {
5057 		/*
5058 		 * Something is on path playing
5059 		 * since this b/w is not possible based
5060 		 * on our BDP (highest rwnd and lowest rtt
5061 		 * we saw in the measurement window).
5062 		 *
5063 		 * Another option here would be to
5064 		 * instead skip the measurement.
5065 		 */
5066 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
5067 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
5068 					   11, __LINE__, NULL, quality);
5069 		bytes_ps = rack->r_ctl.last_max_bw;
5070 	}
5071 	/* We store gp for b/w in bytes per second */
5072 	if (rack->rc_gp_filled == 0) {
5073 		/* Initial measurement */
5074 		if (bytes_ps) {
5075 			rack->r_ctl.gp_bw = bytes_ps;
5076 			rack->rc_gp_filled = 1;
5077 			rack->r_ctl.num_measurements = 1;
5078 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5079 		} else {
5080 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5081 						   rack->r_ctl.rc_app_limited_cnt,
5082 						   0, 0, 10, __LINE__, NULL, quality);
5083 		}
5084 		if (tcp_in_hpts(rack->rc_tp) &&
5085 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5086 			/*
5087 			 * Ok we can't trust the pacer in this case
5088 			 * where we transition from un-paced to paced.
5089 			 * Or for that matter when the burst mitigation
5090 			 * was making a wild guess and got it wrong.
5091 			 * Stop the pacer and clear up all the aggregate
5092 			 * delays etc.
5093 			 */
5094 			tcp_hpts_remove(rack->rc_tp);
5095 			rack->r_ctl.rc_hpts_flags = 0;
5096 			rack->r_ctl.rc_last_output_to = 0;
5097 		}
5098 		did_add = 2;
5099 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5100 		/* Still a small number run an average */
5101 		rack->r_ctl.gp_bw += bytes_ps;
5102 		addpart = rack->r_ctl.num_measurements;
5103 		rack->r_ctl.num_measurements++;
5104 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5105 			/* We have collected enough to move forward */
5106 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5107 		}
5108 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5109 		did_add = 3;
5110 	} else {
5111 		/*
5112 		 * We want to take 1/wma of the goodput and add in to 7/8th
5113 		 * of the old value weighted by the srtt. So if your measurement
5114 		 * period is say 2 SRTT's long you would get 1/4 as the
5115 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
5116 		 *
5117 		 * But we must be careful not to take too much i.e. if the
5118 		 * srtt is say 20ms and the measurement is taken over
5119 		 * 400ms our weight would be 400/20 i.e. 20. On the
5120 		 * other hand if we get a measurement over 1ms with a
5121 		 * 10ms rtt we only want to take a much smaller portion.
5122 		 */
5123 		uint8_t meth;
5124 
5125 		if (rack->r_ctl.num_measurements < 0xff) {
5126 			rack->r_ctl.num_measurements++;
5127 		}
5128 		srtt = (uint64_t)tp->t_srtt;
5129 		if (srtt == 0) {
5130 			/*
5131 			 * Strange why did t_srtt go back to zero?
5132 			 */
5133 			if (rack->r_ctl.rc_rack_min_rtt)
5134 				srtt = rack->r_ctl.rc_rack_min_rtt;
5135 			else
5136 				srtt = HPTS_USEC_IN_MSEC;
5137 		}
5138 		/*
5139 		 * XXXrrs: Note for reviewers, in playing with
5140 		 * dynamic pacing I discovered this GP calculation
5141 		 * as done originally leads to some undesired results.
5142 		 * Basically you can get longer measurements contributing
5143 		 * too much to the WMA. Thus I changed it if you are doing
5144 		 * dynamic adjustments to only do the aportioned adjustment
5145 		 * if we have a very small (time wise) measurement. Longer
5146 		 * measurements just get there weight (defaulting to 1/8)
5147 		 * add to the WMA. We may want to think about changing
5148 		 * this to always do that for both sides i.e. dynamic
5149 		 * and non-dynamic... but considering lots of folks
5150 		 * were playing with this I did not want to change the
5151 		 * calculation per.se. without your thoughts.. Lawerence?
5152 		 * Peter??
5153 		 */
5154 		if (rack->rc_gp_dyn_mul == 0) {
5155 			subpart = rack->r_ctl.gp_bw * utim;
5156 			subpart /= (srtt * 8);
5157 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
5158 				/*
5159 				 * The b/w update takes no more
5160 				 * away then 1/2 our running total
5161 				 * so factor it in.
5162 				 */
5163 				addpart = bytes_ps * utim;
5164 				addpart /= (srtt * 8);
5165 				meth = 1;
5166 			} else {
5167 				/*
5168 				 * Don't allow a single measurement
5169 				 * to account for more than 1/2 of the
5170 				 * WMA. This could happen on a retransmission
5171 				 * where utim becomes huge compared to
5172 				 * srtt (multiple retransmissions when using
5173 				 * the sending rate which factors in all the
5174 				 * transmissions from the first one).
5175 				 */
5176 				subpart = rack->r_ctl.gp_bw / 2;
5177 				addpart = bytes_ps / 2;
5178 				meth = 2;
5179 			}
5180 			rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5181 			resid_bw = rack->r_ctl.gp_bw - subpart;
5182 			rack->r_ctl.gp_bw = resid_bw + addpart;
5183 			did_add = 1;
5184 		} else {
5185 			if ((utim / srtt) <= 1) {
5186 				/*
5187 				 * The b/w update was over a small period
5188 				 * of time. The idea here is to prevent a small
5189 				 * measurement time period from counting
5190 				 * too much. So we scale it based on the
5191 				 * time so it attributes less than 1/rack_wma_divisor
5192 				 * of its measurement.
5193 				 */
5194 				subpart = rack->r_ctl.gp_bw * utim;
5195 				subpart /= (srtt * rack_wma_divisor);
5196 				addpart = bytes_ps * utim;
5197 				addpart /= (srtt * rack_wma_divisor);
5198 				meth = 3;
5199 			} else {
5200 				/*
5201 				 * The scaled measurement was long
5202 				 * enough so lets just add in the
5203 				 * portion of the measurement i.e. 1/rack_wma_divisor
5204 				 */
5205 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5206 				addpart = bytes_ps / rack_wma_divisor;
5207 				meth = 4;
5208 			}
5209 			if ((rack->measure_saw_probe_rtt == 0) ||
5210 		            (bytes_ps > rack->r_ctl.gp_bw)) {
5211 				/*
5212 				 * For probe-rtt we only add it in
5213 				 * if its larger, all others we just
5214 				 * add in.
5215 				 */
5216 				did_add = 1;
5217 				rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5218 				resid_bw = rack->r_ctl.gp_bw - subpart;
5219 				rack->r_ctl.gp_bw = resid_bw + addpart;
5220 			}
5221 		}
5222 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5223 	}
5224 	/*
5225 	 * We only watch the growth of the GP during the initial startup
5226 	 * or first-slowstart that ensues. If we ever needed to watch
5227 	 * growth of gp outside of that period all we need to do is
5228 	 * remove the first clause of this if (rc_initial_ss_comp).
5229 	 */
5230 	if ((rack->rc_initial_ss_comp == 0) &&
5231 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG)) {
5232 		uint64_t gp_est;
5233 
5234 		gp_est = bytes_ps;
5235 		if (tcp_bblogging_on(rack->rc_tp)) {
5236 			union tcp_log_stackspecific log;
5237 			struct timeval tv;
5238 
5239 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5240 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5241 			log.u_bbr.flex1 = rack->r_ctl.current_round;
5242 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
5243 			log.u_bbr.delRate = gp_est;
5244 			log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5245 			log.u_bbr.flex8 = 41;
5246 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5247 					    0, &log, false, NULL, __func__, __LINE__,&tv);
5248 		}
5249 		if ((rack->r_ctl.num_measurements == RACK_REQ_AVG) ||
5250 		    (rack->r_ctl.last_gpest == 0)) {
5251 			/*
5252 			 * The round we get our measurement averaging going
5253 			 * is the base round so it always is the source point
5254 			 * for when we had our first increment. From there on
5255 			 * we only record the round that had a rise.
5256 			 */
5257 			rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5258 			rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5259 		} else if (gp_est >= rack->r_ctl.last_gpest) {
5260 			/*
5261 			 * Test to see if its gone up enough
5262 			 * to set the round count up to now. Note
5263 			 * that on the seeding of the 4th measurement we
5264 			 */
5265 			gp_est *= 1000;
5266 			gp_est /= rack->r_ctl.last_gpest;
5267 			if ((uint32_t)gp_est > rack->r_ctl.gp_gain_req) {
5268 				/*
5269 				 * We went up enough to record the round.
5270 				 */
5271 				if (tcp_bblogging_on(rack->rc_tp)) {
5272 					union tcp_log_stackspecific log;
5273 					struct timeval tv;
5274 
5275 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5276 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5277 					log.u_bbr.flex1 = rack->r_ctl.current_round;
5278 					log.u_bbr.flex2 = (uint32_t)gp_est;
5279 					log.u_bbr.flex3 = rack->r_ctl.gp_gain_req;
5280 					log.u_bbr.delRate = gp_est;
5281 					log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5282 					log.u_bbr.flex8 = 42;
5283 					(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5284 							    0, &log, false, NULL, __func__, __LINE__,&tv);
5285 				}
5286 				rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5287 				if (rack->r_ctl.use_gp_not_last == 1)
5288 					rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5289 				else
5290 					rack->r_ctl.last_gpest = bytes_ps;
5291 			}
5292 		}
5293 	}
5294 	if ((rack->gp_ready == 0) &&
5295 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5296 		/* We have enough measurements now */
5297 		rack->gp_ready = 1;
5298 		if (rack->dgp_on ||
5299 		    rack->rack_hibeta)
5300 			rack_set_cc_pacing(rack);
5301 		if (rack->defer_options)
5302 			rack_apply_deferred_options(rack);
5303 	}
5304 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5305 				   rack_get_bw(rack), 22, did_add, NULL, quality);
5306 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
5307 
5308 	if ((rack->measure_saw_probe_rtt == 0) &&
5309 	    rack->rc_gp_rtt_set) {
5310 		if (rack->rc_skip_timely == 0) {
5311 			rack_update_multiplier(rack, timely_says, bytes_ps,
5312 					       rack->r_ctl.rc_gp_srtt,
5313 					       rack->r_ctl.rc_rtt_diff);
5314 		}
5315 	}
5316 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5317 				   rack_get_bw(rack), 3, line, NULL, quality);
5318 	rack_log_pacing_delay_calc(rack,
5319 				   bytes, /* flex2 */
5320 				   tim, /* flex1 */
5321 				   bytes_ps, /* bw_inuse */
5322 				   rack->r_ctl.gp_bw, /* delRate */
5323 				   rack_get_lt_bw(rack), /* rttProp */
5324 				   20, line, NULL, 0);
5325 	/* reset the gp srtt and setup the new prev */
5326 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5327 	/* Record the lost count for the next measurement */
5328 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5329 skip_measurement:
5330 	/*
5331 	 * We restart our diffs based on the gpsrtt in the
5332 	 * measurement window.
5333 	 */
5334 	rack->rc_gp_rtt_set = 0;
5335 	rack->rc_gp_saw_rec = 0;
5336 	rack->rc_gp_saw_ca = 0;
5337 	rack->rc_gp_saw_ss = 0;
5338 	rack->rc_dragged_bottom = 0;
5339 	if (quality == RACK_QUALITY_HIGH) {
5340 		/*
5341 		 * Gput in the stats world is in kbps where bytes_ps is
5342 		 * bytes per second so we do ((x * 8)/ 1000).
5343 		 */
5344 		gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5345 #ifdef STATS
5346 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5347 					 gput);
5348 		/*
5349 		 * XXXLAS: This is a temporary hack, and should be
5350 		 * chained off VOI_TCP_GPUT when stats(9) grows an
5351 		 * API to deal with chained VOIs.
5352 		 */
5353 		if (tp->t_stats_gput_prev > 0)
5354 			stats_voi_update_abs_s32(tp->t_stats,
5355 						 VOI_TCP_GPUT_ND,
5356 						 ((gput - tp->t_stats_gput_prev) * 100) /
5357 						 tp->t_stats_gput_prev);
5358 #endif
5359 		tp->t_stats_gput_prev = gput;
5360 	}
5361 	tp->t_flags &= ~TF_GPUTINPROG;
5362 	/*
5363 	 * Now are we app limited now and there is space from where we
5364 	 * were to where we want to go?
5365 	 *
5366 	 * We don't do the other case i.e. non-applimited here since
5367 	 * the next send will trigger us picking up the missing data.
5368 	 */
5369 	if (rack->r_ctl.rc_first_appl &&
5370 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
5371 	    rack->r_ctl.rc_app_limited_cnt &&
5372 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5373 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5374 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5375 		/*
5376 		 * Yep there is enough outstanding to make a measurement here.
5377 		 */
5378 		struct rack_sendmap *rsm;
5379 
5380 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5381 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5382 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5383 		rack->app_limited_needs_set = 0;
5384 		tp->gput_seq = th_ack;
5385 		if (rack->in_probe_rtt)
5386 			rack->measure_saw_probe_rtt = 1;
5387 		else if ((rack->measure_saw_probe_rtt) &&
5388 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5389 			rack->measure_saw_probe_rtt = 0;
5390 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5391 			/* There is a full window to gain info from */
5392 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5393 		} else {
5394 			/* We can only measure up to the applimited point */
5395 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5396 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5397 				/*
5398 				 * We don't have enough to make a measurement.
5399 				 */
5400 				tp->t_flags &= ~TF_GPUTINPROG;
5401 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5402 							   0, 0, 0, 6, __LINE__, NULL, quality);
5403 				return;
5404 			}
5405 		}
5406 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
5407 			/*
5408 			 * We will get no more data into the SB
5409 			 * this means we need to have the data available
5410 			 * before we start a measurement.
5411 			 */
5412 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5413 				/* Nope not enough data. */
5414 				return;
5415 			}
5416 		}
5417 		tp->t_flags |= TF_GPUTINPROG;
5418 		/*
5419 		 * Now we need to find the timestamp of the send at tp->gput_seq
5420 		 * for the send based measurement.
5421 		 */
5422 		rack->r_ctl.rc_gp_cumack_ts = 0;
5423 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5424 		if (rsm) {
5425 			/* Ok send-based limit is set */
5426 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5427 				/*
5428 				 * Move back to include the earlier part
5429 				 * so our ack time lines up right (this may
5430 				 * make an overlapping measurement but thats
5431 				 * ok).
5432 				 */
5433 				tp->gput_seq = rsm->r_start;
5434 			}
5435 			if (rsm->r_flags & RACK_ACKED) {
5436 				struct rack_sendmap *nrsm;
5437 
5438 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5439 				tp->gput_seq = rsm->r_end;
5440 				nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5441 				if (nrsm)
5442 					rsm = nrsm;
5443 				else {
5444 					rack->app_limited_needs_set = 1;
5445 				}
5446 			} else
5447 				rack->app_limited_needs_set = 1;
5448 			/* We always go from the first send */
5449 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5450 		} else {
5451 			/*
5452 			 * If we don't find the rsm due to some
5453 			 * send-limit set the current time, which
5454 			 * basically disables the send-limit.
5455 			 */
5456 			struct timeval tv;
5457 
5458 			microuptime(&tv);
5459 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5460 		}
5461 		rack_tend_gp_marks(tp, rack);
5462 		rack_log_pacing_delay_calc(rack,
5463 					   tp->gput_seq,
5464 					   tp->gput_ack,
5465 					   (uintptr_t)rsm,
5466 					   tp->gput_ts,
5467 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5468 					   9,
5469 					   __LINE__, rsm, quality);
5470 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5471 	} else {
5472 		/*
5473 		 * To make sure proper timestamp merging occurs, we need to clear
5474 		 * all GP marks if we don't start a measurement.
5475 		 */
5476 		rack_clear_gp_marks(tp, rack);
5477 	}
5478 }
5479 
5480 /*
5481  * CC wrapper hook functions
5482  */
5483 static void
5484 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5485     uint16_t type, int32_t post_recovery)
5486 {
5487 	uint32_t prior_cwnd, acked;
5488 	struct tcp_log_buffer *lgb = NULL;
5489 	uint8_t labc_to_use, quality;
5490 
5491 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5492 	tp->t_ccv.nsegs = nsegs;
5493 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5494 	if ((post_recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5495 		uint32_t max;
5496 
5497 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5498 		if (tp->t_ccv.bytes_this_ack > max) {
5499 			tp->t_ccv.bytes_this_ack = max;
5500 		}
5501 	}
5502 #ifdef STATS
5503 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5504 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5505 #endif
5506 	if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5507 		/*
5508 		 * We will ack all the data, time to end any
5509 		 * lt_bw_up we have running until something
5510 		 * new is sent. Note we need to use the actual
5511 		 * ack_rcv_time which with pacing may be different.
5512 		 */
5513 		uint64_t tmark;
5514 
5515 		rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5516 		rack->r_ctl.lt_seq = tp->snd_max;
5517 		tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
5518 		if (tmark >= rack->r_ctl.lt_timemark) {
5519 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
5520 		}
5521 		rack->r_ctl.lt_timemark = tmark;
5522 		rack->lt_bw_up = 0;
5523 	}
5524 	quality = RACK_QUALITY_NONE;
5525 	if ((tp->t_flags & TF_GPUTINPROG) &&
5526 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5527 		/* Measure the Goodput */
5528 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5529 	}
5530 	/* Which way our we limited, if not cwnd limited no advance in CA */
5531 	if (tp->snd_cwnd <= tp->snd_wnd)
5532 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
5533 	else
5534 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5535 	if (tp->snd_cwnd > tp->snd_ssthresh) {
5536 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5537 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5538 		/* For the setting of a window past use the actual scwnd we are using */
5539 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5540 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5541 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5542 		}
5543 	} else {
5544 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5545 		tp->t_bytes_acked = 0;
5546 	}
5547 	prior_cwnd = tp->snd_cwnd;
5548 	if ((post_recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5549 	    (rack_client_low_buf && rack->client_bufferlvl &&
5550 	    (rack->client_bufferlvl < rack_client_low_buf)))
5551 		labc_to_use = rack->rc_labc;
5552 	else
5553 		labc_to_use = rack_max_abc_post_recovery;
5554 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5555 		union tcp_log_stackspecific log;
5556 		struct timeval tv;
5557 
5558 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5559 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5560 		log.u_bbr.flex1 = th_ack;
5561 		log.u_bbr.flex2 = tp->t_ccv.flags;
5562 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5563 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5564 		log.u_bbr.flex5 = labc_to_use;
5565 		log.u_bbr.flex6 = prior_cwnd;
5566 		log.u_bbr.flex7 = V_tcp_do_newsack;
5567 		log.u_bbr.flex8 = 1;
5568 		lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5569 				     0, &log, false, NULL, __func__, __LINE__,&tv);
5570 	}
5571 	if (CC_ALGO(tp)->ack_received != NULL) {
5572 		/* XXXLAS: Find a way to live without this */
5573 		tp->t_ccv.curack = th_ack;
5574 		tp->t_ccv.labc = labc_to_use;
5575 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5576 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5577 	}
5578 	if (lgb) {
5579 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5580 	}
5581 	if (rack->r_must_retran) {
5582 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5583 			/*
5584 			 * We now are beyond the rxt point so lets disable
5585 			 * the flag.
5586 			 */
5587 			rack->r_ctl.rc_out_at_rto = 0;
5588 			rack->r_must_retran = 0;
5589 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5590 			/*
5591 			 * Only decrement the rc_out_at_rto if the cwnd advances
5592 			 * at least a whole segment. Otherwise next time the peer
5593 			 * acks, we won't be able to send this generaly happens
5594 			 * when we are in Congestion Avoidance.
5595 			 */
5596 			if (acked <= rack->r_ctl.rc_out_at_rto){
5597 				rack->r_ctl.rc_out_at_rto -= acked;
5598 			} else {
5599 				rack->r_ctl.rc_out_at_rto = 0;
5600 			}
5601 		}
5602 	}
5603 #ifdef STATS
5604 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5605 #endif
5606 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5607 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5608 	}
5609 	if ((rack->rc_initial_ss_comp == 0) &&
5610 	    (tp->snd_cwnd >= tp->snd_ssthresh)) {
5611 		/*
5612 		 * The cwnd has grown beyond ssthresh we have
5613 		 * entered ca and completed our first Slowstart.
5614 		 */
5615 		rack->rc_initial_ss_comp = 1;
5616 	}
5617 }
5618 
5619 static void
5620 tcp_rack_partialack(struct tcpcb *tp)
5621 {
5622 	struct tcp_rack *rack;
5623 
5624 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5625 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5626 	/*
5627 	 * If we are doing PRR and have enough
5628 	 * room to send <or> we are pacing and prr
5629 	 * is disabled we will want to see if we
5630 	 * can send data (by setting r_wanted_output to
5631 	 * true).
5632 	 */
5633 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5634 	    rack->rack_no_prr)
5635 		rack->r_wanted_output = 1;
5636 }
5637 
5638 static void
5639 rack_exit_recovery(struct tcpcb *tp, struct tcp_rack *rack, int how)
5640 {
5641 	/*
5642 	 * Now exit recovery.
5643 	 */
5644 	EXIT_RECOVERY(tp->t_flags);
5645 }
5646 
5647 static void
5648 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
5649 {
5650 	struct tcp_rack *rack;
5651 	uint32_t orig_cwnd;
5652 
5653 	orig_cwnd = tp->snd_cwnd;
5654 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5655 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5656 	/* only alert CC if we alerted when we entered */
5657 	if (CC_ALGO(tp)->post_recovery != NULL) {
5658 		tp->t_ccv.curack = th_ack;
5659 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
5660 		if (tp->snd_cwnd < tp->snd_ssthresh) {
5661 			/*
5662 			 * Rack has burst control and pacing
5663 			 * so lets not set this any lower than
5664 			 * snd_ssthresh per RFC-6582 (option 2).
5665 			 */
5666 			tp->snd_cwnd = tp->snd_ssthresh;
5667 		}
5668 	}
5669 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5670 		union tcp_log_stackspecific log;
5671 		struct timeval tv;
5672 
5673 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5674 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5675 		log.u_bbr.flex1 = th_ack;
5676 		log.u_bbr.flex2 = tp->t_ccv.flags;
5677 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5678 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5679 		log.u_bbr.flex5 = V_tcp_abc_l_var;
5680 		log.u_bbr.flex6 = orig_cwnd;
5681 		log.u_bbr.flex7 = V_tcp_do_newsack;
5682 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
5683 		log.u_bbr.flex8 = 2;
5684 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5685 			       0, &log, false, NULL, __func__, __LINE__, &tv);
5686 	}
5687 	if ((rack->rack_no_prr == 0) &&
5688 	    (rack->no_prr_addback == 0) &&
5689 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
5690 		/*
5691 		 * Suck the next prr cnt back into cwnd, but
5692 		 * only do that if we are not application limited.
5693 		 */
5694 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
5695 			/*
5696 			 * We are allowed to add back to the cwnd the amount we did
5697 			 * not get out if:
5698 			 * a) no_prr_addback is off.
5699 			 * b) we are not app limited
5700 			 * c) we are doing prr
5701 			 * <and>
5702 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
5703 			 */
5704 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
5705 					    rack->r_ctl.rc_prr_sndcnt);
5706 		}
5707 		rack->r_ctl.rc_prr_sndcnt = 0;
5708 		rack_log_to_prr(rack, 1, 0, __LINE__);
5709 	}
5710 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
5711 	tp->snd_recover = tp->snd_una;
5712 	if (rack->r_ctl.dsack_persist) {
5713 		rack->r_ctl.dsack_persist--;
5714 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
5715 			rack->r_ctl.num_dsack = 0;
5716 		}
5717 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
5718 	}
5719 	if (rack->rto_from_rec == 1) {
5720 		rack->rto_from_rec = 0;
5721 		if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
5722 			tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
5723 	}
5724 	rack_exit_recovery(tp, rack, 1);
5725 }
5726 
5727 static void
5728 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
5729 {
5730 	struct tcp_rack *rack;
5731 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
5732 
5733 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5734 #ifdef STATS
5735 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
5736 #endif
5737 	if (IN_RECOVERY(tp->t_flags) == 0) {
5738 		in_rec_at_entry = 0;
5739 		ssthresh_enter = tp->snd_ssthresh;
5740 		cwnd_enter = tp->snd_cwnd;
5741 	} else
5742 		in_rec_at_entry = 1;
5743 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5744 	switch (type) {
5745 	case CC_NDUPACK:
5746 		tp->t_flags &= ~TF_WASFRECOVERY;
5747 		tp->t_flags &= ~TF_WASCRECOVERY;
5748 		if (!IN_FASTRECOVERY(tp->t_flags)) {
5749 			/* Check if this is the end of the initial Start-up i.e. initial slow-start */
5750 			if (rack->rc_initial_ss_comp == 0) {
5751 				/* Yep it is the end of the initial slowstart */
5752 				rack->rc_initial_ss_comp = 1;
5753 			}
5754 			rack->r_ctl.rc_prr_delivered = 0;
5755 			rack->r_ctl.rc_prr_out = 0;
5756 			rack->r_fast_output = 0;
5757 			if (rack->rack_no_prr == 0) {
5758 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5759 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
5760 			}
5761 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
5762 			tp->snd_recover = tp->snd_max;
5763 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5764 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5765 		}
5766 		break;
5767 	case CC_ECN:
5768 		if (!IN_CONGRECOVERY(tp->t_flags) ||
5769 		    /*
5770 		     * Allow ECN reaction on ACK to CWR, if
5771 		     * that data segment was also CE marked.
5772 		     */
5773 		    SEQ_GEQ(ack, tp->snd_recover)) {
5774 			EXIT_CONGRECOVERY(tp->t_flags);
5775 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5776 			rack->r_fast_output = 0;
5777 			tp->snd_recover = tp->snd_max + 1;
5778 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5779 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5780 		}
5781 		break;
5782 	case CC_RTO:
5783 		tp->t_dupacks = 0;
5784 		tp->t_bytes_acked = 0;
5785 		rack->r_fast_output = 0;
5786 		if (IN_RECOVERY(tp->t_flags))
5787 			rack_exit_recovery(tp, rack, 2);
5788 		orig_cwnd = tp->snd_cwnd;
5789 		rack_log_to_prr(rack, 16, orig_cwnd, line);
5790 		if (CC_ALGO(tp)->cong_signal == NULL) {
5791 			/* TSNH */
5792 			tp->snd_ssthresh = max(2,
5793 			    min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5794 			    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5795 			tp->snd_cwnd = ctf_fixed_maxseg(tp);
5796 		}
5797 		if (tp->t_flags2 & TF2_ECN_PERMIT)
5798 			tp->t_flags2 |= TF2_ECN_SND_CWR;
5799 		break;
5800 	case CC_RTO_ERR:
5801 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5802 		/* RTO was unnecessary, so reset everything. */
5803 		tp->snd_cwnd = tp->snd_cwnd_prev;
5804 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
5805 		tp->snd_recover = tp->snd_recover_prev;
5806 		if (tp->t_flags & TF_WASFRECOVERY) {
5807 			ENTER_FASTRECOVERY(tp->t_flags);
5808 			tp->t_flags &= ~TF_WASFRECOVERY;
5809 		}
5810 		if (tp->t_flags & TF_WASCRECOVERY) {
5811 			ENTER_CONGRECOVERY(tp->t_flags);
5812 			tp->t_flags &= ~TF_WASCRECOVERY;
5813 		}
5814 		tp->snd_nxt = tp->snd_max;
5815 		tp->t_badrxtwin = 0;
5816 		break;
5817 	}
5818 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
5819 	    (type != CC_RTO)){
5820 		tp->t_ccv.curack = ack;
5821 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
5822 	}
5823 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5824 		rack_log_to_prr(rack, 15, cwnd_enter, line);
5825 		rack->r_ctl.dsack_byte_cnt = 0;
5826 		rack->r_ctl.retran_during_recovery = 0;
5827 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5828 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5829 		rack->r_ent_rec_ns = 1;
5830 	}
5831 }
5832 
5833 static inline void
5834 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5835 {
5836 	uint32_t i_cwnd;
5837 
5838 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5839 
5840 	if (CC_ALGO(tp)->after_idle != NULL)
5841 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
5842 
5843 	if (tp->snd_cwnd == 1)
5844 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
5845 	else
5846 		i_cwnd = rc_init_window(rack);
5847 
5848 	/*
5849 	 * Being idle is no different than the initial window. If the cc
5850 	 * clamps it down below the initial window raise it to the initial
5851 	 * window.
5852 	 */
5853 	if (tp->snd_cwnd < i_cwnd) {
5854 		tp->snd_cwnd = i_cwnd;
5855 	}
5856 }
5857 
5858 /*
5859  * Indicate whether this ack should be delayed.  We can delay the ack if
5860  * following conditions are met:
5861  *	- There is no delayed ack timer in progress.
5862  *	- Our last ack wasn't a 0-sized window. We never want to delay
5863  *	  the ack that opens up a 0-sized window.
5864  *	- LRO wasn't used for this segment. We make sure by checking that the
5865  *	  segment size is not larger than the MSS.
5866  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
5867  *	  connection.
5868  */
5869 #define DELAY_ACK(tp, tlen)			 \
5870 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
5871 	((tp->t_flags & TF_DELACK) == 0) &&	 \
5872 	(tlen <= tp->t_maxseg) &&		 \
5873 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
5874 
5875 static struct rack_sendmap *
5876 rack_find_lowest_rsm(struct tcp_rack *rack)
5877 {
5878 	struct rack_sendmap *rsm;
5879 
5880 	/*
5881 	 * Walk the time-order transmitted list looking for an rsm that is
5882 	 * not acked. This will be the one that was sent the longest time
5883 	 * ago that is still outstanding.
5884 	 */
5885 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
5886 		if (rsm->r_flags & RACK_ACKED) {
5887 			continue;
5888 		}
5889 		goto finish;
5890 	}
5891 finish:
5892 	return (rsm);
5893 }
5894 
5895 static struct rack_sendmap *
5896 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5897 {
5898 	struct rack_sendmap *prsm;
5899 
5900 	/*
5901 	 * Walk the sequence order list backward until we hit and arrive at
5902 	 * the highest seq not acked. In theory when this is called it
5903 	 * should be the last segment (which it was not).
5904 	 */
5905 	prsm = rsm;
5906 
5907 	TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
5908 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5909 			continue;
5910 		}
5911 		return (prsm);
5912 	}
5913 	return (NULL);
5914 }
5915 
5916 static uint32_t
5917 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts, int line, int log_allowed)
5918 {
5919 	int32_t lro;
5920 	uint32_t thresh;
5921 
5922 	/*
5923 	 * lro is the flag we use to determine if we have seen reordering.
5924 	 * If it gets set we have seen reordering. The reorder logic either
5925 	 * works in one of two ways:
5926 	 *
5927 	 * If reorder-fade is configured, then we track the last time we saw
5928 	 * re-ordering occur. If we reach the point where enough time as
5929 	 * passed we no longer consider reordering has occuring.
5930 	 *
5931 	 * Or if reorder-face is 0, then once we see reordering we consider
5932 	 * the connection to alway be subject to reordering and just set lro
5933 	 * to 1.
5934 	 *
5935 	 * In the end if lro is non-zero we add the extra time for
5936 	 * reordering in.
5937 	 */
5938 	if (srtt == 0)
5939 		srtt = 1;
5940 	if (rack->r_ctl.rc_reorder_ts) {
5941 		if (rack->r_ctl.rc_reorder_fade) {
5942 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5943 				lro = cts - rack->r_ctl.rc_reorder_ts;
5944 				if (lro == 0) {
5945 					/*
5946 					 * No time as passed since the last
5947 					 * reorder, mark it as reordering.
5948 					 */
5949 					lro = 1;
5950 				}
5951 			} else {
5952 				/* Negative time? */
5953 				lro = 0;
5954 			}
5955 			if (lro > rack->r_ctl.rc_reorder_fade) {
5956 				/* Turn off reordering seen too */
5957 				rack->r_ctl.rc_reorder_ts = 0;
5958 				lro = 0;
5959 			}
5960 		} else {
5961 			/* Reodering does not fade */
5962 			lro = 1;
5963 		}
5964 	} else {
5965 		lro = 0;
5966 	}
5967 	if (rack->rc_rack_tmr_std_based == 0) {
5968 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
5969 	} else {
5970 		/* Standards based pkt-delay is 1/4 srtt */
5971 		thresh = srtt +  (srtt >> 2);
5972 	}
5973 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5974 		/* It must be set, if not you get 1/4 rtt */
5975 		if (rack->r_ctl.rc_reorder_shift)
5976 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5977 		else
5978 			thresh += (srtt >> 2);
5979 	}
5980 	if (rack->rc_rack_use_dsack &&
5981 	    lro &&
5982 	    (rack->r_ctl.num_dsack > 0)) {
5983 		/*
5984 		 * We only increase the reordering window if we
5985 		 * have seen reordering <and> we have a DSACK count.
5986 		 */
5987 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5988 		if (log_allowed)
5989 			rack_log_dsack_event(rack, 4, line, srtt, thresh);
5990 	}
5991 	/* SRTT * 2 is the ceiling */
5992 	if (thresh > (srtt * 2)) {
5993 		thresh = srtt * 2;
5994 	}
5995 	/* And we don't want it above the RTO max either */
5996 	if (thresh > rack_rto_max) {
5997 		thresh = rack_rto_max;
5998 	}
5999 	if (log_allowed)
6000 		rack_log_dsack_event(rack, 6, line,  srtt, thresh);
6001 	return (thresh);
6002 }
6003 
6004 static uint32_t
6005 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
6006 		     struct rack_sendmap *rsm, uint32_t srtt)
6007 {
6008 	struct rack_sendmap *prsm;
6009 	uint32_t thresh, len;
6010 	int segsiz;
6011 
6012 	if (srtt == 0)
6013 		srtt = 1;
6014 	if (rack->r_ctl.rc_tlp_threshold)
6015 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
6016 	else
6017 		thresh = (srtt * 2);
6018 
6019 	/* Get the previous sent packet, if any */
6020 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6021 	len = rsm->r_end - rsm->r_start;
6022 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
6023 		/* Exactly like the ID */
6024 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6025 			uint32_t alt_thresh;
6026 			/*
6027 			 * Compensate for delayed-ack with the d-ack time.
6028 			 */
6029 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6030 			if (alt_thresh > thresh)
6031 				thresh = alt_thresh;
6032 		}
6033 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6034 		/* 2.1 behavior */
6035 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6036 		if (prsm && (len <= segsiz)) {
6037 			/*
6038 			 * Two packets outstanding, thresh should be (2*srtt) +
6039 			 * possible inter-packet delay (if any).
6040 			 */
6041 			uint32_t inter_gap = 0;
6042 			int idx, nidx;
6043 
6044 			idx = rsm->r_rtr_cnt - 1;
6045 			nidx = prsm->r_rtr_cnt - 1;
6046 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6047 				/* Yes it was sent later (or at the same time) */
6048 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6049 			}
6050 			thresh += inter_gap;
6051 		} else if (len <= segsiz) {
6052 			/*
6053 			 * Possibly compensate for delayed-ack.
6054 			 */
6055 			uint32_t alt_thresh;
6056 
6057 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6058 			if (alt_thresh > thresh)
6059 				thresh = alt_thresh;
6060 		}
6061 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6062 		/* 2.2 behavior */
6063 		if (len <= segsiz) {
6064 			uint32_t alt_thresh;
6065 			/*
6066 			 * Compensate for delayed-ack with the d-ack time.
6067 			 */
6068 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6069 			if (alt_thresh > thresh)
6070 				thresh = alt_thresh;
6071 		}
6072 	}
6073 	/* Not above an RTO */
6074 	if (thresh > tp->t_rxtcur) {
6075 		thresh = tp->t_rxtcur;
6076 	}
6077 	/* Not above a RTO max */
6078 	if (thresh > rack_rto_max) {
6079 		thresh = rack_rto_max;
6080 	}
6081 	/* Apply user supplied min TLP */
6082 	if (thresh < rack_tlp_min) {
6083 		thresh = rack_tlp_min;
6084 	}
6085 	return (thresh);
6086 }
6087 
6088 static uint32_t
6089 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6090 {
6091 	/*
6092 	 * We want the rack_rtt which is the
6093 	 * last rtt we measured. However if that
6094 	 * does not exist we fallback to the srtt (which
6095 	 * we probably will never do) and then as a last
6096 	 * resort we use RACK_INITIAL_RTO if no srtt is
6097 	 * yet set.
6098 	 */
6099 	if (rack->rc_rack_rtt)
6100 		return (rack->rc_rack_rtt);
6101 	else if (tp->t_srtt == 0)
6102 		return (RACK_INITIAL_RTO);
6103 	return (tp->t_srtt);
6104 }
6105 
6106 static struct rack_sendmap *
6107 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6108 {
6109 	/*
6110 	 * Check to see that we don't need to fall into recovery. We will
6111 	 * need to do so if our oldest transmit is past the time we should
6112 	 * have had an ack.
6113 	 */
6114 	struct tcp_rack *rack;
6115 	struct rack_sendmap *rsm;
6116 	int32_t idx;
6117 	uint32_t srtt, thresh;
6118 
6119 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6120 	if (tqhash_empty(rack->r_ctl.tqh)) {
6121 		return (NULL);
6122 	}
6123 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6124 	if (rsm == NULL)
6125 		return (NULL);
6126 
6127 
6128 	if (rsm->r_flags & RACK_ACKED) {
6129 		rsm = rack_find_lowest_rsm(rack);
6130 		if (rsm == NULL)
6131 			return (NULL);
6132 	}
6133 	idx = rsm->r_rtr_cnt - 1;
6134 	srtt = rack_grab_rtt(tp, rack);
6135 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
6136 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6137 		return (NULL);
6138 	}
6139 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6140 		return (NULL);
6141 	}
6142 	/* Ok if we reach here we are over-due and this guy can be sent */
6143 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6144 	return (rsm);
6145 }
6146 
6147 static uint32_t
6148 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6149 {
6150 	int32_t t;
6151 	int32_t tt;
6152 	uint32_t ret_val;
6153 
6154 	t = (tp->t_srtt + (tp->t_rttvar << 2));
6155 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6156  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6157 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6158 	ret_val = (uint32_t)tt;
6159 	return (ret_val);
6160 }
6161 
6162 static uint32_t
6163 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6164 {
6165 	/*
6166 	 * Start the FR timer, we do this based on getting the first one in
6167 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6168 	 * events we need to stop the running timer (if its running) before
6169 	 * starting the new one.
6170 	 */
6171 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6172 	uint32_t srtt_cur;
6173 	int32_t idx;
6174 	int32_t is_tlp_timer = 0;
6175 	struct rack_sendmap *rsm;
6176 
6177 	if (rack->t_timers_stopped) {
6178 		/* All timers have been stopped none are to run */
6179 		return (0);
6180 	}
6181 	if (rack->rc_in_persist) {
6182 		/* We can't start any timer in persists */
6183 		return (rack_get_persists_timer_val(tp, rack));
6184 	}
6185 	rack->rc_on_min_to = 0;
6186 	if ((tp->t_state < TCPS_ESTABLISHED) ||
6187 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6188 		goto activate_rxt;
6189 	}
6190 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6191 	if ((rsm == NULL) || sup_rack) {
6192 		/* Nothing on the send map or no rack */
6193 activate_rxt:
6194 		time_since_sent = 0;
6195 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6196 		if (rsm) {
6197 			/*
6198 			 * Should we discount the RTX timer any?
6199 			 *
6200 			 * We want to discount it the smallest amount.
6201 			 * If a timer (Rack/TLP or RXT) has gone off more
6202 			 * recently thats the discount we want to use (now - timer time).
6203 			 * If the retransmit of the oldest packet was more recent then
6204 			 * we want to use that (now - oldest-packet-last_transmit_time).
6205 			 *
6206 			 */
6207 			idx = rsm->r_rtr_cnt - 1;
6208 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6209 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6210 			else
6211 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6212 			if (TSTMP_GT(cts, tstmp_touse))
6213 			    time_since_sent = cts - tstmp_touse;
6214 		}
6215 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6216 		    sbavail(&tptosocket(tp)->so_snd)) {
6217 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6218 			to = tp->t_rxtcur;
6219 			if (to > time_since_sent)
6220 				to -= time_since_sent;
6221 			else
6222 				to = rack->r_ctl.rc_min_to;
6223 			if (to == 0)
6224 				to = 1;
6225 			/* Special case for KEEPINIT */
6226 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6227 			    (TP_KEEPINIT(tp) != 0) &&
6228 			    rsm) {
6229 				/*
6230 				 * We have to put a ceiling on the rxt timer
6231 				 * of the keep-init timeout.
6232 				 */
6233 				uint32_t max_time, red;
6234 
6235 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6236 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6237 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6238 					if (red < max_time)
6239 						max_time -= red;
6240 					else
6241 						max_time = 1;
6242 				}
6243 				/* Reduce timeout to the keep value if needed */
6244 				if (max_time < to)
6245 					to = max_time;
6246 			}
6247 			return (to);
6248 		}
6249 		return (0);
6250 	}
6251 	if (rsm->r_flags & RACK_ACKED) {
6252 		rsm = rack_find_lowest_rsm(rack);
6253 		if (rsm == NULL) {
6254 			/* No lowest? */
6255 			goto activate_rxt;
6256 		}
6257 	}
6258 	/* Convert from ms to usecs */
6259 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
6260 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6261 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6262 		if ((tp->t_flags & TF_SENTFIN) &&
6263 		    ((tp->snd_max - tp->snd_una) == 1) &&
6264 		    (rsm->r_flags & RACK_HAS_FIN)) {
6265 			/*
6266 			 * We don't start a rack timer if all we have is a
6267 			 * FIN outstanding.
6268 			 */
6269 			goto activate_rxt;
6270 		}
6271 		if ((rack->use_rack_rr == 0) &&
6272 		    (IN_FASTRECOVERY(tp->t_flags)) &&
6273 		    (rack->rack_no_prr == 0) &&
6274 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
6275 			/*
6276 			 * We are not cheating, in recovery  and
6277 			 * not enough ack's to yet get our next
6278 			 * retransmission out.
6279 			 *
6280 			 * Note that classified attackers do not
6281 			 * get to use the rack-cheat.
6282 			 */
6283 			goto activate_tlp;
6284 		}
6285 		srtt = rack_grab_rtt(tp, rack);
6286 		thresh = rack_calc_thresh_rack(rack, srtt, cts, __LINE__, 1);
6287 		idx = rsm->r_rtr_cnt - 1;
6288 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6289 		if (SEQ_GEQ(exp, cts)) {
6290 			to = exp - cts;
6291 			if (to < rack->r_ctl.rc_min_to) {
6292 				to = rack->r_ctl.rc_min_to;
6293 				if (rack->r_rr_config == 3)
6294 					rack->rc_on_min_to = 1;
6295 			}
6296 		} else {
6297 			to = rack->r_ctl.rc_min_to;
6298 			if (rack->r_rr_config == 3)
6299 				rack->rc_on_min_to = 1;
6300 		}
6301 	} else {
6302 		/* Ok we need to do a TLP not RACK */
6303 activate_tlp:
6304 		if ((rack->rc_tlp_in_progress != 0) &&
6305 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6306 			/*
6307 			 * The previous send was a TLP and we have sent
6308 			 * N TLP's without sending new data.
6309 			 */
6310 			goto activate_rxt;
6311 		}
6312 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6313 		if (rsm == NULL) {
6314 			/* We found no rsm to TLP with. */
6315 			goto activate_rxt;
6316 		}
6317 		if (rsm->r_flags & RACK_HAS_FIN) {
6318 			/* If its a FIN we dont do TLP */
6319 			rsm = NULL;
6320 			goto activate_rxt;
6321 		}
6322 		idx = rsm->r_rtr_cnt - 1;
6323 		time_since_sent = 0;
6324 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6325 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6326 		else
6327 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6328 		if (TSTMP_GT(cts, tstmp_touse))
6329 		    time_since_sent = cts - tstmp_touse;
6330 		is_tlp_timer = 1;
6331 		if (tp->t_srtt) {
6332 			if ((rack->rc_srtt_measure_made == 0) &&
6333 			    (tp->t_srtt == 1)) {
6334 				/*
6335 				 * If another stack as run and set srtt to 1,
6336 				 * then the srtt was 0, so lets use the initial.
6337 				 */
6338 				srtt = RACK_INITIAL_RTO;
6339 			} else {
6340 				srtt_cur = tp->t_srtt;
6341 				srtt = srtt_cur;
6342 			}
6343 		} else
6344 			srtt = RACK_INITIAL_RTO;
6345 		/*
6346 		 * If the SRTT is not keeping up and the
6347 		 * rack RTT has spiked we want to use
6348 		 * the last RTT not the smoothed one.
6349 		 */
6350 		if (rack_tlp_use_greater &&
6351 		    tp->t_srtt &&
6352 		    (srtt < rack_grab_rtt(tp, rack))) {
6353 			srtt = rack_grab_rtt(tp, rack);
6354 		}
6355 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6356 		if (thresh > time_since_sent) {
6357 			to = thresh - time_since_sent;
6358 		} else {
6359 			to = rack->r_ctl.rc_min_to;
6360 			rack_log_alt_to_to_cancel(rack,
6361 						  thresh,		/* flex1 */
6362 						  time_since_sent,	/* flex2 */
6363 						  tstmp_touse,		/* flex3 */
6364 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6365 						  (uint32_t)rsm->r_tim_lastsent[idx],
6366 						  srtt,
6367 						  idx, 99);
6368 		}
6369 		if (to < rack_tlp_min) {
6370 			to = rack_tlp_min;
6371 		}
6372 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
6373 			/*
6374 			 * If the TLP time works out to larger than the max
6375 			 * RTO lets not do TLP.. just RTO.
6376 			 */
6377 			goto activate_rxt;
6378 		}
6379 	}
6380 	if (is_tlp_timer == 0) {
6381 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6382 	} else {
6383 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6384 	}
6385 	if (to == 0)
6386 		to = 1;
6387 	return (to);
6388 }
6389 
6390 static void
6391 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6392 {
6393 	if (rack->rc_in_persist == 0) {
6394 		if (tp->t_flags & TF_GPUTINPROG) {
6395 			/*
6396 			 * Stop the goodput now, the calling of the
6397 			 * measurement function clears the flag.
6398 			 */
6399 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6400 						    RACK_QUALITY_PERSIST);
6401 		}
6402 #ifdef NETFLIX_SHARED_CWND
6403 		if (rack->r_ctl.rc_scw) {
6404 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6405 			rack->rack_scwnd_is_idle = 1;
6406 		}
6407 #endif
6408 		rack->r_ctl.rc_went_idle_time = cts;
6409 		if (rack->r_ctl.rc_went_idle_time == 0)
6410 			rack->r_ctl.rc_went_idle_time = 1;
6411 		if (rack->lt_bw_up) {
6412 			/* Suspend our LT BW measurement */
6413 			uint64_t tmark;
6414 
6415 			rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6416 			rack->r_ctl.lt_seq = snd_una;
6417 			tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
6418 			if (tmark >= rack->r_ctl.lt_timemark) {
6419 				rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6420 			}
6421 			rack->r_ctl.lt_timemark = tmark;
6422 			rack->lt_bw_up = 0;
6423 			rack->r_persist_lt_bw_off = 1;
6424 		}
6425 		rack_timer_cancel(tp, rack, cts, __LINE__);
6426 		rack->r_ctl.persist_lost_ends = 0;
6427 		rack->probe_not_answered = 0;
6428 		rack->forced_ack = 0;
6429 		tp->t_rxtshift = 0;
6430 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6431 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6432 		rack->rc_in_persist = 1;
6433 	}
6434 }
6435 
6436 static void
6437 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6438 {
6439 	if (tcp_in_hpts(rack->rc_tp)) {
6440 		tcp_hpts_remove(rack->rc_tp);
6441 		rack->r_ctl.rc_hpts_flags = 0;
6442 	}
6443 #ifdef NETFLIX_SHARED_CWND
6444 	if (rack->r_ctl.rc_scw) {
6445 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6446 		rack->rack_scwnd_is_idle = 0;
6447 	}
6448 #endif
6449 	if (rack->rc_gp_dyn_mul &&
6450 	    (rack->use_fixed_rate == 0) &&
6451 	    (rack->rc_always_pace)) {
6452 		/*
6453 		 * Do we count this as if a probe-rtt just
6454 		 * finished?
6455 		 */
6456 		uint32_t time_idle, idle_min;
6457 
6458 		time_idle = cts - rack->r_ctl.rc_went_idle_time;
6459 		idle_min = rack_min_probertt_hold;
6460 		if (rack_probertt_gpsrtt_cnt_div) {
6461 			uint64_t extra;
6462 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
6463 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
6464 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
6465 			idle_min += (uint32_t)extra;
6466 		}
6467 		if (time_idle >= idle_min) {
6468 			/* Yes, we count it as a probe-rtt. */
6469 			uint32_t us_cts;
6470 
6471 			us_cts = tcp_get_usecs(NULL);
6472 			if (rack->in_probe_rtt == 0) {
6473 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6474 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
6475 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
6476 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
6477 			} else {
6478 				rack_exit_probertt(rack, us_cts);
6479 			}
6480 		}
6481 	}
6482 	if (rack->r_persist_lt_bw_off) {
6483 		/* Continue where we left off */
6484 		rack->r_ctl.lt_timemark = tcp_get_u64_usecs(NULL);
6485 		rack->lt_bw_up = 1;
6486 		rack->r_persist_lt_bw_off = 0;
6487 	}
6488 	rack->rc_in_persist = 0;
6489 	rack->r_ctl.rc_went_idle_time = 0;
6490 	tp->t_rxtshift = 0;
6491 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6492 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6493 	rack->r_ctl.rc_agg_delayed = 0;
6494 	rack->r_early = 0;
6495 	rack->r_late = 0;
6496 	rack->r_ctl.rc_agg_early = 0;
6497 }
6498 
6499 static void
6500 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
6501 		   struct hpts_diag *diag, struct timeval *tv)
6502 {
6503 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6504 		union tcp_log_stackspecific log;
6505 
6506 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6507 		log.u_bbr.flex1 = diag->p_nxt_slot;
6508 		log.u_bbr.flex2 = diag->p_cur_slot;
6509 		log.u_bbr.flex3 = diag->slot_req;
6510 		log.u_bbr.flex4 = diag->inp_hptsslot;
6511 		log.u_bbr.flex5 = diag->slot_remaining;
6512 		log.u_bbr.flex6 = diag->need_new_to;
6513 		log.u_bbr.flex7 = diag->p_hpts_active;
6514 		log.u_bbr.flex8 = diag->p_on_min_sleep;
6515 		/* Hijack other fields as needed */
6516 		log.u_bbr.epoch = diag->have_slept;
6517 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
6518 		log.u_bbr.pkts_out = diag->co_ret;
6519 		log.u_bbr.applimited = diag->hpts_sleep_time;
6520 		log.u_bbr.delivered = diag->p_prev_slot;
6521 		log.u_bbr.inflight = diag->p_runningslot;
6522 		log.u_bbr.bw_inuse = diag->wheel_slot;
6523 		log.u_bbr.rttProp = diag->wheel_cts;
6524 		log.u_bbr.timeStamp = cts;
6525 		log.u_bbr.delRate = diag->maxslots;
6526 		log.u_bbr.cur_del_rate = diag->p_curtick;
6527 		log.u_bbr.cur_del_rate <<= 32;
6528 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
6529 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6530 		    &rack->rc_inp->inp_socket->so_rcv,
6531 		    &rack->rc_inp->inp_socket->so_snd,
6532 		    BBR_LOG_HPTSDIAG, 0,
6533 		    0, &log, false, tv);
6534 	}
6535 
6536 }
6537 
6538 static void
6539 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
6540 {
6541 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6542 		union tcp_log_stackspecific log;
6543 		struct timeval tv;
6544 
6545 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6546 		log.u_bbr.flex1 = sb->sb_flags;
6547 		log.u_bbr.flex2 = len;
6548 		log.u_bbr.flex3 = sb->sb_state;
6549 		log.u_bbr.flex8 = type;
6550 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6551 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6552 		    &rack->rc_inp->inp_socket->so_rcv,
6553 		    &rack->rc_inp->inp_socket->so_snd,
6554 		    TCP_LOG_SB_WAKE, 0,
6555 		    len, &log, false, &tv);
6556 	}
6557 }
6558 
6559 static void
6560 rack_start_hpts_timer (struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
6561       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
6562 {
6563 	struct hpts_diag diag;
6564 	struct inpcb *inp = tptoinpcb(tp);
6565 	struct timeval tv;
6566 	uint32_t delayed_ack = 0;
6567 	uint32_t hpts_timeout;
6568 	uint32_t entry_slot = slot;
6569 	uint8_t stopped;
6570 	uint32_t left = 0;
6571 	uint32_t us_cts;
6572 
6573 	if ((tp->t_state == TCPS_CLOSED) ||
6574 	    (tp->t_state == TCPS_LISTEN)) {
6575 		return;
6576 	}
6577 	if (tcp_in_hpts(tp)) {
6578 		/* Already on the pacer */
6579 		return;
6580 	}
6581 	stopped = rack->rc_tmr_stopped;
6582 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
6583 		left = rack->r_ctl.rc_timer_exp - cts;
6584 	}
6585 	rack->r_ctl.rc_timer_exp = 0;
6586 	rack->r_ctl.rc_hpts_flags = 0;
6587 	us_cts = tcp_get_usecs(&tv);
6588 	/* Now early/late accounting */
6589 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
6590 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
6591 		/*
6592 		 * We have a early carry over set,
6593 		 * we can always add more time so we
6594 		 * can always make this compensation.
6595 		 *
6596 		 * Note if ack's are allowed to wake us do not
6597 		 * penalize the next timer for being awoke
6598 		 * by an ack aka the rc_agg_early (non-paced mode).
6599 		 */
6600 		slot += rack->r_ctl.rc_agg_early;
6601 		rack->r_early = 0;
6602 		rack->r_ctl.rc_agg_early = 0;
6603 	}
6604 	if ((rack->r_late) &&
6605 	    ((rack->r_use_hpts_min == 0) || (rack->dgp_on == 0))) {
6606 		/*
6607 		 * This is harder, we can
6608 		 * compensate some but it
6609 		 * really depends on what
6610 		 * the current pacing time is.
6611 		 */
6612 		if (rack->r_ctl.rc_agg_delayed >= slot) {
6613 			/*
6614 			 * We can't compensate for it all.
6615 			 * And we have to have some time
6616 			 * on the clock. We always have a min
6617 			 * 10 slots (10 x 10 i.e. 100 usecs).
6618 			 */
6619 			if (slot <= HPTS_TICKS_PER_SLOT) {
6620 				/* We gain delay */
6621 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
6622 				slot = HPTS_TICKS_PER_SLOT;
6623 			} else {
6624 				/* We take off some */
6625 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
6626 				slot = HPTS_TICKS_PER_SLOT;
6627 			}
6628 		} else {
6629 			slot -= rack->r_ctl.rc_agg_delayed;
6630 			rack->r_ctl.rc_agg_delayed = 0;
6631 			/* Make sure we have 100 useconds at minimum */
6632 			if (slot < HPTS_TICKS_PER_SLOT) {
6633 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
6634 				slot = HPTS_TICKS_PER_SLOT;
6635 			}
6636 			if (rack->r_ctl.rc_agg_delayed == 0)
6637 				rack->r_late = 0;
6638 		}
6639 	} else if (rack->r_late) {
6640 		/* r_use_hpts_min is on and so is DGP */
6641 		uint32_t max_red;
6642 
6643 		max_red = (slot * rack->r_ctl.max_reduction) / 100;
6644 		if (max_red >= rack->r_ctl.rc_agg_delayed) {
6645 			slot -= rack->r_ctl.rc_agg_delayed;
6646 			rack->r_ctl.rc_agg_delayed = 0;
6647 		} else {
6648 			slot -= max_red;
6649 			rack->r_ctl.rc_agg_delayed -= max_red;
6650 		}
6651 	}
6652 	if ((rack->r_use_hpts_min == 1) &&
6653 	    (slot > 0) &&
6654 	    (rack->dgp_on == 1)) {
6655 		/*
6656 		 * We are enforcing a min pacing timer
6657 		 * based on our hpts min timeout.
6658 		 */
6659 		uint32_t min;
6660 
6661 		min = get_hpts_min_sleep_time();
6662 		if (min > slot) {
6663 			slot = min;
6664 		}
6665 	}
6666 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
6667 	if (tp->t_flags & TF_DELACK) {
6668 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
6669 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
6670 	}
6671 	if (delayed_ack && ((hpts_timeout == 0) ||
6672 			    (delayed_ack < hpts_timeout)))
6673 		hpts_timeout = delayed_ack;
6674 	else
6675 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6676 	/*
6677 	 * If no timers are going to run and we will fall off the hptsi
6678 	 * wheel, we resort to a keep-alive timer if its configured.
6679 	 */
6680 	if ((hpts_timeout == 0) &&
6681 	    (slot == 0)) {
6682 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6683 		    (tp->t_state <= TCPS_CLOSING)) {
6684 			/*
6685 			 * Ok we have no timer (persists, rack, tlp, rxt  or
6686 			 * del-ack), we don't have segments being paced. So
6687 			 * all that is left is the keepalive timer.
6688 			 */
6689 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6690 				/* Get the established keep-alive time */
6691 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
6692 			} else {
6693 				/*
6694 				 * Get the initial setup keep-alive time,
6695 				 * note that this is probably not going to
6696 				 * happen, since rack will be running a rxt timer
6697 				 * if a SYN of some sort is outstanding. It is
6698 				 * actually handled in rack_timeout_rxt().
6699 				 */
6700 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
6701 			}
6702 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
6703 			if (rack->in_probe_rtt) {
6704 				/*
6705 				 * We want to instead not wake up a long time from
6706 				 * now but to wake up about the time we would
6707 				 * exit probe-rtt and initiate a keep-alive ack.
6708 				 * This will get us out of probe-rtt and update
6709 				 * our min-rtt.
6710 				 */
6711 				hpts_timeout = rack_min_probertt_hold;
6712 			}
6713 		}
6714 	}
6715 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
6716 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
6717 		/*
6718 		 * RACK, TLP, persists and RXT timers all are restartable
6719 		 * based on actions input .. i.e we received a packet (ack
6720 		 * or sack) and that changes things (rw, or snd_una etc).
6721 		 * Thus we can restart them with a new value. For
6722 		 * keep-alive, delayed_ack we keep track of what was left
6723 		 * and restart the timer with a smaller value.
6724 		 */
6725 		if (left < hpts_timeout)
6726 			hpts_timeout = left;
6727 	}
6728 	if (hpts_timeout) {
6729 		/*
6730 		 * Hack alert for now we can't time-out over 2,147,483
6731 		 * seconds (a bit more than 596 hours), which is probably ok
6732 		 * :).
6733 		 */
6734 		if (hpts_timeout > 0x7ffffffe)
6735 			hpts_timeout = 0x7ffffffe;
6736 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
6737 	}
6738 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
6739 	if ((rack->gp_ready == 0) &&
6740 	    (rack->use_fixed_rate == 0) &&
6741 	    (hpts_timeout < slot) &&
6742 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
6743 		/*
6744 		 * We have no good estimate yet for the
6745 		 * old clunky burst mitigation or the
6746 		 * real pacing. And the tlp or rxt is smaller
6747 		 * than the pacing calculation. Lets not
6748 		 * pace that long since we know the calculation
6749 		 * so far is not accurate.
6750 		 */
6751 		slot = hpts_timeout;
6752 	}
6753 	/**
6754 	 * Turn off all the flags for queuing by default. The
6755 	 * flags have important meanings to what happens when
6756 	 * LRO interacts with the transport. Most likely (by default now)
6757 	 * mbuf_queueing and ack compression are on. So the transport
6758 	 * has a couple of flags that control what happens (if those
6759 	 * are not on then these flags won't have any effect since it
6760 	 * won't go through the queuing LRO path).
6761 	 *
6762 	 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
6763 	 *                        pacing output, so don't disturb. But
6764 	 *                        it also means LRO can wake me if there
6765 	 *                        is a SACK arrival.
6766 	 *
6767 	 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
6768 	 *                       with the above flag (QUEUE_READY) and
6769 	 *                       when present it says don't even wake me
6770 	 *                       if a SACK arrives.
6771 	 *
6772 	 * The idea behind these flags is that if we are pacing we
6773 	 * set the MBUF_QUEUE_READY and only get woken up if
6774 	 * a SACK arrives (which could change things) or if
6775 	 * our pacing timer expires. If, however, we have a rack
6776 	 * timer running, then we don't even want a sack to wake
6777 	 * us since the rack timer has to expire before we can send.
6778 	 *
6779 	 * Other cases should usually have none of the flags set
6780 	 * so LRO can call into us.
6781 	 */
6782 	tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
6783 	if (slot) {
6784 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
6785 		rack->r_ctl.rc_last_output_to = us_cts + slot;
6786 		/*
6787 		 * A pacing timer (slot) is being set, in
6788 		 * such a case we cannot send (we are blocked by
6789 		 * the timer). So lets tell LRO that it should not
6790 		 * wake us unless there is a SACK. Note this only
6791 		 * will be effective if mbuf queueing is on or
6792 		 * compressed acks are being processed.
6793 		 */
6794 		tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
6795 		/*
6796 		 * But wait if we have a Rack timer running
6797 		 * even a SACK should not disturb us (with
6798 		 * the exception of r_rr_config 3).
6799 		 */
6800 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
6801 		    (IN_RECOVERY(tp->t_flags))) {
6802 			if (rack->r_rr_config != 3)
6803 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6804 			else if (rack->rc_pace_dnd) {
6805 				/*
6806 				 * When DND is on, we only let a sack
6807 				 * interrupt us if we are not in recovery.
6808 				 *
6809 				 * If DND is off, then we never hit here
6810 				 * and let all sacks wake us up.
6811 				 *
6812 				 */
6813 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6814 			}
6815 		}
6816 		if (rack->rc_ack_can_sendout_data) {
6817 			/*
6818 			 * Ahh but wait, this is that special case
6819 			 * where the pacing timer can be disturbed
6820 			 * backout the changes (used for non-paced
6821 			 * burst limiting).
6822 			 */
6823 			tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
6824 			    TF2_MBUF_QUEUE_READY);
6825 		}
6826 		if ((rack->use_rack_rr) &&
6827 		    (rack->r_rr_config < 2) &&
6828 		    ((hpts_timeout) && (hpts_timeout < slot))) {
6829 			/*
6830 			 * Arrange for the hpts to kick back in after the
6831 			 * t-o if the t-o does not cause a send.
6832 			 */
6833 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6834 						   __LINE__, &diag);
6835 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6836 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6837 		} else {
6838 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
6839 						   __LINE__, &diag);
6840 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6841 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
6842 		}
6843 	} else if (hpts_timeout) {
6844 		/*
6845 		 * With respect to t_flags2(?) here, lets let any new acks wake
6846 		 * us up here. Since we are not pacing (no pacing timer), output
6847 		 * can happen so we should let it. If its a Rack timer, then any inbound
6848 		 * packet probably won't change the sending (we will be blocked)
6849 		 * but it may change the prr stats so letting it in (the set defaults
6850 		 * at the start of this block) are good enough.
6851 		 */
6852 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6853 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6854 					   __LINE__, &diag);
6855 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6856 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6857 	} else {
6858 		/* No timer starting */
6859 #ifdef INVARIANTS
6860 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
6861 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
6862 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
6863 		}
6864 #endif
6865 	}
6866 	rack->rc_tmr_stopped = 0;
6867 	if (slot)
6868 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6869 }
6870 
6871 static void
6872 rack_mark_lost(struct tcpcb *tp,
6873     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
6874 {
6875 	struct rack_sendmap *nrsm;
6876 	uint32_t thresh,  exp;
6877 
6878 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
6879 	nrsm = rsm;
6880 	TAILQ_FOREACH_FROM(nrsm, &rack->r_ctl.rc_tmap, r_tnext) {
6881 		if ((nrsm->r_flags & RACK_SACK_PASSED) == 0) {
6882 			/* Got up to all that were marked sack-passed */
6883 			break;
6884 		}
6885 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
6886 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
6887 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
6888 				/* We now consider it lost */
6889 				nrsm->r_flags |= RACK_WAS_LOST;
6890 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
6891 			} else {
6892 				/* Past here it won't be lost so stop */
6893 				break;
6894 			}
6895 		}
6896 	}
6897 }
6898 
6899 /*
6900  * RACK Timer, here we simply do logging and house keeping.
6901  * the normal rack_output() function will call the
6902  * appropriate thing to check if we need to do a RACK retransmit.
6903  * We return 1, saying don't proceed with rack_output only
6904  * when all timers have been stopped (destroyed PCB?).
6905  */
6906 static int
6907 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6908 {
6909 	/*
6910 	 * This timer simply provides an internal trigger to send out data.
6911 	 * The check_recovery_mode call will see if there are needed
6912 	 * retransmissions, if so we will enter fast-recovery. The output
6913 	 * call may or may not do the same thing depending on sysctl
6914 	 * settings.
6915 	 */
6916 	struct rack_sendmap *rsm;
6917 
6918 	counter_u64_add(rack_to_tot, 1);
6919 	if (rack->r_state && (rack->r_state != tp->t_state))
6920 		rack_set_state(tp, rack);
6921 	rack->rc_on_min_to = 0;
6922 	rsm = rack_check_recovery_mode(tp, cts);
6923 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
6924 	if (rsm) {
6925 		/* We need to stroke any lost that are now declared as lost */
6926 		rack_mark_lost(tp, rack, rsm, cts);
6927 		rack->r_ctl.rc_resend = rsm;
6928 		rack->r_timer_override = 1;
6929 		if (rack->use_rack_rr) {
6930 			/*
6931 			 * Don't accumulate extra pacing delay
6932 			 * we are allowing the rack timer to
6933 			 * over-ride pacing i.e. rrr takes precedence
6934 			 * if the pacing interval is longer than the rrr
6935 			 * time (in other words we get the min pacing
6936 			 * time versus rrr pacing time).
6937 			 */
6938 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6939 		}
6940 	}
6941 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6942 	if (rsm == NULL) {
6943 		/* restart a timer and return 1 */
6944 		rack_start_hpts_timer(rack, tp, cts,
6945 				      0, 0, 0);
6946 		return (1);
6947 	}
6948 	return (0);
6949 }
6950 
6951 
6952 
6953 static void
6954 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6955 {
6956 
6957 	if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
6958 		/*
6959 		 * The trailing space changed, mbufs can grow
6960 		 * at the tail but they can't shrink from
6961 		 * it, KASSERT that. Adjust the orig_m_len to
6962 		 * compensate for this change.
6963 		 */
6964 		KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
6965 			("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
6966 			 rsm->m,
6967 			 rsm,
6968 			 (intmax_t)M_TRAILINGROOM(rsm->m),
6969 			 rsm->orig_t_space,
6970 			 rsm->orig_m_len,
6971 			 rsm->m->m_len));
6972 		rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
6973 		rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
6974 	}
6975 	if (rsm->m->m_len < rsm->orig_m_len) {
6976 		/*
6977 		 * Mbuf shrank, trimmed off the top by an ack, our
6978 		 * offset changes.
6979 		 */
6980 		KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
6981 			("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
6982 			 rsm->m, rsm->m->m_len,
6983 			 rsm, rsm->orig_m_len,
6984 			 rsm->soff));
6985 		if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
6986 			rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6987 		else
6988 			rsm->soff = 0;
6989 		rsm->orig_m_len = rsm->m->m_len;
6990 #ifdef INVARIANTS
6991 	} else if (rsm->m->m_len > rsm->orig_m_len) {
6992 		panic("rsm:%p m:%p m_len grew outside of t_space compensation",
6993 		      rsm, rsm->m);
6994 #endif
6995 	}
6996 }
6997 
6998 static void
6999 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
7000 {
7001 	struct mbuf *m;
7002 	uint32_t soff;
7003 
7004 	if (src_rsm->m &&
7005 	    ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
7006 	     (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
7007 		/* Fix up the orig_m_len and possibly the mbuf offset */
7008 		rack_adjust_orig_mlen(src_rsm);
7009 	}
7010 	m = src_rsm->m;
7011 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
7012 	while (soff >= m->m_len) {
7013 		/* Move out past this mbuf */
7014 		soff -= m->m_len;
7015 		m = m->m_next;
7016 		KASSERT((m != NULL),
7017 			("rsm:%p nrsm:%p hit at soff:%u null m",
7018 			 src_rsm, rsm, soff));
7019 		if (m == NULL) {
7020 			/* This should *not* happen which is why there is a kassert */
7021 			src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7022 					       (src_rsm->r_start - rack->rc_tp->snd_una),
7023 					       &src_rsm->soff);
7024 			src_rsm->orig_m_len = src_rsm->m->m_len;
7025 			src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7026 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7027 					   (rsm->r_start - rack->rc_tp->snd_una),
7028 					   &rsm->soff);
7029 			rsm->orig_m_len = rsm->m->m_len;
7030 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7031 			return;
7032 		}
7033 	}
7034 	rsm->m = m;
7035 	rsm->soff = soff;
7036 	rsm->orig_m_len = m->m_len;
7037 	rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7038 }
7039 
7040 static __inline void
7041 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7042 	       struct rack_sendmap *rsm, uint32_t start)
7043 {
7044 	int idx;
7045 
7046 	nrsm->r_start = start;
7047 	nrsm->r_end = rsm->r_end;
7048 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7049 	nrsm->r_act_rxt_cnt = rsm->r_act_rxt_cnt;
7050 	nrsm->r_flags = rsm->r_flags;
7051 	nrsm->r_dupack = rsm->r_dupack;
7052 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7053 	nrsm->r_rtr_bytes = 0;
7054 	nrsm->r_fas = rsm->r_fas;
7055 	nrsm->r_bas = rsm->r_bas;
7056 	tqhash_update_end(rack->r_ctl.tqh, rsm, nrsm->r_start);
7057 	nrsm->r_just_ret = rsm->r_just_ret;
7058 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7059 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7060 	}
7061 	/* Now if we have SYN flag we keep it on the left edge */
7062 	if (nrsm->r_flags & RACK_HAS_SYN)
7063 		nrsm->r_flags &= ~RACK_HAS_SYN;
7064 	/* Now if we have a FIN flag we keep it on the right edge */
7065 	if (rsm->r_flags & RACK_HAS_FIN)
7066 		rsm->r_flags &= ~RACK_HAS_FIN;
7067 	/* Push bit must go to the right edge as well */
7068 	if (rsm->r_flags & RACK_HAD_PUSH)
7069 		rsm->r_flags &= ~RACK_HAD_PUSH;
7070 	/* Clone over the state of the hw_tls flag */
7071 	nrsm->r_hw_tls = rsm->r_hw_tls;
7072 	/*
7073 	 * Now we need to find nrsm's new location in the mbuf chain
7074 	 * we basically calculate a new offset, which is soff +
7075 	 * how much is left in original rsm. Then we walk out the mbuf
7076 	 * chain to find the righ position, it may be the same mbuf
7077 	 * or maybe not.
7078 	 */
7079 	KASSERT(((rsm->m != NULL) ||
7080 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7081 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7082 	if (rsm->m)
7083 		rack_setup_offset_for_rsm(rack, rsm, nrsm);
7084 }
7085 
7086 static struct rack_sendmap *
7087 rack_merge_rsm(struct tcp_rack *rack,
7088 	       struct rack_sendmap *l_rsm,
7089 	       struct rack_sendmap *r_rsm)
7090 {
7091 	/*
7092 	 * We are merging two ack'd RSM's,
7093 	 * the l_rsm is on the left (lower seq
7094 	 * values) and the r_rsm is on the right
7095 	 * (higher seq value). The simplest way
7096 	 * to merge these is to move the right
7097 	 * one into the left. I don't think there
7098 	 * is any reason we need to try to find
7099 	 * the oldest (or last oldest retransmitted).
7100 	 */
7101 	rack_log_map_chg(rack->rc_tp, rack, NULL,
7102 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7103 	tqhash_update_end(rack->r_ctl.tqh, l_rsm, r_rsm->r_end);
7104 	if (l_rsm->r_dupack < r_rsm->r_dupack)
7105 		l_rsm->r_dupack = r_rsm->r_dupack;
7106 	if (r_rsm->r_rtr_bytes)
7107 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7108 	if (r_rsm->r_in_tmap) {
7109 		/* This really should not happen */
7110 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7111 		r_rsm->r_in_tmap = 0;
7112 	}
7113 
7114 	/* Now the flags */
7115 	if (r_rsm->r_flags & RACK_HAS_FIN)
7116 		l_rsm->r_flags |= RACK_HAS_FIN;
7117 	if (r_rsm->r_flags & RACK_TLP)
7118 		l_rsm->r_flags |= RACK_TLP;
7119 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7120 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7121 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
7122 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7123 		/*
7124 		 * If both are app-limited then let the
7125 		 * free lower the count. If right is app
7126 		 * limited and left is not, transfer.
7127 		 */
7128 		l_rsm->r_flags |= RACK_APP_LIMITED;
7129 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
7130 		if (r_rsm == rack->r_ctl.rc_first_appl)
7131 			rack->r_ctl.rc_first_appl = l_rsm;
7132 	}
7133 	tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7134 	/*
7135 	 * We keep the largest value, which is the newest
7136 	 * send. We do this in case a segment that is
7137 	 * joined together and not part of a GP estimate
7138 	 * later gets expanded into the GP estimate.
7139 	 *
7140 	 * We prohibit the merging of unlike kinds i.e.
7141 	 * all pieces that are in the GP estimate can be
7142 	 * merged and all pieces that are not in a GP estimate
7143 	 * can be merged, but not disimilar pieces. Combine
7144 	 * this with taking the highest here and we should
7145 	 * be ok unless of course the client reneges. Then
7146 	 * all bets are off.
7147 	 */
7148 	if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7149 	   r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7150 		l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7151 	}
7152 	/*
7153 	 * When merging two RSM's we also need to consider the ack time and keep
7154 	 * newest. If the ack gets merged into a measurement then that is the
7155 	 * one we will want to be using.
7156 	 */
7157 	if(l_rsm->r_ack_arrival	 < r_rsm->r_ack_arrival)
7158 		l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7159 
7160 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7161 		/* Transfer the split limit to the map we free */
7162 		r_rsm->r_limit_type = l_rsm->r_limit_type;
7163 		l_rsm->r_limit_type = 0;
7164 	}
7165 	rack_free(rack, r_rsm);
7166 	l_rsm->r_flags |= RACK_MERGED;
7167 	return (l_rsm);
7168 }
7169 
7170 /*
7171  * TLP Timer, here we simply setup what segment we want to
7172  * have the TLP expire on, the normal rack_output() will then
7173  * send it out.
7174  *
7175  * We return 1, saying don't proceed with rack_output only
7176  * when all timers have been stopped (destroyed PCB?).
7177  */
7178 static int
7179 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7180 {
7181 	/*
7182 	 * Tail Loss Probe.
7183 	 */
7184 	struct rack_sendmap *rsm = NULL;
7185 	int insret __diagused;
7186 	struct socket *so = tptosocket(tp);
7187 	uint32_t amm;
7188 	uint32_t out, avail;
7189 	int collapsed_win = 0;
7190 
7191 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7192 		/* Its not time yet */
7193 		return (0);
7194 	}
7195 	if (ctf_progress_timeout_check(tp, true)) {
7196 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7197 		return (-ETIMEDOUT);	/* tcp_drop() */
7198 	}
7199 	/*
7200 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7201 	 * need to figure out how to force a full MSS segment out.
7202 	 */
7203 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7204 	rack->r_ctl.retran_during_recovery = 0;
7205 	rack->r_might_revert = 0;
7206 	rack->r_ctl.dsack_byte_cnt = 0;
7207 	counter_u64_add(rack_tlp_tot, 1);
7208 	if (rack->r_state && (rack->r_state != tp->t_state))
7209 		rack_set_state(tp, rack);
7210 	avail = sbavail(&so->so_snd);
7211 	out = tp->snd_max - tp->snd_una;
7212 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7213 		/* special case, we need a retransmission */
7214 		collapsed_win = 1;
7215 		goto need_retran;
7216 	}
7217 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7218 		rack->r_ctl.dsack_persist--;
7219 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7220 			rack->r_ctl.num_dsack = 0;
7221 		}
7222 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7223 	}
7224 	if ((tp->t_flags & TF_GPUTINPROG) &&
7225 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7226 		/*
7227 		 * If this is the second in a row
7228 		 * TLP and we are doing a measurement
7229 		 * its time to abandon the measurement.
7230 		 * Something is likely broken on
7231 		 * the clients network and measuring a
7232 		 * broken network does us no good.
7233 		 */
7234 		tp->t_flags &= ~TF_GPUTINPROG;
7235 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7236 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7237 					   tp->gput_seq,
7238 					   0, 0, 18, __LINE__, NULL, 0);
7239 	}
7240 	/*
7241 	 * Check our send oldest always settings, and if
7242 	 * there is an oldest to send jump to the need_retran.
7243 	 */
7244 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7245 		goto need_retran;
7246 
7247 	if (avail > out) {
7248 		/* New data is available */
7249 		amm = avail - out;
7250 		if (amm > ctf_fixed_maxseg(tp)) {
7251 			amm = ctf_fixed_maxseg(tp);
7252 			if ((amm + out) > tp->snd_wnd) {
7253 				/* We are rwnd limited */
7254 				goto need_retran;
7255 			}
7256 		} else if (amm < ctf_fixed_maxseg(tp)) {
7257 			/* not enough to fill a MTU */
7258 			goto need_retran;
7259 		}
7260 		if (IN_FASTRECOVERY(tp->t_flags)) {
7261 			/* Unlikely */
7262 			if (rack->rack_no_prr == 0) {
7263 				if (out + amm <= tp->snd_wnd) {
7264 					rack->r_ctl.rc_prr_sndcnt = amm;
7265 					rack->r_ctl.rc_tlp_new_data = amm;
7266 					rack_log_to_prr(rack, 4, 0, __LINE__);
7267 				}
7268 			} else
7269 				goto need_retran;
7270 		} else {
7271 			/* Set the send-new override */
7272 			if (out + amm <= tp->snd_wnd)
7273 				rack->r_ctl.rc_tlp_new_data = amm;
7274 			else
7275 				goto need_retran;
7276 		}
7277 		rack->r_ctl.rc_tlpsend = NULL;
7278 		counter_u64_add(rack_tlp_newdata, 1);
7279 		goto send;
7280 	}
7281 need_retran:
7282 	/*
7283 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
7284 	 * optionally the first un-acked segment.
7285 	 */
7286 	if (collapsed_win == 0) {
7287 		if (rack_always_send_oldest)
7288 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7289 		else {
7290 			rsm = tqhash_max(rack->r_ctl.tqh);
7291 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7292 				rsm = rack_find_high_nonack(rack, rsm);
7293 			}
7294 		}
7295 		if (rsm == NULL) {
7296 #ifdef TCP_BLACKBOX
7297 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7298 #endif
7299 			goto out;
7300 		}
7301 	} else {
7302 		/*
7303 		 * We had a collapsed window, lets find
7304 		 * the point before the collapse.
7305 		 */
7306 		if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7307 			rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7308 		else {
7309 			rsm = tqhash_min(rack->r_ctl.tqh);
7310 		}
7311 		if (rsm == NULL) {
7312 			/* Huh */
7313 			goto out;
7314 		}
7315 	}
7316 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7317 		/*
7318 		 * We need to split this the last segment in two.
7319 		 */
7320 		struct rack_sendmap *nrsm;
7321 
7322 		nrsm = rack_alloc_full_limit(rack);
7323 		if (nrsm == NULL) {
7324 			/*
7325 			 * No memory to split, we will just exit and punt
7326 			 * off to the RXT timer.
7327 			 */
7328 			goto out;
7329 		}
7330 		rack_clone_rsm(rack, nrsm, rsm,
7331 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
7332 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7333 #ifndef INVARIANTS
7334 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7335 #else
7336 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7337 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
7338 			      nrsm, insret, rack, rsm);
7339 		}
7340 #endif
7341 		if (rsm->r_in_tmap) {
7342 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7343 			nrsm->r_in_tmap = 1;
7344 		}
7345 		rsm = nrsm;
7346 	}
7347 	rack->r_ctl.rc_tlpsend = rsm;
7348 send:
7349 	/* Make sure output path knows we are doing a TLP */
7350 	*doing_tlp = 1;
7351 	rack->r_timer_override = 1;
7352 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7353 	return (0);
7354 out:
7355 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7356 	return (0);
7357 }
7358 
7359 /*
7360  * Delayed ack Timer, here we simply need to setup the
7361  * ACK_NOW flag and remove the DELACK flag. From there
7362  * the output routine will send the ack out.
7363  *
7364  * We only return 1, saying don't proceed, if all timers
7365  * are stopped (destroyed PCB?).
7366  */
7367 static int
7368 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7369 {
7370 
7371 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7372 	tp->t_flags &= ~TF_DELACK;
7373 	tp->t_flags |= TF_ACKNOW;
7374 	KMOD_TCPSTAT_INC(tcps_delack);
7375 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7376 	return (0);
7377 }
7378 
7379 static inline int
7380 rack_send_ack_challange(struct tcp_rack *rack)
7381 {
7382 	struct tcptemp *t_template;
7383 
7384 	t_template = tcpip_maketemplate(rack->rc_inp);
7385 	if (t_template) {
7386 		if (rack->forced_ack == 0) {
7387 			rack->forced_ack = 1;
7388 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7389 		} else {
7390 			rack->probe_not_answered = 1;
7391 		}
7392 		tcp_respond(rack->rc_tp, t_template->tt_ipgen,
7393 			    &t_template->tt_t, (struct mbuf *)NULL,
7394 			    rack->rc_tp->rcv_nxt, rack->rc_tp->snd_una - 1, 0);
7395 		free(t_template, M_TEMP);
7396 		/* This does send an ack so kill any D-ack timer */
7397 		if (rack->rc_tp->t_flags & TF_DELACK)
7398 			rack->rc_tp->t_flags &= ~TF_DELACK;
7399 		return(1);
7400 	} else
7401 		return (0);
7402 
7403 }
7404 
7405 /*
7406  * Persists timer, here we simply send the
7407  * same thing as a keepalive will.
7408  * the one byte send.
7409  *
7410  * We only return 1, saying don't proceed, if all timers
7411  * are stopped (destroyed PCB?).
7412  */
7413 static int
7414 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7415 {
7416 	int32_t retval = 1;
7417 
7418 	if (rack->rc_in_persist == 0)
7419 		return (0);
7420 	if (ctf_progress_timeout_check(tp, false)) {
7421 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7422 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7423 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7424 		return (-ETIMEDOUT);	/* tcp_drop() */
7425 	}
7426 	/*
7427 	 * Persistence timer into zero window. Force a byte to be output, if
7428 	 * possible.
7429 	 */
7430 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
7431 	/*
7432 	 * Hack: if the peer is dead/unreachable, we do not time out if the
7433 	 * window is closed.  After a full backoff, drop the connection if
7434 	 * the idle time (no responses to probes) reaches the maximum
7435 	 * backoff that we would use if retransmitting.
7436 	 */
7437 	if (tp->t_rxtshift >= V_tcp_retries &&
7438 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7439 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7440 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7441 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7442 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7443 		retval = -ETIMEDOUT;	/* tcp_drop() */
7444 		goto out;
7445 	}
7446 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7447 	    tp->snd_una == tp->snd_max)
7448 		rack_exit_persist(tp, rack, cts);
7449 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7450 	/*
7451 	 * If the user has closed the socket then drop a persisting
7452 	 * connection after a much reduced timeout.
7453 	 */
7454 	if (tp->t_state > TCPS_CLOSE_WAIT &&
7455 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7456 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7457 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7458 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7459 		retval = -ETIMEDOUT;	/* tcp_drop() */
7460 		goto out;
7461 	}
7462 	if (rack_send_ack_challange(rack)) {
7463 		/* only set it if we were answered */
7464 		if (rack->probe_not_answered) {
7465 			counter_u64_add(rack_persists_loss, 1);
7466 			rack->r_ctl.persist_lost_ends++;
7467 		}
7468 		counter_u64_add(rack_persists_sends, 1);
7469 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7470 	}
7471 	if (tp->t_rxtshift < V_tcp_retries)
7472 		tp->t_rxtshift++;
7473 out:
7474 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7475 	rack_start_hpts_timer(rack, tp, cts,
7476 			      0, 0, 0);
7477 	return (retval);
7478 }
7479 
7480 /*
7481  * If a keepalive goes off, we had no other timers
7482  * happening. We always return 1 here since this
7483  * routine either drops the connection or sends
7484  * out a segment with respond.
7485  */
7486 static int
7487 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7488 {
7489 	struct inpcb *inp = tptoinpcb(tp);
7490 
7491 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7492 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7493 	/*
7494 	 * Keep-alive timer went off; send something or drop connection if
7495 	 * idle for too long.
7496 	 */
7497 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
7498 	if (tp->t_state < TCPS_ESTABLISHED)
7499 		goto dropit;
7500 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7501 	    tp->t_state <= TCPS_CLOSING) {
7502 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7503 			goto dropit;
7504 		/*
7505 		 * Send a packet designed to force a response if the peer is
7506 		 * up and reachable: either an ACK if the connection is
7507 		 * still alive, or an RST if the peer has closed the
7508 		 * connection due to timeout or reboot. Using sequence
7509 		 * number tp->snd_una-1 causes the transmitted zero-length
7510 		 * segment to lie outside the receive window; by the
7511 		 * protocol spec, this requires the correspondent TCP to
7512 		 * respond.
7513 		 */
7514 		KMOD_TCPSTAT_INC(tcps_keepprobe);
7515 		rack_send_ack_challange(rack);
7516 	}
7517 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7518 	return (1);
7519 dropit:
7520 	KMOD_TCPSTAT_INC(tcps_keepdrops);
7521 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7522 	return (-ETIMEDOUT);	/* tcp_drop() */
7523 }
7524 
7525 /*
7526  * Retransmit helper function, clear up all the ack
7527  * flags and take care of important book keeping.
7528  */
7529 static void
7530 rack_remxt_tmr(struct tcpcb *tp)
7531 {
7532 	/*
7533 	 * The retransmit timer went off, all sack'd blocks must be
7534 	 * un-acked.
7535 	 */
7536 	struct rack_sendmap *rsm, *trsm = NULL;
7537 	struct tcp_rack *rack;
7538 
7539 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7540 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7541 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7542 	rack->r_timer_override = 1;
7543 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7544 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7545 	rack->r_late = 0;
7546 	rack->r_early = 0;
7547 	rack->r_ctl.rc_agg_delayed = 0;
7548 	rack->r_ctl.rc_agg_early = 0;
7549 	if (rack->r_state && (rack->r_state != tp->t_state))
7550 		rack_set_state(tp, rack);
7551 	if (tp->t_rxtshift <= rack_rxt_scoreboard_clear_thresh) {
7552 		/*
7553 		 * We do not clear the scoreboard until we have had
7554 		 * more than rack_rxt_scoreboard_clear_thresh time-outs.
7555 		 */
7556 		rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7557 		if (rack->r_ctl.rc_resend != NULL)
7558 			rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7559 
7560 		return;
7561 	}
7562 	/*
7563 	 * Ideally we would like to be able to
7564 	 * mark SACK-PASS on anything not acked here.
7565 	 *
7566 	 * However, if we do that we would burst out
7567 	 * all that data 1ms apart. This would be unwise,
7568 	 * so for now we will just let the normal rxt timer
7569 	 * and tlp timer take care of it.
7570 	 *
7571 	 * Also we really need to stick them back in sequence
7572 	 * order. This way we send in the proper order and any
7573 	 * sacks that come floating in will "re-ack" the data.
7574 	 * To do this we zap the tmap with an INIT and then
7575 	 * walk through and place every rsm in the tail queue
7576 	 * hash table back in its seq ordered place.
7577 	 */
7578 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
7579 
7580 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
7581 		rsm->r_dupack = 0;
7582 		if (rack_verbose_logging)
7583 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7584 		/* We must re-add it back to the tlist */
7585 		if (trsm == NULL) {
7586 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7587 		} else {
7588 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7589 		}
7590 		rsm->r_in_tmap = 1;
7591 		trsm = rsm;
7592 		if (rsm->r_flags & RACK_ACKED)
7593 			rsm->r_flags |= RACK_WAS_ACKED;
7594 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED | RACK_WAS_LOST);
7595 		rsm->r_flags |= RACK_MUST_RXT;
7596 	}
7597 	/* zero the lost since it's all gone */
7598 	rack->r_ctl.rc_considered_lost = 0;
7599 	/* Clear the count (we just un-acked them) */
7600 	rack->r_ctl.rc_sacked = 0;
7601 	rack->r_ctl.rc_sacklast = NULL;
7602 	/* Clear the tlp rtx mark */
7603 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7604 	if (rack->r_ctl.rc_resend != NULL)
7605 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7606 	rack->r_ctl.rc_prr_sndcnt = 0;
7607 	rack_log_to_prr(rack, 6, 0, __LINE__);
7608 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7609 	if (rack->r_ctl.rc_resend != NULL)
7610 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7611 	if (((tp->t_flags & TF_SACK_PERMIT) == 0) &&
7612 	    ((tp->t_flags & TF_SENTFIN) == 0)) {
7613 		/*
7614 		 * For non-sack customers new data
7615 		 * needs to go out as retransmits until
7616 		 * we retransmit up to snd_max.
7617 		 */
7618 		rack->r_must_retran = 1;
7619 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7620 							    rack->r_ctl.rc_sacked);
7621 	}
7622 }
7623 
7624 static void
7625 rack_convert_rtts(struct tcpcb *tp)
7626 {
7627 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7628 	tp->t_rxtcur = RACK_REXMTVAL(tp);
7629 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7630 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7631 	}
7632 	if (tp->t_rxtcur > rack_rto_max) {
7633 		tp->t_rxtcur = rack_rto_max;
7634 	}
7635 }
7636 
7637 static void
7638 rack_cc_conn_init(struct tcpcb *tp)
7639 {
7640 	struct tcp_rack *rack;
7641 	uint32_t srtt;
7642 
7643 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7644 	srtt = tp->t_srtt;
7645 	cc_conn_init(tp);
7646 	/*
7647 	 * Now convert to rack's internal format,
7648 	 * if required.
7649 	 */
7650 	if ((srtt == 0) && (tp->t_srtt != 0))
7651 		rack_convert_rtts(tp);
7652 	/*
7653 	 * We want a chance to stay in slowstart as
7654 	 * we create a connection. TCP spec says that
7655 	 * initially ssthresh is infinite. For our
7656 	 * purposes that is the snd_wnd.
7657 	 */
7658 	if (tp->snd_ssthresh < tp->snd_wnd) {
7659 		tp->snd_ssthresh = tp->snd_wnd;
7660 	}
7661 	/*
7662 	 * We also want to assure a IW worth of
7663 	 * data can get inflight.
7664 	 */
7665 	if (rc_init_window(rack) < tp->snd_cwnd)
7666 		tp->snd_cwnd = rc_init_window(rack);
7667 }
7668 
7669 /*
7670  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7671  * we will setup to retransmit the lowest seq number outstanding.
7672  */
7673 static int
7674 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7675 {
7676 	struct inpcb *inp = tptoinpcb(tp);
7677 	int32_t rexmt;
7678 	int32_t retval = 0;
7679 	bool isipv6;
7680 
7681 	if ((tp->t_flags & TF_GPUTINPROG) &&
7682 	    (tp->t_rxtshift)) {
7683 		/*
7684 		 * We have had a second timeout
7685 		 * measurements on successive rxt's are not profitable.
7686 		 * It is unlikely to be of any use (the network is
7687 		 * broken or the client went away).
7688 		 */
7689 		tp->t_flags &= ~TF_GPUTINPROG;
7690 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7691 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7692 					   tp->gput_seq,
7693 					   0, 0, 18, __LINE__, NULL, 0);
7694 	}
7695 	if (ctf_progress_timeout_check(tp, false)) {
7696 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7697 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7698 		return (-ETIMEDOUT);	/* tcp_drop() */
7699 	}
7700 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7701 	rack->r_ctl.retran_during_recovery = 0;
7702 	rack->rc_ack_required = 1;
7703 	rack->r_ctl.dsack_byte_cnt = 0;
7704 	if (IN_RECOVERY(tp->t_flags) &&
7705 	    (rack->rto_from_rec == 0)) {
7706 		/*
7707 		 * Mark that we had a rto while in recovery
7708 		 * and save the ssthresh so if we go back
7709 		 * into recovery we will have a chance
7710 		 * to slowstart back to the level.
7711 		 */
7712 		rack->rto_from_rec = 1;
7713 		rack->r_ctl.rto_ssthresh = tp->snd_ssthresh;
7714 	}
7715 	if (IN_FASTRECOVERY(tp->t_flags))
7716 		tp->t_flags |= TF_WASFRECOVERY;
7717 	else
7718 		tp->t_flags &= ~TF_WASFRECOVERY;
7719 	if (IN_CONGRECOVERY(tp->t_flags))
7720 		tp->t_flags |= TF_WASCRECOVERY;
7721 	else
7722 		tp->t_flags &= ~TF_WASCRECOVERY;
7723 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7724 	    (tp->snd_una == tp->snd_max)) {
7725 		/* Nothing outstanding .. nothing to do */
7726 		return (0);
7727 	}
7728 	if (rack->r_ctl.dsack_persist) {
7729 		rack->r_ctl.dsack_persist--;
7730 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7731 			rack->r_ctl.num_dsack = 0;
7732 		}
7733 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7734 	}
7735 	/*
7736 	 * Rack can only run one timer  at a time, so we cannot
7737 	 * run a KEEPINIT (gating SYN sending) and a retransmit
7738 	 * timer for the SYN. So if we are in a front state and
7739 	 * have a KEEPINIT timer we need to check the first transmit
7740 	 * against now to see if we have exceeded the KEEPINIT time
7741 	 * (if one is set).
7742 	 */
7743 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7744 	    (TP_KEEPINIT(tp) != 0)) {
7745 		struct rack_sendmap *rsm;
7746 
7747 		rsm = tqhash_min(rack->r_ctl.tqh);
7748 		if (rsm) {
7749 			/* Ok we have something outstanding to test keepinit with */
7750 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7751 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7752 				/* We have exceeded the KEEPINIT time */
7753 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7754 				goto drop_it;
7755 			}
7756 		}
7757 	}
7758 	/*
7759 	 * Retransmission timer went off.  Message has not been acked within
7760 	 * retransmit interval.  Back off to a longer retransmit interval
7761 	 * and retransmit one segment.
7762 	 */
7763 	if ((rack->r_ctl.rc_resend == NULL) ||
7764 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7765 		/*
7766 		 * If the rwnd collapsed on
7767 		 * the one we are retransmitting
7768 		 * it does not count against the
7769 		 * rxt count.
7770 		 */
7771 		tp->t_rxtshift++;
7772 	}
7773 	rack_remxt_tmr(tp);
7774 	if (tp->t_rxtshift > V_tcp_retries) {
7775 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7776 drop_it:
7777 		tp->t_rxtshift = V_tcp_retries;
7778 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7779 		/* XXXGL: previously t_softerror was casted to uint16_t */
7780 		MPASS(tp->t_softerror >= 0);
7781 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7782 		goto out;	/* tcp_drop() */
7783 	}
7784 	if (tp->t_state == TCPS_SYN_SENT) {
7785 		/*
7786 		 * If the SYN was retransmitted, indicate CWND to be limited
7787 		 * to 1 segment in cc_conn_init().
7788 		 */
7789 		tp->snd_cwnd = 1;
7790 	} else if (tp->t_rxtshift == 1) {
7791 		/*
7792 		 * first retransmit; record ssthresh and cwnd so they can be
7793 		 * recovered if this turns out to be a "bad" retransmit. A
7794 		 * retransmit is considered "bad" if an ACK for this segment
7795 		 * is received within RTT/2 interval; the assumption here is
7796 		 * that the ACK was already in flight.  See "On Estimating
7797 		 * End-to-End Network Path Properties" by Allman and Paxson
7798 		 * for more details.
7799 		 */
7800 		tp->snd_cwnd_prev = tp->snd_cwnd;
7801 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
7802 		tp->snd_recover_prev = tp->snd_recover;
7803 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7804 		tp->t_flags |= TF_PREVVALID;
7805 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7806 		tp->t_flags &= ~TF_PREVVALID;
7807 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7808 	if ((tp->t_state == TCPS_SYN_SENT) ||
7809 	    (tp->t_state == TCPS_SYN_RECEIVED))
7810 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7811 	else
7812 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7813 
7814 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7815 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7816 	/*
7817 	 * We enter the path for PLMTUD if connection is established or, if
7818 	 * connection is FIN_WAIT_1 status, reason for the last is that if
7819 	 * amount of data we send is very small, we could send it in couple
7820 	 * of packets and process straight to FIN. In that case we won't
7821 	 * catch ESTABLISHED state.
7822 	 */
7823 #ifdef INET6
7824 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7825 #else
7826 	isipv6 = false;
7827 #endif
7828 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
7829 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7830 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7831 	    ((tp->t_state == TCPS_ESTABLISHED) ||
7832 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
7833 		/*
7834 		 * Idea here is that at each stage of mtu probe (usually,
7835 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
7836 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7837 		 * should take care of that.
7838 		 */
7839 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7840 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7841 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7842 		    tp->t_rxtshift % 2 == 0)) {
7843 			/*
7844 			 * Enter Path MTU Black-hole Detection mechanism: -
7845 			 * Disable Path MTU Discovery (IP "DF" bit). -
7846 			 * Reduce MTU to lower value than what we negotiated
7847 			 * with peer.
7848 			 */
7849 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7850 				/* Record that we may have found a black hole. */
7851 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7852 				/* Keep track of previous MSS. */
7853 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7854 			}
7855 
7856 			/*
7857 			 * Reduce the MSS to blackhole value or to the
7858 			 * default in an attempt to retransmit.
7859 			 */
7860 #ifdef INET6
7861 			if (isipv6 &&
7862 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7863 				/* Use the sysctl tuneable blackhole MSS. */
7864 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7865 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7866 			} else if (isipv6) {
7867 				/* Use the default MSS. */
7868 				tp->t_maxseg = V_tcp_v6mssdflt;
7869 				/*
7870 				 * Disable Path MTU Discovery when we switch
7871 				 * to minmss.
7872 				 */
7873 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7874 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7875 			}
7876 #endif
7877 #if defined(INET6) && defined(INET)
7878 			else
7879 #endif
7880 #ifdef INET
7881 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7882 				/* Use the sysctl tuneable blackhole MSS. */
7883 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
7884 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7885 			} else {
7886 				/* Use the default MSS. */
7887 				tp->t_maxseg = V_tcp_mssdflt;
7888 				/*
7889 				 * Disable Path MTU Discovery when we switch
7890 				 * to minmss.
7891 				 */
7892 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7893 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7894 			}
7895 #endif
7896 		} else {
7897 			/*
7898 			 * If further retransmissions are still unsuccessful
7899 			 * with a lowered MTU, maybe this isn't a blackhole
7900 			 * and we restore the previous MSS and blackhole
7901 			 * detection flags. The limit '6' is determined by
7902 			 * giving each probe stage (1448, 1188, 524) 2
7903 			 * chances to recover.
7904 			 */
7905 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
7906 			    (tp->t_rxtshift >= 6)) {
7907 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
7908 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
7909 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
7910 				if (tp->t_maxseg < V_tcp_mssdflt) {
7911 					/*
7912 					 * The MSS is so small we should not
7913 					 * process incoming SACK's since we are
7914 					 * subject to attack in such a case.
7915 					 */
7916 					tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
7917 				} else {
7918 					tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
7919 				}
7920 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
7921 			}
7922 		}
7923 	}
7924 	/*
7925 	 * Disable RFC1323 and SACK if we haven't got any response to
7926 	 * our third SYN to work-around some broken terminal servers
7927 	 * (most of which have hopefully been retired) that have bad VJ
7928 	 * header compression code which trashes TCP segments containing
7929 	 * unknown-to-them TCP options.
7930 	 */
7931 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
7932 	    (tp->t_rxtshift == 3))
7933 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
7934 	/*
7935 	 * If we backed off this far, our srtt estimate is probably bogus.
7936 	 * Clobber it so we'll take the next rtt measurement as our srtt;
7937 	 * move the current srtt into rttvar to keep the current retransmit
7938 	 * times until then.
7939 	 */
7940 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
7941 #ifdef INET6
7942 		if ((inp->inp_vflag & INP_IPV6) != 0)
7943 			in6_losing(inp);
7944 		else
7945 #endif
7946 			in_losing(inp);
7947 		tp->t_rttvar += tp->t_srtt;
7948 		tp->t_srtt = 0;
7949 	}
7950 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
7951 	tp->snd_recover = tp->snd_max;
7952 	tp->t_flags |= TF_ACKNOW;
7953 	tp->t_rtttime = 0;
7954 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
7955 out:
7956 	return (retval);
7957 }
7958 
7959 static int
7960 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
7961 {
7962 	int32_t ret = 0;
7963 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
7964 
7965 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7966 	    (tp->t_flags & TF_GPUTINPROG)) {
7967 		/*
7968 		 * We have a goodput in progress
7969 		 * and we have entered a late state.
7970 		 * Do we have enough data in the sb
7971 		 * to handle the GPUT request?
7972 		 */
7973 		uint32_t bytes;
7974 
7975 		bytes = tp->gput_ack - tp->gput_seq;
7976 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
7977 			bytes += tp->gput_seq - tp->snd_una;
7978 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
7979 			/*
7980 			 * There are not enough bytes in the socket
7981 			 * buffer that have been sent to cover this
7982 			 * measurement. Cancel it.
7983 			 */
7984 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7985 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
7986 						   tp->gput_seq,
7987 						   0, 0, 18, __LINE__, NULL, 0);
7988 			tp->t_flags &= ~TF_GPUTINPROG;
7989 		}
7990 	}
7991 	if (timers == 0) {
7992 		return (0);
7993 	}
7994 	if (tp->t_state == TCPS_LISTEN) {
7995 		/* no timers on listen sockets */
7996 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7997 			return (0);
7998 		return (1);
7999 	}
8000 	if ((timers & PACE_TMR_RACK) &&
8001 	    rack->rc_on_min_to) {
8002 		/*
8003 		 * For the rack timer when we
8004 		 * are on a min-timeout (which means rrr_conf = 3)
8005 		 * we don't want to check the timer. It may
8006 		 * be going off for a pace and thats ok we
8007 		 * want to send the retransmit (if its ready).
8008 		 *
8009 		 * If its on a normal rack timer (non-min) then
8010 		 * we will check if its expired.
8011 		 */
8012 		goto skip_time_check;
8013 	}
8014 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
8015 		uint32_t left;
8016 
8017 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
8018 			ret = -1;
8019 			rack_log_to_processing(rack, cts, ret, 0);
8020 			return (0);
8021 		}
8022 		if (hpts_calling == 0) {
8023 			/*
8024 			 * A user send or queued mbuf (sack) has called us? We
8025 			 * return 0 and let the pacing guards
8026 			 * deal with it if they should or
8027 			 * should not cause a send.
8028 			 */
8029 			ret = -2;
8030 			rack_log_to_processing(rack, cts, ret, 0);
8031 			return (0);
8032 		}
8033 		/*
8034 		 * Ok our timer went off early and we are not paced false
8035 		 * alarm, go back to sleep. We make sure we don't have
8036 		 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8037 		 * flag in place.
8038 		 */
8039 		rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8040 		ret = -3;
8041 		left = rack->r_ctl.rc_timer_exp - cts;
8042 		tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8043 		rack_log_to_processing(rack, cts, ret, left);
8044 		return (1);
8045 	}
8046 skip_time_check:
8047 	rack->rc_tmr_stopped = 0;
8048 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8049 	if (timers & PACE_TMR_DELACK) {
8050 		ret = rack_timeout_delack(tp, rack, cts);
8051 	} else if (timers & PACE_TMR_RACK) {
8052 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8053 		rack->r_fast_output = 0;
8054 		ret = rack_timeout_rack(tp, rack, cts);
8055 	} else if (timers & PACE_TMR_TLP) {
8056 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8057 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8058 	} else if (timers & PACE_TMR_RXT) {
8059 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8060 		rack->r_fast_output = 0;
8061 		ret = rack_timeout_rxt(tp, rack, cts);
8062 	} else if (timers & PACE_TMR_PERSIT) {
8063 		ret = rack_timeout_persist(tp, rack, cts);
8064 	} else if (timers & PACE_TMR_KEEP) {
8065 		ret = rack_timeout_keepalive(tp, rack, cts);
8066 	}
8067 	rack_log_to_processing(rack, cts, ret, timers);
8068 	return (ret);
8069 }
8070 
8071 static void
8072 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8073 {
8074 	struct timeval tv;
8075 	uint32_t us_cts, flags_on_entry;
8076 	uint8_t hpts_removed = 0;
8077 
8078 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
8079 	us_cts = tcp_get_usecs(&tv);
8080 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8081 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8082 	     ((tp->snd_max - tp->snd_una) == 0))) {
8083 		tcp_hpts_remove(rack->rc_tp);
8084 		hpts_removed = 1;
8085 		/* If we were not delayed cancel out the flag. */
8086 		if ((tp->snd_max - tp->snd_una) == 0)
8087 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8088 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8089 	}
8090 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8091 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8092 		if (tcp_in_hpts(rack->rc_tp) &&
8093 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8094 			/*
8095 			 * Canceling timer's when we have no output being
8096 			 * paced. We also must remove ourselves from the
8097 			 * hpts.
8098 			 */
8099 			tcp_hpts_remove(rack->rc_tp);
8100 			hpts_removed = 1;
8101 		}
8102 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8103 	}
8104 	if (hpts_removed == 0)
8105 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8106 }
8107 
8108 static int
8109 rack_stopall(struct tcpcb *tp)
8110 {
8111 	struct tcp_rack *rack;
8112 
8113 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8114 	rack->t_timers_stopped = 1;
8115 
8116 	tcp_hpts_remove(tp);
8117 
8118 	return (0);
8119 }
8120 
8121 static void
8122 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8123 {
8124 	/*
8125 	 * Assure no timers are running.
8126 	 */
8127 	if (tcp_timer_active(tp, TT_PERSIST)) {
8128 		/* We enter in persists, set the flag appropriately */
8129 		rack->rc_in_persist = 1;
8130 	}
8131 	if (tcp_in_hpts(rack->rc_tp)) {
8132 		tcp_hpts_remove(rack->rc_tp);
8133 	}
8134 }
8135 
8136 static void
8137 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8138     struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz)
8139 {
8140 	int32_t idx;
8141 
8142 	rsm->r_rtr_cnt++;
8143 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8144 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8145 		rsm->r_flags |= RACK_OVERMAX;
8146 	}
8147 	rsm->r_act_rxt_cnt++;
8148 	/* Peg the count/index */
8149 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8150 	rsm->r_dupack = 0;
8151 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8152 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8153 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8154 	}
8155 	if (rsm->r_flags & RACK_WAS_LOST) {
8156 		/*
8157 		 * We retransmitted it putting it back in flight
8158 		 * remove the lost desgination and reduce the
8159 		 * bytes considered lost.
8160 		 */
8161 		rsm->r_flags  &= ~RACK_WAS_LOST;
8162 		KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
8163 			("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
8164 		if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
8165 			rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
8166 		else
8167 			rack->r_ctl.rc_considered_lost = 0;
8168 	}
8169 	idx = rsm->r_rtr_cnt - 1;
8170 	rsm->r_tim_lastsent[idx] = ts;
8171 	/*
8172 	 * Here we don't add in the len of send, since its already
8173 	 * in snduna <->snd_max.
8174 	 */
8175 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
8176 				     rack->r_ctl.rc_sacked);
8177 	if (rsm->r_flags & RACK_ACKED) {
8178 		/* Problably MTU discovery messing with us */
8179 		rsm->r_flags &= ~RACK_ACKED;
8180 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8181 	}
8182 	if (rsm->r_in_tmap) {
8183 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8184 		rsm->r_in_tmap = 0;
8185 	}
8186 	/* Lets make sure it really is in or not the GP window */
8187 	rack_mark_in_gp_win(tp, rsm);
8188 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8189 	rsm->r_in_tmap = 1;
8190 	rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8191 	/* Take off the must retransmit flag, if its on */
8192 	if (rsm->r_flags & RACK_MUST_RXT) {
8193 		if (rack->r_must_retran)
8194 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8195 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8196 			/*
8197 			 * We have retransmitted all we need. Clear
8198 			 * any must retransmit flags.
8199 			 */
8200 			rack->r_must_retran = 0;
8201 			rack->r_ctl.rc_out_at_rto = 0;
8202 		}
8203 		rsm->r_flags &= ~RACK_MUST_RXT;
8204 	}
8205 	/* Remove any collapsed flag */
8206 	rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8207 	if (rsm->r_flags & RACK_SACK_PASSED) {
8208 		/* We have retransmitted due to the SACK pass */
8209 		rsm->r_flags &= ~RACK_SACK_PASSED;
8210 		rsm->r_flags |= RACK_WAS_SACKPASS;
8211 	}
8212 }
8213 
8214 static uint32_t
8215 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8216     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint32_t add_flag, int segsiz)
8217 {
8218 	/*
8219 	 * We (re-)transmitted starting at rsm->r_start for some length
8220 	 * (possibly less than r_end.
8221 	 */
8222 	struct rack_sendmap *nrsm;
8223 	int insret __diagused;
8224 	uint32_t c_end;
8225 	int32_t len;
8226 
8227 	len = *lenp;
8228 	c_end = rsm->r_start + len;
8229 	if (SEQ_GEQ(c_end, rsm->r_end)) {
8230 		/*
8231 		 * We retransmitted the whole piece or more than the whole
8232 		 * slopping into the next rsm.
8233 		 */
8234 		rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8235 		if (c_end == rsm->r_end) {
8236 			*lenp = 0;
8237 			return (0);
8238 		} else {
8239 			int32_t act_len;
8240 
8241 			/* Hangs over the end return whats left */
8242 			act_len = rsm->r_end - rsm->r_start;
8243 			*lenp = (len - act_len);
8244 			return (rsm->r_end);
8245 		}
8246 		/* We don't get out of this block. */
8247 	}
8248 	/*
8249 	 * Here we retransmitted less than the whole thing which means we
8250 	 * have to split this into what was transmitted and what was not.
8251 	 */
8252 	nrsm = rack_alloc_full_limit(rack);
8253 	if (nrsm == NULL) {
8254 		/*
8255 		 * We can't get memory, so lets not proceed.
8256 		 */
8257 		*lenp = 0;
8258 		return (0);
8259 	}
8260 	/*
8261 	 * So here we are going to take the original rsm and make it what we
8262 	 * retransmitted. nrsm will be the tail portion we did not
8263 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8264 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8265 	 * 1, 6 and the new piece will be 6, 11.
8266 	 */
8267 	rack_clone_rsm(rack, nrsm, rsm, c_end);
8268 	nrsm->r_dupack = 0;
8269 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8270 #ifndef INVARIANTS
8271 	(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8272 #else
8273 	if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8274 		panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8275 		      nrsm, insret, rack, rsm);
8276 	}
8277 #endif
8278 	if (rsm->r_in_tmap) {
8279 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8280 		nrsm->r_in_tmap = 1;
8281 	}
8282 	rsm->r_flags &= (~RACK_HAS_FIN);
8283 	rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8284 	/* Log a split of rsm into rsm and nrsm */
8285 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8286 	*lenp = 0;
8287 	return (0);
8288 }
8289 
8290 static void
8291 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8292 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8293 		struct rack_sendmap *hintrsm, uint32_t add_flag, struct mbuf *s_mb,
8294 		uint32_t s_moff, int hw_tls, int segsiz)
8295 {
8296 	struct tcp_rack *rack;
8297 	struct rack_sendmap *rsm, *nrsm;
8298 	int insret __diagused;
8299 
8300 	register uint32_t snd_max, snd_una;
8301 
8302 	/*
8303 	 * Add to the RACK log of packets in flight or retransmitted. If
8304 	 * there is a TS option we will use the TS echoed, if not we will
8305 	 * grab a TS.
8306 	 *
8307 	 * Retransmissions will increment the count and move the ts to its
8308 	 * proper place. Note that if options do not include TS's then we
8309 	 * won't be able to effectively use the ACK for an RTT on a retran.
8310 	 *
8311 	 * Notes about r_start and r_end. Lets consider a send starting at
8312 	 * sequence 1 for 10 bytes. In such an example the r_start would be
8313 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8314 	 * This means that r_end is actually the first sequence for the next
8315 	 * slot (11).
8316 	 *
8317 	 */
8318 	/*
8319 	 * If err is set what do we do XXXrrs? should we not add the thing?
8320 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
8321 	 * i.e. proceed with add ** do this for now.
8322 	 */
8323 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8324 	if (err)
8325 		/*
8326 		 * We don't log errors -- we could but snd_max does not
8327 		 * advance in this case either.
8328 		 */
8329 		return;
8330 
8331 	if (th_flags & TH_RST) {
8332 		/*
8333 		 * We don't log resets and we return immediately from
8334 		 * sending
8335 		 */
8336 		return;
8337 	}
8338 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8339 	snd_una = tp->snd_una;
8340 	snd_max = tp->snd_max;
8341 	if (th_flags & (TH_SYN | TH_FIN)) {
8342 		/*
8343 		 * The call to rack_log_output is made before bumping
8344 		 * snd_max. This means we can record one extra byte on a SYN
8345 		 * or FIN if seq_out is adding more on and a FIN is present
8346 		 * (and we are not resending).
8347 		 */
8348 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8349 			len++;
8350 		if (th_flags & TH_FIN)
8351 			len++;
8352 	}
8353 	if (SEQ_LEQ((seq_out + len), snd_una)) {
8354 		/* Are sending an old segment to induce an ack (keep-alive)? */
8355 		return;
8356 	}
8357 	if (SEQ_LT(seq_out, snd_una)) {
8358 		/* huh? should we panic? */
8359 		uint32_t end;
8360 
8361 		end = seq_out + len;
8362 		seq_out = snd_una;
8363 		if (SEQ_GEQ(end, seq_out))
8364 			len = end - seq_out;
8365 		else
8366 			len = 0;
8367 	}
8368 	if (len == 0) {
8369 		/* We don't log zero window probes */
8370 		return;
8371 	}
8372 	if (IN_FASTRECOVERY(tp->t_flags)) {
8373 		rack->r_ctl.rc_prr_out += len;
8374 	}
8375 	/* First question is it a retransmission or new? */
8376 	if (seq_out == snd_max) {
8377 		/* Its new */
8378 		rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
8379 again:
8380 		rsm = rack_alloc(rack);
8381 		if (rsm == NULL) {
8382 			/*
8383 			 * Hmm out of memory and the tcb got destroyed while
8384 			 * we tried to wait.
8385 			 */
8386 			return;
8387 		}
8388 		if (th_flags & TH_FIN) {
8389 			rsm->r_flags = RACK_HAS_FIN|add_flag;
8390 		} else {
8391 			rsm->r_flags = add_flag;
8392 		}
8393 		if (hw_tls)
8394 			rsm->r_hw_tls = 1;
8395 		rsm->r_tim_lastsent[0] = cts;
8396 		rsm->r_rtr_cnt = 1;
8397  		rsm->r_act_rxt_cnt = 0;
8398 		rsm->r_rtr_bytes = 0;
8399 		if (th_flags & TH_SYN) {
8400 			/* The data space is one beyond snd_una */
8401 			rsm->r_flags |= RACK_HAS_SYN;
8402 		}
8403 		rsm->r_start = seq_out;
8404 		rsm->r_end = rsm->r_start + len;
8405 		rack_mark_in_gp_win(tp, rsm);
8406 		rsm->r_dupack = 0;
8407 		/*
8408 		 * save off the mbuf location that
8409 		 * sndmbuf_noadv returned (which is
8410 		 * where we started copying from)..
8411 		 */
8412 		rsm->m = s_mb;
8413 		rsm->soff = s_moff;
8414 		/*
8415 		 * Here we do add in the len of send, since its not yet
8416 		 * reflected in in snduna <->snd_max
8417 		 */
8418 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8419 					      rack->r_ctl.rc_sacked) +
8420 			      (rsm->r_end - rsm->r_start));
8421 		if ((rack->rc_initial_ss_comp == 0) &&
8422 		    (rack->r_ctl.ss_hi_fs < rsm->r_fas)) {
8423 			   rack->r_ctl.ss_hi_fs = rsm->r_fas;
8424 		}
8425 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8426 		if (rsm->m) {
8427 			if (rsm->m->m_len <= rsm->soff) {
8428 				/*
8429 				 * XXXrrs Question, will this happen?
8430 				 *
8431 				 * If sbsndptr is set at the correct place
8432 				 * then s_moff should always be somewhere
8433 				 * within rsm->m. But if the sbsndptr was
8434 				 * off then that won't be true. If it occurs
8435 				 * we need to walkout to the correct location.
8436 				 */
8437 				struct mbuf *lm;
8438 
8439 				lm = rsm->m;
8440 				while (lm->m_len <= rsm->soff) {
8441 					rsm->soff -= lm->m_len;
8442 					lm = lm->m_next;
8443 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8444 							     __func__, rack, s_moff, s_mb, rsm->soff));
8445 				}
8446 				rsm->m = lm;
8447 			}
8448 			rsm->orig_m_len = rsm->m->m_len;
8449 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8450 		} else {
8451 			rsm->orig_m_len = 0;
8452 			rsm->orig_t_space = 0;
8453 		}
8454 		rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8455 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8456 		/* Log a new rsm */
8457 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8458 #ifndef INVARIANTS
8459 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
8460 #else
8461 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8462 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8463 			      nrsm, insret, rack, rsm);
8464 		}
8465 #endif
8466 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8467 		rsm->r_in_tmap = 1;
8468 		if (rsm->r_flags & RACK_IS_PCM) {
8469 			rack->r_ctl.pcm_i.send_time = cts;
8470 			rack->r_ctl.pcm_i.eseq = rsm->r_end;
8471 			/* First time through we set the start too */
8472 			if (rack->pcm_in_progress == 0)
8473 				rack->r_ctl.pcm_i.sseq = rsm->r_start;
8474 		}
8475 		/*
8476 		 * Special case detection, is there just a single
8477 		 * packet outstanding when we are not in recovery?
8478 		 *
8479 		 * If this is true mark it so.
8480 		 */
8481 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8482 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8483 			struct rack_sendmap *prsm;
8484 
8485 			prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8486 			if (prsm)
8487 				prsm->r_one_out_nr = 1;
8488 		}
8489 		return;
8490 	}
8491 	/*
8492 	 * If we reach here its a retransmission and we need to find it.
8493 	 */
8494 more:
8495 	if (hintrsm && (hintrsm->r_start == seq_out)) {
8496 		rsm = hintrsm;
8497 		hintrsm = NULL;
8498 	} else {
8499 		/* No hints sorry */
8500 		rsm = NULL;
8501 	}
8502 	if ((rsm) && (rsm->r_start == seq_out)) {
8503 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8504 		if (len == 0) {
8505 			return;
8506 		} else {
8507 			goto more;
8508 		}
8509 	}
8510 	/* Ok it was not the last pointer go through it the hard way. */
8511 refind:
8512 	rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8513 	if (rsm) {
8514 		if (rsm->r_start == seq_out) {
8515 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8516 			if (len == 0) {
8517 				return;
8518 			} else {
8519 				goto refind;
8520 			}
8521 		}
8522 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8523 			/* Transmitted within this piece */
8524 			/*
8525 			 * Ok we must split off the front and then let the
8526 			 * update do the rest
8527 			 */
8528 			nrsm = rack_alloc_full_limit(rack);
8529 			if (nrsm == NULL) {
8530 				rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8531 				return;
8532 			}
8533 			/*
8534 			 * copy rsm to nrsm and then trim the front of rsm
8535 			 * to not include this part.
8536 			 */
8537 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
8538 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8539 #ifndef INVARIANTS
8540 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8541 #else
8542 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8543 				panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8544 				      nrsm, insret, rack, rsm);
8545 			}
8546 #endif
8547 			if (rsm->r_in_tmap) {
8548 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8549 				nrsm->r_in_tmap = 1;
8550 			}
8551 			rsm->r_flags &= (~RACK_HAS_FIN);
8552 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8553 			if (len == 0) {
8554 				return;
8555 			} else if (len > 0)
8556 				goto refind;
8557 		}
8558 	}
8559 	/*
8560 	 * Hmm not found in map did they retransmit both old and on into the
8561 	 * new?
8562 	 */
8563 	if (seq_out == tp->snd_max) {
8564 		goto again;
8565 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
8566 #ifdef INVARIANTS
8567 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8568 		       seq_out, len, tp->snd_una, tp->snd_max);
8569 		printf("Starting Dump of all rack entries\n");
8570 		TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8571 			printf("rsm:%p start:%u end:%u\n",
8572 			       rsm, rsm->r_start, rsm->r_end);
8573 		}
8574 		printf("Dump complete\n");
8575 		panic("seq_out not found rack:%p tp:%p",
8576 		      rack, tp);
8577 #endif
8578 	} else {
8579 #ifdef INVARIANTS
8580 		/*
8581 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
8582 		 * flag)
8583 		 */
8584 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8585 		      seq_out, len, tp->snd_max, tp);
8586 #endif
8587 	}
8588 }
8589 
8590 /*
8591  * Record one of the RTT updates from an ack into
8592  * our sample structure.
8593  */
8594 
8595 static void
8596 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8597 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8598 {
8599 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8600 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8601 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8602 	}
8603 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8604 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8605 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8606 	}
8607 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8608 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8609 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
8610 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8611 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8612 	}
8613 	if ((confidence == 1) &&
8614 	    ((rsm == NULL) ||
8615 	     (rsm->r_just_ret) ||
8616 	     (rsm->r_one_out_nr &&
8617 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8618 		/*
8619 		 * If the rsm had a just return
8620 		 * hit it then we can't trust the
8621 		 * rtt measurement for buffer deterimination
8622 		 * Note that a confidence of 2, indicates
8623 		 * SACK'd which overrides the r_just_ret or
8624 		 * the r_one_out_nr. If it was a CUM-ACK and
8625 		 * we had only two outstanding, but get an
8626 		 * ack for only 1. Then that also lowers our
8627 		 * confidence.
8628 		 */
8629 		confidence = 0;
8630 	}
8631 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8632 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8633 		if (rack->r_ctl.rack_rs.confidence == 0) {
8634 			/*
8635 			 * We take anything with no current confidence
8636 			 * saved.
8637 			 */
8638 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8639 			rack->r_ctl.rack_rs.confidence = confidence;
8640 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8641 		} else if (confidence != 0) {
8642 			/*
8643 			 * Once we have a confident number,
8644 			 * we can update it with a smaller
8645 			 * value since this confident number
8646 			 * may include the DSACK time until
8647 			 * the next segment (the second one) arrived.
8648 			 */
8649 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8650 			rack->r_ctl.rack_rs.confidence = confidence;
8651 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8652 		}
8653 	}
8654 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8655 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8656 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8657 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
8658 }
8659 
8660 /*
8661  * Collect new round-trip time estimate
8662  * and update averages and current timeout.
8663  */
8664 static void
8665 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8666 {
8667 	int32_t delta;
8668 	int32_t rtt;
8669 
8670 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8671 		/* No valid sample */
8672 		return;
8673 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8674 		/* We are to use the lowest RTT seen in a single ack */
8675 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8676 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8677 		/* We are to use the highest RTT seen in a single ack */
8678 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8679 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8680 		/* We are to use the average RTT seen in a single ack */
8681 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8682 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8683 	} else {
8684 #ifdef INVARIANTS
8685 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8686 #endif
8687 		return;
8688 	}
8689 	if (rtt == 0)
8690 		rtt = 1;
8691 	if (rack->rc_gp_rtt_set == 0) {
8692 		/*
8693 		 * With no RTT we have to accept
8694 		 * even one we are not confident of.
8695 		 */
8696 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8697 		rack->rc_gp_rtt_set = 1;
8698 	} else if (rack->r_ctl.rack_rs.confidence) {
8699 		/* update the running gp srtt */
8700 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8701 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8702 	}
8703 	if (rack->r_ctl.rack_rs.confidence) {
8704 		/*
8705 		 * record the low and high for highly buffered path computation,
8706 		 * we only do this if we are confident (not a retransmission).
8707 		 */
8708 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8709 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8710 		}
8711 		if (rack->rc_highly_buffered == 0) {
8712 			/*
8713 			 * Currently once we declare a path has
8714 			 * highly buffered there is no going
8715 			 * back, which may be a problem...
8716 			 */
8717 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8718 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8719 						     rack->r_ctl.rc_highest_us_rtt,
8720 						     rack->r_ctl.rc_lowest_us_rtt,
8721 						     RACK_RTTS_SEEHBP);
8722 				rack->rc_highly_buffered = 1;
8723 			}
8724 		}
8725 	}
8726 	if ((rack->r_ctl.rack_rs.confidence) ||
8727 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8728 		/*
8729 		 * If we are highly confident of it <or> it was
8730 		 * never retransmitted we accept it as the last us_rtt.
8731 		 */
8732 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8733 		/* The lowest rtt can be set if its was not retransmited */
8734 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8735 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8736 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
8737 				rack->r_ctl.rc_lowest_us_rtt = 1;
8738 		}
8739 	}
8740 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8741 	if (tp->t_srtt != 0) {
8742 		/*
8743 		 * We keep a simple srtt in microseconds, like our rtt
8744 		 * measurement. We don't need to do any tricks with shifting
8745 		 * etc. Instead we just add in 1/8th of the new measurement
8746 		 * and subtract out 1/8 of the old srtt. We do the same with
8747 		 * the variance after finding the absolute value of the
8748 		 * difference between this sample and the current srtt.
8749 		 */
8750 		delta = tp->t_srtt - rtt;
8751 		/* Take off 1/8th of the current sRTT */
8752 		tp->t_srtt -= (tp->t_srtt >> 3);
8753 		/* Add in 1/8th of the new RTT just measured */
8754 		tp->t_srtt += (rtt >> 3);
8755 		if (tp->t_srtt <= 0)
8756 			tp->t_srtt = 1;
8757 		/* Now lets make the absolute value of the variance */
8758 		if (delta < 0)
8759 			delta = -delta;
8760 		/* Subtract out 1/8th */
8761 		tp->t_rttvar -= (tp->t_rttvar >> 3);
8762 		/* Add in 1/8th of the new variance we just saw */
8763 		tp->t_rttvar += (delta >> 3);
8764 		if (tp->t_rttvar <= 0)
8765 			tp->t_rttvar = 1;
8766 	} else {
8767 		/*
8768 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
8769 		 * variance to half the rtt (so our first retransmit happens
8770 		 * at 3*rtt).
8771 		 */
8772 		tp->t_srtt = rtt;
8773 		tp->t_rttvar = rtt >> 1;
8774 	}
8775 	rack->rc_srtt_measure_made = 1;
8776 	KMOD_TCPSTAT_INC(tcps_rttupdated);
8777 	if (tp->t_rttupdated < UCHAR_MAX)
8778 		tp->t_rttupdated++;
8779 #ifdef STATS
8780 	if (rack_stats_gets_ms_rtt == 0) {
8781 		/* Send in the microsecond rtt used for rxt timeout purposes */
8782 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8783 	} else if (rack_stats_gets_ms_rtt == 1) {
8784 		/* Send in the millisecond rtt used for rxt timeout purposes */
8785 		int32_t ms_rtt;
8786 
8787 		/* Round up */
8788 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8789 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8790 	} else if (rack_stats_gets_ms_rtt == 2) {
8791 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
8792 		int32_t ms_rtt;
8793 
8794 		/* Round up */
8795 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8796 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8797 	}  else {
8798 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
8799 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8800 	}
8801 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8802 #endif
8803 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
8804 	/*
8805 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
8806 	 * way we do the smoothing, srtt and rttvar will each average +1/2
8807 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
8808 	 * tick of rounding and 1 extra tick because of +-1/2 tick
8809 	 * uncertainty in the firing of the timer.  The bias will give us
8810 	 * exactly the 1.5 tick we need.  But, because the bias is
8811 	 * statistical, we have to test that we don't drop below the minimum
8812 	 * feasible timer (which is 2 ticks).
8813 	 */
8814 	tp->t_rxtshift = 0;
8815 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8816 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8817 	rack_log_rtt_sample(rack, rtt);
8818 	tp->t_softerror = 0;
8819 }
8820 
8821 
8822 static void
8823 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8824 {
8825 	/*
8826 	 * Apply to filter the inbound us-rtt at us_cts.
8827 	 */
8828 	uint32_t old_rtt;
8829 
8830 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8831 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8832 			       us_rtt, us_cts);
8833 	if (old_rtt > us_rtt) {
8834 		/* We just hit a new lower rtt time */
8835 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
8836 				     __LINE__, RACK_RTTS_NEWRTT);
8837 		/*
8838 		 * Only count it if its lower than what we saw within our
8839 		 * calculated range.
8840 		 */
8841 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8842 			if (rack_probertt_lower_within &&
8843 			    rack->rc_gp_dyn_mul &&
8844 			    (rack->use_fixed_rate == 0) &&
8845 			    (rack->rc_always_pace)) {
8846 				/*
8847 				 * We are seeing a new lower rtt very close
8848 				 * to the time that we would have entered probe-rtt.
8849 				 * This is probably due to the fact that a peer flow
8850 				 * has entered probe-rtt. Lets go in now too.
8851 				 */
8852 				uint32_t val;
8853 
8854 				val = rack_probertt_lower_within * rack_time_between_probertt;
8855 				val /= 100;
8856 				if ((rack->in_probe_rtt == 0)  &&
8857 				    (rack->rc_skip_timely == 0) &&
8858 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
8859 					rack_enter_probertt(rack, us_cts);
8860 				}
8861 			}
8862 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8863 		}
8864 	}
8865 }
8866 
8867 static int
8868 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8869     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8870 {
8871 	uint32_t us_rtt;
8872 	int32_t i, all;
8873 	uint32_t t, len_acked;
8874 
8875 	if ((rsm->r_flags & RACK_ACKED) ||
8876 	    (rsm->r_flags & RACK_WAS_ACKED))
8877 		/* Already done */
8878 		return (0);
8879 	if (rsm->r_no_rtt_allowed) {
8880 		/* Not allowed */
8881 		return (0);
8882 	}
8883 	if (ack_type == CUM_ACKED) {
8884 		if (SEQ_GT(th_ack, rsm->r_end)) {
8885 			len_acked = rsm->r_end - rsm->r_start;
8886 			all = 1;
8887 		} else {
8888 			len_acked = th_ack - rsm->r_start;
8889 			all = 0;
8890 		}
8891 	} else {
8892 		len_acked = rsm->r_end - rsm->r_start;
8893 		all = 0;
8894 	}
8895 	if (rsm->r_rtr_cnt == 1) {
8896 
8897 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8898 		if ((int)t <= 0)
8899 			t = 1;
8900 		if (!tp->t_rttlow || tp->t_rttlow > t)
8901 			tp->t_rttlow = t;
8902 		if (!rack->r_ctl.rc_rack_min_rtt ||
8903 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8904 			rack->r_ctl.rc_rack_min_rtt = t;
8905 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
8906 				rack->r_ctl.rc_rack_min_rtt = 1;
8907 			}
8908 		}
8909 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8910 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8911 		else
8912 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8913 		if (us_rtt == 0)
8914 			us_rtt = 1;
8915 		if (CC_ALGO(tp)->rttsample != NULL) {
8916 			/* Kick the RTT to the CC */
8917 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8918 		}
8919 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
8920 		if (ack_type == SACKED) {
8921 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
8922 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
8923 		} else {
8924 			/*
8925 			 * We need to setup what our confidence
8926 			 * is in this ack.
8927 			 *
8928 			 * If the rsm was app limited and it is
8929 			 * less than a mss in length (the end
8930 			 * of the send) then we have a gap. If we
8931 			 * were app limited but say we were sending
8932 			 * multiple MSS's then we are more confident
8933 			 * int it.
8934 			 *
8935 			 * When we are not app-limited then we see if
8936 			 * the rsm is being included in the current
8937 			 * measurement, we tell this by the app_limited_needs_set
8938 			 * flag.
8939 			 *
8940 			 * Note that being cwnd blocked is not applimited
8941 			 * as well as the pacing delay between packets which
8942 			 * are sending only 1 or 2 MSS's also will show up
8943 			 * in the RTT. We probably need to examine this algorithm
8944 			 * a bit more and enhance it to account for the delay
8945 			 * between rsm's. We could do that by saving off the
8946 			 * pacing delay of each rsm (in an rsm) and then
8947 			 * factoring that in somehow though for now I am
8948 			 * not sure how :)
8949 			 */
8950 			int calc_conf = 0;
8951 
8952 			if (rsm->r_flags & RACK_APP_LIMITED) {
8953 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
8954 					calc_conf = 0;
8955 				else
8956 					calc_conf = 1;
8957 			} else if (rack->app_limited_needs_set == 0) {
8958 				calc_conf = 1;
8959 			} else {
8960 				calc_conf = 0;
8961 			}
8962 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
8963 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
8964 					    calc_conf, rsm, rsm->r_rtr_cnt);
8965 		}
8966 		if ((rsm->r_flags & RACK_TLP) &&
8967 		    (!IN_FASTRECOVERY(tp->t_flags))) {
8968 			/* Segment was a TLP and our retrans matched */
8969 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
8970 				rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
8971 			}
8972 		}
8973 		if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
8974 		    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
8975 			    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
8976 			/* New more recent rack_tmit_time */
8977 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8978 			if (rack->r_ctl.rc_rack_tmit_time == 0)
8979 				rack->r_ctl.rc_rack_tmit_time = 1;
8980 			rack->rc_rack_rtt = t;
8981 		}
8982 		return (1);
8983 	}
8984 	/*
8985 	 * We clear the soft/rxtshift since we got an ack.
8986 	 * There is no assurance we will call the commit() function
8987 	 * so we need to clear these to avoid incorrect handling.
8988 	 */
8989 	tp->t_rxtshift = 0;
8990 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8991 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
8992 	tp->t_softerror = 0;
8993 	if (to && (to->to_flags & TOF_TS) &&
8994 	    (ack_type == CUM_ACKED) &&
8995 	    (to->to_tsecr) &&
8996 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
8997 		/*
8998 		 * Now which timestamp does it match? In this block the ACK
8999 		 * must be coming from a previous transmission.
9000 		 */
9001 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
9002 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
9003 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9004 				if ((int)t <= 0)
9005 					t = 1;
9006 				if (CC_ALGO(tp)->rttsample != NULL) {
9007 					/*
9008 					 * Kick the RTT to the CC, here
9009 					 * we lie a bit in that we know the
9010 					 * retransmission is correct even though
9011 					 * we retransmitted. This is because
9012 					 * we match the timestamps.
9013 					 */
9014 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
9015 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
9016 					else
9017 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
9018 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9019 				}
9020 				if ((i + 1) < rsm->r_rtr_cnt) {
9021 					/*
9022 					 * The peer ack'd from our previous
9023 					 * transmission. We have a spurious
9024 					 * retransmission and thus we dont
9025 					 * want to update our rack_rtt.
9026 					 *
9027 					 * Hmm should there be a CC revert here?
9028 					 *
9029 					 */
9030 					return (0);
9031 				}
9032 				if (!tp->t_rttlow || tp->t_rttlow > t)
9033 					tp->t_rttlow = t;
9034 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9035 					rack->r_ctl.rc_rack_min_rtt = t;
9036 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
9037 						rack->r_ctl.rc_rack_min_rtt = 1;
9038 					}
9039 				}
9040 				if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9041 				    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9042 					    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9043 					/* New more recent rack_tmit_time */
9044 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9045 					if (rack->r_ctl.rc_rack_tmit_time == 0)
9046 						rack->r_ctl.rc_rack_tmit_time = 1;
9047 					rack->rc_rack_rtt = t;
9048 				}
9049 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9050 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9051 						    rsm->r_rtr_cnt);
9052 				return (1);
9053 			}
9054 		}
9055 		/* If we are logging log out the sendmap */
9056 		if (tcp_bblogging_on(rack->rc_tp)) {
9057 			for (i = 0; i < rsm->r_rtr_cnt; i++) {
9058 				rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9059 			}
9060 		}
9061 		goto ts_not_found;
9062 	} else {
9063 		/*
9064 		 * Ok its a SACK block that we retransmitted. or a windows
9065 		 * machine without timestamps. We can tell nothing from the
9066 		 * time-stamp since its not there or the time the peer last
9067 		 * received a segment that moved forward its cum-ack point.
9068 		 */
9069 ts_not_found:
9070 		i = rsm->r_rtr_cnt - 1;
9071 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9072 		if ((int)t <= 0)
9073 			t = 1;
9074 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9075 			/*
9076 			 * We retransmitted and the ack came back in less
9077 			 * than the smallest rtt we have observed. We most
9078 			 * likely did an improper retransmit as outlined in
9079 			 * 6.2 Step 2 point 2 in the rack-draft so we
9080 			 * don't want to update our rack_rtt. We in
9081 			 * theory (in future) might want to think about reverting our
9082 			 * cwnd state but we won't for now.
9083 			 */
9084 			return (0);
9085 		} else if (rack->r_ctl.rc_rack_min_rtt) {
9086 			/*
9087 			 * We retransmitted it and the retransmit did the
9088 			 * job.
9089 			 */
9090 			if (!rack->r_ctl.rc_rack_min_rtt ||
9091 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9092 				rack->r_ctl.rc_rack_min_rtt = t;
9093 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
9094 					rack->r_ctl.rc_rack_min_rtt = 1;
9095 				}
9096 			}
9097 			if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9098 			    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9099 				    (uint32_t)rsm->r_tim_lastsent[i]))) {
9100 				/* New more recent rack_tmit_time */
9101 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9102 				if (rack->r_ctl.rc_rack_tmit_time == 0)
9103 					rack->r_ctl.rc_rack_tmit_time = 1;
9104 				rack->rc_rack_rtt = t;
9105 			}
9106 			return (1);
9107 		}
9108 	}
9109 	return (0);
9110 }
9111 
9112 /*
9113  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9114  */
9115 static void
9116 rack_log_sack_passed(struct tcpcb *tp,
9117     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
9118 {
9119 	struct rack_sendmap *nrsm;
9120 	uint32_t thresh;
9121 
9122 	/* Get our rxt threshold for lost consideration */
9123 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
9124 	/* Now start looking at rsm's */
9125 	nrsm = rsm;
9126 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9127 	    rack_head, r_tnext) {
9128 		if (nrsm == rsm) {
9129 			/* Skip original segment he is acked */
9130 			continue;
9131 		}
9132 		if (nrsm->r_flags & RACK_ACKED) {
9133 			/*
9134 			 * Skip ack'd segments, though we
9135 			 * should not see these, since tmap
9136 			 * should not have ack'd segments.
9137 			 */
9138 			continue;
9139 		}
9140 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9141 			/*
9142 			 * If the peer dropped the rwnd on
9143 			 * these then we don't worry about them.
9144 			 */
9145 			continue;
9146 		}
9147 		/* Check lost state */
9148 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
9149 			uint32_t exp;
9150 
9151 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
9152 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
9153 				/* We consider it lost */
9154 				nrsm->r_flags |= RACK_WAS_LOST;
9155 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
9156 			}
9157 		}
9158 		if (nrsm->r_flags & RACK_SACK_PASSED) {
9159 			/*
9160 			 * We found one that is already marked
9161 			 * passed, we have been here before and
9162 			 * so all others below this are marked.
9163 			 */
9164 			break;
9165 		}
9166 		nrsm->r_flags |= RACK_SACK_PASSED;
9167 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9168 	}
9169 }
9170 
9171 static void
9172 rack_need_set_test(struct tcpcb *tp,
9173 		   struct tcp_rack *rack,
9174 		   struct rack_sendmap *rsm,
9175 		   tcp_seq th_ack,
9176 		   int line,
9177 		   int use_which)
9178 {
9179 	struct rack_sendmap *s_rsm;
9180 
9181 	if ((tp->t_flags & TF_GPUTINPROG) &&
9182 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9183 		/*
9184 		 * We were app limited, and this ack
9185 		 * butts up or goes beyond the point where we want
9186 		 * to start our next measurement. We need
9187 		 * to record the new gput_ts as here and
9188 		 * possibly update the start sequence.
9189 		 */
9190 		uint32_t seq, ts;
9191 
9192 		if (rsm->r_rtr_cnt > 1) {
9193 			/*
9194 			 * This is a retransmit, can we
9195 			 * really make any assessment at this
9196 			 * point?  We are not really sure of
9197 			 * the timestamp, is it this or the
9198 			 * previous transmission?
9199 			 *
9200 			 * Lets wait for something better that
9201 			 * is not retransmitted.
9202 			 */
9203 			return;
9204 		}
9205 		seq = tp->gput_seq;
9206 		ts = tp->gput_ts;
9207 		rack->app_limited_needs_set = 0;
9208 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9209 		/* Do we start at a new end? */
9210 		if ((use_which == RACK_USE_BEG) &&
9211 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9212 			/*
9213 			 * When we get an ACK that just eats
9214 			 * up some of the rsm, we set RACK_USE_BEG
9215 			 * since whats at r_start (i.e. th_ack)
9216 			 * is left unacked and thats where the
9217 			 * measurement now starts.
9218 			 */
9219 			tp->gput_seq = rsm->r_start;
9220 		}
9221 		if ((use_which == RACK_USE_END) &&
9222 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9223 			/*
9224 			 * We use the end when the cumack
9225 			 * is moving forward and completely
9226 			 * deleting the rsm passed so basically
9227 			 * r_end holds th_ack.
9228 			 *
9229 			 * For SACK's we also want to use the end
9230 			 * since this piece just got sacked and
9231 			 * we want to target anything after that
9232 			 * in our measurement.
9233 			 */
9234 			tp->gput_seq = rsm->r_end;
9235 		}
9236 		if (use_which == RACK_USE_END_OR_THACK) {
9237 			/*
9238 			 * special case for ack moving forward,
9239 			 * not a sack, we need to move all the
9240 			 * way up to where this ack cum-ack moves
9241 			 * to.
9242 			 */
9243 			if (SEQ_GT(th_ack, rsm->r_end))
9244 				tp->gput_seq = th_ack;
9245 			else
9246 				tp->gput_seq = rsm->r_end;
9247 		}
9248 		if (SEQ_LT(tp->gput_seq, tp->snd_max))
9249 			s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9250 		else
9251 			s_rsm = NULL;
9252 		/*
9253 		 * Pick up the correct send time if we can the rsm passed in
9254 		 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9255 		 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9256 		 * find a different seq i.e. the next send up.
9257 		 *
9258 		 * If that has not been sent, s_rsm will be NULL and we must
9259 		 * arrange it so this function will get called again by setting
9260 		 * app_limited_needs_set.
9261 		 */
9262 		if (s_rsm)
9263 			rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9264 		else {
9265 			/* If we hit here we have to have *not* sent tp->gput_seq */
9266 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9267 			/* Set it up so we will go through here again */
9268 			rack->app_limited_needs_set = 1;
9269 		}
9270 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9271 			/*
9272 			 * We moved beyond this guy's range, re-calculate
9273 			 * the new end point.
9274 			 */
9275 			if (rack->rc_gp_filled == 0) {
9276 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9277 			} else {
9278 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9279 			}
9280 		}
9281 		/*
9282 		 * We are moving the goal post, we may be able to clear the
9283 		 * measure_saw_probe_rtt flag.
9284 		 */
9285 		if ((rack->in_probe_rtt == 0) &&
9286 		    (rack->measure_saw_probe_rtt) &&
9287 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9288 			rack->measure_saw_probe_rtt = 0;
9289 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9290 					   seq, tp->gput_seq,
9291 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9292 					    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9293 					   5, line, NULL, 0);
9294 		if (rack->rc_gp_filled &&
9295 		    ((tp->gput_ack - tp->gput_seq) <
9296 		     max(rc_init_window(rack), (MIN_GP_WIN *
9297 						ctf_fixed_maxseg(tp))))) {
9298 			uint32_t ideal_amount;
9299 
9300 			ideal_amount = rack_get_measure_window(tp, rack);
9301 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9302 				/*
9303 				 * There is no sense of continuing this measurement
9304 				 * because its too small to gain us anything we
9305 				 * trust. Skip it and that way we can start a new
9306 				 * measurement quicker.
9307 				 */
9308 				tp->t_flags &= ~TF_GPUTINPROG;
9309 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9310 							   0, 0,
9311 							   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9312 							    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9313 							   6, __LINE__, NULL, 0);
9314 			} else {
9315 				/*
9316 				 * Reset the window further out.
9317 				 */
9318 				tp->gput_ack = tp->gput_seq + ideal_amount;
9319 			}
9320 		}
9321 		rack_tend_gp_marks(tp, rack);
9322 		rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9323 	}
9324 }
9325 
9326 static inline int
9327 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9328 {
9329 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9330 		/* Behind our TLP definition or right at */
9331 		return (0);
9332 	}
9333 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9334 		/* The start is beyond or right at our end of TLP definition */
9335 		return (0);
9336 	}
9337 	/* It has to be a sub-part of the original TLP recorded */
9338 	return (1);
9339 }
9340 
9341 static uint32_t
9342 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9343 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9344 		   uint32_t segsiz)
9345 {
9346 	uint32_t start, end, changed = 0;
9347 	struct rack_sendmap stack_map;
9348 	struct rack_sendmap *rsm, *nrsm, *prev, *next;
9349 	int insret __diagused;
9350 	int32_t used_ref = 1;
9351 	int can_use_hookery = 0;
9352 
9353 	start = sack->start;
9354 	end = sack->end;
9355 	rsm = *prsm;
9356 
9357 do_rest_ofb:
9358 	if ((rsm == NULL) ||
9359 	    (SEQ_LT(end, rsm->r_start)) ||
9360 	    (SEQ_GEQ(start, rsm->r_end)) ||
9361 	    (SEQ_LT(start, rsm->r_start))) {
9362 		/*
9363 		 * We are not in the right spot,
9364 		 * find the correct spot in the tree.
9365 		 */
9366 		used_ref = 0;
9367 		rsm = tqhash_find(rack->r_ctl.tqh, start);
9368 	}
9369 	if (rsm == NULL) {
9370 		/* TSNH */
9371 		goto out;
9372 	}
9373 	/* Ok we have an ACK for some piece of this rsm */
9374 	if (rsm->r_start != start) {
9375 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9376 			/*
9377 			 * Before any splitting or hookery is
9378 			 * done is it a TLP of interest i.e. rxt?
9379 			 */
9380 			if ((rsm->r_flags & RACK_TLP) &&
9381 			    (rsm->r_rtr_cnt > 1)) {
9382 				/*
9383 				 * We are splitting a rxt TLP, check
9384 				 * if we need to save off the start/end
9385 				 */
9386 				if (rack->rc_last_tlp_acked_set &&
9387 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9388 					/*
9389 					 * We already turned this on since we are inside
9390 					 * the previous one was a partially sack now we
9391 					 * are getting another one (maybe all of it).
9392 					 *
9393 					 */
9394 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9395 					/*
9396 					 * Lets make sure we have all of it though.
9397 					 */
9398 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9399 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9400 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9401 								     rack->r_ctl.last_tlp_acked_end);
9402 					}
9403 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9404 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9405 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9406 								     rack->r_ctl.last_tlp_acked_end);
9407 					}
9408 				} else {
9409 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9410 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9411 					rack->rc_last_tlp_past_cumack = 0;
9412 					rack->rc_last_tlp_acked_set = 1;
9413 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9414 				}
9415 			}
9416 			/**
9417 			 * Need to split this in two pieces the before and after,
9418 			 * the before remains in the map, the after must be
9419 			 * added. In other words we have:
9420 			 * rsm        |--------------|
9421 			 * sackblk        |------->
9422 			 * rsm will become
9423 			 *     rsm    |---|
9424 			 * and nrsm will be  the sacked piece
9425 			 *     nrsm       |----------|
9426 			 *
9427 			 * But before we start down that path lets
9428 			 * see if the sack spans over on top of
9429 			 * the next guy and it is already sacked.
9430 			 *
9431 			 */
9432 			/*
9433 			 * Hookery can only be used if the two entries
9434 			 * are in the same bucket and neither one of
9435 			 * them staddle the bucket line.
9436 			 */
9437 			next = tqhash_next(rack->r_ctl.tqh, rsm);
9438 			if (next &&
9439 			    (rsm->bindex == next->bindex) &&
9440 			    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9441 			    ((next->r_flags & RACK_STRADDLE) == 0) &&
9442 			    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9443 			    ((next->r_flags & RACK_IS_PCM) == 0) &&
9444 			    (rsm->r_flags & RACK_IN_GP_WIN) &&
9445 			    (next->r_flags & RACK_IN_GP_WIN))
9446 				can_use_hookery = 1;
9447 			else
9448 				can_use_hookery = 0;
9449 			if (next && can_use_hookery &&
9450 			    (next->r_flags & RACK_ACKED) &&
9451 			    SEQ_GEQ(end, next->r_start)) {
9452 				/**
9453 				 * So the next one is already acked, and
9454 				 * we can thus by hookery use our stack_map
9455 				 * to reflect the piece being sacked and
9456 				 * then adjust the two tree entries moving
9457 				 * the start and ends around. So we start like:
9458 				 *  rsm     |------------|             (not-acked)
9459 				 *  next                 |-----------| (acked)
9460 				 *  sackblk        |-------->
9461 				 *  We want to end like so:
9462 				 *  rsm     |------|                   (not-acked)
9463 				 *  next           |-----------------| (acked)
9464 				 *  nrsm           |-----|
9465 				 * Where nrsm is a temporary stack piece we
9466 				 * use to update all the gizmos.
9467 				 */
9468 				/* Copy up our fudge block */
9469 				nrsm = &stack_map;
9470 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9471 				/* Now adjust our tree blocks */
9472 				tqhash_update_end(rack->r_ctl.tqh, rsm, start);
9473 				next->r_start = start;
9474  				rsm->r_flags |= RACK_SHUFFLED;
9475 				next->r_flags |= RACK_SHUFFLED;
9476 				/* Now we must adjust back where next->m is */
9477 				rack_setup_offset_for_rsm(rack, rsm, next);
9478 				/*
9479 				 * Which timestamp do we keep? It is rather
9480 				 * important in GP measurements to have the
9481 				 * accurate end of the send window.
9482 				 *
9483 				 * We keep the largest value, which is the newest
9484 				 * send. We do this in case a segment that is
9485 				 * joined together and not part of a GP estimate
9486 				 * later gets expanded into the GP estimate.
9487 				 *
9488 				 * We prohibit the merging of unlike kinds i.e.
9489 				 * all pieces that are in the GP estimate can be
9490 				 * merged and all pieces that are not in a GP estimate
9491 				 * can be merged, but not disimilar pieces. Combine
9492 				 * this with taking the highest here and we should
9493 				 * be ok unless of course the client reneges. Then
9494 				 * all bets are off.
9495 				 */
9496 				if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9497 				    nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9498 					next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9499 				/*
9500 				 * And we must keep the newest ack arrival time.
9501 				 */
9502 				if (next->r_ack_arrival <
9503 				    rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9504 					next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9505 
9506 
9507 				/* We don't need to adjust rsm, it did not change */
9508 				/* Clear out the dup ack count of the remainder */
9509 				rsm->r_dupack = 0;
9510 				rsm->r_just_ret = 0;
9511 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9512 				/* Now lets make sure our fudge block is right */
9513 				nrsm->r_start = start;
9514 				/* Now lets update all the stats and such */
9515 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9516 				if (rack->app_limited_needs_set)
9517 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9518 				changed += (nrsm->r_end - nrsm->r_start);
9519 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9520 				if (rsm->r_flags & RACK_WAS_LOST) {
9521 					int my_chg;
9522 
9523 					my_chg = (nrsm->r_end - nrsm->r_start);
9524 					KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9525 						("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9526 					if (my_chg <= rack->r_ctl.rc_considered_lost)
9527 						rack->r_ctl.rc_considered_lost -= my_chg;
9528 					else
9529 						rack->r_ctl.rc_considered_lost = 0;
9530 				}
9531 				if (nrsm->r_flags & RACK_SACK_PASSED) {
9532 					rack->r_ctl.rc_reorder_ts = cts;
9533 					if (rack->r_ctl.rc_reorder_ts == 0)
9534 						rack->r_ctl.rc_reorder_ts = 1;
9535 				}
9536 				/*
9537 				 * Now we want to go up from rsm (the
9538 				 * one left un-acked) to the next one
9539 				 * in the tmap. We do this so when
9540 				 * we walk backwards we include marking
9541 				 * sack-passed on rsm (The one passed in
9542 				 * is skipped since it is generally called
9543 				 * on something sacked before removing it
9544 				 * from the tmap).
9545 				 */
9546 				if (rsm->r_in_tmap) {
9547 					nrsm = TAILQ_NEXT(rsm, r_tnext);
9548 					/*
9549 					 * Now that we have the next
9550 					 * one walk backwards from there.
9551 					 */
9552 					if (nrsm && nrsm->r_in_tmap)
9553 						rack_log_sack_passed(tp, rack, nrsm, cts);
9554 				}
9555 				/* Now are we done? */
9556 				if (SEQ_LT(end, next->r_end) ||
9557 				    (end == next->r_end)) {
9558 					/* Done with block */
9559 					goto out;
9560 				}
9561 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9562 				counter_u64_add(rack_sack_used_next_merge, 1);
9563 				/* Postion for the next block */
9564 				start = next->r_end;
9565 				rsm = tqhash_next(rack->r_ctl.tqh, next);
9566 				if (rsm == NULL)
9567 					goto out;
9568 			} else {
9569 				/**
9570 				 * We can't use any hookery here, so we
9571 				 * need to split the map. We enter like
9572 				 * so:
9573 				 *  rsm      |--------|
9574 				 *  sackblk       |----->
9575 				 * We will add the new block nrsm and
9576 				 * that will be the new portion, and then
9577 				 * fall through after reseting rsm. So we
9578 				 * split and look like this:
9579 				 *  rsm      |----|
9580 				 *  sackblk       |----->
9581 				 *  nrsm          |---|
9582 				 * We then fall through reseting
9583 				 * rsm to nrsm, so the next block
9584 				 * picks it up.
9585 				 */
9586 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9587 				if (nrsm == NULL) {
9588 					/*
9589 					 * failed XXXrrs what can we do but loose the sack
9590 					 * info?
9591 					 */
9592 					goto out;
9593 				}
9594 				counter_u64_add(rack_sack_splits, 1);
9595 				rack_clone_rsm(rack, nrsm, rsm, start);
9596 				rsm->r_just_ret = 0;
9597 #ifndef INVARIANTS
9598 				(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9599 #else
9600 				if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9601 					panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9602 					      nrsm, insret, rack, rsm);
9603 				}
9604 #endif
9605 				if (rsm->r_in_tmap) {
9606 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9607 					nrsm->r_in_tmap = 1;
9608 				}
9609 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9610 				rsm->r_flags &= (~RACK_HAS_FIN);
9611 				/* Position us to point to the new nrsm that starts the sack blk */
9612 				rsm = nrsm;
9613 			}
9614 		} else {
9615 			/* Already sacked this piece */
9616 			counter_u64_add(rack_sack_skipped_acked, 1);
9617 			if (end == rsm->r_end) {
9618 				/* Done with block */
9619 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9620 				goto out;
9621 			} else if (SEQ_LT(end, rsm->r_end)) {
9622 				/* A partial sack to a already sacked block */
9623 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9624 				goto out;
9625 			} else {
9626 				/*
9627 				 * The end goes beyond this guy
9628 				 * reposition the start to the
9629 				 * next block.
9630 				 */
9631 				start = rsm->r_end;
9632 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9633 				if (rsm == NULL)
9634 					goto out;
9635 			}
9636 		}
9637 	}
9638 	if (SEQ_GEQ(end, rsm->r_end)) {
9639 		/**
9640 		 * The end of this block is either beyond this guy or right
9641 		 * at this guy. I.e.:
9642 		 *  rsm ---                 |-----|
9643 		 *  end                     |-----|
9644 		 *  <or>
9645 		 *  end                     |---------|
9646 		 */
9647 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9648 			/*
9649 			 * Is it a TLP of interest?
9650 			 */
9651 			if ((rsm->r_flags & RACK_TLP) &&
9652 			    (rsm->r_rtr_cnt > 1)) {
9653 				/*
9654 				 * We are splitting a rxt TLP, check
9655 				 * if we need to save off the start/end
9656 				 */
9657 				if (rack->rc_last_tlp_acked_set &&
9658 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9659 					/*
9660 					 * We already turned this on since we are inside
9661 					 * the previous one was a partially sack now we
9662 					 * are getting another one (maybe all of it).
9663 					 */
9664 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9665 					/*
9666 					 * Lets make sure we have all of it though.
9667 					 */
9668 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9669 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9670 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9671 								     rack->r_ctl.last_tlp_acked_end);
9672 					}
9673 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9674 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9675 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9676 								     rack->r_ctl.last_tlp_acked_end);
9677 					}
9678 				} else {
9679 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9680 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9681 					rack->rc_last_tlp_past_cumack = 0;
9682 					rack->rc_last_tlp_acked_set = 1;
9683 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9684 				}
9685 			}
9686 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9687 			changed += (rsm->r_end - rsm->r_start);
9688 			/* You get a count for acking a whole segment or more */
9689 			if (rsm->r_flags & RACK_WAS_LOST) {
9690 				int my_chg;
9691 
9692 				my_chg = (rsm->r_end - rsm->r_start);
9693 				rsm->r_flags &= ~RACK_WAS_LOST;
9694 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9695 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9696 				if (my_chg <= rack->r_ctl.rc_considered_lost)
9697 					rack->r_ctl.rc_considered_lost -= my_chg;
9698 				else
9699 					rack->r_ctl.rc_considered_lost = 0;
9700 			}
9701 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9702 			if (rsm->r_in_tmap) /* should be true */
9703 				rack_log_sack_passed(tp, rack, rsm, cts);
9704 			/* Is Reordering occuring? */
9705 			if (rsm->r_flags & RACK_SACK_PASSED) {
9706 				rsm->r_flags &= ~RACK_SACK_PASSED;
9707 				rack->r_ctl.rc_reorder_ts = cts;
9708 				if (rack->r_ctl.rc_reorder_ts == 0)
9709 					rack->r_ctl.rc_reorder_ts = 1;
9710 			}
9711 			if (rack->app_limited_needs_set)
9712 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9713 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9714 			rsm->r_flags |= RACK_ACKED;
9715 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
9716 			if (rsm->r_in_tmap) {
9717 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9718 				rsm->r_in_tmap = 0;
9719 			}
9720 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9721 		} else {
9722 			counter_u64_add(rack_sack_skipped_acked, 1);
9723 		}
9724 		if (end == rsm->r_end) {
9725 			/* This block only - done, setup for next */
9726 			goto out;
9727 		}
9728 		/*
9729 		 * There is more not coverend by this rsm move on
9730 		 * to the next block in the tail queue hash table.
9731 		 */
9732 		nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9733 		start = rsm->r_end;
9734 		rsm = nrsm;
9735 		if (rsm == NULL)
9736 			goto out;
9737 		goto do_rest_ofb;
9738 	}
9739 	/**
9740 	 * The end of this sack block is smaller than
9741 	 * our rsm i.e.:
9742 	 *  rsm ---                 |-----|
9743 	 *  end                     |--|
9744 	 */
9745 	if ((rsm->r_flags & RACK_ACKED) == 0) {
9746 		/*
9747 		 * Is it a TLP of interest?
9748 		 */
9749 		if ((rsm->r_flags & RACK_TLP) &&
9750 		    (rsm->r_rtr_cnt > 1)) {
9751 			/*
9752 			 * We are splitting a rxt TLP, check
9753 			 * if we need to save off the start/end
9754 			 */
9755 			if (rack->rc_last_tlp_acked_set &&
9756 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9757 				/*
9758 				 * We already turned this on since we are inside
9759 				 * the previous one was a partially sack now we
9760 				 * are getting another one (maybe all of it).
9761 				 */
9762 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9763 				/*
9764 				 * Lets make sure we have all of it though.
9765 				 */
9766 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9767 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9768 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9769 							     rack->r_ctl.last_tlp_acked_end);
9770 				}
9771 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9772 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9773 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9774 							     rack->r_ctl.last_tlp_acked_end);
9775 				}
9776 			} else {
9777 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9778 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9779 				rack->rc_last_tlp_past_cumack = 0;
9780 				rack->rc_last_tlp_acked_set = 1;
9781 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9782 			}
9783 		}
9784 		/*
9785 		 * Hookery can only be used if the two entries
9786 		 * are in the same bucket and neither one of
9787 		 * them staddle the bucket line.
9788 		 */
9789 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9790 		if (prev &&
9791 		    (rsm->bindex == prev->bindex) &&
9792 		    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9793 		    ((prev->r_flags & RACK_STRADDLE) == 0) &&
9794 		    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9795 		    ((prev->r_flags & RACK_IS_PCM) == 0) &&
9796 		    (rsm->r_flags & RACK_IN_GP_WIN) &&
9797 		    (prev->r_flags & RACK_IN_GP_WIN))
9798 			can_use_hookery = 1;
9799 		else
9800 			can_use_hookery = 0;
9801 		if (prev && can_use_hookery &&
9802 		    (prev->r_flags & RACK_ACKED)) {
9803 			/**
9804 			 * Goal, we want the right remainder of rsm to shrink
9805 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
9806 			 * We want to expand prev to go all the way
9807 			 * to prev->r_end <- end.
9808 			 * so in the tree we have before:
9809 			 *   prev     |--------|         (acked)
9810 			 *   rsm               |-------| (non-acked)
9811 			 *   sackblk           |-|
9812 			 * We churn it so we end up with
9813 			 *   prev     |----------|       (acked)
9814 			 *   rsm                 |-----| (non-acked)
9815 			 *   nrsm              |-| (temporary)
9816 			 *
9817 			 * Note if either prev/rsm is a TLP we don't
9818 			 * do this.
9819 			 */
9820 			nrsm = &stack_map;
9821 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9822 			tqhash_update_end(rack->r_ctl.tqh, prev, end);
9823 			rsm->r_start = end;
9824 			rsm->r_flags |= RACK_SHUFFLED;
9825 			prev->r_flags |= RACK_SHUFFLED;
9826 			/* Now adjust nrsm (stack copy) to be
9827 			 * the one that is the small
9828 			 * piece that was "sacked".
9829 			 */
9830 			nrsm->r_end = end;
9831 			rsm->r_dupack = 0;
9832 			/*
9833 			 * Which timestamp do we keep? It is rather
9834 			 * important in GP measurements to have the
9835 			 * accurate end of the send window.
9836 			 *
9837 			 * We keep the largest value, which is the newest
9838 			 * send. We do this in case a segment that is
9839 			 * joined together and not part of a GP estimate
9840 			 * later gets expanded into the GP estimate.
9841 			 *
9842 			 * We prohibit the merging of unlike kinds i.e.
9843 			 * all pieces that are in the GP estimate can be
9844 			 * merged and all pieces that are not in a GP estimate
9845 			 * can be merged, but not disimilar pieces. Combine
9846 			 * this with taking the highest here and we should
9847 			 * be ok unless of course the client reneges. Then
9848 			 * all bets are off.
9849 			 */
9850 			if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
9851 			   nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
9852 				prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9853 			}
9854 			/*
9855 			 * And we must keep the newest ack arrival time.
9856 			 */
9857 
9858 			if(prev->r_ack_arrival <
9859 			   rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9860 				prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9861 
9862 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9863 			/*
9864 			 * Now that the rsm has had its start moved forward
9865 			 * lets go ahead and get its new place in the world.
9866 			 */
9867 			rack_setup_offset_for_rsm(rack, prev, rsm);
9868 			/*
9869 			 * Now nrsm is our new little piece
9870 			 * that is acked (which was merged
9871 			 * to prev). Update the rtt and changed
9872 			 * based on that. Also check for reordering.
9873 			 */
9874 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9875 			if (rack->app_limited_needs_set)
9876 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9877 			changed += (nrsm->r_end - nrsm->r_start);
9878 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9879 			if (rsm->r_flags & RACK_WAS_LOST) {
9880 				int my_chg;
9881 
9882 				my_chg = (nrsm->r_end - nrsm->r_start);
9883 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9884 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9885 				if (my_chg <= rack->r_ctl.rc_considered_lost)
9886 					rack->r_ctl.rc_considered_lost -= my_chg;
9887 				else
9888 					rack->r_ctl.rc_considered_lost = 0;
9889 			}
9890 			if (nrsm->r_flags & RACK_SACK_PASSED) {
9891 				rack->r_ctl.rc_reorder_ts = cts;
9892 				if (rack->r_ctl.rc_reorder_ts == 0)
9893 					rack->r_ctl.rc_reorder_ts = 1;
9894 			}
9895 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
9896 			rsm = prev;
9897 			counter_u64_add(rack_sack_used_prev_merge, 1);
9898 		} else {
9899 			/**
9900 			 * This is the case where our previous
9901 			 * block is not acked either, so we must
9902 			 * split the block in two.
9903 			 */
9904 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9905 			if (nrsm == NULL) {
9906 				/* failed rrs what can we do but loose the sack info? */
9907 				goto out;
9908 			}
9909 			if ((rsm->r_flags & RACK_TLP) &&
9910 			    (rsm->r_rtr_cnt > 1)) {
9911 				/*
9912 				 * We are splitting a rxt TLP, check
9913 				 * if we need to save off the start/end
9914 				 */
9915 				if (rack->rc_last_tlp_acked_set &&
9916 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9917 					/*
9918 					 * We already turned this on since this block is inside
9919 					 * the previous one was a partially sack now we
9920 					 * are getting another one (maybe all of it).
9921 					 */
9922 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9923 					/*
9924 					 * Lets make sure we have all of it though.
9925 					 */
9926 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9927 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9928 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9929 								     rack->r_ctl.last_tlp_acked_end);
9930 					}
9931 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9932 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9933 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9934 								     rack->r_ctl.last_tlp_acked_end);
9935 					}
9936 				} else {
9937 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9938 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9939 					rack->rc_last_tlp_acked_set = 1;
9940 					rack->rc_last_tlp_past_cumack = 0;
9941 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9942 				}
9943 			}
9944 			/**
9945 			 * In this case nrsm becomes
9946 			 * nrsm->r_start = end;
9947 			 * nrsm->r_end = rsm->r_end;
9948 			 * which is un-acked.
9949 			 * <and>
9950 			 * rsm->r_end = nrsm->r_start;
9951 			 * i.e. the remaining un-acked
9952 			 * piece is left on the left
9953 			 * hand side.
9954 			 *
9955 			 * So we start like this
9956 			 * rsm      |----------| (not acked)
9957 			 * sackblk  |---|
9958 			 * build it so we have
9959 			 * rsm      |---|         (acked)
9960 			 * nrsm         |------|  (not acked)
9961 			 */
9962 			counter_u64_add(rack_sack_splits, 1);
9963 			rack_clone_rsm(rack, nrsm, rsm, end);
9964 			rsm->r_flags &= (~RACK_HAS_FIN);
9965 			rsm->r_just_ret = 0;
9966 #ifndef INVARIANTS
9967 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9968 #else
9969 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9970 				panic("Insert in tailq_hash of %p fails ret:% rack:%p rsm:%p",
9971 				      nrsm, insret, rack, rsm);
9972 			}
9973 #endif
9974 			if (rsm->r_in_tmap) {
9975 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9976 				nrsm->r_in_tmap = 1;
9977 			}
9978 			nrsm->r_dupack = 0;
9979 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
9980 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9981 			changed += (rsm->r_end - rsm->r_start);
9982 			if (rsm->r_flags & RACK_WAS_LOST) {
9983 				int my_chg;
9984 
9985 				my_chg = (rsm->r_end - rsm->r_start);
9986 				rsm->r_flags &= ~RACK_WAS_LOST;
9987 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9988 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9989 				if (my_chg <= rack->r_ctl.rc_considered_lost)
9990 					rack->r_ctl.rc_considered_lost -= my_chg;
9991 				else
9992 					rack->r_ctl.rc_considered_lost = 0;
9993 			}
9994 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9995 
9996 			if (rsm->r_in_tmap) /* should be true */
9997 				rack_log_sack_passed(tp, rack, rsm, cts);
9998 			/* Is Reordering occuring? */
9999 			if (rsm->r_flags & RACK_SACK_PASSED) {
10000 				rsm->r_flags &= ~RACK_SACK_PASSED;
10001 				rack->r_ctl.rc_reorder_ts = cts;
10002 				if (rack->r_ctl.rc_reorder_ts == 0)
10003 					rack->r_ctl.rc_reorder_ts = 1;
10004 			}
10005 			if (rack->app_limited_needs_set)
10006 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10007 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10008 			rsm->r_flags |= RACK_ACKED;
10009 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
10010 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
10011 			if (rsm->r_in_tmap) {
10012 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10013 				rsm->r_in_tmap = 0;
10014 			}
10015 		}
10016 	} else if (start != end){
10017 		/*
10018 		 * The block was already acked.
10019 		 */
10020 		counter_u64_add(rack_sack_skipped_acked, 1);
10021 	}
10022 out:
10023 	if (rsm &&
10024 	    ((rsm->r_flags & RACK_TLP) == 0) &&
10025 	    (rsm->r_flags & RACK_ACKED)) {
10026 		/*
10027 		 * Now can we merge where we worked
10028 		 * with either the previous or
10029 		 * next block?
10030 		 */
10031 		next = tqhash_next(rack->r_ctl.tqh, rsm);
10032 		while (next) {
10033 			if (next->r_flags & RACK_TLP)
10034 				break;
10035 			/* Only allow merges between ones in or out of GP window */
10036 			if ((next->r_flags & RACK_IN_GP_WIN) &&
10037 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10038 				break;
10039 			}
10040 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10041 			    ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10042 				break;
10043 			}
10044 			if (rsm->bindex != next->bindex)
10045 				break;
10046 			if (rsm->r_flags & RACK_STRADDLE)
10047 				break;
10048 			if (rsm->r_flags & RACK_IS_PCM)
10049 				break;
10050 			if (next->r_flags & RACK_STRADDLE)
10051 				break;
10052 			if (next->r_flags & RACK_IS_PCM)
10053 				break;
10054 			if (next->r_flags & RACK_ACKED) {
10055 				/* yep this and next can be merged */
10056 				rsm = rack_merge_rsm(rack, rsm, next);
10057 				next = tqhash_next(rack->r_ctl.tqh, rsm);
10058 			} else
10059 				break;
10060 		}
10061 		/* Now what about the previous? */
10062 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10063 		while (prev) {
10064 			if (prev->r_flags & RACK_TLP)
10065 				break;
10066 			/* Only allow merges between ones in or out of GP window */
10067 			if ((prev->r_flags & RACK_IN_GP_WIN) &&
10068 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10069 				break;
10070 			}
10071 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10072 			    ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10073 				break;
10074 			}
10075 			if (rsm->bindex != prev->bindex)
10076 				break;
10077 			if (rsm->r_flags & RACK_STRADDLE)
10078 				break;
10079 			if (rsm->r_flags & RACK_IS_PCM)
10080 				break;
10081 			if (prev->r_flags & RACK_STRADDLE)
10082 				break;
10083 			if (prev->r_flags & RACK_IS_PCM)
10084 				break;
10085 			if (prev->r_flags & RACK_ACKED) {
10086 				/* yep the previous and this can be merged */
10087 				rsm = rack_merge_rsm(rack, prev, rsm);
10088 				prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10089 			} else
10090 				break;
10091 		}
10092 	}
10093 	if (used_ref == 0) {
10094 		counter_u64_add(rack_sack_proc_all, 1);
10095 	} else {
10096 		counter_u64_add(rack_sack_proc_short, 1);
10097 	}
10098 	/* Save off the next one for quick reference. */
10099 	nrsm = tqhash_find(rack->r_ctl.tqh, end);
10100 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
10101 	return (changed);
10102 }
10103 
10104 static void inline
10105 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10106 {
10107 	struct rack_sendmap *tmap;
10108 
10109 	tmap = NULL;
10110 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
10111 		/* Its no longer sacked, mark it so */
10112 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10113 #ifdef INVARIANTS
10114 		if (rsm->r_in_tmap) {
10115 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
10116 			      rack, rsm, rsm->r_flags);
10117 		}
10118 #endif
10119 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10120 		/* Rebuild it into our tmap */
10121 		if (tmap == NULL) {
10122 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10123 			tmap = rsm;
10124 		} else {
10125 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10126 			tmap = rsm;
10127 		}
10128 		tmap->r_in_tmap = 1;
10129 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10130 	}
10131 	/*
10132 	 * Now lets possibly clear the sack filter so we start
10133 	 * recognizing sacks that cover this area.
10134 	 */
10135 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10136 
10137 }
10138 
10139 
10140 static void inline
10141 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10142 {
10143 	/*
10144 	 * We look at advancing the end send time for our GP
10145 	 * measurement tracking only as the cumulative acknowledgment
10146 	 * moves forward. You might wonder about this, why not
10147 	 * at every transmission or retransmission within the
10148 	 * GP window update the rc_gp_cumack_ts? Well its rather
10149 	 * nuanced but basically the GP window *may* expand (as
10150 	 * it does below) or worse and harder to track it may shrink.
10151 	 *
10152 	 * This last makes it impossible to track at the time of
10153 	 * the send, since you may set forward your rc_gp_cumack_ts
10154 	 * when you send, because that send *is* in your currently
10155 	 * "guessed" window, but then it shrinks. Now which was
10156 	 * the send time of the last bytes in the window, by the
10157 	 * time you ask that question that part of the sendmap
10158 	 * is freed. So you don't know and you will have too
10159 	 * long of send window. Instead by updating the time
10160 	 * marker only when the cumack advances this assures us
10161 	 * that we will have only the sends in the window of our
10162 	 * GP measurement.
10163 	 *
10164 	 * Another complication from this is the
10165 	 * merging of sendmap entries. During SACK processing this
10166 	 * can happen to conserve the sendmap size. That breaks
10167 	 * everything down in tracking the send window of the GP
10168 	 * estimate. So to prevent that and keep it working with
10169 	 * a tiny bit more limited merging, we only allow like
10170 	 * types to be merged. I.e. if two sends are in the GP window
10171 	 * then its ok to merge them together. If two sends are not
10172 	 * in the GP window its ok to merge them together too. Though
10173 	 * one send in and one send out cannot be merged. We combine
10174 	 * this with never allowing the shrinking of the GP window when
10175 	 * we are in recovery so that we can properly calculate the
10176 	 * sending times.
10177 	 *
10178 	 * This all of course seems complicated, because it is.. :)
10179 	 *
10180 	 * The cum-ack is being advanced upon the sendmap.
10181 	 * If we are not doing a GP estimate don't
10182 	 * proceed.
10183 	 */
10184 	uint64_t ts;
10185 
10186 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
10187 		return;
10188 	/*
10189 	 * If this sendmap entry is going
10190 	 * beyond the measurement window we had picked,
10191 	 * expand the measurement window by that much.
10192 	 */
10193 	if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10194 		tp->gput_ack = rsm->r_end;
10195 	}
10196 	/*
10197 	 * If we have not setup a ack, then we
10198 	 * have no idea if the newly acked pieces
10199 	 * will be "in our seq measurement range". If
10200 	 * it is when we clear the app_limited_needs_set
10201 	 * flag the timestamp will be updated.
10202 	 */
10203 	if (rack->app_limited_needs_set)
10204 		return;
10205 	/*
10206 	 * Finally, we grab out the latest timestamp
10207 	 * that this packet was sent and then see
10208 	 * if:
10209 	 *  a) The packet touches are newly defined GP range.
10210 	 *  b) The time is greater than (newer) than the
10211 	 *     one we currently have. If so we update
10212 	 *     our sending end time window.
10213 	 *
10214 	 * Note we *do not* do this at send time. The reason
10215 	 * is that if you do you *may* pick up a newer timestamp
10216 	 * for a range you are not going to measure. We project
10217 	 * out how far and then sometimes modify that to be
10218 	 * smaller. If that occurs then you will have a send
10219 	 * that does not belong to the range included.
10220 	 */
10221 	if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10222 	    rack->r_ctl.rc_gp_cumack_ts)
10223 		return;
10224 	if (rack_in_gp_window(tp, rsm)) {
10225 		rack->r_ctl.rc_gp_cumack_ts = ts;
10226 		rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10227 			       __LINE__, from, rsm);
10228 	}
10229 }
10230 
10231 static void
10232 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)
10233 {
10234 	struct rack_sendmap *rsm;
10235 	/*
10236 	 * The ACK point is advancing to th_ack, we must drop off
10237 	 * the packets in the rack log and calculate any eligble
10238 	 * RTT's.
10239 	 */
10240 
10241 	if (sack_filter_blks_used(&rack->r_ctl.rack_sf)) {
10242 		/*
10243 		 * If we have some sack blocks in the filter
10244 		 * lets prune them out by calling sfb with no blocks.
10245 		 */
10246 		sack_filter_blks(tp, &rack->r_ctl.rack_sf, NULL, 0, th_ack);
10247 	}
10248 	if (SEQ_GT(th_ack, tp->snd_una)) {
10249 		/* Clear any app ack remembered settings */
10250 		rack->r_ctl.cleared_app_ack = 0;
10251 	}
10252 	rack->r_wanted_output = 1;
10253 	if (SEQ_GT(th_ack, tp->snd_una))
10254 		rack->r_ctl.last_cumack_advance = acktime;
10255 
10256 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
10257 	if ((rack->rc_last_tlp_acked_set == 1)&&
10258 	    (rack->rc_last_tlp_past_cumack == 1) &&
10259 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10260 		/*
10261 		 * We have reached the point where our last rack
10262 		 * tlp retransmit sequence is ahead of the cum-ack.
10263 		 * This can only happen when the cum-ack moves all
10264 		 * the way around (its been a full 2^^31+1 bytes
10265 		 * or more since we sent a retransmitted TLP). Lets
10266 		 * turn off the valid flag since its not really valid.
10267 		 *
10268 		 * Note since sack's also turn on this event we have
10269 		 * a complication, we have to wait to age it out until
10270 		 * the cum-ack is by the TLP before checking which is
10271 		 * what the next else clause does.
10272 		 */
10273 		rack_log_dsack_event(rack, 9, __LINE__,
10274 				     rack->r_ctl.last_tlp_acked_start,
10275 				     rack->r_ctl.last_tlp_acked_end);
10276 		rack->rc_last_tlp_acked_set = 0;
10277 		rack->rc_last_tlp_past_cumack = 0;
10278 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
10279 		   (rack->rc_last_tlp_past_cumack == 0) &&
10280 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10281 		/*
10282 		 * It is safe to start aging TLP's out.
10283 		 */
10284 		rack->rc_last_tlp_past_cumack = 1;
10285 	}
10286 	/* We do the same for the tlp send seq as well */
10287 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10288 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
10289 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
10290 		rack_log_dsack_event(rack, 9, __LINE__,
10291 				     rack->r_ctl.last_sent_tlp_seq,
10292 				     (rack->r_ctl.last_sent_tlp_seq +
10293 				      rack->r_ctl.last_sent_tlp_len));
10294 		rack->rc_last_sent_tlp_seq_valid = 0;
10295 		rack->rc_last_sent_tlp_past_cumack = 0;
10296 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10297 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
10298 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10299 		/*
10300 		 * It is safe to start aging TLP's send.
10301 		 */
10302 		rack->rc_last_sent_tlp_past_cumack = 1;
10303 	}
10304 more:
10305 	rsm = tqhash_min(rack->r_ctl.tqh);
10306 	if (rsm == NULL) {
10307 		if ((th_ack - 1) == tp->iss) {
10308 			/*
10309 			 * For the SYN incoming case we will not
10310 			 * have called tcp_output for the sending of
10311 			 * the SYN, so there will be no map. All
10312 			 * other cases should probably be a panic.
10313 			 */
10314 			return;
10315 		}
10316 		if (tp->t_flags & TF_SENTFIN) {
10317 			/* if we sent a FIN we often will not have map */
10318 			return;
10319 		}
10320 #ifdef INVARIANTS
10321 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u\n",
10322 		      tp,
10323 		      tp->t_state, th_ack, rack,
10324 		      tp->snd_una, tp->snd_max);
10325 #endif
10326 		return;
10327 	}
10328 	if (SEQ_LT(th_ack, rsm->r_start)) {
10329 		/* Huh map is missing this */
10330 #ifdef INVARIANTS
10331 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10332 		       rsm->r_start,
10333 		       th_ack, tp->t_state, rack->r_state);
10334 #endif
10335 		return;
10336 	}
10337 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10338 
10339 	/* Now was it a retransmitted TLP? */
10340 	if ((rsm->r_flags & RACK_TLP) &&
10341 	    (rsm->r_rtr_cnt > 1)) {
10342 		/*
10343 		 * Yes, this rsm was a TLP and retransmitted, remember that
10344 		 * since if a DSACK comes back on this we don't want
10345 		 * to think of it as a reordered segment. This may
10346 		 * get updated again with possibly even other TLPs
10347 		 * in flight, but thats ok. Only when we don't send
10348 		 * a retransmitted TLP for 1/2 the sequences space
10349 		 * will it get turned off (above).
10350 		 */
10351 		if (rack->rc_last_tlp_acked_set &&
10352 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10353 			/*
10354 			 * We already turned this on since the end matches,
10355 			 * the previous one was a partially ack now we
10356 			 * are getting another one (maybe all of it).
10357 			 */
10358 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10359 			/*
10360 			 * Lets make sure we have all of it though.
10361 			 */
10362 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10363 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10364 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10365 						     rack->r_ctl.last_tlp_acked_end);
10366 			}
10367 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10368 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10369 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10370 						     rack->r_ctl.last_tlp_acked_end);
10371 			}
10372 		} else {
10373 			rack->rc_last_tlp_past_cumack = 1;
10374 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10375 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10376 			rack->rc_last_tlp_acked_set = 1;
10377 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10378 		}
10379 	}
10380 	/* Now do we consume the whole thing? */
10381 	rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
10382 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
10383 		/* Its all consumed. */
10384 		uint32_t left;
10385 		uint8_t newly_acked;
10386 
10387 		if (rsm->r_flags & RACK_WAS_LOST) {
10388 			/*
10389 			 * This can happen when we marked it as lost
10390 			 * and yet before retransmitting we get an ack
10391 			 * which can happen due to reordering.
10392 			 */
10393 			rsm->r_flags  &= ~RACK_WAS_LOST;
10394 			KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
10395 				("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10396 			if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
10397 				rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
10398 			else
10399 				rack->r_ctl.rc_considered_lost = 0;
10400 		}
10401 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10402 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10403 		rsm->r_rtr_bytes = 0;
10404 		/*
10405 		 * Record the time of highest cumack sent if its in our measurement
10406 		 * window and possibly bump out the end.
10407 		 */
10408 		rack_rsm_sender_update(rack, tp, rsm, 4);
10409 		tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10410 		if (rsm->r_in_tmap) {
10411 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10412 			rsm->r_in_tmap = 0;
10413 		}
10414 		newly_acked = 1;
10415 		if (rsm->r_flags & RACK_ACKED) {
10416 			/*
10417 			 * It was acked on the scoreboard -- remove
10418 			 * it from total
10419 			 */
10420 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10421 			newly_acked = 0;
10422 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
10423 			/*
10424 			 * There are segments ACKED on the
10425 			 * scoreboard further up. We are seeing
10426 			 * reordering.
10427 			 */
10428 			rsm->r_flags &= ~RACK_SACK_PASSED;
10429 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10430 			rsm->r_flags |= RACK_ACKED;
10431 			rack->r_ctl.rc_reorder_ts = cts;
10432 			if (rack->r_ctl.rc_reorder_ts == 0)
10433 				rack->r_ctl.rc_reorder_ts = 1;
10434 			if (rack->r_ent_rec_ns) {
10435 				/*
10436 				 * We have sent no more, and we saw an sack
10437 				 * then ack arrive.
10438 				 */
10439 				rack->r_might_revert = 1;
10440 			}
10441 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10442 		} else {
10443 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10444 		}
10445 		if ((rsm->r_flags & RACK_TO_REXT) &&
10446 		    (tp->t_flags & TF_RCVD_TSTMP) &&
10447 		    (to->to_flags & TOF_TS) &&
10448 		    (to->to_tsecr != 0) &&
10449 		    (tp->t_flags & TF_PREVVALID)) {
10450 			/*
10451 			 * We can use the timestamp to see
10452 			 * if this retransmission was from the
10453 			 * first transmit. If so we made a mistake.
10454 			 */
10455 			tp->t_flags &= ~TF_PREVVALID;
10456 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10457 				/* The first transmit is what this ack is for */
10458 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10459 			}
10460 		}
10461 		left = th_ack - rsm->r_end;
10462 		if (rack->app_limited_needs_set && newly_acked)
10463 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10464 		/* Free back to zone */
10465 		rack_free(rack, rsm);
10466 		if (left) {
10467 			goto more;
10468 		}
10469 		/* Check for reneging */
10470 		rsm = tqhash_min(rack->r_ctl.tqh);
10471 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10472 			/*
10473 			 * The peer has moved snd_una up to
10474 			 * the edge of this send, i.e. one
10475 			 * that it had previously acked. The only
10476 			 * way that can be true if the peer threw
10477 			 * away data (space issues) that it had
10478 			 * previously sacked (else it would have
10479 			 * given us snd_una up to (rsm->r_end).
10480 			 * We need to undo the acked markings here.
10481 			 *
10482 			 * Note we have to look to make sure th_ack is
10483 			 * our rsm->r_start in case we get an old ack
10484 			 * where th_ack is behind snd_una.
10485 			 */
10486 			rack_peer_reneges(rack, rsm, th_ack);
10487 		}
10488 		return;
10489 	}
10490 	if (rsm->r_flags & RACK_ACKED) {
10491 		/*
10492 		 * It was acked on the scoreboard -- remove it from
10493 		 * total for the part being cum-acked.
10494 		 */
10495 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10496 	} else {
10497 		rack_update_pcm_ack(rack, 1, rsm->r_start, th_ack);
10498 	}
10499 	/* And what about the lost flag? */
10500 	if (rsm->r_flags & RACK_WAS_LOST) {
10501 		/*
10502 		 * This can happen when we marked it as lost
10503 		 * and yet before retransmitting we get an ack
10504 		 * which can happen due to reordering. In this
10505 		 * case its only a partial ack of the send.
10506 		 */
10507 		KASSERT((rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start)),
10508 			("rsm:%p rack:%p rc_considered_lost goes negative th_ack:%u", rsm,  rack, th_ack));
10509 		if (rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start))
10510 			rack->r_ctl.rc_considered_lost -= th_ack - rsm->r_start;
10511 		else
10512 			rack->r_ctl.rc_considered_lost = 0;
10513 	}
10514 	/*
10515 	 * Clear the dup ack count for
10516 	 * the piece that remains.
10517 	 */
10518 	rsm->r_dupack = 0;
10519 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10520 	if (rsm->r_rtr_bytes) {
10521 		/*
10522 		 * It was retransmitted adjust the
10523 		 * sack holes for what was acked.
10524 		 */
10525 		int ack_am;
10526 
10527 		ack_am = (th_ack - rsm->r_start);
10528 		if (ack_am >= rsm->r_rtr_bytes) {
10529 			rack->r_ctl.rc_holes_rxt -= ack_am;
10530 			rsm->r_rtr_bytes -= ack_am;
10531 		}
10532 	}
10533 	/*
10534 	 * Update where the piece starts and record
10535 	 * the time of send of highest cumack sent if
10536 	 * its in our GP range.
10537 	 */
10538 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10539 	/* Now we need to move our offset forward too */
10540 	if (rsm->m &&
10541 	    ((rsm->orig_m_len != rsm->m->m_len) ||
10542 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10543 		/* Fix up the orig_m_len and possibly the mbuf offset */
10544 		rack_adjust_orig_mlen(rsm);
10545 	}
10546 	rsm->soff += (th_ack - rsm->r_start);
10547 	rack_rsm_sender_update(rack, tp, rsm, 5);
10548 	/* The trim will move th_ack into r_start for us */
10549 	tqhash_trim(rack->r_ctl.tqh, th_ack);
10550 	/* Now do we need to move the mbuf fwd too? */
10551 	{
10552 		struct mbuf *m;
10553 		uint32_t soff;
10554 
10555 		m = rsm->m;
10556 		soff = rsm->soff;
10557 		if (m) {
10558 			while (soff >= m->m_len) {
10559 				soff -= m->m_len;
10560 				KASSERT((m->m_next != NULL),
10561 					(" rsm:%p  off:%u soff:%u m:%p",
10562 					 rsm, rsm->soff, soff, m));
10563 				m = m->m_next;
10564 				if (m == NULL) {
10565 					/*
10566 					 * This is a fall-back that prevents a panic. In reality
10567 					 * we should be able to walk the mbuf's and find our place.
10568 					 * At this point snd_una has not been updated with the sbcut() yet
10569 					 * but tqhash_trim did update rsm->r_start so the offset calcuation
10570 					 * should work fine. This is undesirable since we will take cache
10571 					 * hits to access the socket buffer. And even more puzzling is that
10572 					 * it happens occasionally. It should not :(
10573 					 */
10574 					m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
10575 						      (rsm->r_start - tp->snd_una),
10576 						      &soff);
10577 					break;
10578 				}
10579 			}
10580 			/*
10581 			 * Now save in our updated values.
10582 			 */
10583 			rsm->m = m;
10584 			rsm->soff = soff;
10585 			rsm->orig_m_len = rsm->m->m_len;
10586 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10587 		}
10588 	}
10589 	if (rack->app_limited_needs_set &&
10590 	    SEQ_GEQ(th_ack, tp->gput_seq))
10591 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10592 }
10593 
10594 static void
10595 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10596 {
10597 	struct rack_sendmap *rsm;
10598 	int sack_pass_fnd = 0;
10599 
10600 	if (rack->r_might_revert) {
10601 		/*
10602 		 * Ok we have reordering, have not sent anything, we
10603 		 * might want to revert the congestion state if nothing
10604 		 * further has SACK_PASSED on it. Lets check.
10605 		 *
10606 		 * We also get here when we have DSACKs come in for
10607 		 * all the data that we FR'd. Note that a rxt or tlp
10608 		 * timer clears this from happening.
10609 		 */
10610 
10611 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10612 			if (rsm->r_flags & RACK_SACK_PASSED) {
10613 				sack_pass_fnd = 1;
10614 				break;
10615 			}
10616 		}
10617 		if (sack_pass_fnd == 0) {
10618 			/*
10619 			 * We went into recovery
10620 			 * incorrectly due to reordering!
10621 			 */
10622 			int orig_cwnd;
10623 
10624 			rack->r_ent_rec_ns = 0;
10625 			orig_cwnd = tp->snd_cwnd;
10626 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10627 			tp->snd_recover = tp->snd_una;
10628 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10629 			if (IN_RECOVERY(tp->t_flags)) {
10630 				rack_exit_recovery(tp, rack, 3);
10631 				if ((rack->rto_from_rec == 1) && (rack_ssthresh_rest_rto_rec != 0) ){
10632 					/*
10633 					 * We were in recovery, had an RTO
10634 					 * and then re-entered recovery (more sack's arrived)
10635 					 * and we have properly recorded the old ssthresh from
10636 					 * the first recovery. We want to be able to slow-start
10637 					 * back to this level. The ssthresh from the timeout
10638 					 * and then back into recovery will end up most likely
10639 					 * to be min(cwnd=1mss, 2mss). Which makes it basically
10640 					 * so we get no slow-start after our RTO.
10641 					 */
10642 					rack->rto_from_rec = 0;
10643 					if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
10644 						tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
10645 				}
10646 			}
10647 		}
10648 		rack->r_might_revert = 0;
10649 	}
10650 }
10651 
10652 
10653 static int
10654 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10655 {
10656 
10657 	uint32_t am, l_end;
10658 	int was_tlp = 0;
10659 
10660 	if (SEQ_GT(end, start))
10661 		am = end - start;
10662 	else
10663 		am = 0;
10664 	if ((rack->rc_last_tlp_acked_set ) &&
10665 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
10666 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
10667 		/*
10668 		 * The DSACK is because of a TLP which we don't
10669 		 * do anything with the reordering window over since
10670 		 * it was not reordering that caused the DSACK but
10671 		 * our previous retransmit TLP.
10672 		 */
10673 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
10674 		was_tlp = 1;
10675 		goto skip_dsack_round;
10676 	}
10677 	if (rack->rc_last_sent_tlp_seq_valid) {
10678 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
10679 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
10680 		    (SEQ_LEQ(end, l_end))) {
10681 			/*
10682 			 * This dsack is from the last sent TLP, ignore it
10683 			 * for reordering purposes.
10684 			 */
10685 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
10686 			was_tlp = 1;
10687 			goto skip_dsack_round;
10688 		}
10689 	}
10690 	if (rack->rc_dsack_round_seen == 0) {
10691 		rack->rc_dsack_round_seen = 1;
10692 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
10693 		rack->r_ctl.num_dsack++;
10694 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
10695 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
10696 	}
10697 skip_dsack_round:
10698 	/*
10699 	 * We keep track of how many DSACK blocks we get
10700 	 * after a recovery incident.
10701 	 */
10702 	rack->r_ctl.dsack_byte_cnt += am;
10703 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
10704 	    rack->r_ctl.retran_during_recovery &&
10705 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
10706 		/*
10707 		 * False recovery most likely culprit is reordering. If
10708 		 * nothing else is missing we need to revert.
10709 		 */
10710 		rack->r_might_revert = 1;
10711 		rack_handle_might_revert(rack->rc_tp, rack);
10712 		rack->r_might_revert = 0;
10713 		rack->r_ctl.retran_during_recovery = 0;
10714 		rack->r_ctl.dsack_byte_cnt = 0;
10715 	}
10716 	return (was_tlp);
10717 }
10718 
10719 static uint32_t
10720 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
10721 {
10722 	return (((tp->snd_max - snd_una) -
10723 		 (rack->r_ctl.rc_sacked + rack->r_ctl.rc_considered_lost)) + rack->r_ctl.rc_holes_rxt);
10724 }
10725 
10726 static int32_t
10727 rack_compute_pipe(struct tcpcb *tp)
10728 {
10729 	return ((int32_t)do_rack_compute_pipe(tp,
10730 					      (struct tcp_rack *)tp->t_fb_ptr,
10731 					      tp->snd_una));
10732 }
10733 
10734 static void
10735 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
10736 {
10737 	/* Deal with changed and PRR here (in recovery only) */
10738 	uint32_t pipe, snd_una;
10739 
10740 	rack->r_ctl.rc_prr_delivered += changed;
10741 
10742 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
10743 		/*
10744 		 * It is all outstanding, we are application limited
10745 		 * and thus we don't need more room to send anything.
10746 		 * Note we use tp->snd_una here and not th_ack because
10747 		 * the data as yet not been cut from the sb.
10748 		 */
10749 		rack->r_ctl.rc_prr_sndcnt = 0;
10750 		return;
10751 	}
10752 	/* Compute prr_sndcnt */
10753 	if (SEQ_GT(tp->snd_una, th_ack)) {
10754 		snd_una = tp->snd_una;
10755 	} else {
10756 		snd_una = th_ack;
10757 	}
10758 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
10759 	if (pipe > tp->snd_ssthresh) {
10760 		long sndcnt;
10761 
10762 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
10763 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
10764 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
10765 		else {
10766 			rack->r_ctl.rc_prr_sndcnt = 0;
10767 			rack_log_to_prr(rack, 9, 0, __LINE__);
10768 			sndcnt = 0;
10769 		}
10770 		sndcnt++;
10771 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
10772 			sndcnt -= rack->r_ctl.rc_prr_out;
10773 		else
10774 			sndcnt = 0;
10775 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
10776 		rack_log_to_prr(rack, 10, 0, __LINE__);
10777 	} else {
10778 		uint32_t limit;
10779 
10780 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
10781 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
10782 		else
10783 			limit = 0;
10784 		if (changed > limit)
10785 			limit = changed;
10786 		limit += ctf_fixed_maxseg(tp);
10787 		if (tp->snd_ssthresh > pipe) {
10788 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
10789 			rack_log_to_prr(rack, 11, 0, __LINE__);
10790 		} else {
10791 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
10792 			rack_log_to_prr(rack, 12, 0, __LINE__);
10793 		}
10794 	}
10795 }
10796 
10797 static void
10798 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
10799 	     int *dsack_seen, int *sacks_seen)
10800 {
10801 	uint32_t changed;
10802 	struct tcp_rack *rack;
10803 	struct rack_sendmap *rsm;
10804 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
10805 	register uint32_t th_ack;
10806 	int32_t i, j, k, num_sack_blks = 0;
10807 	uint32_t cts, acked, ack_point;
10808 	int loop_start = 0;
10809 	uint32_t tsused;
10810 	uint32_t segsiz;
10811 
10812 
10813 	INP_WLOCK_ASSERT(tptoinpcb(tp));
10814 	if (tcp_get_flags(th) & TH_RST) {
10815 		/* We don't log resets */
10816 		return;
10817 	}
10818 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10819 	cts = tcp_get_usecs(NULL);
10820 	rsm = tqhash_min(rack->r_ctl.tqh);
10821 	changed = 0;
10822 	th_ack = th->th_ack;
10823 	segsiz = ctf_fixed_maxseg(rack->rc_tp);
10824 	if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
10825 		/*
10826 		 * You only get credit for
10827 		 * MSS and greater (and you get extra
10828 		 * credit for larger cum-ack moves).
10829 		 */
10830 		int ac;
10831 
10832 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
10833 		counter_u64_add(rack_ack_total, ac);
10834 	}
10835 	if (SEQ_GT(th_ack, tp->snd_una)) {
10836 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
10837 		tp->t_acktime = ticks;
10838 	}
10839 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
10840 		changed = th_ack - rsm->r_start;
10841 	if (changed) {
10842 		rack_process_to_cumack(tp, rack, th_ack, cts, to,
10843 				       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
10844 	}
10845 	if ((to->to_flags & TOF_SACK) == 0) {
10846 		/* We are done nothing left and no sack. */
10847 		rack_handle_might_revert(tp, rack);
10848 		/*
10849 		 * For cases where we struck a dup-ack
10850 		 * with no SACK, add to the changes so
10851 		 * PRR will work right.
10852 		 */
10853 		if (dup_ack_struck && (changed == 0)) {
10854 			changed += ctf_fixed_maxseg(rack->rc_tp);
10855 		}
10856 		goto out;
10857 	}
10858 	/* Sack block processing */
10859 	if (SEQ_GT(th_ack, tp->snd_una))
10860 		ack_point = th_ack;
10861 	else
10862 		ack_point = tp->snd_una;
10863 	for (i = 0; i < to->to_nsacks; i++) {
10864 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
10865 		      &sack, sizeof(sack));
10866 		sack.start = ntohl(sack.start);
10867 		sack.end = ntohl(sack.end);
10868 		if (SEQ_GT(sack.end, sack.start) &&
10869 		    SEQ_GT(sack.start, ack_point) &&
10870 		    SEQ_LT(sack.start, tp->snd_max) &&
10871 		    SEQ_GT(sack.end, ack_point) &&
10872 		    SEQ_LEQ(sack.end, tp->snd_max)) {
10873 			sack_blocks[num_sack_blks] = sack;
10874 			num_sack_blks++;
10875 		} else if (SEQ_LEQ(sack.start, th_ack) &&
10876 			   SEQ_LEQ(sack.end, th_ack)) {
10877 			int was_tlp;
10878 
10879 			if (dsack_seen != NULL)
10880 				*dsack_seen = 1;
10881 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
10882 			/*
10883 			 * Its a D-SACK block.
10884 			 */
10885 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
10886 		}
10887 	}
10888 	if (rack->rc_dsack_round_seen) {
10889 		/* Is the dsack roound over? */
10890 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
10891 			/* Yes it is */
10892 			rack->rc_dsack_round_seen = 0;
10893 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
10894 		}
10895 	}
10896 	/*
10897 	 * Sort the SACK blocks so we can update the rack scoreboard with
10898 	 * just one pass.
10899 	 */
10900 	num_sack_blks = sack_filter_blks(tp, &rack->r_ctl.rack_sf, sack_blocks,
10901 					 num_sack_blks, th->th_ack);
10902 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
10903 	if (sacks_seen != NULL)
10904 		*sacks_seen = num_sack_blks;
10905 	if (num_sack_blks == 0) {
10906 		/* Nothing to sack, but we need to update counts */
10907 		goto out_with_totals;
10908 	}
10909 	/* Its a sack of some sort */
10910 	if (num_sack_blks < 2) {
10911 		/* Only one, we don't need to sort */
10912 		goto do_sack_work;
10913 	}
10914 	/* Sort the sacks */
10915 	for (i = 0; i < num_sack_blks; i++) {
10916 		for (j = i + 1; j < num_sack_blks; j++) {
10917 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
10918 				sack = sack_blocks[i];
10919 				sack_blocks[i] = sack_blocks[j];
10920 				sack_blocks[j] = sack;
10921 			}
10922 		}
10923 	}
10924 	/*
10925 	 * Now are any of the sack block ends the same (yes some
10926 	 * implementations send these)?
10927 	 */
10928 again:
10929 	if (num_sack_blks == 0)
10930 		goto out_with_totals;
10931 	if (num_sack_blks > 1) {
10932 		for (i = 0; i < num_sack_blks; i++) {
10933 			for (j = i + 1; j < num_sack_blks; j++) {
10934 				if (sack_blocks[i].end == sack_blocks[j].end) {
10935 					/*
10936 					 * Ok these two have the same end we
10937 					 * want the smallest end and then
10938 					 * throw away the larger and start
10939 					 * again.
10940 					 */
10941 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
10942 						/*
10943 						 * The second block covers
10944 						 * more area use that
10945 						 */
10946 						sack_blocks[i].start = sack_blocks[j].start;
10947 					}
10948 					/*
10949 					 * Now collapse out the dup-sack and
10950 					 * lower the count
10951 					 */
10952 					for (k = (j + 1); k < num_sack_blks; k++) {
10953 						sack_blocks[j].start = sack_blocks[k].start;
10954 						sack_blocks[j].end = sack_blocks[k].end;
10955 						j++;
10956 					}
10957 					num_sack_blks--;
10958 					goto again;
10959 				}
10960 			}
10961 		}
10962 	}
10963 do_sack_work:
10964 	/*
10965 	 * First lets look to see if
10966 	 * we have retransmitted and
10967 	 * can use the transmit next?
10968 	 */
10969 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
10970 	if (rsm &&
10971 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
10972 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
10973 		/*
10974 		 * We probably did the FR and the next
10975 		 * SACK in continues as we would expect.
10976 		 */
10977 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, segsiz);
10978 		if (acked) {
10979 			rack->r_wanted_output = 1;
10980 			changed += acked;
10981 		}
10982 		if (num_sack_blks == 1) {
10983 			/*
10984 			 * This is what we would expect from
10985 			 * a normal implementation to happen
10986 			 * after we have retransmitted the FR,
10987 			 * i.e the sack-filter pushes down
10988 			 * to 1 block and the next to be retransmitted
10989 			 * is the sequence in the sack block (has more
10990 			 * are acked). Count this as ACK'd data to boost
10991 			 * up the chances of recovering any false positives.
10992 			 */
10993 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
10994 			counter_u64_add(rack_express_sack, 1);
10995 			goto out_with_totals;
10996 		} else {
10997 			/*
10998 			 * Start the loop through the
10999 			 * rest of blocks, past the first block.
11000 			 */
11001 			loop_start = 1;
11002 		}
11003 	}
11004 	counter_u64_add(rack_sack_total, 1);
11005 	rsm = rack->r_ctl.rc_sacklast;
11006 	for (i = loop_start; i < num_sack_blks; i++) {
11007 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts,  segsiz);
11008 		if (acked) {
11009 			rack->r_wanted_output = 1;
11010 			changed += acked;
11011 		}
11012 	}
11013 out_with_totals:
11014 	if (num_sack_blks > 1) {
11015 		/*
11016 		 * You get an extra stroke if
11017 		 * you have more than one sack-blk, this
11018 		 * could be where we are skipping forward
11019 		 * and the sack-filter is still working, or
11020 		 * it could be an attacker constantly
11021 		 * moving us.
11022 		 */
11023 		counter_u64_add(rack_move_some, 1);
11024 	}
11025 out:
11026 	if (changed) {
11027 		/* Something changed cancel the rack timer */
11028 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11029 	}
11030 	tsused = tcp_get_usecs(NULL);
11031 	rsm = tcp_rack_output(tp, rack, tsused);
11032 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11033 	    rsm &&
11034 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11035 		/* Enter recovery */
11036 		entered_recovery = 1;
11037 		rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
11038 		/*
11039 		 * When we enter recovery we need to assure we send
11040 		 * one packet.
11041 		 */
11042 		if (rack->rack_no_prr == 0) {
11043 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11044 			rack_log_to_prr(rack, 8, 0, __LINE__);
11045 		}
11046 		rack->r_timer_override = 1;
11047 		rack->r_early = 0;
11048 		rack->r_ctl.rc_agg_early = 0;
11049 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
11050 		   rsm &&
11051 		   (rack->r_rr_config == 3)) {
11052 		/*
11053 		 * Assure we can output and we get no
11054 		 * remembered pace time except the retransmit.
11055 		 */
11056 		rack->r_timer_override = 1;
11057 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11058 		rack->r_ctl.rc_resend = rsm;
11059 	}
11060 	if (IN_FASTRECOVERY(tp->t_flags) &&
11061 	    (rack->rack_no_prr == 0) &&
11062 	    (entered_recovery == 0)) {
11063 		rack_update_prr(tp, rack, changed, th_ack);
11064 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11065 		     ((tcp_in_hpts(rack->rc_tp) == 0) &&
11066 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11067 			/*
11068 			 * If you are pacing output you don't want
11069 			 * to override.
11070 			 */
11071 			rack->r_early = 0;
11072 			rack->r_ctl.rc_agg_early = 0;
11073 			rack->r_timer_override = 1;
11074 		}
11075 	}
11076 }
11077 
11078 static void
11079 rack_strike_dupack(struct tcp_rack *rack, tcp_seq th_ack)
11080 {
11081 	struct rack_sendmap *rsm;
11082 
11083 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11084 	while (rsm) {
11085 		/*
11086 		 * We need to skip anything already set
11087 		 * to be retransmitted.
11088 		 */
11089 		if ((rsm->r_dupack >= DUP_ACK_THRESHOLD)  ||
11090 		    (rsm->r_flags & RACK_MUST_RXT)) {
11091 			rsm = TAILQ_NEXT(rsm, r_tnext);
11092 			continue;
11093 		}
11094 		break;
11095 	}
11096 	if (rsm && (rsm->r_dupack < 0xff)) {
11097 		rsm->r_dupack++;
11098 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11099 			struct timeval tv;
11100 			uint32_t cts;
11101 			/*
11102 			 * Here we see if we need to retransmit. For
11103 			 * a SACK type connection if enough time has passed
11104 			 * we will get a return of the rsm. For a non-sack
11105 			 * connection we will get the rsm returned if the
11106 			 * dupack value is 3 or more.
11107 			 */
11108 			cts = tcp_get_usecs(&tv);
11109 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11110 			if (rack->r_ctl.rc_resend != NULL) {
11111 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11112 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11113 							 th_ack,  __LINE__);
11114 				}
11115 				rack->r_wanted_output = 1;
11116 				rack->r_timer_override = 1;
11117 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11118 			}
11119 		} else {
11120 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11121 		}
11122 	}
11123 }
11124 
11125 static void
11126 rack_check_bottom_drag(struct tcpcb *tp,
11127 		       struct tcp_rack *rack,
11128 		       struct socket *so)
11129 {
11130 	/*
11131 	 * So what is dragging bottom?
11132 	 *
11133 	 * Dragging bottom means you were under pacing and had a
11134 	 * delay in processing inbound acks waiting on our pacing
11135 	 * timer to expire. While you were waiting all of the acknowledgments
11136 	 * for the packets you sent have arrived. This means we are pacing
11137 	 * way underneath the bottleneck to the point where our Goodput
11138 	 * measurements stop working, since they require more than one
11139 	 * ack (usually at least 8 packets worth with multiple acks so we can
11140 	 * gauge the inter-ack times). If that occurs we have a real problem
11141 	 * since we are stuck in a hole that we can't get out of without
11142 	 * something speeding us up.
11143 	 *
11144 	 * We also check to see if we are widdling down to just one segment
11145 	 * outstanding. If this occurs and we have room to send in our cwnd/rwnd
11146 	 * then we are adding the delayed ack interval into our measurments and
11147 	 * we need to speed up slightly.
11148 	 */
11149 	uint32_t segsiz, minseg;
11150 
11151 	segsiz = ctf_fixed_maxseg(tp);
11152 	minseg = segsiz;
11153 	if (tp->snd_max == tp->snd_una) {
11154 		/*
11155 		 * We are doing dynamic pacing and we are way
11156 		 * under. Basically everything got acked while
11157 		 * we were still waiting on the pacer to expire.
11158 		 *
11159 		 * This means we need to boost the b/w in
11160 		 * addition to any earlier boosting of
11161 		 * the multiplier.
11162 		 */
11163 		uint64_t lt_bw;
11164 
11165 		tcp_trace_point(rack->rc_tp, TCP_TP_PACED_BOTTOM);
11166 		lt_bw = rack_get_lt_bw(rack);
11167 		rack->rc_dragged_bottom = 1;
11168 		rack_validate_multipliers_at_or_above100(rack);
11169 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11170 		    (rack->dis_lt_bw == 0) &&
11171 		    (rack->use_lesser_lt_bw == 0) &&
11172 		    (lt_bw > 0)) {
11173 			/*
11174 			 * Lets use the long-term b/w we have
11175 			 * been getting as a base.
11176 			 */
11177 			if (rack->rc_gp_filled == 0) {
11178 				if (lt_bw > ONE_POINT_TWO_MEG) {
11179 					/*
11180 					 * If we have no measurement
11181 					 * don't let us set in more than
11182 					 * 1.2Mbps. If we are still too
11183 					 * low after pacing with this we
11184 					 * will hopefully have a max b/w
11185 					 * available to sanity check things.
11186 					 */
11187 					lt_bw = ONE_POINT_TWO_MEG;
11188 				}
11189 				rack->r_ctl.rc_rtt_diff = 0;
11190 				rack->r_ctl.gp_bw = lt_bw;
11191 				rack->rc_gp_filled = 1;
11192 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11193 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11194 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11195 			} else if (lt_bw > rack->r_ctl.gp_bw) {
11196 				rack->r_ctl.rc_rtt_diff = 0;
11197 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11198 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11199 				rack->r_ctl.gp_bw = lt_bw;
11200 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11201 			} else
11202 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
11203 			if ((rack->gp_ready == 0) &&
11204 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11205 				/* We have enough measurements now */
11206 				rack->gp_ready = 1;
11207 				if (rack->dgp_on ||
11208 				    rack->rack_hibeta)
11209 					rack_set_cc_pacing(rack);
11210 				if (rack->defer_options)
11211 					rack_apply_deferred_options(rack);
11212 			}
11213 		} else {
11214 			/*
11215 			 * zero rtt possibly?, settle for just an old increase.
11216 			 */
11217 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
11218 		}
11219 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11220 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11221 					       minseg)) &&
11222 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11223 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11224 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11225 		    (segsiz * rack_req_segs))) {
11226 		/*
11227 		 * We are doing dynamic GP pacing and
11228 		 * we have everything except 1MSS or less
11229 		 * bytes left out. We are still pacing away.
11230 		 * And there is data that could be sent, This
11231 		 * means we are inserting delayed ack time in
11232 		 * our measurements because we are pacing too slow.
11233 		 */
11234 		rack_validate_multipliers_at_or_above100(rack);
11235 		rack->rc_dragged_bottom = 1;
11236 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
11237 	}
11238 }
11239 
11240 #ifdef TCP_REQUEST_TRK
11241 static void
11242 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11243 		struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11244 {
11245 	int do_log;
11246 
11247 	do_log = tcp_bblogging_on(rack->rc_tp);
11248 	if (do_log == 0) {
11249 		if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11250 			return;
11251 		/* We only allow the three below with point logging on */
11252 		if ((mod != HYBRID_LOG_RULES_APP) &&
11253 		    (mod != HYBRID_LOG_RULES_SET) &&
11254 		    (mod != HYBRID_LOG_REQ_COMP))
11255 			return;
11256 
11257 	}
11258 	if (do_log) {
11259 		union tcp_log_stackspecific log;
11260 		struct timeval tv;
11261 
11262 		/* Convert our ms to a microsecond */
11263 		memset(&log, 0, sizeof(log));
11264 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11265 		log.u_bbr.flex1 = seq;
11266 		log.u_bbr.cwnd_gain = line;
11267 		if (cur != NULL) {
11268 			uint64_t off;
11269 
11270 			log.u_bbr.flex2 = cur->start_seq;
11271 			log.u_bbr.flex3 = cur->end_seq;
11272 			log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11273 			log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11274 			log.u_bbr.flex6 = cur->flags;
11275 			log.u_bbr.pkts_out = cur->hybrid_flags;
11276 			log.u_bbr.rttProp = cur->timestamp;
11277 			log.u_bbr.cur_del_rate = cur->cspr;
11278 			log.u_bbr.bw_inuse = cur->start;
11279 			log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11280 			log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11281 			log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11282 			log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11283 			log.u_bbr.inhpts = 1;
11284 #ifdef TCP_REQUEST_TRK
11285 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11286 			log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11287 #endif
11288 		} else {
11289 			log.u_bbr.flex2 = err;
11290 		}
11291 		/*
11292 		 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11293 		 */
11294 		log.u_bbr.flex7 = rack->rc_catch_up;
11295 		log.u_bbr.flex7 <<= 1;
11296 		log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11297 		log.u_bbr.flex7 <<= 1;
11298 		log.u_bbr.flex7 |= rack->dgp_on;
11299 		/*
11300 		 * Compose bbr_state to be a bit wise 0000ADHF
11301 		 * where A is the always_pace flag
11302 		 * where D is the dgp_on flag
11303 		 * where H is the hybrid_mode on flag
11304 		 * where F is the use_fixed_rate flag.
11305 		 */
11306 		log.u_bbr.bbr_state = rack->rc_always_pace;
11307 		log.u_bbr.bbr_state <<= 1;
11308 		log.u_bbr.bbr_state |= rack->dgp_on;
11309 		log.u_bbr.bbr_state <<= 1;
11310 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
11311 		log.u_bbr.bbr_state <<= 1;
11312 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
11313 		log.u_bbr.flex8 = mod;
11314 		log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11315 		log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11316 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11317 		log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11318 		log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11319 		log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11320 		tcp_log_event(rack->rc_tp, NULL,
11321 		    &rack->rc_inp->inp_socket->so_rcv,
11322 		    &rack->rc_inp->inp_socket->so_snd,
11323 		    TCP_HYBRID_PACING_LOG, 0,
11324 	            0, &log, false, NULL, __func__, __LINE__, &tv);
11325 	}
11326 }
11327 #endif
11328 
11329 #ifdef TCP_REQUEST_TRK
11330 static void
11331 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11332 {
11333 	struct tcp_sendfile_track *rc_cur, *orig_ent;
11334 	struct tcpcb *tp;
11335 	int err = 0;
11336 
11337 	orig_ent = rack->r_ctl.rc_last_sft;
11338 	rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
11339 	if (rc_cur == NULL) {
11340 		/* If not in the beginning what about the end piece */
11341 		if (rack->rc_hybrid_mode)
11342 			rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11343 		rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11344 	} else {
11345 		err = 12345;
11346 	}
11347 	/* If we find no parameters we are in straight DGP mode */
11348 	if(rc_cur == NULL) {
11349 		/* None found for this seq, just DGP for now */
11350 		if (rack->rc_hybrid_mode) {
11351 			rack->r_ctl.client_suggested_maxseg = 0;
11352 			rack->rc_catch_up = 0;
11353 			if (rack->cspr_is_fcc == 0)
11354 				rack->r_ctl.bw_rate_cap = 0;
11355 			else
11356 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11357 		}
11358 		if (rack->rc_hybrid_mode) {
11359 			rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11360 		}
11361 		if (rack->r_ctl.rc_last_sft) {
11362 			rack->r_ctl.rc_last_sft = NULL;
11363 		}
11364 		return;
11365 	}
11366 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
11367 		/* This entry was never setup for hybrid pacing on/off etc */
11368 		if (rack->rc_hybrid_mode) {
11369 			rack->r_ctl.client_suggested_maxseg = 0;
11370 			rack->rc_catch_up = 0;
11371 			rack->r_ctl.bw_rate_cap = 0;
11372 		}
11373 		if (rack->r_ctl.rc_last_sft) {
11374 			rack->r_ctl.rc_last_sft = NULL;
11375 		}
11376 		if ((rc_cur->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11377 			rc_cur->flags |= TCP_TRK_TRACK_FLG_FSND;
11378 			rc_cur->first_send = cts;
11379 			rc_cur->sent_at_fs = rack->rc_tp->t_sndbytes;
11380 			rc_cur->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11381 		}
11382 		return;
11383 	}
11384 	/*
11385 	 * Ok if we have a new entry *or* have never
11386 	 * set up an entry we need to proceed. If
11387 	 * we have already set it up this entry we
11388 	 * just continue along with what we already
11389 	 * setup.
11390 	 */
11391 	tp = rack->rc_tp;
11392 	if ((rack->r_ctl.rc_last_sft != NULL) &&
11393 	    (rack->r_ctl.rc_last_sft == rc_cur)) {
11394 		/* Its already in place */
11395 		if (rack->rc_hybrid_mode)
11396 			rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11397 		return;
11398 	}
11399 	if (rack->rc_hybrid_mode == 0) {
11400 		rack->r_ctl.rc_last_sft = rc_cur;
11401 		if (orig_ent) {
11402 			orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11403 			orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11404 			orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11405 		}
11406 		rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11407 		return;
11408 	}
11409 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11410 		/* Compensate for all the header overhead's */
11411 		if (rack->cspr_is_fcc == 0)
11412 			rack->r_ctl.bw_rate_cap	= rack_compensate_for_linerate(rack, rc_cur->cspr);
11413 		else
11414 			rack->r_ctl.fillcw_cap =  rack_compensate_for_linerate(rack, rc_cur->cspr);
11415 	} else {
11416 		if (rack->rc_hybrid_mode) {
11417 			if (rack->cspr_is_fcc == 0)
11418 				rack->r_ctl.bw_rate_cap = 0;
11419 			else
11420 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11421 		}
11422 	}
11423 	if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11424 		rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11425 	else
11426 		rack->r_ctl.client_suggested_maxseg = 0;
11427 	if (rc_cur->timestamp == rack->r_ctl.last_tm_mark) {
11428 		/*
11429 		 * It is the same timestamp as the previous one
11430 		 * add the hybrid flag that will indicate we use
11431 		 * sendtime not arrival time for catch-up mode.
11432 		 */
11433 		rc_cur->hybrid_flags |= TCP_HYBRID_PACING_SENDTIME;
11434 	}
11435 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11436 	    (rc_cur->cspr > 0)) {
11437 		uint64_t len;
11438 
11439 		rack->rc_catch_up = 1;
11440 		/*
11441 		 * Calculate the deadline time, first set the
11442 		 * time to when the request arrived.
11443 		 */
11444 		if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_SENDTIME) {
11445 			/*
11446 			 * For cases where its a duplicate tm (we received more
11447 			 * than one request for a tm) we want to use now, the point
11448 			 * where we are just sending the first bit of the request.
11449 			 */
11450 			rc_cur->deadline = cts;
11451 		} else {
11452 			/*
11453 			 * Here we have a different tm from the last request
11454 			 * so we want to use arrival time as our base.
11455 			 */
11456 			rc_cur->deadline = rc_cur->localtime;
11457 		}
11458 		/*
11459 		 * Next calculate the length and compensate for
11460 		 * TLS if need be.
11461 		 */
11462 		len = rc_cur->end - rc_cur->start;
11463 		if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11464 			/*
11465 			 * This session is doing TLS. Take a swag guess
11466 			 * at the overhead.
11467 			 */
11468 			len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11469 		}
11470 		/*
11471 		 * Now considering the size, and the cspr, what is the time that
11472 		 * would be required at the cspr rate. Here we use the raw
11473 		 * cspr value since the client only looks at the raw data. We
11474 		 * do use len which includes TLS overhead, but not the TCP/IP etc.
11475 		 * That will get made up for in the CU pacing rate set.
11476 		 */
11477 		len *= HPTS_USEC_IN_SEC;
11478 		len /= rc_cur->cspr;
11479 		rc_cur->deadline += len;
11480 	} else {
11481 		rack->rc_catch_up = 0;
11482 		rc_cur->deadline = 0;
11483 	}
11484 	if (rack->r_ctl.client_suggested_maxseg != 0) {
11485 		/*
11486 		 * We need to reset the max pace segs if we have a
11487 		 * client_suggested_maxseg.
11488 		 */
11489 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
11490 	}
11491 	if (orig_ent) {
11492 		orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11493 		orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11494 		orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11495 	}
11496 	rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11497 	/* Remember it for next time and for CU mode */
11498 	rack->r_ctl.rc_last_sft = rc_cur;
11499 	rack->r_ctl.last_tm_mark = rc_cur->timestamp;
11500 }
11501 #endif
11502 
11503 static void
11504 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11505 {
11506 #ifdef TCP_REQUEST_TRK
11507 	struct tcp_sendfile_track *ent;
11508 
11509 	ent = rack->r_ctl.rc_last_sft;
11510 	if ((ent == NULL) ||
11511 	    (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
11512 	    (SEQ_GEQ(seq, ent->end_seq))) {
11513 		/* Time to update the track. */
11514 		rack_set_dgp_hybrid_mode(rack, seq, len, cts);
11515 		ent = rack->r_ctl.rc_last_sft;
11516 	}
11517 	/* Out of all */
11518 	if (ent == NULL) {
11519 		return;
11520 	}
11521 	if (SEQ_LT(ent->end_seq, (seq + len))) {
11522 		/*
11523 		 * This is the case where our end_seq guess
11524 		 * was wrong. This is usually due to TLS having
11525 		 * more bytes then our guess. It could also be the
11526 		 * case that the client sent in two requests closely
11527 		 * and the SB is full of both so we are sending part
11528 		 * of each (end|beg). In such a case lets move this
11529 		 * guys end to match the end of this send. That
11530 		 * way it will complete when all of it is acked.
11531 		 */
11532 		ent->end_seq = (seq + len);
11533 		if (rack->rc_hybrid_mode)
11534 			rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
11535 	}
11536 	/* Now validate we have set the send time of this one */
11537 	if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11538 		ent->flags |= TCP_TRK_TRACK_FLG_FSND;
11539 		ent->first_send = cts;
11540 		ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11541 		ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11542 	}
11543 #endif
11544 }
11545 
11546 static void
11547 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11548 {
11549 	/*
11550 	 * The fast output path is enabled and we
11551 	 * have moved the cumack forward. Lets see if
11552 	 * we can expand forward the fast path length by
11553 	 * that amount. What we would ideally like to
11554 	 * do is increase the number of bytes in the
11555 	 * fast path block (left_to_send) by the
11556 	 * acked amount. However we have to gate that
11557 	 * by two factors:
11558 	 * 1) The amount outstanding and the rwnd of the peer
11559 	 *    (i.e. we don't want to exceed the rwnd of the peer).
11560 	 *    <and>
11561 	 * 2) The amount of data left in the socket buffer (i.e.
11562 	 *    we can't send beyond what is in the buffer).
11563 	 *
11564 	 * Note that this does not take into account any increase
11565 	 * in the cwnd. We will only extend the fast path by
11566 	 * what was acked.
11567 	 */
11568 	uint32_t new_total, gating_val;
11569 
11570 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11571 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11572 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11573 	if (new_total <= gating_val) {
11574 		/* We can increase left_to_send by the acked amount */
11575 		counter_u64_add(rack_extended_rfo, 1);
11576 		rack->r_ctl.fsb.left_to_send = new_total;
11577 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11578 			("rack:%p left_to_send:%u sbavail:%u out:%u",
11579 			 rack, rack->r_ctl.fsb.left_to_send,
11580 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
11581 			 (tp->snd_max - tp->snd_una)));
11582 
11583 	}
11584 }
11585 
11586 static void
11587 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11588 {
11589 	/*
11590 	 * Here any sendmap entry that points to the
11591 	 * beginning mbuf must be adjusted to the correct
11592 	 * offset. This must be called with:
11593 	 * 1) The socket buffer locked
11594 	 * 2) snd_una adjusted to its new position.
11595 	 *
11596 	 * Note that (2) implies rack_ack_received has also
11597 	 * been called and all the sbcut's have been done.
11598 	 *
11599 	 * We grab the first mbuf in the socket buffer and
11600 	 * then go through the front of the sendmap, recalculating
11601 	 * the stored offset for any sendmap entry that has
11602 	 * that mbuf. We must use the sb functions to do this
11603 	 * since its possible an add was done has well as
11604 	 * the subtraction we may have just completed. This should
11605 	 * not be a penalty though, since we just referenced the sb
11606 	 * to go in and trim off the mbufs that we freed (of course
11607 	 * there will be a penalty for the sendmap references though).
11608 	 *
11609 	 * Note also with INVARIANT on, we validate with a KASSERT
11610 	 * that the first sendmap entry has a soff of 0.
11611 	 *
11612 	 */
11613 	struct mbuf *m;
11614 	struct rack_sendmap *rsm;
11615 	tcp_seq snd_una;
11616 #ifdef INVARIANTS
11617 	int first_processed = 0;
11618 #endif
11619 
11620 	snd_una = rack->rc_tp->snd_una;
11621 	SOCKBUF_LOCK_ASSERT(sb);
11622 	m = sb->sb_mb;
11623 	rsm = tqhash_min(rack->r_ctl.tqh);
11624 	if ((rsm == NULL) || (m == NULL)) {
11625 		/* Nothing outstanding */
11626 		return;
11627 	}
11628 	/* The very first RSM's mbuf must point to the head mbuf in the sb */
11629 	KASSERT((rsm->m == m),
11630 		("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
11631 		 rack, sb, rsm));
11632 	while (rsm->m && (rsm->m == m)) {
11633 		/* one to adjust */
11634 #ifdef INVARIANTS
11635 		struct mbuf *tm;
11636 		uint32_t soff;
11637 
11638 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
11639 		if ((rsm->orig_m_len != m->m_len) ||
11640 		    (rsm->orig_t_space != M_TRAILINGROOM(m))){
11641 			rack_adjust_orig_mlen(rsm);
11642 		}
11643 		if (first_processed == 0) {
11644 			KASSERT((rsm->soff == 0),
11645 				("Rack:%p rsm:%p -- rsm at head but soff not zero",
11646 				 rack, rsm));
11647 			first_processed = 1;
11648 		}
11649 		if ((rsm->soff != soff) || (rsm->m != tm)) {
11650 			/*
11651 			 * This is not a fatal error, we anticipate it
11652 			 * might happen (the else code), so we count it here
11653 			 * so that under invariant we can see that it really
11654 			 * does happen.
11655 			 */
11656 			counter_u64_add(rack_adjust_map_bw, 1);
11657 		}
11658 		rsm->m = tm;
11659 		rsm->soff = soff;
11660 		if (tm) {
11661 			rsm->orig_m_len = rsm->m->m_len;
11662 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11663 		} else {
11664 			rsm->orig_m_len = 0;
11665 			rsm->orig_t_space = 0;
11666 		}
11667 #else
11668 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
11669 		if (rsm->m) {
11670 			rsm->orig_m_len = rsm->m->m_len;
11671 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11672 		} else {
11673 			rsm->orig_m_len = 0;
11674 			rsm->orig_t_space = 0;
11675 		}
11676 #endif
11677 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
11678 		if (rsm == NULL)
11679 			break;
11680 	}
11681 }
11682 
11683 #ifdef TCP_REQUEST_TRK
11684 static inline void
11685 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
11686 {
11687 	struct tcp_sendfile_track *ent;
11688 	int i;
11689 
11690 	if ((rack->rc_hybrid_mode == 0) &&
11691 	    (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
11692 		/*
11693 		 * Just do normal completions hybrid pacing is not on
11694 		 * and CLDL is off as well.
11695 		 */
11696 		tcp_req_check_for_comp(rack->rc_tp, th_ack);
11697 		return;
11698 	}
11699 	/*
11700 	 * Originally I was just going to find the th_ack associated
11701 	 * with an entry. But then I realized a large strech ack could
11702 	 * in theory ack two or more requests at once. So instead we
11703 	 * need to find all entries that are completed by th_ack not
11704 	 * just a single entry and do our logging.
11705 	 */
11706 	ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11707 	while (ent != NULL) {
11708 		/*
11709 		 * We may be doing hybrid pacing or CLDL and need more details possibly
11710 		 * so we do it manually instead of calling
11711 		 * tcp_req_check_for_comp()
11712 		 */
11713 		uint64_t laa, tim, data, cbw, ftim;
11714 
11715 		/* Ok this ack frees it */
11716 		rack_log_hybrid(rack, th_ack,
11717 				ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
11718 		rack_log_hybrid_sends(rack, ent, __LINE__);
11719 		/* calculate the time based on the ack arrival */
11720 		data = ent->end - ent->start;
11721 		laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
11722 		if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
11723 			if (ent->first_send > ent->localtime)
11724 				ftim = ent->first_send;
11725 			else
11726 				ftim = ent->localtime;
11727 		} else {
11728 			/* TSNH */
11729 			ftim = ent->localtime;
11730 		}
11731 		if (laa > ent->localtime)
11732 			tim = laa - ftim;
11733 		else
11734 			tim = 0;
11735 		cbw = data * HPTS_USEC_IN_SEC;
11736 		if (tim > 0)
11737 			cbw /= tim;
11738 		else
11739 			cbw = 0;
11740 		rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
11741 		/*
11742 		 * Check to see if we are freeing what we are pointing to send wise
11743 		 * if so be sure to NULL the pointer so we know we are no longer
11744 		 * set to anything.
11745 		 */
11746 		if (ent == rack->r_ctl.rc_last_sft) {
11747 			rack->r_ctl.rc_last_sft = NULL;
11748 			if (rack->rc_hybrid_mode) {
11749 				rack->rc_catch_up = 0;
11750 				if (rack->cspr_is_fcc == 0)
11751 					rack->r_ctl.bw_rate_cap = 0;
11752 				else
11753 					rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11754 				rack->r_ctl.client_suggested_maxseg = 0;
11755 			}
11756 		}
11757 		/* Generate the log that the tcp_netflix call would have */
11758 		tcp_req_log_req_info(rack->rc_tp, ent,
11759 				      i, TCP_TRK_REQ_LOG_FREED, 0, 0);
11760 		/* Free it and see if there is another one */
11761 		tcp_req_free_a_slot(rack->rc_tp, ent);
11762 		ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11763 	}
11764 }
11765 #endif
11766 
11767 
11768 /*
11769  * Return value of 1, we do not need to call rack_process_data().
11770  * return value of 0, rack_process_data can be called.
11771  * For ret_val if its 0 the TCP is locked, if its non-zero
11772  * its unlocked and probably unsafe to touch the TCB.
11773  */
11774 static int
11775 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11776     struct tcpcb *tp, struct tcpopt *to,
11777     uint32_t tiwin, int32_t tlen,
11778     int32_t * ofia, int32_t thflags, int32_t *ret_val, int32_t orig_tlen)
11779 {
11780 	int32_t ourfinisacked = 0;
11781 	int32_t nsegs, acked_amount;
11782 	int32_t acked;
11783 	struct mbuf *mfree;
11784 	struct tcp_rack *rack;
11785 	int32_t under_pacing = 0;
11786 	int32_t post_recovery = 0;
11787 	uint32_t p_cwnd;
11788 
11789 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11790 
11791 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11792 	if (SEQ_GEQ(tp->snd_una, tp->iss + (65535 << tp->snd_scale))) {
11793 		/* Checking SEG.ACK against ISS is definitely redundant. */
11794 		tp->t_flags2 |= TF2_NO_ISS_CHECK;
11795 	}
11796 	if (!V_tcp_insecure_ack) {
11797 		tcp_seq seq_min;
11798 		bool ghost_ack_check;
11799 
11800 		if (tp->t_flags2 & TF2_NO_ISS_CHECK) {
11801 			/* Check for too old ACKs (RFC 5961, Section 5.2). */
11802 			seq_min = tp->snd_una - tp->max_sndwnd;
11803 			ghost_ack_check = false;
11804 		} else {
11805 			if (SEQ_GT(tp->iss + 1, tp->snd_una - tp->max_sndwnd)) {
11806 				/* Checking for ghost ACKs is stricter. */
11807 				seq_min = tp->iss + 1;
11808 				ghost_ack_check = true;
11809 			} else {
11810 				/*
11811 				 * Checking for too old ACKs (RFC 5961,
11812 				 * Section 5.2) is stricter.
11813 				 */
11814 				seq_min = tp->snd_una - tp->max_sndwnd;
11815 				ghost_ack_check = false;
11816 			}
11817 		}
11818 		if (SEQ_LT(th->th_ack, seq_min)) {
11819 			if (ghost_ack_check)
11820 				TCPSTAT_INC(tcps_rcvghostack);
11821 			else
11822 				TCPSTAT_INC(tcps_rcvacktooold);
11823 			/* Send challenge ACK. */
11824 			ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11825 			rack->r_wanted_output = 1;
11826 			return (1);
11827 		}
11828 	}
11829 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
11830 		ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11831 		rack->r_wanted_output = 1;
11832 		return (1);
11833 	}
11834 	if (rack->gp_ready &&
11835 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11836 		under_pacing = 1;
11837 	}
11838 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
11839 		int in_rec, dup_ack_struck = 0;
11840 		int dsack_seen = 0, sacks_seen = 0;
11841 
11842 		in_rec = IN_FASTRECOVERY(tp->t_flags);
11843 		if (rack->rc_in_persist) {
11844 			tp->t_rxtshift = 0;
11845 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11846 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11847 		}
11848 
11849 		if ((th->th_ack == tp->snd_una) &&
11850 		    (tiwin == tp->snd_wnd) &&
11851 		    (orig_tlen == 0) &&
11852 		    ((to->to_flags & TOF_SACK) == 0)) {
11853 			rack_strike_dupack(rack, th->th_ack);
11854 			dup_ack_struck = 1;
11855 		}
11856 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
11857 			     dup_ack_struck, &dsack_seen, &sacks_seen);
11858 
11859 	}
11860 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
11861 		/*
11862 		 * Old ack, behind (or duplicate to) the last one rcv'd
11863 		 * Note: We mark reordering is occuring if its
11864 		 * less than and we have not closed our window.
11865 		 */
11866 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
11867 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
11868 			if (rack->r_ctl.rc_reorder_ts == 0)
11869 				rack->r_ctl.rc_reorder_ts = 1;
11870 		}
11871 		return (0);
11872 	}
11873 	/*
11874 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
11875 	 * something we sent.
11876 	 */
11877 	if (tp->t_flags & TF_NEEDSYN) {
11878 		/*
11879 		 * T/TCP: Connection was half-synchronized, and our SYN has
11880 		 * been ACK'd (so connection is now fully synchronized).  Go
11881 		 * to non-starred state, increment snd_una for ACK of SYN,
11882 		 * and check if we can do window scaling.
11883 		 */
11884 		tp->t_flags &= ~TF_NEEDSYN;
11885 		tp->snd_una++;
11886 		/* Do window scaling? */
11887 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11888 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11889 			tp->rcv_scale = tp->request_r_scale;
11890 			/* Send window already scaled. */
11891 		}
11892 	}
11893 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
11894 
11895 	acked = BYTES_THIS_ACK(tp, th);
11896 	if (acked) {
11897 		/*
11898 		 * Any time we move the cum-ack forward clear
11899 		 * keep-alive tied probe-not-answered. The
11900 		 * persists clears its own on entry.
11901 		 */
11902 		rack->probe_not_answered = 0;
11903 	}
11904 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11905 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11906 	/*
11907 	 * If we just performed our first retransmit, and the ACK arrives
11908 	 * within our recovery window, then it was a mistake to do the
11909 	 * retransmit in the first place.  Recover our original cwnd and
11910 	 * ssthresh, and proceed to transmit where we left off.
11911 	 */
11912 	if ((tp->t_flags & TF_PREVVALID) &&
11913 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11914 		tp->t_flags &= ~TF_PREVVALID;
11915 		if (tp->t_rxtshift == 1 &&
11916 		    (int)(ticks - tp->t_badrxtwin) < 0)
11917 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
11918 	}
11919 	if (acked) {
11920 		/* assure we are not backed off */
11921 		tp->t_rxtshift = 0;
11922 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11923 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11924 		rack->rc_tlp_in_progress = 0;
11925 		rack->r_ctl.rc_tlp_cnt_out = 0;
11926 		/*
11927 		 * If it is the RXT timer we want to
11928 		 * stop it, so we can restart a TLP.
11929 		 */
11930 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11931 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11932 #ifdef TCP_REQUEST_TRK
11933 		rack_req_check_for_comp(rack, th->th_ack);
11934 #endif
11935 	}
11936 	/*
11937 	 * If we have a timestamp reply, update smoothed round trip time. If
11938 	 * no timestamp is present but transmit timer is running and timed
11939 	 * sequence number was acked, update smoothed round trip time. Since
11940 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
11941 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
11942 	 * timer.
11943 	 *
11944 	 * Some boxes send broken timestamp replies during the SYN+ACK
11945 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
11946 	 * and blow up the retransmit timer.
11947 	 */
11948 	/*
11949 	 * If all outstanding data is acked, stop retransmit timer and
11950 	 * remember to restart (more output or persist). If there is more
11951 	 * data to be acked, restart retransmit timer, using current
11952 	 * (possibly backed-off) value.
11953 	 */
11954 	if (acked == 0) {
11955 		if (ofia)
11956 			*ofia = ourfinisacked;
11957 		return (0);
11958 	}
11959 	if (IN_RECOVERY(tp->t_flags)) {
11960 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
11961 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
11962 			tcp_rack_partialack(tp);
11963 		} else {
11964 			rack_post_recovery(tp, th->th_ack);
11965 			post_recovery = 1;
11966 			/*
11967 			 * Grab the segsiz, multiply by 2 and add the snd_cwnd
11968 			 * that is the max the CC should add if we are exiting
11969 			 * recovery and doing a late add.
11970 			 */
11971 			p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
11972 			p_cwnd <<= 1;
11973 			p_cwnd += tp->snd_cwnd;
11974 		}
11975 	} else if ((rack->rto_from_rec == 1) &&
11976 		   SEQ_GEQ(th->th_ack, tp->snd_recover)) {
11977 		/*
11978 		 * We were in recovery, hit a rxt timeout
11979 		 * and never re-entered recovery. The timeout(s)
11980 		 * made up all the lost data. In such a case
11981 		 * we need to clear the rto_from_rec flag.
11982 		 */
11983 		rack->rto_from_rec = 0;
11984 	}
11985 	/*
11986 	 * Let the congestion control algorithm update congestion control
11987 	 * related information. This typically means increasing the
11988 	 * congestion window.
11989 	 */
11990 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, post_recovery);
11991 	if (post_recovery &&
11992 	    (tp->snd_cwnd > p_cwnd)) {
11993 		/* Must be non-newreno (cubic) getting too ahead of itself */
11994 		tp->snd_cwnd = p_cwnd;
11995 	}
11996 	SOCK_SENDBUF_LOCK(so);
11997 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
11998 	tp->snd_wnd -= acked_amount;
11999 	mfree = sbcut_locked(&so->so_snd, acked_amount);
12000 	if ((sbused(&so->so_snd) == 0) &&
12001 	    (acked > acked_amount) &&
12002 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
12003 	    (tp->t_flags & TF_SENTFIN)) {
12004 		/*
12005 		 * We must be sure our fin
12006 		 * was sent and acked (we can be
12007 		 * in FIN_WAIT_1 without having
12008 		 * sent the fin).
12009 		 */
12010 		ourfinisacked = 1;
12011 	}
12012 	tp->snd_una = th->th_ack;
12013 	/* wakeups? */
12014 	if (acked_amount && sbavail(&so->so_snd))
12015 		rack_adjust_sendmap_head(rack, &so->so_snd);
12016 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12017 	/* NB: sowwakeup_locked() does an implicit unlock. */
12018 	sowwakeup_locked(so);
12019 	m_freem(mfree);
12020 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
12021 		tp->snd_recover = tp->snd_una;
12022 
12023 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12024 		tp->snd_nxt = tp->snd_max;
12025 	}
12026 	if (under_pacing &&
12027 	    (rack->use_fixed_rate == 0) &&
12028 	    (rack->in_probe_rtt == 0) &&
12029 	    rack->rc_gp_dyn_mul &&
12030 	    rack->rc_always_pace) {
12031 		/* Check if we are dragging bottom */
12032 		rack_check_bottom_drag(tp, rack, so);
12033 	}
12034 	if (tp->snd_una == tp->snd_max) {
12035 		/* Nothing left outstanding */
12036 		tp->t_flags &= ~TF_PREVVALID;
12037 		if (rack->r_ctl.rc_went_idle_time == 0)
12038 			rack->r_ctl.rc_went_idle_time = 1;
12039 		rack->r_ctl.retran_during_recovery = 0;
12040 		rack->r_ctl.dsack_byte_cnt = 0;
12041 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12042 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12043 			tp->t_acktime = 0;
12044 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12045 		rack->rc_suspicious = 0;
12046 		/* Set need output so persist might get set */
12047 		rack->r_wanted_output = 1;
12048 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12049 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12050 		    (sbavail(&so->so_snd) == 0) &&
12051 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12052 			/*
12053 			 * The socket was gone and the
12054 			 * peer sent data (now or in the past), time to
12055 			 * reset him.
12056 			 */
12057 			*ret_val = 1;
12058 			/* tcp_close will kill the inp pre-log the Reset */
12059 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12060 			tp = tcp_close(tp);
12061 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12062 			return (1);
12063 		}
12064 	}
12065 	if (ofia)
12066 		*ofia = ourfinisacked;
12067 	return (0);
12068 }
12069 
12070 
12071 static void
12072 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12073 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
12074 {
12075 	if (tcp_bblogging_on(rack->rc_tp)) {
12076 		union tcp_log_stackspecific log;
12077 		struct timeval tv;
12078 
12079 		memset(&log, 0, sizeof(log));
12080 		log.u_bbr.flex1 = cnt;
12081 		log.u_bbr.flex2 = split;
12082 		log.u_bbr.flex3 = out;
12083 		log.u_bbr.flex4 = line;
12084 		log.u_bbr.flex5 = rack->r_must_retran;
12085 		log.u_bbr.flex6 = flags;
12086 		log.u_bbr.flex7 = rack->rc_has_collapsed;
12087 		log.u_bbr.flex8 = dir;	/*
12088 					 * 1 is collapsed, 0 is uncollapsed,
12089 					 * 2 is log of a rsm being marked, 3 is a split.
12090 					 */
12091 		if (rsm == NULL)
12092 			log.u_bbr.rttProp = 0;
12093 		else
12094 			log.u_bbr.rttProp = (uintptr_t)rsm;
12095 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12096 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12097 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
12098 		    &rack->rc_inp->inp_socket->so_rcv,
12099 		    &rack->rc_inp->inp_socket->so_snd,
12100 		    TCP_RACK_LOG_COLLAPSE, 0,
12101 		    0, &log, false, &tv);
12102 	}
12103 }
12104 
12105 static void
12106 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12107 {
12108 	/*
12109 	 * Here all we do is mark the collapsed point and set the flag.
12110 	 * This may happen again and again, but there is no
12111 	 * sense splitting our map until we know where the
12112 	 * peer finally lands in the collapse.
12113 	 */
12114 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12115 	if ((rack->rc_has_collapsed == 0) ||
12116 	    (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12117 		counter_u64_add(rack_collapsed_win_seen, 1);
12118 	rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12119 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12120 	rack->rc_has_collapsed = 1;
12121 	rack->r_collapse_point_valid = 1;
12122 	rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12123 }
12124 
12125 static void
12126 rack_un_collapse_window(struct tcp_rack *rack, int line)
12127 {
12128 	struct rack_sendmap *nrsm, *rsm;
12129 	int cnt = 0, split = 0;
12130 	int insret __diagused;
12131 
12132 
12133 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12134 	rack->rc_has_collapsed = 0;
12135 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12136 	if (rsm == NULL) {
12137 		/* Nothing to do maybe the peer ack'ed it all */
12138 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12139 		return;
12140 	}
12141 	/* Now do we need to split this one? */
12142 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12143 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12144 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12145 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12146 		if (nrsm == NULL) {
12147 			/* We can't get a rsm, mark all? */
12148 			nrsm = rsm;
12149 			goto no_split;
12150 		}
12151 		/* Clone it */
12152 		split = 1;
12153 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12154 #ifndef INVARIANTS
12155 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12156 #else
12157 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12158 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
12159 			      nrsm, insret, rack, rsm);
12160 		}
12161 #endif
12162 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12163 				 rack->r_ctl.last_collapse_point, __LINE__);
12164 		if (rsm->r_in_tmap) {
12165 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12166 			nrsm->r_in_tmap = 1;
12167 		}
12168 		/*
12169 		 * Set in the new RSM as the
12170 		 * collapsed starting point
12171 		 */
12172 		rsm = nrsm;
12173 	}
12174 
12175 no_split:
12176 	TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
12177 		cnt++;
12178 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
12179 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12180 		cnt++;
12181 	}
12182 	if (cnt) {
12183 		counter_u64_add(rack_collapsed_win, 1);
12184 	}
12185 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12186 }
12187 
12188 static void
12189 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12190 			int32_t tlen, int32_t tfo_syn)
12191 {
12192 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
12193 		rack_timer_cancel(tp, rack,
12194 				  rack->r_ctl.rc_rcvtime, __LINE__);
12195 		tp->t_flags |= TF_DELACK;
12196 	} else {
12197 		rack->r_wanted_output = 1;
12198 		tp->t_flags |= TF_ACKNOW;
12199 	}
12200 }
12201 
12202 static void
12203 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12204 {
12205 	/*
12206 	 * If fast output is in progress, lets validate that
12207 	 * the new window did not shrink on us and make it
12208 	 * so fast output should end.
12209 	 */
12210 	if (rack->r_fast_output) {
12211 		uint32_t out;
12212 
12213 		/*
12214 		 * Calculate what we will send if left as is
12215 		 * and compare that to our send window.
12216 		 */
12217 		out = ctf_outstanding(tp);
12218 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12219 			/* ok we have an issue */
12220 			if (out >= tp->snd_wnd) {
12221 				/* Turn off fast output the window is met or collapsed */
12222 				rack->r_fast_output = 0;
12223 			} else {
12224 				/* we have some room left */
12225 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12226 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12227 					/* If not at least 1 full segment never mind */
12228 					rack->r_fast_output = 0;
12229 				}
12230 			}
12231 		}
12232 	}
12233 }
12234 
12235 /*
12236  * Return value of 1, the TCB is unlocked and most
12237  * likely gone, return value of 0, the TCP is still
12238  * locked.
12239  */
12240 static int
12241 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12242     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12243     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12244 {
12245 	/*
12246 	 * Update window information. Don't look at window if no ACK: TAC's
12247 	 * send garbage on first SYN.
12248 	 */
12249 	int32_t nsegs;
12250 	int32_t tfo_syn;
12251 	struct tcp_rack *rack;
12252 
12253 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12254 
12255 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12256 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12257 	if ((thflags & TH_ACK) &&
12258 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12259 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12260 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12261 		/* keep track of pure window updates */
12262 		if (tlen == 0 &&
12263 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12264 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12265 		tp->snd_wnd = tiwin;
12266 		rack_validate_fo_sendwin_up(tp, rack);
12267 		tp->snd_wl1 = th->th_seq;
12268 		tp->snd_wl2 = th->th_ack;
12269 		if (tp->snd_wnd > tp->max_sndwnd)
12270 			tp->max_sndwnd = tp->snd_wnd;
12271 		rack->r_wanted_output = 1;
12272 	} else if (thflags & TH_ACK) {
12273 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12274 			tp->snd_wnd = tiwin;
12275 			rack_validate_fo_sendwin_up(tp, rack);
12276 			tp->snd_wl1 = th->th_seq;
12277 			tp->snd_wl2 = th->th_ack;
12278 		}
12279 	}
12280 	if (tp->snd_wnd < ctf_outstanding(tp))
12281 		/* The peer collapsed the window */
12282 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12283 	else if (rack->rc_has_collapsed)
12284 		rack_un_collapse_window(rack, __LINE__);
12285 	if ((rack->r_collapse_point_valid) &&
12286 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12287 		rack->r_collapse_point_valid = 0;
12288 	/* Was persist timer active and now we have window space? */
12289 	if ((rack->rc_in_persist != 0) &&
12290 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12291 				rack->r_ctl.rc_pace_min_segs))) {
12292 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12293 		tp->snd_nxt = tp->snd_max;
12294 		/* Make sure we output to start the timer */
12295 		rack->r_wanted_output = 1;
12296 	}
12297 	/* Do we enter persists? */
12298 	if ((rack->rc_in_persist == 0) &&
12299 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12300 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12301 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12302 	    sbavail(&tptosocket(tp)->so_snd) &&
12303 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12304 		/*
12305 		 * Here the rwnd is less than
12306 		 * the pacing size, we are established,
12307 		 * nothing is outstanding, and there is
12308 		 * data to send. Enter persists.
12309 		 */
12310 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12311 	}
12312 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12313 		m_freem(m);
12314 		return (0);
12315 	}
12316 	/*
12317 	 * don't process the URG bit, ignore them drag
12318 	 * along the up.
12319 	 */
12320 	tp->rcv_up = tp->rcv_nxt;
12321 
12322 	/*
12323 	 * Process the segment text, merging it into the TCP sequencing
12324 	 * queue, and arranging for acknowledgment of receipt if necessary.
12325 	 * This process logically involves adjusting tp->rcv_wnd as data is
12326 	 * presented to the user (this happens in tcp_usrreq.c, case
12327 	 * PRU_RCVD).  If a FIN has already been received on this connection
12328 	 * then we just ignore the text.
12329 	 */
12330 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12331 	    (tp->t_flags & TF_FASTOPEN));
12332 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12333 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12334 		tcp_seq save_start = th->th_seq;
12335 		tcp_seq save_rnxt  = tp->rcv_nxt;
12336 		int     save_tlen  = tlen;
12337 
12338 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12339 		/*
12340 		 * Insert segment which includes th into TCP reassembly
12341 		 * queue with control block tp.  Set thflags to whether
12342 		 * reassembly now includes a segment with FIN.  This handles
12343 		 * the common case inline (segment is the next to be
12344 		 * received on an established connection, and the queue is
12345 		 * empty), avoiding linkage into and removal from the queue
12346 		 * and repetition of various conversions. Set DELACK for
12347 		 * segments received in order, but ack immediately when
12348 		 * segments are out of order (so fast retransmit can work).
12349 		 */
12350 		if (th->th_seq == tp->rcv_nxt &&
12351 		    SEGQ_EMPTY(tp) &&
12352 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
12353 		    tfo_syn)) {
12354 #ifdef NETFLIX_SB_LIMITS
12355 			u_int mcnt, appended;
12356 
12357 			if (so->so_rcv.sb_shlim) {
12358 				mcnt = m_memcnt(m);
12359 				appended = 0;
12360 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12361 				    CFO_NOSLEEP, NULL) == false) {
12362 					counter_u64_add(tcp_sb_shlim_fails, 1);
12363 					m_freem(m);
12364 					return (0);
12365 				}
12366 			}
12367 #endif
12368 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12369 			tp->rcv_nxt += tlen;
12370 			if (tlen &&
12371 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12372 			    (tp->t_fbyte_in == 0)) {
12373 				tp->t_fbyte_in = ticks;
12374 				if (tp->t_fbyte_in == 0)
12375 					tp->t_fbyte_in = 1;
12376 				if (tp->t_fbyte_out && tp->t_fbyte_in)
12377 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12378 			}
12379 			thflags = tcp_get_flags(th) & TH_FIN;
12380 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12381 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12382 			SOCK_RECVBUF_LOCK(so);
12383 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12384 				m_freem(m);
12385 			} else {
12386 				int32_t newsize;
12387 
12388 				if (tlen > 0) {
12389 					newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12390 					if (newsize)
12391 						if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12392 							so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12393 				}
12394 #ifdef NETFLIX_SB_LIMITS
12395 				appended =
12396 #endif
12397 					sbappendstream_locked(&so->so_rcv, m, 0);
12398 			}
12399 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12400 			/* NB: sorwakeup_locked() does an implicit unlock. */
12401 			sorwakeup_locked(so);
12402 #ifdef NETFLIX_SB_LIMITS
12403 			if (so->so_rcv.sb_shlim && appended != mcnt)
12404 				counter_fo_release(so->so_rcv.sb_shlim,
12405 				    mcnt - appended);
12406 #endif
12407 		} else {
12408 			/*
12409 			 * XXX: Due to the header drop above "th" is
12410 			 * theoretically invalid by now.  Fortunately
12411 			 * m_adj() doesn't actually frees any mbufs when
12412 			 * trimming from the head.
12413 			 */
12414 			tcp_seq temp = save_start;
12415 
12416 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
12417 			tp->t_flags |= TF_ACKNOW;
12418 			if (tp->t_flags & TF_WAKESOR) {
12419 				tp->t_flags &= ~TF_WAKESOR;
12420 				/* NB: sorwakeup_locked() does an implicit unlock. */
12421 				sorwakeup_locked(so);
12422 			}
12423 		}
12424 		if ((tp->t_flags & TF_SACK_PERMIT) &&
12425 		    (save_tlen > 0) &&
12426 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
12427 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12428 				/*
12429 				 * DSACK actually handled in the fastpath
12430 				 * above.
12431 				 */
12432 				tcp_update_sack_list(tp, save_start,
12433 				    save_start + save_tlen);
12434 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12435 				if ((tp->rcv_numsacks >= 1) &&
12436 				    (tp->sackblks[0].end == save_start)) {
12437 					/*
12438 					 * Partial overlap, recorded at todrop
12439 					 * above.
12440 					 */
12441 					tcp_update_sack_list(tp,
12442 					    tp->sackblks[0].start,
12443 					    tp->sackblks[0].end);
12444 				} else {
12445 					tcp_update_dsack_list(tp, save_start,
12446 					    save_start + save_tlen);
12447 				}
12448 			} else if (tlen >= save_tlen) {
12449 				/* Update of sackblks. */
12450 				tcp_update_dsack_list(tp, save_start,
12451 				    save_start + save_tlen);
12452 			} else if (tlen > 0) {
12453 				tcp_update_dsack_list(tp, save_start,
12454 				    save_start + tlen);
12455 			}
12456 		}
12457 	} else {
12458 		m_freem(m);
12459 		thflags &= ~TH_FIN;
12460 	}
12461 
12462 	/*
12463 	 * If FIN is received ACK the FIN and let the user know that the
12464 	 * connection is closing.
12465 	 */
12466 	if (thflags & TH_FIN) {
12467 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12468 			/* The socket upcall is handled by socantrcvmore. */
12469 			socantrcvmore(so);
12470 			/*
12471 			 * If connection is half-synchronized (ie NEEDSYN
12472 			 * flag on) then delay ACK, so it may be piggybacked
12473 			 * when SYN is sent. Otherwise, since we received a
12474 			 * FIN then no more input can be expected, send ACK
12475 			 * now.
12476 			 */
12477 			if (tp->t_flags & TF_NEEDSYN) {
12478 				rack_timer_cancel(tp, rack,
12479 				    rack->r_ctl.rc_rcvtime, __LINE__);
12480 				tp->t_flags |= TF_DELACK;
12481 			} else {
12482 				tp->t_flags |= TF_ACKNOW;
12483 			}
12484 			tp->rcv_nxt++;
12485 		}
12486 		switch (tp->t_state) {
12487 			/*
12488 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
12489 			 * CLOSE_WAIT state.
12490 			 */
12491 		case TCPS_SYN_RECEIVED:
12492 			tp->t_starttime = ticks;
12493 			/* FALLTHROUGH */
12494 		case TCPS_ESTABLISHED:
12495 			rack_timer_cancel(tp, rack,
12496 			    rack->r_ctl.rc_rcvtime, __LINE__);
12497 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
12498 			break;
12499 
12500 			/*
12501 			 * If still in FIN_WAIT_1 STATE FIN has not been
12502 			 * acked so enter the CLOSING state.
12503 			 */
12504 		case TCPS_FIN_WAIT_1:
12505 			rack_timer_cancel(tp, rack,
12506 			    rack->r_ctl.rc_rcvtime, __LINE__);
12507 			tcp_state_change(tp, TCPS_CLOSING);
12508 			break;
12509 
12510 			/*
12511 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
12512 			 * starting the time-wait timer, turning off the
12513 			 * other standard timers.
12514 			 */
12515 		case TCPS_FIN_WAIT_2:
12516 			rack_timer_cancel(tp, rack,
12517 			    rack->r_ctl.rc_rcvtime, __LINE__);
12518 			tcp_twstart(tp);
12519 			return (1);
12520 		}
12521 	}
12522 	/*
12523 	 * Return any desired output.
12524 	 */
12525 	if ((tp->t_flags & TF_ACKNOW) ||
12526 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12527 		rack->r_wanted_output = 1;
12528 	}
12529 	return (0);
12530 }
12531 
12532 /*
12533  * Here nothing is really faster, its just that we
12534  * have broken out the fast-data path also just like
12535  * the fast-ack.
12536  */
12537 static int
12538 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12539     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12540     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12541 {
12542 	int32_t nsegs;
12543 	int32_t newsize = 0;	/* automatic sockbuf scaling */
12544 	struct tcp_rack *rack;
12545 #ifdef NETFLIX_SB_LIMITS
12546 	u_int mcnt, appended;
12547 #endif
12548 
12549 	/*
12550 	 * If last ACK falls within this segment's sequence numbers, record
12551 	 * the timestamp. NOTE that the test is modified according to the
12552 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12553 	 */
12554 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12555 		return (0);
12556 	}
12557 	if (tiwin && tiwin != tp->snd_wnd) {
12558 		return (0);
12559 	}
12560 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12561 		return (0);
12562 	}
12563 	if (__predict_false((to->to_flags & TOF_TS) &&
12564 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12565 		return (0);
12566 	}
12567 	if (__predict_false((th->th_ack != tp->snd_una))) {
12568 		return (0);
12569 	}
12570 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12571 		return (0);
12572 	}
12573 	if ((to->to_flags & TOF_TS) != 0 &&
12574 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12575 		tp->ts_recent_age = tcp_ts_getticks();
12576 		tp->ts_recent = to->to_tsval;
12577 	}
12578 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12579 	/*
12580 	 * This is a pure, in-sequence data packet with nothing on the
12581 	 * reassembly queue and we have enough buffer space to take it.
12582 	 */
12583 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12584 
12585 #ifdef NETFLIX_SB_LIMITS
12586 	if (so->so_rcv.sb_shlim) {
12587 		mcnt = m_memcnt(m);
12588 		appended = 0;
12589 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12590 		    CFO_NOSLEEP, NULL) == false) {
12591 			counter_u64_add(tcp_sb_shlim_fails, 1);
12592 			m_freem(m);
12593 			return (1);
12594 		}
12595 	}
12596 #endif
12597 	/* Clean receiver SACK report if present */
12598 	if (tp->rcv_numsacks)
12599 		tcp_clean_sackreport(tp);
12600 	KMOD_TCPSTAT_INC(tcps_preddat);
12601 	tp->rcv_nxt += tlen;
12602 	if (tlen &&
12603 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12604 	    (tp->t_fbyte_in == 0)) {
12605 		tp->t_fbyte_in = ticks;
12606 		if (tp->t_fbyte_in == 0)
12607 			tp->t_fbyte_in = 1;
12608 		if (tp->t_fbyte_out && tp->t_fbyte_in)
12609 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12610 	}
12611 	/*
12612 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12613 	 */
12614 	tp->snd_wl1 = th->th_seq;
12615 	/*
12616 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12617 	 */
12618 	tp->rcv_up = tp->rcv_nxt;
12619 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12620 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12621 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12622 
12623 	/* Add data to socket buffer. */
12624 	SOCK_RECVBUF_LOCK(so);
12625 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12626 		m_freem(m);
12627 	} else {
12628 		/*
12629 		 * Set new socket buffer size. Give up when limit is
12630 		 * reached.
12631 		 */
12632 		if (newsize)
12633 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12634 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12635 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12636 #ifdef NETFLIX_SB_LIMITS
12637 		appended =
12638 #endif
12639 			sbappendstream_locked(&so->so_rcv, m, 0);
12640 		ctf_calc_rwin(so, tp);
12641 	}
12642 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12643 	/* NB: sorwakeup_locked() does an implicit unlock. */
12644 	sorwakeup_locked(so);
12645 #ifdef NETFLIX_SB_LIMITS
12646 	if (so->so_rcv.sb_shlim && mcnt != appended)
12647 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12648 #endif
12649 	rack_handle_delayed_ack(tp, rack, tlen, 0);
12650 	if (tp->snd_una == tp->snd_max)
12651 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12652 	return (1);
12653 }
12654 
12655 /*
12656  * This subfunction is used to try to highly optimize the
12657  * fast path. We again allow window updates that are
12658  * in sequence to remain in the fast-path. We also add
12659  * in the __predict's to attempt to help the compiler.
12660  * Note that if we return a 0, then we can *not* process
12661  * it and the caller should push the packet into the
12662  * slow-path.
12663  */
12664 static int
12665 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12666     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12667     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12668 {
12669 	int32_t acked;
12670 	int32_t nsegs;
12671 	int32_t under_pacing = 0;
12672 	struct tcp_rack *rack;
12673 
12674 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12675 		/* Old ack, behind (or duplicate to) the last one rcv'd */
12676 		return (0);
12677 	}
12678 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
12679 		/* Above what we have sent? */
12680 		return (0);
12681 	}
12682 	if (__predict_false(tiwin == 0)) {
12683 		/* zero window */
12684 		return (0);
12685 	}
12686 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
12687 		/* We need a SYN or a FIN, unlikely.. */
12688 		return (0);
12689 	}
12690 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
12691 		/* Timestamp is behind .. old ack with seq wrap? */
12692 		return (0);
12693 	}
12694 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
12695 		/* Still recovering */
12696 		return (0);
12697 	}
12698 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12699 	if (rack->r_ctl.rc_sacked) {
12700 		/* We have sack holes on our scoreboard */
12701 		return (0);
12702 	}
12703 	/* Ok if we reach here, we can process a fast-ack */
12704 	if (rack->gp_ready &&
12705 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12706 		under_pacing = 1;
12707 	}
12708 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12709 	rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
12710 	/* Did the window get updated? */
12711 	if (tiwin != tp->snd_wnd) {
12712 		tp->snd_wnd = tiwin;
12713 		rack_validate_fo_sendwin_up(tp, rack);
12714 		tp->snd_wl1 = th->th_seq;
12715 		if (tp->snd_wnd > tp->max_sndwnd)
12716 			tp->max_sndwnd = tp->snd_wnd;
12717 	}
12718 	/* Do we exit persists? */
12719 	if ((rack->rc_in_persist != 0) &&
12720 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12721 			       rack->r_ctl.rc_pace_min_segs))) {
12722 		rack_exit_persist(tp, rack, cts);
12723 	}
12724 	/* Do we enter persists? */
12725 	if ((rack->rc_in_persist == 0) &&
12726 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12727 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12728 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12729 	    sbavail(&tptosocket(tp)->so_snd) &&
12730 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12731 		/*
12732 		 * Here the rwnd is less than
12733 		 * the pacing size, we are established,
12734 		 * nothing is outstanding, and there is
12735 		 * data to send. Enter persists.
12736 		 */
12737 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
12738 	}
12739 	/*
12740 	 * If last ACK falls within this segment's sequence numbers, record
12741 	 * the timestamp. NOTE that the test is modified according to the
12742 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12743 	 */
12744 	if ((to->to_flags & TOF_TS) != 0 &&
12745 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12746 		tp->ts_recent_age = tcp_ts_getticks();
12747 		tp->ts_recent = to->to_tsval;
12748 	}
12749 	/*
12750 	 * This is a pure ack for outstanding data.
12751 	 */
12752 	KMOD_TCPSTAT_INC(tcps_predack);
12753 
12754 	/*
12755 	 * "bad retransmit" recovery.
12756 	 */
12757 	if ((tp->t_flags & TF_PREVVALID) &&
12758 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12759 		tp->t_flags &= ~TF_PREVVALID;
12760 		if (tp->t_rxtshift == 1 &&
12761 		    (int)(ticks - tp->t_badrxtwin) < 0)
12762 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12763 	}
12764 	/*
12765 	 * Recalculate the transmit timer / rtt.
12766 	 *
12767 	 * Some boxes send broken timestamp replies during the SYN+ACK
12768 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
12769 	 * and blow up the retransmit timer.
12770 	 */
12771 	acked = BYTES_THIS_ACK(tp, th);
12772 
12773 #ifdef TCP_HHOOK
12774 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
12775 	hhook_run_tcp_est_in(tp, th, to);
12776 #endif
12777 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12778 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12779 	if (acked) {
12780 		struct mbuf *mfree;
12781 
12782 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
12783 		SOCK_SENDBUF_LOCK(so);
12784 		mfree = sbcut_locked(&so->so_snd, acked);
12785 		tp->snd_una = th->th_ack;
12786 		/* Note we want to hold the sb lock through the sendmap adjust */
12787 		rack_adjust_sendmap_head(rack, &so->so_snd);
12788 		/* Wake up the socket if we have room to write more */
12789 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12790 		sowwakeup_locked(so);
12791 		m_freem(mfree);
12792 		tp->t_rxtshift = 0;
12793 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12794 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12795 		rack->rc_tlp_in_progress = 0;
12796 		rack->r_ctl.rc_tlp_cnt_out = 0;
12797 		/*
12798 		 * If it is the RXT timer we want to
12799 		 * stop it, so we can restart a TLP.
12800 		 */
12801 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12802 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12803 
12804 #ifdef TCP_REQUEST_TRK
12805 		rack_req_check_for_comp(rack, th->th_ack);
12806 #endif
12807 	}
12808 	/*
12809 	 * Let the congestion control algorithm update congestion control
12810 	 * related information. This typically means increasing the
12811 	 * congestion window.
12812 	 */
12813 	if (tp->snd_wnd < ctf_outstanding(tp)) {
12814 		/* The peer collapsed the window */
12815 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12816 	} else if (rack->rc_has_collapsed)
12817 		rack_un_collapse_window(rack, __LINE__);
12818 	if ((rack->r_collapse_point_valid) &&
12819 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
12820 		rack->r_collapse_point_valid = 0;
12821 	/*
12822 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
12823 	 */
12824 	tp->snd_wl2 = th->th_ack;
12825 	tp->t_dupacks = 0;
12826 	m_freem(m);
12827 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
12828 
12829 	/*
12830 	 * If all outstanding data are acked, stop retransmit timer,
12831 	 * otherwise restart timer using current (possibly backed-off)
12832 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
12833 	 * If data are ready to send, let tcp_output decide between more
12834 	 * output or persist.
12835 	 */
12836 	if (under_pacing &&
12837 	    (rack->use_fixed_rate == 0) &&
12838 	    (rack->in_probe_rtt == 0) &&
12839 	    rack->rc_gp_dyn_mul &&
12840 	    rack->rc_always_pace) {
12841 		/* Check if we are dragging bottom */
12842 		rack_check_bottom_drag(tp, rack, so);
12843 	}
12844 	if (tp->snd_una == tp->snd_max) {
12845 		tp->t_flags &= ~TF_PREVVALID;
12846 		rack->r_ctl.retran_during_recovery = 0;
12847 		rack->rc_suspicious = 0;
12848 		rack->r_ctl.dsack_byte_cnt = 0;
12849 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12850 		if (rack->r_ctl.rc_went_idle_time == 0)
12851 			rack->r_ctl.rc_went_idle_time = 1;
12852 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12853 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12854 			tp->t_acktime = 0;
12855 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12856 	}
12857 	if (acked && rack->r_fast_output)
12858 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
12859 	if (sbavail(&so->so_snd)) {
12860 		rack->r_wanted_output = 1;
12861 	}
12862 	return (1);
12863 }
12864 
12865 /*
12866  * Return value of 1, the TCB is unlocked and most
12867  * likely gone, return value of 0, the TCP is still
12868  * locked.
12869  */
12870 static int
12871 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
12872     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12873     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12874 {
12875 	int32_t ret_val = 0;
12876 	int32_t orig_tlen = tlen;
12877 	int32_t todrop;
12878 	int32_t ourfinisacked = 0;
12879 	struct tcp_rack *rack;
12880 
12881 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12882 
12883 	ctf_calc_rwin(so, tp);
12884 	/*
12885 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
12886 	 * SYN, drop the input. if seg contains a RST, then drop the
12887 	 * connection. if seg does not contain SYN, then drop it. Otherwise
12888 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
12889 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
12890 	 * contains an ECE and ECN support is enabled, the stream is ECN
12891 	 * capable. if SYN has been acked change to ESTABLISHED else
12892 	 * SYN_RCVD state arrange for segment to be acked (eventually)
12893 	 * continue processing rest of data/controls.
12894 	 */
12895 	if ((thflags & TH_ACK) &&
12896 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
12897 	    SEQ_GT(th->th_ack, tp->snd_max))) {
12898 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
12899 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12900 		return (1);
12901 	}
12902 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
12903 		TCP_PROBE5(connect__refused, NULL, tp,
12904 		    mtod(m, const char *), tp, th);
12905 		tp = tcp_drop(tp, ECONNREFUSED);
12906 		ctf_do_drop(m, tp);
12907 		return (1);
12908 	}
12909 	if (thflags & TH_RST) {
12910 		ctf_do_drop(m, tp);
12911 		return (1);
12912 	}
12913 	if (!(thflags & TH_SYN)) {
12914 		ctf_do_drop(m, tp);
12915 		return (1);
12916 	}
12917 	tp->irs = th->th_seq;
12918 	tcp_rcvseqinit(tp);
12919 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12920 	if (thflags & TH_ACK) {
12921 		int tfo_partial = 0;
12922 
12923 		KMOD_TCPSTAT_INC(tcps_connects);
12924 		soisconnected(so);
12925 #ifdef MAC
12926 		mac_socketpeer_set_from_mbuf(m, so);
12927 #endif
12928 		/* Do window scaling on this connection? */
12929 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12930 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12931 			tp->rcv_scale = tp->request_r_scale;
12932 		}
12933 		tp->rcv_adv += min(tp->rcv_wnd,
12934 		    TCP_MAXWIN << tp->rcv_scale);
12935 		/*
12936 		 * If not all the data that was sent in the TFO SYN
12937 		 * has been acked, resend the remainder right away.
12938 		 */
12939 		if ((tp->t_flags & TF_FASTOPEN) &&
12940 		    (tp->snd_una != tp->snd_max)) {
12941 			/* Was it a partial ack? */
12942 			if (SEQ_LT(th->th_ack, tp->snd_max))
12943 				tfo_partial = 1;
12944 		}
12945 		/*
12946 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
12947 		 * will be turned on later.
12948 		 */
12949 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
12950 			rack_timer_cancel(tp, rack,
12951 					  rack->r_ctl.rc_rcvtime, __LINE__);
12952 			tp->t_flags |= TF_DELACK;
12953 		} else {
12954 			rack->r_wanted_output = 1;
12955 			tp->t_flags |= TF_ACKNOW;
12956 		}
12957 
12958 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
12959 
12960 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
12961 			/*
12962 			 * We advance snd_una for the
12963 			 * fast open case. If th_ack is
12964 			 * acknowledging data beyond
12965 			 * snd_una we can't just call
12966 			 * ack-processing since the
12967 			 * data stream in our send-map
12968 			 * will start at snd_una + 1 (one
12969 			 * beyond the SYN). If its just
12970 			 * equal we don't need to do that
12971 			 * and there is no send_map.
12972 			 */
12973 			tp->snd_una++;
12974 			if (tfo_partial && (SEQ_GT(tp->snd_max, tp->snd_una))) {
12975 				/*
12976 				 * We sent a SYN with data, and thus have a
12977 				 * sendmap entry with a SYN set. Lets find it
12978 				 * and take off the send bit and the byte and
12979 				 * set it up to be what we send (send it next).
12980 				 */
12981 				struct rack_sendmap *rsm;
12982 
12983 				rsm = tqhash_min(rack->r_ctl.tqh);
12984 				if (rsm) {
12985 					if (rsm->r_flags & RACK_HAS_SYN) {
12986 						rsm->r_flags &= ~RACK_HAS_SYN;
12987 						rsm->r_start++;
12988 					}
12989 					rack->r_ctl.rc_resend = rsm;
12990 				}
12991 			}
12992 		}
12993 		/*
12994 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
12995 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
12996 		 */
12997 		tp->t_starttime = ticks;
12998 		if (tp->t_flags & TF_NEEDFIN) {
12999 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
13000 			tp->t_flags &= ~TF_NEEDFIN;
13001 			thflags &= ~TH_SYN;
13002 		} else {
13003 			tcp_state_change(tp, TCPS_ESTABLISHED);
13004 			TCP_PROBE5(connect__established, NULL, tp,
13005 			    mtod(m, const char *), tp, th);
13006 			rack_cc_conn_init(tp);
13007 		}
13008 	} else {
13009 		/*
13010 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
13011 		 * open.  If segment contains CC option and there is a
13012 		 * cached CC, apply TAO test. If it succeeds, connection is *
13013 		 * half-synchronized. Otherwise, do 3-way handshake:
13014 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
13015 		 * there was no CC option, clear cached CC value.
13016 		 */
13017 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
13018 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
13019 	}
13020 	/*
13021 	 * Advance th->th_seq to correspond to first data byte. If data,
13022 	 * trim to stay within window, dropping FIN if necessary.
13023 	 */
13024 	th->th_seq++;
13025 	if (tlen > tp->rcv_wnd) {
13026 		todrop = tlen - tp->rcv_wnd;
13027 		m_adj(m, -todrop);
13028 		tlen = tp->rcv_wnd;
13029 		thflags &= ~TH_FIN;
13030 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13031 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13032 	}
13033 	tp->snd_wl1 = th->th_seq - 1;
13034 	tp->rcv_up = th->th_seq;
13035 	/*
13036 	 * Client side of transaction: already sent SYN and data. If the
13037 	 * remote host used T/TCP to validate the SYN, our data will be
13038 	 * ACK'd; if so, enter normal data segment processing in the middle
13039 	 * of step 5, ack processing. Otherwise, goto step 6.
13040 	 */
13041 	if (thflags & TH_ACK) {
13042 		/* For syn-sent we need to possibly update the rtt */
13043 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13044 			uint32_t t, mcts;
13045 
13046 			mcts = tcp_ts_getticks();
13047 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13048 			if (!tp->t_rttlow || tp->t_rttlow > t)
13049 				tp->t_rttlow = t;
13050 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13051 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13052 			tcp_rack_xmit_timer_commit(rack, tp);
13053 		}
13054 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen))
13055 			return (ret_val);
13056 		/* We may have changed to FIN_WAIT_1 above */
13057 		if (tp->t_state == TCPS_FIN_WAIT_1) {
13058 			/*
13059 			 * In FIN_WAIT_1 STATE in addition to the processing
13060 			 * for the ESTABLISHED state if our FIN is now
13061 			 * acknowledged then enter FIN_WAIT_2.
13062 			 */
13063 			if (ourfinisacked) {
13064 				/*
13065 				 * If we can't receive any more data, then
13066 				 * closing user can proceed. Starting the
13067 				 * timer is contrary to the specification,
13068 				 * but if we don't get a FIN we'll hang
13069 				 * forever.
13070 				 *
13071 				 * XXXjl: we should release the tp also, and
13072 				 * use a compressed state.
13073 				 */
13074 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13075 					soisdisconnected(so);
13076 					tcp_timer_activate(tp, TT_2MSL,
13077 					    (tcp_fast_finwait2_recycle ?
13078 					    tcp_finwait2_timeout :
13079 					    TP_MAXIDLE(tp)));
13080 				}
13081 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13082 			}
13083 		}
13084 	}
13085 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13086 	   tiwin, thflags, nxt_pkt));
13087 }
13088 
13089 /*
13090  * Return value of 1, the TCB is unlocked and most
13091  * likely gone, return value of 0, the TCP is still
13092  * locked.
13093  */
13094 static int
13095 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13096     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13097     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13098 {
13099 	struct tcp_rack *rack;
13100 	int32_t orig_tlen = tlen;
13101 	int32_t ret_val = 0;
13102 	int32_t ourfinisacked = 0;
13103 
13104 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13105 	ctf_calc_rwin(so, tp);
13106 	if ((thflags & TH_RST) ||
13107 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13108 		return (ctf_process_rst(m, th, so, tp));
13109 	if ((thflags & TH_ACK) &&
13110 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13111 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13112 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13113 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13114 		return (1);
13115 	}
13116 	if (tp->t_flags & TF_FASTOPEN) {
13117 		/*
13118 		 * When a TFO connection is in SYN_RECEIVED, the
13119 		 * only valid packets are the initial SYN, a
13120 		 * retransmit/copy of the initial SYN (possibly with
13121 		 * a subset of the original data), a valid ACK, a
13122 		 * FIN, or a RST.
13123 		 */
13124 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13125 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13126 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13127 			return (1);
13128 		} else if (thflags & TH_SYN) {
13129 			/* non-initial SYN is ignored */
13130 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13131 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13132 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13133 				ctf_do_drop(m, NULL);
13134 				return (0);
13135 			}
13136 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13137 			ctf_do_drop(m, NULL);
13138 			return (0);
13139 		}
13140 	}
13141 
13142 	/*
13143 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13144 	 * it's less than ts_recent, drop it.
13145 	 */
13146 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13147 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13148 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13149 			return (ret_val);
13150 	}
13151 	/*
13152 	 * In the SYN-RECEIVED state, validate that the packet belongs to
13153 	 * this connection before trimming the data to fit the receive
13154 	 * window.  Check the sequence number versus IRS since we know the
13155 	 * sequence numbers haven't wrapped.  This is a partial fix for the
13156 	 * "LAND" DoS attack.
13157 	 */
13158 	if (SEQ_LT(th->th_seq, tp->irs)) {
13159 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13160 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13161 		return (1);
13162 	}
13163 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13164 		return (ret_val);
13165 	}
13166 	/*
13167 	 * If last ACK falls within this segment's sequence numbers, record
13168 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13169 	 * from the latest proposal of the tcplw@cray.com list (Braden
13170 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13171 	 * with our earlier PAWS tests, so this check should be solely
13172 	 * predicated on the sequence space of this segment. 3) That we
13173 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13174 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13175 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13176 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13177 	 * p.869. In such cases, we can still calculate the RTT correctly
13178 	 * when RCV.NXT == Last.ACK.Sent.
13179 	 */
13180 	if ((to->to_flags & TOF_TS) != 0 &&
13181 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13182 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13183 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13184 		tp->ts_recent_age = tcp_ts_getticks();
13185 		tp->ts_recent = to->to_tsval;
13186 	}
13187 	tp->snd_wnd = tiwin;
13188 	rack_validate_fo_sendwin_up(tp, rack);
13189 	/*
13190 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13191 	 * is on (half-synchronized state), then queue data for later
13192 	 * processing; else drop segment and return.
13193 	 */
13194 	if ((thflags & TH_ACK) == 0) {
13195 		if (tp->t_flags & TF_FASTOPEN) {
13196 			rack_cc_conn_init(tp);
13197 		}
13198 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13199 		    tiwin, thflags, nxt_pkt));
13200 	}
13201 	KMOD_TCPSTAT_INC(tcps_connects);
13202 	if (tp->t_flags & TF_SONOTCONN) {
13203 		tp->t_flags &= ~TF_SONOTCONN;
13204 		soisconnected(so);
13205 	}
13206 	/* Do window scaling? */
13207 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13208 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13209 		tp->rcv_scale = tp->request_r_scale;
13210 	}
13211 	/*
13212 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
13213 	 * FIN-WAIT-1
13214 	 */
13215 	tp->t_starttime = ticks;
13216 	if ((tp->t_flags & TF_FASTOPEN) && tp->t_tfo_pending) {
13217 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13218 		tp->t_tfo_pending = NULL;
13219 	}
13220 	if (tp->t_flags & TF_NEEDFIN) {
13221 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
13222 		tp->t_flags &= ~TF_NEEDFIN;
13223 	} else {
13224 		tcp_state_change(tp, TCPS_ESTABLISHED);
13225 		TCP_PROBE5(accept__established, NULL, tp,
13226 		    mtod(m, const char *), tp, th);
13227 		/*
13228 		 * TFO connections call cc_conn_init() during SYN
13229 		 * processing.  Calling it again here for such connections
13230 		 * is not harmless as it would undo the snd_cwnd reduction
13231 		 * that occurs when a TFO SYN|ACK is retransmitted.
13232 		 */
13233 		if (!(tp->t_flags & TF_FASTOPEN))
13234 			rack_cc_conn_init(tp);
13235 	}
13236 	/*
13237 	 * Account for the ACK of our SYN prior to
13238 	 * regular ACK processing below, except for
13239 	 * simultaneous SYN, which is handled later.
13240 	 */
13241 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13242 		tp->snd_una++;
13243 	/*
13244 	 * If segment contains data or ACK, will call tcp_reass() later; if
13245 	 * not, do so now to pass queued data to user.
13246 	 */
13247 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
13248 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13249 		    (struct mbuf *)0);
13250 		if (tp->t_flags & TF_WAKESOR) {
13251 			tp->t_flags &= ~TF_WAKESOR;
13252 			/* NB: sorwakeup_locked() does an implicit unlock. */
13253 			sorwakeup_locked(so);
13254 		}
13255 	}
13256 	tp->snd_wl1 = th->th_seq - 1;
13257 	/* For syn-recv we need to possibly update the rtt */
13258 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13259 		uint32_t t, mcts;
13260 
13261 		mcts = tcp_ts_getticks();
13262 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13263 		if (!tp->t_rttlow || tp->t_rttlow > t)
13264 			tp->t_rttlow = t;
13265 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13266 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13267 		tcp_rack_xmit_timer_commit(rack, tp);
13268 	}
13269 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13270 		return (ret_val);
13271 	}
13272 	if (tp->t_state == TCPS_FIN_WAIT_1) {
13273 		/* We could have went to FIN_WAIT_1 (or EST) above */
13274 		/*
13275 		 * In FIN_WAIT_1 STATE in addition to the processing for the
13276 		 * ESTABLISHED state if our FIN is now acknowledged then
13277 		 * enter FIN_WAIT_2.
13278 		 */
13279 		if (ourfinisacked) {
13280 			/*
13281 			 * If we can't receive any more data, then closing
13282 			 * user can proceed. Starting the timer is contrary
13283 			 * to the specification, but if we don't get a FIN
13284 			 * we'll hang forever.
13285 			 *
13286 			 * XXXjl: we should release the tp also, and use a
13287 			 * compressed state.
13288 			 */
13289 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13290 				soisdisconnected(so);
13291 				tcp_timer_activate(tp, TT_2MSL,
13292 				    (tcp_fast_finwait2_recycle ?
13293 				    tcp_finwait2_timeout :
13294 				    TP_MAXIDLE(tp)));
13295 			}
13296 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
13297 		}
13298 	}
13299 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13300 	    tiwin, thflags, nxt_pkt));
13301 }
13302 
13303 /*
13304  * Return value of 1, the TCB is unlocked and most
13305  * likely gone, return value of 0, the TCP is still
13306  * locked.
13307  */
13308 static int
13309 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13310     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13311     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13312 {
13313 	int32_t ret_val = 0;
13314 	int32_t orig_tlen = tlen;
13315 	struct tcp_rack *rack;
13316 
13317 	/*
13318 	 * Header prediction: check for the two common cases of a
13319 	 * uni-directional data xfer.  If the packet has no control flags,
13320 	 * is in-sequence, the window didn't change and we're not
13321 	 * retransmitting, it's a candidate.  If the length is zero and the
13322 	 * ack moved forward, we're the sender side of the xfer.  Just free
13323 	 * the data acked & wake any higher level process that was blocked
13324 	 * waiting for space.  If the length is non-zero and the ack didn't
13325 	 * move, we're the receiver side.  If we're getting packets in-order
13326 	 * (the reassembly queue is empty), add the data toc The socket
13327 	 * buffer and note that we need a delayed ack. Make sure that the
13328 	 * hidden state-flags are also off. Since we check for
13329 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13330 	 */
13331 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13332 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13333 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13334 	    __predict_true(SEGQ_EMPTY(tp)) &&
13335 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
13336 		if (tlen == 0) {
13337 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13338 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13339 				return (0);
13340 			}
13341 		} else {
13342 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13343 			    tiwin, nxt_pkt, iptos)) {
13344 				return (0);
13345 			}
13346 		}
13347 	}
13348 	ctf_calc_rwin(so, tp);
13349 
13350 	if ((thflags & TH_RST) ||
13351 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13352 		return (ctf_process_rst(m, th, so, tp));
13353 
13354 	/*
13355 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13356 	 * synchronized state.
13357 	 */
13358 	if (thflags & TH_SYN) {
13359 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13360 		return (ret_val);
13361 	}
13362 	/*
13363 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13364 	 * it's less than ts_recent, drop it.
13365 	 */
13366 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13367 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13368 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13369 			return (ret_val);
13370 	}
13371 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13372 		return (ret_val);
13373 	}
13374 	/*
13375 	 * If last ACK falls within this segment's sequence numbers, record
13376 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13377 	 * from the latest proposal of the tcplw@cray.com list (Braden
13378 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13379 	 * with our earlier PAWS tests, so this check should be solely
13380 	 * predicated on the sequence space of this segment. 3) That we
13381 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13382 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13383 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13384 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13385 	 * p.869. In such cases, we can still calculate the RTT correctly
13386 	 * when RCV.NXT == Last.ACK.Sent.
13387 	 */
13388 	if ((to->to_flags & TOF_TS) != 0 &&
13389 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13390 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13391 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13392 		tp->ts_recent_age = tcp_ts_getticks();
13393 		tp->ts_recent = to->to_tsval;
13394 	}
13395 	/*
13396 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13397 	 * is on (half-synchronized state), then queue data for later
13398 	 * processing; else drop segment and return.
13399 	 */
13400 	if ((thflags & TH_ACK) == 0) {
13401 		if (tp->t_flags & TF_NEEDSYN) {
13402 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13403 			    tiwin, thflags, nxt_pkt));
13404 
13405 		} else if (tp->t_flags & TF_ACKNOW) {
13406 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13407 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13408 			return (ret_val);
13409 		} else {
13410 			ctf_do_drop(m, NULL);
13411 			return (0);
13412 		}
13413 	}
13414 	/*
13415 	 * Ack processing.
13416 	 */
13417 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13418 		return (ret_val);
13419 	}
13420 	if (sbavail(&so->so_snd)) {
13421 		if (ctf_progress_timeout_check(tp, true)) {
13422 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13423 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13424 			return (1);
13425 		}
13426 	}
13427 	/* State changes only happen in rack_process_data() */
13428 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13429 	    tiwin, thflags, nxt_pkt));
13430 }
13431 
13432 /*
13433  * Return value of 1, the TCB is unlocked and most
13434  * likely gone, return value of 0, the TCP is still
13435  * locked.
13436  */
13437 static int
13438 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13439     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13440     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13441 {
13442 	int32_t ret_val = 0;
13443 	int32_t orig_tlen = tlen;
13444 
13445 	ctf_calc_rwin(so, tp);
13446 	if ((thflags & TH_RST) ||
13447 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13448 		return (ctf_process_rst(m, th, so, tp));
13449 	/*
13450 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13451 	 * synchronized state.
13452 	 */
13453 	if (thflags & TH_SYN) {
13454 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13455 		return (ret_val);
13456 	}
13457 	/*
13458 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13459 	 * it's less than ts_recent, drop it.
13460 	 */
13461 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13462 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13463 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13464 			return (ret_val);
13465 	}
13466 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13467 		return (ret_val);
13468 	}
13469 	/*
13470 	 * If last ACK falls within this segment's sequence numbers, record
13471 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13472 	 * from the latest proposal of the tcplw@cray.com list (Braden
13473 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13474 	 * with our earlier PAWS tests, so this check should be solely
13475 	 * predicated on the sequence space of this segment. 3) That we
13476 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13477 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13478 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13479 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13480 	 * p.869. In such cases, we can still calculate the RTT correctly
13481 	 * when RCV.NXT == Last.ACK.Sent.
13482 	 */
13483 	if ((to->to_flags & TOF_TS) != 0 &&
13484 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13485 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13486 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13487 		tp->ts_recent_age = tcp_ts_getticks();
13488 		tp->ts_recent = to->to_tsval;
13489 	}
13490 	/*
13491 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13492 	 * is on (half-synchronized state), then queue data for later
13493 	 * processing; else drop segment and return.
13494 	 */
13495 	if ((thflags & TH_ACK) == 0) {
13496 		if (tp->t_flags & TF_NEEDSYN) {
13497 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13498 			    tiwin, thflags, nxt_pkt));
13499 
13500 		} else if (tp->t_flags & TF_ACKNOW) {
13501 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13502 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13503 			return (ret_val);
13504 		} else {
13505 			ctf_do_drop(m, NULL);
13506 			return (0);
13507 		}
13508 	}
13509 	/*
13510 	 * Ack processing.
13511 	 */
13512 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13513 		return (ret_val);
13514 	}
13515 	if (sbavail(&so->so_snd)) {
13516 		if (ctf_progress_timeout_check(tp, true)) {
13517 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13518 						tp, tick, PROGRESS_DROP, __LINE__);
13519 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13520 			return (1);
13521 		}
13522 	}
13523 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13524 	    tiwin, thflags, nxt_pkt));
13525 }
13526 
13527 static int
13528 rack_check_data_after_close(struct mbuf *m,
13529     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13530 {
13531 	struct tcp_rack *rack;
13532 
13533 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13534 	if (rack->rc_allow_data_af_clo == 0) {
13535 	close_now:
13536 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13537 		/* tcp_close will kill the inp pre-log the Reset */
13538 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13539 		tp = tcp_close(tp);
13540 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13541 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
13542 		return (1);
13543 	}
13544 	if (sbavail(&so->so_snd) == 0)
13545 		goto close_now;
13546 	/* Ok we allow data that is ignored and a followup reset */
13547 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13548 	tp->rcv_nxt = th->th_seq + *tlen;
13549 	tp->t_flags2 |= TF2_DROP_AF_DATA;
13550 	rack->r_wanted_output = 1;
13551 	*tlen = 0;
13552 	return (0);
13553 }
13554 
13555 /*
13556  * Return value of 1, the TCB is unlocked and most
13557  * likely gone, return value of 0, the TCP is still
13558  * locked.
13559  */
13560 static int
13561 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13562     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13563     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13564 {
13565 	int32_t ret_val = 0;
13566 	int32_t orig_tlen = tlen;
13567 	int32_t ourfinisacked = 0;
13568 
13569 	ctf_calc_rwin(so, tp);
13570 
13571 	if ((thflags & TH_RST) ||
13572 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13573 		return (ctf_process_rst(m, th, so, tp));
13574 	/*
13575 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13576 	 * synchronized state.
13577 	 */
13578 	if (thflags & TH_SYN) {
13579 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13580 		return (ret_val);
13581 	}
13582 	/*
13583 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13584 	 * it's less than ts_recent, drop it.
13585 	 */
13586 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13587 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13588 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13589 			return (ret_val);
13590 	}
13591 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13592 		return (ret_val);
13593 	}
13594 	/*
13595 	 * If new data are received on a connection after the user processes
13596 	 * are gone, then RST the other end.
13597 	 */
13598 	if ((tp->t_flags & TF_CLOSED) && tlen &&
13599 	    rack_check_data_after_close(m, tp, &tlen, th, so))
13600 		return (1);
13601 	/*
13602 	 * If last ACK falls within this segment's sequence numbers, record
13603 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13604 	 * from the latest proposal of the tcplw@cray.com list (Braden
13605 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13606 	 * with our earlier PAWS tests, so this check should be solely
13607 	 * predicated on the sequence space of this segment. 3) That we
13608 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13609 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13610 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13611 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13612 	 * p.869. In such cases, we can still calculate the RTT correctly
13613 	 * when RCV.NXT == Last.ACK.Sent.
13614 	 */
13615 	if ((to->to_flags & TOF_TS) != 0 &&
13616 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13617 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13618 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13619 		tp->ts_recent_age = tcp_ts_getticks();
13620 		tp->ts_recent = to->to_tsval;
13621 	}
13622 	/*
13623 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13624 	 * is on (half-synchronized state), then queue data for later
13625 	 * processing; else drop segment and return.
13626 	 */
13627 	if ((thflags & TH_ACK) == 0) {
13628 		if (tp->t_flags & TF_NEEDSYN) {
13629 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13630 			    tiwin, thflags, nxt_pkt));
13631 		} else if (tp->t_flags & TF_ACKNOW) {
13632 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13633 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13634 			return (ret_val);
13635 		} else {
13636 			ctf_do_drop(m, NULL);
13637 			return (0);
13638 		}
13639 	}
13640 	/*
13641 	 * Ack processing.
13642 	 */
13643 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13644 		return (ret_val);
13645 	}
13646 	if (ourfinisacked) {
13647 		/*
13648 		 * If we can't receive any more data, then closing user can
13649 		 * proceed. Starting the timer is contrary to the
13650 		 * specification, but if we don't get a FIN we'll hang
13651 		 * forever.
13652 		 *
13653 		 * XXXjl: we should release the tp also, and use a
13654 		 * compressed state.
13655 		 */
13656 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13657 			soisdisconnected(so);
13658 			tcp_timer_activate(tp, TT_2MSL,
13659 			    (tcp_fast_finwait2_recycle ?
13660 			    tcp_finwait2_timeout :
13661 			    TP_MAXIDLE(tp)));
13662 		}
13663 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
13664 	}
13665 	if (sbavail(&so->so_snd)) {
13666 		if (ctf_progress_timeout_check(tp, true)) {
13667 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13668 						tp, tick, PROGRESS_DROP, __LINE__);
13669 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13670 			return (1);
13671 		}
13672 	}
13673 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13674 	    tiwin, thflags, nxt_pkt));
13675 }
13676 
13677 /*
13678  * Return value of 1, the TCB is unlocked and most
13679  * likely gone, return value of 0, the TCP is still
13680  * locked.
13681  */
13682 static int
13683 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
13684     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13685     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13686 {
13687 	int32_t ret_val = 0;
13688 	int32_t orig_tlen = tlen;
13689 	int32_t ourfinisacked = 0;
13690 
13691 	ctf_calc_rwin(so, tp);
13692 
13693 	if ((thflags & TH_RST) ||
13694 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13695 		return (ctf_process_rst(m, th, so, tp));
13696 	/*
13697 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13698 	 * synchronized state.
13699 	 */
13700 	if (thflags & TH_SYN) {
13701 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13702 		return (ret_val);
13703 	}
13704 	/*
13705 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13706 	 * it's less than ts_recent, drop it.
13707 	 */
13708 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13709 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13710 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13711 			return (ret_val);
13712 	}
13713 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13714 		return (ret_val);
13715 	}
13716 	/*
13717 	 * If last ACK falls within this segment's sequence numbers, record
13718 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13719 	 * from the latest proposal of the tcplw@cray.com list (Braden
13720 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13721 	 * with our earlier PAWS tests, so this check should be solely
13722 	 * predicated on the sequence space of this segment. 3) That we
13723 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13724 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13725 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13726 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13727 	 * p.869. In such cases, we can still calculate the RTT correctly
13728 	 * when RCV.NXT == Last.ACK.Sent.
13729 	 */
13730 	if ((to->to_flags & TOF_TS) != 0 &&
13731 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13732 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13733 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13734 		tp->ts_recent_age = tcp_ts_getticks();
13735 		tp->ts_recent = to->to_tsval;
13736 	}
13737 	/*
13738 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13739 	 * is on (half-synchronized state), then queue data for later
13740 	 * processing; else drop segment and return.
13741 	 */
13742 	if ((thflags & TH_ACK) == 0) {
13743 		if (tp->t_flags & TF_NEEDSYN) {
13744 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13745 			    tiwin, thflags, nxt_pkt));
13746 		} else if (tp->t_flags & TF_ACKNOW) {
13747 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13748 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13749 			return (ret_val);
13750 		} else {
13751 			ctf_do_drop(m, NULL);
13752 			return (0);
13753 		}
13754 	}
13755 	/*
13756 	 * Ack processing.
13757 	 */
13758 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13759 		return (ret_val);
13760 	}
13761 	if (ourfinisacked) {
13762 		tcp_twstart(tp);
13763 		m_freem(m);
13764 		return (1);
13765 	}
13766 	if (sbavail(&so->so_snd)) {
13767 		if (ctf_progress_timeout_check(tp, true)) {
13768 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13769 						tp, tick, PROGRESS_DROP, __LINE__);
13770 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13771 			return (1);
13772 		}
13773 	}
13774 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13775 	    tiwin, thflags, nxt_pkt));
13776 }
13777 
13778 /*
13779  * Return value of 1, the TCB is unlocked and most
13780  * likely gone, return value of 0, the TCP is still
13781  * locked.
13782  */
13783 static int
13784 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13785     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13786     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13787 {
13788 	int32_t ret_val = 0;
13789 	int32_t orig_tlen;
13790 	int32_t ourfinisacked = 0;
13791 
13792 	ctf_calc_rwin(so, tp);
13793 
13794 	if ((thflags & TH_RST) ||
13795 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13796 		return (ctf_process_rst(m, th, so, tp));
13797 	/*
13798 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13799 	 * synchronized state.
13800 	 */
13801 	if (thflags & TH_SYN) {
13802 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13803 		return (ret_val);
13804 	}
13805 	/*
13806 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13807 	 * it's less than ts_recent, drop it.
13808 	 */
13809 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13810 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13811 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13812 			return (ret_val);
13813 	}
13814 	orig_tlen = tlen;
13815 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13816 		return (ret_val);
13817 	}
13818 	/*
13819 	 * If last ACK falls within this segment's sequence numbers, record
13820 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13821 	 * from the latest proposal of the tcplw@cray.com list (Braden
13822 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13823 	 * with our earlier PAWS tests, so this check should be solely
13824 	 * predicated on the sequence space of this segment. 3) That we
13825 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13826 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13827 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13828 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13829 	 * p.869. In such cases, we can still calculate the RTT correctly
13830 	 * when RCV.NXT == Last.ACK.Sent.
13831 	 */
13832 	if ((to->to_flags & TOF_TS) != 0 &&
13833 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13834 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13835 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13836 		tp->ts_recent_age = tcp_ts_getticks();
13837 		tp->ts_recent = to->to_tsval;
13838 	}
13839 	/*
13840 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13841 	 * is on (half-synchronized state), then queue data for later
13842 	 * processing; else drop segment and return.
13843 	 */
13844 	if ((thflags & TH_ACK) == 0) {
13845 		if (tp->t_flags & TF_NEEDSYN) {
13846 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13847 			    tiwin, thflags, nxt_pkt));
13848 		} else if (tp->t_flags & TF_ACKNOW) {
13849 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13850 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13851 			return (ret_val);
13852 		} else {
13853 			ctf_do_drop(m, NULL);
13854 			return (0);
13855 		}
13856 	}
13857 	/*
13858 	 * case TCPS_LAST_ACK: Ack processing.
13859 	 */
13860 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13861 		return (ret_val);
13862 	}
13863 	if (ourfinisacked) {
13864 		tp = tcp_close(tp);
13865 		ctf_do_drop(m, tp);
13866 		return (1);
13867 	}
13868 	if (sbavail(&so->so_snd)) {
13869 		if (ctf_progress_timeout_check(tp, true)) {
13870 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13871 						tp, tick, PROGRESS_DROP, __LINE__);
13872 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13873 			return (1);
13874 		}
13875 	}
13876 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13877 	    tiwin, thflags, nxt_pkt));
13878 }
13879 
13880 /*
13881  * Return value of 1, the TCB is unlocked and most
13882  * likely gone, return value of 0, the TCP is still
13883  * locked.
13884  */
13885 static int
13886 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
13887     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13888     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13889 {
13890 	int32_t ret_val = 0;
13891 	int32_t orig_tlen = tlen;
13892 	int32_t ourfinisacked = 0;
13893 
13894 	ctf_calc_rwin(so, tp);
13895 
13896 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
13897 	if ((thflags & TH_RST) ||
13898 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13899 		return (ctf_process_rst(m, th, so, tp));
13900 	/*
13901 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13902 	 * synchronized state.
13903 	 */
13904 	if (thflags & TH_SYN) {
13905 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13906 		return (ret_val);
13907 	}
13908 	/*
13909 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13910 	 * it's less than ts_recent, drop it.
13911 	 */
13912 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13913 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13914 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13915 			return (ret_val);
13916 	}
13917 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13918 		return (ret_val);
13919 	}
13920 	/*
13921 	 * If new data are received on a connection after the user processes
13922 	 * are gone, then RST the other end.
13923 	 */
13924 	if ((tp->t_flags & TF_CLOSED) && tlen &&
13925 	    rack_check_data_after_close(m, tp, &tlen, th, so))
13926 		return (1);
13927 	/*
13928 	 * If last ACK falls within this segment's sequence numbers, record
13929 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13930 	 * from the latest proposal of the tcplw@cray.com list (Braden
13931 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13932 	 * with our earlier PAWS tests, so this check should be solely
13933 	 * predicated on the sequence space of this segment. 3) That we
13934 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13935 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13936 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13937 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13938 	 * p.869. In such cases, we can still calculate the RTT correctly
13939 	 * when RCV.NXT == Last.ACK.Sent.
13940 	 */
13941 	if ((to->to_flags & TOF_TS) != 0 &&
13942 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13943 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13944 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13945 		tp->ts_recent_age = tcp_ts_getticks();
13946 		tp->ts_recent = to->to_tsval;
13947 	}
13948 	/*
13949 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13950 	 * is on (half-synchronized state), then queue data for later
13951 	 * processing; else drop segment and return.
13952 	 */
13953 	if ((thflags & TH_ACK) == 0) {
13954 		if (tp->t_flags & TF_NEEDSYN) {
13955 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13956 			    tiwin, thflags, nxt_pkt));
13957 		} else if (tp->t_flags & TF_ACKNOW) {
13958 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13959 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13960 			return (ret_val);
13961 		} else {
13962 			ctf_do_drop(m, NULL);
13963 			return (0);
13964 		}
13965 	}
13966 	/*
13967 	 * Ack processing.
13968 	 */
13969 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13970 		return (ret_val);
13971 	}
13972 	if (sbavail(&so->so_snd)) {
13973 		if (ctf_progress_timeout_check(tp, true)) {
13974 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13975 						tp, tick, PROGRESS_DROP, __LINE__);
13976 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13977 			return (1);
13978 		}
13979 	}
13980 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13981 	    tiwin, thflags, nxt_pkt));
13982 }
13983 
13984 static void inline
13985 rack_clear_rate_sample(struct tcp_rack *rack)
13986 {
13987 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
13988 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
13989 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
13990 }
13991 
13992 static void
13993 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
13994 {
13995 	uint64_t bw_est, rate_wanted;
13996 	int chged = 0;
13997 	uint32_t user_max, orig_min, orig_max;
13998 
13999 #ifdef TCP_REQUEST_TRK
14000 	if (rack->rc_hybrid_mode &&
14001 	    (rack->r_ctl.rc_pace_max_segs != 0) &&
14002 	    (rack_hybrid_allow_set_maxseg == 1) &&
14003 	    (rack->r_ctl.rc_last_sft != NULL)) {
14004 		rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
14005 		return;
14006 	}
14007 #endif
14008 	orig_min = rack->r_ctl.rc_pace_min_segs;
14009 	orig_max = rack->r_ctl.rc_pace_max_segs;
14010 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
14011 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
14012 		chged = 1;
14013 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
14014 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
14015 		if (user_max != rack->r_ctl.rc_pace_max_segs)
14016 			chged = 1;
14017 	}
14018 	if (rack->rc_force_max_seg) {
14019 		rack->r_ctl.rc_pace_max_segs = user_max;
14020 	} else if (rack->use_fixed_rate) {
14021 		bw_est = rack_get_bw(rack);
14022 		if ((rack->r_ctl.crte == NULL) ||
14023 		    (bw_est != rack->r_ctl.crte->rate)) {
14024 			rack->r_ctl.rc_pace_max_segs = user_max;
14025 		} else {
14026 			/* We are pacing right at the hardware rate */
14027 			uint32_t segsiz, pace_one;
14028 
14029 			if (rack_pace_one_seg ||
14030 			    (rack->r_ctl.rc_user_set_min_segs == 1))
14031 				pace_one = 1;
14032 			else
14033 				pace_one = 0;
14034 			segsiz = min(ctf_fixed_maxseg(tp),
14035 				     rack->r_ctl.rc_pace_min_segs);
14036 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14037 				tp, bw_est, segsiz, pace_one,
14038 				rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14039 		}
14040 	} else if (rack->rc_always_pace) {
14041 		if (rack->r_ctl.gp_bw ||
14042 		    rack->r_ctl.init_rate) {
14043 			/* We have a rate of some sort set */
14044 			uint32_t  orig;
14045 
14046 			bw_est = rack_get_bw(rack);
14047 			orig = rack->r_ctl.rc_pace_max_segs;
14048 			if (fill_override)
14049 				rate_wanted = *fill_override;
14050 			else
14051 				rate_wanted = rack_get_gp_est(rack);
14052 			if (rate_wanted) {
14053 				/* We have something */
14054 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14055 										   rate_wanted,
14056 										   ctf_fixed_maxseg(rack->rc_tp));
14057 			} else
14058 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14059 			if (orig != rack->r_ctl.rc_pace_max_segs)
14060 				chged = 1;
14061 		} else if ((rack->r_ctl.gp_bw == 0) &&
14062 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
14063 			/*
14064 			 * If we have nothing limit us to bursting
14065 			 * out IW sized pieces.
14066 			 */
14067 			chged = 1;
14068 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14069 		}
14070 	}
14071 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14072 		chged = 1;
14073 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14074 	}
14075 	if (chged)
14076 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14077 }
14078 
14079 
14080 static void
14081 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14082 {
14083 #ifdef INET6
14084 	struct ip6_hdr *ip6 = NULL;
14085 #endif
14086 #ifdef INET
14087 	struct ip *ip = NULL;
14088 #endif
14089 	struct udphdr *udp = NULL;
14090 
14091 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
14092 #ifdef INET6
14093 	if (rack->r_is_v6) {
14094 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14095 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14096 		if (tp->t_port) {
14097 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14098 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14099 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14100 			udp->uh_dport = tp->t_port;
14101 			rack->r_ctl.fsb.udp = udp;
14102 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14103 		} else
14104 		{
14105 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14106 			rack->r_ctl.fsb.udp = NULL;
14107 		}
14108 		tcpip_fillheaders(rack->rc_inp,
14109 				  tp->t_port,
14110 				  ip6, rack->r_ctl.fsb.th);
14111 		rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14112 	} else
14113 #endif				/* INET6 */
14114 #ifdef INET
14115 	{
14116 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14117 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14118 		if (tp->t_port) {
14119 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14120 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14121 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14122 			udp->uh_dport = tp->t_port;
14123 			rack->r_ctl.fsb.udp = udp;
14124 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14125 		} else
14126 		{
14127 			rack->r_ctl.fsb.udp = NULL;
14128 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14129 		}
14130 		tcpip_fillheaders(rack->rc_inp,
14131 				  tp->t_port,
14132 				  ip, rack->r_ctl.fsb.th);
14133 		rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14134 	}
14135 #endif
14136 	rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14137 	    (long)TCP_MAXWIN << tp->rcv_scale);
14138 	rack->r_fsb_inited = 1;
14139 }
14140 
14141 static int
14142 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14143 {
14144 	/*
14145 	 * Allocate the larger of spaces V6 if available else just
14146 	 * V4 and include udphdr (overbook)
14147 	 */
14148 #ifdef INET6
14149 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14150 #else
14151 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14152 #endif
14153 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14154 					    M_TCPFSB, M_NOWAIT|M_ZERO);
14155 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14156 		return (ENOMEM);
14157 	}
14158 	rack->r_fsb_inited = 0;
14159 	return (0);
14160 }
14161 
14162 static void
14163 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14164 {
14165 	/*
14166 	 * Types of logs (mod value)
14167 	 * 20 - Initial round setup
14168 	 * 21 - Rack declares a new round.
14169 	 */
14170 	struct tcpcb *tp;
14171 
14172 	tp = rack->rc_tp;
14173 	if (tcp_bblogging_on(tp)) {
14174 		union tcp_log_stackspecific log;
14175 		struct timeval tv;
14176 
14177 		memset(&log, 0, sizeof(log));
14178 		log.u_bbr.flex1 = rack->r_ctl.current_round;
14179 		log.u_bbr.flex2 = rack->r_ctl.roundends;
14180 		log.u_bbr.flex3 = high_seq;
14181 		log.u_bbr.flex4 = tp->snd_max;
14182 		log.u_bbr.flex8 = mod;
14183 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14184 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14185 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14186 		TCP_LOG_EVENTP(tp, NULL,
14187 		    &tptosocket(tp)->so_rcv,
14188 		    &tptosocket(tp)->so_snd,
14189 		    TCP_HYSTART, 0,
14190 		    0, &log, false, &tv);
14191 	}
14192 }
14193 
14194 static void
14195 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14196 {
14197 	rack->rack_deferred_inited = 1;
14198 	rack->r_ctl.roundends = tp->snd_max;
14199 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14200 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14201 }
14202 
14203 static void
14204 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14205 {
14206 	/* Retransmit bit controls.
14207 	 *
14208 	 * The setting of these values control one of
14209 	 * three settings you can have and dictate
14210 	 * how rack does retransmissions. Note this
14211 	 * is in *any* mode i.e. pacing on or off DGP
14212 	 * fixed rate pacing, or just bursting rack.
14213 	 *
14214 	 * 1 - Use full sized retransmits i.e. limit
14215 	 *     the size to whatever the pace_max_segments
14216 	 *     size is.
14217 	 *
14218 	 * 2 - Use pacer min granularity as a guide to
14219 	 *     the size combined with the current calculated
14220 	 *     goodput b/w measurement. So for example if
14221 	 *     the goodput is measured at 20Mbps we would
14222 	 *     calculate 8125 (pacer minimum 250usec in
14223 	 *     that b/w) and then round it up to the next
14224 	 *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14225 	 *
14226 	 * 0 - The rack default 1 MSS (anything not 0/1/2
14227 	 *     fall here too if we are setting via rack_init()).
14228 	 *
14229 	 */
14230 	if (ctl == 1) {
14231 		rack->full_size_rxt = 1;
14232 		rack->shape_rxt_to_pacing_min  = 0;
14233 	} else if (ctl == 2) {
14234 		rack->full_size_rxt = 0;
14235 		rack->shape_rxt_to_pacing_min  = 1;
14236 	} else {
14237 		rack->full_size_rxt = 0;
14238 		rack->shape_rxt_to_pacing_min  = 0;
14239 	}
14240 }
14241 
14242 static void
14243 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14244 		  uint32_t flex1,
14245 		  uint32_t flex2,
14246 		  uint32_t flex3)
14247 {
14248 	if (tcp_bblogging_on(rack->rc_tp)) {
14249 		union tcp_log_stackspecific log;
14250 		struct timeval tv;
14251 
14252 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14253 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14254 		log.u_bbr.flex8 = mod;
14255 		log.u_bbr.flex1 = flex1;
14256 		log.u_bbr.flex2 = flex2;
14257 		log.u_bbr.flex3 = flex3;
14258 		tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14259 			       0, &log, false, NULL, __func__, __LINE__, &tv);
14260 	}
14261 }
14262 
14263 static int
14264 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14265 {
14266 	struct tcp_rack *rack;
14267 	struct rack_sendmap *rsm;
14268 	int i;
14269 
14270 
14271 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14272 	switch (reqr->req) {
14273 	case TCP_QUERY_SENDMAP:
14274 		if ((reqr->req_param == tp->snd_max) ||
14275 		    (tp->snd_max == tp->snd_una)){
14276 			/* Unlikely */
14277 			return (0);
14278 		}
14279 		rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14280 		if (rsm == NULL) {
14281 			/* Can't find that seq -- unlikely */
14282 			return (0);
14283 		}
14284 		reqr->sendmap_start = rsm->r_start;
14285 		reqr->sendmap_end = rsm->r_end;
14286 		reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14287 		reqr->sendmap_fas = rsm->r_fas;
14288 		if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14289 			reqr->sendmap_send_cnt = SNDMAP_NRTX;
14290 		for(i=0; i<reqr->sendmap_send_cnt; i++)
14291 			reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14292 		reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14293 		reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14294 		reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14295 		reqr->sendmap_dupacks = rsm->r_dupack;
14296 		rack_log_chg_info(tp, rack, 1,
14297 				  rsm->r_start,
14298 				  rsm->r_end,
14299 				  rsm->r_flags);
14300 		return(1);
14301 		break;
14302 	case TCP_QUERY_TIMERS_UP:
14303 		if (rack->r_ctl.rc_hpts_flags == 0) {
14304 			/* no timers up */
14305 			return (0);
14306 		}
14307 		reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14308 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14309 			reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14310 		}
14311 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14312 			reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14313 		}
14314 		rack_log_chg_info(tp, rack, 2,
14315 				  rack->r_ctl.rc_hpts_flags,
14316 				  rack->r_ctl.rc_last_output_to,
14317 				  rack->r_ctl.rc_timer_exp);
14318 		return (1);
14319 		break;
14320 	case TCP_QUERY_RACK_TIMES:
14321 		/* Reordering items */
14322 		reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14323 		reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14324 		/* Timerstamps and timers */
14325 		reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14326 		reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14327 		reqr->rack_rtt = rack->rc_rack_rtt;
14328 		reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14329 		reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14330 		/* PRR data */
14331 		reqr->rack_sacked = rack->r_ctl.rc_sacked;
14332 		reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14333 		reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14334 		reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14335 		reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14336 		reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14337 		/* TLP and persists info */
14338 		reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14339 		reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14340 		if (rack->rc_in_persist) {
14341 			reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14342 			reqr->rack_in_persist = 1;
14343 		} else {
14344 			reqr->rack_time_went_idle = 0;
14345 			reqr->rack_in_persist = 0;
14346 		}
14347 		if (rack->r_wanted_output)
14348 			reqr->rack_wanted_output = 1;
14349 		else
14350 			reqr->rack_wanted_output = 0;
14351 		return (1);
14352 		break;
14353 	default:
14354 		return (-EINVAL);
14355 	}
14356 }
14357 
14358 static void
14359 rack_switch_failed(struct tcpcb *tp)
14360 {
14361 	/*
14362 	 * This method gets called if a stack switch was
14363 	 * attempted and it failed. We are left
14364 	 * but our hpts timers were stopped and we
14365 	 * need to validate time units and t_flags2.
14366 	 */
14367 	struct tcp_rack *rack;
14368 	struct timeval tv;
14369 	uint32_t cts;
14370 	uint32_t toval;
14371 	struct hpts_diag diag;
14372 
14373 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14374 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14375 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14376 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14377 	else
14378 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14379 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14380 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
14381 	if (tp->t_in_hpts > IHPTS_NONE) {
14382 		/* Strange */
14383 		return;
14384 	}
14385 	cts = tcp_get_usecs(&tv);
14386 	if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14387 		if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14388 			toval = rack->r_ctl.rc_last_output_to - cts;
14389 		} else {
14390 			/* one slot please */
14391 			toval = HPTS_TICKS_PER_SLOT;
14392 		}
14393 	} else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14394 		if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14395 			toval = rack->r_ctl.rc_timer_exp - cts;
14396 		} else {
14397 			/* one slot please */
14398 			toval = HPTS_TICKS_PER_SLOT;
14399 		}
14400 	} else
14401 		toval = HPTS_TICKS_PER_SLOT;
14402 	(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14403 				   __LINE__, &diag);
14404 	rack_log_hpts_diag(rack, cts, &diag, &tv);
14405 }
14406 
14407 static int
14408 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14409 {
14410 	struct rack_sendmap *rsm, *ersm;
14411 	int insret __diagused;
14412 	/*
14413 	 * When initing outstanding, we must be quite careful
14414 	 * to not refer to tp->t_fb_ptr. This has the old rack
14415 	 * pointer in it, not the "new" one (when we are doing
14416 	 * a stack switch).
14417 	 */
14418 
14419 
14420 	if (tp->t_fb->tfb_chg_query == NULL) {
14421 		/* Create a send map for the current outstanding data */
14422 
14423 		rsm = rack_alloc(rack);
14424 		if (rsm == NULL) {
14425 			uma_zfree(rack_pcb_zone, ptr);
14426 			return (ENOMEM);
14427 		}
14428 		rsm->r_no_rtt_allowed = 1;
14429 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14430 		rsm->r_rtr_cnt = 1;
14431 		rsm->r_rtr_bytes = 0;
14432 		if (tp->t_flags & TF_SENTFIN)
14433 			rsm->r_flags |= RACK_HAS_FIN;
14434 		rsm->r_end = tp->snd_max;
14435 		if (tp->snd_una == tp->iss) {
14436 			/* The data space is one beyond snd_una */
14437 			rsm->r_flags |= RACK_HAS_SYN;
14438 			rsm->r_start = tp->iss;
14439 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14440 		} else
14441 			rsm->r_start = tp->snd_una;
14442 		rsm->r_dupack = 0;
14443 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14444 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14445 			if (rsm->m) {
14446 				rsm->orig_m_len = rsm->m->m_len;
14447 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14448 			} else {
14449 				rsm->orig_m_len = 0;
14450 				rsm->orig_t_space = 0;
14451 			}
14452 		} else {
14453 			/*
14454 			 * This can happen if we have a stand-alone FIN or
14455 			 *  SYN.
14456 			 */
14457 			rsm->m = NULL;
14458 			rsm->orig_m_len = 0;
14459 			rsm->orig_t_space = 0;
14460 			rsm->soff = 0;
14461 		}
14462 #ifdef INVARIANTS
14463 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14464 			panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14465 			      insret, rack, rsm);
14466 		}
14467 #else
14468 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14469 #endif
14470 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14471 		rsm->r_in_tmap = 1;
14472 	} else {
14473 		/* We have a query mechanism, lets use it */
14474 		struct tcp_query_resp qr;
14475 		int i;
14476 		tcp_seq at;
14477 
14478 		at = tp->snd_una;
14479 		while (at != tp->snd_max) {
14480 			memset(&qr, 0, sizeof(qr));
14481 			qr.req = TCP_QUERY_SENDMAP;
14482 			qr.req_param = at;
14483 			if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14484 				break;
14485 			/* Move forward */
14486 			at = qr.sendmap_end;
14487 			/* Now lets build the entry for this one */
14488 			rsm = rack_alloc(rack);
14489 			if (rsm == NULL) {
14490 				uma_zfree(rack_pcb_zone, ptr);
14491 				return (ENOMEM);
14492 			}
14493 			memset(rsm, 0, sizeof(struct rack_sendmap));
14494 			/* Now configure the rsm and insert it */
14495 			rsm->r_dupack = qr.sendmap_dupacks;
14496 			rsm->r_start = qr.sendmap_start;
14497 			rsm->r_end = qr.sendmap_end;
14498 			if (qr.sendmap_fas)
14499 				rsm->r_fas = qr.sendmap_end;
14500 			else
14501 				rsm->r_fas = rsm->r_start - tp->snd_una;
14502 			/*
14503 			 * We have carefully aligned the bits
14504 			 * so that all we have to do is copy over
14505 			 * the bits with the mask.
14506 			 */
14507 			rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14508 			rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14509 			rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14510 			rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14511 			for (i=0 ; i<rsm->r_rtr_cnt; i++)
14512 				rsm->r_tim_lastsent[i]	= qr.sendmap_time[i];
14513 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14514 					   (rsm->r_start - tp->snd_una), &rsm->soff);
14515 			if (rsm->m) {
14516 				rsm->orig_m_len = rsm->m->m_len;
14517 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14518 			} else {
14519 				rsm->orig_m_len = 0;
14520 				rsm->orig_t_space = 0;
14521 			}
14522 #ifdef INVARIANTS
14523 			if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14524 				panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14525 				      insret, rack, rsm);
14526 			}
14527 #else
14528 			(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14529 #endif
14530 			if ((rsm->r_flags & RACK_ACKED) == 0)  {
14531 				TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14532 					if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14533 					    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14534 						/*
14535 						 * If the existing ersm was sent at
14536 						 * a later time than the new one, then
14537 						 * the new one should appear ahead of this
14538 						 * ersm.
14539 						 */
14540 						rsm->r_in_tmap = 1;
14541 						TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14542 						break;
14543 					}
14544 				}
14545 				if (rsm->r_in_tmap == 0) {
14546 					/*
14547 					 * Not found so shove it on the tail.
14548 					 */
14549 					TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14550 					rsm->r_in_tmap = 1;
14551 				}
14552  			} else {
14553 				if ((rack->r_ctl.rc_sacklast == NULL) ||
14554 				    (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14555 					rack->r_ctl.rc_sacklast = rsm;
14556 				}
14557 			}
14558 			rack_log_chg_info(tp, rack, 3,
14559 					  rsm->r_start,
14560 					  rsm->r_end,
14561 					  rsm->r_flags);
14562 		}
14563 	}
14564 	return (0);
14565 }
14566 
14567 
14568 static int32_t
14569 rack_init(struct tcpcb *tp, void **ptr)
14570 {
14571 	struct inpcb *inp = tptoinpcb(tp);
14572 	struct tcp_rack *rack = NULL;
14573 	uint32_t iwin, snt, us_cts;
14574 	size_t sz;
14575 	int err, no_query;
14576 
14577 	tcp_hpts_init(tp);
14578 
14579 	/*
14580 	 * First are we the initial or are we a switched stack?
14581 	 * If we are initing via tcp_newtcppcb the ptr passed
14582 	 * will be tp->t_fb_ptr. If its a stack switch that
14583 	 * has a previous stack we can query it will be a local
14584 	 * var that will in the end be set into t_fb_ptr.
14585 	 */
14586 	if (ptr == &tp->t_fb_ptr)
14587 		no_query = 1;
14588 	else
14589 		no_query = 0;
14590 	*ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14591 	if (*ptr == NULL) {
14592 		/*
14593 		 * We need to allocate memory but cant. The INP and INP_INFO
14594 		 * locks and they are recursive (happens during setup. So a
14595 		 * scheme to drop the locks fails :(
14596 		 *
14597 		 */
14598 		return(ENOMEM);
14599 	}
14600 	memset(*ptr, 0, sizeof(struct tcp_rack));
14601 	rack = (struct tcp_rack *)*ptr;
14602 	rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
14603 	if (rack->r_ctl.tqh == NULL) {
14604 		uma_zfree(rack_pcb_zone, rack);
14605 		return(ENOMEM);
14606 	}
14607 	tqhash_init(rack->r_ctl.tqh);
14608 	TAILQ_INIT(&rack->r_ctl.rc_free);
14609 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
14610 	rack->rc_tp = tp;
14611 	rack->rc_inp = inp;
14612 	/* Set the flag */
14613 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
14614 	/* Probably not needed but lets be sure */
14615 	rack_clear_rate_sample(rack);
14616 	/*
14617 	 * Save off the default values, socket options will poke
14618 	 * at these if pacing is not on or we have not yet
14619 	 * reached where pacing is on (gp_ready/fixed enabled).
14620 	 * When they get set into the CC module (when gp_ready
14621 	 * is enabled or we enable fixed) then we will set these
14622 	 * values into the CC and place in here the old values
14623 	 * so we have a restoral. Then we will set the flag
14624 	 * rc_pacing_cc_set. That way whenever we turn off pacing
14625 	 * or switch off this stack, we will know to go restore
14626 	 * the saved values.
14627 	 *
14628 	 * We specifically put into the beta the ecn value for pacing.
14629 	 */
14630 	rack->rc_new_rnd_needed = 1;
14631 	rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
14632 	/* We want abe like behavior as well */
14633 
14634 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
14635 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
14636 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
14637 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
14638 	if (rack_fill_cw_state)
14639 		rack->rc_pace_to_cwnd = 1;
14640 	if (rack_pacing_min_seg)
14641 		rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
14642 	if (use_rack_rr)
14643 		rack->use_rack_rr = 1;
14644 	if (rack_dnd_default) {
14645 		rack->rc_pace_dnd = 1;
14646 	}
14647 	if (V_tcp_delack_enabled)
14648 		tp->t_delayed_ack = 1;
14649 	else
14650 		tp->t_delayed_ack = 0;
14651 #ifdef TCP_ACCOUNTING
14652 	if (rack_tcp_accounting) {
14653 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
14654 	}
14655 #endif
14656 	rack->r_ctl.pcm_i.cnt_alloc = RACK_DEFAULT_PCM_ARRAY;
14657 	sz = (sizeof(struct rack_pcm_stats) * rack->r_ctl.pcm_i.cnt_alloc);
14658 	rack->r_ctl.pcm_s = malloc(sz,M_TCPPCM, M_NOWAIT);
14659 	if (rack->r_ctl.pcm_s == NULL) {
14660 		rack->r_ctl.pcm_i.cnt_alloc = 0;
14661 	}
14662 #ifdef NETFLIX_STATS
14663 	rack->r_ctl.side_chan_dis_mask = tcp_sidechannel_disable_mask;
14664 #endif
14665 	rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
14666 	rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
14667 	if (rack_enable_shared_cwnd)
14668 		rack->rack_enable_scwnd = 1;
14669 	rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
14670 	rack->rc_user_set_max_segs = rack_hptsi_segments;
14671 	rack->r_ctl.max_reduction = rack_max_reduce;
14672 	rack->rc_force_max_seg = 0;
14673 	TAILQ_INIT(&rack->r_ctl.opt_list);
14674 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
14675 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
14676 	if (rack_hibeta_setting) {
14677 		rack->rack_hibeta = 1;
14678 		if ((rack_hibeta_setting >= 50) &&
14679 		    (rack_hibeta_setting <= 100)) {
14680 			rack->r_ctl.rc_saved_beta.beta = rack_hibeta_setting;
14681 			rack->r_ctl.saved_hibeta = rack_hibeta_setting;
14682 		}
14683 	} else {
14684 		rack->r_ctl.saved_hibeta = 50;
14685 	}
14686 	/*
14687 	 * We initialize to all ones so we never match 0
14688 	 * just in case the client sends in 0, it hopefully
14689 	 * will never have all 1's in ms :-)
14690 	 */
14691 	rack->r_ctl.last_tm_mark = 0xffffffffffffffff;
14692 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
14693 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
14694 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
14695 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
14696 	rack->r_ctl.rc_highest_us_rtt = 0;
14697 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
14698 	rack->pcm_enabled = rack_pcm_is_enabled;
14699 	if (rack_fillcw_bw_cap)
14700 		rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
14701 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
14702 	if (rack_use_cmp_acks)
14703 		rack->r_use_cmp_ack = 1;
14704 	if (rack_disable_prr)
14705 		rack->rack_no_prr = 1;
14706 	if (rack_gp_no_rec_chg)
14707 		rack->rc_gp_no_rec_chg = 1;
14708 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
14709 		rack->r_ctl.pacing_method |= RACK_REG_PACING;
14710 		rack->rc_always_pace = 1;
14711 		if (rack->rack_hibeta)
14712 			rack_set_cc_pacing(rack);
14713 	} else
14714 		rack->rc_always_pace = 0;
14715 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
14716 		rack->r_mbuf_queue = 1;
14717 	else
14718 		rack->r_mbuf_queue = 0;
14719 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
14720 	if (rack_limits_scwnd)
14721 		rack->r_limit_scw = 1;
14722 	else
14723 		rack->r_limit_scw = 0;
14724 	rack_init_retransmit_value(rack, rack_rxt_controls);
14725 	rack->rc_labc = V_tcp_abc_l_var;
14726 	if (rack_honors_hpts_min_to)
14727 		rack->r_use_hpts_min = 1;
14728 	if (tp->snd_una != 0) {
14729 		rack->rc_sendvars_notset = 0;
14730 		/*
14731 		 * Make sure any TCP timers are not running.
14732 		 */
14733 		tcp_timer_stop(tp);
14734 	} else {
14735 		/*
14736 		 * Server side, we are called from the
14737 		 * syn-cache. This means none of the
14738 		 * snd_una/max are set yet so we have
14739 		 * to defer this until the first send.
14740 		 */
14741 		rack->rc_sendvars_notset = 1;
14742 	}
14743 
14744 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
14745 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
14746 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
14747 	rack->r_ctl.rc_min_to = rack_min_to;
14748 	microuptime(&rack->r_ctl.act_rcv_time);
14749 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
14750 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
14751 	if (rack_hw_up_only)
14752 		rack->r_up_only = 1;
14753 	if (rack_do_dyn_mul) {
14754 		/* When dynamic adjustment is on CA needs to start at 100% */
14755 		rack->rc_gp_dyn_mul = 1;
14756 		if (rack_do_dyn_mul >= 100)
14757 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
14758 	} else
14759 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
14760 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
14761 	if (rack_timely_off) {
14762 		rack->rc_skip_timely = 1;
14763 	}
14764 	if (rack->rc_skip_timely) {
14765 		rack->r_ctl.rack_per_of_gp_rec = 90;
14766 		rack->r_ctl.rack_per_of_gp_ca = 100;
14767 		rack->r_ctl.rack_per_of_gp_ss = 250;
14768 	}
14769 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
14770 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
14771 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
14772 
14773 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
14774 				rack_probertt_filter_life);
14775 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14776 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
14777 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
14778 	rack->r_ctl.rc_went_idle_time = us_cts;
14779 	rack->r_ctl.rc_time_probertt_starts = 0;
14780 
14781 	rack->r_ctl.gp_rnd_thresh = rack_rnd_cnt_req & 0xff;
14782 	if (rack_rnd_cnt_req  & 0x10000)
14783 		rack->r_ctl.gate_to_fs = 1;
14784 	rack->r_ctl.gp_gain_req = rack_gp_gain_req;
14785 	if ((rack_rnd_cnt_req & 0x100) > 0) {
14786 
14787 	}
14788 	if (rack_dsack_std_based & 0x1) {
14789 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
14790 		rack->rc_rack_tmr_std_based = 1;
14791 	}
14792 	if (rack_dsack_std_based & 0x2) {
14793 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
14794 		rack->rc_rack_use_dsack = 1;
14795 	}
14796 	/* We require at least one measurement, even if the sysctl is 0 */
14797 	if (rack_req_measurements)
14798 		rack->r_ctl.req_measurements = rack_req_measurements;
14799 	else
14800 		rack->r_ctl.req_measurements = 1;
14801 	if (rack_enable_hw_pacing)
14802 		rack->rack_hdw_pace_ena = 1;
14803 	if (rack_hw_rate_caps)
14804 		rack->r_rack_hw_rate_caps = 1;
14805 	if (rack_non_rxt_use_cr)
14806 		rack->rack_rec_nonrxt_use_cr = 1;
14807 	/* Lets setup the fsb block */
14808 	err = rack_init_fsb(tp, rack);
14809 	if (err) {
14810 		uma_zfree(rack_pcb_zone, *ptr);
14811 		*ptr = NULL;
14812 		return (err);
14813 	}
14814 	if (rack_do_hystart) {
14815 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
14816 		if (rack_do_hystart > 1)
14817 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
14818 		if (rack_do_hystart > 2)
14819 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
14820 	}
14821 	/* Log what we will do with queries */
14822 	rack_log_chg_info(tp, rack, 7,
14823 			  no_query, 0, 0);
14824 	if (rack_def_profile)
14825 		rack_set_profile(rack, rack_def_profile);
14826 	/* Cancel the GP measurement in progress */
14827 	tp->t_flags &= ~TF_GPUTINPROG;
14828 	if ((tp->t_state != TCPS_CLOSED) &&
14829 	    (tp->t_state != TCPS_TIME_WAIT)) {
14830 		/*
14831 		 * We are already open, we may
14832 		 * need to adjust a few things.
14833 		 */
14834 		if (SEQ_GT(tp->snd_max, tp->iss))
14835 			snt = tp->snd_max - tp->iss;
14836 		else
14837 			snt = 0;
14838 		iwin = rc_init_window(rack);
14839 		if ((snt < iwin) &&
14840 		    (no_query == 1)) {
14841 			/* We are not past the initial window
14842 			 * on the first init (i.e. a stack switch
14843 			 * has not yet occured) so we need to make
14844 			 * sure cwnd and ssthresh is correct.
14845 			 */
14846 			if (tp->snd_cwnd < iwin)
14847 				tp->snd_cwnd = iwin;
14848 			/*
14849 			 * If we are within the initial window
14850 			 * we want ssthresh to be unlimited. Setting
14851 			 * it to the rwnd (which the default stack does
14852 			 * and older racks) is not really a good idea
14853 			 * since we want to be in SS and grow both the
14854 			 * cwnd and the rwnd (via dynamic rwnd growth). If
14855 			 * we set it to the rwnd then as the peer grows its
14856 			 * rwnd we will be stuck in CA and never hit SS.
14857 			 *
14858 			 * Its far better to raise it up high (this takes the
14859 			 * risk that there as been a loss already, probably
14860 			 * we should have an indicator in all stacks of loss
14861 			 * but we don't), but considering the normal use this
14862 			 * is a risk worth taking. The consequences of not
14863 			 * hitting SS are far worse than going one more time
14864 			 * into it early on (before we have sent even a IW).
14865 			 * It is highly unlikely that we will have had a loss
14866 			 * before getting the IW out.
14867 			 */
14868 			tp->snd_ssthresh = 0xffffffff;
14869 		}
14870 		/*
14871 		 * Any init based on sequence numbers
14872 		 * should be done in the deferred init path
14873 		 * since we can be CLOSED and not have them
14874 		 * inited when rack_init() is called. We
14875 		 * are not closed so lets call it.
14876 		 */
14877 		rack_deferred_init(tp, rack);
14878 	}
14879 	if ((tp->t_state != TCPS_CLOSED) &&
14880 	    (tp->t_state != TCPS_TIME_WAIT) &&
14881 	    (no_query == 0) &&
14882 	    (tp->snd_una != tp->snd_max))  {
14883 		err = rack_init_outstanding(tp, rack, us_cts, *ptr);
14884 		if (err) {
14885 			*ptr = NULL;
14886 			return(err);
14887 		}
14888 	}
14889 	rack_stop_all_timers(tp, rack);
14890 	/* Setup all the t_flags2 */
14891 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14892 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14893 	else
14894 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14895 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14896 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
14897 	/*
14898 	 * Timers in Rack are kept in microseconds so lets
14899 	 * convert any initial incoming variables
14900 	 * from ticks into usecs. Note that we
14901 	 * also change the values of t_srtt and t_rttvar, if
14902 	 * they are non-zero. They are kept with a 5
14903 	 * bit decimal so we have to carefully convert
14904 	 * these to get the full precision.
14905 	 */
14906 	rack_convert_rtts(tp);
14907 	rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
14908 	if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
14909 		/* We do not start any timers on DROPPED connections */
14910 		if (tp->t_fb->tfb_chg_query == NULL) {
14911 			rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14912 		} else {
14913 			struct tcp_query_resp qr;
14914 			int ret;
14915 
14916 			memset(&qr, 0, sizeof(qr));
14917 
14918 			/* Get the misc time stamps and such for rack */
14919 			qr.req = TCP_QUERY_RACK_TIMES;
14920 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14921 			if (ret == 1) {
14922 				rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
14923 				rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
14924 				rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
14925 				rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
14926 				rack->rc_rack_rtt = qr.rack_rtt;
14927 				rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
14928 				rack->r_ctl.rc_sacked = qr.rack_sacked;
14929 				rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
14930 				rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
14931 				rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
14932 				rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
14933 				rack->r_ctl.rc_prr_out = qr.rack_prr_out;
14934 				if (qr.rack_tlp_out) {
14935 					rack->rc_tlp_in_progress = 1;
14936 					rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
14937 				} else {
14938 					rack->rc_tlp_in_progress = 0;
14939 					rack->r_ctl.rc_tlp_cnt_out = 0;
14940 				}
14941 				if (qr.rack_srtt_measured)
14942 					rack->rc_srtt_measure_made = 1;
14943 				if (qr.rack_in_persist == 1) {
14944 					rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
14945 #ifdef NETFLIX_SHARED_CWND
14946 					if (rack->r_ctl.rc_scw) {
14947 						tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
14948 						rack->rack_scwnd_is_idle = 1;
14949 					}
14950 #endif
14951 					rack->r_ctl.persist_lost_ends = 0;
14952 					rack->probe_not_answered = 0;
14953 					rack->forced_ack = 0;
14954 					tp->t_rxtshift = 0;
14955 					rack->rc_in_persist = 1;
14956 					RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
14957 							   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
14958 				}
14959 				if (qr.rack_wanted_output)
14960 					rack->r_wanted_output = 1;
14961 				rack_log_chg_info(tp, rack, 6,
14962 						  qr.rack_min_rtt,
14963 						  qr.rack_rtt,
14964 						  qr.rack_reorder_ts);
14965 			}
14966 			/* Get the old stack timers */
14967 			qr.req_param = 0;
14968 			qr.req = TCP_QUERY_TIMERS_UP;
14969 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14970 			if (ret) {
14971 				/*
14972 				 * non-zero return means we have a timer('s)
14973 				 * to start. Zero means no timer (no keepalive
14974 				 * I suppose).
14975 				 */
14976 				uint32_t tov = 0;
14977 
14978 				rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
14979 				if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
14980 					rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
14981 					if (TSTMP_GT(qr.timer_pacing_to, us_cts))
14982 						tov = qr.timer_pacing_to - us_cts;
14983 					else
14984 						tov = HPTS_TICKS_PER_SLOT;
14985 				}
14986 				if (qr.timer_hpts_flags & PACE_TMR_MASK) {
14987 					rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
14988 					if (tov == 0) {
14989 						if (TSTMP_GT(qr.timer_timer_exp, us_cts))
14990 							tov = qr.timer_timer_exp - us_cts;
14991 						else
14992 							tov = HPTS_TICKS_PER_SLOT;
14993 					}
14994 				}
14995 				rack_log_chg_info(tp, rack, 4,
14996 						  rack->r_ctl.rc_hpts_flags,
14997 						  rack->r_ctl.rc_last_output_to,
14998 						  rack->r_ctl.rc_timer_exp);
14999 				if (tov) {
15000 					struct hpts_diag diag;
15001 
15002 					(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
15003 								   __LINE__, &diag);
15004 					rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
15005 				}
15006 			}
15007 		}
15008 		rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
15009 				     __LINE__, RACK_RTTS_INIT);
15010 	}
15011 	return (0);
15012 }
15013 
15014 static int
15015 rack_handoff_ok(struct tcpcb *tp)
15016 {
15017 	if ((tp->t_state == TCPS_CLOSED) ||
15018 	    (tp->t_state == TCPS_LISTEN)) {
15019 		/* Sure no problem though it may not stick */
15020 		return (0);
15021 	}
15022 	if ((tp->t_state == TCPS_SYN_SENT) ||
15023 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
15024 		/*
15025 		 * We really don't know if you support sack,
15026 		 * you have to get to ESTAB or beyond to tell.
15027 		 */
15028 		return (EAGAIN);
15029 	}
15030 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15031 		/*
15032 		 * Rack will only send a FIN after all data is acknowledged.
15033 		 * So in this case we have more data outstanding. We can't
15034 		 * switch stacks until either all data and only the FIN
15035 		 * is left (in which case rack_init() now knows how
15036 		 * to deal with that) <or> all is acknowledged and we
15037 		 * are only left with incoming data, though why you
15038 		 * would want to switch to rack after all data is acknowledged
15039 		 * I have no idea (rrs)!
15040 		 */
15041 		return (EAGAIN);
15042 	}
15043 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15044 		return (0);
15045 	}
15046 	/*
15047 	 * If we reach here we don't do SACK on this connection so we can
15048 	 * never do rack.
15049 	 */
15050 	return (EINVAL);
15051 }
15052 
15053 static void
15054 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15055 {
15056 
15057 	if (tp->t_fb_ptr) {
15058 		uint32_t cnt_free = 0;
15059 		struct tcp_rack *rack;
15060 		struct rack_sendmap *rsm;
15061 
15062 		tcp_handle_orphaned_packets(tp);
15063 		tp->t_flags &= ~TF_FORCEDATA;
15064 		rack = (struct tcp_rack *)tp->t_fb_ptr;
15065 		rack_log_pacing_delay_calc(rack,
15066 					   0,
15067 					   0,
15068 					   0,
15069 					   rack_get_gp_est(rack), /* delRate */
15070 					   rack_get_lt_bw(rack), /* rttProp */
15071 					   20, __LINE__, NULL, 0);
15072 #ifdef NETFLIX_SHARED_CWND
15073 		if (rack->r_ctl.rc_scw) {
15074 			uint32_t limit;
15075 
15076 			if (rack->r_limit_scw)
15077 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15078 			else
15079 				limit = 0;
15080 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15081 						  rack->r_ctl.rc_scw_index,
15082 						  limit);
15083 			rack->r_ctl.rc_scw = NULL;
15084 		}
15085 #endif
15086 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
15087 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15088 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15089 			rack->r_ctl.fsb.th = NULL;
15090 		}
15091 		if (rack->rc_always_pace == 1) {
15092 			rack_remove_pacing(rack);
15093 		}
15094 		/* Clean up any options if they were not applied */
15095 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15096 			struct deferred_opt_list *dol;
15097 
15098 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15099 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15100 			free(dol, M_TCPDO);
15101 		}
15102 		/* rack does not use force data but other stacks may clear it */
15103 		if (rack->r_ctl.crte != NULL) {
15104 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15105 			rack->rack_hdrw_pacing = 0;
15106 			rack->r_ctl.crte = NULL;
15107 		}
15108 #ifdef TCP_BLACKBOX
15109 		tcp_log_flowend(tp);
15110 #endif
15111 		/*
15112 		 * Lets take a different approach to purging just
15113 		 * get each one and free it like a cum-ack would and
15114 		 * not use a foreach loop.
15115 		 */
15116 		rsm = tqhash_min(rack->r_ctl.tqh);
15117 		while (rsm) {
15118 			tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15119 			rack->r_ctl.rc_num_maps_alloced--;
15120 			uma_zfree(rack_zone, rsm);
15121 			rsm = tqhash_min(rack->r_ctl.tqh);
15122 		}
15123 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15124 		while (rsm) {
15125 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15126 			rack->r_ctl.rc_num_maps_alloced--;
15127 			rack->rc_free_cnt--;
15128 			cnt_free++;
15129 			uma_zfree(rack_zone, rsm);
15130 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15131 		}
15132 		if (rack->r_ctl.pcm_s != NULL) {
15133 			free(rack->r_ctl.pcm_s, M_TCPPCM);
15134 			rack->r_ctl.pcm_s = NULL;
15135 			rack->r_ctl.pcm_i.cnt_alloc = 0;
15136 			rack->r_ctl.pcm_i.cnt = 0;
15137 		}
15138 		if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15139 		    (tcp_bblogging_on(tp))) {
15140 			union tcp_log_stackspecific log;
15141 			struct timeval tv;
15142 
15143 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15144 			log.u_bbr.flex8 = 10;
15145 			log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15146 			log.u_bbr.flex2 = rack->rc_free_cnt;
15147 			log.u_bbr.flex3 = cnt_free;
15148 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15149 			rsm = tqhash_min(rack->r_ctl.tqh);
15150 			log.u_bbr.delRate = (uintptr_t)rsm;
15151 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15152 			log.u_bbr.cur_del_rate = (uintptr_t)rsm;
15153 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15154 			log.u_bbr.pkt_epoch = __LINE__;
15155 			(void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15156 					     0, &log, false, NULL, NULL, 0, &tv);
15157 		}
15158 		KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15159 			("rack:%p num_aloc:%u after freeing all?",
15160 			 rack,
15161 			 rack->r_ctl.rc_num_maps_alloced));
15162 		rack->rc_free_cnt = 0;
15163 		free(rack->r_ctl.tqh, M_TCPFSB);
15164 		rack->r_ctl.tqh = NULL;
15165 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15166 		tp->t_fb_ptr = NULL;
15167 	}
15168 	/* Make sure snd_nxt is correctly set */
15169 	tp->snd_nxt = tp->snd_max;
15170 }
15171 
15172 static void
15173 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15174 {
15175 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15176 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15177 	}
15178 	switch (tp->t_state) {
15179 	case TCPS_SYN_SENT:
15180 		rack->r_state = TCPS_SYN_SENT;
15181 		rack->r_substate = rack_do_syn_sent;
15182 		break;
15183 	case TCPS_SYN_RECEIVED:
15184 		rack->r_state = TCPS_SYN_RECEIVED;
15185 		rack->r_substate = rack_do_syn_recv;
15186 		break;
15187 	case TCPS_ESTABLISHED:
15188 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15189 		rack->r_state = TCPS_ESTABLISHED;
15190 		rack->r_substate = rack_do_established;
15191 		break;
15192 	case TCPS_CLOSE_WAIT:
15193 		rack->r_state = TCPS_CLOSE_WAIT;
15194 		rack->r_substate = rack_do_close_wait;
15195 		break;
15196 	case TCPS_FIN_WAIT_1:
15197 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15198 		rack->r_state = TCPS_FIN_WAIT_1;
15199 		rack->r_substate = rack_do_fin_wait_1;
15200 		break;
15201 	case TCPS_CLOSING:
15202 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15203 		rack->r_state = TCPS_CLOSING;
15204 		rack->r_substate = rack_do_closing;
15205 		break;
15206 	case TCPS_LAST_ACK:
15207 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15208 		rack->r_state = TCPS_LAST_ACK;
15209 		rack->r_substate = rack_do_lastack;
15210 		break;
15211 	case TCPS_FIN_WAIT_2:
15212 		rack->r_state = TCPS_FIN_WAIT_2;
15213 		rack->r_substate = rack_do_fin_wait_2;
15214 		break;
15215 	case TCPS_LISTEN:
15216 	case TCPS_CLOSED:
15217 	case TCPS_TIME_WAIT:
15218 	default:
15219 		break;
15220 	};
15221 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15222 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15223 
15224 }
15225 
15226 static void
15227 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15228 {
15229 	/*
15230 	 * We received an ack, and then did not
15231 	 * call send or were bounced out due to the
15232 	 * hpts was running. Now a timer is up as well, is
15233 	 * it the right timer?
15234 	 */
15235 	struct rack_sendmap *rsm;
15236 	int tmr_up;
15237 
15238 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15239 	if (tcp_in_hpts(rack->rc_tp) == 0) {
15240 		/*
15241 		 * Ok we probably need some timer up, but no
15242 		 * matter what the mask we are not in hpts. We
15243 		 * may have received an old ack and thus did nothing.
15244 		 */
15245 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15246 		rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15247 		return;
15248 	}
15249 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15250 		return;
15251 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15252 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15253 	    (tmr_up == PACE_TMR_RXT)) {
15254 		/* Should be an RXT */
15255 		return;
15256 	}
15257 	if (rsm == NULL) {
15258 		/* Nothing outstanding? */
15259 		if (tp->t_flags & TF_DELACK) {
15260 			if (tmr_up == PACE_TMR_DELACK)
15261 				/* We are supposed to have delayed ack up and we do */
15262 				return;
15263 		} else if (((V_tcp_always_keepalive ||
15264 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15265 			    (tp->t_state <= TCPS_CLOSING)) &&
15266 			   (tmr_up == PACE_TMR_KEEP) &&
15267 			   (tp->snd_max == tp->snd_una)) {
15268 			/* We should have keep alive up and we do */
15269 			return;
15270 		}
15271 	}
15272 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15273 		   ((tmr_up == PACE_TMR_TLP) ||
15274 		    (tmr_up == PACE_TMR_RACK) ||
15275 		    (tmr_up == PACE_TMR_RXT))) {
15276 		/*
15277 		 * Either a Rack, TLP or RXT is fine if  we
15278 		 * have outstanding data.
15279 		 */
15280 		return;
15281 	} else if (tmr_up == PACE_TMR_DELACK) {
15282 		/*
15283 		 * If the delayed ack was going to go off
15284 		 * before the rtx/tlp/rack timer were going to
15285 		 * expire, then that would be the timer in control.
15286 		 * Note we don't check the time here trusting the
15287 		 * code is correct.
15288 		 */
15289 		return;
15290 	}
15291 	/*
15292 	 * Ok the timer originally started is not what we want now.
15293 	 * We will force the hpts to be stopped if any, and restart
15294 	 * with the slot set to what was in the saved slot.
15295 	 */
15296 	if (tcp_in_hpts(rack->rc_tp)) {
15297 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15298 			uint32_t us_cts;
15299 
15300 			us_cts = tcp_get_usecs(NULL);
15301 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15302 				rack->r_early = 1;
15303 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15304 			}
15305 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15306 		}
15307 		tcp_hpts_remove(rack->rc_tp);
15308 	}
15309 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15310 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15311 }
15312 
15313 
15314 static void
15315 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15316 {
15317 	if ((SEQ_LT(tp->snd_wl1, seq) ||
15318 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15319 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15320 		/* keep track of pure window updates */
15321 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15322 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15323 		tp->snd_wnd = tiwin;
15324 		rack_validate_fo_sendwin_up(tp, rack);
15325 		tp->snd_wl1 = seq;
15326 		tp->snd_wl2 = ack;
15327 		if (tp->snd_wnd > tp->max_sndwnd)
15328 			tp->max_sndwnd = tp->snd_wnd;
15329 	    rack->r_wanted_output = 1;
15330 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15331 		tp->snd_wnd = tiwin;
15332 		rack_validate_fo_sendwin_up(tp, rack);
15333 		tp->snd_wl1 = seq;
15334 		tp->snd_wl2 = ack;
15335 	} else {
15336 		/* Not a valid win update */
15337 		return;
15338 	}
15339 	if (tp->snd_wnd > tp->max_sndwnd)
15340 		tp->max_sndwnd = tp->snd_wnd;
15341 	/* Do we exit persists? */
15342 	if ((rack->rc_in_persist != 0) &&
15343 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15344 				rack->r_ctl.rc_pace_min_segs))) {
15345 		rack_exit_persist(tp, rack, cts);
15346 	}
15347 	/* Do we enter persists? */
15348 	if ((rack->rc_in_persist == 0) &&
15349 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15350 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
15351 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15352 	    sbavail(&tptosocket(tp)->so_snd) &&
15353 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15354 		/*
15355 		 * Here the rwnd is less than
15356 		 * the pacing size, we are established,
15357 		 * nothing is outstanding, and there is
15358 		 * data to send. Enter persists.
15359 		 */
15360 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15361 	}
15362 }
15363 
15364 static void
15365 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15366 {
15367 
15368 	if (tcp_bblogging_on(rack->rc_tp)) {
15369 		struct inpcb *inp = tptoinpcb(tp);
15370 		union tcp_log_stackspecific log;
15371 		struct timeval ltv;
15372 		char tcp_hdr_buf[60];
15373 		struct tcphdr *th;
15374 		struct timespec ts;
15375 		uint32_t orig_snd_una;
15376 		uint8_t xx = 0;
15377 
15378 #ifdef TCP_REQUEST_TRK
15379 		struct tcp_sendfile_track *tcp_req;
15380 
15381 		if (SEQ_GT(ae->ack, tp->snd_una)) {
15382 			tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
15383 		} else {
15384 			tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
15385 		}
15386 #endif
15387 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15388 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
15389 		if (rack->rack_no_prr == 0)
15390 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15391 		else
15392 			log.u_bbr.flex1 = 0;
15393 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15394 		log.u_bbr.use_lt_bw <<= 1;
15395 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
15396 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15397 		log.u_bbr.bbr_state = rack->rc_free_cnt;
15398 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15399 		log.u_bbr.pkts_out = tp->t_maxseg;
15400 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15401 		log.u_bbr.flex7 = 1;
15402 		log.u_bbr.lost = ae->flags;
15403 		log.u_bbr.cwnd_gain = ackval;
15404 		log.u_bbr.pacing_gain = 0x2;
15405 		if (ae->flags & TSTMP_HDWR) {
15406 			/* Record the hardware timestamp if present */
15407 			log.u_bbr.flex3 = M_TSTMP;
15408 			ts.tv_sec = ae->timestamp / 1000000000;
15409 			ts.tv_nsec = ae->timestamp % 1000000000;
15410 			ltv.tv_sec = ts.tv_sec;
15411 			ltv.tv_usec = ts.tv_nsec / 1000;
15412 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
15413 		} else if (ae->flags & TSTMP_LRO) {
15414 			/* Record the LRO the arrival timestamp */
15415 			log.u_bbr.flex3 = M_TSTMP_LRO;
15416 			ts.tv_sec = ae->timestamp / 1000000000;
15417 			ts.tv_nsec = ae->timestamp % 1000000000;
15418 			ltv.tv_sec = ts.tv_sec;
15419 			ltv.tv_usec = ts.tv_nsec / 1000;
15420 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
15421 		}
15422 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
15423 		/* Log the rcv time */
15424 		log.u_bbr.delRate = ae->timestamp;
15425 #ifdef TCP_REQUEST_TRK
15426 		log.u_bbr.applimited = tp->t_tcpreq_closed;
15427 		log.u_bbr.applimited <<= 8;
15428 		log.u_bbr.applimited |= tp->t_tcpreq_open;
15429 		log.u_bbr.applimited <<= 8;
15430 		log.u_bbr.applimited |= tp->t_tcpreq_req;
15431 		if (tcp_req) {
15432 			/* Copy out any client req info */
15433 			/* seconds */
15434 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
15435 			/* useconds */
15436 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
15437 			log.u_bbr.rttProp = tcp_req->timestamp;
15438 			log.u_bbr.cur_del_rate = tcp_req->start;
15439 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
15440 				log.u_bbr.flex8 |= 1;
15441 			} else {
15442 				log.u_bbr.flex8 |= 2;
15443 				log.u_bbr.bw_inuse = tcp_req->end;
15444 			}
15445 			log.u_bbr.flex6 = tcp_req->start_seq;
15446 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
15447 				log.u_bbr.flex8 |= 4;
15448 				log.u_bbr.epoch = tcp_req->end_seq;
15449 			}
15450 		}
15451 #endif
15452 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15453 		th = (struct tcphdr *)tcp_hdr_buf;
15454 		th->th_seq = ae->seq;
15455 		th->th_ack = ae->ack;
15456 		th->th_win = ae->win;
15457 		/* Now fill in the ports */
15458 		th->th_sport = inp->inp_fport;
15459 		th->th_dport = inp->inp_lport;
15460 		tcp_set_flags(th, ae->flags);
15461 		/* Now do we have a timestamp option? */
15462 		if (ae->flags & HAS_TSTMP) {
15463 			u_char *cp;
15464 			uint32_t val;
15465 
15466 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15467 			cp = (u_char *)(th + 1);
15468 			*cp = TCPOPT_NOP;
15469 			cp++;
15470 			*cp = TCPOPT_NOP;
15471 			cp++;
15472 			*cp = TCPOPT_TIMESTAMP;
15473 			cp++;
15474 			*cp = TCPOLEN_TIMESTAMP;
15475 			cp++;
15476 			val = htonl(ae->ts_value);
15477 			bcopy((char *)&val,
15478 			      (char *)cp, sizeof(uint32_t));
15479 			val = htonl(ae->ts_echo);
15480 			bcopy((char *)&val,
15481 			      (char *)(cp + 4), sizeof(uint32_t));
15482 		} else
15483 			th->th_off = (sizeof(struct tcphdr) >> 2);
15484 
15485 		/*
15486 		 * For sane logging we need to play a little trick.
15487 		 * If the ack were fully processed we would have moved
15488 		 * snd_una to high_seq, but since compressed acks are
15489 		 * processed in two phases, at this point (logging) snd_una
15490 		 * won't be advanced. So we would see multiple acks showing
15491 		 * the advancement. We can prevent that by "pretending" that
15492 		 * snd_una was advanced and then un-advancing it so that the
15493 		 * logging code has the right value for tlb_snd_una.
15494 		 */
15495 		if (tp->snd_una != high_seq) {
15496 			orig_snd_una = tp->snd_una;
15497 			tp->snd_una = high_seq;
15498 			xx = 1;
15499 		} else
15500 			xx = 0;
15501 		TCP_LOG_EVENTP(tp, th,
15502 			       &tptosocket(tp)->so_rcv,
15503 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15504 			       0, &log, true, &ltv);
15505 		if (xx) {
15506 			tp->snd_una = orig_snd_una;
15507 		}
15508 	}
15509 
15510 }
15511 
15512 static void
15513 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15514 {
15515 	uint32_t us_rtt;
15516 	/*
15517 	 * A persist or keep-alive was forced out, update our
15518 	 * min rtt time. Note now worry about lost responses.
15519 	 * When a subsequent keep-alive or persist times out
15520 	 * and forced_ack is still on, then the last probe
15521 	 * was not responded to. In such cases we have a
15522 	 * sysctl that controls the behavior. Either we apply
15523 	 * the rtt but with reduced confidence (0). Or we just
15524 	 * plain don't apply the rtt estimate. Having data flow
15525 	 * will clear the probe_not_answered flag i.e. cum-ack
15526 	 * move forward <or> exiting and reentering persists.
15527 	 */
15528 
15529 	rack->forced_ack = 0;
15530 	rack->rc_tp->t_rxtshift = 0;
15531 	if ((rack->rc_in_persist &&
15532 	     (tiwin == rack->rc_tp->snd_wnd)) ||
15533 	    (rack->rc_in_persist == 0)) {
15534 		/*
15535 		 * In persists only apply the RTT update if this is
15536 		 * a response to our window probe. And that
15537 		 * means the rwnd sent must match the current
15538 		 * snd_wnd. If it does not, then we got a
15539 		 * window update ack instead. For keepalive
15540 		 * we allow the answer no matter what the window.
15541 		 *
15542 		 * Note that if the probe_not_answered is set then
15543 		 * the forced_ack_ts is the oldest one i.e. the first
15544 		 * probe sent that might have been lost. This assures
15545 		 * us that if we do calculate an RTT it is longer not
15546 		 * some short thing.
15547 		 */
15548 		if (rack->rc_in_persist)
15549 			counter_u64_add(rack_persists_acks, 1);
15550 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15551 		if (us_rtt == 0)
15552 			us_rtt = 1;
15553 		if (rack->probe_not_answered == 0) {
15554 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15555 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15556 		} else {
15557 			/* We have a retransmitted probe here too */
15558 			if (rack_apply_rtt_with_reduced_conf) {
15559 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15560 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15561 			}
15562 		}
15563 	}
15564 }
15565 
15566 static void
15567 rack_new_round_starts(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15568 {
15569 	/*
15570 	 * The next send has occurred mark the end of the round
15571 	 * as when that data gets acknowledged. We can
15572 	 * also do common things we might need to do when
15573 	 * a round begins.
15574 	 */
15575 	rack->r_ctl.roundends = tp->snd_max;
15576 	rack->rc_new_rnd_needed = 0;
15577 	rack_log_hystart_event(rack, tp->snd_max, 4);
15578 }
15579 
15580 
15581 static void
15582 rack_log_pcm(struct tcp_rack *rack, uint8_t mod, uint32_t flex1, uint32_t flex2,
15583 	     uint32_t flex3)
15584 {
15585 	if (tcp_bblogging_on(rack->rc_tp)) {
15586 		union tcp_log_stackspecific log;
15587 		struct timeval tv;
15588 
15589 		(void)tcp_get_usecs(&tv);
15590 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15591 		log.u_bbr.timeStamp = tcp_tv_to_usectick(&tv);
15592 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15593 		log.u_bbr.flex8 = mod;
15594 		log.u_bbr.flex1 = flex1;
15595 		log.u_bbr.flex2 = flex2;
15596 		log.u_bbr.flex3 = flex3;
15597 		log.u_bbr.flex4 = rack_pcm_every_n_rounds;
15598 		log.u_bbr.flex5 = rack->r_ctl.pcm_idle_rounds;
15599 		log.u_bbr.bbr_substate = rack->pcm_needed;
15600 		log.u_bbr.bbr_substate <<= 1;
15601 		log.u_bbr.bbr_substate |= rack->pcm_in_progress;
15602 		log.u_bbr.bbr_substate <<= 1;
15603 		log.u_bbr.bbr_substate |= rack->pcm_enabled; /* bits are NIE for Needed, Inprogress, Enabled */
15604 		(void)tcp_log_event(rack->rc_tp, NULL, NULL, NULL, TCP_PCM_MEASURE, ERRNO_UNK,
15605 				    0, &log, false, NULL, NULL, 0, &tv);
15606 	}
15607 }
15608 
15609 static void
15610 rack_new_round_setup(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15611 {
15612 	/*
15613 	 * The round (current_round) has ended. We now
15614 	 * setup for the next round by incrementing the
15615 	 * round numnber and doing any round specific
15616 	 * things.
15617 	 */
15618 	rack_log_hystart_event(rack, high_seq, 21);
15619 	rack->r_ctl.current_round++;
15620 	/* New round (current_round) begins at next send */
15621 	rack->rc_new_rnd_needed = 1;
15622 	if ((rack->pcm_enabled == 1) &&
15623 	    (rack->pcm_needed == 0) &&
15624 	    (rack->pcm_in_progress == 0)) {
15625 		/*
15626 		 * If we have enabled PCM, then we need to
15627 		 * check if the round has adanced to the state
15628 		 * where one is required.
15629 		 */
15630 		int rnds;
15631 
15632 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
15633 		if ((rnds + rack->r_ctl.pcm_idle_rounds) >= rack_pcm_every_n_rounds) {
15634 			rack->pcm_needed = 1;
15635 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15636 		} else if (rack_verbose_logging) {
15637 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15638 		}
15639 	}
15640 	if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
15641 		/* We have hystart enabled send the round info in */
15642 		if (CC_ALGO(tp)->newround != NULL) {
15643 			CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
15644 		}
15645 	}
15646 	/*
15647 	 * For DGP an initial startup check. We want to validate
15648 	 * that we are not just pushing on slow-start and just
15649 	 * not gaining.. i.e. filling buffers without getting any
15650 	 * boost in b/w during the inital slow-start.
15651 	 */
15652 	if (rack->dgp_on &&
15653 	    (rack->rc_initial_ss_comp == 0) &&
15654 	    (tp->snd_cwnd < tp->snd_ssthresh) &&
15655 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG) &&
15656 	    (rack->r_ctl.gp_rnd_thresh > 0) &&
15657 	    ((rack->r_ctl.current_round - rack->r_ctl.last_rnd_of_gp_rise) >= rack->r_ctl.gp_rnd_thresh)) {
15658 
15659 		/*
15660 		 * We are in the initial SS and we have hd rack_rnd_cnt_req rounds(def:5) where
15661 		 * we have not gained the required amount in the gp_est (120.0% aka 1200). Lets
15662 		 * exit SS.
15663 		 *
15664 		 * Pick up the flight size now as we enter slowstart (not the
15665 		 * cwnd which may be inflated).
15666 		 */
15667 		rack->rc_initial_ss_comp = 1;
15668 
15669 		if (tcp_bblogging_on(rack->rc_tp)) {
15670 			union tcp_log_stackspecific log;
15671 			struct timeval tv;
15672 
15673 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15674 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15675 			log.u_bbr.flex1 = rack->r_ctl.current_round;
15676 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
15677 			log.u_bbr.flex3 = rack->r_ctl.gp_rnd_thresh;
15678 			log.u_bbr.flex4 = rack->r_ctl.gate_to_fs;
15679 			log.u_bbr.flex5 = rack->r_ctl.ss_hi_fs;
15680 			log.u_bbr.flex8 = 40;
15681 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
15682 					    0, &log, false, NULL, __func__, __LINE__,&tv);
15683 		}
15684 		if ((rack->r_ctl.gate_to_fs == 1) &&
15685 		     (tp->snd_cwnd > rack->r_ctl.ss_hi_fs)) {
15686 			tp->snd_cwnd = rack->r_ctl.ss_hi_fs;
15687 		}
15688 		tp->snd_ssthresh = tp->snd_cwnd - 1;
15689 		/* Turn off any fast output running */
15690 		rack->r_fast_output = 0;
15691 	}
15692 }
15693 
15694 static int
15695 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15696 {
15697 	/*
15698 	 * Handle a "special" compressed ack mbuf. Each incoming
15699 	 * ack has only four possible dispositions:
15700 	 *
15701 	 * A) It moves the cum-ack forward
15702 	 * B) It is behind the cum-ack.
15703 	 * C) It is a window-update ack.
15704 	 * D) It is a dup-ack.
15705 	 *
15706 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15707 	 * in the incoming mbuf. We also need to still pay attention
15708 	 * to nxt_pkt since there may be another packet after this
15709 	 * one.
15710 	 */
15711 #ifdef TCP_ACCOUNTING
15712 	uint64_t ts_val;
15713 	uint64_t rdstc;
15714 #endif
15715 	int segsiz;
15716 	struct timespec ts;
15717 	struct tcp_rack *rack;
15718 	struct tcp_ackent *ae;
15719 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15720 	int cnt, i, did_out, ourfinisacked = 0;
15721 	struct tcpopt to_holder, *to = NULL;
15722 #ifdef TCP_ACCOUNTING
15723 	int win_up_req = 0;
15724 #endif
15725 	int nsegs = 0;
15726 	int under_pacing = 0;
15727 	int post_recovery = 0;
15728 #ifdef TCP_ACCOUNTING
15729 	sched_pin();
15730 #endif
15731 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15732 	if (rack->gp_ready &&
15733 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15734 		under_pacing = 1;
15735 
15736 	if (rack->r_state != tp->t_state)
15737 		rack_set_state(tp, rack);
15738 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15739 	    (tp->t_flags & TF_GPUTINPROG)) {
15740 		/*
15741 		 * We have a goodput in progress
15742 		 * and we have entered a late state.
15743 		 * Do we have enough data in the sb
15744 		 * to handle the GPUT request?
15745 		 */
15746 		uint32_t bytes;
15747 
15748 		bytes = tp->gput_ack - tp->gput_seq;
15749 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
15750 			bytes += tp->gput_seq - tp->snd_una;
15751 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15752 			/*
15753 			 * There are not enough bytes in the socket
15754 			 * buffer that have been sent to cover this
15755 			 * measurement. Cancel it.
15756 			 */
15757 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15758 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
15759 						   tp->gput_seq,
15760 						   0, 0, 18, __LINE__, NULL, 0);
15761 			tp->t_flags &= ~TF_GPUTINPROG;
15762 		}
15763 	}
15764 	to = &to_holder;
15765 	to->to_flags = 0;
15766 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15767 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15768 	cnt = m->m_len / sizeof(struct tcp_ackent);
15769 	counter_u64_add(rack_multi_single_eq, cnt);
15770 	high_seq = tp->snd_una;
15771 	the_win = tp->snd_wnd;
15772 	win_seq = tp->snd_wl1;
15773 	win_upd_ack = tp->snd_wl2;
15774 	cts = tcp_tv_to_usectick(tv);
15775 	ms_cts = tcp_tv_to_mssectick(tv);
15776 	rack->r_ctl.rc_rcvtime = cts;
15777 	segsiz = ctf_fixed_maxseg(tp);
15778 	if ((rack->rc_gp_dyn_mul) &&
15779 	    (rack->use_fixed_rate == 0) &&
15780 	    (rack->rc_always_pace)) {
15781 		/* Check in on probertt */
15782 		rack_check_probe_rtt(rack, cts);
15783 	}
15784 	for (i = 0; i < cnt; i++) {
15785 #ifdef TCP_ACCOUNTING
15786 		ts_val = get_cyclecount();
15787 #endif
15788 		rack_clear_rate_sample(rack);
15789 		ae = ((mtod(m, struct tcp_ackent *)) + i);
15790 		if (ae->flags & TH_FIN)
15791 			rack_log_pacing_delay_calc(rack,
15792 						   0,
15793 						   0,
15794 						   0,
15795 						   rack_get_gp_est(rack), /* delRate */
15796 						   rack_get_lt_bw(rack), /* rttProp */
15797 						   20, __LINE__, NULL, 0);
15798 		/* Setup the window */
15799 		tiwin = ae->win << tp->snd_scale;
15800 		if (tiwin > rack->r_ctl.rc_high_rwnd)
15801 			rack->r_ctl.rc_high_rwnd = tiwin;
15802 		/* figure out the type of ack */
15803 		if (SEQ_LT(ae->ack, high_seq)) {
15804 			/* Case B*/
15805 			ae->ack_val_set = ACK_BEHIND;
15806 		} else if (SEQ_GT(ae->ack, high_seq)) {
15807 			/* Case A */
15808 			ae->ack_val_set = ACK_CUMACK;
15809 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
15810 			/* Case D */
15811 			ae->ack_val_set = ACK_DUPACK;
15812 		} else {
15813 			/* Case C */
15814 			ae->ack_val_set = ACK_RWND;
15815 		}
15816 		rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
15817 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
15818 		/* Validate timestamp */
15819 		if (ae->flags & HAS_TSTMP) {
15820 			/* Setup for a timestamp */
15821 			to->to_flags = TOF_TS;
15822 			ae->ts_echo -= tp->ts_offset;
15823 			to->to_tsecr = ae->ts_echo;
15824 			to->to_tsval = ae->ts_value;
15825 			/*
15826 			 * If echoed timestamp is later than the current time, fall back to
15827 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
15828 			 * were used when this connection was established.
15829 			 */
15830 			if (TSTMP_GT(ae->ts_echo, ms_cts))
15831 				to->to_tsecr = 0;
15832 			if (tp->ts_recent &&
15833 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
15834 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
15835 #ifdef TCP_ACCOUNTING
15836 					rdstc = get_cyclecount();
15837 					if (rdstc > ts_val) {
15838 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15839 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15840 						}
15841 					}
15842 #endif
15843 					continue;
15844 				}
15845 			}
15846 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
15847 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
15848 				tp->ts_recent_age = tcp_ts_getticks();
15849 				tp->ts_recent = ae->ts_value;
15850 			}
15851 		} else {
15852 			/* Setup for a no options */
15853 			to->to_flags = 0;
15854 		}
15855 		/* Update the rcv time and perform idle reduction possibly */
15856 		if  (tp->t_idle_reduce &&
15857 		     (tp->snd_max == tp->snd_una) &&
15858 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
15859 			counter_u64_add(rack_input_idle_reduces, 1);
15860 			rack_cc_after_idle(rack, tp);
15861 		}
15862 		tp->t_rcvtime = ticks;
15863 		/* Now what about ECN of a chain of pure ACKs? */
15864 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
15865 			tcp_packets_this_ack(tp, ae->ack),
15866 			ae->codepoint))
15867 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
15868 #ifdef TCP_ACCOUNTING
15869 		/* Count for the specific type of ack in */
15870 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15871 			tp->tcp_cnt_counters[ae->ack_val_set]++;
15872 		}
15873 #endif
15874 		/*
15875 		 * Note how we could move up these in the determination
15876 		 * above, but we don't so that way the timestamp checks (and ECN)
15877 		 * is done first before we do any processing on the ACK.
15878 		 * The non-compressed path through the code has this
15879 		 * weakness (noted by @jtl) that it actually does some
15880 		 * processing before verifying the timestamp information.
15881 		 * We don't take that path here which is why we set
15882 		 * the ack_val_set first, do the timestamp and ecn
15883 		 * processing, and then look at what we have setup.
15884 		 */
15885 		if (ae->ack_val_set == ACK_BEHIND) {
15886 			/*
15887 			 * Case B flag reordering, if window is not closed
15888 			 * or it could be a keep-alive or persists
15889 			 */
15890 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
15891 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15892 				if (rack->r_ctl.rc_reorder_ts == 0)
15893 					rack->r_ctl.rc_reorder_ts = 1;
15894 			}
15895 		} else if (ae->ack_val_set == ACK_DUPACK) {
15896 			/* Case D */
15897 			rack_strike_dupack(rack, ae->ack);
15898 		} else if (ae->ack_val_set == ACK_RWND) {
15899 			/* Case C */
15900 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15901 				ts.tv_sec = ae->timestamp / 1000000000;
15902 				ts.tv_nsec = ae->timestamp % 1000000000;
15903 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15904 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15905 			} else {
15906 				rack->r_ctl.act_rcv_time = *tv;
15907 			}
15908 			if (rack->forced_ack) {
15909 				rack_handle_probe_response(rack, tiwin,
15910 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
15911 			}
15912 #ifdef TCP_ACCOUNTING
15913 			win_up_req = 1;
15914 #endif
15915 			win_upd_ack = ae->ack;
15916 			win_seq = ae->seq;
15917 			the_win = tiwin;
15918 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15919 		} else {
15920 			/* Case A */
15921 			if (SEQ_GT(ae->ack, tp->snd_max)) {
15922 				/*
15923 				 * We just send an ack since the incoming
15924 				 * ack is beyond the largest seq we sent.
15925 				 */
15926 				if ((tp->t_flags & TF_ACKNOW) == 0) {
15927 					ctf_ack_war_checks(tp);
15928 					if (tp->t_flags && TF_ACKNOW)
15929 						rack->r_wanted_output = 1;
15930 				}
15931 			} else {
15932 				nsegs++;
15933 				/* If the window changed setup to update */
15934 				if (tiwin != tp->snd_wnd) {
15935 					win_upd_ack = ae->ack;
15936 					win_seq = ae->seq;
15937 					the_win = tiwin;
15938 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15939 				}
15940 #ifdef TCP_ACCOUNTING
15941 				/* Account for the acks */
15942 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15943 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
15944 				}
15945 #endif
15946 				high_seq = ae->ack;
15947 				/* Setup our act_rcv_time */
15948 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15949 					ts.tv_sec = ae->timestamp / 1000000000;
15950 					ts.tv_nsec = ae->timestamp % 1000000000;
15951 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15952 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15953 				} else {
15954 					rack->r_ctl.act_rcv_time = *tv;
15955 				}
15956 				rack_process_to_cumack(tp, rack, ae->ack, cts, to,
15957 						       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
15958 #ifdef TCP_REQUEST_TRK
15959 				rack_req_check_for_comp(rack, high_seq);
15960 #endif
15961 				if (rack->rc_dsack_round_seen) {
15962 					/* Is the dsack round over? */
15963 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
15964 						/* Yes it is */
15965 						rack->rc_dsack_round_seen = 0;
15966 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
15967 					}
15968 				}
15969 			}
15970 		}
15971 		/* And lets be sure to commit the rtt measurements for this ack */
15972 		tcp_rack_xmit_timer_commit(rack, tp);
15973 #ifdef TCP_ACCOUNTING
15974 		rdstc = get_cyclecount();
15975 		if (rdstc > ts_val) {
15976 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15977 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15978 				if (ae->ack_val_set == ACK_CUMACK)
15979 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
15980 			}
15981 		}
15982 #endif
15983 	}
15984 #ifdef TCP_ACCOUNTING
15985 	ts_val = get_cyclecount();
15986 #endif
15987 	/* Tend to any collapsed window */
15988 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
15989 		/* The peer collapsed the window */
15990 		rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
15991 	} else if (rack->rc_has_collapsed)
15992 		rack_un_collapse_window(rack, __LINE__);
15993 	if ((rack->r_collapse_point_valid) &&
15994 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
15995 		rack->r_collapse_point_valid = 0;
15996 	acked_amount = acked = (high_seq - tp->snd_una);
15997 	if (acked) {
15998 		/*
15999 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
16000 		 * causes issues when we are just going app limited. Lets
16001 		 * instead use SEQ_GT <or> where its equal but more data
16002 		 * is outstanding.
16003 		 *
16004 		 * Also make sure we are on the last ack of a series. We
16005 		 * have to have all the ack's processed in queue to know
16006 		 * if there is something left outstanding.
16007 		 *
16008 		 */
16009 		if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
16010 		    (rack->rc_new_rnd_needed == 0) &&
16011 		    (nxt_pkt == 0)) {
16012 			/*
16013 			 * We have crossed into a new round with
16014 			 * this th_ack value.
16015 			 */
16016 			rack_new_round_setup(tp, rack, high_seq);
16017 		}
16018 		/*
16019 		 * Clear the probe not answered flag
16020 		 * since cum-ack moved forward.
16021 		 */
16022 		rack->probe_not_answered = 0;
16023 		if (tp->t_flags & TF_NEEDSYN) {
16024 			/*
16025 			 * T/TCP: Connection was half-synchronized, and our SYN has
16026 			 * been ACK'd (so connection is now fully synchronized).  Go
16027 			 * to non-starred state, increment snd_una for ACK of SYN,
16028 			 * and check if we can do window scaling.
16029 			 */
16030 			tp->t_flags &= ~TF_NEEDSYN;
16031 			tp->snd_una++;
16032 			acked_amount = acked = (high_seq - tp->snd_una);
16033 		}
16034 		if (acked > sbavail(&so->so_snd))
16035 			acked_amount = sbavail(&so->so_snd);
16036 		if (IN_FASTRECOVERY(tp->t_flags) &&
16037 		    (rack->rack_no_prr == 0))
16038 			rack_update_prr(tp, rack, acked_amount, high_seq);
16039 		if (IN_RECOVERY(tp->t_flags)) {
16040 			if (SEQ_LT(high_seq, tp->snd_recover) &&
16041 			    (SEQ_LT(high_seq, tp->snd_max))) {
16042 				tcp_rack_partialack(tp);
16043 			} else {
16044 				rack_post_recovery(tp, high_seq);
16045 				post_recovery = 1;
16046 			}
16047 		}  else if ((rack->rto_from_rec == 1) &&
16048 			    SEQ_GEQ(high_seq, tp->snd_recover)) {
16049 			/*
16050 			 * We were in recovery, hit a rxt timeout
16051 			 * and never re-entered recovery. The timeout(s)
16052 			 * made up all the lost data. In such a case
16053 			 * we need to clear the rto_from_rec flag.
16054 			 */
16055 			rack->rto_from_rec = 0;
16056 		}
16057 		/* Handle the rack-log-ack part (sendmap) */
16058 		if ((sbused(&so->so_snd) == 0) &&
16059 		    (acked > acked_amount) &&
16060 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16061 		    (tp->t_flags & TF_SENTFIN)) {
16062 			/*
16063 			 * We must be sure our fin
16064 			 * was sent and acked (we can be
16065 			 * in FIN_WAIT_1 without having
16066 			 * sent the fin).
16067 			 */
16068 			ourfinisacked = 1;
16069 			/*
16070 			 * Lets make sure snd_una is updated
16071 			 * since most likely acked_amount = 0 (it
16072 			 * should be).
16073 			 */
16074 			tp->snd_una = high_seq;
16075 		}
16076 		/* Did we make a RTO error? */
16077 		if ((tp->t_flags & TF_PREVVALID) &&
16078 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16079 			tp->t_flags &= ~TF_PREVVALID;
16080 			if (tp->t_rxtshift == 1 &&
16081 			    (int)(ticks - tp->t_badrxtwin) < 0)
16082 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16083 		}
16084 		/* Handle the data in the socket buffer */
16085 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16086 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16087 		if (acked_amount > 0) {
16088 			uint32_t p_cwnd;
16089 			struct mbuf *mfree;
16090 
16091 			if (post_recovery) {
16092 				/*
16093 				 * Grab the segsiz, multiply by 2 and add the snd_cwnd
16094 				 * that is the max the CC should add if we are exiting
16095 				 * recovery and doing a late add.
16096 				 */
16097 				p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16098 				p_cwnd <<= 1;
16099 				p_cwnd += tp->snd_cwnd;
16100 			}
16101 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, post_recovery);
16102 			if (post_recovery && (tp->snd_cwnd > p_cwnd)) {
16103 				/* Must be non-newreno (cubic) getting too ahead of itself */
16104 				tp->snd_cwnd = p_cwnd;
16105 			}
16106 			SOCK_SENDBUF_LOCK(so);
16107 			mfree = sbcut_locked(&so->so_snd, acked_amount);
16108 			tp->snd_una = high_seq;
16109 			/* Note we want to hold the sb lock through the sendmap adjust */
16110 			rack_adjust_sendmap_head(rack, &so->so_snd);
16111 			/* Wake up the socket if we have room to write more */
16112 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16113 			sowwakeup_locked(so);
16114 			m_freem(mfree);
16115 		}
16116 		/* update progress */
16117 		tp->t_acktime = ticks;
16118 		rack_log_progress_event(rack, tp, tp->t_acktime,
16119 					PROGRESS_UPDATE, __LINE__);
16120 		/* Clear out shifts and such */
16121 		tp->t_rxtshift = 0;
16122 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16123 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16124 		rack->rc_tlp_in_progress = 0;
16125 		rack->r_ctl.rc_tlp_cnt_out = 0;
16126 		/* Send recover and snd_nxt must be dragged along */
16127 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
16128 			tp->snd_recover = tp->snd_una;
16129 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
16130 			tp->snd_nxt = tp->snd_max;
16131 		/*
16132 		 * If the RXT timer is running we want to
16133 		 * stop it, so we can restart a TLP (or new RXT).
16134 		 */
16135 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16136 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16137 		tp->snd_wl2 = high_seq;
16138 		tp->t_dupacks = 0;
16139 		if (under_pacing &&
16140 		    (rack->use_fixed_rate == 0) &&
16141 		    (rack->in_probe_rtt == 0) &&
16142 		    rack->rc_gp_dyn_mul &&
16143 		    rack->rc_always_pace) {
16144 			/* Check if we are dragging bottom */
16145 			rack_check_bottom_drag(tp, rack, so);
16146 		}
16147 		if (tp->snd_una == tp->snd_max) {
16148 			tp->t_flags &= ~TF_PREVVALID;
16149 			rack->r_ctl.retran_during_recovery = 0;
16150 			rack->rc_suspicious = 0;
16151 			rack->r_ctl.dsack_byte_cnt = 0;
16152 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16153 			if (rack->r_ctl.rc_went_idle_time == 0)
16154 				rack->r_ctl.rc_went_idle_time = 1;
16155 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16156 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
16157 				tp->t_acktime = 0;
16158 			/* Set so we might enter persists... */
16159 			rack->r_wanted_output = 1;
16160 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16161 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16162 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16163 			    (sbavail(&so->so_snd) == 0) &&
16164 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16165 				/*
16166 				 * The socket was gone and the
16167 				 * peer sent data (not now in the past), time to
16168 				 * reset him.
16169 				 */
16170 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16171 				/* tcp_close will kill the inp pre-log the Reset */
16172 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16173 #ifdef TCP_ACCOUNTING
16174 				rdstc = get_cyclecount();
16175 				if (rdstc > ts_val) {
16176 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16177 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16178 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16179 					}
16180 				}
16181 #endif
16182 				m_freem(m);
16183 				tp = tcp_close(tp);
16184 				if (tp == NULL) {
16185 #ifdef TCP_ACCOUNTING
16186 					sched_unpin();
16187 #endif
16188 					return (1);
16189 				}
16190 				/*
16191 				 * We would normally do drop-with-reset which would
16192 				 * send back a reset. We can't since we don't have
16193 				 * all the needed bits. Instead lets arrange for
16194 				 * a call to tcp_output(). That way since we
16195 				 * are in the closed state we will generate a reset.
16196 				 *
16197 				 * Note if tcp_accounting is on we don't unpin since
16198 				 * we do that after the goto label.
16199 				 */
16200 				goto send_out_a_rst;
16201 			}
16202 			if ((sbused(&so->so_snd) == 0) &&
16203 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16204 			    (tp->t_flags & TF_SENTFIN)) {
16205 				/*
16206 				 * If we can't receive any more data, then closing user can
16207 				 * proceed. Starting the timer is contrary to the
16208 				 * specification, but if we don't get a FIN we'll hang
16209 				 * forever.
16210 				 *
16211 				 */
16212 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16213 					soisdisconnected(so);
16214 					tcp_timer_activate(tp, TT_2MSL,
16215 							   (tcp_fast_finwait2_recycle ?
16216 							    tcp_finwait2_timeout :
16217 							    TP_MAXIDLE(tp)));
16218 				}
16219 				if (ourfinisacked == 0) {
16220 					/*
16221 					 * We don't change to fin-wait-2 if we have our fin acked
16222 					 * which means we are probably in TCPS_CLOSING.
16223 					 */
16224 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
16225 				}
16226 			}
16227 		}
16228 		/* Wake up the socket if we have room to write more */
16229 		if (sbavail(&so->so_snd)) {
16230 			rack->r_wanted_output = 1;
16231 			if (ctf_progress_timeout_check(tp, true)) {
16232 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16233 							tp, tick, PROGRESS_DROP, __LINE__);
16234 				/*
16235 				 * We cheat here and don't send a RST, we should send one
16236 				 * when the pacer drops the connection.
16237 				 */
16238 #ifdef TCP_ACCOUNTING
16239 				rdstc = get_cyclecount();
16240 				if (rdstc > ts_val) {
16241 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16242 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16243 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16244 					}
16245 				}
16246 				sched_unpin();
16247 #endif
16248 				(void)tcp_drop(tp, ETIMEDOUT);
16249 				m_freem(m);
16250 				return (1);
16251 			}
16252 		}
16253 		if (ourfinisacked) {
16254 			switch(tp->t_state) {
16255 			case TCPS_CLOSING:
16256 #ifdef TCP_ACCOUNTING
16257 				rdstc = get_cyclecount();
16258 				if (rdstc > ts_val) {
16259 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16260 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16261 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16262 					}
16263 				}
16264 				sched_unpin();
16265 #endif
16266 				tcp_twstart(tp);
16267 				m_freem(m);
16268 				return (1);
16269 				break;
16270 			case TCPS_LAST_ACK:
16271 #ifdef TCP_ACCOUNTING
16272 				rdstc = get_cyclecount();
16273 				if (rdstc > ts_val) {
16274 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16275 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16276 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16277 					}
16278 				}
16279 				sched_unpin();
16280 #endif
16281 				tp = tcp_close(tp);
16282 				ctf_do_drop(m, tp);
16283 				return (1);
16284 				break;
16285 			case TCPS_FIN_WAIT_1:
16286 #ifdef TCP_ACCOUNTING
16287 				rdstc = get_cyclecount();
16288 				if (rdstc > ts_val) {
16289 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16290 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16291 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16292 					}
16293 				}
16294 #endif
16295 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16296 					soisdisconnected(so);
16297 					tcp_timer_activate(tp, TT_2MSL,
16298 							   (tcp_fast_finwait2_recycle ?
16299 							    tcp_finwait2_timeout :
16300 							    TP_MAXIDLE(tp)));
16301 				}
16302 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
16303 				break;
16304 			default:
16305 				break;
16306 			}
16307 		}
16308 		if (rack->r_fast_output) {
16309 			/*
16310 			 * We re doing fast output.. can we expand that?
16311 			 */
16312 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16313 		}
16314 #ifdef TCP_ACCOUNTING
16315 		rdstc = get_cyclecount();
16316 		if (rdstc > ts_val) {
16317 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16318 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16319 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16320 			}
16321 		}
16322 
16323 	} else if (win_up_req) {
16324 		rdstc = get_cyclecount();
16325 		if (rdstc > ts_val) {
16326 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16327 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16328 			}
16329 		}
16330 #endif
16331 	}
16332 	/* Now is there a next packet, if so we are done */
16333 	m_freem(m);
16334 	did_out = 0;
16335 	if (nxt_pkt) {
16336 #ifdef TCP_ACCOUNTING
16337 		sched_unpin();
16338 #endif
16339 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16340 		return (0);
16341 	}
16342 	rack_handle_might_revert(tp, rack);
16343 	ctf_calc_rwin(so, tp);
16344 	if ((rack->r_wanted_output != 0) ||
16345 	    (rack->r_fast_output != 0) ||
16346 	    (tp->t_flags & TF_ACKNOW )) {
16347 	send_out_a_rst:
16348 		if (tcp_output(tp) < 0) {
16349 #ifdef TCP_ACCOUNTING
16350 			sched_unpin();
16351 #endif
16352 			return (1);
16353 		}
16354 		did_out = 1;
16355 	}
16356 	if (tp->t_flags2 & TF2_HPTS_CALLS)
16357 		tp->t_flags2 &= ~TF2_HPTS_CALLS;
16358 	rack_free_trim(rack);
16359 #ifdef TCP_ACCOUNTING
16360 	sched_unpin();
16361 #endif
16362 	rack_timer_audit(tp, rack, &so->so_snd);
16363 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16364 	return (0);
16365 }
16366 
16367 #define	TCP_LRO_TS_OPTION \
16368     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16369 	  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16370 
16371 static int
16372 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16373     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16374     struct timeval *tv)
16375 {
16376 	struct inpcb *inp = tptoinpcb(tp);
16377 	struct socket *so = tptosocket(tp);
16378 #ifdef TCP_ACCOUNTING
16379 	uint64_t ts_val;
16380 #endif
16381 	int32_t thflags, retval, did_out = 0;
16382 	int32_t way_out = 0;
16383 	/*
16384 	 * cts - is the current time from tv (caller gets ts) in microseconds.
16385 	 * ms_cts - is the current time from tv in milliseconds.
16386 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16387 	 */
16388 	uint32_t cts, us_cts, ms_cts;
16389 	uint32_t tiwin;
16390 	struct timespec ts;
16391 	struct tcpopt to;
16392 	struct tcp_rack *rack;
16393 	struct rack_sendmap *rsm;
16394 	int32_t prev_state = 0;
16395 	int no_output = 0;
16396 	int slot_remaining = 0;
16397 #ifdef TCP_ACCOUNTING
16398 	int ack_val_set = 0xf;
16399 #endif
16400 	int nsegs;
16401 
16402 	NET_EPOCH_ASSERT();
16403 	INP_WLOCK_ASSERT(inp);
16404 
16405 	/*
16406 	 * tv passed from common code is from either M_TSTMP_LRO or
16407 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16408 	 */
16409 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16410 	if (rack->rack_deferred_inited == 0) {
16411 		/*
16412 		 * If we are the connecting socket we will
16413 		 * hit rack_init() when no sequence numbers
16414 		 * are setup. This makes it so we must defer
16415 		 * some initialization. Call that now.
16416 		 */
16417 		rack_deferred_init(tp, rack);
16418 	}
16419 	/*
16420 	 * Check to see if we need to skip any output plans. This
16421 	 * can happen in the non-LRO path where we are pacing and
16422 	 * must process the ack coming in but need to defer sending
16423 	 * anything becase a pacing timer is running.
16424 	 */
16425 	us_cts = tcp_tv_to_usectick(tv);
16426 	if (m->m_flags & M_ACKCMP) {
16427 		/*
16428 		 * All compressed ack's are ack's by definition so
16429 		 * remove any ack required flag and then do the processing.
16430 		 */
16431 		rack->rc_ack_required = 0;
16432 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16433 	}
16434 	thflags = tcp_get_flags(th);
16435 	if ((rack->rc_always_pace == 1) &&
16436 	    (rack->rc_ack_can_sendout_data == 0) &&
16437 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16438 	    (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16439 		/*
16440 		 * Ok conditions are right for queuing the packets
16441 		 * but we do have to check the flags in the inp, it
16442 		 * could be, if a sack is present, we want to be awoken and
16443 		 * so should process the packets.
16444 		 */
16445 		slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16446 		if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
16447 			no_output = 1;
16448 		} else {
16449 			/*
16450 			 * If there is no options, or just a
16451 			 * timestamp option, we will want to queue
16452 			 * the packets. This is the same that LRO does
16453 			 * and will need to change with accurate ECN.
16454 			 */
16455 			uint32_t *ts_ptr;
16456 			int optlen;
16457 
16458 			optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16459 			ts_ptr = (uint32_t *)(th + 1);
16460 			if ((optlen == 0) ||
16461 			    ((optlen == TCPOLEN_TSTAMP_APPA) &&
16462 			     (*ts_ptr == TCP_LRO_TS_OPTION)))
16463 				no_output = 1;
16464 		}
16465 		if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16466 			/*
16467 			 * It is unrealistic to think we can pace in less than
16468 			 * the minimum granularity of the pacer (def:250usec). So
16469 			 * if we have less than that time remaining we should go
16470 			 * ahead and allow output to be "early". We will attempt to
16471 			 * make up for it in any pacing time we try to apply on
16472 			 * the outbound packet.
16473 			 */
16474 			no_output = 0;
16475 		}
16476 	}
16477 	/*
16478 	 * If there is a RST or FIN lets dump out the bw
16479 	 * with a FIN the connection may go on but we
16480 	 * may not.
16481 	 */
16482 	if ((thflags & TH_FIN) || (thflags & TH_RST))
16483 		rack_log_pacing_delay_calc(rack,
16484 					   rack->r_ctl.gp_bw,
16485 					   0,
16486 					   0,
16487 					   rack_get_gp_est(rack), /* delRate */
16488 					   rack_get_lt_bw(rack), /* rttProp */
16489 					   20, __LINE__, NULL, 0);
16490 	if (m->m_flags & M_ACKCMP) {
16491 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16492 	}
16493 	cts = tcp_tv_to_usectick(tv);
16494 	ms_cts =  tcp_tv_to_mssectick(tv);
16495 	nsegs = m->m_pkthdr.lro_nsegs;
16496 	counter_u64_add(rack_proc_non_comp_ack, 1);
16497 #ifdef TCP_ACCOUNTING
16498 	sched_pin();
16499 	if (thflags & TH_ACK)
16500 		ts_val = get_cyclecount();
16501 #endif
16502 	if ((m->m_flags & M_TSTMP) ||
16503 	    (m->m_flags & M_TSTMP_LRO)) {
16504 		mbuf_tstmp2timespec(m, &ts);
16505 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16506 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16507 	} else
16508 		rack->r_ctl.act_rcv_time = *tv;
16509 	kern_prefetch(rack, &prev_state);
16510 	prev_state = 0;
16511 	/*
16512 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
16513 	 * the scale is zero.
16514 	 */
16515 	tiwin = th->th_win << tp->snd_scale;
16516 #ifdef TCP_ACCOUNTING
16517 	if (thflags & TH_ACK) {
16518 		/*
16519 		 * We have a tradeoff here. We can either do what we are
16520 		 * doing i.e. pinning to this CPU and then doing the accounting
16521 		 * <or> we could do a critical enter, setup the rdtsc and cpu
16522 		 * as in below, and then validate we are on the same CPU on
16523 		 * exit. I have choosen to not do the critical enter since
16524 		 * that often will gain you a context switch, and instead lock
16525 		 * us (line above this if) to the same CPU with sched_pin(). This
16526 		 * means we may be context switched out for a higher priority
16527 		 * interupt but we won't be moved to another CPU.
16528 		 *
16529 		 * If this occurs (which it won't very often since we most likely
16530 		 * are running this code in interupt context and only a higher
16531 		 * priority will bump us ... clock?) we will falsely add in
16532 		 * to the time the interupt processing time plus the ack processing
16533 		 * time. This is ok since its a rare event.
16534 		 */
16535 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16536 						    ctf_fixed_maxseg(tp));
16537 	}
16538 #endif
16539 	/*
16540 	 * Parse options on any incoming segment.
16541 	 */
16542 	memset(&to, 0, sizeof(to));
16543 	tcp_dooptions(&to, (u_char *)(th + 1),
16544 	    (th->th_off << 2) - sizeof(struct tcphdr),
16545 	    (thflags & TH_SYN) ? TO_SYN : 0);
16546 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16547 	    __func__));
16548 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16549 	    __func__));
16550 	if (tp->t_flags2 & TF2_PROC_SACK_PROHIBIT) {
16551 		/*
16552 		 * We don't look at sack's from the
16553 		 * peer because the MSS is too small which
16554 		 * can subject us to an attack.
16555 		 */
16556 		to.to_flags &= ~TOF_SACK;
16557 	}
16558 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16559 	    (tp->t_flags & TF_GPUTINPROG)) {
16560 		/*
16561 		 * We have a goodput in progress
16562 		 * and we have entered a late state.
16563 		 * Do we have enough data in the sb
16564 		 * to handle the GPUT request?
16565 		 */
16566 		uint32_t bytes;
16567 
16568 		bytes = tp->gput_ack - tp->gput_seq;
16569 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
16570 			bytes += tp->gput_seq - tp->snd_una;
16571 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16572 			/*
16573 			 * There are not enough bytes in the socket
16574 			 * buffer that have been sent to cover this
16575 			 * measurement. Cancel it.
16576 			 */
16577 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16578 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
16579 						   tp->gput_seq,
16580 						   0, 0, 18, __LINE__, NULL, 0);
16581 			tp->t_flags &= ~TF_GPUTINPROG;
16582 		}
16583 	}
16584 	if (tcp_bblogging_on(rack->rc_tp)) {
16585 		union tcp_log_stackspecific log;
16586 		struct timeval ltv;
16587 #ifdef TCP_REQUEST_TRK
16588 		struct tcp_sendfile_track *tcp_req;
16589 
16590 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
16591 			tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
16592 		} else {
16593 			tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
16594 		}
16595 #endif
16596 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16597 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16598 		if (rack->rack_no_prr == 0)
16599 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16600 		else
16601 			log.u_bbr.flex1 = 0;
16602 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16603 		log.u_bbr.use_lt_bw <<= 1;
16604 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
16605 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16606 		log.u_bbr.bbr_state = rack->rc_free_cnt;
16607 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16608 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16609 		log.u_bbr.flex3 = m->m_flags;
16610 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16611 		log.u_bbr.lost = thflags;
16612 		log.u_bbr.pacing_gain = 0x1;
16613 #ifdef TCP_ACCOUNTING
16614 		log.u_bbr.cwnd_gain = ack_val_set;
16615 #endif
16616 		log.u_bbr.flex7 = 2;
16617 		if (m->m_flags & M_TSTMP) {
16618 			/* Record the hardware timestamp if present */
16619 			mbuf_tstmp2timespec(m, &ts);
16620 			ltv.tv_sec = ts.tv_sec;
16621 			ltv.tv_usec = ts.tv_nsec / 1000;
16622 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
16623 		} else if (m->m_flags & M_TSTMP_LRO) {
16624 			/* Record the LRO the arrival timestamp */
16625 			mbuf_tstmp2timespec(m, &ts);
16626 			ltv.tv_sec = ts.tv_sec;
16627 			ltv.tv_usec = ts.tv_nsec / 1000;
16628 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
16629 		}
16630 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16631 		/* Log the rcv time */
16632 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16633 #ifdef TCP_REQUEST_TRK
16634 		log.u_bbr.applimited = tp->t_tcpreq_closed;
16635 		log.u_bbr.applimited <<= 8;
16636 		log.u_bbr.applimited |= tp->t_tcpreq_open;
16637 		log.u_bbr.applimited <<= 8;
16638 		log.u_bbr.applimited |= tp->t_tcpreq_req;
16639 		if (tcp_req) {
16640 			/* Copy out any client req info */
16641 			/* seconds */
16642 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16643 			/* useconds */
16644 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16645 			log.u_bbr.rttProp = tcp_req->timestamp;
16646 			log.u_bbr.cur_del_rate = tcp_req->start;
16647 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16648 				log.u_bbr.flex8 |= 1;
16649 			} else {
16650 				log.u_bbr.flex8 |= 2;
16651 				log.u_bbr.bw_inuse = tcp_req->end;
16652 			}
16653 			log.u_bbr.flex6 = tcp_req->start_seq;
16654 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16655 				log.u_bbr.flex8 |= 4;
16656 				log.u_bbr.epoch = tcp_req->end_seq;
16657 			}
16658 		}
16659 #endif
16660 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16661 		    tlen, &log, true, &ltv);
16662 	}
16663 	/* Remove ack required flag if set, we have one  */
16664 	if (thflags & TH_ACK)
16665 		rack->rc_ack_required = 0;
16666 	rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16667 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16668 		way_out = 4;
16669 		retval = 0;
16670 		m_freem(m);
16671 		goto done_with_input;
16672 	}
16673 	/*
16674 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
16675 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16676 	 */
16677 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16678 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16679 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16680 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
16681 #ifdef TCP_ACCOUNTING
16682 		sched_unpin();
16683 #endif
16684 		return (1);
16685 	}
16686 	/*
16687 	 * If timestamps were negotiated during SYN/ACK and a
16688 	 * segment without a timestamp is received, silently drop
16689 	 * the segment, unless it is a RST segment or missing timestamps are
16690 	 * tolerated.
16691 	 * See section 3.2 of RFC 7323.
16692 	 */
16693 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16694 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16695 		way_out = 5;
16696 		retval = 0;
16697 		m_freem(m);
16698 		goto done_with_input;
16699 	}
16700 	/*
16701 	 * Segment received on connection. Reset idle time and keep-alive
16702 	 * timer. XXX: This should be done after segment validation to
16703 	 * ignore broken/spoofed segs.
16704 	 */
16705 	if  (tp->t_idle_reduce &&
16706 	     (tp->snd_max == tp->snd_una) &&
16707 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16708 		counter_u64_add(rack_input_idle_reduces, 1);
16709 		rack_cc_after_idle(rack, tp);
16710 	}
16711 	tp->t_rcvtime = ticks;
16712 #ifdef STATS
16713 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16714 #endif
16715 	if (tiwin > rack->r_ctl.rc_high_rwnd)
16716 		rack->r_ctl.rc_high_rwnd = tiwin;
16717 	/*
16718 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16719 	 * this to occur after we've validated the segment.
16720 	 */
16721 	if (tcp_ecn_input_segment(tp, thflags, tlen,
16722 	    tcp_packets_this_ack(tp, th->th_ack),
16723 	    iptos))
16724 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16725 
16726 	/*
16727 	 * If echoed timestamp is later than the current time, fall back to
16728 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16729 	 * were used when this connection was established.
16730 	 */
16731 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16732 		to.to_tsecr -= tp->ts_offset;
16733 		if (TSTMP_GT(to.to_tsecr, ms_cts))
16734 			to.to_tsecr = 0;
16735 	}
16736 	if ((rack->r_rcvpath_rtt_up == 1) &&
16737 	    (to.to_flags & TOF_TS) &&
16738 	    (TSTMP_GEQ(to.to_tsecr, rack->r_ctl.last_rcv_tstmp_for_rtt))) {
16739 		uint32_t rtt = 0;
16740 
16741 		/*
16742 		 * We are receiving only and thus not sending
16743 		 * data to do an RTT. We set a flag when we first
16744 		 * sent this TS to the peer. We now have it back
16745 		 * and have an RTT to share. We log it as a conf
16746 		 * 4, we are not so sure about it.. since we
16747 		 * may have lost an ack.
16748 		 */
16749 		if (TSTMP_GT(cts, rack->r_ctl.last_time_of_arm_rcv))
16750 		    rtt = (cts - rack->r_ctl.last_time_of_arm_rcv);
16751 		rack->r_rcvpath_rtt_up = 0;
16752 		/* Submit and commit the timer */
16753 		if (rtt > 0) {
16754 			tcp_rack_xmit_timer(rack, rtt, 0, rtt, 4, NULL, 1);
16755 			tcp_rack_xmit_timer_commit(rack, tp);
16756 		}
16757 	}
16758 	/*
16759 	 * If its the first time in we need to take care of options and
16760 	 * verify we can do SACK for rack!
16761 	 */
16762 	if (rack->r_state == 0) {
16763 		/* Should be init'd by rack_init() */
16764 		KASSERT(rack->rc_inp != NULL,
16765 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
16766 		if (rack->rc_inp == NULL) {
16767 			rack->rc_inp = inp;
16768 		}
16769 
16770 		/*
16771 		 * Process options only when we get SYN/ACK back. The SYN
16772 		 * case for incoming connections is handled in tcp_syncache.
16773 		 * According to RFC1323 the window field in a SYN (i.e., a
16774 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16775 		 * this is traditional behavior, may need to be cleaned up.
16776 		 */
16777 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16778 			/* Handle parallel SYN for ECN */
16779 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
16780 			if ((to.to_flags & TOF_SCALE) &&
16781 			    (tp->t_flags & TF_REQ_SCALE)) {
16782 				tp->t_flags |= TF_RCVD_SCALE;
16783 				tp->snd_scale = to.to_wscale;
16784 			} else
16785 				tp->t_flags &= ~TF_REQ_SCALE;
16786 			/*
16787 			 * Initial send window.  It will be updated with the
16788 			 * next incoming segment to the scaled value.
16789 			 */
16790 			tp->snd_wnd = th->th_win;
16791 			rack_validate_fo_sendwin_up(tp, rack);
16792 			if ((to.to_flags & TOF_TS) &&
16793 			    (tp->t_flags & TF_REQ_TSTMP)) {
16794 				tp->t_flags |= TF_RCVD_TSTMP;
16795 				tp->ts_recent = to.to_tsval;
16796 				tp->ts_recent_age = cts;
16797 			} else
16798 				tp->t_flags &= ~TF_REQ_TSTMP;
16799 			if (to.to_flags & TOF_MSS) {
16800 				tcp_mss(tp, to.to_mss);
16801 			}
16802 			if ((tp->t_flags & TF_SACK_PERMIT) &&
16803 			    (to.to_flags & TOF_SACKPERM) == 0)
16804 				tp->t_flags &= ~TF_SACK_PERMIT;
16805 			if (tp->t_flags & TF_FASTOPEN) {
16806 				if (to.to_flags & TOF_FASTOPEN) {
16807 					uint16_t mss;
16808 
16809 					if (to.to_flags & TOF_MSS)
16810 						mss = to.to_mss;
16811 					else
16812 						if ((inp->inp_vflag & INP_IPV6) != 0)
16813 							mss = TCP6_MSS;
16814 						else
16815 							mss = TCP_MSS;
16816 					tcp_fastopen_update_cache(tp, mss,
16817 					    to.to_tfo_len, to.to_tfo_cookie);
16818 				} else
16819 					tcp_fastopen_disable_path(tp);
16820 			}
16821 		}
16822 		/*
16823 		 * At this point we are at the initial call. Here we decide
16824 		 * if we are doing RACK or not. We do this by seeing if
16825 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
16826 		 * The code now does do dup-ack counting so if you don't
16827 		 * switch back you won't get rack & TLP, but you will still
16828 		 * get this stack.
16829 		 */
16830 
16831 		if ((rack_sack_not_required == 0) &&
16832 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
16833 			tcp_switch_back_to_default(tp);
16834 			(*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
16835 			    tlen, iptos);
16836 #ifdef TCP_ACCOUNTING
16837 			sched_unpin();
16838 #endif
16839 			return (1);
16840 		}
16841 		tcp_set_hpts(tp);
16842 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
16843 	}
16844 	if (thflags & TH_FIN)
16845 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
16846 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16847 	if ((rack->rc_gp_dyn_mul) &&
16848 	    (rack->use_fixed_rate == 0) &&
16849 	    (rack->rc_always_pace)) {
16850 		/* Check in on probertt */
16851 		rack_check_probe_rtt(rack, cts);
16852 	}
16853 	rack_clear_rate_sample(rack);
16854 	if ((rack->forced_ack) &&
16855 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
16856 		rack_handle_probe_response(rack, tiwin, us_cts);
16857 	}
16858 	/*
16859 	 * This is the one exception case where we set the rack state
16860 	 * always. All other times (timers etc) we must have a rack-state
16861 	 * set (so we assure we have done the checks above for SACK).
16862 	 */
16863 	rack->r_ctl.rc_rcvtime = cts;
16864 	if (rack->r_state != tp->t_state)
16865 		rack_set_state(tp, rack);
16866 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
16867 	    (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
16868 		kern_prefetch(rsm, &prev_state);
16869 	prev_state = rack->r_state;
16870 	if ((thflags & TH_RST) &&
16871 	    ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
16872 	      SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
16873 	     (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
16874 		/* The connection will be killed by a reset check the tracepoint */
16875 		tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
16876 	}
16877 	retval = (*rack->r_substate) (m, th, so,
16878 	    tp, &to, drop_hdrlen,
16879 	    tlen, tiwin, thflags, nxt_pkt, iptos);
16880 	if (retval == 0) {
16881 		/*
16882 		 * If retval is 1 the tcb is unlocked and most likely the tp
16883 		 * is gone.
16884 		 */
16885 		INP_WLOCK_ASSERT(inp);
16886 		if ((rack->rc_gp_dyn_mul) &&
16887 		    (rack->rc_always_pace) &&
16888 		    (rack->use_fixed_rate == 0) &&
16889 		    rack->in_probe_rtt &&
16890 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
16891 			/*
16892 			 * If we are going for target, lets recheck before
16893 			 * we output.
16894 			 */
16895 			rack_check_probe_rtt(rack, cts);
16896 		}
16897 		if (rack->set_pacing_done_a_iw == 0) {
16898 			/* How much has been acked? */
16899 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
16900 				/* We have enough to set in the pacing segment size */
16901 				rack->set_pacing_done_a_iw = 1;
16902 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
16903 			}
16904 		}
16905 		tcp_rack_xmit_timer_commit(rack, tp);
16906 #ifdef TCP_ACCOUNTING
16907 		/*
16908 		 * If we set the ack_val_se to what ack processing we are doing
16909 		 * we also want to track how many cycles we burned. Note
16910 		 * the bits after tcp_output we let be "free". This is because
16911 		 * we are also tracking the tcp_output times as well. Note the
16912 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
16913 		 * 0xf cannot be returned and is what we initialize it too to
16914 		 * indicate we are not doing the tabulations.
16915 		 */
16916 		if (ack_val_set != 0xf) {
16917 			uint64_t crtsc;
16918 
16919 			crtsc = get_cyclecount();
16920 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16921 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
16922 			}
16923 		}
16924 #endif
16925 		if ((nxt_pkt == 0) && (no_output == 0)) {
16926 			if ((rack->r_wanted_output != 0) ||
16927 			    (tp->t_flags & TF_ACKNOW) ||
16928 			    (rack->r_fast_output != 0)) {
16929 
16930 do_output_now:
16931 				if (tcp_output(tp) < 0) {
16932 #ifdef TCP_ACCOUNTING
16933 					sched_unpin();
16934 #endif
16935 					return (1);
16936 				}
16937 				did_out = 1;
16938 			}
16939 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16940 			rack_free_trim(rack);
16941 		} else if ((nxt_pkt == 0) && (tp->t_flags & TF_ACKNOW)) {
16942 			goto do_output_now;
16943 		} else if ((no_output == 1) &&
16944 			   (nxt_pkt == 0)  &&
16945 			   (tcp_in_hpts(rack->rc_tp) == 0)) {
16946 			/*
16947 			 * We are not in hpts and we had a pacing timer up. Use
16948 			 * the remaining time (slot_remaining) to restart the timer.
16949 			 */
16950 			KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
16951 			rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
16952 			rack_free_trim(rack);
16953 		}
16954 		/* Clear the flag, it may have been cleared by output but we may not have  */
16955 		if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
16956 			tp->t_flags2 &= ~TF2_HPTS_CALLS;
16957 		/*
16958 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
16959 		 * causes issues when we are just going app limited. Lets
16960 		 * instead use SEQ_GT <or> where its equal but more data
16961 		 * is outstanding.
16962 		 *
16963 		 * Also make sure we are on the last ack of a series. We
16964 		 * have to have all the ack's processed in queue to know
16965 		 * if there is something left outstanding.
16966 		 */
16967 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
16968 		    (rack->rc_new_rnd_needed == 0) &&
16969 		    (nxt_pkt == 0)) {
16970 			/*
16971 			 * We have crossed into a new round with
16972 			 * the new snd_unae.
16973 			 */
16974 			rack_new_round_setup(tp, rack, tp->snd_una);
16975 		}
16976 		if ((nxt_pkt == 0) &&
16977 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
16978 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
16979 		     (tp->t_flags & TF_DELACK) ||
16980 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
16981 		      (tp->t_state <= TCPS_CLOSING)))) {
16982 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
16983 			if ((tp->snd_max == tp->snd_una) &&
16984 			    ((tp->t_flags & TF_DELACK) == 0) &&
16985 			    (tcp_in_hpts(rack->rc_tp)) &&
16986 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
16987 				/* keep alive not needed if we are hptsi output yet */
16988 				;
16989 			} else {
16990 				int late = 0;
16991 				if (tcp_in_hpts(tp)) {
16992 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
16993 						us_cts = tcp_get_usecs(NULL);
16994 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
16995 							rack->r_early = 1;
16996 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
16997 						} else
16998 							late = 1;
16999 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
17000 					}
17001 					tcp_hpts_remove(tp);
17002 				}
17003 				if (late && (did_out == 0)) {
17004 					/*
17005 					 * We are late in the sending
17006 					 * and we did not call the output
17007 					 * (this probably should not happen).
17008 					 */
17009 					goto do_output_now;
17010 				}
17011 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
17012 			}
17013 			way_out = 1;
17014 		} else if (nxt_pkt == 0) {
17015 			/* Do we have the correct timer running? */
17016 			rack_timer_audit(tp, rack, &so->so_snd);
17017 			way_out = 2;
17018 		}
17019 	done_with_input:
17020 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
17021 		if (did_out)
17022 			rack->r_wanted_output = 0;
17023 	}
17024 
17025 #ifdef TCP_ACCOUNTING
17026 	sched_unpin();
17027 #endif
17028 	return (retval);
17029 }
17030 
17031 static void
17032 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17033     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17034 {
17035 	struct timeval tv;
17036 
17037 	/* First lets see if we have old packets */
17038 	if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17039 		if (ctf_do_queued_segments(tp, 1)) {
17040 			m_freem(m);
17041 			return;
17042 		}
17043 	}
17044 	if (m->m_flags & M_TSTMP_LRO) {
17045 		mbuf_tstmp2timeval(m, &tv);
17046 	} else {
17047 		/* Should not be should we kassert instead? */
17048 		tcp_get_usecs(&tv);
17049 	}
17050 	if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17051 	    &tv) == 0) {
17052 		INP_WUNLOCK(tptoinpcb(tp));
17053 	}
17054 }
17055 
17056 struct rack_sendmap *
17057 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17058 {
17059 	struct rack_sendmap *rsm = NULL;
17060 	int32_t idx;
17061 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
17062 
17063 	/* Return the next guy to be re-transmitted */
17064 	if (tqhash_empty(rack->r_ctl.tqh)) {
17065 		return (NULL);
17066 	}
17067 	if (tp->t_flags & TF_SENTFIN) {
17068 		/* retran the end FIN? */
17069 		return (NULL);
17070 	}
17071 	/* ok lets look at this one */
17072 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17073 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17074 		return (rsm);
17075 	}
17076 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17077 		goto check_it;
17078 	}
17079 	rsm = rack_find_lowest_rsm(rack);
17080 	if (rsm == NULL) {
17081 		return (NULL);
17082 	}
17083 check_it:
17084 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
17085 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17086 		/*
17087 		 * No sack so we automatically do the 3 strikes and
17088 		 * retransmit (no rack timer would be started).
17089 		 */
17090 		return (rsm);
17091 	}
17092 	if (rsm->r_flags & RACK_ACKED) {
17093 		return (NULL);
17094 	}
17095 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17096 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17097 		/* Its not yet ready */
17098 		return (NULL);
17099 	}
17100 	srtt = rack_grab_rtt(tp, rack);
17101 	idx = rsm->r_rtr_cnt - 1;
17102 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17103 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
17104 	if ((tsused == ts_low) ||
17105 	    (TSTMP_LT(tsused, ts_low))) {
17106 		/* No time since sending */
17107 		return (NULL);
17108 	}
17109 	if ((tsused - ts_low) < thresh) {
17110 		/* It has not been long enough yet */
17111 		return (NULL);
17112 	}
17113 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17114 	    ((rsm->r_flags & RACK_SACK_PASSED))) {
17115 		/*
17116 		 * We have passed the dup-ack threshold <or>
17117 		 * a SACK has indicated this is missing.
17118 		 * Note that if you are a declared attacker
17119 		 * it is only the dup-ack threshold that
17120 		 * will cause retransmits.
17121 		 */
17122 		/* log retransmit reason */
17123 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17124 		rack->r_fast_output = 0;
17125 		return (rsm);
17126 	}
17127 	return (NULL);
17128 }
17129 
17130 static void
17131 rack_log_pacing_delay_calc (struct tcp_rack *rack, uint32_t len, uint32_t slot,
17132 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17133 			   int line, struct rack_sendmap *rsm, uint8_t quality)
17134 {
17135 	if (tcp_bblogging_on(rack->rc_tp)) {
17136 		union tcp_log_stackspecific log;
17137 		struct timeval tv;
17138 
17139 		if (rack_verbose_logging == 0) {
17140 			/*
17141 			 * We are not verbose screen out all but
17142 			 * ones we always want.
17143 			 */
17144 			if ((method != 2) &&
17145 			    (method != 3) &&
17146 			    (method != 7) &&
17147 			    (method != 89) &&
17148 			    (method != 14) &&
17149 			    (method != 20)) {
17150 				return;
17151 			}
17152 		}
17153 		memset(&log, 0, sizeof(log));
17154 		log.u_bbr.flex1 = slot;
17155 		log.u_bbr.flex2 = len;
17156 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17157 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17158 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17159 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17160 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17161 		log.u_bbr.use_lt_bw <<= 1;
17162 		log.u_bbr.use_lt_bw |= rack->r_late;
17163 		log.u_bbr.use_lt_bw <<= 1;
17164 		log.u_bbr.use_lt_bw |= rack->r_early;
17165 		log.u_bbr.use_lt_bw <<= 1;
17166 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17167 		log.u_bbr.use_lt_bw <<= 1;
17168 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17169 		log.u_bbr.use_lt_bw <<= 1;
17170 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17171 		log.u_bbr.use_lt_bw <<= 1;
17172 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17173 		log.u_bbr.use_lt_bw <<= 1;
17174 		log.u_bbr.use_lt_bw |= rack->gp_ready;
17175 		log.u_bbr.pkt_epoch = line;
17176 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17177 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17178 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17179 		log.u_bbr.bw_inuse = bw_est;
17180 		log.u_bbr.delRate = bw;
17181 		if (rack->r_ctl.gp_bw == 0)
17182 			log.u_bbr.cur_del_rate = 0;
17183 		else
17184 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
17185 		log.u_bbr.rttProp = len_time;
17186 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17187 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17188 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17189 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17190 			/* We are in slow start */
17191 			log.u_bbr.flex7 = 1;
17192 		} else {
17193 			/* we are on congestion avoidance */
17194 			log.u_bbr.flex7 = 0;
17195 		}
17196 		log.u_bbr.flex8 = method;
17197 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17198 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17199 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17200 		log.u_bbr.cwnd_gain <<= 1;
17201 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17202 		log.u_bbr.cwnd_gain <<= 1;
17203 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17204 		log.u_bbr.bbr_substate = quality;
17205 		log.u_bbr.bbr_state = rack->dgp_on;
17206 		log.u_bbr.bbr_state <<= 1;
17207 		log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17208 		log.u_bbr.bbr_state <<= 2;
17209 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
17210 		    &rack->rc_inp->inp_socket->so_rcv,
17211 		    &rack->rc_inp->inp_socket->so_snd,
17212 		    BBR_LOG_HPTSI_CALC, 0,
17213 		    0, &log, false, &tv);
17214 	}
17215 }
17216 
17217 static uint32_t
17218 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17219 {
17220 	uint32_t new_tso, user_max, pace_one;
17221 
17222 	user_max = rack->rc_user_set_max_segs * mss;
17223 	if (rack->rc_force_max_seg) {
17224 		return (user_max);
17225 	}
17226 	if (rack->use_fixed_rate &&
17227 	    ((rack->r_ctl.crte == NULL) ||
17228 	     (bw != rack->r_ctl.crte->rate))) {
17229 		/* Use the user mss since we are not exactly matched */
17230 		return (user_max);
17231 	}
17232 	if (rack_pace_one_seg ||
17233 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17234 		pace_one = 1;
17235 	else
17236 		pace_one = 0;
17237 
17238 	new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17239 		     pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17240 	if (new_tso > user_max)
17241 		new_tso = user_max;
17242 	if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17243 		if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17244 			new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17245 	}
17246 	if (rack->r_ctl.rc_user_set_min_segs &&
17247 	    ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17248 	    new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17249 	return (new_tso);
17250 }
17251 
17252 static uint64_t
17253 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
17254 {
17255 	uint64_t reduced_win;
17256 	uint32_t gain;
17257 
17258 	if (window_input < rc_init_window(rack)) {
17259 		/*
17260 		 * The cwnd is collapsed to
17261 		 * nearly zero, maybe because of a time-out?
17262 		 * Lets drop back to the lt-bw.
17263 		 */
17264 		reduced_win = rack_get_lt_bw(rack);
17265 		/* Set the flag so the caller knows its a rate and not a reduced window */
17266 		*rate_set = 1;
17267 		gain = 100;
17268 	} else if  (IN_RECOVERY(rack->rc_tp->t_flags)) {
17269 		/*
17270 		 * If we are in recover our cwnd needs to be less for
17271 		 * our pacing consideration.
17272 		 */
17273 		if (rack->rack_hibeta == 0) {
17274 			reduced_win = window_input / 2;
17275 			gain = 50;
17276 		} else {
17277 			reduced_win = window_input * rack->r_ctl.saved_hibeta;
17278 			reduced_win /= 100;
17279 			gain = rack->r_ctl.saved_hibeta;
17280 		}
17281 	} else {
17282 		/*
17283 		 * Apply Timely factor to increase/decrease the
17284 		 * amount we are pacing at.
17285 		 */
17286 		gain = rack_get_output_gain(rack, NULL);
17287 		if (gain > rack_gain_p5_ub) {
17288 			gain = rack_gain_p5_ub;
17289 		}
17290 		reduced_win = window_input * gain;
17291 		reduced_win /= 100;
17292 	}
17293 	if (gain_b != NULL)
17294 		*gain_b = gain;
17295 	/*
17296 	 * What is being returned here is a trimmed down
17297 	 * window values in all cases where rate_set is left
17298 	 * at 0. In one case we actually return the rate (lt_bw).
17299 	 * the "reduced_win" is returned as a slimmed down cwnd that
17300 	 * is then calculated by the caller into a rate when rate_set
17301 	 * is 0.
17302 	 */
17303 	return (reduced_win);
17304 }
17305 
17306 static int32_t
17307 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)
17308 {
17309 	uint64_t lentim, fill_bw;
17310 
17311 	rack->r_via_fill_cw = 0;
17312 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17313 		return (slot);
17314 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17315 		return (slot);
17316 	if (rack->r_ctl.rc_last_us_rtt == 0)
17317 		return (slot);
17318 	if (rack->rc_pace_fill_if_rttin_range &&
17319 	    (rack->r_ctl.rc_last_us_rtt >=
17320 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17321 		/* The rtt is huge, N * smallest, lets not fill */
17322 		return (slot);
17323 	}
17324 	if (rack->r_ctl.fillcw_cap && *rate_wanted >= rack->r_ctl.fillcw_cap)
17325 		return (slot);
17326 	/*
17327 	 * first lets calculate the b/w based on the last us-rtt
17328 	 * and the the smallest send window.
17329 	 */
17330 	fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17331 	if (rack->rc_fillcw_apply_discount) {
17332 		uint32_t rate_set = 0;
17333 
17334 		fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
17335 		if (rate_set) {
17336 			goto at_lt_bw;
17337 		}
17338 	}
17339 	/* Take the rwnd if its smaller */
17340 	if (fill_bw > rack->rc_tp->snd_wnd)
17341 		fill_bw = rack->rc_tp->snd_wnd;
17342 	/* Now lets make it into a b/w */
17343 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17344 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17345 	/* Adjust to any cap */
17346 	if (rack->r_ctl.fillcw_cap && fill_bw >= rack->r_ctl.fillcw_cap)
17347 		fill_bw = rack->r_ctl.fillcw_cap;
17348 
17349 at_lt_bw:
17350 	if (rack_bw_multipler > 0) {
17351 		/*
17352 		 * We want to limit fill-cw to the some multiplier
17353 		 * of the max(lt_bw, gp_est). The normal default
17354 		 * is 0 for off, so a sysctl has enabled it.
17355 		 */
17356 		uint64_t lt_bw, gp, rate;
17357 
17358 		gp = rack_get_gp_est(rack);
17359 		lt_bw = rack_get_lt_bw(rack);
17360 		if (lt_bw > gp)
17361 			rate = lt_bw;
17362 		else
17363 			rate = gp;
17364 		rate *= rack_bw_multipler;
17365 		rate /= 100;
17366 		if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
17367 			union tcp_log_stackspecific log;
17368 			struct timeval tv;
17369 
17370 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17371 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17372 			log.u_bbr.flex1 = rack_bw_multipler;
17373 			log.u_bbr.flex2 = len;
17374 			log.u_bbr.cur_del_rate = gp;
17375 			log.u_bbr.delRate = lt_bw;
17376 			log.u_bbr.bw_inuse = rate;
17377 			log.u_bbr.rttProp = fill_bw;
17378 			log.u_bbr.flex8 = 44;
17379 			tcp_log_event(rack->rc_tp, NULL, NULL, NULL,
17380 				      BBR_LOG_CWND, 0,
17381 				      0, &log, false, NULL,
17382 				      __func__, __LINE__, &tv);
17383 		}
17384 		if (fill_bw > rate)
17385 			fill_bw = rate;
17386 	}
17387 	/* We are below the min b/w */
17388 	if (non_paced)
17389 		*rate_wanted = fill_bw;
17390 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17391 		return (slot);
17392 	rack->r_via_fill_cw = 1;
17393 	if (rack->r_rack_hw_rate_caps &&
17394 	    (rack->r_ctl.crte != NULL)) {
17395 		uint64_t high_rate;
17396 
17397 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17398 		if (fill_bw > high_rate) {
17399 			/* We are capping bw at the highest rate table entry */
17400 			if (*rate_wanted > high_rate) {
17401 				/* The original rate was also capped */
17402 				rack->r_via_fill_cw = 0;
17403 			}
17404 			rack_log_hdwr_pacing(rack,
17405 					     fill_bw, high_rate, __LINE__,
17406 					     0, 3);
17407 			fill_bw = high_rate;
17408 			if (capped)
17409 				*capped = 1;
17410 		}
17411 	} else if ((rack->r_ctl.crte == NULL) &&
17412 		   (rack->rack_hdrw_pacing == 0) &&
17413 		   (rack->rack_hdw_pace_ena) &&
17414 		   rack->r_rack_hw_rate_caps &&
17415 		   (rack->rack_attempt_hdwr_pace == 0) &&
17416 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
17417 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17418 		/*
17419 		 * Ok we may have a first attempt that is greater than our top rate
17420 		 * lets check.
17421 		 */
17422 		uint64_t high_rate;
17423 
17424 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17425 		if (high_rate) {
17426 			if (fill_bw > high_rate) {
17427 				fill_bw = high_rate;
17428 				if (capped)
17429 					*capped = 1;
17430 			}
17431 		}
17432 	}
17433 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17434 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17435 				   fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
17436 		fill_bw = rack->r_ctl.bw_rate_cap;
17437 	}
17438 	/*
17439 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
17440 	 * in an rtt (unless it was capped), what does that
17441 	 * time wise equate too?
17442 	 */
17443 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17444 	lentim /= fill_bw;
17445 	*rate_wanted = fill_bw;
17446 	if (non_paced || (lentim < slot)) {
17447 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17448 					   0, lentim, 12, __LINE__, NULL, 0);
17449 		return ((int32_t)lentim);
17450 	} else
17451 		return (slot);
17452 }
17453 
17454 static int32_t
17455 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz, int line)
17456 {
17457 	uint64_t srtt;
17458 	int32_t slot = 0;
17459 	int can_start_hw_pacing = 1;
17460 	int err;
17461 	int pace_one;
17462 
17463 	if (rack_pace_one_seg ||
17464 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17465 		pace_one = 1;
17466 	else
17467 		pace_one = 0;
17468 	if (rack->rc_always_pace == 0) {
17469 		/*
17470 		 * We use the most optimistic possible cwnd/srtt for
17471 		 * sending calculations. This will make our
17472 		 * calculation anticipate getting more through
17473 		 * quicker then possible. But thats ok we don't want
17474 		 * the peer to have a gap in data sending.
17475 		 */
17476 		uint64_t cwnd, tr_perms = 0;
17477 		int32_t reduce;
17478 
17479 	old_method:
17480 		/*
17481 		 * We keep no precise pacing with the old method
17482 		 * instead we use the pacer to mitigate bursts.
17483 		 */
17484 		if (rack->r_ctl.rc_rack_min_rtt)
17485 			srtt = rack->r_ctl.rc_rack_min_rtt;
17486 		else
17487 			srtt = max(tp->t_srtt, 1);
17488 		if (rack->r_ctl.rc_rack_largest_cwnd)
17489 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17490 		else
17491 			cwnd = rack->r_ctl.cwnd_to_use;
17492 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
17493 		tr_perms = (cwnd * 1000) / srtt;
17494 		if (tr_perms == 0) {
17495 			tr_perms = ctf_fixed_maxseg(tp);
17496 		}
17497 		/*
17498 		 * Calculate how long this will take to drain, if
17499 		 * the calculation comes out to zero, thats ok we
17500 		 * will use send_a_lot to possibly spin around for
17501 		 * more increasing tot_len_this_send to the point
17502 		 * that its going to require a pace, or we hit the
17503 		 * cwnd. Which in that case we are just waiting for
17504 		 * a ACK.
17505 		 */
17506 		slot = len / tr_perms;
17507 		/* Now do we reduce the time so we don't run dry? */
17508 		if (slot && rack_slot_reduction) {
17509 			reduce = (slot / rack_slot_reduction);
17510 			if (reduce < slot) {
17511 				slot -= reduce;
17512 			} else
17513 				slot = 0;
17514 		} else
17515 			reduce = 0;
17516 		slot *= HPTS_USEC_IN_MSEC;
17517 		if (rack->rc_pace_to_cwnd) {
17518 			uint64_t rate_wanted = 0;
17519 
17520 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17521 			rack->rc_ack_can_sendout_data = 1;
17522 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17523 		} else
17524 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17525 		/*******************************************************/
17526 		/* RRS: We insert non-paced call to stats here for len */
17527 		/*******************************************************/
17528 	} else {
17529 		uint64_t bw_est, res, lentim, rate_wanted;
17530 		uint32_t segs, oh;
17531 		int capped = 0;
17532 		int prev_fill;
17533 
17534 		if ((rack->r_rr_config == 1) && rsm) {
17535 			return (rack->r_ctl.rc_min_to);
17536 		}
17537 		if (rack->use_fixed_rate) {
17538 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17539 		} else if ((rack->r_ctl.init_rate == 0) &&
17540 			   (rack->r_ctl.gp_bw == 0)) {
17541 			/* no way to yet do an estimate */
17542 			bw_est = rate_wanted = 0;
17543 		} else if (rack->dgp_on) {
17544 			bw_est = rack_get_bw(rack);
17545 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17546 		} else {
17547 			uint32_t gain, rate_set = 0;
17548 
17549 			rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17550 			rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
17551 			if (rate_set == 0) {
17552 				if (rate_wanted > rack->rc_tp->snd_wnd)
17553 					rate_wanted = rack->rc_tp->snd_wnd;
17554 				/* Now lets make it into a b/w */
17555 				rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
17556 				rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17557 			}
17558 			bw_est = rate_wanted;
17559 			rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
17560 						   rack->r_ctl.cwnd_to_use,
17561 						   rate_wanted, bw_est,
17562 						   rack->r_ctl.rc_last_us_rtt,
17563 						   88, __LINE__, NULL, gain);
17564 		}
17565 		if ((bw_est == 0) || (rate_wanted == 0) ||
17566 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
17567 			/*
17568 			 * No way yet to make a b/w estimate or
17569 			 * our raise is set incorrectly.
17570 			 */
17571 			goto old_method;
17572 		}
17573 		rack_rate_cap_bw(rack, &rate_wanted, &capped);
17574 		/* We need to account for all the overheads */
17575 		segs = (len + segsiz - 1) / segsiz;
17576 		/*
17577 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
17578 		 * and how much data we put in each packet. Yes this
17579 		 * means we may be off if we are larger than 1500 bytes
17580 		 * or smaller. But this just makes us more conservative.
17581 		 */
17582 
17583 		oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17584 		if (rack->r_is_v6) {
17585 #ifdef INET6
17586 			oh += sizeof(struct ip6_hdr);
17587 #endif
17588 		} else {
17589 #ifdef INET
17590 			oh += sizeof(struct ip);
17591 #endif
17592 		}
17593 		/* We add a fixed 14 for the ethernet header */
17594 		oh += 14;
17595 		segs *= oh;
17596 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17597 		res = lentim / rate_wanted;
17598 		slot = (uint32_t)res;
17599 		if (rack_hw_rate_min &&
17600 		    (rate_wanted < rack_hw_rate_min)) {
17601 			can_start_hw_pacing = 0;
17602 			if (rack->r_ctl.crte) {
17603 				/*
17604 				 * Ok we need to release it, we
17605 				 * have fallen too low.
17606 				 */
17607 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17608 				rack->r_ctl.crte = NULL;
17609 				rack->rack_attempt_hdwr_pace = 0;
17610 				rack->rack_hdrw_pacing = 0;
17611 			}
17612 		}
17613 		if (rack->r_ctl.crte &&
17614 		    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17615 			/*
17616 			 * We want more than the hardware can give us,
17617 			 * don't start any hw pacing.
17618 			 */
17619 			can_start_hw_pacing = 0;
17620 			if (rack->r_rack_hw_rate_caps == 0) {
17621 				/*
17622 				 * Ok we need to release it, we
17623 				 * want more than the card can give us and
17624 				 * no rate cap is in place. Set it up so
17625 				 * when we want less we can retry.
17626 				 */
17627 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17628 				rack->r_ctl.crte = NULL;
17629 				rack->rack_attempt_hdwr_pace = 0;
17630 				rack->rack_hdrw_pacing = 0;
17631 			}
17632 		}
17633 		if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17634 			/*
17635 			 * We lost our rate somehow, this can happen
17636 			 * if the interface changed underneath us.
17637 			 */
17638 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17639 			rack->r_ctl.crte = NULL;
17640 			/* Lets re-allow attempting to setup pacing */
17641 			rack->rack_hdrw_pacing = 0;
17642 			rack->rack_attempt_hdwr_pace = 0;
17643 			rack_log_hdwr_pacing(rack,
17644 					     rate_wanted, bw_est, __LINE__,
17645 					     0, 6);
17646 		}
17647 		prev_fill = rack->r_via_fill_cw;
17648 		if ((rack->rc_pace_to_cwnd) &&
17649 		    (capped == 0) &&
17650 		    (rack->dgp_on == 1) &&
17651 		    (rack->use_fixed_rate == 0) &&
17652 		    (rack->in_probe_rtt == 0) &&
17653 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17654 			/*
17655 			 * We want to pace at our rate *or* faster to
17656 			 * fill the cwnd to the max if its not full.
17657 			 */
17658 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17659 			/* Re-check to make sure we are not exceeding our max b/w */
17660 			if ((rack->r_ctl.crte != NULL) &&
17661 			    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17662 				/*
17663 				 * We want more than the hardware can give us,
17664 				 * don't start any hw pacing.
17665 				 */
17666 				can_start_hw_pacing = 0;
17667 				if (rack->r_rack_hw_rate_caps == 0) {
17668 					/*
17669 					 * Ok we need to release it, we
17670 					 * want more than the card can give us and
17671 					 * no rate cap is in place. Set it up so
17672 					 * when we want less we can retry.
17673 					 */
17674 					tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17675 					rack->r_ctl.crte = NULL;
17676 					rack->rack_attempt_hdwr_pace = 0;
17677 					rack->rack_hdrw_pacing = 0;
17678 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17679 				}
17680 			}
17681 		}
17682 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17683 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17684 			if ((rack->rack_hdw_pace_ena) &&
17685 			    (can_start_hw_pacing > 0) &&
17686 			    (rack->rack_hdrw_pacing == 0) &&
17687 			    (rack->rack_attempt_hdwr_pace == 0)) {
17688 				/*
17689 				 * Lets attempt to turn on hardware pacing
17690 				 * if we can.
17691 				 */
17692 				rack->rack_attempt_hdwr_pace = 1;
17693 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17694 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
17695 								       rate_wanted,
17696 								       RS_PACING_GEQ,
17697 								       &err, &rack->r_ctl.crte_prev_rate);
17698 				if (rack->r_ctl.crte) {
17699 					rack->rack_hdrw_pacing = 1;
17700 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17701 													   pace_one, rack->r_ctl.crte,
17702 													   NULL, rack->r_ctl.pace_len_divisor);
17703 					rack_log_hdwr_pacing(rack,
17704 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17705 							     err, 0);
17706 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17707 				} else {
17708 					counter_u64_add(rack_hw_pace_init_fail, 1);
17709 				}
17710 			} else if (rack->rack_hdrw_pacing &&
17711 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17712 				/* Do we need to adjust our rate? */
17713 				const struct tcp_hwrate_limit_table *nrte;
17714 
17715 				if (rack->r_up_only &&
17716 				    (rate_wanted < rack->r_ctl.crte->rate)) {
17717 					/**
17718 					 * We have four possible states here
17719 					 * having to do with the previous time
17720 					 * and this time.
17721 					 *   previous  |  this-time
17722 					 * A)     0      |     0   -- fill_cw not in the picture
17723 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
17724 					 * C)     1      |     1   -- all rates from fill_cw
17725 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
17726 					 *
17727 					 * For case A, C and D we don't allow a drop. But for
17728 					 * case B where we now our on our steady rate we do
17729 					 * allow a drop.
17730 					 *
17731 					 */
17732 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17733 						goto done_w_hdwr;
17734 				}
17735 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
17736 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17737 					if (rack_hw_rate_to_low &&
17738 					    (bw_est < rack_hw_rate_to_low)) {
17739 						/*
17740 						 * The pacing rate is too low for hardware, but
17741 						 * do allow hardware pacing to be restarted.
17742 						 */
17743 						rack_log_hdwr_pacing(rack,
17744 								     bw_est, rack->r_ctl.crte->rate, __LINE__,
17745 								     0, 5);
17746 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17747 						rack->r_ctl.crte = NULL;
17748 						rack->rack_attempt_hdwr_pace = 0;
17749 						rack->rack_hdrw_pacing = 0;
17750 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17751 						goto done_w_hdwr;
17752 					}
17753 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17754 								   rack->rc_tp,
17755 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
17756 								   rate_wanted,
17757 								   RS_PACING_GEQ,
17758 								   &err, &rack->r_ctl.crte_prev_rate);
17759 					if (nrte == NULL) {
17760 						/*
17761 						 * Lost the rate, lets drop hardware pacing
17762 						 * period.
17763 						 */
17764 						rack->rack_hdrw_pacing = 0;
17765 						rack->r_ctl.crte = NULL;
17766 						rack_log_hdwr_pacing(rack,
17767 								     rate_wanted, 0, __LINE__,
17768 								     err, 1);
17769 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17770 						counter_u64_add(rack_hw_pace_lost, 1);
17771 					} else if (nrte != rack->r_ctl.crte) {
17772 						rack->r_ctl.crte = nrte;
17773 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17774 														   segsiz, pace_one, rack->r_ctl.crte,
17775 														   NULL, rack->r_ctl.pace_len_divisor);
17776 						rack_log_hdwr_pacing(rack,
17777 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17778 								     err, 2);
17779 						rack->r_ctl.last_hw_bw_req = rate_wanted;
17780 					}
17781 				} else {
17782 					/* We just need to adjust the segment size */
17783 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17784 					rack_log_hdwr_pacing(rack,
17785 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17786 							     0, 4);
17787 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17788 				}
17789 			}
17790 		}
17791 	done_w_hdwr:
17792 		if (rack_limit_time_with_srtt &&
17793 		    (rack->use_fixed_rate == 0) &&
17794 		    (rack->rack_hdrw_pacing == 0)) {
17795 			/*
17796 			 * Sanity check, we do not allow the pacing delay
17797 			 * to be longer than the SRTT of the path. If it is
17798 			 * a slow path, then adding a packet should increase
17799 			 * the RTT and compensate for this i.e. the srtt will
17800 			 * be greater so the allowed pacing time will be greater.
17801 			 *
17802 			 * Note this restriction is not for where a peak rate
17803 			 * is set, we are doing fixed pacing or hardware pacing.
17804 			 */
17805 			if (rack->rc_tp->t_srtt)
17806 				srtt = rack->rc_tp->t_srtt;
17807 			else
17808 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
17809 			if (srtt < (uint64_t)slot) {
17810 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
17811 				slot = srtt;
17812 			}
17813 		}
17814 		/*******************************************************************/
17815 		/* RRS: We insert paced call to stats here for len and rate_wanted */
17816 		/*******************************************************************/
17817 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
17818 	}
17819 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
17820 		/*
17821 		 * If this rate is seeing enobufs when it
17822 		 * goes to send then either the nic is out
17823 		 * of gas or we are mis-estimating the time
17824 		 * somehow and not letting the queue empty
17825 		 * completely. Lets add to the pacing time.
17826 		 */
17827 		int hw_boost_delay;
17828 
17829 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
17830 		if (hw_boost_delay > rack_enobuf_hw_max)
17831 			hw_boost_delay = rack_enobuf_hw_max;
17832 		else if (hw_boost_delay < rack_enobuf_hw_min)
17833 			hw_boost_delay = rack_enobuf_hw_min;
17834 		slot += hw_boost_delay;
17835 	}
17836 	return (slot);
17837 }
17838 
17839 static void
17840 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
17841     tcp_seq startseq, uint32_t sb_offset)
17842 {
17843 	struct rack_sendmap *my_rsm = NULL;
17844 
17845 	if (tp->t_state < TCPS_ESTABLISHED) {
17846 		/*
17847 		 * We don't start any measurements if we are
17848 		 * not at least established.
17849 		 */
17850 		return;
17851 	}
17852 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
17853 		/*
17854 		 * We will get no more data into the SB
17855 		 * this means we need to have the data available
17856 		 * before we start a measurement.
17857 		 */
17858 
17859 		if (sbavail(&tptosocket(tp)->so_snd) <
17860 		    max(rc_init_window(rack),
17861 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
17862 			/* Nope not enough data */
17863 			return;
17864 		}
17865 	}
17866 	tp->t_flags |= TF_GPUTINPROG;
17867 	rack->r_ctl.rc_gp_cumack_ts = 0;
17868 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
17869 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
17870 	tp->gput_seq = startseq;
17871 	rack->app_limited_needs_set = 0;
17872 	if (rack->in_probe_rtt)
17873 		rack->measure_saw_probe_rtt = 1;
17874 	else if ((rack->measure_saw_probe_rtt) &&
17875 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
17876 		rack->measure_saw_probe_rtt = 0;
17877 	if (rack->rc_gp_filled)
17878 		tp->gput_ts = rack->r_ctl.last_cumack_advance;
17879 	else {
17880 		/* Special case initial measurement */
17881 		struct timeval tv;
17882 
17883 		tp->gput_ts = tcp_get_usecs(&tv);
17884 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
17885 	}
17886 	/*
17887 	 * We take a guess out into the future,
17888 	 * if we have no measurement and no
17889 	 * initial rate, we measure the first
17890 	 * initial-windows worth of data to
17891 	 * speed up getting some GP measurement and
17892 	 * thus start pacing.
17893 	 */
17894 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
17895 		rack->app_limited_needs_set = 1;
17896 		tp->gput_ack = startseq + max(rc_init_window(rack),
17897 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
17898 		rack_log_pacing_delay_calc(rack,
17899 					   tp->gput_seq,
17900 					   tp->gput_ack,
17901 					   0,
17902 					   tp->gput_ts,
17903 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17904 					   9,
17905 					   __LINE__, NULL, 0);
17906 		rack_tend_gp_marks(tp, rack);
17907 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
17908 		return;
17909 	}
17910 	if (sb_offset) {
17911 		/*
17912 		 * We are out somewhere in the sb
17913 		 * can we use the already outstanding data?
17914 		 */
17915 
17916 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
17917 			/*
17918 			 * Yes first one is good and in this case
17919 			 * the tp->gput_ts is correctly set based on
17920 			 * the last ack that arrived (no need to
17921 			 * set things up when an ack comes in).
17922 			 */
17923 			my_rsm = tqhash_min(rack->r_ctl.tqh);
17924 			if ((my_rsm == NULL) ||
17925 			    (my_rsm->r_rtr_cnt != 1)) {
17926 				/* retransmission? */
17927 				goto use_latest;
17928 			}
17929 		} else {
17930 			if (rack->r_ctl.rc_first_appl == NULL) {
17931 				/*
17932 				 * If rc_first_appl is NULL
17933 				 * then the cnt should be 0.
17934 				 * This is probably an error, maybe
17935 				 * a KASSERT would be approprate.
17936 				 */
17937 				goto use_latest;
17938 			}
17939 			/*
17940 			 * If we have a marker pointer to the last one that is
17941 			 * app limited we can use that, but we need to set
17942 			 * things up so that when it gets ack'ed we record
17943 			 * the ack time (if its not already acked).
17944 			 */
17945 			rack->app_limited_needs_set = 1;
17946 			/*
17947 			 * We want to get to the rsm that is either
17948 			 * next with space i.e. over 1 MSS or the one
17949 			 * after that (after the app-limited).
17950 			 */
17951 			my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
17952 			if (my_rsm) {
17953 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
17954 					/* Have to use the next one */
17955 					my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17956 				else {
17957 					/* Use after the first MSS of it is acked */
17958 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
17959 					goto start_set;
17960 				}
17961 			}
17962 			if ((my_rsm == NULL) ||
17963 			    (my_rsm->r_rtr_cnt != 1)) {
17964 				/*
17965 				 * Either its a retransmit or
17966 				 * the last is the app-limited one.
17967 				 */
17968 				goto use_latest;
17969 			}
17970 		}
17971 		tp->gput_seq = my_rsm->r_start;
17972 start_set:
17973 		if (my_rsm->r_flags & RACK_ACKED) {
17974 			/*
17975 			 * This one has been acked use the arrival ack time
17976 			 */
17977 			struct rack_sendmap *nrsm;
17978 
17979 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
17980 			rack->app_limited_needs_set = 0;
17981 			/*
17982 			 * Ok in this path we need to use the r_end now
17983 			 * since this guy is the starting ack.
17984 			 */
17985 			tp->gput_seq = my_rsm->r_end;
17986 			/*
17987 			 * We also need to adjust up the sendtime
17988 			 * to the send of the next data after my_rsm.
17989 			 */
17990 			nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17991 			if (nrsm != NULL)
17992 				my_rsm = nrsm;
17993 			else {
17994 				/*
17995 				 * The next as not been sent, thats the
17996 				 * case for using the latest.
17997 				 */
17998 				goto use_latest;
17999 			}
18000 		}
18001 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18002 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
18003 		rack->r_ctl.rc_gp_cumack_ts = 0;
18004 		if ((rack->r_ctl.cleared_app_ack == 1) &&
18005 		    (SEQ_GEQ(rack->r_ctl.cleared_app_ack, tp->gput_seq))) {
18006 			/*
18007 			 * We just cleared an application limited period
18008 			 * so the next seq out needs to skip the first
18009 			 * ack.
18010 			 */
18011 			rack->app_limited_needs_set = 1;
18012 			rack->r_ctl.cleared_app_ack = 0;
18013 		}
18014 		rack_log_pacing_delay_calc(rack,
18015 					   tp->gput_seq,
18016 					   tp->gput_ack,
18017 					   (uintptr_t)my_rsm,
18018 					   tp->gput_ts,
18019 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18020 					   9,
18021 					   __LINE__, my_rsm, 0);
18022 		/* Now lets make sure all are marked as they should be */
18023 		rack_tend_gp_marks(tp, rack);
18024 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18025 		return;
18026 	}
18027 
18028 use_latest:
18029 	/*
18030 	 * We don't know how long we may have been
18031 	 * idle or if this is the first-send. Lets
18032 	 * setup the flag so we will trim off
18033 	 * the first ack'd data so we get a true
18034 	 * measurement.
18035 	 */
18036 	rack->app_limited_needs_set = 1;
18037 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
18038 	rack->r_ctl.rc_gp_cumack_ts = 0;
18039 	/* Find this guy so we can pull the send time */
18040 	my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
18041 	if (my_rsm) {
18042 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18043 		if (my_rsm->r_flags & RACK_ACKED) {
18044 			/*
18045 			 * Unlikely since its probably what was
18046 			 * just transmitted (but I am paranoid).
18047 			 */
18048 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18049 			rack->app_limited_needs_set = 0;
18050 		}
18051 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18052 			/* This also is unlikely */
18053 			tp->gput_seq = my_rsm->r_start;
18054 		}
18055 	} else {
18056 		/*
18057 		 * TSNH unless we have some send-map limit,
18058 		 * and even at that it should not be hitting
18059 		 * that limit (we should have stopped sending).
18060 		 */
18061 		struct timeval tv;
18062 
18063 		microuptime(&tv);
18064 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18065 	}
18066 	rack_tend_gp_marks(tp, rack);
18067 	rack_log_pacing_delay_calc(rack,
18068 				   tp->gput_seq,
18069 				   tp->gput_ack,
18070 				   (uintptr_t)my_rsm,
18071 				   tp->gput_ts,
18072 				   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18073 				   9, __LINE__, NULL, 0);
18074 	rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18075 }
18076 
18077 static inline uint32_t
18078 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
18079     uint32_t avail, int32_t sb_offset)
18080 {
18081 	uint32_t len;
18082 	uint32_t sendwin;
18083 
18084 	if (tp->snd_wnd > cwnd_to_use)
18085 		sendwin = cwnd_to_use;
18086 	else
18087 		sendwin = tp->snd_wnd;
18088 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
18089 		/* We never want to go over our peers rcv-window */
18090 		len = 0;
18091 	} else {
18092 		uint32_t flight;
18093 
18094 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18095 		if (flight >= sendwin) {
18096 			/*
18097 			 * We have in flight what we are allowed by cwnd (if
18098 			 * it was rwnd blocking it would have hit above out
18099 			 * >= tp->snd_wnd).
18100 			 */
18101 			return (0);
18102 		}
18103 		len = sendwin - flight;
18104 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18105 			/* We would send too much (beyond the rwnd) */
18106 			len = tp->snd_wnd - ctf_outstanding(tp);
18107 		}
18108 		if ((len + sb_offset) > avail) {
18109 			/*
18110 			 * We don't have that much in the SB, how much is
18111 			 * there?
18112 			 */
18113 			len = avail - sb_offset;
18114 		}
18115 	}
18116 	return (len);
18117 }
18118 
18119 static void
18120 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18121 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18122 	     int rsm_is_null, int optlen, int line, uint16_t mode)
18123 {
18124 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18125 		union tcp_log_stackspecific log;
18126 		struct timeval tv;
18127 
18128 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18129 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18130 		log.u_bbr.flex1 = error;
18131 		log.u_bbr.flex2 = flags;
18132 		log.u_bbr.flex3 = rsm_is_null;
18133 		log.u_bbr.flex4 = ipoptlen;
18134 		log.u_bbr.flex5 = tp->rcv_numsacks;
18135 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18136 		log.u_bbr.flex7 = optlen;
18137 		log.u_bbr.flex8 = rack->r_fsb_inited;
18138 		log.u_bbr.applimited = rack->r_fast_output;
18139 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18140 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18141 		log.u_bbr.cwnd_gain = mode;
18142 		log.u_bbr.pkts_out = orig_len;
18143 		log.u_bbr.lt_epoch = len;
18144 		log.u_bbr.delivered = line;
18145 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18146 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18147 		tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18148 			       len, &log, false, NULL, __func__, __LINE__, &tv);
18149 	}
18150 }
18151 
18152 
18153 static struct mbuf *
18154 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18155 		   struct rack_fast_send_blk *fsb,
18156 		   int32_t seglimit, int32_t segsize, int hw_tls)
18157 {
18158 #ifdef KERN_TLS
18159 	struct ktls_session *tls, *ntls;
18160 #ifdef INVARIANTS
18161 	struct mbuf *start;
18162 #endif
18163 #endif
18164 	struct mbuf *m, *n, **np, *smb;
18165 	struct mbuf *top;
18166 	int32_t off, soff;
18167 	int32_t len = *plen;
18168 	int32_t fragsize;
18169 	int32_t len_cp = 0;
18170 	uint32_t mlen, frags;
18171 
18172 	soff = off = the_off;
18173 	smb = m = the_m;
18174 	np = &top;
18175 	top = NULL;
18176 #ifdef KERN_TLS
18177 	if (hw_tls && (m->m_flags & M_EXTPG))
18178 		tls = m->m_epg_tls;
18179 	else
18180 		tls = NULL;
18181 #ifdef INVARIANTS
18182 	start = m;
18183 #endif
18184 #endif
18185 	while (len > 0) {
18186 		if (m == NULL) {
18187 			*plen = len_cp;
18188 			break;
18189 		}
18190 #ifdef KERN_TLS
18191 		if (hw_tls) {
18192 			if (m->m_flags & M_EXTPG)
18193 				ntls = m->m_epg_tls;
18194 			else
18195 				ntls = NULL;
18196 
18197 			/*
18198 			 * Avoid mixing TLS records with handshake
18199 			 * data or TLS records from different
18200 			 * sessions.
18201 			 */
18202 			if (tls != ntls) {
18203 				MPASS(m != start);
18204 				*plen = len_cp;
18205 				break;
18206 			}
18207 		}
18208 #endif
18209 		mlen = min(len, m->m_len - off);
18210 		if (seglimit) {
18211 			/*
18212 			 * For M_EXTPG mbufs, add 3 segments
18213 			 * + 1 in case we are crossing page boundaries
18214 			 * + 2 in case the TLS hdr/trailer are used
18215 			 * It is cheaper to just add the segments
18216 			 * than it is to take the cache miss to look
18217 			 * at the mbuf ext_pgs state in detail.
18218 			 */
18219 			if (m->m_flags & M_EXTPG) {
18220 				fragsize = min(segsize, PAGE_SIZE);
18221 				frags = 3;
18222 			} else {
18223 				fragsize = segsize;
18224 				frags = 0;
18225 			}
18226 
18227 			/* Break if we really can't fit anymore. */
18228 			if ((frags + 1) >= seglimit) {
18229 				*plen =	len_cp;
18230 				break;
18231 			}
18232 
18233 			/*
18234 			 * Reduce size if you can't copy the whole
18235 			 * mbuf. If we can't copy the whole mbuf, also
18236 			 * adjust len so the loop will end after this
18237 			 * mbuf.
18238 			 */
18239 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18240 				mlen = (seglimit - frags - 1) * fragsize;
18241 				len = mlen;
18242 				*plen = len_cp + len;
18243 			}
18244 			frags += howmany(mlen, fragsize);
18245 			if (frags == 0)
18246 				frags++;
18247 			seglimit -= frags;
18248 			KASSERT(seglimit > 0,
18249 			    ("%s: seglimit went too low", __func__));
18250 		}
18251 		n = m_get(M_NOWAIT, m->m_type);
18252 		*np = n;
18253 		if (n == NULL)
18254 			goto nospace;
18255 		n->m_len = mlen;
18256 		soff += mlen;
18257 		len_cp += n->m_len;
18258 		if (m->m_flags & (M_EXT | M_EXTPG)) {
18259 			n->m_data = m->m_data + off;
18260 			mb_dupcl(n, m);
18261 		} else {
18262 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18263 			    (u_int)n->m_len);
18264 		}
18265 		len -= n->m_len;
18266 		off = 0;
18267 		m = m->m_next;
18268 		np = &n->m_next;
18269 		if (len || (soff == smb->m_len)) {
18270 			/*
18271 			 * We have more so we move forward  or
18272 			 * we have consumed the entire mbuf and
18273 			 * len has fell to 0.
18274 			 */
18275 			soff = 0;
18276 			smb = m;
18277 		}
18278 
18279 	}
18280 	if (fsb != NULL) {
18281 		fsb->m = smb;
18282 		fsb->off = soff;
18283 		if (smb) {
18284 			/*
18285 			 * Save off the size of the mbuf. We do
18286 			 * this so that we can recognize when it
18287 			 * has been trimmed by sbcut() as acks
18288 			 * come in.
18289 			 */
18290 			fsb->o_m_len = smb->m_len;
18291 			fsb->o_t_len = M_TRAILINGROOM(smb);
18292 		} else {
18293 			/*
18294 			 * This is the case where the next mbuf went to NULL. This
18295 			 * means with this copy we have sent everything in the sb.
18296 			 * In theory we could clear the fast_output flag, but lets
18297 			 * not since its possible that we could get more added
18298 			 * and acks that call the extend function which would let
18299 			 * us send more.
18300 			 */
18301 			fsb->o_m_len = 0;
18302 			fsb->o_t_len = 0;
18303 		}
18304 	}
18305 	return (top);
18306 nospace:
18307 	if (top)
18308 		m_freem(top);
18309 	return (NULL);
18310 
18311 }
18312 
18313 /*
18314  * This is a copy of m_copym(), taking the TSO segment size/limit
18315  * constraints into account, and advancing the sndptr as it goes.
18316  */
18317 static struct mbuf *
18318 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18319 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18320 {
18321 	struct mbuf *m, *n;
18322 	int32_t soff;
18323 
18324 	m = rack->r_ctl.fsb.m;
18325 	if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18326 		/*
18327 		 * The trailing space changed, mbufs can grow
18328 		 * at the tail but they can't shrink from
18329 		 * it, KASSERT that. Adjust the orig_m_len to
18330 		 * compensate for this change.
18331 		 */
18332 		KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18333 			("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18334 			 m,
18335 			 rack,
18336 			 (intmax_t)M_TRAILINGROOM(m),
18337 			 rack->r_ctl.fsb.o_t_len,
18338 			 rack->r_ctl.fsb.o_m_len,
18339 			 m->m_len));
18340 		rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18341 		rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18342 	}
18343 	if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18344 		/*
18345 		 * Mbuf shrank, trimmed off the top by an ack, our
18346 		 * offset changes.
18347 		 */
18348 		KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18349 			("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18350 			 m, m->m_len,
18351 			 rack, rack->r_ctl.fsb.o_m_len,
18352 			 rack->r_ctl.fsb.off));
18353 
18354 		if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18355 			rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18356 		else
18357 			rack->r_ctl.fsb.off = 0;
18358 		rack->r_ctl.fsb.o_m_len = m->m_len;
18359 #ifdef INVARIANTS
18360 	} else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18361 		panic("rack:%p m:%p m_len grew outside of t_space compensation",
18362 		      rack, m);
18363 #endif
18364 	}
18365 	soff = rack->r_ctl.fsb.off;
18366 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18367 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18368 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18369 				 __FUNCTION__,
18370 				 rack, *plen, m, m->m_len));
18371 	/* Save off the right location before we copy and advance */
18372 	*s_soff = soff;
18373 	*s_mb = rack->r_ctl.fsb.m;
18374 	n = rack_fo_base_copym(m, soff, plen,
18375 			       &rack->r_ctl.fsb,
18376 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18377 	return (n);
18378 }
18379 
18380 /* Log the buffer level */
18381 static void
18382 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18383 		     int len, struct timeval *tv,
18384 		     uint32_t cts)
18385 {
18386 	uint32_t p_rate = 0, p_queue = 0, err = 0;
18387 	union tcp_log_stackspecific log;
18388 
18389 #ifdef RATELIMIT
18390 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18391 	err = in_pcbquery_txrtlmt(rack->rc_inp,	&p_rate);
18392 #endif
18393 	memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18394 	log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18395 	log.u_bbr.flex1 = p_rate;
18396 	log.u_bbr.flex2 = p_queue;
18397 	log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18398 	log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18399 	log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18400 	log.u_bbr.flex7 = 99;
18401 	log.u_bbr.flex8 = 0;
18402 	log.u_bbr.pkts_out = err;
18403 	log.u_bbr.delRate = rack->r_ctl.crte->rate;
18404 	log.u_bbr.timeStamp = cts;
18405 	log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18406 	tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18407 		       len, &log, false, NULL, __func__, __LINE__, tv);
18408 
18409 }
18410 
18411 static uint32_t
18412 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18413 		       struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18414 {
18415 	uint64_t lentime = 0;
18416 #ifdef RATELIMIT
18417 	uint32_t p_rate = 0, p_queue = 0, err;
18418 	union tcp_log_stackspecific log;
18419 	uint64_t bw;
18420 
18421 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18422 	/* Failed or queue is zero */
18423 	if (err || (p_queue == 0)) {
18424 		lentime = 0;
18425 		goto out;
18426 	}
18427 	err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18428 	if (err) {
18429 		lentime = 0;
18430 		goto out;
18431 	}
18432 	/*
18433 	 * If we reach here we have some bytes in
18434 	 * the queue. The number returned is a value
18435 	 * between 0 and 0xffff where ffff is full
18436 	 * and 0 is empty. So how best to make this into
18437 	 * something usable?
18438 	 *
18439 	 * The "safer" way is lets take the b/w gotten
18440 	 * from the query (which should be our b/w rate)
18441 	 * and pretend that a full send (our rc_pace_max_segs)
18442 	 * is outstanding. We factor it so its as if a full
18443 	 * number of our MSS segment is terms of full
18444 	 * ethernet segments are outstanding.
18445 	 */
18446 	bw = p_rate / 8;
18447 	if (bw) {
18448 		lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18449 		lentime *= ETHERNET_SEGMENT_SIZE;
18450 		lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18451 		lentime /= bw;
18452 	} else {
18453 		/* TSNH -- KASSERT? */
18454 		lentime = 0;
18455 	}
18456 out:
18457 	if (tcp_bblogging_on(tp)) {
18458 		memset(&log, 0, sizeof(log));
18459 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18460 		log.u_bbr.flex1 = p_rate;
18461 		log.u_bbr.flex2 = p_queue;
18462 		log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18463 		log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18464 		log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18465 		log.u_bbr.flex7 = 99;
18466 		log.u_bbr.flex8 = 0;
18467 		log.u_bbr.pkts_out = err;
18468 		log.u_bbr.delRate = rack->r_ctl.crte->rate;
18469 		log.u_bbr.cur_del_rate = lentime;
18470 		log.u_bbr.timeStamp = cts;
18471 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18472 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18473 			       len, &log, false, NULL, __func__, __LINE__,tv);
18474 	}
18475 #endif
18476 	return ((uint32_t)lentime);
18477 }
18478 
18479 static int
18480 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18481 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18482 {
18483 	/*
18484 	 * Enter the fast retransmit path. We are given that a sched_pin is
18485 	 * in place (if accounting is compliled in) and the cycle count taken
18486 	 * at the entry is in the ts_val. The concept her is that the rsm
18487 	 * now holds the mbuf offsets and such so we can directly transmit
18488 	 * without a lot of overhead, the len field is already set for
18489 	 * us to prohibit us from sending too much (usually its 1MSS).
18490 	 */
18491 	struct ip *ip = NULL;
18492 	struct udphdr *udp = NULL;
18493 	struct tcphdr *th = NULL;
18494 	struct mbuf *m = NULL;
18495 	struct inpcb *inp;
18496 	uint8_t *cpto;
18497 	struct tcp_log_buffer *lgb;
18498 #ifdef TCP_ACCOUNTING
18499 	uint64_t crtsc;
18500 	int cnt_thru = 1;
18501 #endif
18502 	struct tcpopt to;
18503 	u_char opt[TCP_MAXOLEN];
18504 	uint32_t hdrlen, optlen;
18505 	int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18506 	uint16_t flags;
18507 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
18508 	uint32_t if_hw_tsomaxsegsize;
18509 	int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18510 
18511 #ifdef INET6
18512 	struct ip6_hdr *ip6 = NULL;
18513 
18514 	if (rack->r_is_v6) {
18515 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18516 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18517 	} else
18518 #endif				/* INET6 */
18519 	{
18520 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18521 		hdrlen = sizeof(struct tcpiphdr);
18522 	}
18523 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18524 		goto failed;
18525 	}
18526 	if (doing_tlp) {
18527 		/* Its a TLP add the flag, it may already be there but be sure */
18528 		rsm->r_flags |= RACK_TLP;
18529 	} else {
18530 		/* If it was a TLP it is not not on this retransmit */
18531 		rsm->r_flags &= ~RACK_TLP;
18532 	}
18533 	startseq = rsm->r_start;
18534 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18535 	inp = rack->rc_inp;
18536 	to.to_flags = 0;
18537 	flags = tcp_outflags[tp->t_state];
18538 	if (flags & (TH_SYN|TH_RST)) {
18539 		goto failed;
18540 	}
18541 	if (rsm->r_flags & RACK_HAS_FIN) {
18542 		/* We can't send a FIN here */
18543 		goto failed;
18544 	}
18545 	if (flags & TH_FIN) {
18546 		/* We never send a FIN */
18547 		flags &= ~TH_FIN;
18548 	}
18549 	if (tp->t_flags & TF_RCVD_TSTMP) {
18550 		to.to_tsval = ms_cts + tp->ts_offset;
18551 		to.to_tsecr = tp->ts_recent;
18552 		to.to_flags = TOF_TS;
18553 	}
18554 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18555 	/* TCP-MD5 (RFC2385). */
18556 	if (tp->t_flags & TF_SIGNATURE)
18557 		to.to_flags |= TOF_SIGNATURE;
18558 #endif
18559 	optlen = tcp_addoptions(&to, opt);
18560 	hdrlen += optlen;
18561 	udp = rack->r_ctl.fsb.udp;
18562 	if (udp)
18563 		hdrlen += sizeof(struct udphdr);
18564 	if (rack->r_ctl.rc_pace_max_segs)
18565 		max_val = rack->r_ctl.rc_pace_max_segs;
18566 	else if (rack->rc_user_set_max_segs)
18567 		max_val = rack->rc_user_set_max_segs * segsiz;
18568 	else
18569 		max_val = len;
18570 	if ((tp->t_flags & TF_TSO) &&
18571 	    V_tcp_do_tso &&
18572 	    (len > segsiz) &&
18573 	    (tp->t_port == 0))
18574 		tso = 1;
18575 #ifdef INET6
18576 	if (MHLEN < hdrlen + max_linkhdr)
18577 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18578 	else
18579 #endif
18580 		m = m_gethdr(M_NOWAIT, MT_DATA);
18581 	if (m == NULL)
18582 		goto failed;
18583 	m->m_data += max_linkhdr;
18584 	m->m_len = hdrlen;
18585 	th = rack->r_ctl.fsb.th;
18586 	/* Establish the len to send */
18587 	if (len > max_val)
18588 		len = max_val;
18589 	if ((tso) && (len + optlen > segsiz)) {
18590 		uint32_t if_hw_tsomax;
18591 		int32_t max_len;
18592 
18593 		/* extract TSO information */
18594 		if_hw_tsomax = tp->t_tsomax;
18595 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18596 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18597 		/*
18598 		 * Check if we should limit by maximum payload
18599 		 * length:
18600 		 */
18601 		if (if_hw_tsomax != 0) {
18602 			/* compute maximum TSO length */
18603 			max_len = (if_hw_tsomax - hdrlen -
18604 				   max_linkhdr);
18605 			if (max_len <= 0) {
18606 				goto failed;
18607 			} else if (len > max_len) {
18608 				len = max_len;
18609 			}
18610 		}
18611 		if (len <= segsiz) {
18612 			/*
18613 			 * In case there are too many small fragments don't
18614 			 * use TSO:
18615 			 */
18616 			tso = 0;
18617 		}
18618 	} else {
18619 		tso = 0;
18620 	}
18621 	if ((tso == 0) && (len > segsiz))
18622 		len = segsiz;
18623 	(void)tcp_get_usecs(tv);
18624 	if ((len == 0) ||
18625 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
18626 		goto failed;
18627 	}
18628 	th->th_seq = htonl(rsm->r_start);
18629 	th->th_ack = htonl(tp->rcv_nxt);
18630 	/*
18631 	 * The PUSH bit should only be applied
18632 	 * if the full retransmission is made. If
18633 	 * we are sending less than this is the
18634 	 * left hand edge and should not have
18635 	 * the PUSH bit.
18636 	 */
18637 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
18638 	    (len == (rsm->r_end - rsm->r_start)))
18639 		flags |= TH_PUSH;
18640 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18641 	if (th->th_win == 0) {
18642 		tp->t_sndzerowin++;
18643 		tp->t_flags |= TF_RXWIN0SENT;
18644 	} else
18645 		tp->t_flags &= ~TF_RXWIN0SENT;
18646 	if (rsm->r_flags & RACK_TLP) {
18647 		/*
18648 		 * TLP should not count in retran count, but
18649 		 * in its own bin
18650 		 */
18651 		counter_u64_add(rack_tlp_retran, 1);
18652 		counter_u64_add(rack_tlp_retran_bytes, len);
18653 	} else {
18654 		tp->t_sndrexmitpack++;
18655 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18656 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18657 	}
18658 #ifdef STATS
18659 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18660 				 len);
18661 #endif
18662 	if (rsm->m == NULL)
18663 		goto failed;
18664 	if (rsm->m &&
18665 	    ((rsm->orig_m_len != rsm->m->m_len) ||
18666 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18667 		/* Fix up the orig_m_len and possibly the mbuf offset */
18668 		rack_adjust_orig_mlen(rsm);
18669 	}
18670 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18671 	if (len <= segsiz) {
18672 		/*
18673 		 * Must have ran out of mbufs for the copy
18674 		 * shorten it to no longer need tso. Lets
18675 		 * not put on sendalot since we are low on
18676 		 * mbufs.
18677 		 */
18678 		tso = 0;
18679 	}
18680 	if ((m->m_next == NULL) || (len <= 0)){
18681 		goto failed;
18682 	}
18683 	if (udp) {
18684 		if (rack->r_is_v6)
18685 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
18686 		else
18687 			ulen = hdrlen + len - sizeof(struct ip);
18688 		udp->uh_ulen = htons(ulen);
18689 	}
18690 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18691 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18692 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18693 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
18694 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18695 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18696 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18697 #ifdef INET6
18698 		if (rack->r_is_v6) {
18699 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18700 		    ip6->ip6_flow |= htonl(ect << 20);
18701 		}
18702 		else
18703 #endif
18704 		{
18705 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
18706 		    ip->ip_tos |= ect;
18707 		}
18708 	}
18709 	if (rack->r_ctl.crte != NULL) {
18710 		/* See if we can send via the hw queue */
18711 		slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18712 		/* If there is nothing in queue (no pacing time) we can send via the hw queue */
18713 		if (slot == 0)
18714 			ip_sendflag = 0;
18715 	}
18716 	tcp_set_flags(th, flags);
18717 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18718 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18719 	if (to.to_flags & TOF_SIGNATURE) {
18720 		/*
18721 		 * Calculate MD5 signature and put it into the place
18722 		 * determined before.
18723 		 * NOTE: since TCP options buffer doesn't point into
18724 		 * mbuf's data, calculate offset and use it.
18725 		 */
18726 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18727 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18728 			/*
18729 			 * Do not send segment if the calculation of MD5
18730 			 * digest has failed.
18731 			 */
18732 			goto failed;
18733 		}
18734 	}
18735 #endif
18736 #ifdef INET6
18737 	if (rack->r_is_v6) {
18738 		if (tp->t_port) {
18739 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18740 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18741 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18742 			th->th_sum = htons(0);
18743 			UDPSTAT_INC(udps_opackets);
18744 		} else {
18745 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18746 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18747 			th->th_sum = in6_cksum_pseudo(ip6,
18748 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18749 						      0);
18750 		}
18751 	}
18752 #endif
18753 #if defined(INET6) && defined(INET)
18754 	else
18755 #endif
18756 #ifdef INET
18757 	{
18758 		if (tp->t_port) {
18759 			m->m_pkthdr.csum_flags = CSUM_UDP;
18760 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18761 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18762 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18763 			th->th_sum = htons(0);
18764 			UDPSTAT_INC(udps_opackets);
18765 		} else {
18766 			m->m_pkthdr.csum_flags = CSUM_TCP;
18767 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18768 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18769 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18770 									IPPROTO_TCP + len + optlen));
18771 		}
18772 		/* IP version must be set here for ipv4/ipv6 checking later */
18773 		KASSERT(ip->ip_v == IPVERSION,
18774 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18775 	}
18776 #endif
18777 	if (tso) {
18778 		/*
18779 		 * Here we use segsiz since we have no added options besides
18780 		 * any standard timestamp options (no DSACKs or SACKS are sent
18781 		 * via either fast-path).
18782 		 */
18783 		KASSERT(len > segsiz,
18784 			("%s: len <= tso_segsz tp:%p", __func__, tp));
18785 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18786 		m->m_pkthdr.tso_segsz = segsiz;
18787 	}
18788 #ifdef INET6
18789 	if (rack->r_is_v6) {
18790 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18791 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18792 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18793 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18794 		else
18795 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18796 	}
18797 #endif
18798 #if defined(INET) && defined(INET6)
18799 	else
18800 #endif
18801 #ifdef INET
18802 	{
18803 		ip->ip_len = htons(m->m_pkthdr.len);
18804 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
18805 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18806 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18807 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18808 				ip->ip_off |= htons(IP_DF);
18809 			}
18810 		} else {
18811 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18812 		}
18813 	}
18814 #endif
18815 	if (doing_tlp == 0) {
18816 		/* Set we retransmitted */
18817 		rack->rc_gp_saw_rec = 1;
18818 	} else {
18819 		/* Its a TLP set ca or ss */
18820 		if (tp->snd_cwnd > tp->snd_ssthresh) {
18821 			/* Set we sent in CA */
18822 			rack->rc_gp_saw_ca = 1;
18823 		} else {
18824 			/* Set we sent in SS */
18825 			rack->rc_gp_saw_ss = 1;
18826 		}
18827 	}
18828 	/* Time to copy in our header */
18829 	cpto = mtod(m, uint8_t *);
18830 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18831 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18832 	if (optlen) {
18833 		bcopy(opt, th + 1, optlen);
18834 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18835 	} else {
18836 		th->th_off = sizeof(struct tcphdr) >> 2;
18837 	}
18838 	if (tcp_bblogging_on(rack->rc_tp)) {
18839 		union tcp_log_stackspecific log;
18840 
18841 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18842 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18843 			counter_u64_add(rack_collapsed_win_rxt, 1);
18844 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18845 		}
18846 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18847 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18848 		if (rack->rack_no_prr)
18849 			log.u_bbr.flex1 = 0;
18850 		else
18851 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18852 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18853 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18854 		log.u_bbr.flex4 = max_val;
18855 		/* Save off the early/late values */
18856 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18857 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18858 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18859 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
18860 		if (doing_tlp == 0)
18861 			log.u_bbr.flex8 = 1;
18862 		else
18863 			log.u_bbr.flex8 = 2;
18864 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18865 		log.u_bbr.flex7 = 55;
18866 		log.u_bbr.pkts_out = tp->t_maxseg;
18867 		log.u_bbr.timeStamp = cts;
18868 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18869 		if (rsm->r_rtr_cnt > 0) {
18870 			/*
18871 			 * When we have a retransmit we want to log the
18872 			 * burst at send and flight at send from before.
18873 			 */
18874 			log.u_bbr.flex5 = rsm->r_fas;
18875 			log.u_bbr.bbr_substate = rsm->r_bas;
18876 		} else {
18877 			/*
18878 			 * This is currently unlikely until we do the
18879 			 * packet pair probes but I will add it for completeness.
18880 			 */
18881 			log.u_bbr.flex5 = log.u_bbr.inflight;
18882 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
18883 		}
18884 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
18885 		log.u_bbr.delivered = 0;
18886 		log.u_bbr.rttProp = (uintptr_t)rsm;
18887 		log.u_bbr.delRate = rsm->r_flags;
18888 		log.u_bbr.delRate <<= 31;
18889 		log.u_bbr.delRate |= rack->r_must_retran;
18890 		log.u_bbr.delRate <<= 1;
18891 		log.u_bbr.delRate |= 1;
18892 		log.u_bbr.pkt_epoch = __LINE__;
18893 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
18894 				     len, &log, false, NULL, __func__, __LINE__, tv);
18895 	} else
18896 		lgb = NULL;
18897 	if ((rack->r_ctl.crte != NULL) &&
18898 	    tcp_bblogging_on(tp)) {
18899 		rack_log_queue_level(tp, rack, len, tv, cts);
18900 	}
18901 #ifdef INET6
18902 	if (rack->r_is_v6) {
18903 		error = ip6_output(m, inp->in6p_outputopts,
18904 				   &inp->inp_route6,
18905 				   ip_sendflag, NULL, NULL, inp);
18906 	}
18907 	else
18908 #endif
18909 #ifdef INET
18910 	{
18911 		error = ip_output(m, NULL,
18912 				  &inp->inp_route,
18913 				  ip_sendflag, 0, inp);
18914 	}
18915 #endif
18916 	m = NULL;
18917 	if (lgb) {
18918 		lgb->tlb_errno = error;
18919 		lgb = NULL;
18920 	}
18921 	/* Move snd_nxt to snd_max so we don't have false retransmissions */
18922 	tp->snd_nxt = tp->snd_max;
18923 	if (error) {
18924 		goto failed;
18925 	} else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
18926 		rack->rc_hw_nobuf = 0;
18927 		rack->r_ctl.rc_agg_delayed = 0;
18928 		rack->r_early = 0;
18929 		rack->r_late = 0;
18930 		rack->r_ctl.rc_agg_early = 0;
18931 	}
18932 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
18933 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
18934 	if (doing_tlp) {
18935 		rack->rc_tlp_in_progress = 1;
18936 		rack->r_ctl.rc_tlp_cnt_out++;
18937 	}
18938 	if (error == 0) {
18939 		counter_u64_add(rack_total_bytes, len);
18940 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
18941 		if (doing_tlp) {
18942 			rack->rc_last_sent_tlp_past_cumack = 0;
18943 			rack->rc_last_sent_tlp_seq_valid = 1;
18944 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18945 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18946 		}
18947 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18948 			rack->r_ctl.rc_prr_sndcnt -= len;
18949 		else
18950 			rack->r_ctl.rc_prr_sndcnt = 0;
18951 	}
18952 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18953 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
18954 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18955 		rack->r_ctl.retran_during_recovery += len;
18956 	{
18957 		int idx;
18958 
18959 		idx = (len / segsiz) + 3;
18960 		if (idx >= TCP_MSS_ACCT_ATIMER)
18961 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18962 		else
18963 			counter_u64_add(rack_out_size[idx], 1);
18964 	}
18965 	if (tp->t_rtttime == 0) {
18966 		tp->t_rtttime = ticks;
18967 		tp->t_rtseq = startseq;
18968 		KMOD_TCPSTAT_INC(tcps_segstimed);
18969 	}
18970 	counter_u64_add(rack_fto_rsm_send, 1);
18971 	if (error && (error == ENOBUFS)) {
18972 		if (rack->r_ctl.crte != NULL) {
18973 			tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
18974 			if (tcp_bblogging_on(rack->rc_tp))
18975 				rack_log_queue_level(tp, rack, len, tv, cts);
18976 		} else
18977 			tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
18978 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18979 		if (rack->rc_enobuf < 0x7f)
18980 			rack->rc_enobuf++;
18981 		if (slot < (10 * HPTS_USEC_IN_MSEC))
18982 			slot = 10 * HPTS_USEC_IN_MSEC;
18983 		if (rack->r_ctl.crte != NULL) {
18984 			counter_u64_add(rack_saw_enobuf_hw, 1);
18985 			tcp_rl_log_enobuf(rack->r_ctl.crte);
18986 		}
18987 		counter_u64_add(rack_saw_enobuf, 1);
18988 	} else {
18989 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz, __LINE__);
18990 	}
18991 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
18992 #ifdef TCP_ACCOUNTING
18993 	crtsc = get_cyclecount();
18994 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18995 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
18996 	}
18997 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18998 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
18999 	}
19000 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19001 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
19002 	}
19003 	sched_unpin();
19004 #endif
19005 	return (0);
19006 failed:
19007 	if (m)
19008 		m_free(m);
19009 	return (-1);
19010 }
19011 
19012 static void
19013 rack_sndbuf_autoscale(struct tcp_rack *rack)
19014 {
19015 	/*
19016 	 * Automatic sizing of send socket buffer.  Often the send buffer
19017 	 * size is not optimally adjusted to the actual network conditions
19018 	 * at hand (delay bandwidth product).  Setting the buffer size too
19019 	 * small limits throughput on links with high bandwidth and high
19020 	 * delay (eg. trans-continental/oceanic links).  Setting the
19021 	 * buffer size too big consumes too much real kernel memory,
19022 	 * especially with many connections on busy servers.
19023 	 *
19024 	 * The criteria to step up the send buffer one notch are:
19025 	 *  1. receive window of remote host is larger than send buffer
19026 	 *     (with a fudge factor of 5/4th);
19027 	 *  2. send buffer is filled to 7/8th with data (so we actually
19028 	 *     have data to make use of it);
19029 	 *  3. send buffer fill has not hit maximal automatic size;
19030 	 *  4. our send window (slow start and cogestion controlled) is
19031 	 *     larger than sent but unacknowledged data in send buffer.
19032 	 *
19033 	 * Note that the rack version moves things much faster since
19034 	 * we want to avoid hitting cache lines in the rack_fast_output()
19035 	 * path so this is called much less often and thus moves
19036 	 * the SB forward by a percentage.
19037 	 */
19038 	struct socket *so;
19039 	struct tcpcb *tp;
19040 	uint32_t sendwin, scaleup;
19041 
19042 	tp = rack->rc_tp;
19043 	so = rack->rc_inp->inp_socket;
19044 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
19045 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
19046 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
19047 		    sbused(&so->so_snd) >=
19048 		    (so->so_snd.sb_hiwat / 8 * 7) &&
19049 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
19050 		    sendwin >= (sbused(&so->so_snd) -
19051 		    (tp->snd_max - tp->snd_una))) {
19052 			if (rack_autosndbuf_inc)
19053 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
19054 			else
19055 				scaleup = V_tcp_autosndbuf_inc;
19056 			if (scaleup < V_tcp_autosndbuf_inc)
19057 				scaleup = V_tcp_autosndbuf_inc;
19058 			scaleup += so->so_snd.sb_hiwat;
19059 			if (scaleup > V_tcp_autosndbuf_max)
19060 				scaleup = V_tcp_autosndbuf_max;
19061 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
19062 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
19063 		}
19064 	}
19065 }
19066 
19067 static int
19068 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
19069 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
19070 {
19071 	/*
19072 	 * Enter to do fast output. We are given that the sched_pin is
19073 	 * in place (if accounting is compiled in) and the cycle count taken
19074 	 * at entry is in place in ts_val. The idea here is that
19075 	 * we know how many more bytes needs to be sent (presumably either
19076 	 * during pacing or to fill the cwnd and that was greater than
19077 	 * the max-burst). We have how much to send and all the info we
19078 	 * need to just send.
19079 	 */
19080 #ifdef INET
19081 	struct ip *ip = NULL;
19082 #endif
19083 	struct udphdr *udp = NULL;
19084 	struct tcphdr *th = NULL;
19085 	struct mbuf *m, *s_mb;
19086 	struct inpcb *inp;
19087 	uint8_t *cpto;
19088 	struct tcp_log_buffer *lgb;
19089 #ifdef TCP_ACCOUNTING
19090 	uint64_t crtsc;
19091 #endif
19092 	struct tcpopt to;
19093 	u_char opt[TCP_MAXOLEN];
19094 	uint32_t hdrlen, optlen;
19095 #ifdef TCP_ACCOUNTING
19096 	int cnt_thru = 1;
19097 #endif
19098 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19099 	uint16_t flags;
19100 	uint32_t s_soff;
19101 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
19102 	uint32_t if_hw_tsomaxsegsize;
19103 	uint32_t add_flag = RACK_SENT_FP;
19104 #ifdef INET6
19105 	struct ip6_hdr *ip6 = NULL;
19106 
19107 	if (rack->r_is_v6) {
19108 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19109 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19110 	} else
19111 #endif				/* INET6 */
19112 	{
19113 #ifdef INET
19114 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19115 		hdrlen = sizeof(struct tcpiphdr);
19116 #endif
19117 	}
19118 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19119 		m = NULL;
19120 		goto failed;
19121 	}
19122 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19123 	startseq = tp->snd_max;
19124 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19125 	inp = rack->rc_inp;
19126 	len = rack->r_ctl.fsb.left_to_send;
19127 	to.to_flags = 0;
19128 	flags = rack->r_ctl.fsb.tcp_flags;
19129 	if (tp->t_flags & TF_RCVD_TSTMP) {
19130 		to.to_tsval = ms_cts + tp->ts_offset;
19131 		to.to_tsecr = tp->ts_recent;
19132 		to.to_flags = TOF_TS;
19133 	}
19134 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19135 	/* TCP-MD5 (RFC2385). */
19136 	if (tp->t_flags & TF_SIGNATURE)
19137 		to.to_flags |= TOF_SIGNATURE;
19138 #endif
19139 	optlen = tcp_addoptions(&to, opt);
19140 	hdrlen += optlen;
19141 	udp = rack->r_ctl.fsb.udp;
19142 	if (udp)
19143 		hdrlen += sizeof(struct udphdr);
19144 	if (rack->r_ctl.rc_pace_max_segs)
19145 		max_val = rack->r_ctl.rc_pace_max_segs;
19146 	else if (rack->rc_user_set_max_segs)
19147 		max_val = rack->rc_user_set_max_segs * segsiz;
19148 	else
19149 		max_val = len;
19150 	if ((tp->t_flags & TF_TSO) &&
19151 	    V_tcp_do_tso &&
19152 	    (len > segsiz) &&
19153 	    (tp->t_port == 0))
19154 		tso = 1;
19155 again:
19156 #ifdef INET6
19157 	if (MHLEN < hdrlen + max_linkhdr)
19158 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19159 	else
19160 #endif
19161 		m = m_gethdr(M_NOWAIT, MT_DATA);
19162 	if (m == NULL)
19163 		goto failed;
19164 	m->m_data += max_linkhdr;
19165 	m->m_len = hdrlen;
19166 	th = rack->r_ctl.fsb.th;
19167 	/* Establish the len to send */
19168 	if (len > max_val)
19169 		len = max_val;
19170 	if ((tso) && (len + optlen > segsiz)) {
19171 		uint32_t if_hw_tsomax;
19172 		int32_t max_len;
19173 
19174 		/* extract TSO information */
19175 		if_hw_tsomax = tp->t_tsomax;
19176 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19177 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19178 		/*
19179 		 * Check if we should limit by maximum payload
19180 		 * length:
19181 		 */
19182 		if (if_hw_tsomax != 0) {
19183 			/* compute maximum TSO length */
19184 			max_len = (if_hw_tsomax - hdrlen -
19185 				   max_linkhdr);
19186 			if (max_len <= 0) {
19187 				goto failed;
19188 			} else if (len > max_len) {
19189 				len = max_len;
19190 			}
19191 		}
19192 		if (len <= segsiz) {
19193 			/*
19194 			 * In case there are too many small fragments don't
19195 			 * use TSO:
19196 			 */
19197 			tso = 0;
19198 		}
19199 	} else {
19200 		tso = 0;
19201 	}
19202 	if ((tso == 0) && (len > segsiz))
19203 		len = segsiz;
19204 	(void)tcp_get_usecs(tv);
19205 	if ((len == 0) ||
19206 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
19207 		goto failed;
19208 	}
19209 	sb_offset = tp->snd_max - tp->snd_una;
19210 	th->th_seq = htonl(tp->snd_max);
19211 	th->th_ack = htonl(tp->rcv_nxt);
19212 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19213 	if (th->th_win == 0) {
19214 		tp->t_sndzerowin++;
19215 		tp->t_flags |= TF_RXWIN0SENT;
19216 	} else
19217 		tp->t_flags &= ~TF_RXWIN0SENT;
19218 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
19219 	KMOD_TCPSTAT_INC(tcps_sndpack);
19220 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19221 #ifdef STATS
19222 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19223 				 len);
19224 #endif
19225 	if (rack->r_ctl.fsb.m == NULL)
19226 		goto failed;
19227 
19228 	/* s_mb and s_soff are saved for rack_log_output */
19229 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19230 				    &s_mb, &s_soff);
19231 	if (len <= segsiz) {
19232 		/*
19233 		 * Must have ran out of mbufs for the copy
19234 		 * shorten it to no longer need tso. Lets
19235 		 * not put on sendalot since we are low on
19236 		 * mbufs.
19237 		 */
19238 		tso = 0;
19239 	}
19240 	if (rack->r_ctl.fsb.rfo_apply_push &&
19241 	    (len == rack->r_ctl.fsb.left_to_send)) {
19242 		tcp_set_flags(th, flags | TH_PUSH);
19243 		add_flag |= RACK_HAD_PUSH;
19244 	}
19245 	if ((m->m_next == NULL) || (len <= 0)){
19246 		goto failed;
19247 	}
19248 	if (udp) {
19249 		if (rack->r_is_v6)
19250 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
19251 		else
19252 			ulen = hdrlen + len - sizeof(struct ip);
19253 		udp->uh_ulen = htons(ulen);
19254 	}
19255 	m->m_pkthdr.rcvif = (struct ifnet *)0;
19256 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
19257 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19258 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
19259 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19260 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
19261 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19262 #ifdef INET6
19263 		if (rack->r_is_v6) {
19264 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19265 			ip6->ip6_flow |= htonl(ect << 20);
19266 		}
19267 		else
19268 #endif
19269 		{
19270 #ifdef INET
19271 			ip->ip_tos &= ~IPTOS_ECN_MASK;
19272 			ip->ip_tos |= ect;
19273 #endif
19274 		}
19275 	}
19276 	tcp_set_flags(th, flags);
19277 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
19278 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19279 	if (to.to_flags & TOF_SIGNATURE) {
19280 		/*
19281 		 * Calculate MD5 signature and put it into the place
19282 		 * determined before.
19283 		 * NOTE: since TCP options buffer doesn't point into
19284 		 * mbuf's data, calculate offset and use it.
19285 		 */
19286 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
19287 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
19288 			/*
19289 			 * Do not send segment if the calculation of MD5
19290 			 * digest has failed.
19291 			 */
19292 			goto failed;
19293 		}
19294 	}
19295 #endif
19296 #ifdef INET6
19297 	if (rack->r_is_v6) {
19298 		if (tp->t_port) {
19299 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19300 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19301 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19302 			th->th_sum = htons(0);
19303 			UDPSTAT_INC(udps_opackets);
19304 		} else {
19305 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19306 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19307 			th->th_sum = in6_cksum_pseudo(ip6,
19308 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19309 						      0);
19310 		}
19311 	}
19312 #endif
19313 #if defined(INET6) && defined(INET)
19314 	else
19315 #endif
19316 #ifdef INET
19317 	{
19318 		if (tp->t_port) {
19319 			m->m_pkthdr.csum_flags = CSUM_UDP;
19320 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19321 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19322 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19323 			th->th_sum = htons(0);
19324 			UDPSTAT_INC(udps_opackets);
19325 		} else {
19326 			m->m_pkthdr.csum_flags = CSUM_TCP;
19327 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19328 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
19329 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19330 									IPPROTO_TCP + len + optlen));
19331 		}
19332 		/* IP version must be set here for ipv4/ipv6 checking later */
19333 		KASSERT(ip->ip_v == IPVERSION,
19334 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
19335 	}
19336 #endif
19337 	if (tso) {
19338 		/*
19339 		 * Here we use segsiz since we have no added options besides
19340 		 * any standard timestamp options (no DSACKs or SACKS are sent
19341 		 * via either fast-path).
19342 		 */
19343 		KASSERT(len > segsiz,
19344 			("%s: len <= tso_segsz tp:%p", __func__, tp));
19345 		m->m_pkthdr.csum_flags |= CSUM_TSO;
19346 		m->m_pkthdr.tso_segsz = segsiz;
19347 	}
19348 #ifdef INET6
19349 	if (rack->r_is_v6) {
19350 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19351 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19352 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19353 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19354 		else
19355 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19356 	}
19357 #endif
19358 #if defined(INET) && defined(INET6)
19359 	else
19360 #endif
19361 #ifdef INET
19362 	{
19363 		ip->ip_len = htons(m->m_pkthdr.len);
19364 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19365 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19366 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19367 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19368 				ip->ip_off |= htons(IP_DF);
19369 			}
19370 		} else {
19371 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19372 		}
19373 	}
19374 #endif
19375 	if (tp->snd_cwnd > tp->snd_ssthresh) {
19376 		/* Set we sent in CA */
19377 		rack->rc_gp_saw_ca = 1;
19378 	} else {
19379 		/* Set we sent in SS */
19380 		rack->rc_gp_saw_ss = 1;
19381 	}
19382 	/* Time to copy in our header */
19383 	cpto = mtod(m, uint8_t *);
19384 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19385 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19386 	if (optlen) {
19387 		bcopy(opt, th + 1, optlen);
19388 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19389 	} else {
19390 		th->th_off = sizeof(struct tcphdr) >> 2;
19391 	}
19392 	if ((rack->r_ctl.crte != NULL) &&
19393 	    tcp_bblogging_on(tp)) {
19394 		rack_log_queue_level(tp, rack, len, tv, cts);
19395 	}
19396 	if (tcp_bblogging_on(rack->rc_tp)) {
19397 		union tcp_log_stackspecific log;
19398 
19399 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19400 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19401 		if (rack->rack_no_prr)
19402 			log.u_bbr.flex1 = 0;
19403 		else
19404 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19405 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19406 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19407 		log.u_bbr.flex4 = max_val;
19408 		/* Save off the early/late values */
19409 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19410 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19411 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19412 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19413 		log.u_bbr.flex8 = 0;
19414 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19415 		log.u_bbr.flex7 = 44;
19416 		log.u_bbr.pkts_out = tp->t_maxseg;
19417 		log.u_bbr.timeStamp = cts;
19418 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19419 		log.u_bbr.flex5 = log.u_bbr.inflight;
19420 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19421 		log.u_bbr.delivered = 0;
19422 		log.u_bbr.rttProp = 0;
19423 		log.u_bbr.delRate = rack->r_must_retran;
19424 		log.u_bbr.delRate <<= 1;
19425 		log.u_bbr.pkt_epoch = __LINE__;
19426 		/* For fast output no retrans so just inflight and how many mss we send */
19427 		log.u_bbr.flex5 = log.u_bbr.inflight;
19428 		log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19429 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19430 				     len, &log, false, NULL, __func__, __LINE__, tv);
19431 	} else
19432 		lgb = NULL;
19433 #ifdef INET6
19434 	if (rack->r_is_v6) {
19435 		error = ip6_output(m, inp->in6p_outputopts,
19436 				   &inp->inp_route6,
19437 				   0, NULL, NULL, inp);
19438 	}
19439 #endif
19440 #if defined(INET) && defined(INET6)
19441 	else
19442 #endif
19443 #ifdef INET
19444 	{
19445 		error = ip_output(m, NULL,
19446 				  &inp->inp_route,
19447 				  0, 0, inp);
19448 	}
19449 #endif
19450 	if (lgb) {
19451 		lgb->tlb_errno = error;
19452 		lgb = NULL;
19453 	}
19454 	if (error) {
19455 		*send_err = error;
19456 		m = NULL;
19457 		goto failed;
19458 	} else if (rack->rc_hw_nobuf) {
19459 		rack->rc_hw_nobuf = 0;
19460 		rack->r_ctl.rc_agg_delayed = 0;
19461 		rack->r_early = 0;
19462 		rack->r_late = 0;
19463 		rack->r_ctl.rc_agg_early = 0;
19464 	}
19465 	if ((error == 0) && (rack->lt_bw_up == 0)) {
19466 		/* Unlikely */
19467 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
19468 		rack->r_ctl.lt_seq = tp->snd_una;
19469 		rack->lt_bw_up = 1;
19470 	} else if ((error == 0) &&
19471 		   (((tp->snd_max + len) - rack->r_ctl.lt_seq) > 0x7fffffff)) {
19472 		/*
19473 		 * Need to record what we have since we are
19474 		 * approaching seq wrap.
19475 		 */
19476 		struct timeval tv;
19477 		uint64_t tmark;
19478 
19479 		rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
19480 		rack->r_ctl.lt_seq = tp->snd_una;
19481 		tmark = tcp_get_u64_usecs(&tv);
19482 		if (tmark > rack->r_ctl.lt_timemark) {
19483 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
19484 			rack->r_ctl.lt_timemark = tmark;
19485 		}
19486 	}
19487 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19488 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19489 	if (tp->snd_una == tp->snd_max) {
19490 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
19491 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19492 		tp->t_acktime = ticks;
19493 	}
19494 	counter_u64_add(rack_total_bytes, len);
19495 	tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19496 
19497 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19498 	tot_len += len;
19499 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
19500 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19501 	tp->snd_max += len;
19502 	tp->snd_nxt = tp->snd_max;
19503 	if (rack->rc_new_rnd_needed) {
19504 		rack_new_round_starts(tp, rack, tp->snd_max);
19505 	}
19506 	{
19507 		int idx;
19508 
19509 		idx = (len / segsiz) + 3;
19510 		if (idx >= TCP_MSS_ACCT_ATIMER)
19511 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19512 		else
19513 			counter_u64_add(rack_out_size[idx], 1);
19514 	}
19515 	if (len <= rack->r_ctl.fsb.left_to_send)
19516 		rack->r_ctl.fsb.left_to_send -= len;
19517 	else
19518 		rack->r_ctl.fsb.left_to_send = 0;
19519 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
19520 		rack->r_fast_output = 0;
19521 		rack->r_ctl.fsb.left_to_send = 0;
19522 		/* At the end of fast_output scale up the sb */
19523 		SOCK_SENDBUF_LOCK(rack->rc_inp->inp_socket);
19524 		rack_sndbuf_autoscale(rack);
19525 		SOCK_SENDBUF_UNLOCK(rack->rc_inp->inp_socket);
19526 	}
19527 	if (tp->t_rtttime == 0) {
19528 		tp->t_rtttime = ticks;
19529 		tp->t_rtseq = startseq;
19530 		KMOD_TCPSTAT_INC(tcps_segstimed);
19531 	}
19532 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19533 	    (max_val > len) &&
19534 	    (tso == 0)) {
19535 		max_val -= len;
19536 		len = segsiz;
19537 		th = rack->r_ctl.fsb.th;
19538 #ifdef TCP_ACCOUNTING
19539 		cnt_thru++;
19540 #endif
19541 		goto again;
19542 	}
19543 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19544 	counter_u64_add(rack_fto_send, 1);
19545 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz, __LINE__);
19546 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
19547 #ifdef TCP_ACCOUNTING
19548 	crtsc = get_cyclecount();
19549 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19550 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19551 	}
19552 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19553 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19554 	}
19555 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19556 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
19557 	}
19558 	sched_unpin();
19559 #endif
19560 	return (0);
19561 failed:
19562 	if (m)
19563 		m_free(m);
19564 	rack->r_fast_output = 0;
19565 	return (-1);
19566 }
19567 
19568 static inline void
19569 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19570 		       struct sockbuf *sb,
19571 		       int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19572 		       bool hw_tls,
19573 		       uint16_t flags)
19574 {
19575 	rack->r_fast_output = 1;
19576 	rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19577 	rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19578 	rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19579 	rack->r_ctl.fsb.tcp_flags = flags;
19580 	rack->r_ctl.fsb.left_to_send = orig_len - len;
19581 	if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19582 		/* Less than a full sized pace, lets not  */
19583 		rack->r_fast_output = 0;
19584 		return;
19585 	} else {
19586 		/* Round down to the nearest pace_max_seg */
19587 		rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19588 	}
19589 	if (hw_tls)
19590 		rack->r_ctl.fsb.hw_tls = 1;
19591 	else
19592 		rack->r_ctl.fsb.hw_tls = 0;
19593 	KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19594 		("rack:%p left_to_send:%u sbavail:%u out:%u",
19595 		 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19596 		 (tp->snd_max - tp->snd_una)));
19597 	if (rack->r_ctl.fsb.left_to_send < segsiz)
19598 		rack->r_fast_output = 0;
19599 	else {
19600 		if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19601 			rack->r_ctl.fsb.rfo_apply_push = 1;
19602 		else
19603 			rack->r_ctl.fsb.rfo_apply_push = 0;
19604 	}
19605 }
19606 
19607 static uint32_t
19608 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19609 {
19610 	uint64_t min_time;
19611 	uint32_t maxlen;
19612 
19613 	min_time = (uint64_t)get_hpts_min_sleep_time();
19614 	maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19615 	maxlen = roundup(maxlen, segsiz);
19616 	return (maxlen);
19617 }
19618 
19619 static struct rack_sendmap *
19620 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19621 {
19622 	struct rack_sendmap *rsm = NULL;
19623 	int thresh;
19624 
19625 restart:
19626 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19627 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19628 		/* Nothing, strange turn off validity  */
19629 		rack->r_collapse_point_valid = 0;
19630 		return (NULL);
19631 	}
19632 	/* Can we send it yet? */
19633 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19634 		/*
19635 		 * Receiver window has not grown enough for
19636 		 * the segment to be put on the wire.
19637 		 */
19638 		return (NULL);
19639 	}
19640 	if (rsm->r_flags & RACK_ACKED) {
19641 		/*
19642 		 * It has been sacked, lets move to the
19643 		 * next one if possible.
19644 		 */
19645 		rack->r_ctl.last_collapse_point = rsm->r_end;
19646 		/* Are we done? */
19647 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19648 			    rack->r_ctl.high_collapse_point)) {
19649 			rack->r_collapse_point_valid = 0;
19650 			return (NULL);
19651 		}
19652 		goto restart;
19653 	}
19654 	/* Now has it been long enough ? */
19655 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts, __LINE__, 1);
19656 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19657 		rack_log_collapse(rack, rsm->r_start,
19658 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19659 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
19660 		return (rsm);
19661 	}
19662 	/* Not enough time */
19663 	rack_log_collapse(rack, rsm->r_start,
19664 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19665 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
19666 	return (NULL);
19667 }
19668 
19669 static inline void
19670 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19671 {
19672 	if ((rack->full_size_rxt == 0) &&
19673 	    (rack->shape_rxt_to_pacing_min == 0) &&
19674 	    (*len >= segsiz)) {
19675 		*len = segsiz;
19676 	} else if (rack->shape_rxt_to_pacing_min &&
19677 		 rack->gp_ready) {
19678 		/* We use pacing min as shaping len req */
19679 		uint32_t maxlen;
19680 
19681 		maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19682 		if (*len > maxlen)
19683 			*len = maxlen;
19684 	} else {
19685 		/*
19686 		 * The else is full_size_rxt is on so send it all
19687 		 * note we do need to check this for exceeding
19688 		 * our max segment size due to the fact that
19689 		 * we do sometimes merge chunks together i.e.
19690 		 * we cannot just assume that we will never have
19691 		 * a chunk greater than pace_max_seg
19692 		 */
19693 		if (*len > pace_max_seg)
19694 			*len = pace_max_seg;
19695 	}
19696 }
19697 
19698 static int
19699 rack_output(struct tcpcb *tp)
19700 {
19701 	struct socket *so;
19702 	uint32_t recwin;
19703 	uint32_t sb_offset, s_moff = 0;
19704 	int32_t len, error = 0;
19705 	uint16_t flags;
19706 	struct mbuf *m, *s_mb = NULL;
19707 	struct mbuf *mb;
19708 	uint32_t if_hw_tsomaxsegcount = 0;
19709 	uint32_t if_hw_tsomaxsegsize;
19710 	int32_t segsiz, minseg;
19711 	long tot_len_this_send = 0;
19712 #ifdef INET
19713 	struct ip *ip = NULL;
19714 #endif
19715 	struct udphdr *udp = NULL;
19716 	struct tcp_rack *rack;
19717 	struct tcphdr *th;
19718 	uint8_t pass = 0;
19719 	uint8_t mark = 0;
19720 	uint8_t check_done = 0;
19721 	uint8_t wanted_cookie = 0;
19722 	u_char opt[TCP_MAXOLEN];
19723 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
19724 	uint32_t rack_seq;
19725 
19726 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19727 	unsigned ipsec_optlen = 0;
19728 
19729 #endif
19730 	int32_t idle, sendalot;
19731 	uint32_t tot_idle;
19732 	int32_t sub_from_prr = 0;
19733 	volatile int32_t sack_rxmit;
19734 	struct rack_sendmap *rsm = NULL;
19735 	int32_t tso, mtu;
19736 	struct tcpopt to;
19737 	int32_t slot = 0;
19738 	int32_t sup_rack = 0;
19739 	uint32_t cts, ms_cts, delayed, early;
19740 	uint32_t add_flag = RACK_SENT_SP;
19741 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19742 	uint8_t doing_tlp = 0;
19743 	uint32_t cwnd_to_use, pace_max_seg;
19744 	int32_t do_a_prefetch = 0;
19745 	int32_t prefetch_rsm = 0;
19746 	int32_t orig_len = 0;
19747 	struct timeval tv;
19748 	int32_t prefetch_so_done = 0;
19749 	struct tcp_log_buffer *lgb;
19750 	struct inpcb *inp = tptoinpcb(tp);
19751 	struct sockbuf *sb;
19752 	uint64_t ts_val = 0;
19753 #ifdef TCP_ACCOUNTING
19754 	uint64_t crtsc;
19755 #endif
19756 #ifdef INET6
19757 	struct ip6_hdr *ip6 = NULL;
19758 	int32_t isipv6;
19759 #endif
19760 	bool hpts_calling, hw_tls = false;
19761 
19762 	NET_EPOCH_ASSERT();
19763 	INP_WLOCK_ASSERT(inp);
19764 
19765 	/* setup and take the cache hits here */
19766 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19767 #ifdef TCP_ACCOUNTING
19768 	sched_pin();
19769 	ts_val = get_cyclecount();
19770 #endif
19771 	hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
19772 	tp->t_flags2 &= ~TF2_HPTS_CALLS;
19773 #ifdef TCP_OFFLOAD
19774 	if (tp->t_flags & TF_TOE) {
19775 #ifdef TCP_ACCOUNTING
19776 		sched_unpin();
19777 #endif
19778 		return (tcp_offload_output(tp));
19779 	}
19780 #endif
19781 	if (rack->rack_deferred_inited == 0) {
19782 		/*
19783 		 * If we are the connecting socket we will
19784 		 * hit rack_init() when no sequence numbers
19785 		 * are setup. This makes it so we must defer
19786 		 * some initialization. Call that now.
19787 		 */
19788 		rack_deferred_init(tp, rack);
19789 	}
19790 	/*
19791 	 * For TFO connections in SYN_RECEIVED, only allow the initial
19792 	 * SYN|ACK and those sent by the retransmit timer.
19793 	 */
19794 	if ((tp->t_flags & TF_FASTOPEN) &&
19795 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
19796 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
19797 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
19798 #ifdef TCP_ACCOUNTING
19799 		sched_unpin();
19800 #endif
19801 		return (0);
19802 	}
19803 #ifdef INET6
19804 	if (rack->r_state) {
19805 		/* Use the cache line loaded if possible */
19806 		isipv6 = rack->r_is_v6;
19807 	} else {
19808 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
19809 	}
19810 #endif
19811 	early = 0;
19812 	cts = tcp_get_usecs(&tv);
19813 	ms_cts = tcp_tv_to_mssectick(&tv);
19814 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
19815 	    tcp_in_hpts(rack->rc_tp)) {
19816 		/*
19817 		 * We are on the hpts for some timer but not hptsi output.
19818 		 * Remove from the hpts unconditionally.
19819 		 */
19820 		rack_timer_cancel(tp, rack, cts, __LINE__);
19821 	}
19822 	/* Are we pacing and late? */
19823 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19824 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
19825 		/* We are delayed */
19826 		delayed = cts - rack->r_ctl.rc_last_output_to;
19827 	} else {
19828 		delayed = 0;
19829 	}
19830 	/* Do the timers, which may override the pacer */
19831 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
19832 		int retval;
19833 
19834 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
19835 					     &doing_tlp);
19836 		if (retval != 0) {
19837 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
19838 #ifdef TCP_ACCOUNTING
19839 			sched_unpin();
19840 #endif
19841 			/*
19842 			 * If timers want tcp_drop(), then pass error out,
19843 			 * otherwise suppress it.
19844 			 */
19845 			return (retval < 0 ? retval : 0);
19846 		}
19847 	}
19848 	if (rack->rc_in_persist) {
19849 		if (tcp_in_hpts(rack->rc_tp) == 0) {
19850 			/* Timer is not running */
19851 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19852 		}
19853 #ifdef TCP_ACCOUNTING
19854 		sched_unpin();
19855 #endif
19856 		return (0);
19857 	}
19858 	if ((rack->rc_ack_required == 1) &&
19859 	    (rack->r_timer_override == 0)){
19860 		/* A timeout occurred and no ack has arrived */
19861 		if (tcp_in_hpts(rack->rc_tp) == 0) {
19862 			/* Timer is not running */
19863 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19864 		}
19865 #ifdef TCP_ACCOUNTING
19866 		sched_unpin();
19867 #endif
19868 		return (0);
19869 	}
19870 	if ((rack->r_timer_override) ||
19871 	    (rack->rc_ack_can_sendout_data) ||
19872 	    (delayed) ||
19873 	    (tp->t_state < TCPS_ESTABLISHED)) {
19874 		rack->rc_ack_can_sendout_data = 0;
19875 		if (tcp_in_hpts(rack->rc_tp))
19876 			tcp_hpts_remove(rack->rc_tp);
19877 	} else if (tcp_in_hpts(rack->rc_tp)) {
19878 		/*
19879 		 * On the hpts you can't pass even if ACKNOW is on, we will
19880 		 * when the hpts fires.
19881 		 */
19882 #ifdef TCP_ACCOUNTING
19883 		crtsc = get_cyclecount();
19884 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19885 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
19886 		}
19887 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19888 			tp->tcp_cnt_counters[SND_BLOCKED]++;
19889 		}
19890 		sched_unpin();
19891 #endif
19892 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
19893 		return (0);
19894 	}
19895 	/* Finish out both pacing early and late accounting */
19896 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19897 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
19898 		early = rack->r_ctl.rc_last_output_to - cts;
19899 	} else
19900 		early = 0;
19901 	if (delayed && (rack->rc_always_pace == 1)) {
19902 		rack->r_ctl.rc_agg_delayed += delayed;
19903 		rack->r_late = 1;
19904 	} else if (early && (rack->rc_always_pace == 1)) {
19905 		rack->r_ctl.rc_agg_early += early;
19906 		rack->r_early = 1;
19907 	} else if (rack->rc_always_pace == 0) {
19908 		/* Non-paced we are not late */
19909 		rack->r_ctl.rc_agg_delayed = rack->r_ctl.rc_agg_early = 0;
19910 		rack->r_early = rack->r_late = 0;
19911 	}
19912 	/* Now that early/late accounting is done turn off the flag */
19913 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
19914 	rack->r_wanted_output = 0;
19915 	rack->r_timer_override = 0;
19916 	if ((tp->t_state != rack->r_state) &&
19917 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
19918 		rack_set_state(tp, rack);
19919 	}
19920 	if ((rack->r_fast_output) &&
19921 	    (doing_tlp == 0) &&
19922 	    (tp->rcv_numsacks == 0)) {
19923 		int ret;
19924 
19925 		error = 0;
19926 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19927 		if (ret >= 0)
19928 			return(ret);
19929 		else if (error) {
19930 			inp = rack->rc_inp;
19931 			so = inp->inp_socket;
19932 			sb = &so->so_snd;
19933 			goto nomore;
19934 		}
19935 	}
19936 	inp = rack->rc_inp;
19937 	/*
19938 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
19939 	 * only allow the initial SYN or SYN|ACK and those sent
19940 	 * by the retransmit timer.
19941 	 */
19942 	if ((tp->t_flags & TF_FASTOPEN) &&
19943 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
19944 	     (tp->t_state == TCPS_SYN_SENT)) &&
19945 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
19946 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
19947 		rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19948 #ifdef TCP_ACCOUNTING
19949 		sched_unpin();
19950 #endif
19951 		return (0);
19952 	}
19953 	/*
19954 	 * Determine length of data that should be transmitted, and flags
19955 	 * that will be used. If there is some data or critical controls
19956 	 * (SYN, RST) to send, then transmit; otherwise, investigate
19957 	 * further.
19958 	 */
19959 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
19960 	if (tp->t_idle_reduce) {
19961 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
19962 			rack_cc_after_idle(rack, tp);
19963 	}
19964 	tp->t_flags &= ~TF_LASTIDLE;
19965 	if (idle) {
19966 		if (tp->t_flags & TF_MORETOCOME) {
19967 			tp->t_flags |= TF_LASTIDLE;
19968 			idle = 0;
19969 		}
19970 	}
19971 	if ((tp->snd_una == tp->snd_max) &&
19972 	    rack->r_ctl.rc_went_idle_time &&
19973 	    (cts > rack->r_ctl.rc_went_idle_time)) {
19974 		tot_idle = (cts - rack->r_ctl.rc_went_idle_time);
19975 		if (tot_idle > rack_min_probertt_hold) {
19976 			/* Count as a probe rtt */
19977 			if (rack->in_probe_rtt == 0) {
19978 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
19979 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
19980 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
19981 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
19982 			} else {
19983 				rack_exit_probertt(rack, cts);
19984 			}
19985 		}
19986 	} else
19987 		tot_idle = 0;
19988 	if (rack_use_fsb &&
19989 	    (rack->r_ctl.fsb.tcp_ip_hdr) &&
19990 	    (rack->r_fsb_inited == 0) &&
19991 	    (rack->r_state != TCPS_CLOSED))
19992 		rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
19993 	if (rack->rc_sendvars_notset == 1) {
19994 		rack->rc_sendvars_notset = 0;
19995 		/*
19996 		 * Make sure any TCP timers (keep-alive) is not running.
19997 		 */
19998 		tcp_timer_stop(tp);
19999 	}
20000 	if ((rack->rack_no_prr == 1) &&
20001 	    (rack->rc_always_pace == 0)) {
20002 		/*
20003 		 * Sanity check before sending, if we have
20004 		 * no-pacing enabled and prr is turned off that
20005 		 * is a logistics error. Correct this by turnning
20006 		 * prr back on. A user *must* set some form of
20007 		 * pacing in order to turn PRR off. We do this
20008 		 * in the output path so that we can avoid socket
20009 		 * option ordering issues that would occur if we
20010 		 * tried to do it while setting rack_no_prr on.
20011 		 */
20012 		rack->rack_no_prr = 0;
20013 	}
20014 	if ((rack->pcm_enabled == 1) &&
20015 	    (rack->pcm_needed == 0) &&
20016 	    (tot_idle > 0)) {
20017 		/*
20018 		 * We have been idle some micro seconds. We need
20019 		 * to factor this in to see if a PCM is needed.
20020 		 */
20021 		uint32_t rtts_idle, rnds;
20022 
20023 		if (tp->t_srtt)
20024 			rtts_idle = tot_idle / tp->t_srtt;
20025 		else
20026 			rtts_idle = 0;
20027 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
20028 		rack->r_ctl.pcm_idle_rounds += rtts_idle;
20029 		if ((rnds + rack->r_ctl.pcm_idle_rounds)  >= rack_pcm_every_n_rounds) {
20030 			rack->pcm_needed = 1;
20031 			rack_log_pcm(rack, 8, rack->r_ctl.last_pcm_round, rtts_idle, rack->r_ctl.current_round );
20032 		}
20033 	}
20034 again:
20035 	sendalot = 0;
20036 	cts = tcp_get_usecs(&tv);
20037 	ms_cts = tcp_tv_to_mssectick(&tv);
20038 	tso = 0;
20039 	mtu = 0;
20040 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
20041 	minseg = segsiz;
20042 	if (rack->r_ctl.rc_pace_max_segs == 0)
20043 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
20044 	else
20045 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
20046 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
20047 	    (rack->r_ctl.pcm_max_seg == 0)) {
20048 		/*
20049 		 * We set in our first send so we know that the ctf_fixed_maxseg
20050 		 * has been fully set. If we do it in rack_init() we most likely
20051 		 * see 512 bytes so we end up at 5120, not desirable.
20052 		 */
20053 		rack->r_ctl.pcm_max_seg = rc_init_window(rack);
20054 		if (rack->r_ctl.pcm_max_seg < (ctf_fixed_maxseg(tp) * 10)) {
20055 			/*
20056 			 * Assure our initial PCM probe is at least 10 MSS.
20057 			 */
20058 			rack->r_ctl.pcm_max_seg = ctf_fixed_maxseg(tp) * 10;
20059 		}
20060 	}
20061 	if ((rack->r_ctl.pcm_max_seg != 0)  && (rack->pcm_needed == 1)) {
20062 		uint32_t rw_avail, cwa;
20063 
20064 		if (tp->snd_wnd > ctf_outstanding(tp))
20065 			rw_avail = tp->snd_wnd - ctf_outstanding(tp);
20066 		else
20067 			rw_avail = 0;
20068 		if (tp->snd_cwnd > ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked))
20069 			cwa = tp->snd_cwnd -ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20070 		else
20071 			cwa = 0;
20072 		if ((cwa >= rack->r_ctl.pcm_max_seg) &&
20073 		    (rw_avail > rack->r_ctl.pcm_max_seg)) {
20074 			/* Raise up the max seg for this trip through */
20075 			pace_max_seg = rack->r_ctl.pcm_max_seg;
20076 			/* Disable any fast output */
20077 			rack->r_fast_output = 0;
20078 		}
20079 		if (rack_verbose_logging) {
20080 			rack_log_pcm(rack, 4,
20081 				     cwa, rack->r_ctl.pcm_max_seg, rw_avail);
20082 		}
20083 	}
20084 	sb_offset = tp->snd_max - tp->snd_una;
20085 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20086 	flags = tcp_outflags[tp->t_state];
20087 	while (rack->rc_free_cnt < rack_free_cache) {
20088 		rsm = rack_alloc(rack);
20089 		if (rsm == NULL) {
20090 			if (hpts_calling)
20091 				/* Retry in a ms */
20092 				slot = (1 * HPTS_USEC_IN_MSEC);
20093 			so = inp->inp_socket;
20094 			sb = &so->so_snd;
20095 			goto just_return_nolock;
20096 		}
20097 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
20098 		rack->rc_free_cnt++;
20099 		rsm = NULL;
20100 	}
20101 	sack_rxmit = 0;
20102 	len = 0;
20103 	rsm = NULL;
20104 	if (flags & TH_RST) {
20105 		SOCK_SENDBUF_LOCK(inp->inp_socket);
20106 		so = inp->inp_socket;
20107 		sb = &so->so_snd;
20108 		goto send;
20109 	}
20110 	if (rack->r_ctl.rc_resend) {
20111 		/* Retransmit timer */
20112 		rsm = rack->r_ctl.rc_resend;
20113 		rack->r_ctl.rc_resend = NULL;
20114 		len = rsm->r_end - rsm->r_start;
20115 		sack_rxmit = 1;
20116 		sendalot = 0;
20117 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20118 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20119 			 __func__, __LINE__,
20120 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20121 		sb_offset = rsm->r_start - tp->snd_una;
20122 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20123 	} else if (rack->r_collapse_point_valid &&
20124 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
20125 		/*
20126 		 * If an RSM is returned then enough time has passed
20127 		 * for us to retransmit it. Move up the collapse point,
20128 		 * since this rsm has its chance to retransmit now.
20129 		 */
20130 		tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
20131 		rack->r_ctl.last_collapse_point = rsm->r_end;
20132 		/* Are we done? */
20133 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
20134 			    rack->r_ctl.high_collapse_point))
20135 			rack->r_collapse_point_valid = 0;
20136 		sack_rxmit = 1;
20137 		/* We are not doing a TLP */
20138 		doing_tlp = 0;
20139 		len = rsm->r_end - rsm->r_start;
20140 		sb_offset = rsm->r_start - tp->snd_una;
20141 		sendalot = 0;
20142 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20143 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
20144 		/* We have a retransmit that takes precedence */
20145 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
20146 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
20147 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
20148 			/* Enter recovery if not induced by a time-out */
20149 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
20150 		}
20151 #ifdef INVARIANTS
20152 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
20153 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
20154 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
20155 		}
20156 #endif
20157 		len = rsm->r_end - rsm->r_start;
20158 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20159 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20160 			 __func__, __LINE__,
20161 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20162 		sb_offset = rsm->r_start - tp->snd_una;
20163 		sendalot = 0;
20164 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20165 		if (len > 0) {
20166 			sack_rxmit = 1;
20167 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
20168 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
20169 					 min(len, segsiz));
20170 		}
20171 	} else if (rack->r_ctl.rc_tlpsend) {
20172 		/* Tail loss probe */
20173 		long cwin;
20174 		long tlen;
20175 
20176 		/*
20177 		 * Check if we can do a TLP with a RACK'd packet
20178 		 * this can happen if we are not doing the rack
20179 		 * cheat and we skipped to a TLP and it
20180 		 * went off.
20181 		 */
20182 		rsm = rack->r_ctl.rc_tlpsend;
20183 		/* We are doing a TLP make sure the flag is preent */
20184 		rsm->r_flags |= RACK_TLP;
20185 		rack->r_ctl.rc_tlpsend = NULL;
20186 		sack_rxmit = 1;
20187 		tlen = rsm->r_end - rsm->r_start;
20188 		if (tlen > segsiz)
20189 			tlen = segsiz;
20190 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20191 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20192 			 __func__, __LINE__,
20193 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20194 		sb_offset = rsm->r_start - tp->snd_una;
20195 		cwin = min(tp->snd_wnd, tlen);
20196 		len = cwin;
20197 	}
20198 	if (rack->r_must_retran &&
20199 	    (doing_tlp == 0) &&
20200 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20201 	    (rsm == NULL)) {
20202 		/*
20203 		 * There are two different ways that we
20204 		 * can get into this block:
20205 		 * a) This is a non-sack connection, we had a time-out
20206 		 *    and thus r_must_retran was set and everything
20207 		 *    left outstanding as been marked for retransmit.
20208 		 * b) The MTU of the path shrank, so that everything
20209 		 *    was marked to be retransmitted with the smaller
20210 		 *    mtu and r_must_retran was set.
20211 		 *
20212 		 * This means that we expect the sendmap (outstanding)
20213 		 * to all be marked must. We can use the tmap to
20214 		 * look at them.
20215 		 *
20216 		 */
20217 		int sendwin, flight;
20218 
20219 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20220 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20221 		if (flight >= sendwin) {
20222 			/*
20223 			 * We can't send yet.
20224 			 */
20225 			so = inp->inp_socket;
20226 			sb = &so->so_snd;
20227 			goto just_return_nolock;
20228 		}
20229 		/*
20230 		 * This is the case a/b mentioned above. All
20231 		 * outstanding/not-acked should be marked.
20232 		 * We can use the tmap to find them.
20233 		 */
20234 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20235 		if (rsm == NULL) {
20236 			/* TSNH */
20237 			rack->r_must_retran = 0;
20238 			rack->r_ctl.rc_out_at_rto = 0;
20239 			so = inp->inp_socket;
20240 			sb = &so->so_snd;
20241 			goto just_return_nolock;
20242 		}
20243 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20244 			/*
20245 			 * The first one does not have the flag, did we collapse
20246 			 * further up in our list?
20247 			 */
20248 			rack->r_must_retran = 0;
20249 			rack->r_ctl.rc_out_at_rto = 0;
20250 			rsm = NULL;
20251 			sack_rxmit = 0;
20252 		} else {
20253 			sack_rxmit = 1;
20254 			len = rsm->r_end - rsm->r_start;
20255 			sb_offset = rsm->r_start - tp->snd_una;
20256 			sendalot = 0;
20257 			if ((rack->full_size_rxt == 0) &&
20258 			    (rack->shape_rxt_to_pacing_min == 0) &&
20259 			    (len >= segsiz))
20260 				len = segsiz;
20261 			else if (rack->shape_rxt_to_pacing_min &&
20262 				 rack->gp_ready) {
20263 				/* We use pacing min as shaping len req */
20264 				uint32_t maxlen;
20265 
20266 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20267 				if (len > maxlen)
20268 					len = maxlen;
20269 			}
20270 			/*
20271 			 * Delay removing the flag RACK_MUST_RXT so
20272 			 * that the fastpath for retransmit will
20273 			 * work with this rsm.
20274 			 */
20275 		}
20276 	}
20277 	/*
20278 	 * Enforce a connection sendmap count limit if set
20279 	 * as long as we are not retransmiting.
20280 	 */
20281 	if ((rsm == NULL) &&
20282 	    (V_tcp_map_entries_limit > 0) &&
20283 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20284 		counter_u64_add(rack_to_alloc_limited, 1);
20285 		if (!rack->alloc_limit_reported) {
20286 			rack->alloc_limit_reported = 1;
20287 			counter_u64_add(rack_alloc_limited_conns, 1);
20288 		}
20289 		so = inp->inp_socket;
20290 		sb = &so->so_snd;
20291 		goto just_return_nolock;
20292 	}
20293 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20294 		/* we are retransmitting the fin */
20295 		len--;
20296 		if (len) {
20297 			/*
20298 			 * When retransmitting data do *not* include the
20299 			 * FIN. This could happen from a TLP probe.
20300 			 */
20301 			flags &= ~TH_FIN;
20302 		}
20303 	}
20304 	if (rsm && rack->r_fsb_inited &&
20305 	    rack_use_rsm_rfo &&
20306 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20307 		int ret;
20308 
20309 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20310 		if (ret == 0)
20311 			return (0);
20312 	}
20313 	so = inp->inp_socket;
20314 	sb = &so->so_snd;
20315 	if (do_a_prefetch == 0) {
20316 		kern_prefetch(sb, &do_a_prefetch);
20317 		do_a_prefetch = 1;
20318 	}
20319 #ifdef NETFLIX_SHARED_CWND
20320 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20321 	    rack->rack_enable_scwnd) {
20322 		/* We are doing cwnd sharing */
20323 		if (rack->gp_ready &&
20324 		    (rack->rack_attempted_scwnd == 0) &&
20325 		    (rack->r_ctl.rc_scw == NULL) &&
20326 		    tp->t_lib) {
20327 			/* The pcbid is in, lets make an attempt */
20328 			counter_u64_add(rack_try_scwnd, 1);
20329 			rack->rack_attempted_scwnd = 1;
20330 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20331 								   &rack->r_ctl.rc_scw_index,
20332 								   segsiz);
20333 		}
20334 		if (rack->r_ctl.rc_scw &&
20335 		    (rack->rack_scwnd_is_idle == 1) &&
20336 		    sbavail(&so->so_snd)) {
20337 			/* we are no longer out of data */
20338 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20339 			rack->rack_scwnd_is_idle = 0;
20340 		}
20341 		if (rack->r_ctl.rc_scw) {
20342 			/* First lets update and get the cwnd */
20343 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20344 										       rack->r_ctl.rc_scw_index,
20345 										       tp->snd_cwnd, tp->snd_wnd, segsiz);
20346 		}
20347 	}
20348 #endif
20349 	/*
20350 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
20351 	 * state flags.
20352 	 */
20353 	if (tp->t_flags & TF_NEEDFIN)
20354 		flags |= TH_FIN;
20355 	if (tp->t_flags & TF_NEEDSYN)
20356 		flags |= TH_SYN;
20357 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20358 		void *end_rsm;
20359 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20360 		if (end_rsm)
20361 			kern_prefetch(end_rsm, &prefetch_rsm);
20362 		prefetch_rsm = 1;
20363 	}
20364 	SOCK_SENDBUF_LOCK(so);
20365 	if ((sack_rxmit == 0) &&
20366 	    (TCPS_HAVEESTABLISHED(tp->t_state) ||
20367 	    (tp->t_flags & TF_FASTOPEN))) {
20368 		/*
20369 		 * We are not retransmitting (sack_rxmit is 0) so we
20370 		 * are sending new data. This is always based on snd_max.
20371 		 * Now in theory snd_max may be equal to snd_una, if so
20372 		 * then nothing is outstanding and the offset would be 0.
20373 		 */
20374 		uint32_t avail;
20375 
20376 		avail = sbavail(sb);
20377 		if (SEQ_GT(tp->snd_max, tp->snd_una) && avail)
20378 			sb_offset = tp->snd_max - tp->snd_una;
20379 		else
20380 			sb_offset = 0;
20381 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20382 			if (rack->r_ctl.rc_tlp_new_data) {
20383 				/* TLP is forcing out new data */
20384 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20385 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20386 				}
20387 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20388 					if (tp->snd_wnd > sb_offset)
20389 						len = tp->snd_wnd - sb_offset;
20390 					else
20391 						len = 0;
20392 				} else {
20393 					len = rack->r_ctl.rc_tlp_new_data;
20394 				}
20395 				rack->r_ctl.rc_tlp_new_data = 0;
20396 			}  else {
20397 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20398 			}
20399 			if ((rack->r_ctl.crte == NULL) &&
20400 			    IN_FASTRECOVERY(tp->t_flags) &&
20401 			    (rack->full_size_rxt == 0) &&
20402 			    (rack->shape_rxt_to_pacing_min == 0) &&
20403 			    (len > segsiz)) {
20404 				/*
20405 				 * For prr=off, we need to send only 1 MSS
20406 				 * at a time. We do this because another sack could
20407 				 * be arriving that causes us to send retransmits and
20408 				 * we don't want to be on a long pace due to a larger send
20409 				 * that keeps us from sending out the retransmit.
20410 				 */
20411 				len = segsiz;
20412 			} else if (rack->shape_rxt_to_pacing_min &&
20413 				   rack->gp_ready) {
20414 				/* We use pacing min as shaping len req */
20415 				uint32_t maxlen;
20416 
20417 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20418 				if (len > maxlen)
20419 					len = maxlen;
20420 			}/* The else is full_size_rxt is on so send it all */
20421 		} else {
20422 			uint32_t outstanding;
20423 			/*
20424 			 * We are inside of a Fast recovery episode, this
20425 			 * is caused by a SACK or 3 dup acks. At this point
20426 			 * we have sent all the retransmissions and we rely
20427 			 * on PRR to dictate what we will send in the form of
20428 			 * new data.
20429 			 */
20430 
20431 			outstanding = tp->snd_max - tp->snd_una;
20432 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20433 				if (tp->snd_wnd > outstanding) {
20434 					len = tp->snd_wnd - outstanding;
20435 					/* Check to see if we have the data */
20436 					if ((sb_offset + len) > avail) {
20437 						/* It does not all fit */
20438 						if (avail > sb_offset)
20439 							len = avail - sb_offset;
20440 						else
20441 							len = 0;
20442 					}
20443 				} else {
20444 					len = 0;
20445 				}
20446 			} else if (avail > sb_offset) {
20447 				len = avail - sb_offset;
20448 			} else {
20449 				len = 0;
20450 			}
20451 			if (len > 0) {
20452 				if (len > rack->r_ctl.rc_prr_sndcnt) {
20453 					len = rack->r_ctl.rc_prr_sndcnt;
20454 				}
20455 				if (len > 0) {
20456 					sub_from_prr = 1;
20457 				}
20458 			}
20459 			if (len > segsiz) {
20460 				/*
20461 				 * We should never send more than a MSS when
20462 				 * retransmitting or sending new data in prr
20463 				 * mode unless the override flag is on. Most
20464 				 * likely the PRR algorithm is not going to
20465 				 * let us send a lot as well :-)
20466 				 */
20467 				if (rack->r_ctl.rc_prr_sendalot == 0) {
20468 					len = segsiz;
20469 				}
20470 			} else if (len < segsiz) {
20471 				/*
20472 				 * Do we send any? The idea here is if the
20473 				 * send empty's the socket buffer we want to
20474 				 * do it. However if not then lets just wait
20475 				 * for our prr_sndcnt to get bigger.
20476 				 */
20477 				long leftinsb;
20478 
20479 				leftinsb = sbavail(sb) - sb_offset;
20480 				if (leftinsb > len) {
20481 					/* This send does not empty the sb */
20482 					len = 0;
20483 				}
20484 			}
20485 		}
20486 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20487 		/*
20488 		 * If you have not established
20489 		 * and are not doing FAST OPEN
20490 		 * no data please.
20491 		 */
20492 		if ((sack_rxmit == 0) &&
20493 		    !(tp->t_flags & TF_FASTOPEN)) {
20494 			len = 0;
20495 			sb_offset = 0;
20496 		}
20497 	}
20498 	if (prefetch_so_done == 0) {
20499 		kern_prefetch(so, &prefetch_so_done);
20500 		prefetch_so_done = 1;
20501 	}
20502 	orig_len = len;
20503 	/*
20504 	 * Lop off SYN bit if it has already been sent.  However, if this is
20505 	 * SYN-SENT state and if segment contains data and if we don't know
20506 	 * that foreign host supports TAO, suppress sending segment.
20507 	 */
20508 	if ((flags & TH_SYN) &&
20509 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
20510 	    ((sack_rxmit == 0) &&
20511 	     (tp->t_rxtshift == 0))) {
20512 		/*
20513 		 * When sending additional segments following a TFO SYN|ACK,
20514 		 * do not include the SYN bit.
20515 		 */
20516 		if ((tp->t_flags & TF_FASTOPEN) &&
20517 		    (tp->t_state == TCPS_SYN_RECEIVED))
20518 			flags &= ~TH_SYN;
20519 	}
20520 	/*
20521 	 * Be careful not to send data and/or FIN on SYN segments. This
20522 	 * measure is needed to prevent interoperability problems with not
20523 	 * fully conformant TCP implementations.
20524 	 */
20525 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20526 		len = 0;
20527 		flags &= ~TH_FIN;
20528 	}
20529 	/*
20530 	 * On TFO sockets, ensure no data is sent in the following cases:
20531 	 *
20532 	 *  - When retransmitting SYN|ACK on a passively-created socket
20533 	 *
20534 	 *  - When retransmitting SYN on an actively created socket
20535 	 *
20536 	 *  - When sending a zero-length cookie (cookie request) on an
20537 	 *    actively created socket
20538 	 *
20539 	 *  - When the socket is in the CLOSED state (RST is being sent)
20540 	 */
20541 	if ((tp->t_flags & TF_FASTOPEN) &&
20542 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20543 	     ((tp->t_state == TCPS_SYN_SENT) &&
20544 	      (tp->t_tfo_client_cookie_len == 0)) ||
20545 	     (flags & TH_RST))) {
20546 		sack_rxmit = 0;
20547 		len = 0;
20548 	}
20549 	/* Without fast-open there should never be data sent on a SYN */
20550 	if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
20551 		len = 0;
20552 	}
20553 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20554 		/* We only send 1 MSS if we have a DSACK block */
20555 		add_flag |= RACK_SENT_W_DSACK;
20556 		len = segsiz;
20557 	}
20558 	if (len <= 0) {
20559 		/*
20560 		 * We have nothing to send, or the window shrank, or
20561 		 * is closed, do we need to go into persists?
20562 		 */
20563 		len = 0;
20564 		if ((tp->snd_wnd == 0) &&
20565 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20566 		    (tp->snd_una == tp->snd_max) &&
20567 		    (sb_offset < (int)sbavail(sb))) {
20568 			rack_enter_persist(tp, rack, cts, tp->snd_una);
20569 		}
20570 	} else if ((rsm == NULL) &&
20571 		   (doing_tlp == 0) &&
20572 		   (len < pace_max_seg)) {
20573 		/*
20574 		 * We are not sending a maximum sized segment for
20575 		 * some reason. Should we not send anything (think
20576 		 * sws or persists)?
20577 		 */
20578 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20579 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20580 		    (len < minseg) &&
20581 		    (len < (int)(sbavail(sb) - sb_offset))) {
20582 			/*
20583 			 * Here the rwnd is less than
20584 			 * the minimum pacing size, this is not a retransmit,
20585 			 * we are established and
20586 			 * the send is not the last in the socket buffer
20587 			 * we send nothing, and we may enter persists
20588 			 * if nothing is outstanding.
20589 			 */
20590 			len = 0;
20591 			if (tp->snd_max == tp->snd_una) {
20592 				/*
20593 				 * Nothing out we can
20594 				 * go into persists.
20595 				 */
20596 				rack_enter_persist(tp, rack, cts, tp->snd_una);
20597 			}
20598 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20599 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20600 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20601 			   (len < minseg)) {
20602 			/*
20603 			 * Here we are not retransmitting, and
20604 			 * the cwnd is not so small that we could
20605 			 * not send at least a min size (rxt timer
20606 			 * not having gone off), We have 2 segments or
20607 			 * more already in flight, its not the tail end
20608 			 * of the socket buffer  and the cwnd is blocking
20609 			 * us from sending out a minimum pacing segment size.
20610 			 * Lets not send anything.
20611 			 */
20612 			len = 0;
20613 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20614 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20615 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20616 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20617 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
20618 			/*
20619 			 * Here we have a send window but we have
20620 			 * filled it up and we can't send another pacing segment.
20621 			 * We also have in flight more than 2 segments
20622 			 * and we are not completing the sb i.e. we allow
20623 			 * the last bytes of the sb to go out even if
20624 			 * its not a full pacing segment.
20625 			 */
20626 			len = 0;
20627 		} else if ((rack->r_ctl.crte != NULL) &&
20628 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20629 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20630 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20631 			   (len < (int)(sbavail(sb) - sb_offset))) {
20632 			/*
20633 			 * Here we are doing hardware pacing, this is not a TLP,
20634 			 * we are not sending a pace max segment size, there is rwnd
20635 			 * room to send at least N pace_max_seg, the cwnd is greater
20636 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
20637 			 * more segments in flight and its not the tail of the socket buffer.
20638 			 *
20639 			 * We don't want to send instead we need to get more ack's in to
20640 			 * allow us to send a full pacing segment. Normally, if we are pacing
20641 			 * about the right speed, we should have finished our pacing
20642 			 * send as most of the acks have come back if we are at the
20643 			 * right rate. This is a bit fuzzy since return path delay
20644 			 * can delay the acks, which is why we want to make sure we
20645 			 * have cwnd space to have a bit more than a max pace segments in flight.
20646 			 *
20647 			 * If we have not gotten our acks back we are pacing at too high a
20648 			 * rate delaying will not hurt and will bring our GP estimate down by
20649 			 * injecting the delay. If we don't do this we will send
20650 			 * 2 MSS out in response to the acks being clocked in which
20651 			 * defeats the point of hw-pacing (i.e. to help us get
20652 			 * larger TSO's out).
20653 			 */
20654 			len = 0;
20655 		}
20656 
20657 	}
20658 	/* len will be >= 0 after this point. */
20659 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20660 	rack_sndbuf_autoscale(rack);
20661 	/*
20662 	 * Decide if we can use TCP Segmentation Offloading (if supported by
20663 	 * hardware).
20664 	 *
20665 	 * TSO may only be used if we are in a pure bulk sending state.  The
20666 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20667 	 * options prevent using TSO.  With TSO the TCP header is the same
20668 	 * (except for the sequence number) for all generated packets.  This
20669 	 * makes it impossible to transmit any options which vary per
20670 	 * generated segment or packet.
20671 	 *
20672 	 * IPv4 handling has a clear separation of ip options and ip header
20673 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20674 	 * the right thing below to provide length of just ip options and thus
20675 	 * checking for ipoptlen is enough to decide if ip options are present.
20676 	 */
20677 	ipoptlen = 0;
20678 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20679 	/*
20680 	 * Pre-calculate here as we save another lookup into the darknesses
20681 	 * of IPsec that way and can actually decide if TSO is ok.
20682 	 */
20683 #ifdef INET6
20684 	if (isipv6 && IPSEC_ENABLED(ipv6))
20685 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20686 #ifdef INET
20687 	else
20688 #endif
20689 #endif				/* INET6 */
20690 #ifdef INET
20691 		if (IPSEC_ENABLED(ipv4))
20692 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20693 #endif				/* INET */
20694 #endif
20695 
20696 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20697 	ipoptlen += ipsec_optlen;
20698 #endif
20699 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20700 	    (tp->t_port == 0) &&
20701 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
20702 	    sack_rxmit == 0 &&
20703 	    ipoptlen == 0)
20704 		tso = 1;
20705 	{
20706 		uint32_t outstanding __unused;
20707 
20708 		outstanding = tp->snd_max - tp->snd_una;
20709 		if (tp->t_flags & TF_SENTFIN) {
20710 			/*
20711 			 * If we sent a fin, snd_max is 1 higher than
20712 			 * snd_una
20713 			 */
20714 			outstanding--;
20715 		}
20716 		if (sack_rxmit) {
20717 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20718 				flags &= ~TH_FIN;
20719 		}
20720 	}
20721 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20722 		      (long)TCP_MAXWIN << tp->rcv_scale);
20723 
20724 	/*
20725 	 * Sender silly window avoidance.   We transmit under the following
20726 	 * conditions when len is non-zero:
20727 	 *
20728 	 * - We have a full segment (or more with TSO) - This is the last
20729 	 * buffer in a write()/send() and we are either idle or running
20730 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
20731 	 * then 1/2 the maximum send window's worth of data (receiver may be
20732 	 * limited the window size) - we need to retransmit
20733 	 */
20734 	if (len) {
20735 		if (len >= segsiz) {
20736 			goto send;
20737 		}
20738 		/*
20739 		 * NOTE! on localhost connections an 'ack' from the remote
20740 		 * end may occur synchronously with the output and cause us
20741 		 * to flush a buffer queued with moretocome.  XXX
20742 		 *
20743 		 */
20744 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
20745 		    (idle || (tp->t_flags & TF_NODELAY)) &&
20746 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20747 		    (tp->t_flags & TF_NOPUSH) == 0) {
20748 			pass = 2;
20749 			goto send;
20750 		}
20751 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
20752 			pass = 22;
20753 			goto send;
20754 		}
20755 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20756 			pass = 4;
20757 			goto send;
20758 		}
20759 		if (sack_rxmit) {
20760 			pass = 6;
20761 			goto send;
20762 		}
20763 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20764 		    (ctf_outstanding(tp) < (segsiz * 2))) {
20765 			/*
20766 			 * We have less than two MSS outstanding (delayed ack)
20767 			 * and our rwnd will not let us send a full sized
20768 			 * MSS. Lets go ahead and let this small segment
20769 			 * out because we want to try to have at least two
20770 			 * packets inflight to not be caught by delayed ack.
20771 			 */
20772 			pass = 12;
20773 			goto send;
20774 		}
20775 	}
20776 	/*
20777 	 * Sending of standalone window updates.
20778 	 *
20779 	 * Window updates are important when we close our window due to a
20780 	 * full socket buffer and are opening it again after the application
20781 	 * reads data from it.  Once the window has opened again and the
20782 	 * remote end starts to send again the ACK clock takes over and
20783 	 * provides the most current window information.
20784 	 *
20785 	 * We must avoid the silly window syndrome whereas every read from
20786 	 * the receive buffer, no matter how small, causes a window update
20787 	 * to be sent.  We also should avoid sending a flurry of window
20788 	 * updates when the socket buffer had queued a lot of data and the
20789 	 * application is doing small reads.
20790 	 *
20791 	 * Prevent a flurry of pointless window updates by only sending an
20792 	 * update when we can increase the advertized window by more than
20793 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
20794 	 * full or is very small be more aggressive and send an update
20795 	 * whenever we can increase by two mss sized segments. In all other
20796 	 * situations the ACK's to new incoming data will carry further
20797 	 * window increases.
20798 	 *
20799 	 * Don't send an independent window update if a delayed ACK is
20800 	 * pending (it will get piggy-backed on it) or the remote side
20801 	 * already has done a half-close and won't send more data.  Skip
20802 	 * this if the connection is in T/TCP half-open state.
20803 	 */
20804 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20805 	    !(tp->t_flags & TF_DELACK) &&
20806 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
20807 		/*
20808 		 * "adv" is the amount we could increase the window, taking
20809 		 * into account that we are limited by TCP_MAXWIN <<
20810 		 * tp->rcv_scale.
20811 		 */
20812 		int32_t adv;
20813 		int oldwin;
20814 
20815 		adv = recwin;
20816 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20817 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
20818 			if (adv > oldwin)
20819 				adv -= oldwin;
20820 			else {
20821 				/* We can't increase the window */
20822 				adv = 0;
20823 			}
20824 		} else
20825 			oldwin = 0;
20826 
20827 		/*
20828 		 * If the new window size ends up being the same as or less
20829 		 * than the old size when it is scaled, then don't force
20830 		 * a window update.
20831 		 */
20832 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20833 			goto dontupdate;
20834 
20835 		if (adv >= (int32_t)(2 * segsiz) &&
20836 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20837 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20838 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20839 			pass = 7;
20840 			goto send;
20841 		}
20842 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20843 			pass = 23;
20844 			goto send;
20845 		}
20846 	}
20847 dontupdate:
20848 
20849 	/*
20850 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
20851 	 * is also a catch-all for the retransmit timer timeout case.
20852 	 */
20853 	if (tp->t_flags & TF_ACKNOW) {
20854 		pass = 8;
20855 		goto send;
20856 	}
20857 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
20858 		pass = 9;
20859 		goto send;
20860 	}
20861 	/*
20862 	 * If our state indicates that FIN should be sent and we have not
20863 	 * yet done so, then we need to send.
20864 	 */
20865 	if ((flags & TH_FIN) &&
20866 	    (tp->snd_max == tp->snd_una)) {
20867 		pass = 11;
20868 		goto send;
20869 	}
20870 	/*
20871 	 * No reason to send a segment, just return.
20872 	 */
20873 just_return:
20874 	SOCK_SENDBUF_UNLOCK(so);
20875 just_return_nolock:
20876 	{
20877 		int app_limited = CTF_JR_SENT_DATA;
20878 
20879 		if ((tp->t_flags & TF_FASTOPEN) == 0 &&
20880 		    (flags & TH_FIN) &&
20881 		    (len == 0) &&
20882 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
20883 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
20884 			/*
20885 			 * Ok less than or right at a MSS is
20886 			 * outstanding. The original FreeBSD stack would
20887 			 * have sent a FIN, which can speed things up for
20888 			 * a transactional application doing a MSG_WAITALL.
20889 			 * To speed things up since we do *not* send a FIN
20890 			 * if data is outstanding, we send a "challenge ack".
20891 			 * The idea behind that is instead of having to have
20892 			 * the peer wait for the delayed-ack timer to run off
20893 			 * we send an ack that makes the peer send us an ack.
20894 			 */
20895 			rack_send_ack_challange(rack);
20896 		}
20897 		if (tot_len_this_send > 0) {
20898 			rack->r_ctl.fsb.recwin = recwin;
20899 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz, __LINE__);
20900 			if ((error == 0) &&
20901 			    rack_use_rfo &&
20902 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
20903 			    (ipoptlen == 0) &&
20904 			    rack->r_fsb_inited &&
20905 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
20906 			    ((IN_RECOVERY(tp->t_flags)) == 0) &&
20907 			    (rack->r_must_retran == 0) &&
20908 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
20909 			    (len > 0) && (orig_len > 0) &&
20910 			    (orig_len > len) &&
20911 			    ((orig_len - len) >= segsiz) &&
20912 			    ((optlen == 0) ||
20913 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
20914 				/* We can send at least one more MSS using our fsb */
20915 				rack_setup_fast_output(tp, rack, sb, len, orig_len,
20916 						       segsiz, pace_max_seg, hw_tls, flags);
20917 			} else
20918 				rack->r_fast_output = 0;
20919 			rack_log_fsb(rack, tp, so, flags,
20920 				     ipoptlen, orig_len, len, 0,
20921 				     1, optlen, __LINE__, 1);
20922 			/* Assure when we leave that snd_nxt will point to top */
20923 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
20924 				tp->snd_nxt = tp->snd_max;
20925 		} else {
20926 			int end_window = 0;
20927 			uint32_t seq = tp->gput_ack;
20928 
20929 			rsm = tqhash_max(rack->r_ctl.tqh);
20930 			if (rsm) {
20931 				/*
20932 				 * Mark the last sent that we just-returned (hinting
20933 				 * that delayed ack may play a role in any rtt measurement).
20934 				 */
20935 				rsm->r_just_ret = 1;
20936 			}
20937 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
20938 			rack->r_ctl.rc_agg_delayed = 0;
20939 			rack->r_early = 0;
20940 			rack->r_late = 0;
20941 			rack->r_ctl.rc_agg_early = 0;
20942 			if ((ctf_outstanding(tp) +
20943 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
20944 				 minseg)) >= tp->snd_wnd) {
20945 				/* We are limited by the rwnd */
20946 				app_limited = CTF_JR_RWND_LIMITED;
20947 				if (IN_FASTRECOVERY(tp->t_flags))
20948 					rack->r_ctl.rc_prr_sndcnt = 0;
20949 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
20950 				/* We are limited by whats available -- app limited */
20951 				app_limited = CTF_JR_APP_LIMITED;
20952 				if (IN_FASTRECOVERY(tp->t_flags))
20953 					rack->r_ctl.rc_prr_sndcnt = 0;
20954 			} else if ((idle == 0) &&
20955 				   ((tp->t_flags & TF_NODELAY) == 0) &&
20956 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20957 				   (len < segsiz)) {
20958 				/*
20959 				 * No delay is not on and the
20960 				 * user is sending less than 1MSS. This
20961 				 * brings out SWS avoidance so we
20962 				 * don't send. Another app-limited case.
20963 				 */
20964 				app_limited = CTF_JR_APP_LIMITED;
20965 			} else if (tp->t_flags & TF_NOPUSH) {
20966 				/*
20967 				 * The user has requested no push of
20968 				 * the last segment and we are
20969 				 * at the last segment. Another app
20970 				 * limited case.
20971 				 */
20972 				app_limited = CTF_JR_APP_LIMITED;
20973 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
20974 				/* Its the cwnd */
20975 				app_limited = CTF_JR_CWND_LIMITED;
20976 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
20977 				   (rack->rack_no_prr == 0) &&
20978 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
20979 				app_limited = CTF_JR_PRR;
20980 			} else {
20981 				/* Now why here are we not sending? */
20982 #ifdef NOW
20983 #ifdef INVARIANTS
20984 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
20985 #endif
20986 #endif
20987 				app_limited = CTF_JR_ASSESSING;
20988 			}
20989 			/*
20990 			 * App limited in some fashion, for our pacing GP
20991 			 * measurements we don't want any gap (even cwnd).
20992 			 * Close  down the measurement window.
20993 			 */
20994 			if (rack_cwnd_block_ends_measure &&
20995 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
20996 			     (app_limited == CTF_JR_PRR))) {
20997 				/*
20998 				 * The reason we are not sending is
20999 				 * the cwnd (or prr). We have been configured
21000 				 * to end the measurement window in
21001 				 * this case.
21002 				 */
21003 				end_window = 1;
21004 			} else if (rack_rwnd_block_ends_measure &&
21005 				   (app_limited == CTF_JR_RWND_LIMITED)) {
21006 				/*
21007 				 * We are rwnd limited and have been
21008 				 * configured to end the measurement
21009 				 * window in this case.
21010 				 */
21011 				end_window = 1;
21012 			} else if (app_limited == CTF_JR_APP_LIMITED) {
21013 				/*
21014 				 * A true application limited period, we have
21015 				 * ran out of data.
21016 				 */
21017 				end_window = 1;
21018 			} else if (app_limited == CTF_JR_ASSESSING) {
21019 				/*
21020 				 * In the assessing case we hit the end of
21021 				 * the if/else and had no known reason
21022 				 * This will panic us under invariants..
21023 				 *
21024 				 * If we get this out in logs we need to
21025 				 * investagate which reason we missed.
21026 				 */
21027 				end_window = 1;
21028 			}
21029 			if (end_window) {
21030 				uint8_t log = 0;
21031 
21032 				/* Adjust the Gput measurement */
21033 				if ((tp->t_flags & TF_GPUTINPROG) &&
21034 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
21035 					tp->gput_ack = tp->snd_max;
21036 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
21037 						/*
21038 						 * There is not enough to measure.
21039 						 */
21040 						tp->t_flags &= ~TF_GPUTINPROG;
21041 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
21042 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
21043 									   tp->gput_seq,
21044 									   0, 0, 18, __LINE__, NULL, 0);
21045 					} else
21046 						log = 1;
21047 				}
21048 				/* Mark the last packet has app limited */
21049 				rsm = tqhash_max(rack->r_ctl.tqh);
21050 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
21051 					if (rack->r_ctl.rc_app_limited_cnt == 0)
21052 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
21053 					else {
21054 						/*
21055 						 * Go out to the end app limited and mark
21056 						 * this new one as next and move the end_appl up
21057 						 * to this guy.
21058 						 */
21059 						if (rack->r_ctl.rc_end_appl)
21060 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
21061 						rack->r_ctl.rc_end_appl = rsm;
21062 					}
21063 					rsm->r_flags |= RACK_APP_LIMITED;
21064 					rack->r_ctl.rc_app_limited_cnt++;
21065 				}
21066 				if (log)
21067 					rack_log_pacing_delay_calc(rack,
21068 								   rack->r_ctl.rc_app_limited_cnt, seq,
21069 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
21070 			}
21071 		}
21072 		/* Check if we need to go into persists or not */
21073 		if ((tp->snd_max == tp->snd_una) &&
21074 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
21075 		    sbavail(sb) &&
21076 		    (sbavail(sb) > tp->snd_wnd) &&
21077 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
21078 			/* Yes lets make sure to move to persist before timer-start */
21079 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
21080 		}
21081 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
21082 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
21083 	}
21084 #ifdef NETFLIX_SHARED_CWND
21085 	if ((sbavail(sb) == 0) &&
21086 	    rack->r_ctl.rc_scw) {
21087 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21088 		rack->rack_scwnd_is_idle = 1;
21089 	}
21090 #endif
21091 #ifdef TCP_ACCOUNTING
21092 	if (tot_len_this_send > 0) {
21093 		crtsc = get_cyclecount();
21094 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21095 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
21096 		}
21097 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21098 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
21099 		}
21100 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21101 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
21102 		}
21103 	} else {
21104 		crtsc = get_cyclecount();
21105 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21106 			tp->tcp_cnt_counters[SND_LIMITED]++;
21107 		}
21108 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21109 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
21110 		}
21111 	}
21112 	sched_unpin();
21113 #endif
21114 	return (0);
21115 
21116 send:
21117 	if ((rack->r_ctl.crte != NULL) &&
21118 	    (rsm == NULL) &&
21119 	    ((rack->rc_hw_nobuf == 1) ||
21120 	     (rack_hw_check_queue && (check_done == 0)))) {
21121 		/*
21122 		 * We only want to do this once with the hw_check_queue,
21123 		 * for the enobuf case we would only do it once if
21124 		 * we come around to again, the flag will be clear.
21125 		 */
21126 		check_done = 1;
21127 		slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
21128 		if (slot) {
21129 			rack->r_ctl.rc_agg_delayed = 0;
21130 			rack->r_ctl.rc_agg_early = 0;
21131 			rack->r_early = 0;
21132 			rack->r_late = 0;
21133 			SOCK_SENDBUF_UNLOCK(so);
21134 			goto skip_all_send;
21135 		}
21136 	}
21137 	if (rsm || sack_rxmit)
21138 		counter_u64_add(rack_nfto_resend, 1);
21139 	else
21140 		counter_u64_add(rack_non_fto_send, 1);
21141 	if ((flags & TH_FIN) &&
21142 	    sbavail(sb)) {
21143 		/*
21144 		 * We do not transmit a FIN
21145 		 * with data outstanding. We
21146 		 * need to make it so all data
21147 		 * is acked first.
21148 		 */
21149 		flags &= ~TH_FIN;
21150 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21151 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21152 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
21153 			/*
21154 			 * Ok less than or right at a MSS is
21155 			 * outstanding. The original FreeBSD stack would
21156 			 * have sent a FIN, which can speed things up for
21157 			 * a transactional application doing a MSG_WAITALL.
21158 			 * To speed things up since we do *not* send a FIN
21159 			 * if data is outstanding, we send a "challenge ack".
21160 			 * The idea behind that is instead of having to have
21161 			 * the peer wait for the delayed-ack timer to run off
21162 			 * we send an ack that makes the peer send us an ack.
21163 			 */
21164 			rack_send_ack_challange(rack);
21165 		}
21166 	}
21167 	/* Enforce stack imposed max seg size if we have one */
21168 	if (pace_max_seg &&
21169 	    (len > pace_max_seg)) {
21170 		mark = 1;
21171 		len = pace_max_seg;
21172 	}
21173 	if ((rsm == NULL) &&
21174 	    (rack->pcm_in_progress == 0) &&
21175 	    (rack->r_ctl.pcm_max_seg > 0) &&
21176 	    (len >= rack->r_ctl.pcm_max_seg)) {
21177 		/* It is large enough for a measurement */
21178 		add_flag |= RACK_IS_PCM;
21179 		rack_log_pcm(rack, 5, len, rack->r_ctl.pcm_max_seg,  add_flag);
21180 	} else if (rack_verbose_logging) {
21181 		rack_log_pcm(rack, 6, len, rack->r_ctl.pcm_max_seg,  add_flag);
21182 	}
21183 
21184 	SOCKBUF_LOCK_ASSERT(sb);
21185 	if (len > 0) {
21186 		if (len >= segsiz)
21187 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
21188 		else
21189 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
21190 	}
21191 	/*
21192 	 * Before ESTABLISHED, force sending of initial options unless TCP
21193 	 * set not to do any options. NOTE: we assume that the IP/TCP header
21194 	 * plus TCP options always fit in a single mbuf, leaving room for a
21195 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
21196 	 * + optlen <= MCLBYTES
21197 	 */
21198 	optlen = 0;
21199 #ifdef INET6
21200 	if (isipv6)
21201 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21202 	else
21203 #endif
21204 		hdrlen = sizeof(struct tcpiphdr);
21205 
21206 	/*
21207 	 * Ok what seq are we sending from. If we have
21208 	 * no rsm to use, then we look at various bits,
21209 	 * if we are putting out a SYN it will be ISS.
21210 	 * If we are retransmitting a FIN it will
21211 	 * be snd_max-1 else its snd_max.
21212 	 */
21213 	if (rsm == NULL) {
21214 		if (flags & TH_SYN)
21215 			rack_seq = tp->iss;
21216 		else if ((flags & TH_FIN) &&
21217 			 (tp->t_flags & TF_SENTFIN))
21218 			rack_seq = tp->snd_max - 1;
21219 		else
21220 			rack_seq = tp->snd_max;
21221 	} else {
21222 		rack_seq = rsm->r_start;
21223 	}
21224 	/*
21225 	 * Compute options for segment. We only have to care about SYN and
21226 	 * established connection segments.  Options for SYN-ACK segments
21227 	 * are handled in TCP syncache.
21228 	 */
21229 	to.to_flags = 0;
21230 	if ((tp->t_flags & TF_NOOPT) == 0) {
21231 		/* Maximum segment size. */
21232 		if (flags & TH_SYN) {
21233 			to.to_mss = tcp_mssopt(&inp->inp_inc);
21234 			if (tp->t_port)
21235 				to.to_mss -= V_tcp_udp_tunneling_overhead;
21236 			to.to_flags |= TOF_MSS;
21237 
21238 			/*
21239 			 * On SYN or SYN|ACK transmits on TFO connections,
21240 			 * only include the TFO option if it is not a
21241 			 * retransmit, as the presence of the TFO option may
21242 			 * have caused the original SYN or SYN|ACK to have
21243 			 * been dropped by a middlebox.
21244 			 */
21245 			if ((tp->t_flags & TF_FASTOPEN) &&
21246 			    (tp->t_rxtshift == 0)) {
21247 				if (tp->t_state == TCPS_SYN_RECEIVED) {
21248 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21249 					to.to_tfo_cookie =
21250 						(u_int8_t *)&tp->t_tfo_cookie.server;
21251 					to.to_flags |= TOF_FASTOPEN;
21252 					wanted_cookie = 1;
21253 				} else if (tp->t_state == TCPS_SYN_SENT) {
21254 					to.to_tfo_len =
21255 						tp->t_tfo_client_cookie_len;
21256 					to.to_tfo_cookie =
21257 						tp->t_tfo_cookie.client;
21258 					to.to_flags |= TOF_FASTOPEN;
21259 					wanted_cookie = 1;
21260 					/*
21261 					 * If we wind up having more data to
21262 					 * send with the SYN than can fit in
21263 					 * one segment, don't send any more
21264 					 * until the SYN|ACK comes back from
21265 					 * the other end.
21266 					 */
21267 					sendalot = 0;
21268 				}
21269 			}
21270 		}
21271 		/* Window scaling. */
21272 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21273 			to.to_wscale = tp->request_r_scale;
21274 			to.to_flags |= TOF_SCALE;
21275 		}
21276 		/* Timestamps. */
21277 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
21278 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21279 			uint32_t ts_to_use;
21280 
21281 			if ((rack->r_rcvpath_rtt_up == 1) &&
21282 			    (ms_cts == rack->r_ctl.last_rcv_tstmp_for_rtt)) {
21283 				/*
21284 				 * When we are doing a rcv_rtt probe all
21285 				 * other timestamps use the next msec. This
21286 				 * is safe since our previous ack is in the
21287 				 * air and we will just have a few more
21288 				 * on the next ms. This assures that only
21289 				 * the one ack has the ms_cts that was on
21290 				 * our ack-probe.
21291 				 */
21292 				ts_to_use = ms_cts + 1;
21293 			} else {
21294 				ts_to_use = ms_cts;
21295 			}
21296 			to.to_tsval = ts_to_use + tp->ts_offset;
21297 			to.to_tsecr = tp->ts_recent;
21298 			to.to_flags |= TOF_TS;
21299 			if ((len == 0) &&
21300 			    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
21301 			    ((ms_cts - rack->r_ctl.last_rcv_tstmp_for_rtt) > RCV_PATH_RTT_MS) &&
21302 			    (tp->snd_una == tp->snd_max) &&
21303 			    (flags & TH_ACK) &&
21304 			    (sbavail(sb) == 0) &&
21305 			    (rack->r_ctl.current_round != 0) &&
21306 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
21307 			    (rack->r_rcvpath_rtt_up == 0)) {
21308 				rack->r_ctl.last_rcv_tstmp_for_rtt = ms_cts;
21309 				rack->r_ctl.last_time_of_arm_rcv = cts;
21310 				rack->r_rcvpath_rtt_up = 1;
21311 				/* Subtract 1 from seq to force a response */
21312 				rack_seq--;
21313 			}
21314 		}
21315 		/* Set receive buffer autosizing timestamp. */
21316 		if (tp->rfbuf_ts == 0 &&
21317 		    (so->so_rcv.sb_flags & SB_AUTOSIZE)) {
21318 			tp->rfbuf_ts = ms_cts;
21319 		}
21320 		/* Selective ACK's. */
21321 		if (tp->t_flags & TF_SACK_PERMIT) {
21322 			if (flags & TH_SYN)
21323 				to.to_flags |= TOF_SACKPERM;
21324 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21325 				 tp->rcv_numsacks > 0) {
21326 				to.to_flags |= TOF_SACK;
21327 				to.to_nsacks = tp->rcv_numsacks;
21328 				to.to_sacks = (u_char *)tp->sackblks;
21329 			}
21330 		}
21331 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21332 		/* TCP-MD5 (RFC2385). */
21333 		if (tp->t_flags & TF_SIGNATURE)
21334 			to.to_flags |= TOF_SIGNATURE;
21335 #endif
21336 
21337 		/* Processing the options. */
21338 		hdrlen += optlen = tcp_addoptions(&to, opt);
21339 		/*
21340 		 * If we wanted a TFO option to be added, but it was unable
21341 		 * to fit, ensure no data is sent.
21342 		 */
21343 		if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
21344 		    !(to.to_flags & TOF_FASTOPEN))
21345 			len = 0;
21346 	}
21347 	if (tp->t_port) {
21348 		if (V_tcp_udp_tunneling_port == 0) {
21349 			/* The port was removed?? */
21350 			SOCK_SENDBUF_UNLOCK(so);
21351 #ifdef TCP_ACCOUNTING
21352 			crtsc = get_cyclecount();
21353 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21354 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21355 			}
21356 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21357 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21358 			}
21359 			sched_unpin();
21360 #endif
21361 			return (EHOSTUNREACH);
21362 		}
21363 		hdrlen += sizeof(struct udphdr);
21364 	}
21365 #ifdef INET6
21366 	if (isipv6)
21367 		ipoptlen = ip6_optlen(inp);
21368 	else
21369 #endif
21370 		if (inp->inp_options)
21371 			ipoptlen = inp->inp_options->m_len -
21372 				offsetof(struct ipoption, ipopt_list);
21373 		else
21374 			ipoptlen = 0;
21375 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21376 	ipoptlen += ipsec_optlen;
21377 #endif
21378 
21379 	/*
21380 	 * Adjust data length if insertion of options will bump the packet
21381 	 * length beyond the t_maxseg length. Clear the FIN bit because we
21382 	 * cut off the tail of the segment.
21383 	 */
21384 	if (len + optlen + ipoptlen > tp->t_maxseg) {
21385 		if (tso) {
21386 			uint32_t if_hw_tsomax;
21387 			uint32_t moff;
21388 			int32_t max_len;
21389 
21390 			/* extract TSO information */
21391 			if_hw_tsomax = tp->t_tsomax;
21392 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21393 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21394 			KASSERT(ipoptlen == 0,
21395 				("%s: TSO can't do IP options", __func__));
21396 
21397 			/*
21398 			 * Check if we should limit by maximum payload
21399 			 * length:
21400 			 */
21401 			if (if_hw_tsomax != 0) {
21402 				/* compute maximum TSO length */
21403 				max_len = (if_hw_tsomax - hdrlen -
21404 					   max_linkhdr);
21405 				if (max_len <= 0) {
21406 					len = 0;
21407 				} else if (len > max_len) {
21408 					sendalot = 1;
21409 					len = max_len;
21410 					mark = 2;
21411 				}
21412 			}
21413 			/*
21414 			 * Prevent the last segment from being fractional
21415 			 * unless the send sockbuf can be emptied:
21416 			 */
21417 			max_len = (tp->t_maxseg - optlen);
21418 			if ((sb_offset + len) < sbavail(sb)) {
21419 				moff = len % (u_int)max_len;
21420 				if (moff != 0) {
21421 					mark = 3;
21422 					len -= moff;
21423 				}
21424 			}
21425 			/*
21426 			 * In case there are too many small fragments don't
21427 			 * use TSO:
21428 			 */
21429 			if (len <= max_len) {
21430 				mark = 4;
21431 				tso = 0;
21432 			}
21433 			/*
21434 			 * Send the FIN in a separate segment after the bulk
21435 			 * sending is done. We don't trust the TSO
21436 			 * implementations to clear the FIN flag on all but
21437 			 * the last segment.
21438 			 */
21439 			if (tp->t_flags & TF_NEEDFIN) {
21440 				sendalot = 4;
21441 			}
21442 		} else {
21443 			mark = 5;
21444 			if (optlen + ipoptlen >= tp->t_maxseg) {
21445 				/*
21446 				 * Since we don't have enough space to put
21447 				 * the IP header chain and the TCP header in
21448 				 * one packet as required by RFC 7112, don't
21449 				 * send it. Also ensure that at least one
21450 				 * byte of the payload can be put into the
21451 				 * TCP segment.
21452 				 */
21453 				SOCK_SENDBUF_UNLOCK(so);
21454 				error = EMSGSIZE;
21455 				sack_rxmit = 0;
21456 				goto out;
21457 			}
21458 			len = tp->t_maxseg - optlen - ipoptlen;
21459 			sendalot = 5;
21460 		}
21461 	} else {
21462 		tso = 0;
21463 		mark = 6;
21464 	}
21465 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21466 		("%s: len > IP_MAXPACKET", __func__));
21467 #ifdef DIAGNOSTIC
21468 #ifdef INET6
21469 	if (max_linkhdr + hdrlen > MCLBYTES)
21470 #else
21471 		if (max_linkhdr + hdrlen > MHLEN)
21472 #endif
21473 			panic("tcphdr too big");
21474 #endif
21475 
21476 	/*
21477 	 * This KASSERT is here to catch edge cases at a well defined place.
21478 	 * Before, those had triggered (random) panic conditions further
21479 	 * down.
21480 	 */
21481 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21482 	if ((len == 0) &&
21483 	    (flags & TH_FIN) &&
21484 	    (sbused(sb))) {
21485 		/*
21486 		 * We have outstanding data, don't send a fin by itself!.
21487 		 *
21488 		 * Check to see if we need to send a challenge ack.
21489 		 */
21490 		if ((sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21491 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
21492 			/*
21493 			 * Ok less than or right at a MSS is
21494 			 * outstanding. The original FreeBSD stack would
21495 			 * have sent a FIN, which can speed things up for
21496 			 * a transactional application doing a MSG_WAITALL.
21497 			 * To speed things up since we do *not* send a FIN
21498 			 * if data is outstanding, we send a "challenge ack".
21499 			 * The idea behind that is instead of having to have
21500 			 * the peer wait for the delayed-ack timer to run off
21501 			 * we send an ack that makes the peer send us an ack.
21502 			 */
21503 			rack_send_ack_challange(rack);
21504 		}
21505 		goto just_return;
21506 	}
21507 	/*
21508 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
21509 	 * and initialize the header from the template for sends on this
21510 	 * connection.
21511 	 */
21512 	hw_tls = tp->t_nic_ktls_xmit != 0;
21513 	if (len) {
21514 		uint32_t max_val;
21515 		uint32_t moff;
21516 
21517 		if (pace_max_seg)
21518 			max_val = pace_max_seg;
21519 		else
21520 			max_val = len;
21521 		/*
21522 		 * We allow a limit on sending with hptsi.
21523 		 */
21524 		if (len > max_val) {
21525 			mark = 7;
21526 			len = max_val;
21527 		}
21528 #ifdef INET6
21529 		if (MHLEN < hdrlen + max_linkhdr)
21530 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21531 		else
21532 #endif
21533 			m = m_gethdr(M_NOWAIT, MT_DATA);
21534 
21535 		if (m == NULL) {
21536 			SOCK_SENDBUF_UNLOCK(so);
21537 			error = ENOBUFS;
21538 			sack_rxmit = 0;
21539 			goto out;
21540 		}
21541 		m->m_data += max_linkhdr;
21542 		m->m_len = hdrlen;
21543 
21544 		/*
21545 		 * Start the m_copy functions from the closest mbuf to the
21546 		 * sb_offset in the socket buffer chain.
21547 		 */
21548 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
21549 		s_mb = mb;
21550 		s_moff = moff;
21551 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21552 			m_copydata(mb, moff, (int)len,
21553 				   mtod(m, caddr_t)+hdrlen);
21554 			/*
21555 			 * If we are not retransmitting advance the
21556 			 * sndptr to help remember the next place in
21557 			 * the sb.
21558 			 */
21559 			if (rsm == NULL)
21560 				sbsndptr_adv(sb, mb, len);
21561 			m->m_len += len;
21562 		} else {
21563 			struct sockbuf *msb;
21564 
21565 			/*
21566 			 * If we are not retransmitting pass in msb so
21567 			 * the socket buffer can be advanced. Otherwise
21568 			 * set it to NULL if its a retransmission since
21569 			 * we don't want to change the sb remembered
21570 			 * location.
21571 			 */
21572 			if (rsm == NULL)
21573 				msb = sb;
21574 			else
21575 				msb = NULL;
21576 			m->m_next = tcp_m_copym(
21577 				mb, moff, &len,
21578 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21579 				((rsm == NULL) ? hw_tls : 0)
21580 #ifdef NETFLIX_COPY_ARGS
21581 				, &s_mb, &s_moff
21582 #endif
21583 				);
21584 			if (len <= (tp->t_maxseg - optlen)) {
21585 				/*
21586 				 * Must have ran out of mbufs for the copy
21587 				 * shorten it to no longer need tso. Lets
21588 				 * not put on sendalot since we are low on
21589 				 * mbufs.
21590 				 */
21591 				tso = 0;
21592 			}
21593 			if (m->m_next == NULL) {
21594 				SOCK_SENDBUF_UNLOCK(so);
21595 				(void)m_free(m);
21596 				error = ENOBUFS;
21597 				sack_rxmit = 0;
21598 				goto out;
21599 			}
21600 		}
21601 		if (sack_rxmit) {
21602 			if (rsm && (rsm->r_flags & RACK_TLP)) {
21603 				/*
21604 				 * TLP should not count in retran count, but
21605 				 * in its own bin
21606 				 */
21607 				counter_u64_add(rack_tlp_retran, 1);
21608 				counter_u64_add(rack_tlp_retran_bytes, len);
21609 			} else {
21610 				tp->t_sndrexmitpack++;
21611 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21612 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21613 			}
21614 #ifdef STATS
21615 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21616 						 len);
21617 #endif
21618 		} else {
21619 			KMOD_TCPSTAT_INC(tcps_sndpack);
21620 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21621 #ifdef STATS
21622 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21623 						 len);
21624 #endif
21625 		}
21626 		/*
21627 		 * If we're sending everything we've got, set PUSH. (This
21628 		 * will keep happy those implementations which only give
21629 		 * data to the user when a buffer fills or a PUSH comes in.)
21630 		 */
21631 		if (sb_offset + len == sbused(sb) &&
21632 		    sbused(sb) &&
21633 		    !(flags & TH_SYN)) {
21634 			flags |= TH_PUSH;
21635 			add_flag |= RACK_HAD_PUSH;
21636 		}
21637 
21638 		SOCK_SENDBUF_UNLOCK(so);
21639 	} else {
21640 		SOCK_SENDBUF_UNLOCK(so);
21641 		if (tp->t_flags & TF_ACKNOW)
21642 			KMOD_TCPSTAT_INC(tcps_sndacks);
21643 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
21644 			KMOD_TCPSTAT_INC(tcps_sndctrl);
21645 		else
21646 			KMOD_TCPSTAT_INC(tcps_sndwinup);
21647 
21648 		m = m_gethdr(M_NOWAIT, MT_DATA);
21649 		if (m == NULL) {
21650 			error = ENOBUFS;
21651 			sack_rxmit = 0;
21652 			goto out;
21653 		}
21654 #ifdef INET6
21655 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21656 		    MHLEN >= hdrlen) {
21657 			M_ALIGN(m, hdrlen);
21658 		} else
21659 #endif
21660 			m->m_data += max_linkhdr;
21661 		m->m_len = hdrlen;
21662 	}
21663 	SOCK_SENDBUF_UNLOCK_ASSERT(so);
21664 	m->m_pkthdr.rcvif = (struct ifnet *)0;
21665 #ifdef MAC
21666 	mac_inpcb_create_mbuf(inp, m);
21667 #endif
21668 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
21669 #ifdef INET6
21670 		if (isipv6)
21671 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21672 		else
21673 #endif				/* INET6 */
21674 #ifdef INET
21675 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21676 #endif
21677 		th = rack->r_ctl.fsb.th;
21678 		udp = rack->r_ctl.fsb.udp;
21679 		if (udp) {
21680 #ifdef INET6
21681 			if (isipv6)
21682 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21683 			else
21684 #endif				/* INET6 */
21685 				ulen = hdrlen + len - sizeof(struct ip);
21686 			udp->uh_ulen = htons(ulen);
21687 		}
21688 	} else {
21689 #ifdef INET6
21690 		if (isipv6) {
21691 			ip6 = mtod(m, struct ip6_hdr *);
21692 			if (tp->t_port) {
21693 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21694 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21695 				udp->uh_dport = tp->t_port;
21696 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21697 				udp->uh_ulen = htons(ulen);
21698 				th = (struct tcphdr *)(udp + 1);
21699 			} else
21700 				th = (struct tcphdr *)(ip6 + 1);
21701 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
21702 		} else
21703 #endif				/* INET6 */
21704 		{
21705 #ifdef INET
21706 			ip = mtod(m, struct ip *);
21707 			if (tp->t_port) {
21708 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21709 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21710 				udp->uh_dport = tp->t_port;
21711 				ulen = hdrlen + len - sizeof(struct ip);
21712 				udp->uh_ulen = htons(ulen);
21713 				th = (struct tcphdr *)(udp + 1);
21714 			} else
21715 				th = (struct tcphdr *)(ip + 1);
21716 			tcpip_fillheaders(inp, tp->t_port, ip, th);
21717 #endif
21718 		}
21719 	}
21720 	/*
21721 	 * If we are starting a connection, send ECN setup SYN packet. If we
21722 	 * are on a retransmit, we may resend those bits a number of times
21723 	 * as per RFC 3168.
21724 	 */
21725 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21726 		flags |= tcp_ecn_output_syn_sent(tp);
21727 	}
21728 	/* Also handle parallel SYN for ECN */
21729 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
21730 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21731 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21732 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21733 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
21734 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21735 #ifdef INET6
21736 		if (isipv6) {
21737 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21738 			ip6->ip6_flow |= htonl(ect << 20);
21739 		}
21740 		else
21741 #endif
21742 		{
21743 #ifdef INET
21744 			ip->ip_tos &= ~IPTOS_ECN_MASK;
21745 			ip->ip_tos |= ect;
21746 #endif
21747 		}
21748 	}
21749 	th->th_seq = htonl(rack_seq);
21750 	th->th_ack = htonl(tp->rcv_nxt);
21751 	tcp_set_flags(th, flags);
21752 	/*
21753 	 * Calculate receive window.  Don't shrink window, but avoid silly
21754 	 * window syndrome.
21755 	 * If a RST segment is sent, advertise a window of zero.
21756 	 */
21757 	if (flags & TH_RST) {
21758 		recwin = 0;
21759 	} else {
21760 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21761 		    recwin < (long)segsiz) {
21762 			recwin = 0;
21763 		}
21764 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21765 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21766 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21767 	}
21768 
21769 	/*
21770 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21771 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
21772 	 * handled in syncache.
21773 	 */
21774 	if (flags & TH_SYN)
21775 		th->th_win = htons((u_short)
21776 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21777 	else {
21778 		/* Avoid shrinking window with window scaling. */
21779 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
21780 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21781 	}
21782 	/*
21783 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21784 	 * window.  This may cause the remote transmitter to stall.  This
21785 	 * flag tells soreceive() to disable delayed acknowledgements when
21786 	 * draining the buffer.  This can occur if the receiver is
21787 	 * attempting to read more data than can be buffered prior to
21788 	 * transmitting on the connection.
21789 	 */
21790 	if (th->th_win == 0) {
21791 		tp->t_sndzerowin++;
21792 		tp->t_flags |= TF_RXWIN0SENT;
21793 	} else
21794 		tp->t_flags &= ~TF_RXWIN0SENT;
21795 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
21796 	/* Now are we using fsb?, if so copy the template data to the mbuf */
21797 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21798 		uint8_t *cpto;
21799 
21800 		cpto = mtod(m, uint8_t *);
21801 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21802 		/*
21803 		 * We have just copied in:
21804 		 * IP/IP6
21805 		 * <optional udphdr>
21806 		 * tcphdr (no options)
21807 		 *
21808 		 * We need to grab the correct pointers into the mbuf
21809 		 * for both the tcp header, and possibly the udp header (if tunneling).
21810 		 * We do this by using the offset in the copy buffer and adding it
21811 		 * to the mbuf base pointer (cpto).
21812 		 */
21813 #ifdef INET6
21814 		if (isipv6)
21815 			ip6 = mtod(m, struct ip6_hdr *);
21816 		else
21817 #endif				/* INET6 */
21818 #ifdef INET
21819 			ip = mtod(m, struct ip *);
21820 #endif
21821 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21822 		/* If we have a udp header lets set it into the mbuf as well */
21823 		if (udp)
21824 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21825 	}
21826 	if (optlen) {
21827 		bcopy(opt, th + 1, optlen);
21828 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21829 	}
21830 	/*
21831 	 * Put TCP length in extended header, and then checksum extended
21832 	 * header and data.
21833 	 */
21834 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
21835 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21836 	if (to.to_flags & TOF_SIGNATURE) {
21837 		/*
21838 		 * Calculate MD5 signature and put it into the place
21839 		 * determined before.
21840 		 * NOTE: since TCP options buffer doesn't point into
21841 		 * mbuf's data, calculate offset and use it.
21842 		 */
21843 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21844 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21845 			/*
21846 			 * Do not send segment if the calculation of MD5
21847 			 * digest has failed.
21848 			 */
21849 			goto out;
21850 		}
21851 	}
21852 #endif
21853 #ifdef INET6
21854 	if (isipv6) {
21855 		/*
21856 		 * ip6_plen is not need to be filled now, and will be filled
21857 		 * in ip6_output.
21858 		 */
21859 		if (tp->t_port) {
21860 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21861 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21862 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21863 			th->th_sum = htons(0);
21864 			UDPSTAT_INC(udps_opackets);
21865 		} else {
21866 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21867 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21868 			th->th_sum = in6_cksum_pseudo(ip6,
21869 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21870 						      0);
21871 		}
21872 	}
21873 #endif
21874 #if defined(INET6) && defined(INET)
21875 	else
21876 #endif
21877 #ifdef INET
21878 	{
21879 		if (tp->t_port) {
21880 			m->m_pkthdr.csum_flags = CSUM_UDP;
21881 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21882 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21883 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21884 			th->th_sum = htons(0);
21885 			UDPSTAT_INC(udps_opackets);
21886 		} else {
21887 			m->m_pkthdr.csum_flags = CSUM_TCP;
21888 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21889 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
21890 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21891 									IPPROTO_TCP + len + optlen));
21892 		}
21893 		/* IP version must be set here for ipv4/ipv6 checking later */
21894 		KASSERT(ip->ip_v == IPVERSION,
21895 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
21896 	}
21897 #endif
21898 	/*
21899 	 * Enable TSO and specify the size of the segments. The TCP pseudo
21900 	 * header checksum is always provided. XXX: Fixme: This is currently
21901 	 * not the case for IPv6.
21902 	 */
21903 	if (tso) {
21904 		/*
21905 		 * Here we must use t_maxseg and the optlen since
21906 		 * the optlen may include SACK's (or DSACK).
21907 		 */
21908 		KASSERT(len > tp->t_maxseg - optlen,
21909 			("%s: len <= tso_segsz", __func__));
21910 		m->m_pkthdr.csum_flags |= CSUM_TSO;
21911 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21912 	}
21913 	KASSERT(len + hdrlen == m_length(m, NULL),
21914 		("%s: mbuf chain different than expected: %d + %u != %u",
21915 		 __func__, len, hdrlen, m_length(m, NULL)));
21916 
21917 #ifdef TCP_HHOOK
21918 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21919 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
21920 #endif
21921 	if ((rack->r_ctl.crte != NULL) &&
21922 	    (rack->rc_hw_nobuf == 0) &&
21923 	    tcp_bblogging_on(tp)) {
21924 		rack_log_queue_level(tp, rack, len, &tv, cts);
21925 	}
21926 	/* We're getting ready to send; log now. */
21927 	if (tcp_bblogging_on(rack->rc_tp)) {
21928 		union tcp_log_stackspecific log;
21929 
21930 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
21931 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
21932 		if (rack->rack_no_prr)
21933 			log.u_bbr.flex1 = 0;
21934 		else
21935 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21936 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
21937 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
21938 		log.u_bbr.flex4 = orig_len;
21939 		/* Save off the early/late values */
21940 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
21941 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
21942 		log.u_bbr.bw_inuse = rack_get_bw(rack);
21943 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
21944 		log.u_bbr.flex8 = 0;
21945 		if (rsm) {
21946 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
21947 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
21948 				counter_u64_add(rack_collapsed_win_rxt, 1);
21949 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
21950 			}
21951 			if (doing_tlp)
21952 				log.u_bbr.flex8 = 2;
21953 			else
21954 				log.u_bbr.flex8 = 1;
21955 		} else {
21956 			if (doing_tlp)
21957 				log.u_bbr.flex8 = 3;
21958 		}
21959 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
21960 		log.u_bbr.flex7 = mark;
21961 		log.u_bbr.flex7 <<= 8;
21962 		log.u_bbr.flex7 |= pass;
21963 		log.u_bbr.pkts_out = tp->t_maxseg;
21964 		log.u_bbr.timeStamp = cts;
21965 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
21966 		if (rsm && (rsm->r_rtr_cnt > 0)) {
21967 			/*
21968 			 * When we have a retransmit we want to log the
21969 			 * burst at send and flight at send from before.
21970 			 */
21971 			log.u_bbr.flex5 = rsm->r_fas;
21972 			log.u_bbr.bbr_substate = rsm->r_bas;
21973 		} else {
21974 			/*
21975 			 * New transmits we log in flex5 the inflight again as
21976 			 * well as the number of segments in our send in the
21977 			 * substate field.
21978 			 */
21979 			log.u_bbr.flex5 = log.u_bbr.inflight;
21980 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
21981 		}
21982 		log.u_bbr.lt_epoch = cwnd_to_use;
21983 		log.u_bbr.delivered = sendalot;
21984 		log.u_bbr.rttProp = (uintptr_t)rsm;
21985 		log.u_bbr.pkt_epoch = __LINE__;
21986 		if (rsm) {
21987 			log.u_bbr.delRate = rsm->r_flags;
21988 			log.u_bbr.delRate <<= 31;
21989 			log.u_bbr.delRate |= rack->r_must_retran;
21990 			log.u_bbr.delRate <<= 1;
21991 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21992 		} else {
21993 			log.u_bbr.delRate = rack->r_must_retran;
21994 			log.u_bbr.delRate <<= 1;
21995 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21996 		}
21997 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
21998 				    len, &log, false, NULL, __func__, __LINE__, &tv);
21999 	} else
22000 		lgb = NULL;
22001 
22002 	/*
22003 	 * Fill in IP length and desired time to live and send to IP level.
22004 	 * There should be a better way to handle ttl and tos; we could keep
22005 	 * them in the template, but need a way to checksum without them.
22006 	 */
22007 	/*
22008 	 * m->m_pkthdr.len should have been set before cksum calcuration,
22009 	 * because in6_cksum() need it.
22010 	 */
22011 #ifdef INET6
22012 	if (isipv6) {
22013 		/*
22014 		 * we separately set hoplimit for every segment, since the
22015 		 * user might want to change the value via setsockopt. Also,
22016 		 * desired default hop limit might be changed via Neighbor
22017 		 * Discovery.
22018 		 */
22019 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
22020 
22021 		/*
22022 		 * Set the packet size here for the benefit of DTrace
22023 		 * probes. ip6_output() will set it properly; it's supposed
22024 		 * to include the option header lengths as well.
22025 		 */
22026 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
22027 
22028 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
22029 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22030 		else
22031 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22032 
22033 		if (tp->t_state == TCPS_SYN_SENT)
22034 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
22035 
22036 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
22037 		/* TODO: IPv6 IP6TOS_ECT bit on */
22038 		error = ip6_output(m,
22039 				   inp->in6p_outputopts,
22040 				   &inp->inp_route6,
22041 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
22042 				   NULL, NULL, inp);
22043 
22044 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
22045 			mtu = inp->inp_route6.ro_nh->nh_mtu;
22046 	}
22047 #endif				/* INET6 */
22048 #if defined(INET) && defined(INET6)
22049 	else
22050 #endif
22051 #ifdef INET
22052 	{
22053 		ip->ip_len = htons(m->m_pkthdr.len);
22054 #ifdef INET6
22055 		if (inp->inp_vflag & INP_IPV6PROTO)
22056 			ip->ip_ttl = in6_selecthlim(inp, NULL);
22057 #endif				/* INET6 */
22058 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
22059 		/*
22060 		 * If we do path MTU discovery, then we set DF on every
22061 		 * packet. This might not be the best thing to do according
22062 		 * to RFC3390 Section 2. However the tcp hostcache migitates
22063 		 * the problem so it affects only the first tcp connection
22064 		 * with a host.
22065 		 *
22066 		 * NB: Don't set DF on small MTU/MSS to have a safe
22067 		 * fallback.
22068 		 */
22069 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
22070 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22071 			if (tp->t_port == 0 || len < V_tcp_minmss) {
22072 				ip->ip_off |= htons(IP_DF);
22073 			}
22074 		} else {
22075 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22076 		}
22077 
22078 		if (tp->t_state == TCPS_SYN_SENT)
22079 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
22080 
22081 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
22082 
22083 		error = ip_output(m,
22084 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22085 				  inp->inp_options,
22086 #else
22087 				  NULL,
22088 #endif
22089 				  &inp->inp_route,
22090 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
22091 				  inp);
22092 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
22093 			mtu = inp->inp_route.ro_nh->nh_mtu;
22094 	}
22095 #endif				/* INET */
22096 	if (lgb) {
22097 		lgb->tlb_errno = error;
22098 		lgb = NULL;
22099 	}
22100 
22101 out:
22102 	/*
22103 	 * In transmit state, time the transmission and arrange for the
22104 	 * retransmit.  In persist state, just set snd_max.
22105 	 */
22106 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
22107 			rack_to_usec_ts(&tv),
22108 			rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
22109 	if (error == 0) {
22110 		if (add_flag & RACK_IS_PCM) {
22111 			/* We just launched a PCM */
22112 			/* rrs here log */
22113 			rack->pcm_in_progress = 1;
22114 			rack->pcm_needed = 0;
22115 			rack_log_pcm(rack, 7, len, rack->r_ctl.pcm_max_seg,  add_flag);
22116 		}
22117 		if (rsm == NULL) {
22118 			if (rack->lt_bw_up == 0) {
22119 				rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
22120 				rack->r_ctl.lt_seq = tp->snd_una;
22121 				rack->lt_bw_up = 1;
22122 			} else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
22123 				/*
22124 				 * Need to record what we have since we are
22125 				 * approaching seq wrap.
22126 				 */
22127 				uint64_t tmark;
22128 
22129 				rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
22130 				rack->r_ctl.lt_seq = tp->snd_una;
22131 				tmark = tcp_get_u64_usecs(&tv);
22132 				if (tmark > rack->r_ctl.lt_timemark) {
22133 					rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
22134 					rack->r_ctl.lt_timemark = tmark;
22135 				}
22136 			}
22137 		}
22138 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
22139 		counter_u64_add(rack_total_bytes, len);
22140 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
22141 		if (rsm && doing_tlp) {
22142 			rack->rc_last_sent_tlp_past_cumack = 0;
22143 			rack->rc_last_sent_tlp_seq_valid = 1;
22144 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
22145 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
22146 		}
22147 		if (rack->rc_hw_nobuf) {
22148 			rack->rc_hw_nobuf = 0;
22149 			rack->r_ctl.rc_agg_delayed = 0;
22150 			rack->r_early = 0;
22151 			rack->r_late = 0;
22152 			rack->r_ctl.rc_agg_early = 0;
22153 		}
22154 		if (rsm && (doing_tlp == 0)) {
22155 			/* Set we retransmitted */
22156 			rack->rc_gp_saw_rec = 1;
22157 		} else {
22158 			if (cwnd_to_use > tp->snd_ssthresh) {
22159 				/* Set we sent in CA */
22160 				rack->rc_gp_saw_ca = 1;
22161 			} else {
22162 				/* Set we sent in SS */
22163 				rack->rc_gp_saw_ss = 1;
22164 			}
22165 		}
22166 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22167 		    (tp->t_flags & TF_SACK_PERMIT) &&
22168 		    tp->rcv_numsacks > 0)
22169 			tcp_clean_dsack_blocks(tp);
22170 		tot_len_this_send += len;
22171 		if (len == 0) {
22172 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
22173 		} else {
22174 			int idx;
22175 
22176 			idx = (len / segsiz) + 3;
22177 			if (idx >= TCP_MSS_ACCT_ATIMER)
22178 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
22179 			else
22180 				counter_u64_add(rack_out_size[idx], 1);
22181 		}
22182 	}
22183 	if ((rack->rack_no_prr == 0) &&
22184 	    sub_from_prr &&
22185 	    (error == 0)) {
22186 		if (rack->r_ctl.rc_prr_sndcnt >= len)
22187 			rack->r_ctl.rc_prr_sndcnt -= len;
22188 		else
22189 			rack->r_ctl.rc_prr_sndcnt = 0;
22190 	}
22191 	sub_from_prr = 0;
22192 	if (doing_tlp) {
22193 		/* Make sure the TLP is added */
22194 		add_flag |= RACK_TLP;
22195 	} else if (rsm) {
22196 		/* If its a resend without TLP then it must not have the flag */
22197 		rsm->r_flags &= ~RACK_TLP;
22198 	}
22199 
22200 
22201 	if ((error == 0) &&
22202 	    (len > 0) &&
22203 	    (tp->snd_una == tp->snd_max))
22204 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
22205 
22206 	{
22207 		/*
22208 		 * This block is not associated with the above error == 0 test.
22209 		 * It is used to advance snd_max if we have a new transmit.
22210 		 */
22211 		tcp_seq startseq = tp->snd_max;
22212 
22213 
22214 		if (rsm && (doing_tlp == 0))
22215 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
22216 		if (error)
22217 			/* We don't log or do anything with errors */
22218 			goto nomore;
22219 		if (doing_tlp == 0) {
22220 			if (rsm == NULL) {
22221 				/*
22222 				 * Not a retransmission of some
22223 				 * sort, new data is going out so
22224 				 * clear our TLP count and flag.
22225 				 */
22226 				rack->rc_tlp_in_progress = 0;
22227 				rack->r_ctl.rc_tlp_cnt_out = 0;
22228 			}
22229 		} else {
22230 			/*
22231 			 * We have just sent a TLP, mark that it is true
22232 			 * and make sure our in progress is set so we
22233 			 * continue to check the count.
22234 			 */
22235 			rack->rc_tlp_in_progress = 1;
22236 			rack->r_ctl.rc_tlp_cnt_out++;
22237 		}
22238 		/*
22239 		 * If we are retransmitting we are done, snd_max
22240 		 * does not get updated.
22241 		 */
22242 		if (sack_rxmit)
22243 			goto nomore;
22244 		if ((tp->snd_una == tp->snd_max) && (len > 0)) {
22245 			/*
22246 			 * Update the time we just added data since
22247 			 * nothing was outstanding.
22248 			 */
22249 			rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22250 			tp->t_acktime = ticks;
22251 		}
22252 		/*
22253 		 * Now for special SYN/FIN handling.
22254 		 */
22255 		if (flags & (TH_SYN | TH_FIN)) {
22256 			if ((flags & TH_SYN) &&
22257 			    ((tp->t_flags & TF_SENTSYN) == 0)) {
22258 				tp->snd_max++;
22259 				tp->t_flags |= TF_SENTSYN;
22260 			}
22261 			if ((flags & TH_FIN) &&
22262 			    ((tp->t_flags & TF_SENTFIN) == 0)) {
22263 				tp->snd_max++;
22264 				tp->t_flags |= TF_SENTFIN;
22265 			}
22266 		}
22267 		tp->snd_max += len;
22268 		if (rack->rc_new_rnd_needed) {
22269 			rack_new_round_starts(tp, rack, tp->snd_max);
22270 		}
22271 		/*
22272 		 * Time this transmission if not a retransmission and
22273 		 * not currently timing anything.
22274 		 * This is only relevant in case of switching back to
22275 		 * the base stack.
22276 		 */
22277 		if (tp->t_rtttime == 0) {
22278 			tp->t_rtttime = ticks;
22279 			tp->t_rtseq = startseq;
22280 			KMOD_TCPSTAT_INC(tcps_segstimed);
22281 		}
22282 		if (len &&
22283 		    ((tp->t_flags & TF_GPUTINPROG) == 0))
22284 			rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22285 		/*
22286 		 * If we are doing FO we need to update the mbuf position and subtract
22287 		 * this happens when the peer sends us duplicate information and
22288 		 * we thus want to send a DSACK.
22289 		 *
22290 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22291 		 * turned off? If not then we are going to echo multiple DSACK blocks
22292 		 * out (with the TSO), which we should not be doing.
22293 		 */
22294 		if (rack->r_fast_output && len) {
22295 			if (rack->r_ctl.fsb.left_to_send > len)
22296 				rack->r_ctl.fsb.left_to_send -= len;
22297 			else
22298 				rack->r_ctl.fsb.left_to_send = 0;
22299 			if (rack->r_ctl.fsb.left_to_send < segsiz)
22300 				rack->r_fast_output = 0;
22301 			if (rack->r_fast_output) {
22302 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22303 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22304 				rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22305 			}
22306 		}
22307 		if (rack_pcm_blast == 0) {
22308 			if ((orig_len > len) &&
22309 			    (add_flag & RACK_IS_PCM) &&
22310 			    (len < pace_max_seg) &&
22311 			    ((pace_max_seg - len) > segsiz)) {
22312 				/*
22313 				 * We are doing a PCM measurement and we did
22314 				 * not get enough data in the TSO to meet the
22315 				 * burst requirement.
22316 				 */
22317 				uint32_t n_len;
22318 
22319 				n_len = (orig_len - len);
22320 				orig_len -= len;
22321 				pace_max_seg -= len;
22322 				len = n_len;
22323 				sb_offset = tp->snd_max - tp->snd_una;
22324 				/* Re-lock for the next spin */
22325 				SOCK_SENDBUF_LOCK(so);
22326 				goto send;
22327 			}
22328 		} else {
22329 			if ((orig_len > len) &&
22330 			    (add_flag & RACK_IS_PCM) &&
22331 			    ((orig_len - len) > segsiz)) {
22332 				/*
22333 				 * We are doing a PCM measurement and we did
22334 				 * not get enough data in the TSO to meet the
22335 				 * burst requirement.
22336 				 */
22337 				uint32_t n_len;
22338 
22339 				n_len = (orig_len - len);
22340 				orig_len -= len;
22341 				len = n_len;
22342 				sb_offset = tp->snd_max - tp->snd_una;
22343 				/* Re-lock for the next spin */
22344 				SOCK_SENDBUF_LOCK(so);
22345 				goto send;
22346 			}
22347 		}
22348 	}
22349 nomore:
22350 	if (error) {
22351 		rack->r_ctl.rc_agg_delayed = 0;
22352 		rack->r_early = 0;
22353 		rack->r_late = 0;
22354 		rack->r_ctl.rc_agg_early = 0;
22355 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
22356 		/*
22357 		 * Failures do not advance the seq counter above. For the
22358 		 * case of ENOBUFS we will fall out and retry in 1ms with
22359 		 * the hpts. Everything else will just have to retransmit
22360 		 * with the timer.
22361 		 *
22362 		 * In any case, we do not want to loop around for another
22363 		 * send without a good reason.
22364 		 */
22365 		sendalot = 0;
22366 		switch (error) {
22367 		case EPERM:
22368 		case EACCES:
22369 			tp->t_softerror = error;
22370 #ifdef TCP_ACCOUNTING
22371 			crtsc = get_cyclecount();
22372 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22373 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22374 			}
22375 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22376 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22377 			}
22378 			sched_unpin();
22379 #endif
22380 			return (error);
22381 		case ENOBUFS:
22382 			/*
22383 			 * Pace us right away to retry in a some
22384 			 * time
22385 			 */
22386 			if (rack->r_ctl.crte != NULL) {
22387 				tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22388 				if (tcp_bblogging_on(rack->rc_tp))
22389 					rack_log_queue_level(tp, rack, len, &tv, cts);
22390 			} else
22391 				tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22392 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22393 			if (rack->rc_enobuf < 0x7f)
22394 				rack->rc_enobuf++;
22395 			if (slot < (10 * HPTS_USEC_IN_MSEC))
22396 				slot = 10 * HPTS_USEC_IN_MSEC;
22397 			if (rack->r_ctl.crte != NULL) {
22398 				counter_u64_add(rack_saw_enobuf_hw, 1);
22399 				tcp_rl_log_enobuf(rack->r_ctl.crte);
22400 			}
22401 			counter_u64_add(rack_saw_enobuf, 1);
22402 			goto enobufs;
22403 		case EMSGSIZE:
22404 			/*
22405 			 * For some reason the interface we used initially
22406 			 * to send segments changed to another or lowered
22407 			 * its MTU. If TSO was active we either got an
22408 			 * interface without TSO capabilits or TSO was
22409 			 * turned off. If we obtained mtu from ip_output()
22410 			 * then update it and try again.
22411 			 */
22412 			if (tso)
22413 				tp->t_flags &= ~TF_TSO;
22414 			if (mtu != 0) {
22415 				int saved_mtu;
22416 
22417 				saved_mtu = tp->t_maxseg;
22418 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
22419 				if (saved_mtu > tp->t_maxseg) {
22420 					goto again;
22421 				}
22422 			}
22423 			slot = 10 * HPTS_USEC_IN_MSEC;
22424 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22425 #ifdef TCP_ACCOUNTING
22426 			crtsc = get_cyclecount();
22427 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22428 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22429 			}
22430 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22431 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22432 			}
22433 			sched_unpin();
22434 #endif
22435 			return (error);
22436 		case ENETUNREACH:
22437 			counter_u64_add(rack_saw_enetunreach, 1);
22438 			/* FALLTHROUGH */
22439 		case EHOSTDOWN:
22440 		case EHOSTUNREACH:
22441 		case ENETDOWN:
22442 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
22443 				tp->t_softerror = error;
22444 				error = 0;
22445 			}
22446 			/* FALLTHROUGH */
22447 		default:
22448 			slot = 10 * HPTS_USEC_IN_MSEC;
22449 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22450 #ifdef TCP_ACCOUNTING
22451 			crtsc = get_cyclecount();
22452 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22453 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22454 			}
22455 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22456 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22457 			}
22458 			sched_unpin();
22459 #endif
22460 			return (error);
22461 		}
22462 	} else {
22463 		rack->rc_enobuf = 0;
22464 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22465 			rack->r_ctl.retran_during_recovery += len;
22466 	}
22467 	KMOD_TCPSTAT_INC(tcps_sndtotal);
22468 
22469 	/*
22470 	 * Data sent (as far as we can tell). If this advertises a larger
22471 	 * window than any other segment, then remember the size of the
22472 	 * advertised window. Any pending ACK has now been sent.
22473 	 */
22474 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22475 		tp->rcv_adv = tp->rcv_nxt + recwin;
22476 
22477 	tp->last_ack_sent = tp->rcv_nxt;
22478 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22479 enobufs:
22480 	if (sendalot) {
22481 		/* Do we need to turn off sendalot? */
22482 		if (pace_max_seg &&
22483 		    (tot_len_this_send >= pace_max_seg)) {
22484 			/* We hit our max. */
22485 			sendalot = 0;
22486 		}
22487 	}
22488 	if ((error == 0) && (flags & TH_FIN))
22489 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22490 	if (flags & TH_RST) {
22491 		/*
22492 		 * We don't send again after sending a RST.
22493 		 */
22494 		slot = 0;
22495 		sendalot = 0;
22496 		if (error == 0)
22497 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22498 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22499 		/*
22500 		 * Get our pacing rate, if an error
22501 		 * occurred in sending (ENOBUF) we would
22502 		 * hit the else if with slot preset. Other
22503 		 * errors return.
22504 		 */
22505 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz, __LINE__);
22506 	}
22507 	/* We have sent clear the flag */
22508 	rack->r_ent_rec_ns = 0;
22509 	if (rack->r_must_retran) {
22510 		if (rsm) {
22511 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22512 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22513 				/*
22514 				 * We have retransmitted all.
22515 				 */
22516 				rack->r_must_retran = 0;
22517 				rack->r_ctl.rc_out_at_rto = 0;
22518 			}
22519 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22520 			/*
22521 			 * Sending new data will also kill
22522 			 * the loop.
22523 			 */
22524 			rack->r_must_retran = 0;
22525 			rack->r_ctl.rc_out_at_rto = 0;
22526 		}
22527 	}
22528 	rack->r_ctl.fsb.recwin = recwin;
22529 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22530 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22531 		/*
22532 		 * We hit an RTO and now have past snd_max at the RTO
22533 		 * clear all the WAS flags.
22534 		 */
22535 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22536 	}
22537 	if (slot) {
22538 		/* set the rack tcb into the slot N */
22539 		if ((error == 0) &&
22540 		    rack_use_rfo &&
22541 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22542 		    (rsm == NULL) &&
22543 		    (ipoptlen == 0) &&
22544 		    rack->r_fsb_inited &&
22545 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22546 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22547 		    (rack->r_must_retran == 0) &&
22548 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22549 		    (len > 0) && (orig_len > 0) &&
22550 		    (orig_len > len) &&
22551 		    ((orig_len - len) >= segsiz) &&
22552 		    ((optlen == 0) ||
22553 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22554 			/* We can send at least one more MSS using our fsb */
22555 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22556 					       segsiz, pace_max_seg, hw_tls, flags);
22557 		} else
22558 			rack->r_fast_output = 0;
22559 		rack_log_fsb(rack, tp, so, flags,
22560 			     ipoptlen, orig_len, len, error,
22561 			     (rsm == NULL), optlen, __LINE__, 2);
22562 	} else if (sendalot) {
22563 		int ret;
22564 
22565 		sack_rxmit = 0;
22566 		if ((error == 0) &&
22567 		    rack_use_rfo &&
22568 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22569 		    (rsm == NULL) &&
22570 		    (ipoptlen == 0) &&
22571 		    (rack->r_must_retran == 0) &&
22572 		    rack->r_fsb_inited &&
22573 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22574 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22575 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22576 		    (len > 0) && (orig_len > 0) &&
22577 		    (orig_len > len) &&
22578 		    ((orig_len - len) >= segsiz) &&
22579 		    ((optlen == 0) ||
22580 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22581 			/* we can use fast_output for more */
22582 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22583 					       segsiz, pace_max_seg, hw_tls, flags);
22584 			if (rack->r_fast_output) {
22585 				error = 0;
22586 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
22587 				if (ret >= 0)
22588 					return (ret);
22589 			        else if (error)
22590 					goto nomore;
22591 
22592 			}
22593 		}
22594 		goto again;
22595 	}
22596 skip_all_send:
22597 	/* Assure when we leave that snd_nxt will point to top */
22598 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22599 		tp->snd_nxt = tp->snd_max;
22600 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22601 #ifdef TCP_ACCOUNTING
22602 	crtsc = get_cyclecount() - ts_val;
22603 	if (tot_len_this_send) {
22604 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22605 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
22606 		}
22607 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22608 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22609 		}
22610 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22611 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22612 		}
22613 	} else {
22614 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22615 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
22616 		}
22617 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22618 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22619 		}
22620 	}
22621 	sched_unpin();
22622 #endif
22623 	if (error == ENOBUFS)
22624 		error = 0;
22625 	return (error);
22626 }
22627 
22628 static void
22629 rack_update_seg(struct tcp_rack *rack)
22630 {
22631 	uint32_t orig_val;
22632 
22633 	orig_val = rack->r_ctl.rc_pace_max_segs;
22634 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22635 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
22636 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22637 }
22638 
22639 static void
22640 rack_mtu_change(struct tcpcb *tp)
22641 {
22642 	/*
22643 	 * The MSS may have changed
22644 	 */
22645 	struct tcp_rack *rack;
22646 	struct rack_sendmap *rsm;
22647 
22648 	rack = (struct tcp_rack *)tp->t_fb_ptr;
22649 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22650 		/*
22651 		 * The MTU has changed we need to resend everything
22652 		 * since all we have sent is lost. We first fix
22653 		 * up the mtu though.
22654 		 */
22655 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
22656 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
22657 		rack_remxt_tmr(tp);
22658 		rack->r_fast_output = 0;
22659 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22660 						rack->r_ctl.rc_sacked);
22661 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22662 		rack->r_must_retran = 1;
22663 		/* Mark all inflight to needing to be rxt'd */
22664 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22665 			rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22666 		}
22667 	}
22668 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22669 	/* We don't use snd_nxt to retransmit */
22670 	tp->snd_nxt = tp->snd_max;
22671 }
22672 
22673 static int
22674 rack_set_dgp(struct tcp_rack *rack)
22675 {
22676 	if (rack->dgp_on == 1)
22677 		return(0);
22678 	if ((rack->use_fixed_rate == 1) &&
22679 	    (rack->rc_always_pace == 1)) {
22680 		/*
22681 		 * We are already pacing another
22682 		 * way.
22683 		 */
22684 		return (EBUSY);
22685 	}
22686 	if (rack->rc_always_pace == 1) {
22687 		rack_remove_pacing(rack);
22688 	}
22689 	if (tcp_incr_dgp_pacing_cnt() == 0)
22690 		return (ENOSPC);
22691 	rack->r_ctl.pacing_method |= RACK_DGP_PACING;
22692 	rack->rc_fillcw_apply_discount = 0;
22693 	rack->dgp_on = 1;
22694 	rack->rc_always_pace = 1;
22695 	rack->rc_pace_dnd = 1;
22696 	rack->use_fixed_rate = 0;
22697 	if (rack->gp_ready)
22698 		rack_set_cc_pacing(rack);
22699 	rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22700 	rack->rack_attempt_hdwr_pace = 0;
22701 	/* rxt settings */
22702 	rack->full_size_rxt = 1;
22703 	rack->shape_rxt_to_pacing_min  = 0;
22704 	/* cmpack=1 */
22705 	rack->r_use_cmp_ack = 1;
22706 	if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22707 	    rack->r_use_cmp_ack)
22708 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22709 	/* scwnd=1 */
22710 	rack->rack_enable_scwnd = 1;
22711 	/* dynamic=100 */
22712 	rack->rc_gp_dyn_mul = 1;
22713 	/* gp_inc_ca */
22714 	rack->r_ctl.rack_per_of_gp_ca = 100;
22715 	/* rrr_conf=3 */
22716 	rack->r_rr_config = 3;
22717 	/* npush=2 */
22718 	rack->r_ctl.rc_no_push_at_mrtt = 2;
22719 	/* fillcw=1 */
22720 	rack->rc_pace_to_cwnd = 1;
22721 	rack->rc_pace_fill_if_rttin_range = 0;
22722 	rack->rtt_limit_mul = 0;
22723 	/* noprr=1 */
22724 	rack->rack_no_prr = 1;
22725 	/* lscwnd=1 */
22726 	rack->r_limit_scw = 1;
22727 	/* gp_inc_rec */
22728 	rack->r_ctl.rack_per_of_gp_rec = 90;
22729 	return (0);
22730 }
22731 
22732 static int
22733 rack_set_profile(struct tcp_rack *rack, int prof)
22734 {
22735 	int err = EINVAL;
22736 	if (prof == 1) {
22737 		/*
22738 		 * Profile 1 is "standard" DGP. It ignores
22739 		 * client buffer level.
22740 		 */
22741 		err = rack_set_dgp(rack);
22742 		if (err)
22743 			return (err);
22744 	} else if (prof == 6) {
22745 		err = rack_set_dgp(rack);
22746 		if (err)
22747 			return (err);
22748 		/*
22749 		 * Profile 6 tweaks DGP so that it will apply to
22750 		 * fill-cw the same settings that profile5 does
22751 		 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
22752 		 */
22753 		rack->rc_fillcw_apply_discount = 1;
22754 	} else if (prof == 0) {
22755 		/* This changes things back to the default settings */
22756 		if (rack->rc_always_pace == 1) {
22757 			rack_remove_pacing(rack);
22758 		} else {
22759 			/* Make sure any stray flags are off */
22760 			rack->dgp_on = 0;
22761 			rack->rc_hybrid_mode = 0;
22762 			rack->use_fixed_rate = 0;
22763 		}
22764 		err = 0;
22765 		if (rack_fill_cw_state)
22766 			rack->rc_pace_to_cwnd = 1;
22767 		else
22768 			rack->rc_pace_to_cwnd = 0;
22769 
22770 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22771 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
22772 			rack->rc_always_pace = 1;
22773 			if (rack->rack_hibeta)
22774 				rack_set_cc_pacing(rack);
22775 		} else
22776 			rack->rc_always_pace = 0;
22777 		if (rack_dsack_std_based & 0x1) {
22778 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22779 			rack->rc_rack_tmr_std_based = 1;
22780 		}
22781 		if (rack_dsack_std_based & 0x2) {
22782 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
22783 			rack->rc_rack_use_dsack = 1;
22784 		}
22785 		if (rack_use_cmp_acks)
22786 			rack->r_use_cmp_ack = 1;
22787 		else
22788 			rack->r_use_cmp_ack = 0;
22789 		if (rack_disable_prr)
22790 			rack->rack_no_prr = 1;
22791 		else
22792 			rack->rack_no_prr = 0;
22793 		if (rack_gp_no_rec_chg)
22794 			rack->rc_gp_no_rec_chg = 1;
22795 		else
22796 			rack->rc_gp_no_rec_chg = 0;
22797 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22798 			rack->r_mbuf_queue = 1;
22799 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22800 				rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22801 			rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22802 		} else {
22803 			rack->r_mbuf_queue = 0;
22804 			rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22805 		}
22806 		if (rack_enable_shared_cwnd)
22807 			rack->rack_enable_scwnd = 1;
22808 		else
22809 			rack->rack_enable_scwnd = 0;
22810 		if (rack_do_dyn_mul) {
22811 			/* When dynamic adjustment is on CA needs to start at 100% */
22812 			rack->rc_gp_dyn_mul = 1;
22813 			if (rack_do_dyn_mul >= 100)
22814 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22815 		} else {
22816 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22817 			rack->rc_gp_dyn_mul = 0;
22818 		}
22819 		rack->r_rr_config = 0;
22820 		rack->r_ctl.rc_no_push_at_mrtt = 0;
22821 		rack->rc_pace_fill_if_rttin_range = 0;
22822 		rack->rtt_limit_mul = 0;
22823 
22824 		if (rack_enable_hw_pacing)
22825 			rack->rack_hdw_pace_ena = 1;
22826 		else
22827 			rack->rack_hdw_pace_ena = 0;
22828 		if (rack_disable_prr)
22829 			rack->rack_no_prr = 1;
22830 		else
22831 			rack->rack_no_prr = 0;
22832 		if (rack_limits_scwnd)
22833 			rack->r_limit_scw  = 1;
22834 		else
22835 			rack->r_limit_scw  = 0;
22836 		rack_init_retransmit_value(rack, rack_rxt_controls);
22837 		err = 0;
22838 	}
22839 	return (err);
22840 }
22841 
22842 static int
22843 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22844 {
22845 	struct deferred_opt_list *dol;
22846 
22847 	dol = malloc(sizeof(struct deferred_opt_list),
22848 		     M_TCPDO, M_NOWAIT|M_ZERO);
22849 	if (dol == NULL) {
22850 		/*
22851 		 * No space yikes -- fail out..
22852 		 */
22853 		return (0);
22854 	}
22855 	dol->optname = sopt_name;
22856 	dol->optval = loptval;
22857 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22858 	return (1);
22859 }
22860 
22861 static int
22862 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22863 {
22864 #ifdef TCP_REQUEST_TRK
22865 	struct tcp_sendfile_track *sft;
22866 	struct timeval tv;
22867 	tcp_seq seq;
22868 	int err;
22869 
22870 	microuptime(&tv);
22871 
22872 	/* Make sure no fixed rate is on */
22873 	rack->use_fixed_rate = 0;
22874 	rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22875 	rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22876 	rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22877 	/* Now allocate or find our entry that will have these settings */
22878 	sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
22879 	if (sft == NULL) {
22880 		rack->rc_tp->tcp_hybrid_error++;
22881 		/* no space, where would it have gone? */
22882 		seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22883 		rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22884 		return (ENOSPC);
22885 	}
22886 	/* mask our internal flags */
22887 	hybrid->hybrid_flags &= TCP_HYBRID_PACING_USER_MASK;
22888 	/* The seq will be snd_una + everything in the buffer */
22889 	seq = sft->start_seq;
22890 	if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22891 		/* Disabling hybrid pacing */
22892 		if (rack->rc_hybrid_mode) {
22893 			rack_set_profile(rack, 0);
22894 			rack->rc_tp->tcp_hybrid_stop++;
22895 		}
22896 		rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22897 		return (0);
22898 	}
22899 	if (rack->dgp_on == 0) {
22900 		/*
22901 		 * If we have not yet turned DGP on, do so
22902 		 * now setting pure DGP mode, no buffer level
22903 		 * response.
22904 		 */
22905 		if ((err = rack_set_profile(rack, 1)) != 0){
22906 			/* Failed to turn pacing on */
22907 			rack->rc_tp->tcp_hybrid_error++;
22908 			rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22909 			return (err);
22910 		}
22911 	}
22912 	/*
22913 	 * Now we must switch to hybrid mode as well which also
22914 	 * means moving to regular pacing.
22915 	 */
22916 	if (rack->rc_hybrid_mode == 0) {
22917 		/* First time */
22918 		if (tcp_can_enable_pacing()) {
22919 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
22920 			rack->rc_hybrid_mode = 1;
22921 		} else {
22922 			return (ENOSPC);
22923 		}
22924 		if (rack->r_ctl.pacing_method & RACK_DGP_PACING) {
22925 			/*
22926 			 * This should be true.
22927 			 */
22928 			tcp_dec_dgp_pacing_cnt();
22929 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
22930 		}
22931 	}
22932 	/* Now set in our flags */
22933 	sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
22934 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22935 		sft->cspr = hybrid->cspr;
22936 	else
22937 		sft->cspr = 0;
22938 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22939 		sft->hint_maxseg = hybrid->hint_maxseg;
22940 	else
22941 		sft->hint_maxseg = 0;
22942 	rack->rc_tp->tcp_hybrid_start++;
22943 	rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
22944 	return (0);
22945 #else
22946 	return (ENOTSUP);
22947 #endif
22948 }
22949 
22950 static int
22951 rack_stack_information(struct tcpcb *tp, struct stack_specific_info *si)
22952 {
22953 	/* We pulled a SSI info log out what was there */
22954 	si->bytes_transmitted = tp->t_sndbytes;
22955 	si->bytes_retransmitted = tp->t_snd_rxt_bytes;
22956 	return (0);
22957 }
22958 
22959 static int
22960 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
22961 		    uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
22962 
22963 {
22964 	struct epoch_tracker et;
22965 	struct sockopt sopt;
22966 	struct cc_newreno_opts opt;
22967 	uint64_t val;
22968 	int error = 0;
22969 	uint16_t ca, ss;
22970 
22971 	switch (sopt_name) {
22972 	case TCP_RACK_SET_RXT_OPTIONS:
22973 		if (optval <= 2) {
22974 			rack_init_retransmit_value(rack, optval);
22975 		} else {
22976 			/*
22977 			 * You must send in 0, 1 or 2 all else is
22978 			 * invalid.
22979 			 */
22980 			error = EINVAL;
22981 		}
22982 		break;
22983 	case TCP_RACK_DSACK_OPT:
22984 		RACK_OPTS_INC(tcp_rack_dsack_opt);
22985 		if (optval & 0x1) {
22986 			rack->rc_rack_tmr_std_based = 1;
22987 		} else {
22988 			rack->rc_rack_tmr_std_based = 0;
22989 		}
22990 		if (optval & 0x2) {
22991 			rack->rc_rack_use_dsack = 1;
22992 		} else {
22993 			rack->rc_rack_use_dsack = 0;
22994 		}
22995 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
22996 		break;
22997 	case TCP_RACK_PACING_DIVISOR:
22998 		RACK_OPTS_INC(tcp_rack_pacing_divisor);
22999 		if (optval == 0) {
23000 			rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
23001 		} else {
23002 			if (optval < RL_MIN_DIVISOR)
23003 				rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
23004 			else
23005 				rack->r_ctl.pace_len_divisor = optval;
23006 		}
23007 		break;
23008 	case TCP_RACK_HI_BETA:
23009 		RACK_OPTS_INC(tcp_rack_hi_beta);
23010 		if (optval > 0) {
23011 			rack->rack_hibeta = 1;
23012 			if ((optval >= 50) &&
23013 			    (optval <= 100)) {
23014 				/*
23015 				 * User wants to set a custom beta.
23016 				 */
23017 				rack->r_ctl.saved_hibeta = optval;
23018 				if (rack->rc_pacing_cc_set)
23019 					rack_undo_cc_pacing(rack);
23020 				rack->r_ctl.rc_saved_beta.beta = optval;
23021 			}
23022 			if (rack->rc_pacing_cc_set == 0)
23023 				rack_set_cc_pacing(rack);
23024 		} else {
23025 			rack->rack_hibeta = 0;
23026 			if (rack->rc_pacing_cc_set)
23027 				rack_undo_cc_pacing(rack);
23028 		}
23029 		break;
23030 	case TCP_RACK_PACING_BETA:
23031 		error = EINVAL;
23032 		break;
23033 	case TCP_RACK_TIMER_SLOP:
23034 		RACK_OPTS_INC(tcp_rack_timer_slop);
23035 		rack->r_ctl.timer_slop = optval;
23036 		if (rack->rc_tp->t_srtt) {
23037 			/*
23038 			 * If we have an SRTT lets update t_rxtcur
23039 			 * to have the new slop.
23040 			 */
23041 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
23042 					   rack_rto_min, rack_rto_max,
23043 					   rack->r_ctl.timer_slop);
23044 		}
23045 		break;
23046 	case TCP_RACK_PACING_BETA_ECN:
23047 		RACK_OPTS_INC(tcp_rack_beta_ecn);
23048 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23049 			/* This only works for newreno. */
23050 			error = EINVAL;
23051 			break;
23052 		}
23053 		if (rack->rc_pacing_cc_set) {
23054 			/*
23055 			 * Set them into the real CC module
23056 			 * whats in the rack pcb is the old values
23057 			 * to be used on restoral/
23058 			 */
23059 			sopt.sopt_dir = SOPT_SET;
23060 			opt.name = CC_NEWRENO_BETA_ECN;
23061 			opt.val = optval;
23062 			if (CC_ALGO(tp)->ctl_output != NULL)
23063 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23064 			else
23065 				error = ENOENT;
23066 		} else {
23067 			/*
23068 			 * Not pacing yet so set it into our local
23069 			 * rack pcb storage.
23070 			 */
23071 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
23072 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
23073 		}
23074 		break;
23075 	case TCP_DEFER_OPTIONS:
23076 		RACK_OPTS_INC(tcp_defer_opt);
23077 		if (optval) {
23078 			if (rack->gp_ready) {
23079 				/* Too late */
23080 				error = EINVAL;
23081 				break;
23082 			}
23083 			rack->defer_options = 1;
23084 		} else
23085 			rack->defer_options = 0;
23086 		break;
23087 	case TCP_RACK_MEASURE_CNT:
23088 		RACK_OPTS_INC(tcp_rack_measure_cnt);
23089 		if (optval && (optval <= 0xff)) {
23090 			rack->r_ctl.req_measurements = optval;
23091 		} else
23092 			error = EINVAL;
23093 		break;
23094 	case TCP_REC_ABC_VAL:
23095 		RACK_OPTS_INC(tcp_rec_abc_val);
23096 		if (optval > 0)
23097 			rack->r_use_labc_for_rec = 1;
23098 		else
23099 			rack->r_use_labc_for_rec = 0;
23100 		break;
23101 	case TCP_RACK_ABC_VAL:
23102 		RACK_OPTS_INC(tcp_rack_abc_val);
23103 		if ((optval > 0) && (optval < 255))
23104 			rack->rc_labc = optval;
23105 		else
23106 			error = EINVAL;
23107 		break;
23108 	case TCP_HDWR_UP_ONLY:
23109 		RACK_OPTS_INC(tcp_pacing_up_only);
23110 		if (optval)
23111 			rack->r_up_only = 1;
23112 		else
23113 			rack->r_up_only = 0;
23114 		break;
23115 	case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
23116 		RACK_OPTS_INC(tcp_fillcw_rate_cap);
23117 		rack->r_ctl.fillcw_cap = loptval;
23118 		break;
23119 	case TCP_PACING_RATE_CAP:
23120 		RACK_OPTS_INC(tcp_pacing_rate_cap);
23121 		if ((rack->dgp_on == 1) &&
23122 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23123 			/*
23124 			 * If we are doing DGP we need to switch
23125 			 * to using the pacing limit.
23126 			 */
23127 			if (tcp_can_enable_pacing() == 0) {
23128 				error = ENOSPC;
23129 				break;
23130 			}
23131 			/*
23132 			 * Now change up the flags and counts to be correct.
23133 			 */
23134 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
23135 			tcp_dec_dgp_pacing_cnt();
23136 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23137 		}
23138 		rack->r_ctl.bw_rate_cap = loptval;
23139 		break;
23140 	case TCP_HYBRID_PACING:
23141 		if (hybrid == NULL) {
23142 			error = EINVAL;
23143 			break;
23144 		}
23145 		if (rack->r_ctl.side_chan_dis_mask & HYBRID_DIS_MASK) {
23146 			error = EPERM;
23147 			break;
23148 		}
23149 		error = process_hybrid_pacing(rack, hybrid);
23150 		break;
23151 	case TCP_SIDECHAN_DIS:			/*  URL:scodm */
23152 		if (optval)
23153 			rack->r_ctl.side_chan_dis_mask = optval;
23154 		else
23155 			rack->r_ctl.side_chan_dis_mask = 0;
23156 		break;
23157 	case TCP_RACK_PROFILE:
23158 		RACK_OPTS_INC(tcp_profile);
23159 		error = rack_set_profile(rack, optval);
23160 		break;
23161 	case TCP_USE_CMP_ACKS:
23162 		RACK_OPTS_INC(tcp_use_cmp_acks);
23163 		if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
23164 			/* You can't turn it off once its on! */
23165 			error = EINVAL;
23166 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
23167 			rack->r_use_cmp_ack = 1;
23168 			rack->r_mbuf_queue = 1;
23169 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23170 		}
23171 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
23172 			tp->t_flags2 |= TF2_MBUF_ACKCMP;
23173 		break;
23174 	case TCP_SHARED_CWND_TIME_LIMIT:
23175 		RACK_OPTS_INC(tcp_lscwnd);
23176 		if (optval)
23177 			rack->r_limit_scw = 1;
23178 		else
23179 			rack->r_limit_scw = 0;
23180 		break;
23181 	case TCP_RACK_DGP_IN_REC:
23182 		error = EINVAL;
23183 		break;
23184  	case TCP_RACK_PACE_TO_FILL:
23185 		RACK_OPTS_INC(tcp_fillcw);
23186 		if (optval == 0)
23187 			rack->rc_pace_to_cwnd = 0;
23188 		else {
23189 			rack->rc_pace_to_cwnd = 1;
23190 		}
23191 		if ((optval >= rack_gp_rtt_maxmul) &&
23192 		    rack_gp_rtt_maxmul &&
23193 		    (optval < 0xf)) {
23194 			rack->rc_pace_fill_if_rttin_range = 1;
23195 			rack->rtt_limit_mul = optval;
23196 		} else {
23197 			rack->rc_pace_fill_if_rttin_range = 0;
23198 			rack->rtt_limit_mul = 0;
23199 		}
23200 		break;
23201 	case TCP_RACK_NO_PUSH_AT_MAX:
23202 		RACK_OPTS_INC(tcp_npush);
23203 		if (optval == 0)
23204 			rack->r_ctl.rc_no_push_at_mrtt = 0;
23205 		else if (optval < 0xff)
23206 			rack->r_ctl.rc_no_push_at_mrtt = optval;
23207 		else
23208 			error = EINVAL;
23209 		break;
23210 	case TCP_SHARED_CWND_ENABLE:
23211 		RACK_OPTS_INC(tcp_rack_scwnd);
23212 		if (optval == 0)
23213 			rack->rack_enable_scwnd = 0;
23214 		else
23215 			rack->rack_enable_scwnd = 1;
23216 		break;
23217 	case TCP_RACK_MBUF_QUEUE:
23218 		/* Now do we use the LRO mbuf-queue feature */
23219 		RACK_OPTS_INC(tcp_rack_mbufq);
23220 		if (optval || rack->r_use_cmp_ack)
23221 			rack->r_mbuf_queue = 1;
23222 		else
23223 			rack->r_mbuf_queue = 0;
23224 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23225 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23226 		else
23227 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23228 		break;
23229 	case TCP_RACK_NONRXT_CFG_RATE:
23230 		RACK_OPTS_INC(tcp_rack_cfg_rate);
23231 		if (optval == 0)
23232 			rack->rack_rec_nonrxt_use_cr = 0;
23233 		else
23234 			rack->rack_rec_nonrxt_use_cr = 1;
23235 		break;
23236 	case TCP_NO_PRR:
23237 		RACK_OPTS_INC(tcp_rack_noprr);
23238 		if (optval == 0)
23239 			rack->rack_no_prr = 0;
23240 		else if (optval == 1)
23241 			rack->rack_no_prr = 1;
23242 		else if (optval == 2)
23243 			rack->no_prr_addback = 1;
23244 		else
23245 			error = EINVAL;
23246 		break;
23247 	case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
23248 		if (optval > 0)
23249 			rack->cspr_is_fcc = 1;
23250 		else
23251 			rack->cspr_is_fcc = 0;
23252 		break;
23253 	case TCP_TIMELY_DYN_ADJ:
23254 		RACK_OPTS_INC(tcp_timely_dyn);
23255 		if (optval == 0)
23256 			rack->rc_gp_dyn_mul = 0;
23257 		else {
23258 			rack->rc_gp_dyn_mul = 1;
23259 			if (optval >= 100) {
23260 				/*
23261 				 * If the user sets something 100 or more
23262 				 * its the gp_ca value.
23263 				 */
23264 				rack->r_ctl.rack_per_of_gp_ca  = optval;
23265 			}
23266 		}
23267 		break;
23268 	case TCP_RACK_DO_DETECTION:
23269 		error = EINVAL;
23270 		break;
23271 	case TCP_RACK_TLP_USE:
23272 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
23273 			error = EINVAL;
23274 			break;
23275 		}
23276 		RACK_OPTS_INC(tcp_tlp_use);
23277 		rack->rack_tlp_threshold_use = optval;
23278 		break;
23279 	case TCP_RACK_TLP_REDUCE:
23280 		/* RACK TLP cwnd reduction (bool) */
23281 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
23282 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
23283 		break;
23284 		/*  Pacing related ones */
23285 	case TCP_RACK_PACE_ALWAYS:
23286 		/*
23287 		 * zero is old rack method, 1 is new
23288 		 * method using a pacing rate.
23289 		 */
23290 		RACK_OPTS_INC(tcp_rack_pace_always);
23291 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23292 			error = EPERM;
23293 			break;
23294 		}
23295 		if (optval > 0) {
23296 			if (rack->rc_always_pace) {
23297 				error = EALREADY;
23298 				break;
23299 			} else if (tcp_can_enable_pacing()) {
23300 				rack->r_ctl.pacing_method |= RACK_REG_PACING;
23301 				rack->rc_always_pace = 1;
23302 				if (rack->rack_hibeta)
23303 					rack_set_cc_pacing(rack);
23304 			}
23305 			else {
23306 				error = ENOSPC;
23307 				break;
23308 			}
23309 		} else {
23310 			if (rack->rc_always_pace == 1) {
23311 				rack_remove_pacing(rack);
23312 			}
23313 		}
23314 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23315 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23316 		else
23317 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23318 		/* A rate may be set irate or other, if so set seg size */
23319 		rack_update_seg(rack);
23320 		break;
23321 	case TCP_BBR_RACK_INIT_RATE:
23322 		RACK_OPTS_INC(tcp_initial_rate);
23323 		val = optval;
23324 		/* Change from kbits per second to bytes per second */
23325 		val *= 1000;
23326 		val /= 8;
23327 		rack->r_ctl.init_rate = val;
23328 		if (rack->rc_always_pace)
23329 			rack_update_seg(rack);
23330 		break;
23331 	case TCP_BBR_IWINTSO:
23332 		error = EINVAL;
23333 		break;
23334 	case TCP_RACK_FORCE_MSEG:
23335 		RACK_OPTS_INC(tcp_rack_force_max_seg);
23336 		if (optval)
23337 			rack->rc_force_max_seg = 1;
23338 		else
23339 			rack->rc_force_max_seg = 0;
23340 		break;
23341 	case TCP_RACK_PACE_MIN_SEG:
23342 		RACK_OPTS_INC(tcp_rack_min_seg);
23343 		rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23344 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23345 		break;
23346 	case TCP_RACK_PACE_MAX_SEG:
23347 		/* Max segments size in a pace in bytes */
23348 		RACK_OPTS_INC(tcp_rack_max_seg);
23349 		if ((rack->dgp_on == 1) &&
23350 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23351 			/*
23352 			 * If we set a max-seg and are doing DGP then
23353 			 * we now fall under the pacing limits not the
23354 			 * DGP ones.
23355 			 */
23356 			if (tcp_can_enable_pacing() == 0) {
23357 				error = ENOSPC;
23358 				break;
23359 			}
23360 			/*
23361 			 * Now change up the flags and counts to be correct.
23362 			 */
23363 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
23364 			tcp_dec_dgp_pacing_cnt();
23365 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23366 		}
23367 		if (optval <= MAX_USER_SET_SEG)
23368 			rack->rc_user_set_max_segs = optval;
23369 		else
23370 			rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
23371 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23372 		break;
23373 	case TCP_RACK_PACE_RATE_REC:
23374 		/* Set the fixed pacing rate in Bytes per second ca */
23375 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23376 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23377 			error = EPERM;
23378 			break;
23379 		}
23380 		if (rack->dgp_on) {
23381 			/*
23382 			 * We are already pacing another
23383 			 * way.
23384 			 */
23385 			error = EBUSY;
23386 			break;
23387 		}
23388 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23389 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23390 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23391 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23392 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23393 		rack->use_fixed_rate = 1;
23394 		if (rack->rack_hibeta)
23395 			rack_set_cc_pacing(rack);
23396 		rack_log_pacing_delay_calc(rack,
23397 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23398 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23399 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23400 					   __LINE__, NULL,0);
23401 		break;
23402 
23403 	case TCP_RACK_PACE_RATE_SS:
23404 		/* Set the fixed pacing rate in Bytes per second ca */
23405 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23406 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23407 			error = EPERM;
23408 			break;
23409 		}
23410 		if (rack->dgp_on) {
23411 			/*
23412 			 * We are already pacing another
23413 			 * way.
23414 			 */
23415 			error = EBUSY;
23416 			break;
23417 		}
23418 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23419 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23420 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23421 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23422 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23423 		rack->use_fixed_rate = 1;
23424 		if (rack->rack_hibeta)
23425 			rack_set_cc_pacing(rack);
23426 		rack_log_pacing_delay_calc(rack,
23427 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23428 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23429 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23430 					   __LINE__, NULL, 0);
23431 		break;
23432 
23433 	case TCP_RACK_PACE_RATE_CA:
23434 		/* Set the fixed pacing rate in Bytes per second ca */
23435 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23436 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23437 			error = EPERM;
23438 			break;
23439 		}
23440 		if (rack->dgp_on) {
23441 			/*
23442 			 * We are already pacing another
23443 			 * way.
23444 			 */
23445 			error = EBUSY;
23446 			break;
23447 		}
23448 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23449 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23450 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23451 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23452 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23453 		rack->use_fixed_rate = 1;
23454 		if (rack->rack_hibeta)
23455 			rack_set_cc_pacing(rack);
23456 		rack_log_pacing_delay_calc(rack,
23457 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23458 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23459 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23460 					   __LINE__, NULL, 0);
23461 		break;
23462 	case TCP_RACK_GP_INCREASE_REC:
23463 		RACK_OPTS_INC(tcp_gp_inc_rec);
23464 		rack->r_ctl.rack_per_of_gp_rec = optval;
23465 		rack_log_pacing_delay_calc(rack,
23466 					   rack->r_ctl.rack_per_of_gp_ss,
23467 					   rack->r_ctl.rack_per_of_gp_ca,
23468 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23469 					   __LINE__, NULL, 0);
23470 		break;
23471 	case TCP_RACK_GP_INCREASE_CA:
23472 		RACK_OPTS_INC(tcp_gp_inc_ca);
23473 		ca = optval;
23474 		if (ca < 100) {
23475 			/*
23476 			 * We don't allow any reduction
23477 			 * over the GP b/w.
23478 			 */
23479 			error = EINVAL;
23480 			break;
23481 		}
23482 		rack->r_ctl.rack_per_of_gp_ca = ca;
23483 		rack_log_pacing_delay_calc(rack,
23484 					   rack->r_ctl.rack_per_of_gp_ss,
23485 					   rack->r_ctl.rack_per_of_gp_ca,
23486 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23487 					   __LINE__, NULL, 0);
23488 		break;
23489 	case TCP_RACK_GP_INCREASE_SS:
23490 		RACK_OPTS_INC(tcp_gp_inc_ss);
23491 		ss = optval;
23492 		if (ss < 100) {
23493 			/*
23494 			 * We don't allow any reduction
23495 			 * over the GP b/w.
23496 			 */
23497 			error = EINVAL;
23498 			break;
23499 		}
23500 		rack->r_ctl.rack_per_of_gp_ss = ss;
23501 		rack_log_pacing_delay_calc(rack,
23502 					   rack->r_ctl.rack_per_of_gp_ss,
23503 					   rack->r_ctl.rack_per_of_gp_ca,
23504 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23505 					   __LINE__, NULL, 0);
23506 		break;
23507 	case TCP_RACK_RR_CONF:
23508 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23509 		if (optval && optval <= 3)
23510 			rack->r_rr_config = optval;
23511 		else
23512 			rack->r_rr_config = 0;
23513 		break;
23514 	case TCP_PACING_DND:			/*  URL:dnd */
23515 		if (optval > 0)
23516 			rack->rc_pace_dnd = 1;
23517 		else
23518 			rack->rc_pace_dnd = 0;
23519 		break;
23520 	case TCP_HDWR_RATE_CAP:
23521 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
23522 		if (optval) {
23523 			if (rack->r_rack_hw_rate_caps == 0)
23524 				rack->r_rack_hw_rate_caps = 1;
23525 			else
23526 				error = EALREADY;
23527 		} else {
23528 			rack->r_rack_hw_rate_caps = 0;
23529 		}
23530 		break;
23531 	case TCP_DGP_UPPER_BOUNDS:
23532 	{
23533 		uint8_t val;
23534 		val = optval & 0x0000ff;
23535 		rack->r_ctl.rack_per_upper_bound_ca = val;
23536 		val = (optval >> 16) & 0x0000ff;
23537 		rack->r_ctl.rack_per_upper_bound_ss = val;
23538 		break;
23539 	}
23540 	case TCP_SS_EEXIT:			/*  URL:eexit */
23541 		if (optval > 0) {
23542 			rack->r_ctl.gp_rnd_thresh =  optval & 0x0ff;
23543 			if (optval & 0x10000) {
23544 				rack->r_ctl.gate_to_fs = 1;
23545 			} else {
23546 				rack->r_ctl.gate_to_fs = 0;
23547 			}
23548 			if (optval & 0x20000) {
23549 				rack->r_ctl.use_gp_not_last = 1;
23550 			} else {
23551 				rack->r_ctl.use_gp_not_last = 0;
23552 			}
23553 			if (optval & 0xfffc0000) {
23554 				uint32_t v;
23555 
23556 				v = (optval >> 18) & 0x00003fff;
23557 				if (v >= 1000)
23558 					rack->r_ctl.gp_gain_req = v;
23559 			}
23560 		} else {
23561 			/* We do not do ss early exit at all */
23562 			rack->rc_initial_ss_comp = 1;
23563 			rack->r_ctl.gp_rnd_thresh = 0;
23564 		}
23565 		break;
23566 	case TCP_RACK_SPLIT_LIMIT:
23567 		RACK_OPTS_INC(tcp_split_limit);
23568 		rack->r_ctl.rc_split_limit = optval;
23569 		break;
23570 	case TCP_BBR_HDWR_PACE:
23571 		RACK_OPTS_INC(tcp_hdwr_pacing);
23572 		if (optval){
23573 			if (rack->rack_hdrw_pacing == 0) {
23574 				rack->rack_hdw_pace_ena = 1;
23575 				rack->rack_attempt_hdwr_pace = 0;
23576 			} else
23577 				error = EALREADY;
23578 		} else {
23579 			rack->rack_hdw_pace_ena = 0;
23580 #ifdef RATELIMIT
23581 			if (rack->r_ctl.crte != NULL) {
23582 				rack->rack_hdrw_pacing = 0;
23583 				rack->rack_attempt_hdwr_pace = 0;
23584 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23585 				rack->r_ctl.crte = NULL;
23586 			}
23587 #endif
23588 		}
23589 		break;
23590 		/*  End Pacing related ones */
23591 	case TCP_RACK_PRR_SENDALOT:
23592 		/* Allow PRR to send more than one seg */
23593 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
23594 		rack->r_ctl.rc_prr_sendalot = optval;
23595 		break;
23596 	case TCP_RACK_MIN_TO:
23597 		/* Minimum time between rack t-o's in ms */
23598 		RACK_OPTS_INC(tcp_rack_min_to);
23599 		rack->r_ctl.rc_min_to = optval;
23600 		break;
23601 	case TCP_RACK_EARLY_SEG:
23602 		/* If early recovery max segments */
23603 		RACK_OPTS_INC(tcp_rack_early_seg);
23604 		rack->r_ctl.rc_early_recovery_segs = optval;
23605 		break;
23606 	case TCP_RACK_ENABLE_HYSTART:
23607 	{
23608 		if (optval) {
23609 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23610 			if (rack_do_hystart > RACK_HYSTART_ON)
23611 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23612 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23613 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23614 		} else {
23615 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23616 		}
23617 	}
23618 	break;
23619 	case TCP_RACK_REORD_THRESH:
23620 		/* RACK reorder threshold (shift amount) */
23621 		RACK_OPTS_INC(tcp_rack_reord_thresh);
23622 		if ((optval > 0) && (optval < 31))
23623 			rack->r_ctl.rc_reorder_shift = optval;
23624 		else
23625 			error = EINVAL;
23626 		break;
23627 	case TCP_RACK_REORD_FADE:
23628 		/* Does reordering fade after ms time */
23629 		RACK_OPTS_INC(tcp_rack_reord_fade);
23630 		rack->r_ctl.rc_reorder_fade = optval;
23631 		break;
23632 	case TCP_RACK_TLP_THRESH:
23633 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
23634 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
23635 		if (optval)
23636 			rack->r_ctl.rc_tlp_threshold = optval;
23637 		else
23638 			error = EINVAL;
23639 		break;
23640 	case TCP_BBR_USE_RACK_RR:
23641 		RACK_OPTS_INC(tcp_rack_rr);
23642 		if (optval)
23643 			rack->use_rack_rr = 1;
23644 		else
23645 			rack->use_rack_rr = 0;
23646 		break;
23647 	case TCP_RACK_PKT_DELAY:
23648 		/* RACK added ms i.e. rack-rtt + reord + N */
23649 		RACK_OPTS_INC(tcp_rack_pkt_delay);
23650 		rack->r_ctl.rc_pkt_delay = optval;
23651 		break;
23652 	case TCP_DELACK:
23653 		RACK_OPTS_INC(tcp_rack_delayed_ack);
23654 		if (optval == 0)
23655 			tp->t_delayed_ack = 0;
23656 		else
23657 			tp->t_delayed_ack = 1;
23658 		if (tp->t_flags & TF_DELACK) {
23659 			tp->t_flags &= ~TF_DELACK;
23660 			tp->t_flags |= TF_ACKNOW;
23661 			NET_EPOCH_ENTER(et);
23662 			rack_output(tp);
23663 			NET_EPOCH_EXIT(et);
23664 		}
23665 		break;
23666 
23667 	case TCP_BBR_RACK_RTT_USE:
23668 		RACK_OPTS_INC(tcp_rack_rtt_use);
23669 		if ((optval != USE_RTT_HIGH) &&
23670 		    (optval != USE_RTT_LOW) &&
23671 		    (optval != USE_RTT_AVG))
23672 			error = EINVAL;
23673 		else
23674 			rack->r_ctl.rc_rate_sample_method = optval;
23675 		break;
23676 	case TCP_HONOR_HPTS_MIN:
23677 		RACK_OPTS_INC(tcp_honor_hpts);
23678 		if (optval) {
23679 			rack->r_use_hpts_min = 1;
23680 			/*
23681 			 * Must be between 2 - 80% to be a reduction else
23682 			 * we keep the default (10%).
23683 			 */
23684 			if ((optval > 1) && (optval <= 80)) {
23685 				rack->r_ctl.max_reduction = optval;
23686 			}
23687 		} else
23688 			rack->r_use_hpts_min = 0;
23689 		break;
23690 	case TCP_REC_IS_DYN:			/*  URL:dynrec */
23691 		RACK_OPTS_INC(tcp_dyn_rec);
23692 		if (optval)
23693 			rack->rc_gp_no_rec_chg = 1;
23694 		else
23695 			rack->rc_gp_no_rec_chg = 0;
23696 		break;
23697 	case TCP_NO_TIMELY:
23698 		RACK_OPTS_INC(tcp_notimely);
23699 		if (optval) {
23700 			rack->rc_skip_timely = 1;
23701 			rack->r_ctl.rack_per_of_gp_rec = 90;
23702 			rack->r_ctl.rack_per_of_gp_ca = 100;
23703 			rack->r_ctl.rack_per_of_gp_ss = 250;
23704 		} else {
23705 			rack->rc_skip_timely = 0;
23706 		}
23707 		break;
23708 	case TCP_GP_USE_LTBW:
23709 		if (optval == 0) {
23710 			rack->use_lesser_lt_bw = 0;
23711 			rack->dis_lt_bw = 1;
23712 		} else if (optval == 1) {
23713 			rack->use_lesser_lt_bw = 1;
23714 			rack->dis_lt_bw = 0;
23715 		} else if (optval == 2) {
23716 			rack->use_lesser_lt_bw = 0;
23717 			rack->dis_lt_bw = 0;
23718 		}
23719 		break;
23720 	case TCP_DATA_AFTER_CLOSE:
23721 		RACK_OPTS_INC(tcp_data_after_close);
23722 		if (optval)
23723 			rack->rc_allow_data_af_clo = 1;
23724 		else
23725 			rack->rc_allow_data_af_clo = 0;
23726 		break;
23727 	default:
23728 		break;
23729 	}
23730 	tcp_log_socket_option(tp, sopt_name, optval, error);
23731 	return (error);
23732 }
23733 
23734 static void
23735 rack_inherit(struct tcpcb *tp, struct inpcb *parent)
23736 {
23737 	/*
23738 	 * A new connection has been created (tp) and
23739 	 * the parent is the inpcb given. We want to
23740 	 * apply a read-lock to the parent (we are already
23741 	 * holding a write lock on the tp) and copy anything
23742 	 * out of the rack specific data as long as its tfb is
23743 	 * the same as ours i.e. we are the same stack. Otherwise
23744 	 * we just return.
23745 	 */
23746 	struct tcpcb *par;
23747 	struct tcp_rack *dest, *src;
23748 	int cnt = 0;
23749 
23750 	par = intotcpcb(parent);
23751 	if (par->t_fb != tp->t_fb) {
23752 		/* Not the same stack */
23753 		tcp_log_socket_option(tp, 0, 0, 1);
23754 		return;
23755 	}
23756 	/* Ok if we reach here lets setup the two rack pointers */
23757 	dest = (struct tcp_rack *)tp->t_fb_ptr;
23758 	src = (struct tcp_rack *)par->t_fb_ptr;
23759 	if ((src == NULL) || (dest == NULL)) {
23760 		/* Huh? */
23761 		tcp_log_socket_option(tp, 0, 0, 2);
23762 		return;
23763 	}
23764 	/* Now copy out anything we wish to inherit i.e. things in socket-options */
23765 	/* TCP_RACK_PROFILE we can't know but we can set DGP if its on */
23766 	if ((src->dgp_on) && (dest->dgp_on == 0)) {
23767 		/* Profile 1 had to be set via sock opt */
23768 		rack_set_dgp(dest);
23769 		cnt++;
23770 	}
23771 	/* TCP_RACK_SET_RXT_OPTIONS */
23772 	if (dest->full_size_rxt != src->full_size_rxt) {
23773 		dest->full_size_rxt = src->full_size_rxt;
23774 		cnt++;
23775 	}
23776 	if (dest->shape_rxt_to_pacing_min  != src->shape_rxt_to_pacing_min) {
23777 		dest->shape_rxt_to_pacing_min = src->shape_rxt_to_pacing_min;
23778 		cnt++;
23779 	}
23780 	/* TCP_RACK_DSACK_OPT */
23781 	if (dest->rc_rack_tmr_std_based != src->rc_rack_tmr_std_based) {
23782 		dest->rc_rack_tmr_std_based = src->rc_rack_tmr_std_based;
23783 		cnt++;
23784 	}
23785 	if (dest->rc_rack_use_dsack != src->rc_rack_use_dsack) {
23786 		dest->rc_rack_use_dsack = src->rc_rack_use_dsack;
23787 		cnt++;
23788 	}
23789 	/* TCP_RACK_PACING_DIVISOR */
23790 	if (dest->r_ctl.pace_len_divisor != src->r_ctl.pace_len_divisor) {
23791 		dest->r_ctl.pace_len_divisor = src->r_ctl.pace_len_divisor;
23792 		cnt++;
23793 	}
23794 	/* TCP_RACK_HI_BETA */
23795 	if (src->rack_hibeta != dest->rack_hibeta) {
23796 		cnt++;
23797 		if (src->rack_hibeta) {
23798 			dest->r_ctl.rc_saved_beta.beta = src->r_ctl.rc_saved_beta.beta;
23799 			dest->rack_hibeta = 1;
23800 		} else {
23801 			dest->rack_hibeta = 0;
23802 		}
23803 	}
23804 	/* TCP_RACK_TIMER_SLOP */
23805 	if (dest->r_ctl.timer_slop != src->r_ctl.timer_slop) {
23806 		dest->r_ctl.timer_slop = src->r_ctl.timer_slop;
23807 		cnt++;
23808 	}
23809 	/* TCP_RACK_PACING_BETA_ECN */
23810 	if (dest->r_ctl.rc_saved_beta.beta_ecn != src->r_ctl.rc_saved_beta.beta_ecn) {
23811 		dest->r_ctl.rc_saved_beta.beta_ecn = src->r_ctl.rc_saved_beta.beta_ecn;
23812 		cnt++;
23813 	}
23814 	if (dest->r_ctl.rc_saved_beta.newreno_flags != src->r_ctl.rc_saved_beta.newreno_flags) {
23815 		dest->r_ctl.rc_saved_beta.newreno_flags = src->r_ctl.rc_saved_beta.newreno_flags;
23816 		cnt++;
23817 	}
23818 	/* We do not do TCP_DEFER_OPTIONS */
23819 	/* TCP_RACK_MEASURE_CNT */
23820 	if (dest->r_ctl.req_measurements != src->r_ctl.req_measurements) {
23821 		dest->r_ctl.req_measurements = src->r_ctl.req_measurements;
23822 		cnt++;
23823 	}
23824 	/* TCP_HDWR_UP_ONLY */
23825 	if (dest->r_up_only != src->r_up_only) {
23826 		dest->r_up_only = src->r_up_only;
23827 		cnt++;
23828 	}
23829 	/* TCP_FILLCW_RATE_CAP */
23830 	if (dest->r_ctl.fillcw_cap != src->r_ctl.fillcw_cap) {
23831 		dest->r_ctl.fillcw_cap = src->r_ctl.fillcw_cap;
23832 		cnt++;
23833 	}
23834 	/* TCP_PACING_RATE_CAP */
23835 	if (dest->r_ctl.bw_rate_cap != src->r_ctl.bw_rate_cap) {
23836 		dest->r_ctl.bw_rate_cap = src->r_ctl.bw_rate_cap;
23837 		cnt++;
23838 	}
23839 	/* A listener can't set TCP_HYBRID_PACING */
23840 	/* TCP_SIDECHAN_DIS */
23841 	if (dest->r_ctl.side_chan_dis_mask != src->r_ctl.side_chan_dis_mask) {
23842 		dest->r_ctl.side_chan_dis_mask = src->r_ctl.side_chan_dis_mask;
23843 		cnt++;
23844 	}
23845 	/* TCP_SHARED_CWND_TIME_LIMIT */
23846 	if (dest->r_limit_scw != src->r_limit_scw) {
23847 		dest->r_limit_scw = src->r_limit_scw;
23848 		cnt++;
23849 	}
23850 	/* TCP_RACK_PACE_TO_FILL */
23851 	if (dest->rc_pace_to_cwnd != src->rc_pace_to_cwnd) {
23852 		dest->rc_pace_to_cwnd = src->rc_pace_to_cwnd;
23853 		cnt++;
23854 	}
23855 	if (dest->rc_pace_fill_if_rttin_range != src->rc_pace_fill_if_rttin_range) {
23856 		dest->rc_pace_fill_if_rttin_range = src->rc_pace_fill_if_rttin_range;
23857 		cnt++;
23858 	}
23859 	if (dest->rtt_limit_mul != src->rtt_limit_mul) {
23860 		dest->rtt_limit_mul = src->rtt_limit_mul;
23861 		cnt++;
23862 	}
23863 	/* TCP_RACK_NO_PUSH_AT_MAX */
23864 	if (dest->r_ctl.rc_no_push_at_mrtt != src->r_ctl.rc_no_push_at_mrtt) {
23865 		dest->r_ctl.rc_no_push_at_mrtt = src->r_ctl.rc_no_push_at_mrtt;
23866 		cnt++;
23867 	}
23868 	/* TCP_SHARED_CWND_ENABLE */
23869 	if (dest->rack_enable_scwnd != src->rack_enable_scwnd) {
23870 		dest->rack_enable_scwnd = src->rack_enable_scwnd;
23871 		cnt++;
23872 	}
23873 	/* TCP_USE_CMP_ACKS */
23874 	if (dest->r_use_cmp_ack != src->r_use_cmp_ack) {
23875 		dest->r_use_cmp_ack = src->r_use_cmp_ack;
23876 		cnt++;
23877 	}
23878 
23879 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23880 		dest->r_mbuf_queue = src->r_mbuf_queue;
23881 		cnt++;
23882 	}
23883 	/* TCP_RACK_MBUF_QUEUE */
23884 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23885 		dest->r_mbuf_queue = src->r_mbuf_queue;
23886 		cnt++;
23887 	}
23888 	if  (dest->r_mbuf_queue || dest->rc_always_pace || dest->r_use_cmp_ack) {
23889 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23890 	} else {
23891 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23892 	}
23893 	if (dest->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) {
23894 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
23895 	}
23896 	/* TCP_RACK_NONRXT_CFG_RATE */
23897 	if (dest->rack_rec_nonrxt_use_cr != src->rack_rec_nonrxt_use_cr) {
23898 		dest->rack_rec_nonrxt_use_cr = src->rack_rec_nonrxt_use_cr;
23899 		cnt++;
23900 	}
23901 	/* TCP_NO_PRR */
23902 	if (dest->rack_no_prr != src->rack_no_prr) {
23903 		dest->rack_no_prr = src->rack_no_prr;
23904 		cnt++;
23905 	}
23906 	if (dest->no_prr_addback != src->no_prr_addback) {
23907 		dest->no_prr_addback = src->no_prr_addback;
23908 		cnt++;
23909 	}
23910 	/* RACK_CSPR_IS_FCC */
23911 	if (dest->cspr_is_fcc != src->cspr_is_fcc) {
23912 		dest->cspr_is_fcc = src->cspr_is_fcc;
23913 		cnt++;
23914 	}
23915 	/* TCP_TIMELY_DYN_ADJ */
23916 	if (dest->rc_gp_dyn_mul != src->rc_gp_dyn_mul) {
23917 		dest->rc_gp_dyn_mul = src->rc_gp_dyn_mul;
23918 		cnt++;
23919 	}
23920 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23921 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23922 		cnt++;
23923 	}
23924 	/* TCP_RACK_TLP_USE */
23925 	if (dest->rack_tlp_threshold_use != src->rack_tlp_threshold_use) {
23926 		dest->rack_tlp_threshold_use = src->rack_tlp_threshold_use;
23927 		cnt++;
23928 	}
23929 	/* we don't allow inheritence of TCP_RACK_PACE_ALWAYS */
23930 	/* TCP_BBR_RACK_INIT_RATE */
23931 	if (dest->r_ctl.init_rate != src->r_ctl.init_rate) {
23932 		dest->r_ctl.init_rate = src->r_ctl.init_rate;
23933 		cnt++;
23934 	}
23935 	/* TCP_RACK_FORCE_MSEG */
23936 	if (dest->rc_force_max_seg != src->rc_force_max_seg) {
23937 		dest->rc_force_max_seg = src->rc_force_max_seg;
23938 		cnt++;
23939 	}
23940 	/* TCP_RACK_PACE_MIN_SEG */
23941 	if (dest->r_ctl.rc_user_set_min_segs != src->r_ctl.rc_user_set_min_segs) {
23942 		dest->r_ctl.rc_user_set_min_segs = src->r_ctl.rc_user_set_min_segs;
23943 		cnt++;
23944 	}
23945 	/* we don't allow TCP_RACK_PACE_MAX_SEG */
23946 	/* TCP_RACK_PACE_RATE_REC, TCP_RACK_PACE_RATE_SS,  TCP_RACK_PACE_RATE_CA */
23947 	if (dest->r_ctl.rc_fixed_pacing_rate_ca != src->r_ctl.rc_fixed_pacing_rate_ca) {
23948 		dest->r_ctl.rc_fixed_pacing_rate_ca = src->r_ctl.rc_fixed_pacing_rate_ca;
23949 		cnt++;
23950 	}
23951 	if (dest->r_ctl.rc_fixed_pacing_rate_ss != src->r_ctl.rc_fixed_pacing_rate_ss) {
23952 		dest->r_ctl.rc_fixed_pacing_rate_ss = src->r_ctl.rc_fixed_pacing_rate_ss;
23953 		cnt++;
23954 	}
23955 	if (dest->r_ctl.rc_fixed_pacing_rate_rec != src->r_ctl.rc_fixed_pacing_rate_rec) {
23956 		dest->r_ctl.rc_fixed_pacing_rate_rec = src->r_ctl.rc_fixed_pacing_rate_rec;
23957 		cnt++;
23958 	}
23959 	/* TCP_RACK_GP_INCREASE_REC, TCP_RACK_GP_INCREASE_CA, TCP_RACK_GP_INCREASE_SS */
23960 	if (dest->r_ctl.rack_per_of_gp_rec != src->r_ctl.rack_per_of_gp_rec) {
23961 		dest->r_ctl.rack_per_of_gp_rec = src->r_ctl.rack_per_of_gp_rec;
23962 		cnt++;
23963 	}
23964 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23965 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23966 		cnt++;
23967 	}
23968 
23969 	if (dest->r_ctl.rack_per_of_gp_ss != src->r_ctl.rack_per_of_gp_ss) {
23970 		dest->r_ctl.rack_per_of_gp_ss = src->r_ctl.rack_per_of_gp_ss;
23971 		cnt++;
23972 	}
23973 	/* TCP_RACK_RR_CONF */
23974 	if (dest->r_rr_config != src->r_rr_config) {
23975 		dest->r_rr_config = src->r_rr_config;
23976 		cnt++;
23977 	}
23978 	/* TCP_PACING_DND */
23979 	if (dest->rc_pace_dnd != src->rc_pace_dnd) {
23980 		dest->rc_pace_dnd = src->rc_pace_dnd;
23981 		cnt++;
23982 	}
23983 	/* TCP_HDWR_RATE_CAP */
23984 	if (dest->r_rack_hw_rate_caps != src->r_rack_hw_rate_caps) {
23985 		dest->r_rack_hw_rate_caps = src->r_rack_hw_rate_caps;
23986 		cnt++;
23987 	}
23988 	/* TCP_DGP_UPPER_BOUNDS */
23989 	if (dest->r_ctl.rack_per_upper_bound_ca != src->r_ctl.rack_per_upper_bound_ca) {
23990 		dest->r_ctl.rack_per_upper_bound_ca = src->r_ctl.rack_per_upper_bound_ca;
23991 		cnt++;
23992 	}
23993 	if (dest->r_ctl.rack_per_upper_bound_ss != src->r_ctl.rack_per_upper_bound_ss) {
23994 		dest->r_ctl.rack_per_upper_bound_ss = src->r_ctl.rack_per_upper_bound_ss;
23995 		cnt++;
23996 	}
23997 	/* TCP_SS_EEXIT */
23998 	if (dest->r_ctl.gp_rnd_thresh != src->r_ctl.gp_rnd_thresh) {
23999 		dest->r_ctl.gp_rnd_thresh = src->r_ctl.gp_rnd_thresh;
24000 		cnt++;
24001 	}
24002 	if (dest->r_ctl.gate_to_fs != src->r_ctl.gate_to_fs) {
24003 		dest->r_ctl.gate_to_fs = src->r_ctl.gate_to_fs;
24004 		cnt++;
24005 	}
24006 	if (dest->r_ctl.use_gp_not_last != src->r_ctl.use_gp_not_last) {
24007 		dest->r_ctl.use_gp_not_last = src->r_ctl.use_gp_not_last;
24008 		cnt++;
24009 	}
24010 	if (dest->r_ctl.gp_gain_req != src->r_ctl.gp_gain_req) {
24011 		dest->r_ctl.gp_gain_req = src->r_ctl.gp_gain_req;
24012 		cnt++;
24013 	}
24014 	/* TCP_BBR_HDWR_PACE */
24015 	if (dest->rack_hdw_pace_ena != src->rack_hdw_pace_ena) {
24016 		dest->rack_hdw_pace_ena = src->rack_hdw_pace_ena;
24017 		cnt++;
24018 	}
24019 	if (dest->rack_attempt_hdwr_pace != src->rack_attempt_hdwr_pace) {
24020 		dest->rack_attempt_hdwr_pace = src->rack_attempt_hdwr_pace;
24021 		cnt++;
24022 	}
24023 	/* TCP_RACK_PRR_SENDALOT */
24024 	if (dest->r_ctl.rc_prr_sendalot != src->r_ctl.rc_prr_sendalot) {
24025 		dest->r_ctl.rc_prr_sendalot = src->r_ctl.rc_prr_sendalot;
24026 		cnt++;
24027 	}
24028 	/* TCP_RACK_MIN_TO */
24029 	if (dest->r_ctl.rc_min_to != src->r_ctl.rc_min_to) {
24030 		dest->r_ctl.rc_min_to = src->r_ctl.rc_min_to;
24031 		cnt++;
24032 	}
24033 	/* TCP_RACK_EARLY_SEG */
24034 	if (dest->r_ctl.rc_early_recovery_segs != src->r_ctl.rc_early_recovery_segs) {
24035 		dest->r_ctl.rc_early_recovery_segs = src->r_ctl.rc_early_recovery_segs;
24036 		cnt++;
24037 	}
24038 	/* TCP_RACK_ENABLE_HYSTART */
24039 	if (par->t_ccv.flags != tp->t_ccv.flags) {
24040 		cnt++;
24041 		if (par->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24042 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
24043 			if (rack_do_hystart > RACK_HYSTART_ON)
24044 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
24045 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
24046 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
24047 		} else {
24048 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
24049 		}
24050 	}
24051 	/* TCP_RACK_REORD_THRESH */
24052 	if (dest->r_ctl.rc_reorder_shift != src->r_ctl.rc_reorder_shift) {
24053 		dest->r_ctl.rc_reorder_shift = src->r_ctl.rc_reorder_shift;
24054 		cnt++;
24055 	}
24056 	/* TCP_RACK_REORD_FADE */
24057 	if (dest->r_ctl.rc_reorder_fade != src->r_ctl.rc_reorder_fade) {
24058 		dest->r_ctl.rc_reorder_fade = src->r_ctl.rc_reorder_fade;
24059 		cnt++;
24060 	}
24061 	/* TCP_RACK_TLP_THRESH */
24062 	if (dest->r_ctl.rc_tlp_threshold != src->r_ctl.rc_tlp_threshold) {
24063 		dest->r_ctl.rc_tlp_threshold = src->r_ctl.rc_tlp_threshold;
24064 		cnt++;
24065 	}
24066 	/* TCP_BBR_USE_RACK_RR */
24067 	if (dest->use_rack_rr != src->use_rack_rr) {
24068 		dest->use_rack_rr = src->use_rack_rr;
24069 		cnt++;
24070 	}
24071 	/* TCP_RACK_PKT_DELAY */
24072 	if (dest->r_ctl.rc_pkt_delay != src->r_ctl.rc_pkt_delay) {
24073 		dest->r_ctl.rc_pkt_delay = src->r_ctl.rc_pkt_delay;
24074 		cnt++;
24075 	}
24076 	/* TCP_DELACK will get copied via the main code if applicable */
24077 	/* TCP_BBR_RACK_RTT_USE */
24078 	if (dest->r_ctl.rc_rate_sample_method != src->r_ctl.rc_rate_sample_method) {
24079 		dest->r_ctl.rc_rate_sample_method = src->r_ctl.rc_rate_sample_method;
24080 		cnt++;
24081 	}
24082 	/* TCP_HONOR_HPTS_MIN */
24083 	if (dest->r_use_hpts_min != src->r_use_hpts_min) {
24084 		dest->r_use_hpts_min = src->r_use_hpts_min;
24085 		cnt++;
24086 	}
24087 	if (dest->r_ctl.max_reduction != src->r_ctl.max_reduction) {
24088 		dest->r_ctl.max_reduction = src->r_ctl.max_reduction;
24089 		cnt++;
24090 	}
24091 	/* TCP_REC_IS_DYN */
24092 	if (dest->rc_gp_no_rec_chg != src->rc_gp_no_rec_chg) {
24093 		dest->rc_gp_no_rec_chg = src->rc_gp_no_rec_chg;
24094 		cnt++;
24095 	}
24096 	if (dest->rc_skip_timely != src->rc_skip_timely) {
24097 		dest->rc_skip_timely = src->rc_skip_timely;
24098 		cnt++;
24099 	}
24100 	/* TCP_DATA_AFTER_CLOSE */
24101 	if (dest->rc_allow_data_af_clo != src->rc_allow_data_af_clo) {
24102 		dest->rc_allow_data_af_clo = src->rc_allow_data_af_clo;
24103 		cnt++;
24104 	}
24105 	/* TCP_GP_USE_LTBW */
24106 	if (src->use_lesser_lt_bw != dest->use_lesser_lt_bw) {
24107 		dest->use_lesser_lt_bw = src->use_lesser_lt_bw;
24108 		cnt++;
24109 	}
24110 	if (dest->dis_lt_bw != src->dis_lt_bw) {
24111 		dest->dis_lt_bw = src->dis_lt_bw;
24112 		cnt++;
24113 	}
24114 	tcp_log_socket_option(tp, 0, cnt, 0);
24115 }
24116 
24117 
24118 static void
24119 rack_apply_deferred_options(struct tcp_rack *rack)
24120 {
24121 	struct deferred_opt_list *dol, *sdol;
24122 	uint32_t s_optval;
24123 
24124 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
24125 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
24126 		/* Disadvantage of deferal is you loose the error return */
24127 		s_optval = (uint32_t)dol->optval;
24128 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
24129 		free(dol, M_TCPDO);
24130 	}
24131 }
24132 
24133 static void
24134 rack_hw_tls_change(struct tcpcb *tp, int chg)
24135 {
24136 	/* Update HW tls state */
24137 	struct tcp_rack *rack;
24138 
24139 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24140 	if (chg)
24141 		rack->r_ctl.fsb.hw_tls = 1;
24142 	else
24143 		rack->r_ctl.fsb.hw_tls = 0;
24144 }
24145 
24146 static int
24147 rack_pru_options(struct tcpcb *tp, int flags)
24148 {
24149 	if (flags & PRUS_OOB)
24150 		return (EOPNOTSUPP);
24151 	return (0);
24152 }
24153 
24154 static bool
24155 rack_wake_check(struct tcpcb *tp)
24156 {
24157 	struct tcp_rack *rack;
24158 	struct timeval tv;
24159 	uint32_t cts;
24160 
24161 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24162 	if (rack->r_ctl.rc_hpts_flags) {
24163 		cts = tcp_get_usecs(&tv);
24164 		if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
24165 			/*
24166 			 * Pacing timer is up, check if we are ready.
24167 			 */
24168 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
24169 				return (true);
24170 		} else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
24171 			/*
24172 			 * A timer is up, check if we are ready.
24173 			 */
24174 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
24175 				return (true);
24176 		}
24177 	}
24178 	return (false);
24179 }
24180 
24181 static struct tcp_function_block __tcp_rack = {
24182 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
24183 	.tfb_tcp_output = rack_output,
24184 	.tfb_do_queued_segments = ctf_do_queued_segments,
24185 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
24186 	.tfb_tcp_do_segment = rack_do_segment,
24187 	.tfb_tcp_ctloutput = rack_ctloutput,
24188 	.tfb_tcp_fb_init = rack_init,
24189 	.tfb_tcp_fb_fini = rack_fini,
24190 	.tfb_tcp_timer_stop_all = rack_stopall,
24191 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
24192 	.tfb_tcp_handoff_ok = rack_handoff_ok,
24193 	.tfb_tcp_mtu_chg = rack_mtu_change,
24194 	.tfb_pru_options = rack_pru_options,
24195 	.tfb_hwtls_change = rack_hw_tls_change,
24196 	.tfb_chg_query = rack_chg_query,
24197 	.tfb_switch_failed = rack_switch_failed,
24198 	.tfb_early_wake_check = rack_wake_check,
24199 	.tfb_compute_pipe = rack_compute_pipe,
24200 	.tfb_stack_info = rack_stack_information,
24201 	.tfb_inherit = rack_inherit,
24202 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP | TCP_FUNC_DEFAULT_OK,
24203 
24204 };
24205 
24206 /*
24207  * rack_ctloutput() must drop the inpcb lock before performing copyin on
24208  * socket option arguments.  When it re-acquires the lock after the copy, it
24209  * has to revalidate that the connection is still valid for the socket
24210  * option.
24211  */
24212 static int
24213 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24214 {
24215 	struct inpcb *inp = tptoinpcb(tp);
24216 #ifdef INET
24217 	struct ip *ip;
24218 #endif
24219 	struct tcp_rack *rack;
24220 	struct tcp_hybrid_req hybrid;
24221 	uint64_t loptval;
24222 	int32_t error = 0, optval;
24223 
24224 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24225 	if (rack == NULL) {
24226 		INP_WUNLOCK(inp);
24227 		return (EINVAL);
24228 	}
24229 #ifdef INET
24230 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
24231 #endif
24232 
24233 	switch (sopt->sopt_level) {
24234 #ifdef INET6
24235 	case IPPROTO_IPV6:
24236 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
24237 		switch (sopt->sopt_name) {
24238 		case IPV6_USE_MIN_MTU:
24239 			tcp6_use_min_mtu(tp);
24240 			break;
24241 		}
24242 		INP_WUNLOCK(inp);
24243 		return (0);
24244 #endif
24245 #ifdef INET
24246 	case IPPROTO_IP:
24247 		switch (sopt->sopt_name) {
24248 		case IP_TOS:
24249 			/*
24250 			 * The DSCP codepoint has changed, update the fsb.
24251 			 */
24252 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
24253 			break;
24254 		case IP_TTL:
24255 			/*
24256 			 * The TTL has changed, update the fsb.
24257 			 */
24258 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
24259 			break;
24260 		}
24261 		INP_WUNLOCK(inp);
24262 		return (0);
24263 #endif
24264 #ifdef SO_PEERPRIO
24265 	case SOL_SOCKET:
24266 		switch (sopt->sopt_name) {
24267 		case SO_PEERPRIO:			/*  SC-URL:bs */
24268 			/* Already read in and sanity checked in sosetopt(). */
24269 			if (inp->inp_socket) {
24270 				rack->client_bufferlvl = inp->inp_socket->so_peerprio;
24271 			}
24272 			break;
24273 		}
24274 		INP_WUNLOCK(inp);
24275 		return (0);
24276 #endif
24277 	case IPPROTO_TCP:
24278 		switch (sopt->sopt_name) {
24279 		case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
24280 		/*  Pacing related ones */
24281 		case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
24282 		case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
24283 		case TCP_RACK_PACE_MIN_SEG:		/*  URL:pace_min_seg */
24284 		case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
24285 		case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
24286 		case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
24287 		case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
24288 		case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
24289 		case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
24290 		case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
24291 		case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
24292 		case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
24293 		case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
24294 		case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
24295 		case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
24296 		case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
24297 		case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
24298 		case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
24299 		case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
24300 			/* End pacing related */
24301 		case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
24302 		case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
24303 		case TCP_RACK_MIN_TO:			/*  URL:min_to */
24304 		case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
24305 		case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
24306 		case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
24307 		case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
24308 		case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
24309 		case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
24310 		case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
24311 		case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
24312 		case TCP_NO_PRR:			/*  URL:noprr */
24313 		case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
24314 		case TCP_DATA_AFTER_CLOSE:		/*  no URL */
24315 		case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
24316 		case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
24317 		case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
24318 		case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
24319 		case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
24320 		case TCP_RACK_PROFILE:			/*  URL:profile */
24321 		case TCP_SIDECHAN_DIS:			/*  URL:scodm */
24322 		case TCP_HYBRID_PACING:			/*  URL:pacing=hybrid */
24323 		case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
24324 		case TCP_RACK_ABC_VAL:			/*  URL:labc */
24325 		case TCP_REC_ABC_VAL:			/*  URL:reclabc */
24326 		case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
24327 		case TCP_DEFER_OPTIONS:			/*  URL:defer */
24328 		case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
24329 		case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
24330 		case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
24331 		case TCP_RACK_SET_RXT_OPTIONS:		/*  URL:rxtsz */
24332 		case TCP_RACK_HI_BETA:			/*  URL:hibeta */
24333 		case TCP_RACK_SPLIT_LIMIT:		/*  URL:split */
24334 		case TCP_SS_EEXIT:			/*  URL:eexit */
24335 		case TCP_DGP_UPPER_BOUNDS:		/*  URL:upper */
24336 		case TCP_RACK_PACING_DIVISOR:		/*  URL:divisor */
24337 		case TCP_PACING_DND:			/*  URL:dnd */
24338 		case TCP_NO_TIMELY:			/*  URL:notimely */
24339 		case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
24340 		case TCP_HONOR_HPTS_MIN:		/*  URL:hptsmin */
24341 		case TCP_REC_IS_DYN:			/*  URL:dynrec */
24342 		case TCP_GP_USE_LTBW:			/*  URL:useltbw */
24343 			goto process_opt;
24344 			break;
24345 		default:
24346 			/* Filter off all unknown options to the base stack */
24347 			return (tcp_default_ctloutput(tp, sopt));
24348 			break;
24349 		}
24350 	default:
24351 		INP_WUNLOCK(inp);
24352 		return (0);
24353 	}
24354 process_opt:
24355 	INP_WUNLOCK(inp);
24356 	if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24357 	    (sopt->sopt_name == TCP_FILLCW_RATE_CAP)) {
24358 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
24359 		/*
24360 		 * We truncate it down to 32 bits for the socket-option trace this
24361 		 * means rates > 34Gbps won't show right, but thats probably ok.
24362 		 */
24363 		optval = (uint32_t)loptval;
24364 	} else if (sopt->sopt_name == TCP_HYBRID_PACING) {
24365 		error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
24366 	} else {
24367 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
24368 		/* Save it in 64 bit form too */
24369 		loptval = optval;
24370 	}
24371 	if (error)
24372 		return (error);
24373 	INP_WLOCK(inp);
24374 	if (tp->t_fb != &__tcp_rack) {
24375 		INP_WUNLOCK(inp);
24376 		return (ENOPROTOOPT);
24377 	}
24378 	if (rack->defer_options && (rack->gp_ready == 0) &&
24379 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
24380 	    (sopt->sopt_name != TCP_HYBRID_PACING) &&
24381 	    (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
24382 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
24383 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
24384 		/* Options are being deferred */
24385 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
24386 			INP_WUNLOCK(inp);
24387 			return (0);
24388 		} else {
24389 			/* No memory to defer, fail */
24390 			INP_WUNLOCK(inp);
24391 			return (ENOMEM);
24392 		}
24393 	}
24394 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
24395 	INP_WUNLOCK(inp);
24396 	return (error);
24397 }
24398 
24399 static void
24400 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
24401 {
24402 
24403 	INP_WLOCK_ASSERT(tptoinpcb(tp));
24404 	bzero(ti, sizeof(*ti));
24405 
24406 	ti->tcpi_state = tp->t_state;
24407 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
24408 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
24409 	if (tp->t_flags & TF_SACK_PERMIT)
24410 		ti->tcpi_options |= TCPI_OPT_SACK;
24411 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
24412 		ti->tcpi_options |= TCPI_OPT_WSCALE;
24413 		ti->tcpi_snd_wscale = tp->snd_scale;
24414 		ti->tcpi_rcv_wscale = tp->rcv_scale;
24415 	}
24416 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
24417 		ti->tcpi_options |= TCPI_OPT_ECN;
24418 	if (tp->t_flags & TF_FASTOPEN)
24419 		ti->tcpi_options |= TCPI_OPT_TFO;
24420 	/* still kept in ticks is t_rcvtime */
24421 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
24422 	/* Since we hold everything in precise useconds this is easy */
24423 	ti->tcpi_rtt = tp->t_srtt;
24424 	ti->tcpi_rttvar = tp->t_rttvar;
24425 	ti->tcpi_rto = tp->t_rxtcur;
24426 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
24427 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
24428 	/*
24429 	 * FreeBSD-specific extension fields for tcp_info.
24430 	 */
24431 	ti->tcpi_rcv_space = tp->rcv_wnd;
24432 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
24433 	ti->tcpi_snd_wnd = tp->snd_wnd;
24434 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
24435 	ti->tcpi_snd_nxt = tp->snd_nxt;
24436 	ti->tcpi_snd_mss = tp->t_maxseg;
24437 	ti->tcpi_rcv_mss = tp->t_maxseg;
24438 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
24439 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
24440 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
24441 	ti->tcpi_total_tlp = tp->t_sndtlppack;
24442 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
24443 	ti->tcpi_rttmin = tp->t_rttlow;
24444 #ifdef NETFLIX_STATS
24445 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
24446 #endif
24447 #ifdef TCP_OFFLOAD
24448 	if (tp->t_flags & TF_TOE) {
24449 		ti->tcpi_options |= TCPI_OPT_TOE;
24450 		tcp_offload_tcp_info(tp, ti);
24451 	}
24452 #endif
24453 }
24454 
24455 static int
24456 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24457 {
24458 	struct inpcb *inp = tptoinpcb(tp);
24459 	struct tcp_rack *rack;
24460 	int32_t error, optval;
24461 	uint64_t val, loptval;
24462 	struct	tcp_info ti;
24463 	/*
24464 	 * Because all our options are either boolean or an int, we can just
24465 	 * pull everything into optval and then unlock and copy. If we ever
24466 	 * add a option that is not a int, then this will have quite an
24467 	 * impact to this routine.
24468 	 */
24469 	error = 0;
24470 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24471 	if (rack == NULL) {
24472 		INP_WUNLOCK(inp);
24473 		return (EINVAL);
24474 	}
24475 	switch (sopt->sopt_name) {
24476 	case TCP_INFO:
24477 		/* First get the info filled */
24478 		rack_fill_info(tp, &ti);
24479 		/* Fix up the rtt related fields if needed */
24480 		INP_WUNLOCK(inp);
24481 		error = sooptcopyout(sopt, &ti, sizeof ti);
24482 		return (error);
24483 	/*
24484 	 * Beta is the congestion control value for NewReno that influences how
24485 	 * much of a backoff happens when loss is detected. It is normally set
24486 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
24487 	 * when you exit recovery.
24488 	 */
24489 	case TCP_RACK_PACING_BETA:
24490 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24491 			error = EINVAL;
24492 		else if (rack->rc_pacing_cc_set == 0)
24493 			optval = rack->r_ctl.rc_saved_beta.beta;
24494 		else {
24495 			/*
24496 			 * Reach out into the CC data and report back what
24497 			 * I have previously set. Yeah it looks hackish but
24498 			 * we don't want to report the saved values.
24499 			 */
24500 			if (tp->t_ccv.cc_data)
24501 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
24502 			else
24503 				error = EINVAL;
24504 		}
24505 		break;
24506 	/*
24507 	 * Beta_ecn is the congestion control value for NewReno that influences how
24508 	 * much of a backoff happens when a ECN mark is detected. It is normally set
24509 	 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
24510 	 * you exit recovery. Note that classic ECN has a beta of 50, it is only
24511 	 * ABE Ecn that uses this "less" value, but we do too with pacing :)
24512 	 */
24513 	case TCP_RACK_PACING_BETA_ECN:
24514 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24515 			error = EINVAL;
24516 		else if (rack->rc_pacing_cc_set == 0)
24517 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
24518 		else {
24519 			/*
24520 			 * Reach out into the CC data and report back what
24521 			 * I have previously set. Yeah it looks hackish but
24522 			 * we don't want to report the saved values.
24523 			 */
24524 			if (tp->t_ccv.cc_data)
24525 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
24526 			else
24527 				error = EINVAL;
24528 		}
24529 		break;
24530 	case TCP_RACK_DSACK_OPT:
24531 		optval = 0;
24532 		if (rack->rc_rack_tmr_std_based) {
24533 			optval |= 1;
24534 		}
24535 		if (rack->rc_rack_use_dsack) {
24536 			optval |= 2;
24537 		}
24538 		break;
24539 	case TCP_RACK_ENABLE_HYSTART:
24540 	{
24541 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24542 			optval = RACK_HYSTART_ON;
24543 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
24544 				optval = RACK_HYSTART_ON_W_SC;
24545 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
24546 				optval = RACK_HYSTART_ON_W_SC_C;
24547 		} else {
24548 			optval = RACK_HYSTART_OFF;
24549 		}
24550 	}
24551 	break;
24552 	case TCP_RACK_DGP_IN_REC:
24553 		error = EINVAL;
24554 		break;
24555 	case TCP_RACK_HI_BETA:
24556 		optval = rack->rack_hibeta;
24557 		break;
24558 	case TCP_DEFER_OPTIONS:
24559 		optval = rack->defer_options;
24560 		break;
24561 	case TCP_RACK_MEASURE_CNT:
24562 		optval = rack->r_ctl.req_measurements;
24563 		break;
24564 	case TCP_REC_ABC_VAL:
24565 		optval = rack->r_use_labc_for_rec;
24566 		break;
24567 	case TCP_RACK_ABC_VAL:
24568 		optval = rack->rc_labc;
24569 		break;
24570 	case TCP_HDWR_UP_ONLY:
24571 		optval= rack->r_up_only;
24572 		break;
24573 	case TCP_FILLCW_RATE_CAP:
24574 		loptval = rack->r_ctl.fillcw_cap;
24575 		break;
24576 	case TCP_PACING_RATE_CAP:
24577 		loptval = rack->r_ctl.bw_rate_cap;
24578 		break;
24579 	case TCP_RACK_PROFILE:
24580 		/* You cannot retrieve a profile, its write only */
24581 		error = EINVAL;
24582 		break;
24583 	case TCP_SIDECHAN_DIS:
24584 		optval = rack->r_ctl.side_chan_dis_mask;
24585 		break;
24586 	case TCP_HYBRID_PACING:
24587 		/* You cannot retrieve hybrid pacing information, its write only */
24588 		error = EINVAL;
24589 		break;
24590 	case TCP_USE_CMP_ACKS:
24591 		optval = rack->r_use_cmp_ack;
24592 		break;
24593 	case TCP_RACK_PACE_TO_FILL:
24594 		optval = rack->rc_pace_to_cwnd;
24595 		break;
24596 	case TCP_RACK_NO_PUSH_AT_MAX:
24597 		optval = rack->r_ctl.rc_no_push_at_mrtt;
24598 		break;
24599 	case TCP_SHARED_CWND_ENABLE:
24600 		optval = rack->rack_enable_scwnd;
24601 		break;
24602 	case TCP_RACK_NONRXT_CFG_RATE:
24603 		optval = rack->rack_rec_nonrxt_use_cr;
24604 		break;
24605 	case TCP_NO_PRR:
24606 		if (rack->rack_no_prr  == 1)
24607 			optval = 1;
24608 		else if (rack->no_prr_addback == 1)
24609 			optval = 2;
24610 		else
24611 			optval = 0;
24612 		break;
24613 	case TCP_GP_USE_LTBW:
24614 		if (rack->dis_lt_bw) {
24615 			/* It is not used */
24616 			optval = 0;
24617 		} else if (rack->use_lesser_lt_bw) {
24618 			/* we use min() */
24619 			optval = 1;
24620 		} else {
24621 			/* we use max() */
24622 			optval = 2;
24623 		}
24624 		break;
24625 	case TCP_RACK_DO_DETECTION:
24626 		error = EINVAL;
24627 		break;
24628 	case TCP_RACK_MBUF_QUEUE:
24629 		/* Now do we use the LRO mbuf-queue feature */
24630 		optval = rack->r_mbuf_queue;
24631 		break;
24632 	case RACK_CSPR_IS_FCC:
24633 		optval = rack->cspr_is_fcc;
24634 		break;
24635 	case TCP_TIMELY_DYN_ADJ:
24636 		optval = rack->rc_gp_dyn_mul;
24637 		break;
24638 	case TCP_BBR_IWINTSO:
24639 		error = EINVAL;
24640 		break;
24641 	case TCP_RACK_TLP_REDUCE:
24642 		/* RACK TLP cwnd reduction (bool) */
24643 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
24644 		break;
24645 	case TCP_BBR_RACK_INIT_RATE:
24646 		val = rack->r_ctl.init_rate;
24647 		/* convert to kbits per sec */
24648 		val *= 8;
24649 		val /= 1000;
24650 		optval = (uint32_t)val;
24651 		break;
24652 	case TCP_RACK_FORCE_MSEG:
24653 		optval = rack->rc_force_max_seg;
24654 		break;
24655 	case TCP_RACK_PACE_MIN_SEG:
24656 		optval = rack->r_ctl.rc_user_set_min_segs;
24657 		break;
24658 	case TCP_RACK_PACE_MAX_SEG:
24659 		/* Max segments in a pace */
24660 		optval = rack->rc_user_set_max_segs;
24661 		break;
24662 	case TCP_RACK_PACE_ALWAYS:
24663 		/* Use the always pace method */
24664 		optval = rack->rc_always_pace;
24665 		break;
24666 	case TCP_RACK_PRR_SENDALOT:
24667 		/* Allow PRR to send more than one seg */
24668 		optval = rack->r_ctl.rc_prr_sendalot;
24669 		break;
24670 	case TCP_RACK_MIN_TO:
24671 		/* Minimum time between rack t-o's in ms */
24672 		optval = rack->r_ctl.rc_min_to;
24673 		break;
24674 	case TCP_RACK_SPLIT_LIMIT:
24675 		optval = rack->r_ctl.rc_split_limit;
24676 		break;
24677 	case TCP_RACK_EARLY_SEG:
24678 		/* If early recovery max segments */
24679 		optval = rack->r_ctl.rc_early_recovery_segs;
24680 		break;
24681 	case TCP_RACK_REORD_THRESH:
24682 		/* RACK reorder threshold (shift amount) */
24683 		optval = rack->r_ctl.rc_reorder_shift;
24684 		break;
24685 	case TCP_SS_EEXIT:
24686 		if (rack->r_ctl.gp_rnd_thresh) {
24687 			uint32_t v;
24688 
24689 			v = rack->r_ctl.gp_gain_req;
24690 			v <<= 17;
24691 			optval = v | (rack->r_ctl.gp_rnd_thresh & 0xff);
24692 			if (rack->r_ctl.gate_to_fs == 1)
24693 				optval |= 0x10000;
24694 		} else
24695 			optval = 0;
24696 		break;
24697 	case TCP_RACK_REORD_FADE:
24698 		/* Does reordering fade after ms time */
24699 		optval = rack->r_ctl.rc_reorder_fade;
24700 		break;
24701 	case TCP_BBR_USE_RACK_RR:
24702 		/* Do we use the rack cheat for rxt */
24703 		optval = rack->use_rack_rr;
24704 		break;
24705 	case TCP_RACK_RR_CONF:
24706 		optval = rack->r_rr_config;
24707 		break;
24708 	case TCP_HDWR_RATE_CAP:
24709 		optval = rack->r_rack_hw_rate_caps;
24710 		break;
24711 	case TCP_BBR_HDWR_PACE:
24712 		optval = rack->rack_hdw_pace_ena;
24713 		break;
24714 	case TCP_RACK_TLP_THRESH:
24715 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
24716 		optval = rack->r_ctl.rc_tlp_threshold;
24717 		break;
24718 	case TCP_RACK_PKT_DELAY:
24719 		/* RACK added ms i.e. rack-rtt + reord + N */
24720 		optval = rack->r_ctl.rc_pkt_delay;
24721 		break;
24722 	case TCP_RACK_TLP_USE:
24723 		optval = rack->rack_tlp_threshold_use;
24724 		break;
24725 	case TCP_PACING_DND:
24726 		optval = rack->rc_pace_dnd;
24727 		break;
24728 	case TCP_RACK_PACE_RATE_CA:
24729 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
24730 		break;
24731 	case TCP_RACK_PACE_RATE_SS:
24732 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
24733 		break;
24734 	case TCP_RACK_PACE_RATE_REC:
24735 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
24736 		break;
24737 	case TCP_DGP_UPPER_BOUNDS:
24738 		optval = rack->r_ctl.rack_per_upper_bound_ss;
24739 		optval <<= 16;
24740 		optval |= rack->r_ctl.rack_per_upper_bound_ca;
24741 		break;
24742 	case TCP_RACK_GP_INCREASE_SS:
24743 		optval = rack->r_ctl.rack_per_of_gp_ca;
24744 		break;
24745 	case TCP_RACK_GP_INCREASE_CA:
24746 		optval = rack->r_ctl.rack_per_of_gp_ss;
24747 		break;
24748 	case TCP_RACK_PACING_DIVISOR:
24749 		optval = rack->r_ctl.pace_len_divisor;
24750 		break;
24751 	case TCP_BBR_RACK_RTT_USE:
24752 		optval = rack->r_ctl.rc_rate_sample_method;
24753 		break;
24754 	case TCP_DELACK:
24755 		optval = tp->t_delayed_ack;
24756 		break;
24757 	case TCP_DATA_AFTER_CLOSE:
24758 		optval = rack->rc_allow_data_af_clo;
24759 		break;
24760 	case TCP_SHARED_CWND_TIME_LIMIT:
24761 		optval = rack->r_limit_scw;
24762 		break;
24763 	case TCP_HONOR_HPTS_MIN:
24764 		if (rack->r_use_hpts_min)
24765 			optval = rack->r_ctl.max_reduction;
24766 		else
24767 			optval = 0;
24768 		break;
24769 	case TCP_REC_IS_DYN:
24770 		optval = rack->rc_gp_no_rec_chg;
24771 		break;
24772 	case TCP_NO_TIMELY:
24773 		optval = rack->rc_skip_timely;
24774 		break;
24775 	case TCP_RACK_TIMER_SLOP:
24776 		optval = rack->r_ctl.timer_slop;
24777 		break;
24778 	default:
24779 		return (tcp_default_ctloutput(tp, sopt));
24780 		break;
24781 	}
24782 	INP_WUNLOCK(inp);
24783 	if (error == 0) {
24784 		if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24785 		    (sopt->sopt_name == TCP_FILLCW_RATE_CAP))
24786 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
24787 		else
24788 			error = sooptcopyout(sopt, &optval, sizeof optval);
24789 	}
24790 	return (error);
24791 }
24792 
24793 static int
24794 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
24795 {
24796 	if (sopt->sopt_dir == SOPT_SET) {
24797 		return (rack_set_sockopt(tp, sopt));
24798 	} else if (sopt->sopt_dir == SOPT_GET) {
24799 		return (rack_get_sockopt(tp, sopt));
24800 	} else {
24801 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
24802 	}
24803 }
24804 
24805 static const char *rack_stack_names[] = {
24806 	__XSTRING(STACKNAME),
24807 #ifdef STACKALIAS
24808 	__XSTRING(STACKALIAS),
24809 #endif
24810 };
24811 
24812 static int
24813 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
24814 {
24815 	memset(mem, 0, size);
24816 	return (0);
24817 }
24818 
24819 static void
24820 rack_dtor(void *mem, int32_t size, void *arg)
24821 {
24822 
24823 }
24824 
24825 static bool rack_mod_inited = false;
24826 
24827 static int
24828 tcp_addrack(module_t mod, int32_t type, void *data)
24829 {
24830 	int32_t err = 0;
24831 	int num_stacks;
24832 
24833 	switch (type) {
24834 	case MOD_LOAD:
24835 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24836 		    sizeof(struct rack_sendmap),
24837 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24838 
24839 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24840 		    sizeof(struct tcp_rack),
24841 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24842 
24843 		sysctl_ctx_init(&rack_sysctl_ctx);
24844 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24845 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24846 		    OID_AUTO,
24847 #ifdef STACKALIAS
24848 		    __XSTRING(STACKALIAS),
24849 #else
24850 		    __XSTRING(STACKNAME),
24851 #endif
24852 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24853 		    "");
24854 		if (rack_sysctl_root == NULL) {
24855 			printf("Failed to add sysctl node\n");
24856 			err = EFAULT;
24857 			goto free_uma;
24858 		}
24859 		rack_init_sysctls();
24860 		num_stacks = nitems(rack_stack_names);
24861 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24862 		    rack_stack_names, &num_stacks);
24863 		if (err) {
24864 			printf("Failed to register %s stack name for "
24865 			    "%s module\n", rack_stack_names[num_stacks],
24866 			    __XSTRING(MODNAME));
24867 			sysctl_ctx_free(&rack_sysctl_ctx);
24868 free_uma:
24869 			uma_zdestroy(rack_zone);
24870 			uma_zdestroy(rack_pcb_zone);
24871 			rack_counter_destroy();
24872 			printf("Failed to register rack module -- err:%d\n", err);
24873 			return (err);
24874 		}
24875 		tcp_lro_reg_mbufq();
24876 		rack_mod_inited = true;
24877 		break;
24878 	case MOD_QUIESCE:
24879 		err = deregister_tcp_functions(&__tcp_rack, true, false);
24880 		break;
24881 	case MOD_UNLOAD:
24882 		err = deregister_tcp_functions(&__tcp_rack, false, true);
24883 		if (err == EBUSY)
24884 			break;
24885 		if (rack_mod_inited) {
24886 			uma_zdestroy(rack_zone);
24887 			uma_zdestroy(rack_pcb_zone);
24888 			sysctl_ctx_free(&rack_sysctl_ctx);
24889 			rack_counter_destroy();
24890 			rack_mod_inited = false;
24891 		}
24892 		tcp_lro_dereg_mbufq();
24893 		err = 0;
24894 		break;
24895 	default:
24896 		return (EOPNOTSUPP);
24897 	}
24898 	return (err);
24899 }
24900 
24901 static moduledata_t tcp_rack = {
24902 	.name = __XSTRING(MODNAME),
24903 	.evhand = tcp_addrack,
24904 	.priv = 0
24905 };
24906 
24907 MODULE_VERSION(MODNAME, 1);
24908 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24909 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24910 
24911 #endif /* #if !defined(INET) && !defined(INET6) */
24912