xref: /freebsd/sys/netinet/tcp_stacks/bbr.c (revision 580744621f33383027108364dcadad718df46ffe) 
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  * Author: Randall Stewart <rrs@netflix.com>
28  * This work is based on the ACM Queue paper
29  * BBR - Congestion Based Congestion Control
30  * and also numerous discussions with Neal, Yuchung and Van.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_tcpdebug.h"
40 #include "opt_ratelimit.h"
41 #include "opt_kern_tls.h"
42 #include <sys/param.h>
43 #include <sys/arb.h>
44 #include <sys/module.h>
45 #include <sys/kernel.h>
46 #ifdef TCP_HHOOK
47 #include <sys/hhook.h>
48 #endif
49 #include <sys/malloc.h>
50 #include <sys/mbuf.h>
51 #include <sys/proc.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #ifdef KERN_TLS
55 #include <sys/ktls.h>
56 #endif
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
59 #ifdef STATS
60 #include <sys/qmath.h>
61 #include <sys/tree.h>
62 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
63 #endif
64 #include <sys/refcount.h>
65 #include <sys/queue.h>
66 #include <sys/eventhandler.h>
67 #include <sys/smp.h>
68 #include <sys/kthread.h>
69 #include <sys/lock.h>
70 #include <sys/mutex.h>
71 #include <sys/tim_filter.h>
72 #include <sys/time.h>
73 #include <sys/protosw.h>
74 #include <vm/uma.h>
75 #include <sys/kern_prefetch.h>
76 
77 #include <net/route.h>
78 #include <net/route/nhop.h>
79 #include <net/vnet.h>
80 
81 #define TCPSTATES		/* for logging */
82 
83 #include <netinet/in.h>
84 #include <netinet/in_kdtrace.h>
85 #include <netinet/in_pcb.h>
86 #include <netinet/ip.h>
87 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
88 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
89 #include <netinet/ip_var.h>
90 #include <netinet/ip6.h>
91 #include <netinet6/in6_pcb.h>
92 #include <netinet6/ip6_var.h>
93 #define	TCPOUTFLAGS
94 #include <netinet/tcp.h>
95 #include <netinet/tcp_fsm.h>
96 #include <netinet/tcp_seq.h>
97 #include <netinet/tcp_timer.h>
98 #include <netinet/tcp_var.h>
99 #include <netinet/tcpip.h>
100 #include <netinet/tcp_hpts.h>
101 #include <netinet/cc/cc.h>
102 #include <netinet/tcp_log_buf.h>
103 #include <netinet/tcp_ratelimit.h>
104 #include <netinet/tcp_lro.h>
105 #ifdef TCPDEBUG
106 #include <netinet/tcp_debug.h>
107 #endif				/* TCPDEBUG */
108 #ifdef TCP_OFFLOAD
109 #include <netinet/tcp_offload.h>
110 #endif
111 #ifdef INET6
112 #include <netinet6/tcp6_var.h>
113 #endif
114 #include <netinet/tcp_fastopen.h>
115 
116 #include <netipsec/ipsec_support.h>
117 #include <net/if.h>
118 #include <net/if_var.h>
119 #include <net/ethernet.h>
120 
121 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
122 #include <netipsec/ipsec.h>
123 #include <netipsec/ipsec6.h>
124 #endif				/* IPSEC */
125 
126 #include <netinet/udp.h>
127 #include <netinet/udp_var.h>
128 #include <machine/in_cksum.h>
129 
130 #ifdef MAC
131 #include <security/mac/mac_framework.h>
132 #endif
133 
134 #include "sack_filter.h"
135 #include "tcp_bbr.h"
136 #include "rack_bbr_common.h"
137 uma_zone_t bbr_zone;
138 uma_zone_t bbr_pcb_zone;
139 
140 struct sysctl_ctx_list bbr_sysctl_ctx;
141 struct sysctl_oid *bbr_sysctl_root;
142 
143 #define	TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \
144 	(tv) = (value); \
145 	if ((u_long)(tv) < (u_long)(tvmin)) \
146 		(tv) = (tvmin); \
147 	if ((u_long)(tv) > (u_long)(tvmax)) \
148 		(tv) = (tvmax); \
149 } while(0)
150 
151 /*#define BBR_INVARIANT 1*/
152 
153 /*
154  * initial window
155  */
156 static uint32_t bbr_def_init_win = 10;
157 static int32_t bbr_persist_min = 250000;	/* 250ms */
158 static int32_t bbr_persist_max = 1000000;	/* 1 Second */
159 static int32_t bbr_cwnd_may_shrink = 0;
160 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP;
161 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT;
162 static int32_t bbr_hardware_pacing_limit = 8000;
163 static int32_t bbr_quanta = 3;	/* How much extra quanta do we get? */
164 static int32_t bbr_no_retran = 0;
165 
166 
167 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */
168 static int32_t bbr_max_net_error_cnt = 10;
169 /* Should the following be dynamic too -- loss wise */
170 static int32_t bbr_rtt_gain_thresh = 0;
171 /* Measurement controls */
172 static int32_t bbr_use_google_algo = 1;
173 static int32_t bbr_ts_limiting = 1;
174 static int32_t bbr_ts_can_raise = 0;
175 static int32_t bbr_do_red = 600;
176 static int32_t bbr_red_scale = 20000;
177 static int32_t bbr_red_mul = 1;
178 static int32_t bbr_red_div = 2;
179 static int32_t bbr_red_growth_restrict = 1;
180 static int32_t  bbr_target_is_bbunit = 0;
181 static int32_t bbr_drop_limit = 0;
182 /*
183  * How much gain do we need to see to
184  * stay in startup?
185  */
186 static int32_t bbr_marks_rxt_sack_passed = 0;
187 static int32_t bbr_start_exit = 25;
188 static int32_t bbr_low_start_exit = 25;	/* When we are in reduced gain */
189 static int32_t bbr_startup_loss_thresh = 2000;	/* 20.00% loss */
190 static int32_t bbr_hptsi_max_mul = 1;	/* These two mul/div assure a min pacing */
191 static int32_t bbr_hptsi_max_div = 2;	/* time, 0 means turned off. We need this
192 					 * if we go back ever to where the pacer
193 					 * has priority over timers.
194 					 */
195 static int32_t bbr_policer_call_from_rack_to = 0;
196 static int32_t bbr_policer_detection_enabled = 1;
197 static int32_t bbr_min_measurements_req = 1;	/* We need at least 2
198 						 * measurments before we are
199 						 * "good" note that 2 == 1.
200 						 * This is because we use a >
201 						 * comparison. This means if
202 						 * min_measure was 0, it takes
203 						 * num-measures > min(0) and
204 						 * you get 1 measurement and
205 						 * you are good. Set to 1, you
206 						 * have to have two
207 						 * measurements (this is done
208 						 * to prevent it from being ok
209 						 * to have no measurements). */
210 static int32_t bbr_no_pacing_until = 4;
211 
212 static int32_t bbr_min_usec_delta = 20000;	/* 20,000 usecs */
213 static int32_t bbr_min_peer_delta = 20;		/* 20 units */
214 static int32_t bbr_delta_percent = 150;		/* 15.0 % */
215 
216 static int32_t bbr_target_cwnd_mult_limit = 8;
217 /*
218  * bbr_cwnd_min_val is the number of
219  * segments we hold to in the RTT probe
220  * state typically 4.
221  */
222 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS;
223 
224 
225 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS;
226 
227 static int32_t bbr_gain_to_target = 1;
228 static int32_t bbr_gain_gets_extra_too = 1;
229 /*
230  * bbr_high_gain is the 2/ln(2) value we need
231  * to double the sending rate in startup. This
232  * is used for both cwnd and hptsi gain's.
233  */
234 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1;
235 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1;
236 static int32_t bbr_use_lower_gain_in_startup = 1;
237 
238 /* thresholds for reduction on drain in sub-states/drain */
239 static int32_t bbr_drain_rtt = BBR_SRTT;
240 static int32_t bbr_drain_floor = 88;
241 static int32_t google_allow_early_out = 1;
242 static int32_t google_consider_lost = 1;
243 static int32_t bbr_drain_drop_mul = 4;
244 static int32_t bbr_drain_drop_div = 5;
245 static int32_t bbr_rand_ot = 50;
246 static int32_t bbr_can_force_probertt = 0;
247 static int32_t bbr_can_adjust_probertt = 1;
248 static int32_t bbr_probertt_sets_rtt = 0;
249 static int32_t bbr_can_use_ts_for_rtt = 1;
250 static int32_t bbr_is_ratio = 0;
251 static int32_t bbr_sub_drain_app_limit = 1;
252 static int32_t bbr_prtt_slam_cwnd = 1;
253 static int32_t bbr_sub_drain_slam_cwnd = 1;
254 static int32_t bbr_slam_cwnd_in_main_drain = 1;
255 static int32_t bbr_filter_len_sec = 6;	/* How long does the rttProp filter
256 					 * hold */
257 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4);
258 /*
259  * bbr_drain_gain is the reverse of the high_gain
260  * designed to drain back out the standing queue
261  * that is formed in startup by causing a larger
262  * hptsi gain and thus drainging the packets
263  * in flight.
264  */
265 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885;
266 static int32_t bbr_rttprobe_gain = 192;
267 
268 /*
269  * The cwnd_gain is the default cwnd gain applied when
270  * calculating a target cwnd. Note that the cwnd is
271  * a secondary factor in the way BBR works (see the
272  * paper and think about it, it will take some time).
273  * Basically the hptsi_gain spreads the packets out
274  * so you never get more than BDP to the peer even
275  * if the cwnd is high. In our implemenation that
276  * means in non-recovery/retransmission scenarios
277  * cwnd will never be reached by the flight-size.
278  */
279 static int32_t bbr_cwnd_gain = BBR_UNIT * 2;
280 static int32_t bbr_tlp_type_to_use = BBR_SRTT;
281 static int32_t bbr_delack_time = 100000;	/* 100ms in useconds */
282 static int32_t bbr_sack_not_required = 0;	/* set to one to allow non-sack to use bbr */
283 static int32_t bbr_initial_bw_bps = 62500;	/* 500kbps in bytes ps */
284 static int32_t bbr_ignore_data_after_close = 1;
285 static int16_t bbr_hptsi_gain[] = {
286 	(BBR_UNIT *5 / 4),
287 	(BBR_UNIT * 3 / 4),
288 	BBR_UNIT,
289 	BBR_UNIT,
290 	BBR_UNIT,
291 	BBR_UNIT,
292 	BBR_UNIT,
293 	BBR_UNIT
294 };
295 int32_t bbr_use_rack_resend_cheat = 1;
296 int32_t bbr_sends_full_iwnd = 1;
297 
298 #define BBR_HPTSI_GAIN_MAX 8
299 /*
300  * The BBR module incorporates a number of
301  * TCP ideas that have been put out into the IETF
302  * over the last few years:
303  * - Yuchung Cheng's RACK TCP (for which its named) that
304  *    will stop us using the number of dup acks and instead
305  *    use time as the gage of when we retransmit.
306  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
307  *    of Dukkipati et.al.
308  * - Van Jacobson's et.al BBR.
309  *
310  * RACK depends on SACK, so if an endpoint arrives that
311  * cannot do SACK the state machine below will shuttle the
312  * connection back to using the "default" TCP stack that is
313  * in FreeBSD.
314  *
315  * To implement BBR and RACK the original TCP stack was first decomposed
316  * into a functional state machine with individual states
317  * for each of the possible TCP connection states. The do_segement
318  * functions role in life is to mandate the connection supports SACK
319  * initially and then assure that the RACK state matches the conenction
320  * state before calling the states do_segment function. Data processing
321  * of inbound segments also now happens in the hpts_do_segment in general
322  * with only one exception. This is so we can keep the connection on
323  * a single CPU.
324  *
325  * Each state is simplified due to the fact that the original do_segment
326  * has been decomposed and we *know* what state we are in (no
327  * switches on the state) and all tests for SACK are gone. This
328  * greatly simplifies what each state does.
329  *
330  * TCP output is also over-written with a new version since it
331  * must maintain the new rack scoreboard and has had hptsi
332  * integrated as a requirment. Still todo is to eliminate the
333  * use of the callout_() system and use the hpts for all
334  * timers as well.
335  */
336 static uint32_t bbr_rtt_probe_time = 200000;	/* 200ms in micro seconds */
337 static uint32_t bbr_rtt_probe_cwndtarg = 4;	/* How many mss's outstanding */
338 static const int32_t bbr_min_req_free = 2;	/* The min we must have on the
339 						 * free list */
340 static int32_t bbr_tlp_thresh = 1;
341 static int32_t bbr_reorder_thresh = 2;
342 static int32_t bbr_reorder_fade = 60000000;	/* 0 - never fade, def
343 						 * 60,000,000 - 60 seconds */
344 static int32_t bbr_pkt_delay = 1000;
345 static int32_t bbr_min_to = 1000;	/* Number of usec's minimum timeout */
346 static int32_t bbr_incr_timers = 1;
347 
348 static int32_t bbr_tlp_min = 10000;	/* 10ms in usecs */
349 static int32_t bbr_delayed_ack_time = 200000;	/* 200ms in usecs */
350 static int32_t bbr_exit_startup_at_loss = 1;
351 
352 /*
353  * bbr_lt_bw_ratio is 1/8th
354  * bbr_lt_bw_diff is  < 4 Kbit/sec
355  */
356 static uint64_t bbr_lt_bw_diff = 4000 / 8;	/* In bytes per second */
357 static uint64_t bbr_lt_bw_ratio = 8;	/* For 1/8th */
358 static uint32_t bbr_lt_bw_max_rtts = 48;	/* How many rtt's do we use
359 						 * the lt_bw for */
360 static uint32_t bbr_lt_intvl_min_rtts = 4;	/* Min num of RTT's to measure
361 						 * lt_bw */
362 static int32_t bbr_lt_intvl_fp = 0;		/* False positive epoch diff */
363 static int32_t bbr_lt_loss_thresh = 196;	/* Lost vs delivered % */
364 static int32_t bbr_lt_fd_thresh = 100;		/* false detection % */
365 
366 static int32_t bbr_verbose_logging = 0;
367 /*
368  * Currently regular tcp has a rto_min of 30ms
369  * the backoff goes 12 times so that ends up
370  * being a total of 122.850 seconds before a
371  * connection is killed.
372  */
373 static int32_t bbr_rto_min_ms = 30;	/* 30ms same as main freebsd */
374 static int32_t bbr_rto_max_sec = 4;	/* 4 seconds */
375 
376 /****************************************************/
377 /* DEFAULT TSO SIZING  (cpu performance impacting)  */
378 /****************************************************/
379 /* What amount is our formula using to get TSO size */
380 static int32_t bbr_hptsi_per_second = 1000;
381 
382 /*
383  * For hptsi under bbr_cross_over connections what is delay
384  * target 7ms (in usec) combined with a seg_max of 2
385  * gets us close to identical google behavior in
386  * TSO size selection (possibly more 1MSS sends).
387  */
388 static int32_t bbr_hptsi_segments_delay_tar = 7000;
389 
390 /* Does pacing delay include overhead's in its time calculations? */
391 static int32_t bbr_include_enet_oh = 0;
392 static int32_t bbr_include_ip_oh = 1;
393 static int32_t bbr_include_tcp_oh = 1;
394 static int32_t bbr_google_discount = 10;
395 
396 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */
397 static int32_t bbr_state_is_pkt_epoch = 0;
398 static int32_t bbr_state_drain_2_tar = 1;
399 /* What is the max the 0 - bbr_cross_over MBPS TSO target
400  * can reach using our delay target. Note that this
401  * value becomes the floor for the cross over
402  * algorithm.
403  */
404 static int32_t bbr_hptsi_segments_max = 2;
405 static int32_t bbr_hptsi_segments_floor = 1;
406 static int32_t bbr_hptsi_utter_max = 0;
407 
408 /* What is the min the 0 - bbr_cross-over MBPS  TSO target can be */
409 static int32_t bbr_hptsi_bytes_min = 1460;
410 static int32_t bbr_all_get_min = 0;
411 
412 /* Cross over point from algo-a to algo-b */
413 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS;
414 
415 /* Do we deal with our restart state? */
416 static int32_t bbr_uses_idle_restart = 0;
417 static int32_t bbr_idle_restart_threshold = 100000;	/* 100ms in useconds */
418 
419 /* Do we allow hardware pacing? */
420 static int32_t bbr_allow_hdwr_pacing = 0;
421 static int32_t bbr_hdwr_pace_adjust = 2;	/* multipler when we calc the tso size */
422 static int32_t bbr_hdwr_pace_floor = 1;
423 static int32_t bbr_hdwr_pacing_delay_cnt = 10;
424 
425 /****************************************************/
426 static int32_t bbr_resends_use_tso = 0;
427 static int32_t bbr_tlp_max_resend = 2;
428 static int32_t bbr_sack_block_limit = 128;
429 
430 #define  BBR_MAX_STAT 19
431 counter_u64_t bbr_state_time[BBR_MAX_STAT];
432 counter_u64_t bbr_state_lost[BBR_MAX_STAT];
433 counter_u64_t bbr_state_resend[BBR_MAX_STAT];
434 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE];
435 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE];
436 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE];
437 counter_u64_t bbr_flows_whdwr_pacing;
438 counter_u64_t bbr_flows_nohdwr_pacing;
439 
440 counter_u64_t bbr_nohdwr_pacing_enobuf;
441 counter_u64_t bbr_hdwr_pacing_enobuf;
442 
443 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr);
444 
445 /*
446  * Static defintions we need for forward declarations.
447  */
448 static uint32_t
449 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain,
450     uint32_t useconds_time, uint64_t bw);
451 static uint32_t
452 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain);
453 static void
454      bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win);
455 static void
456 bbr_set_probebw_gains(struct tcp_bbr *bbr,  uint32_t cts, uint32_t losses);
457 static void
458 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line,
459 		    int dolog);
460 static uint32_t
461 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain);
462 static void
463 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch,
464 		 int32_t pkt_epoch, uint32_t losses);
465 static uint32_t
466 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm);
467 static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp);
468 static uint32_t
469 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
470     struct bbr_sendmap *rsm, uint32_t srtt,
471     uint32_t cts);
472 static void
473 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts,
474     int32_t line);
475 static void
476      bbr_set_state_target(struct tcp_bbr *bbr, int line);
477 static void
478      bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line);
479 
480 static void
481      bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line);
482 
483 static void
484      tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts);
485 
486 static void
487      bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts);
488 
489 static void
490      bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt,
491 			 uint32_t line, uint8_t is_start, uint16_t set);
492 
493 static struct bbr_sendmap *
494             bbr_find_lowest_rsm(struct tcp_bbr *bbr);
495 static __inline uint32_t
496 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type);
497 static void
498      bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which);
499 
500 static void
501 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
502     uint32_t thresh, uint32_t to);
503 static void
504      bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag);
505 
506 static void
507 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot,
508     uint32_t del_by, uint32_t cts, uint32_t sloton, uint32_t prev_delay);
509 
510 static void
511 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr,
512     uint32_t cts, int32_t line);
513 static void
514      bbr_stop_all_timers(struct tcpcb *tp);
515 static void
516      bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts);
517 static void
518      bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts);
519 static void
520      bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts);
521 
522 
523 static void
524 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
525     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod);
526 
527 static inline uint8_t
528 bbr_state_val(struct tcp_bbr *bbr)
529 {
530 	return(bbr->rc_bbr_substate);
531 }
532 
533 static inline uint32_t
534 get_min_cwnd(struct tcp_bbr *bbr)
535 {
536 	int mss;
537 
538 	mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
539 	if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED)
540 		return (bbr_cwnd_min_val_hs * mss);
541 	else
542 		return (bbr_cwnd_min_val * mss);
543 }
544 
545 static uint32_t
546 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr)
547 {
548 	uint64_t srtt, var;
549 	uint64_t ret_val;
550 
551 	bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
552 	if (tp->t_srtt == 0) {
553 		srtt = (uint64_t)BBR_INITIAL_RTO;
554 		var = 0;
555 	} else {
556 		srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
557 		var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT);
558 	}
559 	TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]),
560 	    bbr_persist_min, bbr_persist_max);
561 	return ((uint32_t)ret_val);
562 }
563 
564 static uint32_t
565 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
566 {
567 	/*
568 	 * Start the FR timer, we do this based on getting the first one in
569 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
570 	 * events we need to stop the running timer (if its running) before
571 	 * starting the new one.
572 	 */
573 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
574 	int32_t idx;
575 	int32_t is_tlp_timer = 0;
576 	struct bbr_sendmap *rsm;
577 
578 	if (bbr->rc_all_timers_stopped) {
579 		/* All timers have been stopped none are to run */
580 		return (0);
581 	}
582 	if (bbr->rc_in_persist) {
583 		/* We can't start any timer in persists */
584 		return (bbr_get_persists_timer_val(tp, bbr));
585 	}
586 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
587 	if ((rsm == NULL) ||
588 	    ((tp->t_flags & TF_SACK_PERMIT) == 0) ||
589 	    (tp->t_state < TCPS_ESTABLISHED)) {
590 		/* Nothing on the send map */
591 activate_rxt:
592 		if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
593 			uint64_t tov;
594 
595 			time_since_sent = 0;
596 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
597 			if (rsm) {
598 				idx = rsm->r_rtr_cnt - 1;
599 				if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
600 					tstmp_touse = rsm->r_tim_lastsent[idx];
601 				else
602 					tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
603 				if (TSTMP_GT(tstmp_touse, cts))
604 				    time_since_sent = cts - tstmp_touse;
605 			}
606 			bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
607 			if (tp->t_srtt == 0)
608 				tov = BBR_INITIAL_RTO;
609 			else
610 				tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) +
611 				    ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT);
612 			if (tp->t_rxtshift)
613 				tov *= tcp_backoff[tp->t_rxtshift];
614 			if (tov > time_since_sent)
615 				tov -= time_since_sent;
616 			else
617 				tov = bbr->r_ctl.rc_min_to;
618 			TCPT_RANGESET_NOSLOP(to, tov,
619 			    (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC),
620 			    (bbr->rc_max_rto_sec * USECS_IN_SECOND));
621 			bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to);
622 			return (to);
623 		}
624 		return (0);
625 	}
626 	if (rsm->r_flags & BBR_ACKED) {
627 		rsm = bbr_find_lowest_rsm(bbr);
628 		if (rsm == NULL) {
629 			/* No lowest? */
630 			goto activate_rxt;
631 		}
632 	}
633 	/* Convert from ms to usecs */
634 	if (rsm->r_flags & BBR_SACK_PASSED) {
635 		if ((tp->t_flags & TF_SENTFIN) &&
636 		    ((tp->snd_max - tp->snd_una) == 1) &&
637 		    (rsm->r_flags & BBR_HAS_FIN)) {
638 			/*
639 			 * We don't start a bbr rack timer if all we have is
640 			 * a FIN outstanding.
641 			 */
642 			goto activate_rxt;
643 		}
644 		srtt = bbr_get_rtt(bbr, BBR_RTT_RACK);
645 		thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm);
646 		idx = rsm->r_rtr_cnt - 1;
647 		exp = rsm->r_tim_lastsent[idx] + thresh;
648 		if (SEQ_GEQ(exp, cts)) {
649 			to = exp - cts;
650 			if (to < bbr->r_ctl.rc_min_to) {
651 				to = bbr->r_ctl.rc_min_to;
652 			}
653 		} else {
654 			to = bbr->r_ctl.rc_min_to;
655 		}
656 	} else {
657 		/* Ok we need to do a TLP not RACK */
658 		if (bbr->rc_tlp_in_progress != 0) {
659 			/*
660 			 * The previous send was a TLP.
661 			 */
662 			goto activate_rxt;
663 		}
664 		rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
665 		if (rsm == NULL) {
666 			/* We found no rsm to TLP with. */
667 			goto activate_rxt;
668 		}
669 		if (rsm->r_flags & BBR_HAS_FIN) {
670 			/* If its a FIN we don't do TLP */
671 			rsm = NULL;
672 			goto activate_rxt;
673 		}
674 		time_since_sent = 0;
675 		idx = rsm->r_rtr_cnt - 1;
676 		if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
677 			tstmp_touse = rsm->r_tim_lastsent[idx];
678 		else
679 			tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
680 		if (TSTMP_GT(tstmp_touse, cts))
681 		    time_since_sent = cts - tstmp_touse;
682 		is_tlp_timer = 1;
683 		srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use);
684 		thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts);
685 		if (thresh > time_since_sent)
686 			to = thresh - time_since_sent;
687 		else
688 			to = bbr->r_ctl.rc_min_to;
689 		if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
690 			/*
691 			 * If the TLP time works out to larger than the max
692 			 * RTO lets not do TLP.. just RTO.
693 			 */
694 			goto activate_rxt;
695 		}
696 		if ((bbr->rc_tlp_rtx_out == 1) &&
697 		    (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) {
698 			/*
699 			 * Second retransmit of the same TLP
700 			 * lets not.
701 			 */
702 			bbr->rc_tlp_rtx_out = 0;
703 			goto activate_rxt;
704 		}
705 		if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) {
706 			/*
707 			 * The tail is no longer the last one I did a probe
708 			 * on
709 			 */
710 			bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
711 			bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
712 		}
713 	}
714 	if (is_tlp_timer == 0) {
715 		BBR_STAT_INC(bbr_to_arm_rack);
716 		bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
717 	} else {
718 		bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to);
719 		if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
720 			/*
721 			 * We have exceeded how many times we can retran the
722 			 * current TLP timer, switch to the RTO timer.
723 			 */
724 			goto activate_rxt;
725 		} else {
726 			BBR_STAT_INC(bbr_to_arm_tlp);
727 			bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
728 		}
729 	}
730 	return (to);
731 }
732 
733 static inline int32_t
734 bbr_minseg(struct tcp_bbr *bbr)
735 {
736 	return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options);
737 }
738 
739 static void
740 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len)
741 {
742 	struct inpcb *inp;
743 	struct hpts_diag diag;
744 	uint32_t delayed_ack = 0;
745 	uint32_t left = 0;
746 	uint32_t hpts_timeout;
747 	uint8_t stopped;
748 	int32_t delay_calc = 0;
749 	uint32_t prev_delay = 0;
750 
751 	inp = tp->t_inpcb;
752 	if (inp->inp_in_hpts) {
753 		/* A previous call is already set up */
754 		return;
755 	}
756 	if ((tp->t_state == TCPS_CLOSED) ||
757 	    (tp->t_state == TCPS_LISTEN)) {
758 		return;
759 	}
760 	stopped = bbr->rc_tmr_stopped;
761 	if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
762 		left = bbr->r_ctl.rc_timer_exp - cts;
763 	}
764 	bbr->r_ctl.rc_hpts_flags = 0;
765 	bbr->r_ctl.rc_timer_exp = 0;
766 	prev_delay = bbr->r_ctl.rc_last_delay_val;
767 	if (bbr->r_ctl.rc_last_delay_val &&
768 	    (slot == 0)) {
769 		/*
770 		 * If a previous pacer delay was in place we
771 		 * are not coming from the output side (where
772 		 * we calculate a delay, more likely a timer).
773 		 */
774 		slot = bbr->r_ctl.rc_last_delay_val;
775 		if (TSTMP_GT(cts, bbr->rc_pacer_started)) {
776 			/* Compensate for time passed  */
777 			delay_calc = cts - bbr->rc_pacer_started;
778 			if (delay_calc <= slot)
779 				slot -= delay_calc;
780 		}
781 	}
782 	/* Do we have early to make up for by pushing out the pacing time? */
783 	if (bbr->r_agg_early_set) {
784 		bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2);
785 		slot += bbr->r_ctl.rc_agg_early;
786 		bbr->r_ctl.rc_agg_early = 0;
787 		bbr->r_agg_early_set = 0;
788 	}
789 	/* Are we running a total debt that needs to be compensated for? */
790 	if (bbr->r_ctl.rc_hptsi_agg_delay) {
791 		if (slot > bbr->r_ctl.rc_hptsi_agg_delay) {
792 			/* We nuke the delay */
793 			slot -= bbr->r_ctl.rc_hptsi_agg_delay;
794 			bbr->r_ctl.rc_hptsi_agg_delay = 0;
795 		} else {
796 			/* We nuke some of the delay, put in a minimal 100usecs  */
797 			bbr->r_ctl.rc_hptsi_agg_delay -= slot;
798 			bbr->r_ctl.rc_last_delay_val = slot = 100;
799 		}
800 	}
801 	bbr->r_ctl.rc_last_delay_val = slot;
802 	hpts_timeout = bbr_timer_start(tp, bbr, cts);
803 	if (tp->t_flags & TF_DELACK) {
804 		if (bbr->rc_in_persist == 0) {
805 			delayed_ack = bbr_delack_time;
806 		} else {
807 			/*
808 			 * We are in persists and have
809 			 * gotten a new data element.
810 			 */
811 			if (hpts_timeout > bbr_delack_time) {
812 				/*
813 				 * Lets make the persists timer (which acks)
814 				 * be the smaller of hpts_timeout and bbr_delack_time.
815 				 */
816 				hpts_timeout = bbr_delack_time;
817 			}
818 		}
819 	}
820 	if (delayed_ack &&
821 	    ((hpts_timeout == 0) ||
822 	     (delayed_ack < hpts_timeout))) {
823 		/* We need a Delayed ack timer */
824 		bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
825 		hpts_timeout = delayed_ack;
826 	}
827 	if (slot) {
828 		/* Mark that we have a pacing timer up */
829 		BBR_STAT_INC(bbr_paced_segments);
830 		bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
831 	}
832 	/*
833 	 * If no timers are going to run and we will fall off thfe hptsi
834 	 * wheel, we resort to a keep-alive timer if its configured.
835 	 */
836 	if ((hpts_timeout == 0) &&
837 	    (slot == 0)) {
838 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
839 		    (tp->t_state <= TCPS_CLOSING)) {
840 			/*
841 			 * Ok we have no timer (persists, rack, tlp, rxt  or
842 			 * del-ack), we don't have segments being paced. So
843 			 * all that is left is the keepalive timer.
844 			 */
845 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
846 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
847 			} else {
848 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
849 			}
850 			bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
851 		}
852 	}
853 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
854 	    (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
855 		/*
856 		 * RACK, TLP, persists and RXT timers all are restartable
857 		 * based on actions input .. i.e we received a packet (ack
858 		 * or sack) and that changes things (rw, or snd_una etc).
859 		 * Thus we can restart them with a new value. For
860 		 * keep-alive, delayed_ack we keep track of what was left
861 		 * and restart the timer with a smaller value.
862 		 */
863 		if (left < hpts_timeout)
864 			hpts_timeout = left;
865 	}
866 	if (bbr->r_ctl.rc_incr_tmrs && slot &&
867 	    (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
868 		/*
869 		 * If configured to do so, and the timer is either
870 		 * the TLP or RXT timer, we need to increase the timeout
871 		 * by the pacing time. Consider the bottleneck at my
872 		 * machine as an example, we are sending something
873 		 * to start a TLP on. The last packet won't be emitted
874 		 * fully until the pacing time (the bottleneck will hold
875 		 * the data in place). Once the packet is emitted that
876 		 * is when we want to start waiting for the TLP. This
877 		 * is most evident with hardware pacing (where the nic
878 		 * is holding the packet(s) before emitting). But it
879 		 * can also show up in the network so we do it for all
880 		 * cases. Technically we would take off one packet from
881 		 * this extra delay but this is easier and being more
882 		 * conservative is probably better.
883 		 */
884 		hpts_timeout += slot;
885 	}
886 	if (hpts_timeout) {
887 		/*
888 		 * Hack alert for now we can't time-out over 2147 seconds (a
889 		 * bit more than 35min)
890 		 */
891 		if (hpts_timeout > 0x7ffffffe)
892 			hpts_timeout = 0x7ffffffe;
893 		bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
894 	} else
895 		bbr->r_ctl.rc_timer_exp = 0;
896 	if ((slot) &&
897 	    (bbr->rc_use_google ||
898 	     bbr->output_error_seen ||
899 	     (slot <= hpts_timeout))  ) {
900 		/*
901 		 * Tell LRO that it can queue packets while
902 		 * we pace.
903 		 */
904 		bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
905 		if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
906 		    (bbr->rc_cwnd_limited == 0)) {
907 			/*
908 			 * If we are not cwnd limited and we
909 			 * are running a rack timer we put on
910 			 * the do not disturbe even for sack.
911 			 */
912 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
913 		} else
914 			inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
915 		bbr->rc_pacer_started = cts;
916 
917 		(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
918 					   __LINE__, &diag);
919 		bbr->rc_timer_first = 0;
920 		bbr->bbr_timer_src = frm;
921 		bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1);
922 		bbr_log_hpts_diag(bbr, cts, &diag);
923 	} else if (hpts_timeout) {
924 		(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
925 					   __LINE__, &diag);
926 		/*
927 		 * We add the flag here as well if the slot is set,
928 		 * since hpts will call in to clear the queue first before
929 		 * calling the output routine (which does our timers).
930 		 * We don't want to set the flag if its just a timer
931 		 * else the arrival of data might (that causes us
932 		 * to send more) might get delayed. Imagine being
933 		 * on a keep-alive timer and a request comes in for
934 		 * more data.
935 		 */
936 		if (slot)
937 			bbr->rc_pacer_started = cts;
938 		if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
939 		    (bbr->rc_cwnd_limited == 0)) {
940 			/*
941 			 * For a rack timer, don't wake us even
942 			 * if a sack arrives as long as we are
943 			 * not cwnd limited.
944 			 */
945 			bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
946 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
947 		} else {
948 			/* All other timers wake us up */
949 			bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
950 			inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
951 		}
952 		bbr->bbr_timer_src = frm;
953 		bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0);
954 		bbr_log_hpts_diag(bbr, cts, &diag);
955 		bbr->rc_timer_first = 1;
956 	}
957 	bbr->rc_tmr_stopped = 0;
958 	bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay);
959 }
960 
961 static void
962 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb)
963 {
964 	/*
965 	 * We received an ack, and then did not call send or were bounced
966 	 * out due to the hpts was running. Now a timer is up as well, is it
967 	 * the right timer?
968 	 */
969 	struct inpcb *inp;
970 	struct bbr_sendmap *rsm;
971 	uint32_t hpts_timeout;
972 	int tmr_up;
973 
974 	tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
975 	if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
976 		return;
977 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
978 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
979 	    (tmr_up == PACE_TMR_RXT)) {
980 		/* Should be an RXT */
981 		return;
982 	}
983 	inp = bbr->rc_inp;
984 	if (rsm == NULL) {
985 		/* Nothing outstanding? */
986 		if (tp->t_flags & TF_DELACK) {
987 			if (tmr_up == PACE_TMR_DELACK)
988 				/*
989 				 * We are supposed to have delayed ack up
990 				 * and we do
991 				 */
992 				return;
993 		} else if (sbavail(&inp->inp_socket->so_snd) &&
994 		    (tmr_up == PACE_TMR_RXT)) {
995 			/*
996 			 * if we hit enobufs then we would expect the
997 			 * possiblity of nothing outstanding and the RXT up
998 			 * (and the hptsi timer).
999 			 */
1000 			return;
1001 		} else if (((V_tcp_always_keepalive ||
1002 			    inp->inp_socket->so_options & SO_KEEPALIVE) &&
1003 			    (tp->t_state <= TCPS_CLOSING)) &&
1004 			    (tmr_up == PACE_TMR_KEEP) &&
1005 		    (tp->snd_max == tp->snd_una)) {
1006 			/* We should have keep alive up and we do */
1007 			return;
1008 		}
1009 	}
1010 	if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) {
1011 		if ((tp->t_flags & TF_SENTFIN) &&
1012 		    ((tp->snd_max - tp->snd_una) == 1) &&
1013 		    (rsm->r_flags & BBR_HAS_FIN)) {
1014 			/* needs to be a RXT */
1015 			if (tmr_up == PACE_TMR_RXT)
1016 				return;
1017 			else
1018 				goto wrong_timer;
1019 		} else if (tmr_up == PACE_TMR_RACK)
1020 			return;
1021 		else
1022 			goto wrong_timer;
1023 	} else if (rsm && (tmr_up == PACE_TMR_RACK)) {
1024 		/* Rack timer has priority if we have data out */
1025 		return;
1026 	} else if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1027 		    ((tmr_up == PACE_TMR_TLP) ||
1028 	    (tmr_up == PACE_TMR_RXT))) {
1029 		/*
1030 		 * Either a TLP or RXT is fine if no sack-passed is in place
1031 		 * and data is outstanding.
1032 		 */
1033 		return;
1034 	} else if (tmr_up == PACE_TMR_DELACK) {
1035 		/*
1036 		 * If the delayed ack was going to go off before the
1037 		 * rtx/tlp/rack timer were going to expire, then that would
1038 		 * be the timer in control. Note we don't check the time
1039 		 * here trusting the code is correct.
1040 		 */
1041 		return;
1042 	}
1043 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1044 	    ((tmr_up == PACE_TMR_RXT) ||
1045 	     (tmr_up == PACE_TMR_TLP) ||
1046 	     (tmr_up == PACE_TMR_RACK))) {
1047 		/*
1048 		 * We have outstanding data and
1049 		 * we *do* have a RACK, TLP or RXT
1050 		 * timer running. We won't restart
1051 		 * anything here since thats probably ok we
1052 		 * will get called with some timer here shortly.
1053 		 */
1054 		return;
1055 	}
1056 	/*
1057 	 * Ok the timer originally started is not what we want now. We will
1058 	 * force the hpts to be stopped if any, and restart with the slot
1059 	 * set to what was in the saved slot.
1060 	 */
1061 wrong_timer:
1062 	if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) {
1063 		if (inp->inp_in_hpts)
1064 			tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
1065 		bbr_timer_cancel(bbr, __LINE__, cts);
1066 		bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val,
1067 		    0);
1068 	} else {
1069 		/*
1070 		 * Output is hptsi so we just need to switch the type of
1071 		 * timer. We don't bother with keep-alive, since when we
1072 		 * jump through the output, it will start the keep-alive if
1073 		 * nothing is sent.
1074 		 *
1075 		 * We only need a delayed-ack added and or the hpts_timeout.
1076 		 */
1077 		hpts_timeout = bbr_timer_start(tp, bbr, cts);
1078 		if (tp->t_flags & TF_DELACK) {
1079 			if (hpts_timeout == 0) {
1080 				hpts_timeout = bbr_delack_time;
1081 				bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1082 			}
1083 			else if (hpts_timeout > bbr_delack_time) {
1084 				hpts_timeout = bbr_delack_time;
1085 				bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1086 			}
1087 		}
1088 		if (hpts_timeout) {
1089 			if (hpts_timeout > 0x7ffffffe)
1090 				hpts_timeout = 0x7ffffffe;
1091 			bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
1092 		}
1093 	}
1094 }
1095 
1096 int32_t bbr_clear_lost = 0;
1097 
1098 /*
1099  * Considers the two time values now (cts) and earlier.
1100  * If cts is smaller than earlier, we could have
1101  * had a sequence wrap (our counter wraps every
1102  * 70 min or so) or it could be just clock skew
1103  * getting us two differnt time values. Clock skew
1104  * will show up within 10ms or so. So in such
1105  * a case (where cts is behind earlier time by
1106  * less than 10ms) we return 0. Otherwise we
1107  * return the true difference between them.
1108  */
1109 static inline uint32_t
1110 bbr_calc_time(uint32_t cts, uint32_t earlier_time) {
1111 	/*
1112 	 * Given two timestamps, the current time stamp cts, and some other
1113 	 * time-stamp taken in theory earlier return the difference. The
1114 	 * trick is here sometimes locking will get the other timestamp
1115 	 * after the cts. If this occurs we need to return 0.
1116 	 */
1117 	if (TSTMP_GEQ(cts, earlier_time))
1118 		return (cts - earlier_time);
1119 	/*
1120 	 * cts is behind earlier_time if its less than 10ms consider it 0.
1121 	 * If its more than 10ms difference then we had a time wrap. Else
1122 	 * its just the normal locking foo. I wonder if we should not go to
1123 	 * 64bit TS and get rid of this issue.
1124 	 */
1125 	if (TSTMP_GEQ((cts + 10000), earlier_time))
1126 		return (0);
1127 	/*
1128 	 * Ok the time must have wrapped. So we need to answer a large
1129 	 * amount of time, which the normal subtraction should do.
1130 	 */
1131 	return (cts - earlier_time);
1132 }
1133 
1134 
1135 
1136 static int
1137 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS)
1138 {
1139 	uint32_t stat;
1140 	int32_t error;
1141 
1142 	error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t));
1143 	if (error || req->newptr == NULL)
1144 		return error;
1145 
1146 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
1147 	if (error)
1148 		return (error);
1149 	if (stat == 1) {
1150 #ifdef BBR_INVARIANTS
1151 		printf("Clearing BBR lost counters\n");
1152 #endif
1153 		COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT);
1154 		COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT);
1155 		COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT);
1156 	} else if (stat == 2) {
1157 #ifdef BBR_INVARIANTS
1158 		printf("Clearing BBR option counters\n");
1159 #endif
1160 		COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE);
1161 	} else if (stat == 3) {
1162 #ifdef BBR_INVARIANTS
1163 		printf("Clearing BBR stats counters\n");
1164 #endif
1165 		COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE);
1166 	} else if (stat == 4) {
1167 #ifdef BBR_INVARIANTS
1168 		printf("Clearing BBR out-size counters\n");
1169 #endif
1170 		COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE);
1171 	}
1172 	bbr_clear_lost = 0;
1173 	return (0);
1174 }
1175 
1176 static void
1177 bbr_init_sysctls(void)
1178 {
1179 	struct sysctl_oid *bbr_probertt;
1180 	struct sysctl_oid *bbr_hptsi;
1181 	struct sysctl_oid *bbr_measure;
1182 	struct sysctl_oid *bbr_cwnd;
1183 	struct sysctl_oid *bbr_timeout;
1184 	struct sysctl_oid *bbr_states;
1185 	struct sysctl_oid *bbr_startup;
1186 	struct sysctl_oid *bbr_policer;
1187 
1188 	/* Probe rtt controls */
1189 	bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1190 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1191 	    OID_AUTO,
1192 	    "probertt",
1193 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1194 	    "");
1195 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1196 	    SYSCTL_CHILDREN(bbr_probertt),
1197 	    OID_AUTO, "gain", CTLFLAG_RW,
1198 	    &bbr_rttprobe_gain, 192,
1199 	    "What is the filter gain drop in probe_rtt (0=disable)?");
1200 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1201 	    SYSCTL_CHILDREN(bbr_probertt),
1202 	    OID_AUTO, "cwnd", CTLFLAG_RW,
1203 	    &bbr_rtt_probe_cwndtarg, 4,
1204 	    "How many mss's are outstanding during probe-rtt");
1205 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1206 	    SYSCTL_CHILDREN(bbr_probertt),
1207 	    OID_AUTO, "int", CTLFLAG_RW,
1208 	    &bbr_rtt_probe_limit, 4000000,
1209 	    "If RTT has not shrank in this many micro-seconds enter probe-rtt");
1210 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1211 	    SYSCTL_CHILDREN(bbr_probertt),
1212 	    OID_AUTO, "mintime", CTLFLAG_RW,
1213 	    &bbr_rtt_probe_time, 200000,
1214 	    "How many microseconds in probe-rtt");
1215 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1216 	    SYSCTL_CHILDREN(bbr_probertt),
1217 	    OID_AUTO, "filter_len_sec", CTLFLAG_RW,
1218 	    &bbr_filter_len_sec, 6,
1219 	    "How long in seconds does the rttProp filter run?");
1220 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1221 	    SYSCTL_CHILDREN(bbr_probertt),
1222 	    OID_AUTO, "drain_rtt", CTLFLAG_RW,
1223 	    &bbr_drain_rtt, BBR_SRTT,
1224 	    "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?");
1225 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1226 	    SYSCTL_CHILDREN(bbr_probertt),
1227 	    OID_AUTO, "can_force", CTLFLAG_RW,
1228 	    &bbr_can_force_probertt, 0,
1229 	    "If we keep setting new low rtt's but delay going in probe-rtt can we force in??");
1230 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1231 	    SYSCTL_CHILDREN(bbr_probertt),
1232 	    OID_AUTO, "enter_sets_force", CTLFLAG_RW,
1233 	    &bbr_probertt_sets_rtt, 0,
1234 	    "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?");
1235 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1236 	    SYSCTL_CHILDREN(bbr_probertt),
1237 	    OID_AUTO, "can_adjust", CTLFLAG_RW,
1238 	    &bbr_can_adjust_probertt, 1,
1239 	    "Can we dynamically adjust the probe-rtt limits and times?");
1240 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1241 	    SYSCTL_CHILDREN(bbr_probertt),
1242 	    OID_AUTO, "is_ratio", CTLFLAG_RW,
1243 	    &bbr_is_ratio, 0,
1244 	    "is the limit to filter a ratio?");
1245 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1246 	    SYSCTL_CHILDREN(bbr_probertt),
1247 	    OID_AUTO, "use_cwnd", CTLFLAG_RW,
1248 	    &bbr_prtt_slam_cwnd, 0,
1249 	    "Should we set/recover cwnd?");
1250 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1251 	    SYSCTL_CHILDREN(bbr_probertt),
1252 	    OID_AUTO, "can_use_ts", CTLFLAG_RW,
1253 	    &bbr_can_use_ts_for_rtt, 1,
1254 	    "Can we use the ms timestamp if available for retransmistted rtt calculations?");
1255 
1256 	/* Pacing controls */
1257 	bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1258 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1259 	    OID_AUTO,
1260 	    "pacing",
1261 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1262 	    "");
1263 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1264 	    SYSCTL_CHILDREN(bbr_hptsi),
1265 	    OID_AUTO, "hw_pacing", CTLFLAG_RW,
1266 	    &bbr_allow_hdwr_pacing, 1,
1267 	    "Do we allow hardware pacing?");
1268 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1269 	    SYSCTL_CHILDREN(bbr_hptsi),
1270 	    OID_AUTO, "hw_pacing_limit", CTLFLAG_RW,
1271 	    &bbr_hardware_pacing_limit, 4000,
1272 	    "Do we have a limited number of connections for pacing chelsio (0=no limit)?");
1273 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1274 	    SYSCTL_CHILDREN(bbr_hptsi),
1275 	    OID_AUTO, "hw_pacing_adj", CTLFLAG_RW,
1276 	    &bbr_hdwr_pace_adjust, 2,
1277 	    "Multiplier to calculated tso size?");
1278 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1279 	    SYSCTL_CHILDREN(bbr_hptsi),
1280 	    OID_AUTO, "hw_pacing_floor", CTLFLAG_RW,
1281 	    &bbr_hdwr_pace_floor, 1,
1282 	    "Do we invoke the hardware pacing floor?");
1283 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1284 	    SYSCTL_CHILDREN(bbr_hptsi),
1285 	    OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW,
1286 	    &bbr_hdwr_pacing_delay_cnt, 10,
1287 	    "How many packets must be sent after hdwr pacing is enabled");
1288 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1289 	    SYSCTL_CHILDREN(bbr_hptsi),
1290 	    OID_AUTO, "bw_cross", CTLFLAG_RW,
1291 	    &bbr_cross_over, 3000000,
1292 	    "What is the point where we cross over to linux like TSO size set");
1293 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1294 	    SYSCTL_CHILDREN(bbr_hptsi),
1295 	    OID_AUTO, "seg_deltarg", CTLFLAG_RW,
1296 	    &bbr_hptsi_segments_delay_tar, 7000,
1297 	    "What is the worse case delay target for hptsi < 48Mbp connections");
1298 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1299 	    SYSCTL_CHILDREN(bbr_hptsi),
1300 	    OID_AUTO, "enet_oh", CTLFLAG_RW,
1301 	    &bbr_include_enet_oh, 0,
1302 	    "Do we include the ethernet overhead in calculating pacing delay?");
1303 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1304 	    SYSCTL_CHILDREN(bbr_hptsi),
1305 	    OID_AUTO, "ip_oh", CTLFLAG_RW,
1306 	    &bbr_include_ip_oh, 1,
1307 	    "Do we include the IP overhead in calculating pacing delay?");
1308 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1309 	    SYSCTL_CHILDREN(bbr_hptsi),
1310 	    OID_AUTO, "tcp_oh", CTLFLAG_RW,
1311 	    &bbr_include_tcp_oh, 0,
1312 	    "Do we include the TCP overhead in calculating pacing delay?");
1313 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1314 	    SYSCTL_CHILDREN(bbr_hptsi),
1315 	    OID_AUTO, "google_discount", CTLFLAG_RW,
1316 	    &bbr_google_discount, 10,
1317 	    "What is the default google discount percentage wise for pacing (11 = 1.1%%)?");
1318 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1319 	    SYSCTL_CHILDREN(bbr_hptsi),
1320 	    OID_AUTO, "all_get_min", CTLFLAG_RW,
1321 	    &bbr_all_get_min, 0,
1322 	    "If you are less than a MSS do you just get the min?");
1323 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1324 	    SYSCTL_CHILDREN(bbr_hptsi),
1325 	    OID_AUTO, "tso_min", CTLFLAG_RW,
1326 	    &bbr_hptsi_bytes_min, 1460,
1327 	    "For 0 -> 24Mbps what is floor number of segments for TSO");
1328 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1329 	    SYSCTL_CHILDREN(bbr_hptsi),
1330 	    OID_AUTO, "seg_tso_max", CTLFLAG_RW,
1331 	    &bbr_hptsi_segments_max, 6,
1332 	    "For 0 -> 24Mbps what is top number of segments for TSO");
1333 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1334 	    SYSCTL_CHILDREN(bbr_hptsi),
1335 	    OID_AUTO, "seg_floor", CTLFLAG_RW,
1336 	    &bbr_hptsi_segments_floor, 1,
1337 	    "Minimum TSO size we will fall too in segments");
1338 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1339 	    SYSCTL_CHILDREN(bbr_hptsi),
1340 	    OID_AUTO, "utter_max", CTLFLAG_RW,
1341 	    &bbr_hptsi_utter_max, 0,
1342 	    "The absolute maximum that any pacing (outside of hardware) can be");
1343 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1344 	    SYSCTL_CHILDREN(bbr_hptsi),
1345 	    OID_AUTO, "seg_divisor", CTLFLAG_RW,
1346 	    &bbr_hptsi_per_second, 100,
1347 	    "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps ");
1348 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1349 	    SYSCTL_CHILDREN(bbr_hptsi),
1350 	    OID_AUTO, "srtt_mul", CTLFLAG_RW,
1351 	    &bbr_hptsi_max_mul, 1,
1352 	    "The multiplier for pace len max");
1353 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1354 	    SYSCTL_CHILDREN(bbr_hptsi),
1355 	    OID_AUTO, "srtt_div", CTLFLAG_RW,
1356 	    &bbr_hptsi_max_div, 2,
1357 	    "The divisor for pace len max");
1358 	/* Measurement controls */
1359 	bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1360 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1361 	    OID_AUTO,
1362 	    "measure",
1363 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1364 	    "Measurement controls");
1365 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1366 	    SYSCTL_CHILDREN(bbr_measure),
1367 	    OID_AUTO, "min_i_bw", CTLFLAG_RW,
1368 	    &bbr_initial_bw_bps, 62500,
1369 	    "Minimum initial b/w in bytes per second");
1370 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1371 	    SYSCTL_CHILDREN(bbr_measure),
1372 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1373 	    &bbr_sack_not_required, 0,
1374 	    "Do we allow bbr to run on connections not supporting SACK?");
1375 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1376 	    SYSCTL_CHILDREN(bbr_measure),
1377 	    OID_AUTO, "use_google", CTLFLAG_RW,
1378 	    &bbr_use_google_algo, 0,
1379 	    "Use has close to google V1.0 has possible?");
1380 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1381 	    SYSCTL_CHILDREN(bbr_measure),
1382 	    OID_AUTO, "ts_limiting", CTLFLAG_RW,
1383 	    &bbr_ts_limiting, 1,
1384 	    "Do we attempt to use the peers timestamp to limit b/w caculations?");
1385 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1386 	    SYSCTL_CHILDREN(bbr_measure),
1387 	    OID_AUTO, "ts_can_raise", CTLFLAG_RW,
1388 	    &bbr_ts_can_raise, 0,
1389 	    "Can we raise the b/w via timestamp b/w calculation?");
1390 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1391 	    SYSCTL_CHILDREN(bbr_measure),
1392 	    OID_AUTO, "ts_delta", CTLFLAG_RW,
1393 	    &bbr_min_usec_delta, 20000,
1394 	    "How long in usec between ts of our sends in ts validation code?");
1395 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1396 	    SYSCTL_CHILDREN(bbr_measure),
1397 	    OID_AUTO, "ts_peer_delta", CTLFLAG_RW,
1398 	    &bbr_min_peer_delta, 20,
1399 	    "What min numerical value should be between the peer deltas?");
1400 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1401 	    SYSCTL_CHILDREN(bbr_measure),
1402 	    OID_AUTO, "ts_delta_percent", CTLFLAG_RW,
1403 	    &bbr_delta_percent, 150,
1404 	    "What percentage (150 = 15.0) do we allow variance for?");
1405 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1406 	    SYSCTL_CHILDREN(bbr_measure),
1407 	    OID_AUTO, "min_measure_good_bw", CTLFLAG_RW,
1408 	    &bbr_min_measurements_req, 1,
1409 	    "What is the minimum measurment count we need before we switch to our b/w estimate");
1410 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1411 	    SYSCTL_CHILDREN(bbr_measure),
1412 	    OID_AUTO, "min_measure_before_pace", CTLFLAG_RW,
1413 	    &bbr_no_pacing_until, 4,
1414 	    "How many pkt-epoch's (0 is off) do we need before pacing is on?");
1415 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1416 	    SYSCTL_CHILDREN(bbr_measure),
1417 	    OID_AUTO, "quanta", CTLFLAG_RW,
1418 	    &bbr_quanta, 2,
1419 	    "Extra quanta to add when calculating the target (ID section 4.2.3.2).");
1420 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1421 	    SYSCTL_CHILDREN(bbr_measure),
1422 	    OID_AUTO, "noretran", CTLFLAG_RW,
1423 	    &bbr_no_retran, 0,
1424 	    "Should google mode not use retransmission measurements for the b/w estimation?");
1425 	/* State controls */
1426 	bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1427 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1428 	    OID_AUTO,
1429 	    "states",
1430 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1431 	    "State controls");
1432 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1433 	    SYSCTL_CHILDREN(bbr_states),
1434 	    OID_AUTO, "idle_restart", CTLFLAG_RW,
1435 	    &bbr_uses_idle_restart, 0,
1436 	    "Do we use a new special idle_restart state to ramp back up quickly?");
1437 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1438 	    SYSCTL_CHILDREN(bbr_states),
1439 	    OID_AUTO, "idle_restart_threshold", CTLFLAG_RW,
1440 	    &bbr_idle_restart_threshold, 100000,
1441 	    "How long must we be idle before we restart??");
1442 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1443 	    SYSCTL_CHILDREN(bbr_states),
1444 	    OID_AUTO, "use_pkt_epoch", CTLFLAG_RW,
1445 	    &bbr_state_is_pkt_epoch, 0,
1446 	    "Do we use a pkt-epoch for substate if 0 rttProp?");
1447 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1448 	    SYSCTL_CHILDREN(bbr_states),
1449 	    OID_AUTO, "startup_rtt_gain", CTLFLAG_RW,
1450 	    &bbr_rtt_gain_thresh, 0,
1451 	    "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?");
1452 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1453 	    SYSCTL_CHILDREN(bbr_states),
1454 	    OID_AUTO, "drain_floor", CTLFLAG_RW,
1455 	    &bbr_drain_floor, 88,
1456 	    "What is the lowest we can drain (pg) too?");
1457 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1458 	    SYSCTL_CHILDREN(bbr_states),
1459 	    OID_AUTO, "drain_2_target", CTLFLAG_RW,
1460 	    &bbr_state_drain_2_tar, 1,
1461 	    "Do we drain to target in drain substate?");
1462 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1463 	    SYSCTL_CHILDREN(bbr_states),
1464 	    OID_AUTO, "gain_2_target", CTLFLAG_RW,
1465 	    &bbr_gain_to_target, 1,
1466 	    "Does probe bw gain to target??");
1467 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1468 	    SYSCTL_CHILDREN(bbr_states),
1469 	    OID_AUTO, "gain_extra_time", CTLFLAG_RW,
1470 	    &bbr_gain_gets_extra_too, 1,
1471 	    "Does probe bw gain get the extra time too?");
1472 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1473 	    SYSCTL_CHILDREN(bbr_states),
1474 	    OID_AUTO, "ld_div", CTLFLAG_RW,
1475 	    &bbr_drain_drop_div, 5,
1476 	    "Long drain drop divider?");
1477 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1478 	    SYSCTL_CHILDREN(bbr_states),
1479 	    OID_AUTO, "ld_mul", CTLFLAG_RW,
1480 	    &bbr_drain_drop_mul, 4,
1481 	    "Long drain drop multiplier?");
1482 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1483 	    SYSCTL_CHILDREN(bbr_states),
1484 	    OID_AUTO, "rand_ot_disc", CTLFLAG_RW,
1485 	    &bbr_rand_ot, 50,
1486 	    "Random discount of the ot?");
1487 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1488 	    SYSCTL_CHILDREN(bbr_states),
1489 	    OID_AUTO, "dr_filter_life", CTLFLAG_RW,
1490 	    &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT,
1491 	    "How many packet-epochs does the b/w delivery rate last?");
1492 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1493 	    SYSCTL_CHILDREN(bbr_states),
1494 	    OID_AUTO, "subdrain_applimited", CTLFLAG_RW,
1495 	    &bbr_sub_drain_app_limit, 0,
1496 	    "Does our sub-state drain invoke app limited if its long?");
1497 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1498 	    SYSCTL_CHILDREN(bbr_states),
1499 	    OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW,
1500 	    &bbr_sub_drain_slam_cwnd, 0,
1501 	    "Should we set/recover cwnd for sub-state drain?");
1502 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1503 	    SYSCTL_CHILDREN(bbr_states),
1504 	    OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW,
1505 	    &bbr_slam_cwnd_in_main_drain, 0,
1506 	    "Should we set/recover cwnd for main-state drain?");
1507 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1508 	    SYSCTL_CHILDREN(bbr_states),
1509 	    OID_AUTO, "google_gets_earlyout", CTLFLAG_RW,
1510 	    &google_allow_early_out, 1,
1511 	    "Should we allow google probe-bw/drain to exit early at flight target?");
1512 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1513 	    SYSCTL_CHILDREN(bbr_states),
1514 	    OID_AUTO, "google_exit_loss", CTLFLAG_RW,
1515 	    &google_consider_lost, 1,
1516 	    "Should we have losses exit gain of probebw in google mode??");
1517 	/* Startup controls */
1518 	bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1519 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1520 	    OID_AUTO,
1521 	    "startup",
1522 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1523 	    "Startup controls");
1524 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1525 	    SYSCTL_CHILDREN(bbr_startup),
1526 	    OID_AUTO, "cheat_iwnd", CTLFLAG_RW,
1527 	    &bbr_sends_full_iwnd, 1,
1528 	    "Do we not pace but burst out initial windows has our TSO size?");
1529 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1530 	    SYSCTL_CHILDREN(bbr_startup),
1531 	    OID_AUTO, "loss_threshold", CTLFLAG_RW,
1532 	    &bbr_startup_loss_thresh, 2000,
1533 	    "In startup what is the loss threshold in a pe that will exit us from startup?");
1534 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1535 	    SYSCTL_CHILDREN(bbr_startup),
1536 	    OID_AUTO, "use_lowerpg", CTLFLAG_RW,
1537 	    &bbr_use_lower_gain_in_startup, 1,
1538 	    "Should we use a lower hptsi gain if we see loss in startup?");
1539 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1540 	    SYSCTL_CHILDREN(bbr_startup),
1541 	    OID_AUTO, "gain", CTLFLAG_RW,
1542 	    &bbr_start_exit, 25,
1543 	    "What gain percent do we need to see to stay in startup??");
1544 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1545 	    SYSCTL_CHILDREN(bbr_startup),
1546 	    OID_AUTO, "low_gain", CTLFLAG_RW,
1547 	    &bbr_low_start_exit, 15,
1548 	    "What gain percent do we need to see to stay in the lower gain startup??");
1549 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1550 	    SYSCTL_CHILDREN(bbr_startup),
1551 	    OID_AUTO, "loss_exit", CTLFLAG_RW,
1552 	    &bbr_exit_startup_at_loss, 1,
1553 	    "Should we exit startup at loss in an epoch if we are not gaining?");
1554 	/* CWND controls */
1555 	bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1556 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1557 	    OID_AUTO,
1558 	    "cwnd",
1559 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1560 	    "Cwnd controls");
1561 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1562 	    SYSCTL_CHILDREN(bbr_cwnd),
1563 	    OID_AUTO, "tar_rtt", CTLFLAG_RW,
1564 	    &bbr_cwndtarget_rtt_touse, 0,
1565 	    "Target cwnd rtt measurment to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?");
1566 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1567 	    SYSCTL_CHILDREN(bbr_cwnd),
1568 	    OID_AUTO, "may_shrink", CTLFLAG_RW,
1569 	    &bbr_cwnd_may_shrink, 0,
1570 	    "Can the cwnd shrink if it would grow to more than the target?");
1571 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1572 	    SYSCTL_CHILDREN(bbr_cwnd),
1573 	    OID_AUTO, "max_target_limit", CTLFLAG_RW,
1574 	    &bbr_target_cwnd_mult_limit, 8,
1575 	    "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?");
1576 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1577 	    SYSCTL_CHILDREN(bbr_cwnd),
1578 	    OID_AUTO, "highspeed_min", CTLFLAG_RW,
1579 	    &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS,
1580 	    "What is the high-speed min cwnd (rttProp under 1ms)");
1581 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1582 	    SYSCTL_CHILDREN(bbr_cwnd),
1583 	    OID_AUTO, "lowspeed_min", CTLFLAG_RW,
1584 	    &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS,
1585 	    "What is the min cwnd (rttProp > 1ms)");
1586 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1587 	    SYSCTL_CHILDREN(bbr_cwnd),
1588 	    OID_AUTO, "initwin", CTLFLAG_RW,
1589 	    &bbr_def_init_win, 10,
1590 	    "What is the BBR initial window, if 0 use tcp version");
1591 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1592 	    SYSCTL_CHILDREN(bbr_cwnd),
1593 	    OID_AUTO, "do_loss_red", CTLFLAG_RW,
1594 	    &bbr_do_red, 600,
1595 	    "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?");
1596 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1597 	    SYSCTL_CHILDREN(bbr_cwnd),
1598 	    OID_AUTO, "red_scale", CTLFLAG_RW,
1599 	    &bbr_red_scale, 20000,
1600 	    "What RTT do we scale with?");
1601 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1602 	    SYSCTL_CHILDREN(bbr_cwnd),
1603 	    OID_AUTO, "red_growslow", CTLFLAG_RW,
1604 	    &bbr_red_growth_restrict, 1,
1605 	    "Do we restrict cwnd growth for whats in flight?");
1606 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1607 	    SYSCTL_CHILDREN(bbr_cwnd),
1608 	    OID_AUTO, "red_div", CTLFLAG_RW,
1609 	    &bbr_red_div, 2,
1610 	    "If we reduce whats the divisor?");
1611 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1612 	    SYSCTL_CHILDREN(bbr_cwnd),
1613 	    OID_AUTO, "red_mul", CTLFLAG_RW,
1614 	    &bbr_red_mul, 1,
1615 	    "If we reduce whats the mulitiplier?");
1616 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1617 	    SYSCTL_CHILDREN(bbr_cwnd),
1618 	    OID_AUTO, "target_is_unit", CTLFLAG_RW,
1619 	    &bbr_target_is_bbunit, 0,
1620 	    "Is the state target the pacing_gain or BBR_UNIT?");
1621 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1622 	    SYSCTL_CHILDREN(bbr_cwnd),
1623 	    OID_AUTO, "drop_limit", CTLFLAG_RW,
1624 	    &bbr_drop_limit, 0,
1625 	    "Number of segments limit for drop (0=use min_cwnd w/flight)?");
1626 
1627         /* Timeout controls */
1628 	bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1629 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1630 	    OID_AUTO,
1631 	    "timeout",
1632 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1633 	    "Time out controls");
1634 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1635 	    SYSCTL_CHILDREN(bbr_timeout),
1636 	    OID_AUTO, "delack", CTLFLAG_RW,
1637 	    &bbr_delack_time, 100000,
1638 	    "BBR's delayed ack time");
1639 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1640 	    SYSCTL_CHILDREN(bbr_timeout),
1641 	    OID_AUTO, "tlp_uses", CTLFLAG_RW,
1642 	    &bbr_tlp_type_to_use, 3,
1643 	    "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt");
1644 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1645 	    SYSCTL_CHILDREN(bbr_timeout),
1646 	    OID_AUTO, "persmin", CTLFLAG_RW,
1647 	    &bbr_persist_min, 250000,
1648 	    "What is the minimum time in microseconds between persists");
1649 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1650 	    SYSCTL_CHILDREN(bbr_timeout),
1651 	    OID_AUTO, "persmax", CTLFLAG_RW,
1652 	    &bbr_persist_max, 1000000,
1653 	    "What is the largest delay in microseconds between persists");
1654 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1655 	    SYSCTL_CHILDREN(bbr_timeout),
1656 	    OID_AUTO, "tlp_minto", CTLFLAG_RW,
1657 	    &bbr_tlp_min, 10000,
1658 	    "TLP Min timeout in usecs");
1659 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1660 	    SYSCTL_CHILDREN(bbr_timeout),
1661 	    OID_AUTO, "tlp_dack_time", CTLFLAG_RW,
1662 	    &bbr_delayed_ack_time, 200000,
1663 	    "TLP delayed ack compensation value");
1664 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1665 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1666 	    OID_AUTO, "minrto", CTLFLAG_RW,
1667 	    &bbr_rto_min_ms, 30,
1668 	    "Minimum RTO in ms");
1669 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1670 	    SYSCTL_CHILDREN(bbr_timeout),
1671 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1672 	    &bbr_rto_max_sec, 4,
1673 	    "Maxiumum RTO in seconds -- should be at least as large as min_rto");
1674 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1675 	    SYSCTL_CHILDREN(bbr_timeout),
1676 	    OID_AUTO, "tlp_retry", CTLFLAG_RW,
1677 	    &bbr_tlp_max_resend, 2,
1678 	    "How many times does TLP retry a single segment or multiple with no ACK");
1679 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1680 	    SYSCTL_CHILDREN(bbr_timeout),
1681 	    OID_AUTO, "minto", CTLFLAG_RW,
1682 	    &bbr_min_to, 1000,
1683 	    "Minimum rack timeout in useconds");
1684 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1685 	    SYSCTL_CHILDREN(bbr_timeout),
1686 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1687 	    &bbr_pkt_delay, 1000,
1688 	    "Extra RACK time (in useconds) besides reordering thresh");
1689 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1690 	    SYSCTL_CHILDREN(bbr_timeout),
1691 	    OID_AUTO, "incr_tmrs", CTLFLAG_RW,
1692 	    &bbr_incr_timers, 1,
1693 	    "Increase the RXT/TLP timer by the pacing time used?");
1694 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1695 	    SYSCTL_CHILDREN(bbr_timeout),
1696 	    OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW,
1697 	    &bbr_marks_rxt_sack_passed, 0,
1698 	    "Mark sack passed on all those not ack'd when a RXT hits?");
1699 	/* Policer controls */
1700 	bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1701 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1702 	    OID_AUTO,
1703 	    "policer",
1704 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1705 	    "Policer controls");
1706 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1707 	    SYSCTL_CHILDREN(bbr_policer),
1708 	    OID_AUTO, "detect_enable", CTLFLAG_RW,
1709 	    &bbr_policer_detection_enabled, 1,
1710 	    "Is policer detection enabled??");
1711 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1712 	    SYSCTL_CHILDREN(bbr_policer),
1713 	    OID_AUTO, "min_pes", CTLFLAG_RW,
1714 	    &bbr_lt_intvl_min_rtts, 4,
1715 	    "Minimum number of PE's?");
1716 	SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1717 	    SYSCTL_CHILDREN(bbr_policer),
1718 	    OID_AUTO, "bwdiff", CTLFLAG_RW,
1719 	    &bbr_lt_bw_diff, (4000/8),
1720 	    "Minimal bw diff?");
1721 	SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1722 	    SYSCTL_CHILDREN(bbr_policer),
1723 	    OID_AUTO, "bwratio", CTLFLAG_RW,
1724 	    &bbr_lt_bw_ratio, 8,
1725 	    "Minimal bw diff?");
1726 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1727 	    SYSCTL_CHILDREN(bbr_policer),
1728 	    OID_AUTO, "from_rack_rxt", CTLFLAG_RW,
1729 	    &bbr_policer_call_from_rack_to, 0,
1730 	    "Do we call the policer detection code from a rack-timeout?");
1731 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1732 	    SYSCTL_CHILDREN(bbr_policer),
1733 	    OID_AUTO, "false_postive", CTLFLAG_RW,
1734 	    &bbr_lt_intvl_fp, 0,
1735 	    "What packet epoch do we do false-postive detection at (0=no)?");
1736 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1737 	    SYSCTL_CHILDREN(bbr_policer),
1738 	    OID_AUTO, "loss_thresh", CTLFLAG_RW,
1739 	    &bbr_lt_loss_thresh, 196,
1740 	    "Loss threshold 196 = 19.6%?");
1741 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1742 	    SYSCTL_CHILDREN(bbr_policer),
1743 	    OID_AUTO, "false_postive_thresh", CTLFLAG_RW,
1744 	    &bbr_lt_fd_thresh, 100,
1745 	    "What percentage is the false detection threshold (150=15.0)?");
1746 	/* All the rest */
1747 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1748 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1749 	    OID_AUTO, "cheat_rxt", CTLFLAG_RW,
1750 	    &bbr_use_rack_resend_cheat, 0,
1751 	    "Do we burst 1ms between sends on retransmissions (like rack)?");
1752 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1753 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1754 	    OID_AUTO, "error_paceout", CTLFLAG_RW,
1755 	    &bbr_error_base_paceout, 10000,
1756 	    "When we hit an error what is the min to pace out in usec's?");
1757 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1758 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1759 	    OID_AUTO, "kill_paceout", CTLFLAG_RW,
1760 	    &bbr_max_net_error_cnt, 10,
1761 	    "When we hit this many errors in a row, kill the session?");
1762 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1763 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1764 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1765 	    &bbr_ignore_data_after_close, 1,
1766 	    "Do we hold off sending a RST until all pending data is ack'd");
1767 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1768 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1769 	    OID_AUTO, "resend_use_tso", CTLFLAG_RW,
1770 	    &bbr_resends_use_tso, 0,
1771 	    "Can resends use TSO?");
1772 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1773 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1774 	    OID_AUTO, "sblklimit", CTLFLAG_RW,
1775 	    &bbr_sack_block_limit, 128,
1776 	    "When do we start ignoring small sack blocks");
1777 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1778 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1779 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1780 	    &bbr_verbose_logging, 0,
1781 	    "Should BBR black box logging be verbose");
1782 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1783 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1784 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1785 	    &bbr_reorder_thresh, 2,
1786 	    "What factor for rack will be added when seeing reordering (shift right)");
1787 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1788 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1789 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1790 	    &bbr_reorder_fade, 0,
1791 	    "Does reorder detection fade, if so how many ms (0 means never)");
1792 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1793 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1794 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1795 	    &bbr_tlp_thresh, 1,
1796 	    "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1797 	/* Stats and counters */
1798 	/* The pacing counters for hdwr/software can't be in the array */
1799 	bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1800 	bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1801 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1802 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1803 	    OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD,
1804 	    &bbr_hdwr_pacing_enobuf,
1805 	    "Total number of enobufs for hardware paced flows");
1806 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1807 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1808 	    OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD,
1809 	    &bbr_nohdwr_pacing_enobuf,
1810 	    "Total number of enobufs for non-hardware paced flows");
1811 
1812 
1813 	bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK);
1814 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1815 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1816 	    OID_AUTO, "hdwr_pacing", CTLFLAG_RD,
1817 	    &bbr_flows_whdwr_pacing,
1818 	    "Total number of hardware paced flows");
1819 	bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK);
1820 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1821 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1822 	    OID_AUTO, "software_pacing", CTLFLAG_RD,
1823 	    &bbr_flows_nohdwr_pacing,
1824 	    "Total number of software paced flows");
1825 	COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK);
1826 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1827 	    OID_AUTO, "stats", CTLFLAG_RD,
1828 	    bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats");
1829 	COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK);
1830 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1831 	    OID_AUTO, "opts", CTLFLAG_RD,
1832 	    bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats");
1833 	COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK);
1834 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1835 	    OID_AUTO, "lost", CTLFLAG_RD,
1836 	    bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur");
1837 	COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK);
1838 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1839 	    OID_AUTO, "stateresend", CTLFLAG_RD,
1840 	    bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend");
1841 	COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK);
1842 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1843 	    OID_AUTO, "statetime", CTLFLAG_RD,
1844 	    bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states");
1845 	COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1846 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1847 	    OID_AUTO, "outsize", CTLFLAG_RD,
1848 	    bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls");
1849 	SYSCTL_ADD_PROC(&bbr_sysctl_ctx,
1850 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1851 	    OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1852 	    &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters");
1853 }
1854 
1855 static void
1856 bbr_counter_destroy(void)
1857 {
1858 	COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE);
1859 	COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE);
1860 	COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE);
1861 	COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT);
1862 	COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT);
1863 	COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT);
1864 	counter_u64_free(bbr_nohdwr_pacing_enobuf);
1865 	counter_u64_free(bbr_hdwr_pacing_enobuf);
1866 	counter_u64_free(bbr_flows_whdwr_pacing);
1867 	counter_u64_free(bbr_flows_nohdwr_pacing);
1868 
1869 }
1870 
1871 static __inline void
1872 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts)
1873 {
1874 	memset(l, 0, sizeof(union tcp_log_stackspecific));
1875 	l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate;
1876 	l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate);
1877 	l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
1878 	l->bw_inuse = bbr_get_bw(bbr);
1879 	l->inflight = ctf_flight_size(bbr->rc_tp,
1880 			  (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
1881 	l->applimited = bbr->r_ctl.r_app_limited_until;
1882 	l->delivered = bbr->r_ctl.rc_delivered;
1883 	l->timeStamp = cts;
1884 	l->lost = bbr->r_ctl.rc_lost;
1885 	l->bbr_state = bbr->rc_bbr_state;
1886 	l->bbr_substate = bbr_state_val(bbr);
1887 	l->epoch = bbr->r_ctl.rc_rtt_epoch;
1888 	l->lt_epoch = bbr->r_ctl.rc_lt_epoch;
1889 	l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
1890 	l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain;
1891 	l->inhpts = bbr->rc_inp->inp_in_hpts;
1892 	l->ininput = bbr->rc_inp->inp_in_input;
1893 	l->use_lt_bw = bbr->rc_lt_use_bw;
1894 	l->pkts_out = bbr->r_ctl.rc_flight_at_input;
1895 	l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch;
1896 }
1897 
1898 static void
1899 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason)
1900 {
1901 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1902 		union tcp_log_stackspecific log;
1903 
1904 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1905 		log.u_bbr.flex1 = 0;
1906 		log.u_bbr.flex2 = 0;
1907 		log.u_bbr.flex5 = 0;
1908 		log.u_bbr.flex3 = 0;
1909 		log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate;
1910 		log.u_bbr.flex7 = reason;
1911 		log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt;
1912 		log.u_bbr.flex8 = 0;
1913 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1914 		    &bbr->rc_inp->inp_socket->so_rcv,
1915 		    &bbr->rc_inp->inp_socket->so_snd,
1916 		    BBR_LOG_BW_RED_EV, 0,
1917 		    0, &log, false, &bbr->rc_tv);
1918 	}
1919 }
1920 
1921 static void
1922 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count)
1923 {
1924 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1925 		union tcp_log_stackspecific log;
1926 
1927 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1928 		log.u_bbr.flex1 = seq;
1929 		log.u_bbr.flex2 = count;
1930 		log.u_bbr.flex8 = mode;
1931 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1932 		    &bbr->rc_inp->inp_socket->so_rcv,
1933 		    &bbr->rc_inp->inp_socket->so_snd,
1934 		    BBR_LOG_LOWGAIN, 0,
1935 		    0, &log, false, &bbr->rc_tv);
1936 	}
1937 }
1938 
1939 
1940 
1941 static void
1942 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling,
1943     uint8_t reason, uint32_t p_maxseg, int len)
1944 {
1945 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1946 		union tcp_log_stackspecific log;
1947 
1948 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1949 		log.u_bbr.flex1 = p_maxseg;
1950 		log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags;
1951 		log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
1952 		log.u_bbr.flex4 = reason;
1953 		log.u_bbr.flex5 = bbr->rc_in_persist;
1954 		log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val;
1955 		log.u_bbr.flex7 = p_maxseg;
1956 		log.u_bbr.flex8 = bbr->rc_in_persist;
1957 		log.u_bbr.pkts_out = 0;
1958 		log.u_bbr.applimited = len;
1959 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1960 		    &bbr->rc_inp->inp_socket->so_rcv,
1961 		    &bbr->rc_inp->inp_socket->so_snd,
1962 		    BBR_LOG_JUSTRET, 0,
1963 		    tlen, &log, false, &bbr->rc_tv);
1964 	}
1965 }
1966 
1967 
1968 static void
1969 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq)
1970 {
1971 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1972 		union tcp_log_stackspecific log;
1973 
1974 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1975 		log.u_bbr.flex1 = seq;
1976 		log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
1977 		log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start;
1978 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1979 		    &bbr->rc_inp->inp_socket->so_rcv,
1980 		    &bbr->rc_inp->inp_socket->so_snd,
1981 		    BBR_LOG_ENTREC, 0,
1982 		    0, &log, false, &bbr->rc_tv);
1983 	}
1984 }
1985 
1986 static void
1987 bbr_log_msgsize_fail(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t len, uint32_t maxseg, uint32_t mtu, int32_t csum_flags, int32_t tso, uint32_t cts)
1988 {
1989 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1990 		union tcp_log_stackspecific log;
1991 
1992 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1993 		log.u_bbr.flex1 = tso;
1994 		log.u_bbr.flex2 = maxseg;
1995 		log.u_bbr.flex3 = mtu;
1996 		log.u_bbr.flex4 = csum_flags;
1997 		TCP_LOG_EVENTP(tp, NULL,
1998 		    &bbr->rc_inp->inp_socket->so_rcv,
1999 		    &bbr->rc_inp->inp_socket->so_snd,
2000 		    BBR_LOG_MSGSIZE, 0,
2001 		    0, &log, false, &bbr->rc_tv);
2002 	}
2003 }
2004 
2005 static void
2006 bbr_log_flowend(struct tcp_bbr *bbr)
2007 {
2008 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2009 		union tcp_log_stackspecific log;
2010 		struct sockbuf *r, *s;
2011 		struct timeval tv;
2012 
2013 		if (bbr->rc_inp->inp_socket) {
2014 			r = &bbr->rc_inp->inp_socket->so_rcv;
2015 			s = &bbr->rc_inp->inp_socket->so_snd;
2016 		} else {
2017 			r = s = NULL;
2018 		}
2019 		bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv));
2020 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2021 		    r, s,
2022 		    TCP_LOG_FLOWEND, 0,
2023 		    0, &log, false, &tv);
2024 	}
2025 }
2026 
2027 static void
2028 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line,
2029     uint32_t lost, uint32_t del)
2030 {
2031 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2032 		union tcp_log_stackspecific log;
2033 
2034 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2035 		log.u_bbr.flex1 = lost;
2036 		log.u_bbr.flex2 = del;
2037 		log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw;
2038 		log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt;
2039 		log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2040 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2041 		log.u_bbr.flex7 = line;
2042 		log.u_bbr.flex8 = 0;
2043 		log.u_bbr.inflight = bbr->r_ctl.r_measurement_count;
2044 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2045 		    &bbr->rc_inp->inp_socket->so_rcv,
2046 		    &bbr->rc_inp->inp_socket->so_snd,
2047 		    BBR_LOG_PKT_EPOCH, 0,
2048 		    0, &log, false, &bbr->rc_tv);
2049 	}
2050 }
2051 
2052 static void
2053 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time)
2054 {
2055 	if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2056 		union tcp_log_stackspecific log;
2057 
2058 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2059 		log.u_bbr.flex1 = bbr->r_ctl.rc_lost;
2060 		log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat;
2061 		log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat;
2062 		log.u_bbr.flex7 = line;
2063 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2064 		    &bbr->rc_inp->inp_socket->so_rcv,
2065 		    &bbr->rc_inp->inp_socket->so_snd,
2066 		    BBR_LOG_TIME_EPOCH, 0,
2067 		    0, &log, false, &bbr->rc_tv);
2068 	}
2069 }
2070 
2071 static void
2072 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth)
2073 {
2074 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2075 		union tcp_log_stackspecific log;
2076 
2077 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2078 		log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2079 		log.u_bbr.flex2 = new_tar;
2080 		log.u_bbr.flex3 = line;
2081 		log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2082 		log.u_bbr.flex5 = bbr_quanta;
2083 		log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs;
2084 		log.u_bbr.flex7 = bbr->rc_last_options;
2085 		log.u_bbr.flex8 = meth;
2086 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2087 		    &bbr->rc_inp->inp_socket->so_rcv,
2088 		    &bbr->rc_inp->inp_socket->so_snd,
2089 		    BBR_LOG_STATE_TARGET, 0,
2090 		    0, &log, false, &bbr->rc_tv);
2091 	}
2092 
2093 }
2094 
2095 static void
2096 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2097 {
2098 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2099 		union tcp_log_stackspecific log;
2100 
2101 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2102 		log.u_bbr.flex1 = line;
2103 		log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2104 		log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2105 		if (bbr_state_is_pkt_epoch)
2106 			log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
2107 		else
2108 			log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP);
2109 		log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2110 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2111 		log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000);
2112 		log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra;
2113 		log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state;
2114 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2115 		    &bbr->rc_inp->inp_socket->so_rcv,
2116 		    &bbr->rc_inp->inp_socket->so_snd,
2117 		    BBR_LOG_STATE, 0,
2118 		    0, &log, false, &bbr->rc_tv);
2119 	}
2120 }
2121 
2122 static void
2123 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
2124 		    uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond)
2125 {
2126 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2127 		union tcp_log_stackspecific log;
2128 
2129 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2130 		log.u_bbr.flex1 = line;
2131 		log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2132 		log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt;
2133 		log.u_bbr.flex4 = applied;
2134 		log.u_bbr.flex5 = rtt;
2135 		log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2136 		log.u_bbr.flex7 = cond;
2137 		log.u_bbr.flex8 = reas;
2138 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2139 		    &bbr->rc_inp->inp_socket->so_rcv,
2140 		    &bbr->rc_inp->inp_socket->so_snd,
2141 		    BBR_LOG_RTT_SHRINKS, 0,
2142 		    0, &log, false, &bbr->rc_tv);
2143 	}
2144 }
2145 
2146 static void
2147 bbr_log_type_exit_rec(struct tcp_bbr *bbr)
2148 {
2149 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2150 		union tcp_log_stackspecific log;
2151 
2152 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2153 		log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start;
2154 		log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
2155 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2156 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2157 		    &bbr->rc_inp->inp_socket->so_rcv,
2158 		    &bbr->rc_inp->inp_socket->so_snd,
2159 		    BBR_LOG_EXITREC, 0,
2160 		    0, &log, false, &bbr->rc_tv);
2161 	}
2162 }
2163 
2164 static void
2165 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg,
2166     uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line)
2167 {
2168 	if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2169 		union tcp_log_stackspecific log;
2170 
2171 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2172 		log.u_bbr.flex1 = line;
2173 		log.u_bbr.flex2 = prev_acked;
2174 		log.u_bbr.flex3 = bytes_this_ack;
2175 		log.u_bbr.flex4 = chg;
2176 		log.u_bbr.flex5 = th_ack;
2177 		log.u_bbr.flex6 = target;
2178 		log.u_bbr.flex8 = meth;
2179 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2180 		    &bbr->rc_inp->inp_socket->so_rcv,
2181 		    &bbr->rc_inp->inp_socket->so_snd,
2182 		    BBR_LOG_CWND, 0,
2183 		    0, &log, false, &bbr->rc_tv);
2184 	}
2185 }
2186 
2187 static void
2188 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin)
2189 {
2190 	/*
2191 	 * Log the rtt sample we are applying to the srtt algorithm in
2192 	 * useconds.
2193 	 */
2194 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2195 		union tcp_log_stackspecific log;
2196 
2197 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2198 		log.u_bbr.flex1 = rtt;
2199 		log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time;
2200 		log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay;
2201 		log.u_bbr.flex4 = bbr->rc_tp->ts_offset;
2202 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2203 		log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv);
2204 		log.u_bbr.flex6 = tsin;
2205 		log.u_bbr.flex7 = 0;
2206 		log.u_bbr.flex8 = bbr->rc_ack_was_delayed;
2207 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2208 		    &bbr->rc_inp->inp_socket->so_rcv,
2209 		    &bbr->rc_inp->inp_socket->so_snd,
2210 		    TCP_LOG_RTT, 0,
2211 		    0, &log, false, &bbr->rc_tv);
2212 	}
2213 }
2214 
2215 static void
2216 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit)
2217 {
2218 	if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2219 		union tcp_log_stackspecific log;
2220 
2221 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2222 		log.u_bbr.flex1 = time_in;
2223 		log.u_bbr.flex2 = line;
2224 		log.u_bbr.flex8 = enter_exit;
2225 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2226 		    &bbr->rc_inp->inp_socket->so_rcv,
2227 		    &bbr->rc_inp->inp_socket->so_snd,
2228 		    BBR_LOG_PERSIST, 0,
2229 		    0, &log, false, &bbr->rc_tv);
2230 	}
2231 }
2232 static void
2233 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts)
2234 {
2235 	if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2236 		union tcp_log_stackspecific log;
2237 
2238 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2239 		log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age;
2240 		log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2241 		log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2242 		log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time;
2243 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2244 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2245 		    &bbr->rc_inp->inp_socket->so_rcv,
2246 		    &bbr->rc_inp->inp_socket->so_snd,
2247 		    BBR_LOG_ACKCLEAR, 0,
2248 		    0, &log, false, &bbr->rc_tv);
2249 	}
2250 }
2251 
2252 static void
2253 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen,
2254 		  uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m)
2255 {
2256 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2257 		union tcp_log_stackspecific log;
2258 		struct timeval tv;
2259 
2260 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2261 		log.u_bbr.flex1 = nsegs;
2262 		log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes;
2263 		if (m) {
2264 			struct timespec ts;
2265 
2266 			log.u_bbr.flex3 = m->m_flags;
2267 			if (m->m_flags & M_TSTMP) {
2268 				mbuf_tstmp2timespec(m, &ts);
2269 				tv.tv_sec = ts.tv_sec;
2270 				tv.tv_usec = ts.tv_nsec / 1000;
2271 				log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv);
2272 			} else {
2273 				log.u_bbr.lt_epoch = 0;
2274 			}
2275 			if (m->m_flags & M_TSTMP_LRO) {
2276 				tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000;
2277 				tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000;
2278 				log.u_bbr.flex5 = tcp_tv_to_usectick(&tv);
2279 			} else {
2280 				/* No arrival timestamp */
2281 				log.u_bbr.flex5 = 0;
2282 			}
2283 
2284 			log.u_bbr.pkts_out = tcp_get_usecs(&tv);
2285 		} else {
2286 			log.u_bbr.flex3 = 0;
2287 			log.u_bbr.flex5 = 0;
2288 			log.u_bbr.flex6 = 0;
2289 			log.u_bbr.pkts_out = 0;
2290 		}
2291 		log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state;
2292 		log.u_bbr.flex7 = bbr->r_wanted_output;
2293 		log.u_bbr.flex8 = bbr->rc_in_persist;
2294 		TCP_LOG_EVENTP(bbr->rc_tp, th,
2295 		    &bbr->rc_inp->inp_socket->so_rcv,
2296 		    &bbr->rc_inp->inp_socket->so_snd,
2297 		    TCP_LOG_IN, 0,
2298 		    tlen, &log, true, &bbr->rc_tv);
2299 	}
2300 }
2301 
2302 static void
2303 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out)
2304 {
2305 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2306 		union tcp_log_stackspecific log;
2307 
2308 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2309 		log.u_bbr.flex1 = did_out;
2310 		log.u_bbr.flex2 = nxt_pkt;
2311 		log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val;
2312 		log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2313 		log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp;
2314 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes;
2315 		log.u_bbr.flex7 = bbr->r_wanted_output;
2316 		log.u_bbr.flex8 = bbr->rc_in_persist;
2317 		log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay;
2318 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2319 		    &bbr->rc_inp->inp_socket->so_rcv,
2320 		    &bbr->rc_inp->inp_socket->so_snd,
2321 		    BBR_LOG_DOSEG_DONE, 0,
2322 		    0, &log, true, &bbr->rc_tv);
2323 	}
2324 }
2325 
2326 static void
2327 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts,
2328     int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz)
2329 {
2330 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2331 		union tcp_log_stackspecific log;
2332 
2333 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2334 		log.u_bbr.flex1 = line;
2335 		log.u_bbr.flex2 = o_len;
2336 		log.u_bbr.flex3 = segcnt;
2337 		log.u_bbr.flex4 = segsiz;
2338 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2339 		    &bbr->rc_inp->inp_socket->so_rcv,
2340 		    &bbr->rc_inp->inp_socket->so_snd,
2341 		    BBR_LOG_ENOBUF_JMP, ENOBUFS,
2342 		    len, &log, true, &bbr->rc_tv);
2343 	}
2344 }
2345 
2346 static void
2347 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling)
2348 {
2349 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2350 		union tcp_log_stackspecific log;
2351 
2352 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2353 		log.u_bbr.flex1 = timers;
2354 		log.u_bbr.flex2 = ret;
2355 		log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
2356 		log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2357 		log.u_bbr.flex5 = cts;
2358 		log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2359 		log.u_bbr.flex8 = hpts_calling;
2360 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2361 		    &bbr->rc_inp->inp_socket->so_rcv,
2362 		    &bbr->rc_inp->inp_socket->so_snd,
2363 		    BBR_LOG_TO_PROCESS, 0,
2364 		    0, &log, false, &bbr->rc_tv);
2365 	}
2366 }
2367 
2368 static void
2369 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num)
2370 {
2371 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2372 		union tcp_log_stackspecific log;
2373 		uint64_t ar;
2374 
2375 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2376 		log.u_bbr.flex1 = bbr->bbr_timer_src;
2377 		log.u_bbr.flex2 = 0;
2378 		log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2379 		ar = (uint64_t)(bbr->r_ctl.rc_resend);
2380 		ar >>= 32;
2381 		ar &= 0x00000000ffffffff;
2382 		log.u_bbr.flex4 = (uint32_t)ar;
2383 		ar = (uint64_t)bbr->r_ctl.rc_resend;
2384 		ar &= 0x00000000ffffffff;
2385 		log.u_bbr.flex5 = (uint32_t)ar;
2386 		log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2387 		log.u_bbr.flex8 = to_num;
2388 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2389 		    &bbr->rc_inp->inp_socket->so_rcv,
2390 		    &bbr->rc_inp->inp_socket->so_snd,
2391 		    BBR_LOG_RTO, 0,
2392 		    0, &log, false, &bbr->rc_tv);
2393 	}
2394 }
2395 
2396 static void
2397 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason)
2398 {
2399 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2400 		union tcp_log_stackspecific log;
2401 
2402 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2403 		log.u_bbr.flex1 = flex1;
2404 		log.u_bbr.flex2 = flex2;
2405 		log.u_bbr.flex3 = flex3;
2406 		log.u_bbr.flex4 = 0;
2407 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2408 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2409 		log.u_bbr.flex8 = reason;
2410 		log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw;
2411 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2412 		    &bbr->rc_inp->inp_socket->so_rcv,
2413 		    &bbr->rc_inp->inp_socket->so_snd,
2414 		    BBR_LOG_REDUCE, 0,
2415 		    0, &log, false, &bbr->rc_tv);
2416 	}
2417 }
2418 
2419 static void
2420 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag)
2421 {
2422 	if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2423 		union tcp_log_stackspecific log;
2424 
2425 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2426 		log.u_bbr.flex1 = diag->p_nxt_slot;
2427 		log.u_bbr.flex2 = diag->p_cur_slot;
2428 		log.u_bbr.flex3 = diag->slot_req;
2429 		log.u_bbr.flex4 = diag->inp_hptsslot;
2430 		log.u_bbr.flex5 = diag->slot_remaining;
2431 		log.u_bbr.flex6 = diag->need_new_to;
2432 		log.u_bbr.flex7 = diag->p_hpts_active;
2433 		log.u_bbr.flex8 = diag->p_on_min_sleep;
2434 		/* Hijack other fields as needed  */
2435 		log.u_bbr.epoch = diag->have_slept;
2436 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
2437 		log.u_bbr.pkts_out = diag->co_ret;
2438 		log.u_bbr.applimited = diag->hpts_sleep_time;
2439 		log.u_bbr.delivered = diag->p_prev_slot;
2440 		log.u_bbr.inflight = diag->p_runningtick;
2441 		log.u_bbr.bw_inuse = diag->wheel_tick;
2442 		log.u_bbr.rttProp = diag->wheel_cts;
2443 		log.u_bbr.delRate = diag->maxticks;
2444 		log.u_bbr.cur_del_rate = diag->p_curtick;
2445 		log.u_bbr.cur_del_rate <<= 32;
2446 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
2447 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2448 		    &bbr->rc_inp->inp_socket->so_rcv,
2449 		    &bbr->rc_inp->inp_socket->so_snd,
2450 		    BBR_LOG_HPTSDIAG, 0,
2451 		    0, &log, false, &bbr->rc_tv);
2452 	}
2453 }
2454 
2455 static void
2456 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
2457     uint32_t thresh, uint32_t to)
2458 {
2459 	if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2460 		union tcp_log_stackspecific log;
2461 
2462 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2463 		log.u_bbr.flex1 = bbr->rc_tp->t_rttvar;
2464 		log.u_bbr.flex2 = time_since_sent;
2465 		log.u_bbr.flex3 = srtt;
2466 		log.u_bbr.flex4 = thresh;
2467 		log.u_bbr.flex5 = to;
2468 		log.u_bbr.flex6 = bbr->rc_tp->t_srtt;
2469 		log.u_bbr.flex8 = mode;
2470 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2471 		    &bbr->rc_inp->inp_socket->so_rcv,
2472 		    &bbr->rc_inp->inp_socket->so_snd,
2473 		    BBR_LOG_TIMERPREP, 0,
2474 		    0, &log, false, &bbr->rc_tv);
2475 	}
2476 }
2477 
2478 static void
2479 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
2480     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod)
2481 {
2482 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2483 		union tcp_log_stackspecific log;
2484 
2485 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2486 		log.u_bbr.flex1 = usecs;
2487 		log.u_bbr.flex2 = len;
2488 		log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff);
2489 		log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff);
2490 		if (override)
2491 			log.u_bbr.flex5 = (1 << 2);
2492 		else
2493 			log.u_bbr.flex5 = 0;
2494 		log.u_bbr.flex6 = override;
2495 		log.u_bbr.flex7 = gain;
2496 		log.u_bbr.flex8 = mod;
2497 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2498 		    &bbr->rc_inp->inp_socket->so_rcv,
2499 		    &bbr->rc_inp->inp_socket->so_snd,
2500 		    BBR_LOG_HPTSI_CALC, 0,
2501 		    len, &log, false, &bbr->rc_tv);
2502 	}
2503 }
2504 
2505 static void
2506 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2507 {
2508 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2509 		union tcp_log_stackspecific log;
2510 
2511 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2512 
2513 		log.u_bbr.flex1 = bbr->bbr_timer_src;
2514 		log.u_bbr.flex2 = to;
2515 		log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2516 		log.u_bbr.flex4 = slot;
2517 		log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot;
2518 		log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2519 		log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2;
2520 		log.u_bbr.flex8 = which;
2521 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2522 		    &bbr->rc_inp->inp_socket->so_rcv,
2523 		    &bbr->rc_inp->inp_socket->so_snd,
2524 		    BBR_LOG_TIMERSTAR, 0,
2525 		    0, &log, false, &bbr->rc_tv);
2526 	}
2527 }
2528 
2529 static void
2530 bbr_log_thresh_choice(struct tcp_bbr *bbr, uint32_t cts, uint32_t thresh, uint32_t lro, uint32_t srtt, struct bbr_sendmap *rsm, uint8_t frm)
2531 {
2532 	if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2533 		union tcp_log_stackspecific log;
2534 
2535 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2536 		log.u_bbr.flex1 = thresh;
2537 		log.u_bbr.flex2 = lro;
2538 		log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts;
2539 		log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
2540 		log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2541 		log.u_bbr.flex6 = srtt;
2542 		log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift;
2543 		log.u_bbr.flex8 = frm;
2544 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2545 		    &bbr->rc_inp->inp_socket->so_rcv,
2546 		    &bbr->rc_inp->inp_socket->so_snd,
2547 		    BBR_LOG_THRESH_CALC, 0,
2548 		    0, &log, false, &bbr->rc_tv);
2549 	}
2550 }
2551 
2552 static void
2553 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed)
2554 {
2555 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2556 		union tcp_log_stackspecific log;
2557 
2558 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2559 		log.u_bbr.flex1 = line;
2560 		log.u_bbr.flex2 = bbr->bbr_timer_src;
2561 		log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2562 		log.u_bbr.flex4 = bbr->rc_in_persist;
2563 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2564 		log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2565 		log.u_bbr.flex8 = hpts_removed;
2566 		log.u_bbr.pkts_out = bbr->rc_pacer_started;
2567 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2568 		    &bbr->rc_inp->inp_socket->so_rcv,
2569 		    &bbr->rc_inp->inp_socket->so_snd,
2570 		    BBR_LOG_TIMERCANC, 0,
2571 		    0, &log, false, &bbr->rc_tv);
2572 	}
2573 }
2574 
2575 
2576 static void
2577 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta)
2578 {
2579 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2580 		union tcp_log_stackspecific log;
2581 
2582 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2583 		log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio;
2584 		log.u_bbr.flex2 = (peer_delta >> 32);
2585 		log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff);
2586 		log.u_bbr.flex4 = (delta >> 32);
2587 		log.u_bbr.flex5 = (delta & 0x00000000ffffffff);
2588 		log.u_bbr.flex7 = bbr->rc_ts_clock_set;
2589 		log.u_bbr.flex8 = bbr->rc_ts_cant_be_used;
2590 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2591 		    &bbr->rc_inp->inp_socket->so_rcv,
2592 		    &bbr->rc_inp->inp_socket->so_snd,
2593 		    BBR_LOG_TSTMP_VAL, 0,
2594 		    0, &log, false, &bbr->rc_tv);
2595 
2596 	}
2597 }
2598 
2599 static void
2600 bbr_log_type_tsosize(struct tcp_bbr *bbr, uint32_t cts, uint32_t tsosz, uint32_t tls, uint32_t old_val, uint32_t maxseg, int hdwr)
2601 {
2602 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2603 		union tcp_log_stackspecific log;
2604 
2605 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2606 		log.u_bbr.flex1 = tsosz;
2607 		log.u_bbr.flex2 = tls;
2608 		log.u_bbr.flex3 = tcp_min_hptsi_time;
2609 		log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min;
2610 		log.u_bbr.flex5 = old_val;
2611 		log.u_bbr.flex6 = maxseg;
2612 		log.u_bbr.flex7 = bbr->rc_no_pacing;
2613 		log.u_bbr.flex7 <<= 1;
2614 		log.u_bbr.flex7 |= bbr->rc_past_init_win;
2615 		if (hdwr)
2616 			log.u_bbr.flex8 = 0x80 | bbr->rc_use_google;
2617 		else
2618 			log.u_bbr.flex8 = bbr->rc_use_google;
2619 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2620 		    &bbr->rc_inp->inp_socket->so_rcv,
2621 		    &bbr->rc_inp->inp_socket->so_snd,
2622 		    BBR_LOG_BBRTSO, 0,
2623 		    0, &log, false, &bbr->rc_tv);
2624 	}
2625 }
2626 
2627 static void
2628 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm,
2629 		      uint32_t flags, uint32_t line)
2630 {
2631 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2632 		union tcp_log_stackspecific log;
2633 
2634 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2635 		log.u_bbr.flex1 = line;
2636 		log.u_bbr.flex2 = rsm->r_start;
2637 		log.u_bbr.flex3 = rsm->r_end;
2638 		log.u_bbr.flex4 = rsm->r_delivered;
2639 		log.u_bbr.flex5 = rsm->r_rtr_cnt;
2640 		log.u_bbr.flex6 = rsm->r_dupack;
2641 		log.u_bbr.flex7 = rsm->r_tim_lastsent[0];
2642 		log.u_bbr.flex8 = rsm->r_flags;
2643 		/* Hijack the pkts_out fids */
2644 		log.u_bbr.applimited = flags;
2645 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2646 		    &bbr->rc_inp->inp_socket->so_rcv,
2647 		    &bbr->rc_inp->inp_socket->so_snd,
2648 		    BBR_RSM_CLEARED, 0,
2649 		    0, &log, false, &bbr->rc_tv);
2650 	}
2651 }
2652 
2653 static void
2654 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts,
2655     uint32_t flex3, uint32_t flex2, uint32_t flex5,
2656     uint32_t flex6, uint32_t pkts_out, int flex7,
2657     uint32_t flex4, uint32_t flex1)
2658 {
2659 
2660 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2661 		union tcp_log_stackspecific log;
2662 
2663 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2664 		log.u_bbr.flex1 = flex1;
2665 		log.u_bbr.flex2 = flex2;
2666 		log.u_bbr.flex3 = flex3;
2667 		log.u_bbr.flex4 = flex4;
2668 		log.u_bbr.flex5 = flex5;
2669 		log.u_bbr.flex6 = flex6;
2670 		log.u_bbr.flex7 = flex7;
2671 		/* Hijack the pkts_out fids */
2672 		log.u_bbr.pkts_out = pkts_out;
2673 		log.u_bbr.flex8 = flex8;
2674 		if (bbr->rc_ack_was_delayed)
2675 			log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay;
2676 		else
2677 			log.u_bbr.epoch = 0;
2678 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2679 		    &bbr->rc_inp->inp_socket->so_rcv,
2680 		    &bbr->rc_inp->inp_socket->so_snd,
2681 		    BBR_LOG_BBRUPD, 0,
2682 		    flex2, &log, false, &bbr->rc_tv);
2683 	}
2684 }
2685 
2686 
2687 static void
2688 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason,
2689 	uint32_t newbw, uint32_t obw, uint32_t diff,
2690 	uint32_t tim)
2691 {
2692 	if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2693 		union tcp_log_stackspecific log;
2694 
2695 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2696 		log.u_bbr.flex1 = reason;
2697 		log.u_bbr.flex2 = newbw;
2698 		log.u_bbr.flex3 = obw;
2699 		log.u_bbr.flex4 = diff;
2700 		log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost;
2701 		log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del;
2702 		log.u_bbr.flex7 = bbr->rc_lt_is_sampling;
2703 		log.u_bbr.pkts_out = tim;
2704 		log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw;
2705 		if (bbr->rc_lt_use_bw == 0)
2706 			log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
2707 		else
2708 			log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
2709 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2710 		    &bbr->rc_inp->inp_socket->so_rcv,
2711 		    &bbr->rc_inp->inp_socket->so_snd,
2712 		    BBR_LOG_BWSAMP, 0,
2713 		    0, &log, false, &bbr->rc_tv);
2714 	}
2715 }
2716 
2717 static inline void
2718 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line)
2719 {
2720 	if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2721 		union tcp_log_stackspecific log;
2722 
2723 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2724 		log.u_bbr.flex1 = line;
2725 		log.u_bbr.flex2 = tick;
2726 		log.u_bbr.flex3 = tp->t_maxunacktime;
2727 		log.u_bbr.flex4 = tp->t_acktime;
2728 		log.u_bbr.flex8 = event;
2729 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2730 		    &bbr->rc_inp->inp_socket->so_rcv,
2731 		    &bbr->rc_inp->inp_socket->so_snd,
2732 		    BBR_LOG_PROGRESS, 0,
2733 		    0, &log, false, &bbr->rc_tv);
2734 	}
2735 }
2736 
2737 static void
2738 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp,
2739 			 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts,
2740 			 int error)
2741 {
2742 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2743 		union tcp_log_stackspecific log;
2744 
2745 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2746 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2747 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2748 		log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2749 		log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2750 		log.u_bbr.bw_inuse = rate;
2751 		log.u_bbr.flex5 = line;
2752 		log.u_bbr.flex6 = error;
2753 		log.u_bbr.flex8 = bbr->skip_gain;
2754 		log.u_bbr.flex8 <<= 1;
2755 		log.u_bbr.flex8 |= bbr->gain_is_limited;
2756 		log.u_bbr.flex8 <<= 1;
2757 		log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing;
2758 		log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
2759 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2760 		    &bbr->rc_inp->inp_socket->so_rcv,
2761 		    &bbr->rc_inp->inp_socket->so_snd,
2762 		    BBR_LOG_HDWR_PACE, 0,
2763 		    0, &log, false, &bbr->rc_tv);
2764 	}
2765 }
2766 
2767 static void
2768 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t line, uint32_t prev_delay)
2769 {
2770 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2771 		union tcp_log_stackspecific log;
2772 
2773 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2774 		log.u_bbr.flex1 = slot;
2775 		log.u_bbr.flex2 = del_by;
2776 		log.u_bbr.flex3 = prev_delay;
2777 		log.u_bbr.flex4 = line;
2778 		log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val;
2779 		log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay;
2780 		log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags);
2781 		log.u_bbr.flex8 = bbr->rc_in_persist;
2782 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2783 		    &bbr->rc_inp->inp_socket->so_rcv,
2784 		    &bbr->rc_inp->inp_socket->so_snd,
2785 		    BBR_LOG_BBRSND, 0,
2786 		    len, &log, false, &bbr->rc_tv);
2787 	}
2788 }
2789 
2790 static void
2791 bbr_log_type_bbrrttprop(struct tcp_bbr *bbr, uint32_t t, uint32_t end, uint32_t tsconv, uint32_t cts, int32_t match, uint32_t seq, uint8_t flags)
2792 {
2793 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2794 		union tcp_log_stackspecific log;
2795 
2796 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2797 		log.u_bbr.flex1 = bbr->r_ctl.rc_delivered;
2798 		log.u_bbr.flex2 = 0;
2799 		log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt;
2800 		log.u_bbr.flex4 = end;
2801 		log.u_bbr.flex5 = seq;
2802 		log.u_bbr.flex6 = t;
2803 		log.u_bbr.flex7 = match;
2804 		log.u_bbr.flex8 = flags;
2805 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2806 		    &bbr->rc_inp->inp_socket->so_rcv,
2807 		    &bbr->rc_inp->inp_socket->so_snd,
2808 		    BBR_LOG_BBRRTT, 0,
2809 		    0, &log, false, &bbr->rc_tv);
2810 	}
2811 }
2812 
2813 static void
2814 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method)
2815 {
2816 	if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2817 		union tcp_log_stackspecific log;
2818 
2819 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2820 		log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2821 		log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
2822 		log.u_bbr.flex3 = bbr->r_ctl.gain_epoch;
2823 		log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2824 		log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs;
2825 		log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight;
2826 		log.u_bbr.flex7 = 0;
2827 		log.u_bbr.flex8 = entry_method;
2828 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2829 		    &bbr->rc_inp->inp_socket->so_rcv,
2830 		    &bbr->rc_inp->inp_socket->so_snd,
2831 		    BBR_LOG_EXIT_GAIN, 0,
2832 		    0, &log, false, &bbr->rc_tv);
2833 	}
2834 }
2835 
2836 static void
2837 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired)
2838 {
2839 	if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2840 		union tcp_log_stackspecific log;
2841 
2842 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2843 		/* R-HU */
2844 		log.u_bbr.flex1 = 0;
2845 		log.u_bbr.flex2 = 0;
2846 		log.u_bbr.flex3 = 0;
2847 		log.u_bbr.flex4 = 0;
2848 		log.u_bbr.flex7 = 0;
2849 		log.u_bbr.flex8 = settings_desired;
2850 
2851 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2852 		    &bbr->rc_inp->inp_socket->so_rcv,
2853 		    &bbr->rc_inp->inp_socket->so_snd,
2854 		    BBR_LOG_SETTINGS_CHG, 0,
2855 		    0, &log, false, &bbr->rc_tv);
2856 	}
2857 }
2858 
2859 /*
2860  * Returns the bw from the our filter.
2861  */
2862 static inline uint64_t
2863 bbr_get_full_bw(struct tcp_bbr *bbr)
2864 {
2865 	uint64_t bw;
2866 
2867 	bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2868 
2869 	return (bw);
2870 }
2871 
2872 static inline void
2873 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2874 {
2875 	uint64_t calclr;
2876 	uint32_t lost, del;
2877 
2878 	if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch)
2879 		lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch;
2880 	else
2881 		lost = 0;
2882 	del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del;
2883 	if (lost == 0)  {
2884 		calclr = 0;
2885 	} else if (del) {
2886 		calclr = lost;
2887 		calclr *= (uint64_t)1000;
2888 		calclr /= (uint64_t)del;
2889 	} else {
2890 		/* Nothing delivered? 100.0% loss */
2891 		calclr = 1000;
2892 	}
2893 	bbr->r_ctl.rc_pkt_epoch_loss_rate =  (uint32_t)calclr;
2894 	if (IN_RECOVERY(bbr->rc_tp->t_flags))
2895 		bbr->r_ctl.recovery_lr += (uint32_t)calclr;
2896 	bbr->r_ctl.rc_pkt_epoch++;
2897 	if (bbr->rc_no_pacing &&
2898 	    (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) {
2899 		bbr->rc_no_pacing = 0;
2900 		tcp_bbr_tso_size_check(bbr, cts);
2901 	}
2902 	bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time);
2903 	bbr->r_ctl.rc_pkt_epoch_time = cts;
2904 	/* What was our loss rate */
2905 	bbr_log_pkt_epoch(bbr, cts, line, lost, del);
2906 	bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered;
2907 	bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost;
2908 }
2909 
2910 static inline void
2911 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2912 {
2913 	uint32_t epoch_time;
2914 
2915 	/* Tick the RTT clock */
2916 	bbr->r_ctl.rc_rtt_epoch++;
2917 	epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start;
2918 	bbr_log_time_epoch(bbr, cts, line, epoch_time);
2919 	bbr->r_ctl.rc_rcv_epoch_start = cts;
2920 }
2921 
2922 
2923 static inline void
2924 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked)
2925 {
2926 	if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) {
2927 		bbr->rc_is_pkt_epoch_now = 1;
2928 	}
2929 }
2930 
2931 /*
2932  * Returns the bw from either the b/w filter
2933  * or from the lt_bw (if the connection is being
2934  * policed).
2935  */
2936 static inline uint64_t
2937 __bbr_get_bw(struct tcp_bbr *bbr)
2938 {
2939 	uint64_t bw, min_bw;
2940 	uint64_t rtt;
2941 	int gm_measure_cnt = 1;
2942 
2943 	/*
2944 	 * For startup we make, like google, a
2945 	 * minimum b/w. This is generated from the
2946 	 * IW and the rttProp. We do fall back to srtt
2947 	 * if for some reason (initial handshake) we don't
2948 	 * have a rttProp. We, in the worst case, fall back
2949 	 * to the configured min_bw (rc_initial_hptsi_bw).
2950 	 */
2951 	if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
2952 		/* Attempt first to use rttProp */
2953 		rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2954 		if (rtt && (rtt < 0xffffffff)) {
2955 measure:
2956 			min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2957 				((uint64_t)1000000);
2958 			min_bw /= rtt;
2959 			if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2960 				min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2961 			}
2962 
2963 		} else if (bbr->rc_tp->t_srtt != 0) {
2964 			/* No rttProp, use srtt? */
2965 			rtt = bbr_get_rtt(bbr, BBR_SRTT);
2966 			goto measure;
2967 		} else {
2968 			min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2969 		}
2970 	} else
2971 		min_bw = 0;
2972 
2973 	if ((bbr->rc_past_init_win == 0) &&
2974 	    (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp)))
2975 		bbr->rc_past_init_win = 1;
2976 	if ((bbr->rc_use_google)  && (bbr->r_ctl.r_measurement_count >= 1))
2977 		gm_measure_cnt = 0;
2978 	if (gm_measure_cnt &&
2979 	    ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) ||
2980 	     (bbr->rc_past_init_win == 0))) {
2981 		/* For google we use our guess rate until we get 1 measurement */
2982 
2983 use_initial_window:
2984 		rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2985 		if (rtt && (rtt < 0xffffffff)) {
2986 			/*
2987 			 * We have an RTT measurment. Use that in
2988 			 * combination with our initial window to calculate
2989 			 * a b/w.
2990 			 */
2991 			bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2992 				((uint64_t)1000000);
2993 			bw /= rtt;
2994 			if (bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2995 				bw = bbr->r_ctl.rc_initial_hptsi_bw;
2996 			}
2997 		} else {
2998 			/* Drop back to the 40 and punt to a default */
2999 			bw = bbr->r_ctl.rc_initial_hptsi_bw;
3000 		}
3001 		if (bw < 1)
3002 			/* Probably should panic */
3003 			bw = 1;
3004 		if (bw > min_bw)
3005 			return (bw);
3006 		else
3007 			return (min_bw);
3008 	}
3009 	if (bbr->rc_lt_use_bw)
3010 		bw = bbr->r_ctl.rc_lt_bw;
3011 	else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0))
3012 		bw = bbr->r_ctl.red_bw;
3013 	else
3014 		bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3015 	if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) {
3016 		/*
3017 		 * Enforce user set rate limit, keep in mind that
3018 		 * t_peakrate_thr is in B/s already
3019 		 */
3020 		bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw);
3021 	}
3022 	if (bw == 0) {
3023 		/* We should not be at 0, go to the initial window then  */
3024 		goto use_initial_window;
3025 	}
3026 	if (bw < 1)
3027 		/* Probably should panic */
3028 		bw = 1;
3029 	if (bw < min_bw)
3030 		bw = min_bw;
3031 	return (bw);
3032 }
3033 
3034 static inline uint64_t
3035 bbr_get_bw(struct tcp_bbr *bbr)
3036 {
3037 	uint64_t bw;
3038 
3039 	bw = __bbr_get_bw(bbr);
3040 	return (bw);
3041 }
3042 
3043 static inline void
3044 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts)
3045 {
3046 	bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch;
3047 	bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time;
3048 	bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3049 	bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3050 }
3051 
3052 static inline void
3053 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts)
3054 {
3055 	bbr->rc_lt_is_sampling = 0;
3056 	bbr->rc_lt_use_bw = 0;
3057 	bbr->r_ctl.rc_lt_bw = 0;
3058 	bbr_reset_lt_bw_interval(bbr, cts);
3059 }
3060 
3061 static inline void
3062 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin)
3063 {
3064 	uint64_t diff;
3065 
3066 	/* Do we have a previous sample? */
3067 	if (bbr->r_ctl.rc_lt_bw) {
3068 		/* Get the diff in bytes per second */
3069 		if (bbr->r_ctl.rc_lt_bw > bw)
3070 			diff = bbr->r_ctl.rc_lt_bw - bw;
3071 		else
3072 			diff = bw - bbr->r_ctl.rc_lt_bw;
3073 		if ((diff <= bbr_lt_bw_diff) ||
3074 		    (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) {
3075 			/* Consider us policed */
3076 			uint32_t saved_bw;
3077 
3078 			saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw;
3079 			bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2;	/* average of two */
3080 			bbr->rc_lt_use_bw = 1;
3081 			bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
3082 			/*
3083 			 * Use pkt based epoch for measuring length of
3084 			 * policer up
3085 			 */
3086 			bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch;
3087 			/*
3088 			 * reason 4 is we need to start consider being
3089 			 * policed
3090 			 */
3091 			bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin);
3092 			return;
3093 		}
3094 	}
3095 	bbr->r_ctl.rc_lt_bw = bw;
3096 	bbr_reset_lt_bw_interval(bbr, cts);
3097 	bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin);
3098 }
3099 
3100 /*
3101  * RRS: Copied from user space!
3102  * Calculate a uniformly distributed random number less than upper_bound
3103  * avoiding "modulo bias".
3104  *
3105  * Uniformity is achieved by generating new random numbers until the one
3106  * returned is outside the range [0, 2**32 % upper_bound).  This
3107  * guarantees the selected random number will be inside
3108  * [2**32 % upper_bound, 2**32) which maps back to [0, upper_bound)
3109  * after reduction modulo upper_bound.
3110  */
3111 static uint32_t
3112 arc4random_uniform(uint32_t upper_bound)
3113 {
3114 	uint32_t r, min;
3115 
3116 	if (upper_bound < 2)
3117 		return 0;
3118 
3119 	/* 2**32 % x == (2**32 - x) % x */
3120 	min = -upper_bound % upper_bound;
3121 
3122 	/*
3123 	 * This could theoretically loop forever but each retry has
3124 	 * p > 0.5 (worst case, usually far better) of selecting a
3125 	 * number inside the range we need, so it should rarely need
3126 	 * to re-roll.
3127 	 */
3128 	for (;;) {
3129 		r = arc4random();
3130 		if (r >= min)
3131 			break;
3132 	}
3133 
3134 	return r % upper_bound;
3135 }
3136 
3137 static void
3138 bbr_randomize_extra_state_time(struct tcp_bbr *bbr)
3139 {
3140 	uint32_t ran, deduct;
3141 
3142 	ran = arc4random_uniform(bbr_rand_ot);
3143 	if (ran) {
3144 		deduct = bbr->r_ctl.rc_level_state_extra / ran;
3145 		bbr->r_ctl.rc_level_state_extra -= deduct;
3146 	}
3147 }
3148 /*
3149  * Return randomly the starting state
3150  * to use in probebw.
3151  */
3152 static uint8_t
3153 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts)
3154 {
3155 	uint32_t ran;
3156 	uint8_t ret_val;
3157 
3158 	/* Initialize the offset to 0 */
3159 	bbr->r_ctl.rc_exta_time_gd = 0;
3160 	bbr->rc_hit_state_1 = 0;
3161 	bbr->r_ctl.rc_level_state_extra = 0;
3162 	ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1));
3163 	/*
3164 	 * The math works funny here :) the return value is used to set the
3165 	 * substate and then the state change is called which increments by
3166 	 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when
3167 	 * we fully enter the state. Note that the (8 - 1 - ran) assures that
3168 	 * we return 1 - 7, so we dont return 0 and end up starting in
3169 	 * state 1 (DRAIN).
3170 	 */
3171 	ret_val = BBR_SUBSTATE_COUNT - 1 - ran;
3172 	/* Set an epoch */
3173 	if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP))
3174 		bbr_set_epoch(bbr, cts, __LINE__);
3175 
3176 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
3177 	return (ret_val);
3178 }
3179 
3180 static void
3181 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected)
3182 {
3183 	uint32_t diff, d_time;
3184 	uint64_t del_time, bw, lost, delivered;
3185 
3186 	if (bbr->r_use_policer == 0)
3187 		return;
3188 	if (bbr->rc_lt_use_bw) {
3189 		/* We are using lt bw do we stop yet? */
3190 		diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
3191 		if (diff > bbr_lt_bw_max_rtts) {
3192 			/* Reset it all */
3193 reset_all:
3194 			bbr_reset_lt_bw_sampling(bbr, cts);
3195 			if (bbr->rc_filled_pipe) {
3196 				bbr_set_epoch(bbr, cts, __LINE__);
3197 				bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
3198 				bbr_substate_change(bbr, cts, __LINE__, 0);
3199 				bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
3200 				bbr_log_type_statechange(bbr, cts, __LINE__);
3201 			} else {
3202 				/*
3203 				 * This should not happen really
3204 				 * unless we remove the startup/drain
3205 				 * restrictions above.
3206 				 */
3207 				bbr->rc_bbr_state = BBR_STATE_STARTUP;
3208 				bbr_set_epoch(bbr, cts, __LINE__);
3209 				bbr->r_ctl.rc_bbr_state_time = cts;
3210 				bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
3211 				bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
3212 				bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
3213 				bbr_set_state_target(bbr, __LINE__);
3214 				bbr_log_type_statechange(bbr, cts, __LINE__);
3215 			}
3216 			/* reason 0 is to stop using lt-bw */
3217 			bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0);
3218 			return;
3219 		}
3220 		if (bbr_lt_intvl_fp == 0) {
3221 			/* Not doing false-postive detection */
3222 			return;
3223 		}
3224 		/* False positive detection */
3225 		if (diff == bbr_lt_intvl_fp) {
3226 			/* At bbr_lt_intvl_fp we record the lost */
3227 			bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3228 			bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3229 		} else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) {
3230 			/* Now is our loss rate still high? */
3231 			lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3232 			delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3233 			if ((delivered == 0) ||
3234 			    (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) {
3235 				/* No still below our threshold */
3236 				bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0);
3237 			} else {
3238 				/* Yikes its still high, it must be a false positive */
3239 				bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0);
3240 				goto reset_all;
3241 			}
3242 		}
3243 		return;
3244 	}
3245 	/*
3246 	 * Wait for the first loss before sampling, to let the policer
3247 	 * exhaust its tokens and estimate the steady-state rate allowed by
3248 	 * the policer. Starting samples earlier includes bursts that
3249 	 * over-estimate the bw.
3250 	 */
3251 	if (bbr->rc_lt_is_sampling == 0) {
3252 		/* reason 1 is to begin doing the sampling  */
3253 		if (loss_detected == 0)
3254 			return;
3255 		bbr_reset_lt_bw_interval(bbr, cts);
3256 		bbr->rc_lt_is_sampling = 1;
3257 		bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0);
3258 		return;
3259 	}
3260 	/* Now how long were we delivering long term last> */
3261 	if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time))
3262 		d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time;
3263 	else
3264 		d_time = 0;
3265 
3266 	/* To avoid underestimates, reset sampling if we run out of data. */
3267 	if (bbr->r_ctl.r_app_limited_until) {
3268 		/* Can not measure in app-limited state */
3269 		bbr_reset_lt_bw_sampling(bbr, cts);
3270 		/* reason 2 is to reset sampling due to app limits  */
3271 		bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time);
3272 		return;
3273 	}
3274 	diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
3275 	if (diff < bbr_lt_intvl_min_rtts) {
3276 		/*
3277 		 * need more samples (we don't
3278 		 * start on a round like linux so
3279 		 * we need 1 more).
3280 		 */
3281 		/* 6 is not_enough time or no-loss */
3282 		bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3283 		return;
3284 	}
3285 	if (diff > (4 * bbr_lt_intvl_min_rtts)) {
3286 		/*
3287 		 * For now if we wait too long, reset all sampling. We need
3288 		 * to do some research here, its possible that we should
3289 		 * base this on how much loss as occurred.. something like
3290 		 * if its under 10% (or some thresh) reset all otherwise
3291 		 * don't.  Thats for phase II I guess.
3292 		 */
3293 		bbr_reset_lt_bw_sampling(bbr, cts);
3294  		/* reason 3 is to reset sampling due too long of sampling */
3295 		bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3296 		return;
3297 	}
3298 	/*
3299 	 * End sampling interval when a packet is lost, so we estimate the
3300 	 * policer tokens were exhausted. Stopping the sampling before the
3301 	 * tokens are exhausted under-estimates the policed rate.
3302 	 */
3303 	if (loss_detected == 0) {
3304 		/* 6 is not_enough time or no-loss */
3305 		bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3306 		return;
3307 	}
3308 	/* Calculate packets lost and delivered in sampling interval. */
3309 	lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3310 	delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3311 	if ((delivered == 0) ||
3312 	    (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) {
3313 		bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time);
3314 		return;
3315 	}
3316 	if (d_time < 1000) {
3317 		/* Not enough time. wait */
3318 		/* 6 is not_enough time or no-loss */
3319 		bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3320 		return;
3321 	}
3322 	if (d_time >= (0xffffffff / USECS_IN_MSEC)) {
3323 		/* Too long */
3324 		bbr_reset_lt_bw_sampling(bbr, cts);
3325  		/* reason 3 is to reset sampling due too long of sampling */
3326 		bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3327 		return;
3328 	}
3329 	del_time = d_time;
3330 	bw = delivered;
3331 	bw *= (uint64_t)USECS_IN_SECOND;
3332 	bw /= del_time;
3333 	bbr_lt_bw_samp_done(bbr, bw, cts, d_time);
3334 }
3335 
3336 /*
3337  * Allocate a sendmap from our zone.
3338  */
3339 static struct bbr_sendmap *
3340 bbr_alloc(struct tcp_bbr *bbr)
3341 {
3342 	struct bbr_sendmap *rsm;
3343 
3344 	BBR_STAT_INC(bbr_to_alloc);
3345 	rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO));
3346 	if (rsm) {
3347 		bbr->r_ctl.rc_num_maps_alloced++;
3348 		return (rsm);
3349 	}
3350 	if (bbr->r_ctl.rc_free_cnt) {
3351 		BBR_STAT_INC(bbr_to_alloc_emerg);
3352 		rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
3353 		TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
3354 		bbr->r_ctl.rc_free_cnt--;
3355 		return (rsm);
3356 	}
3357 	BBR_STAT_INC(bbr_to_alloc_failed);
3358 	return (NULL);
3359 }
3360 
3361 static struct bbr_sendmap *
3362 bbr_alloc_full_limit(struct tcp_bbr *bbr)
3363 {
3364 	if ((V_tcp_map_entries_limit > 0) &&
3365 	    (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3366 		BBR_STAT_INC(bbr_alloc_limited);
3367 		if (!bbr->alloc_limit_reported) {
3368 			bbr->alloc_limit_reported = 1;
3369 			BBR_STAT_INC(bbr_alloc_limited_conns);
3370 		}
3371 		return (NULL);
3372 	}
3373 	return (bbr_alloc(bbr));
3374 }
3375 
3376 
3377 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3378 static struct bbr_sendmap *
3379 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type)
3380 {
3381 	struct bbr_sendmap *rsm;
3382 
3383 	if (limit_type) {
3384 		/* currently there is only one limit type */
3385 		if (V_tcp_map_split_limit > 0 &&
3386 		    bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
3387 			BBR_STAT_INC(bbr_split_limited);
3388 			if (!bbr->alloc_limit_reported) {
3389 				bbr->alloc_limit_reported = 1;
3390 				BBR_STAT_INC(bbr_alloc_limited_conns);
3391 			}
3392 			return (NULL);
3393 		}
3394 	}
3395 
3396 	/* allocate and mark in the limit type, if set */
3397 	rsm = bbr_alloc(bbr);
3398 	if (rsm != NULL && limit_type) {
3399 		rsm->r_limit_type = limit_type;
3400 		bbr->r_ctl.rc_num_split_allocs++;
3401 	}
3402 	return (rsm);
3403 }
3404 
3405 static void
3406 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
3407 {
3408 	if (rsm->r_limit_type) {
3409 		/* currently there is only one limit type */
3410 		bbr->r_ctl.rc_num_split_allocs--;
3411 	}
3412 	if (rsm->r_is_smallmap)
3413 		bbr->r_ctl.rc_num_small_maps_alloced--;
3414 	if (bbr->r_ctl.rc_tlp_send == rsm)
3415 		bbr->r_ctl.rc_tlp_send = NULL;
3416 	if (bbr->r_ctl.rc_resend == rsm) {
3417 		bbr->r_ctl.rc_resend = NULL;
3418 	}
3419 	if (bbr->r_ctl.rc_next == rsm)
3420 		bbr->r_ctl.rc_next = NULL;
3421 	if (bbr->r_ctl.rc_sacklast == rsm)
3422 		bbr->r_ctl.rc_sacklast = NULL;
3423 	if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
3424 		memset(rsm, 0, sizeof(struct bbr_sendmap));
3425 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
3426 		rsm->r_limit_type = 0;
3427 		bbr->r_ctl.rc_free_cnt++;
3428 		return;
3429 	}
3430 	bbr->r_ctl.rc_num_maps_alloced--;
3431 	uma_zfree(bbr_zone, rsm);
3432 }
3433 
3434 /*
3435  * Returns the BDP.
3436  */
3437 static uint64_t
3438 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) {
3439 	/*
3440 	 * Calculate the bytes in flight needed given the bw (in bytes per
3441 	 * second) and the specifyed rtt in useconds. We need to put out the
3442 	 * returned value per RTT to match that rate. Gain will normaly
3443 	 * raise it up from there.
3444 	 *
3445 	 * This should not overflow as long as the bandwidth is below 1
3446 	 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller
3447 	 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30).
3448 	 */
3449 	uint64_t usec_per_sec;
3450 
3451 	usec_per_sec = USECS_IN_SECOND;
3452 	return ((rtt * bw) / usec_per_sec);
3453 }
3454 
3455 /*
3456  * Return the initial cwnd.
3457  */
3458 static uint32_t
3459 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp)
3460 {
3461 	uint32_t i_cwnd;
3462 
3463 	if (bbr->rc_init_win) {
3464 		i_cwnd = bbr->rc_init_win * tp->t_maxseg;
3465 	} else if (V_tcp_initcwnd_segments)
3466 		i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg),
3467 		    max(2 * tp->t_maxseg, 14600));
3468 	else if (V_tcp_do_rfc3390)
3469 		i_cwnd = min(4 * tp->t_maxseg,
3470 		    max(2 * tp->t_maxseg, 4380));
3471 	else {
3472 		/* Per RFC5681 Section 3.1 */
3473 		if (tp->t_maxseg > 2190)
3474 			i_cwnd = 2 * tp->t_maxseg;
3475 		else if (tp->t_maxseg > 1095)
3476 			i_cwnd = 3 * tp->t_maxseg;
3477 		else
3478 			i_cwnd = 4 * tp->t_maxseg;
3479 	}
3480 	return (i_cwnd);
3481 }
3482 
3483 /*
3484  * Given a specified gain, return the target
3485  * cwnd based on that gain.
3486  */
3487 static uint32_t
3488 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw)
3489 {
3490 	uint64_t bdp, rtt;
3491 	uint32_t cwnd;
3492 
3493 	if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) ||
3494 	    (bbr_get_full_bw(bbr) == 0)) {
3495 		/* No measurements yet */
3496 		return (bbr_initial_cwnd(bbr, bbr->rc_tp));
3497 	}
3498 	/*
3499 	 * Get bytes per RTT needed (rttProp is normally in
3500 	 * bbr_cwndtarget_rtt_touse)
3501 	 */
3502 	rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse);
3503 	/* Get the bdp from the two values */
3504 	bdp = bbr_get_bw_delay_prod(rtt, bw);
3505 	/* Now apply the gain */
3506 	cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT));
3507 
3508 	return (cwnd);
3509 }
3510 
3511 static uint32_t
3512 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain)
3513 {
3514 	uint32_t cwnd, mss;
3515 
3516 	mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
3517 	/* Get the base cwnd with gain rounded to a mss */
3518 	cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss);
3519 	/*
3520 	 * Add in N (2 default since we do not have a
3521 	 * fq layer to trap packets in) quanta's per the I-D
3522 	 * section 4.2.3.2 quanta adjust.
3523 	 */
3524 	cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs);
3525 	if (bbr->rc_use_google) {
3526 		if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3527 		   (bbr_state_val(bbr) == BBR_SUB_GAIN)) {
3528 			/*
3529 			 * The linux implementation adds
3530 			 * an extra 2 x mss in gain cycle which
3531 			 * is documented no-where except in the code.
3532 			 * so we add more for Neal undocumented feature
3533 			 */
3534 			cwnd += 2 * mss;
3535 		}
3536  		if ((cwnd / mss) & 0x1) {
3537 			/* Round up for odd num mss */
3538 			cwnd += mss;
3539 		}
3540 	}
3541 	/* Are we below the min cwnd? */
3542 	if (cwnd < get_min_cwnd(bbr))
3543 		return (get_min_cwnd(bbr));
3544 	return (cwnd);
3545 }
3546 
3547 static uint16_t
3548 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain)
3549 {
3550 	if (gain < 1)
3551 		gain = 1;
3552 	return (gain);
3553 }
3554 
3555 static uint32_t
3556 bbr_get_header_oh(struct tcp_bbr *bbr)
3557 {
3558 	int seg_oh;
3559 
3560 	seg_oh = 0;
3561 	if (bbr->r_ctl.rc_inc_tcp_oh) {
3562 		/* Do we include TCP overhead? */
3563 		seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr));
3564 	}
3565 	if (bbr->r_ctl.rc_inc_ip_oh) {
3566 		/* Do we include IP overhead? */
3567 #ifdef INET6
3568 		if (bbr->r_is_v6)
3569 			seg_oh += sizeof(struct ip6_hdr);
3570 		else
3571 #endif
3572 #ifdef INET
3573 			seg_oh += sizeof(struct ip);
3574 #endif
3575 	}
3576 	if (bbr->r_ctl.rc_inc_enet_oh) {
3577 		/* Do we include the ethernet overhead?  */
3578 		seg_oh += sizeof(struct ether_header);
3579 	}
3580 	return(seg_oh);
3581 }
3582 
3583 
3584 static uint32_t
3585 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw)
3586 {
3587 	uint64_t divor, res, tim;
3588 
3589 	if (useconds_time == 0)
3590 		return (0);
3591 	gain = bbr_gain_adjust(bbr, gain);
3592 	divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT;
3593 	tim = useconds_time;
3594 	res = (tim * bw * gain) / divor;
3595 	if (res == 0)
3596 		res = 1;
3597 	return ((uint32_t)res);
3598 }
3599 
3600 /*
3601  * Given a gain and a length return the delay in useconds that
3602  * should be used to evenly space out packets
3603  * on the connection (based on the gain factor).
3604  */
3605 static uint32_t
3606 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog)
3607 {
3608 	uint64_t bw, lentim, res;
3609 	uint32_t usecs, srtt, over = 0;
3610 	uint32_t seg_oh, num_segs, maxseg;
3611 
3612 	if (len == 0)
3613 		return (0);
3614 
3615 	maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
3616 	num_segs = (len + maxseg - 1) / maxseg;
3617 	if (bbr->rc_use_google == 0) {
3618 		seg_oh = bbr_get_header_oh(bbr);
3619 		len += (num_segs * seg_oh);
3620 	}
3621 	gain = bbr_gain_adjust(bbr, gain);
3622 	bw = bbr_get_bw(bbr);
3623 	if (bbr->rc_use_google) {
3624 		uint64_t cbw;
3625 
3626 		/*
3627 		 * Reduce the b/w by the google discount
3628 		 * factor 10 = 1%.
3629 		 */
3630 		cbw = bw *  (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount);
3631 		cbw /= (uint64_t)1000;
3632 		/* We don't apply a discount if it results in 0 */
3633 		if (cbw > 0)
3634 			bw = cbw;
3635 	}
3636 	lentim = ((uint64_t)len *
3637 		  (uint64_t)USECS_IN_SECOND *
3638 		  (uint64_t)BBR_UNIT);
3639 	res = lentim / ((uint64_t)gain * bw);
3640 	if (res == 0)
3641 		res = 1;
3642 	usecs = (uint32_t)res;
3643 	srtt = bbr_get_rtt(bbr, BBR_SRTT);
3644 	if (bbr_hptsi_max_mul && bbr_hptsi_max_div &&
3645 	    (bbr->rc_use_google == 0) &&
3646 	    (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) {
3647 		/*
3648 		 * We cannot let the delay be more than 1/2 the srtt time.
3649 		 * Otherwise we cannot pace out or send properly.
3650 		 */
3651 		over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div;
3652 		BBR_STAT_INC(bbr_hpts_min_time);
3653 	}
3654 	if (!nolog)
3655 		bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1);
3656 	return (usecs);
3657 }
3658 
3659 static void
3660 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack,
3661 		 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses)
3662 {
3663 	INP_WLOCK_ASSERT(tp->t_inpcb);
3664 	uint64_t bw;
3665 	uint32_t cwnd, target_cwnd, saved_bytes, maxseg;
3666 	int32_t meth;
3667 
3668 #ifdef STATS
3669 	if ((tp->t_flags & TF_GPUTINPROG) &&
3670 	    SEQ_GEQ(th->th_ack, tp->gput_ack)) {
3671 		/*
3672 		 * Strech acks and compressed acks will cause this to
3673 		 * oscillate but we are doing it the same way as the main
3674 		 * stack so it will be compariable (though possibly not
3675 		 * ideal).
3676 		 */
3677 		int32_t cgput;
3678 		int64_t gput, time_stamp;
3679 
3680 		gput = (int64_t) (th->th_ack - tp->gput_seq) * 8;
3681 		time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000));
3682 		cgput = gput / time_stamp;
3683 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3684 					 cgput);
3685 		if (tp->t_stats_gput_prev > 0)
3686 			stats_voi_update_abs_s32(tp->t_stats,
3687 						 VOI_TCP_GPUT_ND,
3688 						 ((gput - tp->t_stats_gput_prev) * 100) /
3689 						 tp->t_stats_gput_prev);
3690 		tp->t_flags &= ~TF_GPUTINPROG;
3691 		tp->t_stats_gput_prev = cgput;
3692 	}
3693 #endif
3694 	if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3695 	    ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
3696 		/* We don't change anything in probe-rtt */
3697 		return;
3698 	}
3699 	maxseg = tp->t_maxseg - bbr->rc_last_options;
3700 	saved_bytes = bytes_this_ack;
3701 	bytes_this_ack += sack_changed;
3702 	if (bytes_this_ack > prev_acked) {
3703 		bytes_this_ack -= prev_acked;
3704 		/*
3705 		 * A byte ack'd gives us a full mss
3706 		 * to be like linux i.e. they count packets.
3707 		 */
3708 		if ((bytes_this_ack < maxseg) && bbr->rc_use_google)
3709 			bytes_this_ack = maxseg;
3710 	} else {
3711 		/* Unlikely */
3712 		bytes_this_ack = 0;
3713 	}
3714 	cwnd = tp->snd_cwnd;
3715 	bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3716 	if (bw)
3717 		target_cwnd = bbr_get_target_cwnd(bbr,
3718 						  bw,
3719 						  (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain);
3720 	else
3721 		target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp);
3722 	if (IN_RECOVERY(tp->t_flags) &&
3723 	    (bbr->bbr_prev_in_rec == 0)) {
3724 		/*
3725 		 * We are entering recovery and
3726 		 * thus packet conservation.
3727 		 */
3728 		bbr->pkt_conservation = 1;
3729 		bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime;
3730 		cwnd = ctf_flight_size(tp,
3731 				       (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3732 			bytes_this_ack;
3733 	}
3734 	if (IN_RECOVERY(tp->t_flags)) {
3735 		uint32_t flight;
3736 
3737 		bbr->bbr_prev_in_rec = 1;
3738 		if (cwnd > losses) {
3739 			cwnd -= losses;
3740 			if (cwnd < maxseg)
3741 				cwnd = maxseg;
3742 		} else
3743 			cwnd = maxseg;
3744 		flight = ctf_flight_size(tp,
3745 					 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3746 		bbr_log_type_cwndupd(bbr, flight, 0,
3747 				     losses, 10, 0, 0, line);
3748 		if (bbr->pkt_conservation) {
3749 			uint32_t time_in;
3750 
3751 			if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start))
3752 				time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start;
3753 			else
3754 				time_in = 0;
3755 
3756 			if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
3757 				/* Clear packet conservation after an rttProp */
3758 				bbr->pkt_conservation = 0;
3759 			} else {
3760 				if ((flight + bytes_this_ack) > cwnd)
3761 					cwnd = flight + bytes_this_ack;
3762 				if (cwnd < get_min_cwnd(bbr))
3763 					cwnd = get_min_cwnd(bbr);
3764 				tp->snd_cwnd = cwnd;
3765 				bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed,
3766 						     prev_acked, 1, target_cwnd, th->th_ack, line);
3767 				return;
3768 			}
3769 		}
3770 	} else
3771 		bbr->bbr_prev_in_rec = 0;
3772 	if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) {
3773 		bbr->r_ctl.restrict_growth--;
3774 		if (bytes_this_ack > maxseg)
3775 			bytes_this_ack = maxseg;
3776 	}
3777 	if (bbr->rc_filled_pipe) {
3778 		/*
3779 		 * Here we have exited startup and filled the pipe. We will
3780 		 * thus allow the cwnd to shrink to the target. We hit here
3781 		 * mostly.
3782 		 */
3783 		uint32_t s_cwnd;
3784 
3785 		meth = 2;
3786 		s_cwnd = min((cwnd + bytes_this_ack), target_cwnd);
3787 		if (s_cwnd > cwnd)
3788 			cwnd = s_cwnd;
3789 		else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing)
3790 			cwnd = s_cwnd;
3791 	} else {
3792 		/*
3793 		 * Here we are still in startup, we increase cwnd by what
3794 		 * has been acked.
3795 		 */
3796 		if ((cwnd < target_cwnd) ||
3797 		    (bbr->rc_past_init_win == 0)) {
3798 			meth = 3;
3799 			cwnd += bytes_this_ack;
3800 		} else {
3801 			/*
3802 			 * Method 4 means we are at target so no gain in
3803 			 * startup and past the initial window.
3804 			 */
3805 			meth = 4;
3806 		}
3807 	}
3808 	tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr));
3809 	bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line);
3810 }
3811 
3812 static void
3813 tcp_bbr_partialack(struct tcpcb *tp)
3814 {
3815 	struct tcp_bbr *bbr;
3816 
3817 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3818 	INP_WLOCK_ASSERT(tp->t_inpcb);
3819 	if (ctf_flight_size(tp,
3820 		(bbr->r_ctl.rc_sacked  + bbr->r_ctl.rc_lost_bytes)) <=
3821 	    tp->snd_cwnd) {
3822 		bbr->r_wanted_output = 1;
3823 	}
3824 }
3825 
3826 static void
3827 bbr_post_recovery(struct tcpcb *tp)
3828 {
3829 	struct tcp_bbr *bbr;
3830 	uint32_t  flight;
3831 
3832 	INP_WLOCK_ASSERT(tp->t_inpcb);
3833 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3834 	/*
3835 	 * Here we just exit recovery.
3836 	 */
3837 	EXIT_RECOVERY(tp->t_flags);
3838 	/* Lock in our b/w reduction for the specified number of pkt-epochs */
3839 	bbr->r_recovery_bw = 0;
3840 	tp->snd_recover = tp->snd_una;
3841 	tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3842 	bbr->pkt_conservation = 0;
3843 	if (bbr->rc_use_google == 0) {
3844 		/*
3845 		 * For non-google mode lets
3846 		 * go ahead and make sure we clear
3847 		 * the recovery state so if we
3848 		 * bounce back in to recovery we
3849 		 * will do PC.
3850 		 */
3851 		bbr->bbr_prev_in_rec = 0;
3852 	}
3853 	bbr_log_type_exit_rec(bbr);
3854 	if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3855 		tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
3856 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__);
3857 	} else {
3858 		/* For probe-rtt case lets fix up its saved_cwnd */
3859 		if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) {
3860 			bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent;
3861 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__);
3862 		}
3863 	}
3864 	flight = ctf_flight_size(tp,
3865 		     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3866 	if ((bbr->rc_use_google == 0) &&
3867 	    bbr_do_red) {
3868 		uint64_t val, lr2use;
3869 		uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd;
3870 		uint32_t *cwnd_p;
3871 
3872 		if (bbr_get_rtt(bbr, BBR_SRTT)) {
3873 			val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000);
3874 			val /= bbr_get_rtt(bbr, BBR_SRTT);
3875 			ratio = (uint32_t)val;
3876 		} else
3877 			ratio = 1000;
3878 
3879 		bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div,
3880 				     bbr->r_ctl.recovery_lr, 21,
3881 				     ratio,
3882 				     bbr->r_ctl.rc_red_cwnd_pe,
3883 				     __LINE__);
3884 		if ((ratio < bbr_do_red) || (bbr_do_red == 0))
3885 			goto done;
3886 		if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3887 		     bbr_prtt_slam_cwnd) ||
3888 		    (bbr_sub_drain_slam_cwnd &&
3889 		     (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3890 		     bbr->rc_hit_state_1 &&
3891 		     (bbr_state_val(bbr) == BBR_SUB_DRAIN)) ||
3892 		    ((bbr->rc_bbr_state == BBR_STATE_DRAIN) &&
3893 		     bbr_slam_cwnd_in_main_drain)) {
3894 			/*
3895 			 * Here we must poke at the saved cwnd
3896 			 * as well as the cwnd.
3897 			 */
3898 			cwnd = bbr->r_ctl.rc_saved_cwnd;
3899 			cwnd_p = &bbr->r_ctl.rc_saved_cwnd;
3900 		} else {
3901  			cwnd = tp->snd_cwnd;
3902 			cwnd_p = &tp->snd_cwnd;
3903 		}
3904 		maxseg = tp->t_maxseg - bbr->rc_last_options;
3905 		/* Add the overall lr with the recovery lr */
3906 		if (bbr->r_ctl.rc_lost == 0)
3907 			lr2use = 0;
3908 		else if (bbr->r_ctl.rc_delivered == 0)
3909 			lr2use = 1000;
3910 		else {
3911 			lr2use = bbr->r_ctl.rc_lost * 1000;
3912 			lr2use /= bbr->r_ctl.rc_delivered;
3913 		}
3914 		lr2use += bbr->r_ctl.recovery_lr;
3915 		acks_inflight = (flight / (maxseg * 2));
3916 		if (bbr_red_scale) {
3917 			lr2use *= bbr_get_rtt(bbr, BBR_SRTT);
3918 			lr2use /= bbr_red_scale;
3919 			if ((bbr_red_growth_restrict) &&
3920 			    ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1))
3921 			    bbr->r_ctl.restrict_growth += acks_inflight;
3922 		}
3923 		if (lr2use) {
3924 			val = (uint64_t)cwnd * lr2use;
3925 			val /= 1000;
3926 			if (cwnd > val)
3927 				newcwnd = roundup((cwnd - val), maxseg);
3928 			else
3929 				newcwnd = maxseg;
3930 		} else {
3931 			val = (uint64_t)cwnd * (uint64_t)bbr_red_mul;
3932 			val /= (uint64_t)bbr_red_div;
3933 			newcwnd = roundup((uint32_t)val, maxseg);
3934 		}
3935 		/* with standard delayed acks how many acks can I expect? */
3936 		if (bbr_drop_limit == 0) {
3937 			/*
3938 			 * Anticpate how much we will
3939 			 * raise the cwnd based on the acks.
3940 			 */
3941 			if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) {
3942 				/* We do enforce the min (with the acks) */
3943 				newcwnd = (get_min_cwnd(bbr) - acks_inflight);
3944 			}
3945 		} else {
3946 			/*
3947 			 * A strict drop limit of N is is inplace
3948 			 */
3949 			if (newcwnd < (bbr_drop_limit * maxseg)) {
3950 				newcwnd = bbr_drop_limit * maxseg;
3951 			}
3952 		}
3953 		/* For the next N acks do we restrict the growth */
3954 		*cwnd_p = newcwnd;
3955 		if (tp->snd_cwnd > newcwnd)
3956 			tp->snd_cwnd = newcwnd;
3957 		bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22,
3958 				     (uint32_t)lr2use,
3959 				     bbr_get_rtt(bbr, BBR_SRTT), __LINE__);
3960 		bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch;
3961 	}
3962 done:
3963 	bbr->r_ctl.recovery_lr = 0;
3964 	if (flight <= tp->snd_cwnd) {
3965 		bbr->r_wanted_output = 1;
3966 	}
3967 	tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3968 }
3969 
3970 static void
3971 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts)
3972 {
3973 	bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3974 	/* Limit the drop in b/w to 1/2 our current filter. */
3975 	if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate)
3976 		bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate;
3977 	if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2))
3978 		bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2;
3979 	tcp_bbr_tso_size_check(bbr, cts);
3980 }
3981 
3982 static void
3983 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm)
3984 {
3985 	struct tcp_bbr *bbr;
3986 
3987 	INP_WLOCK_ASSERT(tp->t_inpcb);
3988 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3989 	switch (type) {
3990 	case CC_NDUPACK:
3991 		if (!IN_RECOVERY(tp->t_flags)) {
3992 			tp->snd_recover = tp->snd_max;
3993 			/* Start a new epoch */
3994 			bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
3995 			if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) {
3996 				/*
3997 				 * Move forward the lt epoch
3998 				 * so it won't count the truncated
3999 				 * epoch.
4000 				 */
4001 				bbr->r_ctl.rc_lt_epoch++;
4002 			}
4003 			if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
4004 				/*
4005 				 * Just like the policer detection code
4006 				 * if we are in startup we must push
4007 				 * forward the last startup epoch
4008 				 * to hide the truncated PE.
4009 				 */
4010 				bbr->r_ctl.rc_bbr_last_startup_epoch++;
4011 			}
4012 			bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd;
4013 			ENTER_RECOVERY(tp->t_flags);
4014 			bbr->rc_tlp_rtx_out = 0;
4015 			bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate;
4016 			tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
4017 			if (bbr->rc_inp->inp_in_hpts &&
4018 			    ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) {
4019 				/*
4020 				 * When we enter recovery, we need to restart
4021 				 * any timers. This may mean we gain an agg
4022 				 * early, which will be made up for at the last
4023 				 * rxt out.
4024 				 */
4025 				bbr->rc_timer_first = 1;
4026 				bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
4027 			}
4028 			/*
4029 			 * Calculate a new cwnd based on to the current
4030 			 * delivery rate with no gain. We get the bdp
4031 			 * without gaining it up like we normally would and
4032 			 * we use the last cur_del_rate.
4033 			 */
4034 			if ((bbr->rc_use_google == 0) &&
4035 			    (bbr->r_ctl.bbr_rttprobe_gain_val ||
4036 			     (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) {
4037 				tp->snd_cwnd = ctf_flight_size(tp,
4038 					           (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
4039 					(tp->t_maxseg - bbr->rc_last_options);
4040 				if (tp->snd_cwnd < get_min_cwnd(bbr)) {
4041 					/* We always gate to min cwnd */
4042 					tp->snd_cwnd = get_min_cwnd(bbr);
4043 				}
4044 				bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__);
4045 			}
4046 			bbr_log_type_enter_rec(bbr, rsm->r_start);
4047 		}
4048 		break;
4049 	case CC_RTO_ERR:
4050 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4051 		/* RTO was unnecessary, so reset everything. */
4052 		bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime);
4053 		if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
4054 			tp->snd_cwnd = tp->snd_cwnd_prev;
4055 			tp->snd_ssthresh = tp->snd_ssthresh_prev;
4056 			tp->snd_recover = tp->snd_recover_prev;
4057 			tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
4058 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__);
4059 		}
4060 		tp->t_badrxtwin = 0;
4061 		break;
4062 	}
4063 }
4064 
4065 /*
4066  * Indicate whether this ack should be delayed.  We can delay the ack if
4067  * following conditions are met:
4068  *	- There is no delayed ack timer in progress.
4069  *	- Our last ack wasn't a 0-sized window. We never want to delay
4070  *	  the ack that opens up a 0-sized window.
4071  *	- LRO wasn't used for this segment. We make sure by checking that the
4072  *	  segment size is not larger than the MSS.
4073  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
4074  *	  connection.
4075  *	- The data being acked is less than a full segment (a stretch ack
4076  *        of more than a segment we should ack.
4077  *      - nsegs is 1 (if its more than that we received more than 1 ack).
4078  */
4079 #define DELAY_ACK(tp, bbr, nsegs)				\
4080 	(((tp->t_flags & TF_RXWIN0SENT) == 0) &&		\
4081 	 ((tp->t_flags & TF_DELACK) == 0) && 		 	\
4082 	 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) &&	\
4083 	 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4084 
4085 /*
4086  * Return the lowest RSM in the map of
4087  * packets still in flight that is not acked.
4088  * This should normally find on the first one
4089  * since we remove packets from the send
4090  * map after they are marked ACKED.
4091  */
4092 static struct bbr_sendmap *
4093 bbr_find_lowest_rsm(struct tcp_bbr *bbr)
4094 {
4095 	struct bbr_sendmap *rsm;
4096 
4097 	/*
4098 	 * Walk the time-order transmitted list looking for an rsm that is
4099 	 * not acked. This will be the one that was sent the longest time
4100 	 * ago that is still outstanding.
4101 	 */
4102 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) {
4103 		if (rsm->r_flags & BBR_ACKED) {
4104 			continue;
4105 		}
4106 		goto finish;
4107 	}
4108 finish:
4109 	return (rsm);
4110 }
4111 
4112 static struct bbr_sendmap *
4113 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
4114 {
4115 	struct bbr_sendmap *prsm;
4116 
4117 	/*
4118 	 * Walk the sequence order list backward until we hit and arrive at
4119 	 * the highest seq not acked. In theory when this is called it
4120 	 * should be the last segment (which it was not).
4121 	 */
4122 	prsm = rsm;
4123 	TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4124 		if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) {
4125 			continue;
4126 		}
4127 		return (prsm);
4128 	}
4129 	return (NULL);
4130 }
4131 
4132 /*
4133  * Returns to the caller the number of microseconds that
4134  * the packet can be outstanding before we think we
4135  * should have had an ack returned.
4136  */
4137 static uint32_t
4138 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm)
4139 {
4140 	/*
4141 	 * lro is the flag we use to determine if we have seen reordering.
4142 	 * If it gets set we have seen reordering. The reorder logic either
4143 	 * works in one of two ways:
4144 	 *
4145 	 * If reorder-fade is configured, then we track the last time we saw
4146 	 * re-ordering occur. If we reach the point where enough time as
4147 	 * passed we no longer consider reordering has occuring.
4148 	 *
4149 	 * Or if reorder-face is 0, then once we see reordering we consider
4150 	 * the connection to alway be subject to reordering and just set lro
4151 	 * to 1.
4152 	 *
4153 	 * In the end if lro is non-zero we add the extra time for
4154 	 * reordering in.
4155 	 */
4156 	int32_t lro;
4157 	uint32_t thresh, t_rxtcur;
4158 
4159 	if (srtt == 0)
4160 		srtt = 1;
4161 	if (bbr->r_ctl.rc_reorder_ts) {
4162 		if (bbr->r_ctl.rc_reorder_fade) {
4163 			if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) {
4164 				lro = cts - bbr->r_ctl.rc_reorder_ts;
4165 				if (lro == 0) {
4166 					/*
4167 					 * No time as passed since the last
4168 					 * reorder, mark it as reordering.
4169 					 */
4170 					lro = 1;
4171 				}
4172 			} else {
4173 				/* Negative time? */
4174 				lro = 0;
4175 			}
4176 			if (lro > bbr->r_ctl.rc_reorder_fade) {
4177 				/* Turn off reordering seen too */
4178 				bbr->r_ctl.rc_reorder_ts = 0;
4179 				lro = 0;
4180 			}
4181 		} else {
4182 			/* Reodering does not fade */
4183 			lro = 1;
4184 		}
4185 	} else {
4186 		lro = 0;
4187 	}
4188 	thresh = srtt + bbr->r_ctl.rc_pkt_delay;
4189 	if (lro) {
4190 		/* It must be set, if not you get 1/4 rtt */
4191 		if (bbr->r_ctl.rc_reorder_shift)
4192 			thresh += (srtt >> bbr->r_ctl.rc_reorder_shift);
4193 		else
4194 			thresh += (srtt >> 2);
4195 	} else {
4196 		thresh += 1000;
4197 	}
4198 	/* We don't let the rack timeout be above a RTO */
4199 	if ((bbr->rc_tp)->t_srtt == 0)
4200 		t_rxtcur = BBR_INITIAL_RTO;
4201 	else
4202 		t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
4203 	if (thresh > t_rxtcur) {
4204 		thresh = t_rxtcur;
4205 	}
4206 	/* And we don't want it above the RTO max either */
4207 	if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4208 		thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4209 	}
4210 	bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK);
4211 	return (thresh);
4212 }
4213 
4214 /*
4215  * Return to the caller the amount of time in mico-seconds
4216  * that should be used for the TLP timer from the last
4217  * send time of this packet.
4218  */
4219 static uint32_t
4220 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
4221     struct bbr_sendmap *rsm, uint32_t srtt,
4222     uint32_t cts)
4223 {
4224 	uint32_t thresh, len, maxseg, t_rxtcur;
4225 	struct bbr_sendmap *prsm;
4226 
4227 	if (srtt == 0)
4228 		srtt = 1;
4229 	if (bbr->rc_tlp_threshold)
4230 		thresh = srtt + (srtt / bbr->rc_tlp_threshold);
4231 	else
4232 		thresh = (srtt * 2);
4233 	maxseg = tp->t_maxseg - bbr->rc_last_options;
4234 	/* Get the previous sent packet, if any  */
4235 	len = rsm->r_end - rsm->r_start;
4236 
4237 	/* 2.1 behavior */
4238 	prsm = TAILQ_PREV(rsm, bbr_head, r_tnext);
4239 	if (prsm && (len <= maxseg)) {
4240 		/*
4241 		 * Two packets outstanding, thresh should be (2*srtt) +
4242 		 * possible inter-packet delay (if any).
4243 		 */
4244 		uint32_t inter_gap = 0;
4245 		int idx, nidx;
4246 
4247 		idx = rsm->r_rtr_cnt - 1;
4248 		nidx = prsm->r_rtr_cnt - 1;
4249 		if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4250 			/* Yes it was sent later (or at the same time) */
4251 			inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4252 		}
4253 		thresh += inter_gap;
4254 	} else if (len <= maxseg) {
4255 		/*
4256 		 * Possibly compensate for delayed-ack.
4257 		 */
4258 		uint32_t alt_thresh;
4259 
4260 		alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time;
4261 		if (alt_thresh > thresh)
4262 			thresh = alt_thresh;
4263 	}
4264 	/* Not above the current  RTO */
4265 	if (tp->t_srtt == 0)
4266 		t_rxtcur = BBR_INITIAL_RTO;
4267 	else
4268 		t_rxtcur = TICKS_2_USEC(tp->t_rxtcur);
4269 
4270 	bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP);
4271 	/* Not above an RTO */
4272 	if (thresh > t_rxtcur) {
4273 		thresh = t_rxtcur;
4274 	}
4275 	/* Not above a RTO max */
4276 	if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4277 		thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4278 	}
4279 	/* And now apply the user TLP min */
4280 	if (thresh < bbr_tlp_min) {
4281 		thresh = bbr_tlp_min;
4282 	}
4283 	return (thresh);
4284 }
4285 
4286 /*
4287  * Return one of three RTTs to use (in microseconds).
4288  */
4289 static __inline uint32_t
4290 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type)
4291 {
4292 	uint32_t f_rtt;
4293 	uint32_t srtt;
4294 
4295 	f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
4296 	if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) {
4297 		/* We have no rtt at all */
4298 		if (bbr->rc_tp->t_srtt == 0)
4299 			f_rtt = BBR_INITIAL_RTO;
4300 		else
4301 			f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4302 		/*
4303 		 * Since we don't know how good the rtt is apply a
4304 		 * delayed-ack min
4305 		 */
4306 		if (f_rtt < bbr_delayed_ack_time) {
4307 			f_rtt = bbr_delayed_ack_time;
4308 		}
4309 	}
4310 	/* Take the filter version or last measured pkt-rtt */
4311 	if (rtt_type == BBR_RTT_PROP) {
4312 		srtt = f_rtt;
4313 	} else if (rtt_type == BBR_RTT_PKTRTT) {
4314 		if (bbr->r_ctl.rc_pkt_epoch_rtt) {
4315 			srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
4316 		} else {
4317 			/* No pkt rtt yet */
4318 			srtt = f_rtt;
4319 		}
4320 	} else if (rtt_type == BBR_RTT_RACK) {
4321 		srtt = bbr->r_ctl.rc_last_rtt;
4322 		/* We need to add in any internal delay for our timer */
4323 		if (bbr->rc_ack_was_delayed)
4324 			srtt += bbr->r_ctl.rc_ack_hdwr_delay;
4325 	} else if (rtt_type == BBR_SRTT) {
4326 		srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4327 	} else {
4328 		/* TSNH */
4329 		srtt = f_rtt;
4330 #ifdef BBR_INVARIANTS
4331 		panic("Unknown rtt request type %d", rtt_type);
4332 #endif
4333 	}
4334 	return (srtt);
4335 }
4336 
4337 static int
4338 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts)
4339 {
4340 	uint32_t thresh;
4341 
4342 
4343 	thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK),
4344 				      cts, rsm);
4345 	if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) {
4346 		/* It is lost (past time) */
4347 		return (1);
4348 	}
4349 	return (0);
4350 }
4351 
4352 /*
4353  * Return a sendmap if we need to retransmit something.
4354  */
4355 static struct bbr_sendmap *
4356 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4357 {
4358 	/*
4359 	 * Check to see that we don't need to fall into recovery. We will
4360 	 * need to do so if our oldest transmit is past the time we should
4361 	 * have had an ack.
4362 	 */
4363 
4364 	struct bbr_sendmap *rsm;
4365 	int32_t idx;
4366 
4367 	if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) {
4368 		/* Nothing outstanding that we know of */
4369 		return (NULL);
4370 	}
4371 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
4372 	if (rsm == NULL) {
4373 		/* Nothing in the transmit map */
4374 		return (NULL);
4375 	}
4376 	if (tp->t_flags & TF_SENTFIN) {
4377 		/* Fin restricted, don't find anything once a fin is sent */
4378 		return (NULL);
4379 	}
4380 	if (rsm->r_flags & BBR_ACKED) {
4381 		/*
4382 		 * Ok the first one is acked (this really should not happen
4383 		 * since we remove the from the tmap once they are acked)
4384 		 */
4385 		rsm = bbr_find_lowest_rsm(bbr);
4386 		if (rsm == NULL)
4387 			return (NULL);
4388 	}
4389 	idx = rsm->r_rtr_cnt - 1;
4390 	if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) {
4391 		/* Send timestamp is the same or less? can't be ready */
4392 		return (NULL);
4393 	}
4394 	/* Get our RTT time */
4395 	if (bbr_is_lost(bbr, rsm, cts) &&
4396 	    ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
4397 	     (rsm->r_flags & BBR_SACK_PASSED))) {
4398 		if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4399 			rsm->r_flags |= BBR_MARKED_LOST;
4400 			bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4401 			bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4402 		}
4403 		bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm);
4404 #ifdef BBR_INVARIANTS
4405 		if ((rsm->r_end - rsm->r_start) == 0)
4406 			panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm);
4407 #endif
4408 		return (rsm);
4409 	}
4410 	return (NULL);
4411 }
4412 
4413 /*
4414  * RACK Timer, here we simply do logging and house keeping.
4415  * the normal bbr_output_wtime() function will call the
4416  * appropriate thing to check if we need to do a RACK retransmit.
4417  * We return 1, saying don't proceed with bbr_output_wtime only
4418  * when all timers have been stopped (destroyed PCB?).
4419  */
4420 static int
4421 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4422 {
4423 	/*
4424 	 * This timer simply provides an internal trigger to send out data.
4425 	 * The check_recovery_mode call will see if there are needed
4426 	 * retransmissions, if so we will enter fast-recovery. The output
4427 	 * call may or may not do the same thing depending on sysctl
4428 	 * settings.
4429 	 */
4430 	uint32_t lost;
4431 
4432 	if (bbr->rc_all_timers_stopped) {
4433 		return (1);
4434 	}
4435 	if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4436 		/* Its not time yet */
4437 		return (0);
4438 	}
4439 	BBR_STAT_INC(bbr_to_tot);
4440 	lost = bbr->r_ctl.rc_lost;
4441 	if (bbr->r_state && (bbr->r_state != tp->t_state))
4442 		bbr_set_state(tp, bbr, 0);
4443 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK);
4444 	if (bbr->r_ctl.rc_resend == NULL) {
4445 		/* Lets do the check here */
4446 		bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
4447 	}
4448 	if (bbr_policer_call_from_rack_to)
4449 		bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4450 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
4451 	return (0);
4452 }
4453 
4454 static __inline void
4455 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start)
4456 {
4457 	int idx;
4458 
4459 	nrsm->r_start = start;
4460 	nrsm->r_end = rsm->r_end;
4461 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
4462 	nrsm->r_flags = rsm->r_flags;
4463 	/* We don't transfer forward the SYN flag */
4464 	nrsm->r_flags &= ~BBR_HAS_SYN;
4465 	/* We move forward the FIN flag, not that this should happen */
4466 	rsm->r_flags &= ~BBR_HAS_FIN;
4467 	nrsm->r_dupack = rsm->r_dupack;
4468 	nrsm->r_rtr_bytes = 0;
4469 	nrsm->r_is_gain = rsm->r_is_gain;
4470 	nrsm->r_is_drain = rsm->r_is_drain;
4471 	nrsm->r_delivered = rsm->r_delivered;
4472 	nrsm->r_ts_valid = rsm->r_ts_valid;
4473 	nrsm->r_del_ack_ts = rsm->r_del_ack_ts;
4474 	nrsm->r_del_time = rsm->r_del_time;
4475 	nrsm->r_app_limited = rsm->r_app_limited;
4476 	nrsm->r_first_sent_time = rsm->r_first_sent_time;
4477 	nrsm->r_flight_at_send = rsm->r_flight_at_send;
4478 	/* We split a piece the lower section looses any just_ret flag. */
4479 	nrsm->r_bbr_state = rsm->r_bbr_state;
4480 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
4481 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
4482 	}
4483 	rsm->r_end = nrsm->r_start;
4484 	idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
4485 	idx /= 8;
4486 	/* Check if we got too small */
4487 	if ((rsm->r_is_smallmap == 0) &&
4488 	    ((rsm->r_end - rsm->r_start) <= idx)) {
4489 		bbr->r_ctl.rc_num_small_maps_alloced++;
4490 		rsm->r_is_smallmap = 1;
4491 	}
4492 	/* Check the new one as well */
4493 	if ((nrsm->r_end - nrsm->r_start) <= idx) {
4494 		bbr->r_ctl.rc_num_small_maps_alloced++;
4495 		nrsm->r_is_smallmap = 1;
4496 	}
4497 }
4498 
4499 static int
4500 bbr_sack_mergable(struct bbr_sendmap *at,
4501 		  uint32_t start, uint32_t end)
4502 {
4503 	/*
4504 	 * Given a sack block defined by
4505 	 * start and end, and a current postion
4506 	 * at. Return 1 if either side of at
4507 	 * would show that the block is mergable
4508 	 * to that side. A block to be mergable
4509 	 * must have overlap with the start/end
4510 	 * and be in the SACK'd state.
4511 	 */
4512 	struct bbr_sendmap *l_rsm;
4513 	struct bbr_sendmap *r_rsm;
4514 
4515 	/* first get the either side blocks */
4516 	l_rsm = TAILQ_PREV(at, bbr_head, r_next);
4517 	r_rsm = TAILQ_NEXT(at, r_next);
4518 	if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) {
4519 		/* Potentially mergeable */
4520 		if ((l_rsm->r_end == start) ||
4521 		    (SEQ_LT(start, l_rsm->r_end) &&
4522 		     SEQ_GT(end, l_rsm->r_end))) {
4523 			    /*
4524 			     * map blk   |------|
4525 			     * sack blk         |------|
4526 			     * <or>
4527 			     * map blk   |------|
4528 			     * sack blk      |------|
4529 			     */
4530 			    return (1);
4531 		    }
4532 	}
4533 	if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) {
4534 		/* Potentially mergeable */
4535 		if ((r_rsm->r_start == end) ||
4536 		    (SEQ_LT(start, r_rsm->r_start) &&
4537 		     SEQ_GT(end, r_rsm->r_start))) {
4538 			/*
4539 			 * map blk          |---------|
4540 			 * sack blk    |----|
4541 			 * <or>
4542 			 * map blk          |---------|
4543 			 * sack blk    |-------|
4544 			 */
4545 			return (1);
4546 		}
4547 	}
4548 	return (0);
4549 }
4550 
4551 static struct bbr_sendmap *
4552 bbr_merge_rsm(struct tcp_bbr *bbr,
4553 	      struct bbr_sendmap *l_rsm,
4554 	      struct bbr_sendmap *r_rsm)
4555 {
4556 	/*
4557 	 * We are merging two ack'd RSM's,
4558 	 * the l_rsm is on the left (lower seq
4559 	 * values) and the r_rsm is on the right
4560 	 * (higher seq value). The simplest way
4561 	 * to merge these is to move the right
4562 	 * one into the left. I don't think there
4563 	 * is any reason we need to try to find
4564 	 * the oldest (or last oldest retransmitted).
4565 	 */
4566 	l_rsm->r_end = r_rsm->r_end;
4567 	if (l_rsm->r_dupack < r_rsm->r_dupack)
4568 		l_rsm->r_dupack = r_rsm->r_dupack;
4569 	if (r_rsm->r_rtr_bytes)
4570 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
4571 	if (r_rsm->r_in_tmap) {
4572 		/* This really should not happen */
4573 		TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext);
4574 	}
4575 	if (r_rsm->r_app_limited)
4576 		l_rsm->r_app_limited = r_rsm->r_app_limited;
4577 	/* Now the flags */
4578 	if (r_rsm->r_flags & BBR_HAS_FIN)
4579 		l_rsm->r_flags |= BBR_HAS_FIN;
4580 	if (r_rsm->r_flags & BBR_TLP)
4581 		l_rsm->r_flags |= BBR_TLP;
4582 	if (r_rsm->r_flags & BBR_RWND_COLLAPSED)
4583 		l_rsm->r_flags |= BBR_RWND_COLLAPSED;
4584 	if (r_rsm->r_flags & BBR_MARKED_LOST) {
4585 		/* This really should not happen */
4586 		bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start;
4587 	}
4588 	TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next);
4589 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
4590 		/* Transfer the split limit to the map we free */
4591 		r_rsm->r_limit_type = l_rsm->r_limit_type;
4592 		l_rsm->r_limit_type = 0;
4593 	}
4594 	bbr_free(bbr, r_rsm);
4595 	return(l_rsm);
4596 }
4597 
4598 /*
4599  * TLP Timer, here we simply setup what segment we want to
4600  * have the TLP expire on, the normal bbr_output_wtime() will then
4601  * send it out.
4602  *
4603  * We return 1, saying don't proceed with bbr_output_wtime only
4604  * when all timers have been stopped (destroyed PCB?).
4605  */
4606 static int
4607 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4608 {
4609 	/*
4610 	 * Tail Loss Probe.
4611 	 */
4612 	struct bbr_sendmap *rsm = NULL;
4613 	struct socket *so;
4614 	uint32_t amm;
4615 	uint32_t out, avail;
4616 	uint32_t maxseg;
4617 	int collapsed_win = 0;
4618 
4619 	if (bbr->rc_all_timers_stopped) {
4620 		return (1);
4621 	}
4622 	if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4623 		/* Its not time yet */
4624 		return (0);
4625 	}
4626 	if (ctf_progress_timeout_check(tp, true)) {
4627 		bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4628 		tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4629 		return (1);
4630 	}
4631 	/* Did we somehow get into persists? */
4632 	if (bbr->rc_in_persist) {
4633 		return (0);
4634 	}
4635 	if (bbr->r_state && (bbr->r_state != tp->t_state))
4636 		bbr_set_state(tp, bbr, 0);
4637 	BBR_STAT_INC(bbr_tlp_tot);
4638 	maxseg = tp->t_maxseg - bbr->rc_last_options;
4639 #ifdef KERN_TLS
4640 	if (bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
4641 		/*
4642 		 * For hardware TLS we do *not* want to send
4643 		 * new data.
4644 		 */
4645 		goto need_retran;
4646 	}
4647 #endif
4648 	/*
4649 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
4650 	 * need to figure out how to force a full MSS segment out.
4651 	 */
4652 	so = tp->t_inpcb->inp_socket;
4653 	avail = sbavail(&so->so_snd);
4654 	out = ctf_outstanding(tp);
4655 	if (out > tp->snd_wnd) {
4656 		/* special case, we need a retransmission */
4657 		collapsed_win = 1;
4658 		goto need_retran;
4659 	}
4660 	if (avail > out) {
4661 		/* New data is available */
4662 		amm = avail - out;
4663 		if (amm > maxseg) {
4664 			amm = maxseg;
4665 		} else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) {
4666 			/* not enough to fill a MTU and no-delay is off */
4667 			goto need_retran;
4668 		}
4669 		/* Set the send-new override */
4670 		if ((out + amm) <= tp->snd_wnd) {
4671 			bbr->rc_tlp_new_data = 1;
4672 		} else {
4673 			goto need_retran;
4674 		}
4675 		bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4676 		bbr->r_ctl.rc_last_tlp_seq = tp->snd_max;
4677 		bbr->r_ctl.rc_tlp_send = NULL;
4678 		/* cap any slots */
4679 		BBR_STAT_INC(bbr_tlp_newdata);
4680 		goto send;
4681 	}
4682 need_retran:
4683 	/*
4684 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
4685 	 * optionally the first un-acked segment.
4686 	 */
4687 	if (collapsed_win == 0) {
4688 		rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
4689 		if (rsm && (BBR_ACKED | BBR_HAS_FIN)) {
4690 			rsm = bbr_find_high_nonack(bbr, rsm);
4691 		}
4692 		if (rsm == NULL) {
4693 			goto restore;
4694 		}
4695 	} else {
4696 		/*
4697 		 * We must find the last segment
4698 		 * that was acceptable by the client.
4699 		 */
4700 		TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4701 			if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) {
4702 				/* Found one */
4703 				break;
4704 			}
4705 		}
4706 		if (rsm == NULL) {
4707 			/* None? if so send the first */
4708 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4709 			if (rsm == NULL)
4710 				goto restore;
4711 		}
4712 	}
4713 	if ((rsm->r_end - rsm->r_start) > maxseg) {
4714 		/*
4715 		 * We need to split this the last segment in two.
4716 		 */
4717 		struct bbr_sendmap *nrsm;
4718 
4719 		nrsm = bbr_alloc_full_limit(bbr);
4720 		if (nrsm == NULL) {
4721 			/*
4722 			 * We can't get memory to split, we can either just
4723 			 * not split it. Or retransmit the whole piece, lets
4724 			 * do the large send (BTLP :-) ).
4725 			 */
4726 			goto go_for_it;
4727 		}
4728 		bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg));
4729 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
4730 		if (rsm->r_in_tmap) {
4731 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
4732 			nrsm->r_in_tmap = 1;
4733 		}
4734 		rsm->r_flags &= (~BBR_HAS_FIN);
4735 		rsm = nrsm;
4736 	}
4737 go_for_it:
4738 	bbr->r_ctl.rc_tlp_send = rsm;
4739 	bbr->rc_tlp_rtx_out = 1;
4740 	if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) {
4741 		bbr->r_ctl.rc_tlp_seg_send_cnt++;
4742 		tp->t_rxtshift++;
4743 	} else {
4744 		bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
4745 		bbr->r_ctl.rc_tlp_seg_send_cnt = 1;
4746 	}
4747 send:
4748 	if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
4749 		/*
4750 		 * Can't [re]/transmit a segment we have retranmitted the
4751 		 * max times. We need the retransmit timer to take over.
4752 		 */
4753 restore:
4754 		bbr->rc_tlp_new_data = 0;
4755 		bbr->r_ctl.rc_tlp_send = NULL;
4756 		if (rsm)
4757 			rsm->r_flags &= ~BBR_TLP;
4758 		BBR_STAT_INC(bbr_tlp_retran_fail);
4759 		return (0);
4760 	} else if (rsm) {
4761 		rsm->r_flags |= BBR_TLP;
4762 	}
4763 	if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) &&
4764 	    (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) {
4765 		/*
4766 		 * We have retransmitted to many times for TLP. Switch to
4767 		 * the regular RTO timer
4768 		 */
4769 		goto restore;
4770 	}
4771 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP);
4772 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
4773 	return (0);
4774 }
4775 
4776 /*
4777  * Delayed ack Timer, here we simply need to setup the
4778  * ACK_NOW flag and remove the DELACK flag. From there
4779  * the output routine will send the ack out.
4780  *
4781  * We only return 1, saying don't proceed, if all timers
4782  * are stopped (destroyed PCB?).
4783  */
4784 static int
4785 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4786 {
4787 	if (bbr->rc_all_timers_stopped) {
4788 		return (1);
4789 	}
4790 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK);
4791 	tp->t_flags &= ~TF_DELACK;
4792 	tp->t_flags |= TF_ACKNOW;
4793 	KMOD_TCPSTAT_INC(tcps_delack);
4794 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4795 	return (0);
4796 }
4797 
4798 /*
4799  * Here we send a KEEP-ALIVE like probe to the
4800  * peer, we do not send data.
4801  *
4802  * We only return 1, saying don't proceed, if all timers
4803  * are stopped (destroyed PCB?).
4804  */
4805 static int
4806 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4807 {
4808 	struct tcptemp *t_template;
4809 	int32_t retval = 1;
4810 
4811 	if (bbr->rc_all_timers_stopped) {
4812 		return (1);
4813 	}
4814 	if (bbr->rc_in_persist == 0)
4815 		return (0);
4816 	KASSERT(tp->t_inpcb != NULL,
4817 	    ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
4818 	/*
4819 	 * Persistence timer into zero window. Force a byte to be output, if
4820 	 * possible.
4821 	 */
4822 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST);
4823 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
4824 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
4825 	/*
4826 	 * Have we exceeded the user specified progress time?
4827 	 */
4828 	if (ctf_progress_timeout_check(tp, true)) {
4829 		bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4830 		tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4831 		goto out;
4832 	}
4833 	/*
4834 	 * Hack: if the peer is dead/unreachable, we do not time out if the
4835 	 * window is closed.  After a full backoff, drop the connection if
4836 	 * the idle time (no responses to probes) reaches the maximum
4837 	 * backoff that we would use if retransmitting.
4838 	 */
4839 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
4840 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
4841 	    ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
4842 		KMOD_TCPSTAT_INC(tcps_persistdrop);
4843 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4844 		tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4845 		goto out;
4846 	}
4847 	if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) &&
4848 	    tp->snd_una == tp->snd_max) {
4849 		bbr_exit_persist(tp, bbr, cts, __LINE__);
4850 		retval = 0;
4851 		goto out;
4852 	}
4853 	/*
4854 	 * If the user has closed the socket then drop a persisting
4855 	 * connection after a much reduced timeout.
4856 	 */
4857 	if (tp->t_state > TCPS_CLOSE_WAIT &&
4858 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
4859 		KMOD_TCPSTAT_INC(tcps_persistdrop);
4860 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4861 		tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4862 		goto out;
4863 	}
4864 	t_template = tcpip_maketemplate(bbr->rc_inp);
4865 	if (t_template) {
4866 		tcp_respond(tp, t_template->tt_ipgen,
4867 			    &t_template->tt_t, (struct mbuf *)NULL,
4868 			    tp->rcv_nxt, tp->snd_una - 1, 0);
4869 		/* This sends an ack */
4870 		if (tp->t_flags & TF_DELACK)
4871 			tp->t_flags &= ~TF_DELACK;
4872 		free(t_template, M_TEMP);
4873 	}
4874 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
4875 		tp->t_rxtshift++;
4876 	bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0);
4877 out:
4878 	return (retval);
4879 }
4880 
4881 /*
4882  * If a keepalive goes off, we had no other timers
4883  * happening. We always return 1 here since this
4884  * routine either drops the connection or sends
4885  * out a segment with respond.
4886  */
4887 static int
4888 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4889 {
4890 	struct tcptemp *t_template;
4891 	struct inpcb *inp;
4892 
4893 	if (bbr->rc_all_timers_stopped) {
4894 		return (1);
4895 	}
4896 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
4897 	inp = tp->t_inpcb;
4898 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP);
4899 	/*
4900 	 * Keep-alive timer went off; send something or drop connection if
4901 	 * idle for too long.
4902 	 */
4903 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
4904 	if (tp->t_state < TCPS_ESTABLISHED)
4905 		goto dropit;
4906 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4907 	    tp->t_state <= TCPS_CLOSING) {
4908 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
4909 			goto dropit;
4910 		/*
4911 		 * Send a packet designed to force a response if the peer is
4912 		 * up and reachable: either an ACK if the connection is
4913 		 * still alive, or an RST if the peer has closed the
4914 		 * connection due to timeout or reboot. Using sequence
4915 		 * number tp->snd_una-1 causes the transmitted zero-length
4916 		 * segment to lie outside the receive window; by the
4917 		 * protocol spec, this requires the correspondent TCP to
4918 		 * respond.
4919 		 */
4920 		KMOD_TCPSTAT_INC(tcps_keepprobe);
4921 		t_template = tcpip_maketemplate(inp);
4922 		if (t_template) {
4923 			tcp_respond(tp, t_template->tt_ipgen,
4924 			    &t_template->tt_t, (struct mbuf *)NULL,
4925 			    tp->rcv_nxt, tp->snd_una - 1, 0);
4926 			free(t_template, M_TEMP);
4927 		}
4928 	}
4929 	bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0);
4930 	return (1);
4931 dropit:
4932 	KMOD_TCPSTAT_INC(tcps_keepdrops);
4933 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
4934 	tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4935 	return (1);
4936 }
4937 
4938 /*
4939  * Retransmit helper function, clear up all the ack
4940  * flags and take care of important book keeping.
4941  */
4942 static void
4943 bbr_remxt_tmr(struct tcpcb *tp)
4944 {
4945 	/*
4946 	 * The retransmit timer went off, all sack'd blocks must be
4947 	 * un-acked.
4948 	 */
4949 	struct bbr_sendmap *rsm, *trsm = NULL;
4950 	struct tcp_bbr *bbr;
4951 	uint32_t cts, lost;
4952 
4953 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
4954 	cts = tcp_get_usecs(&bbr->rc_tv);
4955 	lost = bbr->r_ctl.rc_lost;
4956 	if (bbr->r_state && (bbr->r_state != tp->t_state))
4957 		bbr_set_state(tp, bbr, 0);
4958 
4959 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
4960 		if (rsm->r_flags & BBR_ACKED) {
4961 			uint32_t old_flags;
4962 
4963 			rsm->r_dupack = 0;
4964 			if (rsm->r_in_tmap == 0) {
4965 				/* We must re-add it back to the tlist */
4966 				if (trsm == NULL) {
4967 					TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
4968 				} else {
4969 					TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext);
4970 				}
4971 				rsm->r_in_tmap = 1;
4972 			}
4973 			old_flags = rsm->r_flags;
4974 			rsm->r_flags |= BBR_RXT_CLEARED;
4975 			rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS);
4976 			bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
4977 		} else {
4978 			if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4979 				bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4980 				bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4981 			}
4982 			if (bbr_marks_rxt_sack_passed) {
4983 				/*
4984 				 * With this option, we will rack out
4985 				 * in 1ms increments the rest of the packets.
4986 				 */
4987 				rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST;
4988 				rsm->r_flags &= ~BBR_WAS_SACKPASS;
4989 			} else {
4990 				/*
4991 				 * With this option we only mark them lost
4992 				 * and remove all sack'd markings. We will run
4993 				 * another RXT or a TLP. This will cause
4994 				 * us to eventually send more based on what
4995 				 * ack's come in.
4996 				 */
4997 				rsm->r_flags |= BBR_MARKED_LOST;
4998 				rsm->r_flags &= ~BBR_WAS_SACKPASS;
4999 				rsm->r_flags &= ~BBR_SACK_PASSED;
5000 			}
5001 		}
5002 		trsm = rsm;
5003 	}
5004 	bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map);
5005 	/* Clear the count (we just un-acked them) */
5006 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR);
5007 	bbr->rc_tlp_new_data = 0;
5008 	bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
5009 	/* zap the behindness on a rxt */
5010 	bbr->r_ctl.rc_hptsi_agg_delay = 0;
5011 	bbr->r_agg_early_set = 0;
5012 	bbr->r_ctl.rc_agg_early = 0;
5013 	bbr->rc_tlp_rtx_out = 0;
5014 	bbr->r_ctl.rc_sacked = 0;
5015 	bbr->r_ctl.rc_sacklast = NULL;
5016 	bbr->r_timer_override = 1;
5017 	bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
5018 }
5019 
5020 /*
5021  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
5022  * we will setup to retransmit the lowest seq number outstanding.
5023  */
5024 static int
5025 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
5026 {
5027 	int32_t rexmt;
5028 	int32_t retval = 0;
5029 	bool isipv6;
5030 
5031 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
5032 	if (bbr->rc_all_timers_stopped) {
5033 		return (1);
5034 	}
5035 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
5036 	    (tp->snd_una == tp->snd_max)) {
5037 		/* Nothing outstanding .. nothing to do */
5038 		return (0);
5039 	}
5040 	/*
5041 	 * Retransmission timer went off.  Message has not been acked within
5042 	 * retransmit interval.  Back off to a longer retransmit interval
5043 	 * and retransmit one segment.
5044 	 */
5045 	if (ctf_progress_timeout_check(tp, true)) {
5046 		retval = 1;
5047 		bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
5048 		tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
5049 		goto out;
5050 	}
5051 	bbr_remxt_tmr(tp);
5052 	if ((bbr->r_ctl.rc_resend == NULL) ||
5053 	    ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) {
5054 		/*
5055 		 * If the rwnd collapsed on
5056 		 * the one we are retransmitting
5057 		 * it does not count against the
5058 		 * rxt count.
5059 		 */
5060 		tp->t_rxtshift++;
5061 	}
5062 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
5063 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
5064 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
5065 		retval = 1;
5066 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5067 		tcp_set_inp_to_drop(bbr->rc_inp,
5068 		    (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
5069 		goto out;
5070 	}
5071 	if (tp->t_state == TCPS_SYN_SENT) {
5072 		/*
5073 		 * If the SYN was retransmitted, indicate CWND to be limited
5074 		 * to 1 segment in cc_conn_init().
5075 		 */
5076 		tp->snd_cwnd = 1;
5077 	} else if (tp->t_rxtshift == 1) {
5078 		/*
5079 		 * first retransmit; record ssthresh and cwnd so they can be
5080 		 * recovered if this turns out to be a "bad" retransmit. A
5081 		 * retransmit is considered "bad" if an ACK for this segment
5082 		 * is received within RTT/2 interval; the assumption here is
5083 		 * that the ACK was already in flight.  See "On Estimating
5084 		 * End-to-End Network Path Properties" by Allman and Paxson
5085 		 * for more details.
5086 		 */
5087 		tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5088 		if (!IN_RECOVERY(tp->t_flags)) {
5089 			tp->snd_cwnd_prev = tp->snd_cwnd;
5090 			tp->snd_ssthresh_prev = tp->snd_ssthresh;
5091 			tp->snd_recover_prev = tp->snd_recover;
5092 			tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5093 			tp->t_flags |= TF_PREVVALID;
5094 		} else {
5095 			tp->t_flags &= ~TF_PREVVALID;
5096 		}
5097 		tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5098 	} else {
5099 		tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5100 		tp->t_flags &= ~TF_PREVVALID;
5101 	}
5102 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
5103 	if ((tp->t_state == TCPS_SYN_SENT) ||
5104 	    (tp->t_state == TCPS_SYN_RECEIVED))
5105 		rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift];
5106 	else
5107 		rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5108 	TCPT_RANGESET(tp->t_rxtcur, rexmt,
5109 	    MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms),
5110 	    MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
5111 	/*
5112 	 * We enter the path for PLMTUD if connection is established or, if
5113 	 * connection is FIN_WAIT_1 status, reason for the last is that if
5114 	 * amount of data we send is very small, we could send it in couple
5115 	 * of packets and process straight to FIN. In that case we won't
5116 	 * catch ESTABLISHED state.
5117 	 */
5118 #ifdef INET6
5119 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
5120 #else
5121 	isipv6 = false;
5122 #endif
5123 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
5124 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
5125 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
5126 	    ((tp->t_state == TCPS_ESTABLISHED) ||
5127 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
5128 
5129 		/*
5130 		 * Idea here is that at each stage of mtu probe (usually,
5131 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
5132 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5133 		 * should take care of that.
5134 		 */
5135 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5136 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5137 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5138 		    tp->t_rxtshift % 2 == 0)) {
5139 			/*
5140 			 * Enter Path MTU Black-hole Detection mechanism: -
5141 			 * Disable Path MTU Discovery (IP "DF" bit). -
5142 			 * Reduce MTU to lower value than what we negotiated
5143 			 * with peer.
5144 			 */
5145 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5146 				/*
5147 				 * Record that we may have found a black
5148 				 * hole.
5149 				 */
5150 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5151 				/* Keep track of previous MSS. */
5152 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5153 			}
5154 			/*
5155 			 * Reduce the MSS to blackhole value or to the
5156 			 * default in an attempt to retransmit.
5157 			 */
5158 #ifdef INET6
5159 			isipv6 = bbr->r_is_v6;
5160 			if (isipv6 &&
5161 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5162 				/* Use the sysctl tuneable blackhole MSS. */
5163 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5164 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5165 			} else if (isipv6) {
5166 				/* Use the default MSS. */
5167 				tp->t_maxseg = V_tcp_v6mssdflt;
5168 				/*
5169 				 * Disable Path MTU Discovery when we switch
5170 				 * to minmss.
5171 				 */
5172 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5173 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5174 			}
5175 #endif
5176 #if defined(INET6) && defined(INET)
5177 			else
5178 #endif
5179 #ifdef INET
5180 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5181 				/* Use the sysctl tuneable blackhole MSS. */
5182 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5183 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5184 			} else {
5185 				/* Use the default MSS. */
5186 				tp->t_maxseg = V_tcp_mssdflt;
5187 				/*
5188 				 * Disable Path MTU Discovery when we switch
5189 				 * to minmss.
5190 				 */
5191 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5192 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5193 			}
5194 #endif
5195 		} else {
5196 			/*
5197 			 * If further retransmissions are still unsuccessful
5198 			 * with a lowered MTU, maybe this isn't a blackhole
5199 			 * and we restore the previous MSS and blackhole
5200 			 * detection flags. The limit '6' is determined by
5201 			 * giving each probe stage (1448, 1188, 524) 2
5202 			 * chances to recover.
5203 			 */
5204 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5205 			    (tp->t_rxtshift >= 6)) {
5206 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5207 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5208 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5209 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5210 			}
5211 		}
5212 	}
5213 	/*
5214 	 * Disable RFC1323 and SACK if we haven't got any response to our
5215 	 * third SYN to work-around some broken terminal servers (most of
5216 	 * which have hopefully been retired) that have bad VJ header
5217 	 * compression code which trashes TCP segments containing
5218 	 * unknown-to-them TCP options.
5219 	 */
5220 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
5221 	    (tp->t_rxtshift == 3))
5222 		tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT);
5223 	/*
5224 	 * If we backed off this far, our srtt estimate is probably bogus.
5225 	 * Clobber it so we'll take the next rtt measurement as our srtt;
5226 	 * move the current srtt into rttvar to keep the current retransmit
5227 	 * times until then.
5228 	 */
5229 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5230 #ifdef INET6
5231 		if (bbr->r_is_v6)
5232 			in6_losing(tp->t_inpcb);
5233 		else
5234 #endif
5235 			in_losing(tp->t_inpcb);
5236 		tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5237 		tp->t_srtt = 0;
5238 	}
5239 	sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
5240 	tp->snd_recover = tp->snd_max;
5241 	tp->t_flags |= TF_ACKNOW;
5242 	tp->t_rtttime = 0;
5243 out:
5244 	return (retval);
5245 }
5246 
5247 static int
5248 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling)
5249 {
5250 	int32_t ret = 0;
5251 	int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5252 
5253 	if (timers == 0) {
5254 		return (0);
5255 	}
5256 	if (tp->t_state == TCPS_LISTEN) {
5257 		/* no timers on listen sockets */
5258 		if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5259 			return (0);
5260 		return (1);
5261 	}
5262 	if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
5263 		uint32_t left;
5264 
5265 		if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5266 			ret = -1;
5267 			bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5268 			return (0);
5269 		}
5270 		if (hpts_calling == 0) {
5271 			ret = -2;
5272 			bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5273 			return (0);
5274 		}
5275 		/*
5276 		 * Ok our timer went off early and we are not paced false
5277 		 * alarm, go back to sleep.
5278 		 */
5279 		left = bbr->r_ctl.rc_timer_exp - cts;
5280 		ret = -3;
5281 		bbr_log_to_processing(bbr, cts, ret, left, hpts_calling);
5282 		tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left));
5283 		return (1);
5284 	}
5285 	bbr->rc_tmr_stopped = 0;
5286 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5287 	if (timers & PACE_TMR_DELACK) {
5288 		ret = bbr_timeout_delack(tp, bbr, cts);
5289 	} else if (timers & PACE_TMR_PERSIT) {
5290 		ret = bbr_timeout_persist(tp, bbr, cts);
5291 	} else if (timers & PACE_TMR_RACK) {
5292 		bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5293 		ret = bbr_timeout_rack(tp, bbr, cts);
5294 	} else if (timers & PACE_TMR_TLP) {
5295 		bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5296 		ret = bbr_timeout_tlp(tp, bbr, cts);
5297 	} else if (timers & PACE_TMR_RXT) {
5298 		bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5299 		ret = bbr_timeout_rxt(tp, bbr, cts);
5300 	} else if (timers & PACE_TMR_KEEP) {
5301 		ret = bbr_timeout_keepalive(tp, bbr, cts);
5302 	}
5303 	bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling);
5304 	return (ret);
5305 }
5306 
5307 static void
5308 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts)
5309 {
5310 	if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5311 		uint8_t hpts_removed = 0;
5312 
5313 		if (bbr->rc_inp->inp_in_hpts &&
5314 		    (bbr->rc_timer_first == 1)) {
5315 			/*
5316 			 * If we are canceling timer's when we have the
5317 			 * timer ahead of the output being paced. We also
5318 			 * must remove ourselves from the hpts.
5319 			 */
5320 			hpts_removed = 1;
5321 			tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
5322 			if (bbr->r_ctl.rc_last_delay_val) {
5323 				/* Update the last hptsi delay too */
5324 				uint32_t time_since_send;
5325 
5326 				if (TSTMP_GT(cts, bbr->rc_pacer_started))
5327 					time_since_send = cts - bbr->rc_pacer_started;
5328 				else
5329 					time_since_send = 0;
5330 				if (bbr->r_ctl.rc_last_delay_val > time_since_send) {
5331 					/* Cut down our slot time */
5332 					bbr->r_ctl.rc_last_delay_val -= time_since_send;
5333 				} else {
5334 					bbr->r_ctl.rc_last_delay_val = 0;
5335 				}
5336 				bbr->rc_pacer_started = cts;
5337 			}
5338 		}
5339 		bbr->rc_timer_first = 0;
5340 		bbr_log_to_cancel(bbr, line, cts, hpts_removed);
5341 		bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5342 		bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5343 	}
5344 }
5345 
5346 static void
5347 bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type)
5348 {
5349 	struct tcp_bbr *bbr;
5350 
5351 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
5352 	bbr->rc_all_timers_stopped = 1;
5353 	return;
5354 }
5355 
5356 /*
5357  * stop all timers always returning 0.
5358  */
5359 static int
5360 bbr_stopall(struct tcpcb *tp)
5361 {
5362 	return (0);
5363 }
5364 
5365 static void
5366 bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
5367 {
5368 	return;
5369 }
5370 
5371 /*
5372  * return true if a bbr timer (rack or tlp) is active.
5373  */
5374 static int
5375 bbr_timer_active(struct tcpcb *tp, uint32_t timer_type)
5376 {
5377 	return (0);
5378 }
5379 
5380 static uint32_t
5381 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts)
5382 {
5383 	struct bbr_sendmap *rsm;
5384 
5385 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
5386 	if ((rsm == NULL) || (u_rsm == rsm))
5387 		return (cts);
5388 	return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
5389 }
5390 
5391 static void
5392 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr,
5393      struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time)
5394 {
5395 	int32_t idx;
5396 
5397 	rsm->r_rtr_cnt++;
5398 	rsm->r_dupack = 0;
5399 	if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) {
5400 		rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS;
5401 		rsm->r_flags |= BBR_OVERMAX;
5402 	}
5403 	if (rsm->r_flags & BBR_RWND_COLLAPSED) {
5404 		/* Take off the collapsed flag at rxt */
5405 		rsm->r_flags &= ~BBR_RWND_COLLAPSED;
5406 	}
5407 	if (rsm->r_flags & BBR_MARKED_LOST) {
5408 		/* We have retransmitted, its no longer lost */
5409 		rsm->r_flags &= ~BBR_MARKED_LOST;
5410 		bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
5411 	}
5412 	if (rsm->r_flags & BBR_RXT_CLEARED) {
5413 		/*
5414 		 * We hit a RXT timer on it and
5415 		 * we cleared the "acked" flag.
5416 		 * We now have it going back into
5417 		 * flight, we can remove the cleared
5418 		 * flag and possibly do accounting on
5419 		 * this piece.
5420 		 */
5421 		rsm->r_flags &= ~BBR_RXT_CLEARED;
5422 	}
5423 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) {
5424 		bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5425 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5426 	}
5427 	idx = rsm->r_rtr_cnt - 1;
5428 	rsm->r_tim_lastsent[idx] = cts;
5429 	rsm->r_pacing_delay = pacing_time;
5430 	rsm->r_delivered = bbr->r_ctl.rc_delivered;
5431 	rsm->r_ts_valid = bbr->rc_ts_valid;
5432 	if (bbr->rc_ts_valid)
5433 		rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5434 	if (bbr->r_ctl.r_app_limited_until)
5435 		rsm->r_app_limited = 1;
5436 	else
5437 		rsm->r_app_limited = 0;
5438 	if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5439 		rsm->r_bbr_state = bbr_state_val(bbr);
5440 	else
5441 		rsm->r_bbr_state = 8;
5442 	if (rsm->r_flags & BBR_ACKED) {
5443 		/* Problably MTU discovery messing with us */
5444 		uint32_t old_flags;
5445 
5446 		old_flags = rsm->r_flags;
5447 		rsm->r_flags &= ~BBR_ACKED;
5448 		bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
5449 		bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5450 		if (bbr->r_ctl.rc_sacked == 0)
5451 			bbr->r_ctl.rc_sacklast = NULL;
5452 	}
5453 	if (rsm->r_in_tmap) {
5454 		TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5455 	}
5456 	TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5457 	rsm->r_in_tmap = 1;
5458 	if (rsm->r_flags & BBR_SACK_PASSED) {
5459 		/* We have retransmitted due to the SACK pass */
5460 		rsm->r_flags &= ~BBR_SACK_PASSED;
5461 		rsm->r_flags |= BBR_WAS_SACKPASS;
5462 	}
5463 	rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5464 	rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5465 						(bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5466 	bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
5467 	if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5468 		rsm->r_is_gain = 1;
5469 		rsm->r_is_drain = 0;
5470 	} else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5471 		rsm->r_is_drain = 1;
5472 		rsm->r_is_gain = 0;
5473 	} else {
5474 		rsm->r_is_drain = 0;
5475 		rsm->r_is_gain = 0;
5476 	}
5477 	rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */
5478 }
5479 
5480 /*
5481  * Returns 0, or the sequence where we stopped
5482  * updating. We also update the lenp to be the amount
5483  * of data left.
5484  */
5485 
5486 static uint32_t
5487 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr,
5488     struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time)
5489 {
5490 	/*
5491 	 * We (re-)transmitted starting at rsm->r_start for some length
5492 	 * (possibly less than r_end.
5493 	 */
5494 	struct bbr_sendmap *nrsm;
5495 	uint32_t c_end;
5496 	int32_t len;
5497 
5498 	len = *lenp;
5499 	c_end = rsm->r_start + len;
5500 	if (SEQ_GEQ(c_end, rsm->r_end)) {
5501 		/*
5502 		 * We retransmitted the whole piece or more than the whole
5503 		 * slopping into the next rsm.
5504 		 */
5505 		bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5506 		if (c_end == rsm->r_end) {
5507 			*lenp = 0;
5508 			return (0);
5509 		} else {
5510 			int32_t act_len;
5511 
5512 			/* Hangs over the end return whats left */
5513 			act_len = rsm->r_end - rsm->r_start;
5514 			*lenp = (len - act_len);
5515 			return (rsm->r_end);
5516 		}
5517 		/* We don't get out of this block. */
5518 	}
5519 	/*
5520 	 * Here we retransmitted less than the whole thing which means we
5521 	 * have to split this into what was transmitted and what was not.
5522 	 */
5523 	nrsm = bbr_alloc_full_limit(bbr);
5524 	if (nrsm == NULL) {
5525 		*lenp = 0;
5526 		return (0);
5527 	}
5528 	/*
5529 	 * So here we are going to take the original rsm and make it what we
5530 	 * retransmitted. nrsm will be the tail portion we did not
5531 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
5532 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
5533 	 * 1, 6 and the new piece will be 6, 11.
5534 	 */
5535 	bbr_clone_rsm(bbr, nrsm, rsm, c_end);
5536 	TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
5537 	nrsm->r_dupack = 0;
5538 	if (rsm->r_in_tmap) {
5539 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5540 		nrsm->r_in_tmap = 1;
5541 	}
5542 	rsm->r_flags &= (~BBR_HAS_FIN);
5543 	bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5544 	*lenp = 0;
5545 	return (0);
5546 }
5547 
5548 static uint64_t
5549 bbr_get_hardware_rate(struct tcp_bbr *bbr)
5550 {
5551 	uint64_t bw;
5552 
5553 	bw = bbr_get_bw(bbr);
5554 	bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN];
5555 	bw /= (uint64_t)BBR_UNIT;
5556 	return(bw);
5557 }
5558 
5559 static void
5560 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts,
5561 		       uint64_t act_rate, uint64_t rate_wanted)
5562 {
5563 	/*
5564 	 * We could not get a full gains worth
5565 	 * of rate.
5566 	 */
5567 	if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) {
5568 		/* we can't even get the real rate */
5569 		uint64_t red;
5570 
5571 		bbr->skip_gain = 1;
5572 		bbr->gain_is_limited = 0;
5573 		red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate;
5574 		if (red)
5575 			filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts);
5576 	} else {
5577 		/* We can use a lower gain */
5578 		bbr->skip_gain = 0;
5579 		bbr->gain_is_limited = 1;
5580 	}
5581 }
5582 
5583 static void
5584 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts)
5585 {
5586 	const struct tcp_hwrate_limit_table *nrte;
5587 	int error, rate = -1;
5588 
5589 	if (bbr->r_ctl.crte == NULL)
5590 		return;
5591 	if ((bbr->rc_inp->inp_route.ro_nh == NULL) ||
5592 	    (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) {
5593 		/* Lost our routes? */
5594 		/* Clear the way for a re-attempt */
5595 		bbr->bbr_attempt_hdwr_pace = 0;
5596 lost_rate:
5597 		bbr->gain_is_limited = 0;
5598 		bbr->skip_gain = 0;
5599 		bbr->bbr_hdrw_pacing = 0;
5600 		counter_u64_add(bbr_flows_whdwr_pacing, -1);
5601 		counter_u64_add(bbr_flows_nohdwr_pacing, 1);
5602 		tcp_bbr_tso_size_check(bbr, cts);
5603 		return;
5604 	}
5605 	rate = bbr_get_hardware_rate(bbr);
5606 	nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte,
5607 				   bbr->rc_tp,
5608 				   bbr->rc_inp->inp_route.ro_nh->nh_ifp,
5609 				   rate,
5610 				   (RS_PACING_GEQ|RS_PACING_SUB_OK),
5611 				   &error);
5612 	if (nrte == NULL) {
5613 		goto lost_rate;
5614 	}
5615 	if (nrte != bbr->r_ctl.crte) {
5616 		bbr->r_ctl.crte = nrte;
5617 		if (error == 0)  {
5618 			BBR_STAT_INC(bbr_hdwr_rl_mod_ok);
5619 			if (bbr->r_ctl.crte->rate < rate) {
5620 				/* We have a problem */
5621 				bbr_setup_less_of_rate(bbr, cts,
5622 						       bbr->r_ctl.crte->rate, rate);
5623 			} else {
5624 				/* We are good */
5625 				bbr->gain_is_limited = 0;
5626 				bbr->skip_gain = 0;
5627 			}
5628 		} else {
5629 			/* A failure should release the tag */
5630 			BBR_STAT_INC(bbr_hdwr_rl_mod_fail);
5631 			bbr->gain_is_limited = 0;
5632 			bbr->skip_gain = 0;
5633 			bbr->bbr_hdrw_pacing = 0;
5634 		}
5635 		bbr_type_log_hdwr_pacing(bbr,
5636 					 bbr->r_ctl.crte->ptbl->rs_ifp,
5637 					 rate,
5638 					 ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate),
5639 					 __LINE__,
5640 					 cts,
5641 					 error);
5642 	}
5643 }
5644 
5645 static void
5646 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts)
5647 {
5648 	/*
5649 	 * If we have hardware pacing support
5650 	 * we need to factor that in for our
5651 	 * TSO size.
5652 	 */
5653 	const struct tcp_hwrate_limit_table *rlp;
5654 	uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay;
5655 
5656 	if ((bbr->bbr_hdrw_pacing == 0) ||
5657 	    (IN_RECOVERY(bbr->rc_tp->t_flags)) ||
5658 	    (bbr->r_ctl.crte == NULL))
5659 		return;
5660 	if (bbr->hw_pacing_set == 0) {
5661 		/* Not yet by the hdwr pacing count delay */
5662 		return;
5663 	}
5664 	if (bbr_hdwr_pace_adjust == 0) {
5665 		/* No adjustment */
5666 		return;
5667 	}
5668 	rlp = bbr->r_ctl.crte;
5669 	if (bbr->rc_tp->t_maxseg > bbr->rc_last_options)
5670 		maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5671 	else
5672 		maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5673 	/*
5674 	 * So lets first get the
5675 	 * time we will take between
5676 	 * TSO sized sends currently without
5677 	 * hardware help.
5678 	 */
5679 	cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT,
5680 		        bbr->r_ctl.rc_pace_max_segs, cts, 1);
5681 	hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg;
5682 	hdwr_delay *= rlp->time_between;
5683 	if (cur_delay > hdwr_delay)
5684 		delta = cur_delay - hdwr_delay;
5685 	else
5686 		delta = 0;
5687 	bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay,
5688 			     (bbr->r_ctl.rc_pace_max_segs / maxseg),
5689 			     1);
5690 	if (delta &&
5691 	    (delta < (max(rlp->time_between,
5692 			  bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) {
5693 		/*
5694 		 * Now lets divide by the pacing
5695 		 * time between each segment the
5696 		 * hardware sends rounding up and
5697 		 * derive a bytes from that. We multiply
5698 		 * that by bbr_hdwr_pace_adjust to get
5699 		 * more bang for our buck.
5700 		 *
5701 		 * The goal is to have the software pacer
5702 		 * waiting no more than an additional
5703 		 * pacing delay if we can (without the
5704 		 * compensation i.e. x bbr_hdwr_pace_adjust).
5705 		 */
5706 		seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between),
5707 			     (bbr->r_ctl.rc_pace_max_segs/maxseg));
5708 		seg_sz *= bbr_hdwr_pace_adjust;
5709 		if (bbr_hdwr_pace_floor &&
5710 		    (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5711 			/* Currently hardware paces
5712 			 * out rs_min_seg segments at a time.
5713 			 * We need to make sure we always send at least
5714 			 * a full burst of bbr_hdwr_pace_floor down.
5715 			 */
5716 			seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5717 		}
5718 		seg_sz *= maxseg;
5719 	} else if (delta == 0) {
5720 		/*
5721 		 * The highest pacing rate is
5722 		 * above our b/w gained. This means
5723 		 * we probably are going quite fast at
5724 		 * the hardware highest rate. Lets just multiply
5725 		 * the calculated TSO size by the
5726 		 * multiplier factor (its probably
5727 		 * 4 segments in the default config for
5728 		 * mlx).
5729 		 */
5730 		seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust;
5731 		if (bbr_hdwr_pace_floor &&
5732 		    (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5733 			/* Currently hardware paces
5734 			 * out rs_min_seg segments at a time.
5735 			 * We need to make sure we always send at least
5736 			 * a full burst of bbr_hdwr_pace_floor down.
5737 			 */
5738 			seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5739 		}
5740 	} else {
5741 		/*
5742 		 * The pacing time difference is so
5743 		 * big that the hardware will
5744 		 * pace out more rapidly then we
5745 		 * really want and then we
5746 		 * will have a long delay. Lets just keep
5747 		 * the same TSO size so its as if
5748 		 * we were not using hdwr pacing (we
5749 		 * just gain a bit of spacing from the
5750 		 * hardware if seg_sz > 1).
5751 		 */
5752 		seg_sz = bbr->r_ctl.rc_pace_max_segs;
5753 	}
5754 	if (seg_sz > bbr->r_ctl.rc_pace_max_segs)
5755 		new_tso = seg_sz;
5756 	else
5757 		new_tso = bbr->r_ctl.rc_pace_max_segs;
5758 	if (new_tso >= (PACE_MAX_IP_BYTES-maxseg))
5759 		new_tso = PACE_MAX_IP_BYTES - maxseg;
5760 
5761 	if (new_tso != bbr->r_ctl.rc_pace_max_segs) {
5762 		bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0);
5763 		bbr->r_ctl.rc_pace_max_segs = new_tso;
5764 	}
5765 }
5766 
5767 static void
5768 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts)
5769 {
5770 	uint64_t bw;
5771 	uint32_t old_tso = 0, new_tso;
5772 	uint32_t maxseg, bytes;
5773 	uint32_t tls_seg=0;
5774 	/*
5775 	 * Google/linux uses the following algorithm to determine
5776 	 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18):
5777 	 *
5778 	 *  bytes = bw_in_bytes_per_second / 1000
5779 	 *  bytes = min(bytes, 64k)
5780 	 *  tso_segs = bytes / MSS
5781 	 *  if (bw < 1.2Mbs)
5782 	 *      min_tso_segs = 1
5783 	 *  else
5784 	 *	min_tso_segs = 2
5785 	 * tso_segs = max(tso_segs, min_tso_segs)
5786 	 *
5787 	 * * Note apply a device specific limit (we apply this in the
5788 	 *   tcp_m_copym).
5789 	 * Note that before the initial measurement is made google bursts out
5790 	 * a full iwnd just like new-reno/cubic.
5791 	 *
5792 	 * We do not use this algorithm. Instead we
5793 	 * use a two phased approach:
5794 	 *
5795 	 *  if ( bw <= per-tcb-cross-over)
5796 	 *     goal_tso =  calculate how much with this bw we
5797 	 *                 can send in goal-time seconds.
5798 	 *     if (goal_tso > mss)
5799 	 *         seg = goal_tso / mss
5800 	 *         tso = seg * mss
5801 	 *     else
5802          *         tso = mss
5803 	 *     if (tso > per-tcb-max)
5804 	 *         tso = per-tcb-max
5805 	 *  else if ( bw > 512Mbps)
5806 	 *     tso = max-tso (64k/mss)
5807 	 *  else
5808 	 *     goal_tso = bw / per-tcb-divsor
5809 	 *     seg = (goal_tso + mss-1)/mss
5810 	 *     tso = seg * mss
5811 	 *
5812 	 * if (tso < per-tcb-floor)
5813 	 *    tso = per-tcb-floor
5814 	 * if (tso > per-tcb-utter_max)
5815 	 *    tso = per-tcb-utter_max
5816 	 *
5817 	 * Note the default per-tcb-divisor is 1000 (same as google).
5818 	 * the goal cross over is 30Mbps however. To recreate googles
5819 	 * algorithm you need to set:
5820 	 *
5821 	 * cross-over = 23,168,000 bps
5822 	 * goal-time = 18000
5823 	 * per-tcb-max = 2
5824 	 * per-tcb-divisor = 1000
5825 	 * per-tcb-floor = 1
5826 	 *
5827 	 * This will get you "google bbr" behavior with respect to tso size.
5828 	 *
5829 	 * Note we do set anything TSO size until we are past the initial
5830 	 * window. Before that we gnerally use either a single MSS
5831 	 * or we use the full IW size (so we burst a IW at a time)
5832 	 * Also note that Hardware-TLS is special and does alternate
5833 	 * things to minimize PCI Bus Bandwidth use.
5834 	 */
5835 
5836 	if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) {
5837 		maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5838 	} else {
5839 		maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5840 	}
5841 #ifdef KERN_TLS
5842 	if (bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) {
5843 		tls_seg =  ctf_get_opt_tls_size(bbr->rc_inp->inp_socket, bbr->rc_tp->snd_wnd);
5844 		bbr->r_ctl.rc_pace_min_segs = (tls_seg + bbr->rc_last_options);
5845 	}
5846 #endif
5847 	old_tso = bbr->r_ctl.rc_pace_max_segs;
5848 	if (bbr->rc_past_init_win == 0) {
5849 		/*
5850 		 * Not enough data has been acknowledged to make a
5851 		 * judgement unless we are hardware TLS. Set up
5852 		 * the initial TSO based on if we are sending a
5853 		 * full IW at once or not.
5854 		 */
5855 		if (bbr->rc_use_google)
5856 			bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2);
5857 		else if (bbr->bbr_init_win_cheat)
5858 			bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp);
5859 		else
5860 			bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5861 		if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg)
5862 			bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg;
5863 #ifdef KERN_TLS
5864 		if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) && tls_seg) {
5865 			/*
5866 			 * For hardware TLS we set our min to the tls_seg size.
5867 			 */
5868 			bbr->r_ctl.rc_pace_max_segs = tls_seg;
5869 			bbr->r_ctl.rc_pace_min_segs = tls_seg + bbr->rc_last_options;
5870 		}
5871 #endif
5872 		if (bbr->r_ctl.rc_pace_max_segs == 0) {
5873 			bbr->r_ctl.rc_pace_max_segs = maxseg;
5874 		}
5875 		bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0);
5876 #ifdef KERN_TLS
5877 		if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) == 0)
5878 #endif
5879 			bbr_adjust_for_hw_pacing(bbr, cts);
5880 		return;
5881 	}
5882 	/**
5883 	 * Now lets set the TSO goal based on our delivery rate in
5884 	 * bytes per second. Note we only do this if
5885 	 * we have acked at least the initial cwnd worth of data.
5886 	 */
5887 	bw = bbr_get_bw(bbr);
5888 	if (IN_RECOVERY(bbr->rc_tp->t_flags) &&
5889 	     (bbr->rc_use_google == 0)) {
5890 		/* We clamp to one MSS in recovery */
5891 		new_tso = maxseg;
5892 	} else if (bbr->rc_use_google) {
5893 		int min_tso_segs;
5894 
5895 		/* Google considers the gain too */
5896 		if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) {
5897 			bw *= bbr->r_ctl.rc_bbr_hptsi_gain;
5898 			bw /= BBR_UNIT;
5899 		}
5900 		bytes = bw / 1024;
5901 		if (bytes > (64 * 1024))
5902 			bytes = 64 * 1024;
5903 		new_tso = bytes / maxseg;
5904 		if (bw < ONE_POINT_TWO_MEG)
5905 			min_tso_segs = 1;
5906 		else
5907 			min_tso_segs = 2;
5908 		if (new_tso < min_tso_segs)
5909 			new_tso = min_tso_segs;
5910 		new_tso *= maxseg;
5911 	} else if (bbr->rc_no_pacing) {
5912 		new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg;
5913 	} else if (bw <= bbr->r_ctl.bbr_cross_over) {
5914 		/*
5915 		 * Calculate the worse case b/w TSO if we are inserting no
5916 		 * more than a delay_target number of TSO's.
5917 		 */
5918 		uint32_t tso_len, min_tso;
5919 
5920 		tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw);
5921 		if (tso_len > maxseg) {
5922 			new_tso = tso_len / maxseg;
5923 			if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max)
5924 				new_tso = bbr->r_ctl.bbr_hptsi_segments_max;
5925 			new_tso *= maxseg;
5926 		} else {
5927 			/*
5928 			 * less than a full sized frame yikes.. long rtt or
5929 			 * low bw?
5930 			 */
5931 			min_tso = bbr_minseg(bbr);
5932 			if ((tso_len > min_tso) && (bbr_all_get_min == 0))
5933 				new_tso = rounddown(tso_len, min_tso);
5934 			else
5935 				new_tso = min_tso;
5936 		}
5937 	} else if (bw > FIVETWELVE_MBPS) {
5938 		/*
5939 		 * This guy is so fast b/w wise that we can TSO as large as
5940 		 * possible of segments that the NIC will allow.
5941 		 */
5942 		new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5943 	} else {
5944 		/*
5945 		 * This formula is based on attempting to send a segment or
5946 		 * more every bbr_hptsi_per_second. The default is 1000
5947 		 * which means you are targeting what you can send every 1ms
5948 		 * based on the peers bw.
5949 		 *
5950 		 * If the number drops to say 500, then you are looking more
5951 		 * at 2ms and you will raise how much we send in a single
5952 		 * TSO thus saving CPU (less bbr_output_wtime() calls). The
5953 		 * trade off of course is you will send more at once and
5954 		 * thus tend to clump up the sends into larger "bursts"
5955 		 * building a queue.
5956 		 */
5957 		bw /= bbr->r_ctl.bbr_hptsi_per_second;
5958 		new_tso = roundup(bw, (uint64_t)maxseg);
5959 		/*
5960 		 * Gate the floor to match what our lower than 48Mbps
5961 		 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus
5962 		 * becomes the floor for this calculation.
5963 		 */
5964 		if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg))
5965 			new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg);
5966 	}
5967 	if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor)))
5968 		new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor;
5969 	if (new_tso > PACE_MAX_IP_BYTES)
5970 		new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5971 	/* Enforce an utter maximum if we are not HW-TLS */
5972 #ifdef KERN_TLS
5973 	if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) == 0)
5974 #endif
5975 		if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) {
5976 			new_tso = bbr->r_ctl.bbr_utter_max * maxseg;
5977 		}
5978 #ifdef KERN_TLS
5979 	if (tls_seg) {
5980 		/*
5981 		 * Lets move the output size
5982 		 * up to 1 or more TLS record sizes.
5983 		 */
5984 		uint32_t temp;
5985 
5986 		temp = roundup(new_tso, tls_seg);
5987 		new_tso = temp;
5988 		/* Back down if needed to under a full frame */
5989 		while (new_tso > PACE_MAX_IP_BYTES)
5990 			new_tso -= tls_seg;
5991 	}
5992 #endif
5993 	if (old_tso != new_tso) {
5994 		/* Only log changes */
5995 		bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0);
5996 		bbr->r_ctl.rc_pace_max_segs = new_tso;
5997 	}
5998 #ifdef KERN_TLS
5999 	if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) &&
6000 	     tls_seg) {
6001 		bbr->r_ctl.rc_pace_min_segs = tls_seg + bbr->rc_last_options;
6002 	} else
6003 #endif
6004 		/* We have hardware pacing and not hardware TLS! */
6005 		bbr_adjust_for_hw_pacing(bbr, cts);
6006 }
6007 
6008 static void
6009 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len,
6010     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t cts,
6011     struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc,
6012     struct sockbuf *sb)
6013 {
6014 
6015 	struct bbr_sendmap *rsm, *nrsm;
6016 	register uint32_t snd_max, snd_una;
6017 	uint32_t pacing_time;
6018 	/*
6019 	 * Add to the RACK log of packets in flight or retransmitted. If
6020 	 * there is a TS option we will use the TS echoed, if not we will
6021 	 * grab a TS.
6022 	 *
6023 	 * Retransmissions will increment the count and move the ts to its
6024 	 * proper place. Note that if options do not include TS's then we
6025 	 * won't be able to effectively use the ACK for an RTT on a retran.
6026 	 *
6027 	 * Notes about r_start and r_end. Lets consider a send starting at
6028 	 * sequence 1 for 10 bytes. In such an example the r_start would be
6029 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
6030 	 * This means that r_end is actually the first sequence for the next
6031 	 * slot (11).
6032 	 *
6033 	 */
6034 	INP_WLOCK_ASSERT(tp->t_inpcb);
6035 	if (err) {
6036 		/*
6037 		 * We don't log errors -- we could but snd_max does not
6038 		 * advance in this case either.
6039 		 */
6040 		return;
6041 	}
6042 	if (th_flags & TH_RST) {
6043 		/*
6044 		 * We don't log resets and we return immediately from
6045 		 * sending
6046 		 */
6047 		*abandon = 1;
6048 		return;
6049 	}
6050 	snd_una = tp->snd_una;
6051 	if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) {
6052 		/*
6053 		 * The call to bbr_log_output is made before bumping
6054 		 * snd_max. This means we can record one extra byte on a SYN
6055 		 * or FIN if seq_out is adding more on and a FIN is present
6056 		 * (and we are not resending).
6057 		 */
6058 		if (th_flags & TH_SYN)
6059 			len++;
6060 		if (th_flags & TH_FIN)
6061 			len++;
6062 	}
6063 	if (SEQ_LEQ((seq_out + len), snd_una)) {
6064 		/* Are sending an old segment to induce an ack (keep-alive)? */
6065 		return;
6066 	}
6067 	if (SEQ_LT(seq_out, snd_una)) {
6068 		/* huh? should we panic? */
6069 		uint32_t end;
6070 
6071 		end = seq_out + len;
6072 		seq_out = snd_una;
6073 		len = end - seq_out;
6074 	}
6075 	snd_max = tp->snd_max;
6076 	if (len == 0) {
6077 		/* We don't log zero window probes */
6078 		return;
6079 	}
6080 	pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1);
6081 	/* First question is it a retransmission? */
6082 	if (seq_out == snd_max) {
6083 again:
6084 		rsm = bbr_alloc(bbr);
6085 		if (rsm == NULL) {
6086 			return;
6087 		}
6088 		rsm->r_flags = 0;
6089 		if (th_flags & TH_SYN)
6090 			rsm->r_flags |= BBR_HAS_SYN;
6091 		if (th_flags & TH_FIN)
6092 			rsm->r_flags |= BBR_HAS_FIN;
6093 		rsm->r_tim_lastsent[0] = cts;
6094 		rsm->r_rtr_cnt = 1;
6095 		rsm->r_rtr_bytes = 0;
6096 		rsm->r_start = seq_out;
6097 		rsm->r_end = rsm->r_start + len;
6098 		rsm->r_dupack = 0;
6099 		rsm->r_delivered = bbr->r_ctl.rc_delivered;
6100 		rsm->r_pacing_delay = pacing_time;
6101 		rsm->r_ts_valid = bbr->rc_ts_valid;
6102 		if (bbr->rc_ts_valid)
6103 			rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
6104 		rsm->r_del_time = bbr->r_ctl.rc_del_time;
6105 		if (bbr->r_ctl.r_app_limited_until)
6106 			rsm->r_app_limited = 1;
6107 		else
6108 			rsm->r_app_limited = 0;
6109 		rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
6110 		rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
6111 						(bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
6112 		/*
6113 		 * Here we must also add in this rsm since snd_max
6114 		 * is updated after we return from a new send.
6115 		 */
6116 		rsm->r_flight_at_send += len;
6117 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
6118 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
6119 		rsm->r_in_tmap = 1;
6120 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
6121 			rsm->r_bbr_state = bbr_state_val(bbr);
6122 		else
6123 			rsm->r_bbr_state = 8;
6124 		if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
6125 			rsm->r_is_gain = 1;
6126 			rsm->r_is_drain = 0;
6127 		} else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
6128 			rsm->r_is_drain = 1;
6129 			rsm->r_is_gain = 0;
6130 		} else {
6131 			rsm->r_is_drain = 0;
6132 			rsm->r_is_gain = 0;
6133 		}
6134 		return;
6135 	}
6136 	/*
6137 	 * If we reach here its a retransmission and we need to find it.
6138 	 */
6139 more:
6140 	if (hintrsm && (hintrsm->r_start == seq_out)) {
6141 		rsm = hintrsm;
6142 		hintrsm = NULL;
6143 	} else if (bbr->r_ctl.rc_next) {
6144 		/* We have a hint from a previous run */
6145 		rsm = bbr->r_ctl.rc_next;
6146 	} else {
6147 		/* No hints sorry */
6148 		rsm = NULL;
6149 	}
6150 	if ((rsm) && (rsm->r_start == seq_out)) {
6151 		/*
6152 		 * We used rc_next or hintrsm  to retransmit, hopefully the
6153 		 * likely case.
6154 		 */
6155 		seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6156 		if (len == 0) {
6157 			return;
6158 		} else {
6159 			goto more;
6160 		}
6161 	}
6162 	/* Ok it was not the last pointer go through it the hard way. */
6163 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6164 		if (rsm->r_start == seq_out) {
6165 			seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6166 			bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
6167 			if (len == 0) {
6168 				return;
6169 			} else {
6170 				continue;
6171 			}
6172 		}
6173 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6174 			/* Transmitted within this piece */
6175 			/*
6176 			 * Ok we must split off the front and then let the
6177 			 * update do the rest
6178 			 */
6179 			nrsm = bbr_alloc_full_limit(bbr);
6180 			if (nrsm == NULL) {
6181 				bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
6182 				return;
6183 			}
6184 			/*
6185 			 * copy rsm to nrsm and then trim the front of rsm
6186 			 * to not include this part.
6187 			 */
6188 			bbr_clone_rsm(bbr, nrsm, rsm, seq_out);
6189 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
6190 			if (rsm->r_in_tmap) {
6191 				TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6192 				nrsm->r_in_tmap = 1;
6193 			}
6194 			rsm->r_flags &= (~BBR_HAS_FIN);
6195 			seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time);
6196 			if (len == 0) {
6197 				return;
6198 			}
6199 		}
6200 	}
6201 	/*
6202 	 * Hmm not found in map did they retransmit both old and on into the
6203 	 * new?
6204 	 */
6205 	if (seq_out == tp->snd_max) {
6206 		goto again;
6207 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
6208 #ifdef BBR_INVARIANTS
6209 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6210 		    seq_out, len, tp->snd_una, tp->snd_max);
6211 		printf("Starting Dump of all rack entries\n");
6212 		TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6213 			printf("rsm:%p start:%u end:%u\n",
6214 			    rsm, rsm->r_start, rsm->r_end);
6215 		}
6216 		printf("Dump complete\n");
6217 		panic("seq_out not found rack:%p tp:%p",
6218 		    bbr, tp);
6219 #endif
6220 	} else {
6221 #ifdef BBR_INVARIANTS
6222 		/*
6223 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
6224 		 * flag)
6225 		 */
6226 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6227 		    seq_out, len, tp->snd_max, tp);
6228 #endif
6229 	}
6230 }
6231 
6232 static void
6233 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt)
6234 {
6235 	/*
6236 	 * Collapse timeout back the cum-ack moved.
6237 	 */
6238 	tp->t_rxtshift = 0;
6239 	tp->t_softerror = 0;
6240 }
6241 
6242 
6243 static void
6244 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin)
6245 {
6246 	bbr->rtt_valid = 1;
6247 	bbr->r_ctl.cur_rtt = rtt_usecs;
6248 	bbr->r_ctl.ts_in = tsin;
6249 	if (rsm_send_time)
6250 		bbr->r_ctl.cur_rtt_send_time = rsm_send_time;
6251 }
6252 
6253 static void
6254 bbr_make_timestamp_determination(struct tcp_bbr *bbr)
6255 {
6256 	/**
6257 	 * We have in our bbr control:
6258 	 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp).
6259 	 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts).
6260 	 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts)
6261 	 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time)
6262 	 *
6263 	 * Now we can calculate the time between the sends by doing:
6264 	 *
6265 	 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts
6266 	 *
6267 	 * And the peer's time between receiving them by doing:
6268 	 *
6269 	 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp
6270 	 *
6271 	 * We want to figure out if the timestamp values are in msec, 10msec or usec.
6272 	 * We also may find that we can't use the timestamps if say we see
6273 	 * that the peer_delta indicates that though we may have taken 10ms to
6274 	 * pace out the data, it only saw 1ms between the two packets. This would
6275 	 * indicate that somewhere on the path is a batching entity that is giving
6276 	 * out time-slices of the actual b/w. This would mean we could not use
6277 	 * reliably the peers timestamps.
6278 	 *
6279 	 * We expect delta > peer_delta initially. Until we figure out the
6280 	 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio.
6281 	 * If we place 1000 there then its a ms vs our usec. If we place 10000 there
6282 	 * then its 10ms vs our usec. If the peer is running a usec clock we would
6283 	 * put a 1 there. If the value is faster then ours, we will disable the
6284 	 * use of timestamps (though we could revist this later if we find it to be not
6285 	 * just an isolated one or two flows)).
6286 	 *
6287 	 * To detect the batching middle boxes we will come up with our compensation and
6288 	 * if with it in place, we find the peer is drastically off (by some margin) in
6289 	 * the smaller direction, then we will assume the worst case and disable use of timestamps.
6290 	 *
6291 	 */
6292 	uint64_t delta, peer_delta, delta_up;
6293 
6294 	delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts;
6295 	if (delta < bbr_min_usec_delta) {
6296 		/*
6297 		 * Have not seen a min amount of time
6298 		 * between our send times so we can
6299 		 * make a determination of the timestamp
6300 		 * yet.
6301 		 */
6302 		return;
6303 	}
6304 	peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp;
6305 	if (peer_delta < bbr_min_peer_delta) {
6306 		/*
6307 		 * We may have enough in the form of
6308 		 * our delta but the peers number
6309 		 * has not changed that much. It could
6310 		 * be its clock ratio is such that
6311 		 * we need more data (10ms tick) or
6312 		 * there may be other compression scenarios
6313 		 * going on. In any event we need the
6314 		 * spread to be larger.
6315 		 */
6316 		return;
6317 	}
6318 	/* Ok lets first see which way our delta is going */
6319 	if (peer_delta > delta) {
6320 		/* Very unlikely, the peer without
6321 		 * compensation shows that it saw
6322 		 * the two sends arrive further apart
6323 		 * then we saw then in micro-seconds.
6324 		 */
6325 		if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) {
6326 			/* well it looks like the peer is a micro-second clock. */
6327 			bbr->rc_ts_clock_set = 1;
6328 			bbr->r_ctl.bbr_peer_tsratio = 1;
6329 		} else {
6330 			bbr->rc_ts_cant_be_used = 1;
6331 			bbr->rc_ts_clock_set = 1;
6332 		}
6333 		return;
6334 	}
6335 	/* Ok we know that the peer_delta is smaller than our send distance */
6336 	bbr->rc_ts_clock_set = 1;
6337 	/* First question is it within the percentage that they are using usec time? */
6338 	delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent;
6339 	if ((peer_delta + delta_up) >= delta) {
6340 		/* Its a usec clock */
6341 		bbr->r_ctl.bbr_peer_tsratio = 1;
6342 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6343 		return;
6344 	}
6345 	/* Ok if not usec, what about 10usec (though unlikely)? */
6346 	delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent;
6347 	if (((peer_delta * 10) + delta_up) >= delta) {
6348 		bbr->r_ctl.bbr_peer_tsratio = 10;
6349 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6350 		return;
6351 	}
6352 	/* And what about 100usec (though again unlikely)? */
6353 	delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent;
6354 	if (((peer_delta * 100) + delta_up) >= delta) {
6355 		bbr->r_ctl.bbr_peer_tsratio = 100;
6356 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6357 		return;
6358 	}
6359 	/* And how about 1 msec (the most likely one)? */
6360 	delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent;
6361 	if (((peer_delta * 1000) + delta_up) >= delta) {
6362 		bbr->r_ctl.bbr_peer_tsratio = 1000;
6363 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6364 		return;
6365 	}
6366 	/* Ok if not msec could it be 10 msec? */
6367 	delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent;
6368 	if (((peer_delta * 10000) + delta_up) >= delta) {
6369 		bbr->r_ctl.bbr_peer_tsratio = 10000;
6370 		return;
6371 	}
6372 	/* If we fall down here the clock tick so slowly we can't use it */
6373 	bbr->rc_ts_cant_be_used = 1;
6374 	bbr->r_ctl.bbr_peer_tsratio = 0;
6375 	bbr_log_tstmp_validation(bbr, peer_delta, delta);
6376 }
6377 
6378 /*
6379  * Collect new round-trip time estimate
6380  * and update averages and current timeout.
6381  */
6382 static void
6383 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts)
6384 {
6385 	int32_t delta;
6386 	uint32_t rtt, tsin;
6387 	int32_t rtt_ticks;
6388 
6389 
6390 	if (bbr->rtt_valid == 0)
6391 		/* No valid sample */
6392 		return;
6393 
6394 	rtt = bbr->r_ctl.cur_rtt;
6395 	tsin = bbr->r_ctl.ts_in;
6396 	if (bbr->rc_prtt_set_ts) {
6397 		/*
6398 		 * We are to force feed the rttProp filter due
6399 		 * to an entry into PROBE_RTT. This assures
6400 		 * that the times are sync'd between when we
6401 		 * go into PROBE_RTT and the filter expiration.
6402 		 *
6403 		 * Google does not use a true filter, so they do
6404 		 * this implicitly since they only keep one value
6405 		 * and when they enter probe-rtt they update the
6406 		 * value to the newest rtt.
6407 		 */
6408 		uint32_t rtt_prop;
6409 
6410 		bbr->rc_prtt_set_ts = 0;
6411 		rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
6412 		if (rtt > rtt_prop)
6413 			filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts);
6414 		else
6415 			apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6416 	}
6417 	if (bbr->rc_ack_was_delayed)
6418 		rtt += bbr->r_ctl.rc_ack_hdwr_delay;
6419 
6420 	if (rtt < bbr->r_ctl.rc_lowest_rtt)
6421 		bbr->r_ctl.rc_lowest_rtt = rtt;
6422 	bbr_log_rtt_sample(bbr, rtt, tsin);
6423 	if (bbr->r_init_rtt) {
6424 		/*
6425 		 * The initial rtt is not-trusted, nuke it and lets get
6426 		 * our first valid measurement in.
6427 		 */
6428 		bbr->r_init_rtt = 0;
6429 		tp->t_srtt = 0;
6430 	}
6431 	if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) {
6432 		/*
6433 		 * So we have not yet figured out
6434 		 * what the peers TSTMP value is
6435 		 * in (most likely ms). We need a
6436 		 * series of cum-ack's to determine
6437 		 * this reliably.
6438 		 */
6439 		if (bbr->rc_ack_is_cumack) {
6440 			if (bbr->rc_ts_data_set) {
6441 				/* Lets attempt to determine the timestamp granularity. */
6442 				bbr_make_timestamp_determination(bbr);
6443 			} else {
6444 				bbr->rc_ts_data_set = 1;
6445 				bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts;
6446 				bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time;
6447 			}
6448 		} else {
6449 			/*
6450 			 * We have to have consecutive acks
6451 			 * reset any "filled" state to none.
6452 			 */
6453 			bbr->rc_ts_data_set = 0;
6454 		}
6455 	}
6456 	/* Round it up */
6457 	rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1)));
6458 	if (rtt_ticks == 0)
6459 		rtt_ticks = 1;
6460 	if (tp->t_srtt != 0) {
6461 		/*
6462 		 * srtt is stored as fixed point with 5 bits after the
6463 		 * binary point (i.e., scaled by 8).  The following magic is
6464 		 * equivalent to the smoothing algorithm in rfc793 with an
6465 		 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6466 		 * Adjust rtt to origin 0.
6467 		 */
6468 
6469 		delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT)
6470 		    - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6471 
6472 		tp->t_srtt += delta;
6473 		if (tp->t_srtt <= 0)
6474 			tp->t_srtt = 1;
6475 
6476 		/*
6477 		 * We accumulate a smoothed rtt variance (actually, a
6478 		 * smoothed mean difference), then set the retransmit timer
6479 		 * to smoothed rtt + 4 times the smoothed variance. rttvar
6480 		 * is stored as fixed point with 4 bits after the binary
6481 		 * point (scaled by 16).  The following is equivalent to
6482 		 * rfc793 smoothing with an alpha of .75 (rttvar =
6483 		 * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
6484 		 * wired-in beta.
6485 		 */
6486 		if (delta < 0)
6487 			delta = -delta;
6488 		delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6489 		tp->t_rttvar += delta;
6490 		if (tp->t_rttvar <= 0)
6491 			tp->t_rttvar = 1;
6492 		if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
6493 			tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6494 	} else {
6495 		/*
6496 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
6497 		 * variance to half the rtt (so our first retransmit happens
6498 		 * at 3*rtt).
6499 		 */
6500 		tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT;
6501 		tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1);
6502 		tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6503 	}
6504 	KMOD_TCPSTAT_INC(tcps_rttupdated);
6505 	tp->t_rttupdated++;
6506 #ifdef STATS
6507 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks));
6508 #endif
6509 	/*
6510 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
6511 	 * way we do the smoothing, srtt and rttvar will each average +1/2
6512 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
6513 	 * tick of rounding and 1 extra tick because of +-1/2 tick
6514 	 * uncertainty in the firing of the timer.  The bias will give us
6515 	 * exactly the 1.5 tick we need.  But, because the bias is
6516 	 * statistical, we have to test that we don't drop below the minimum
6517 	 * feasible timer (which is 2 ticks).
6518 	 */
6519 	TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6520 	    max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2),
6521 	    MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
6522 
6523 	/*
6524 	 * We received an ack for a packet that wasn't retransmitted; it is
6525 	 * probably safe to discard any error indications we've received
6526 	 * recently.  This isn't quite right, but close enough for now (a
6527 	 * route might have failed after we sent a segment, and the return
6528 	 * path might not be symmetrical).
6529 	 */
6530 	tp->t_softerror = 0;
6531 	rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
6532 	if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt)
6533 		bbr->r_ctl.bbr_smallest_srtt_this_state = rtt;
6534 }
6535 
6536 static void
6537 bbr_earlier_retran(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm,
6538 		   uint32_t t, uint32_t cts, int ack_type)
6539 {
6540 	/*
6541 	 * For this RSM, we acknowledged the data from a previous
6542 	 * transmission, not the last one we made. This means we did a false
6543 	 * retransmit.
6544 	 */
6545 	if (rsm->r_flags & BBR_HAS_FIN) {
6546 		/*
6547 		 * The sending of the FIN often is multiple sent when we
6548 		 * have everything outstanding ack'd. We ignore this case
6549 		 * since its over now.
6550 		 */
6551 		return;
6552 	}
6553 	if (rsm->r_flags & BBR_TLP) {
6554 		/*
6555 		 * We expect TLP's to have this occur often
6556 		 */
6557 		bbr->rc_tlp_rtx_out = 0;
6558 		return;
6559 	}
6560 	if (ack_type != BBR_CUM_ACKED) {
6561 		/*
6562 		 * If it was not a cum-ack we
6563 		 * don't really know for sure since
6564 		 * the timestamp could be from some
6565 		 * other transmission.
6566 		 */
6567 		return;
6568 	}
6569 
6570 	if (rsm->r_flags & BBR_WAS_SACKPASS) {
6571 		/*
6572 		 * We retransmitted based on a sack and the earlier
6573 		 * retransmission ack'd it - re-ordering is occuring.
6574 		 */
6575 		BBR_STAT_INC(bbr_reorder_seen);
6576 		bbr->r_ctl.rc_reorder_ts = cts;
6577 	}
6578 	/* Back down the loss count */
6579 	if (rsm->r_flags & BBR_MARKED_LOST) {
6580 		bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
6581 		bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
6582 		rsm->r_flags &= ~BBR_MARKED_LOST;
6583 		if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
6584 			/* LT sampling also needs adjustment */
6585 			bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
6586 	}
6587 	/***** RRS HERE ************************/
6588 	/* Do we need to do this???            */
6589 	/* bbr_reset_lt_bw_sampling(bbr, cts); */
6590 	/***** RRS HERE ************************/
6591 	BBR_STAT_INC(bbr_badfr);
6592 	BBR_STAT_ADD(bbr_badfr_bytes, (rsm->r_end - rsm->r_start));
6593 }
6594 
6595 
6596 static void
6597 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line)
6598 {
6599 	bbr->r_ctl.rc_rtt_shrinks = cts;
6600 	if (bbr_can_force_probertt &&
6601 	    (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
6602 	    ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
6603 		/*
6604 		 * We should enter probe-rtt its been too long
6605 		 * since we have been there.
6606 		 */
6607 		bbr_enter_probe_rtt(bbr, cts, __LINE__);
6608 	} else
6609 		bbr_check_probe_rtt_limits(bbr, cts);
6610 }
6611 
6612 static void
6613 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts)
6614 {
6615 	uint64_t orig_bw;
6616 
6617 	if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) {
6618 		/* We never apply a zero measurment */
6619 		bbr_log_type_bbrupd(bbr, 20, cts, 0, 0,
6620 				    0, 0, 0, 0, 0, 0);
6621 		return;
6622 	}
6623 	if (bbr->r_ctl.r_measurement_count < 0xffffffff)
6624 		bbr->r_ctl.r_measurement_count++;
6625 	orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
6626 	apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch);
6627 	bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw,
6628 			    (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate),
6629 			    0, 0, 0, 0, 0, 0);
6630 	if (orig_bw &&
6631 	    (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) {
6632 		if (bbr->bbr_hdrw_pacing) {
6633 			/*
6634 			 * Apply a new rate to the hardware
6635 			 * possibly.
6636 			 */
6637 			bbr_update_hardware_pacing_rate(bbr, cts);
6638 		}
6639 		bbr_set_state_target(bbr, __LINE__);
6640 		tcp_bbr_tso_size_check(bbr, cts);
6641 		if (bbr->r_recovery_bw)  {
6642 			bbr_setup_red_bw(bbr, cts);
6643 			bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW);
6644 		}
6645 	} else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate))
6646 		tcp_bbr_tso_size_check(bbr, cts);
6647 }
6648 
6649 static void
6650 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6651 {
6652 	if (bbr->rc_in_persist == 0) {
6653 		/* We log only when not in persist */
6654 		/* Translate to a Bytes Per Second */
6655 		uint64_t tim, bw, ts_diff, ts_bw;
6656 		uint32_t upper, lower, delivered;
6657 
6658 		if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6659 			tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6660 		else
6661 			tim = 1;
6662 		/*
6663 		 * Now that we have processed the tim (skipping the sample
6664 		 * or possibly updating the time, go ahead and
6665 		 * calculate the cdr.
6666 		 */
6667 		delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6668 		bw = (uint64_t)delivered;
6669 		bw *= (uint64_t)USECS_IN_SECOND;
6670 		bw /= tim;
6671 		if (bw == 0) {
6672 			/* We must have a calculatable amount */
6673 			return;
6674 		}
6675 		upper = (bw >> 32) & 0x00000000ffffffff;
6676 		lower = bw & 0x00000000ffffffff;
6677 		/*
6678 		 * If we are using this b/w shove it in now so we
6679 		 * can see in the trace viewer if it gets over-ridden.
6680 		 */
6681 		if (rsm->r_ts_valid &&
6682 		    bbr->rc_ts_valid &&
6683 		    bbr->rc_ts_clock_set &&
6684 		    (bbr->rc_ts_cant_be_used == 0) &&
6685 		    bbr->rc_use_ts_limit) {
6686 			ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1);
6687 			ts_diff *= bbr->r_ctl.bbr_peer_tsratio;
6688 			if ((delivered == 0) ||
6689 			    (rtt < 1000)) {
6690 				/* Can't use the ts */
6691 				bbr_log_type_bbrupd(bbr, 61, cts,
6692 						    ts_diff,
6693 						    bbr->r_ctl.last_inbound_ts,
6694 						    rsm->r_del_ack_ts, 0,
6695 						    0, 0, 0, delivered);
6696 			} else {
6697 				ts_bw = (uint64_t)delivered;
6698 				ts_bw *= (uint64_t)USECS_IN_SECOND;
6699 				ts_bw /= ts_diff;
6700 				bbr_log_type_bbrupd(bbr, 62, cts,
6701 						    (ts_bw >> 32),
6702 						    (ts_bw & 0xffffffff), 0, 0,
6703 						    0, 0, ts_diff, delivered);
6704 				if ((bbr->ts_can_raise) &&
6705 				    (ts_bw > bw)) {
6706 					bbr_log_type_bbrupd(bbr, 8, cts,
6707 							    delivered,
6708 							    ts_diff,
6709 							    (bw >> 32),
6710 							    (bw & 0x00000000ffffffff),
6711 							    0, 0, 0, 0);
6712 					bw = ts_bw;
6713 				} else if (ts_bw && (ts_bw < bw)) {
6714 					bbr_log_type_bbrupd(bbr, 7, cts,
6715 							    delivered,
6716 							    ts_diff,
6717 							    (bw >> 32),
6718 							    (bw & 0x00000000ffffffff),
6719 							    0, 0, 0, 0);
6720 					bw = ts_bw;
6721 				}
6722 			}
6723 		}
6724 		if (rsm->r_first_sent_time &&
6725 		    TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6726 			uint64_t sbw, sti;
6727 			/*
6728 			 * We use what was in flight at the time of our
6729 			 * send  and the size of this send to figure
6730 			 * out what we have been sending at (amount).
6731 			 * For the time we take from the time of
6732 			 * the send of the first send outstanding
6733 			 * until this send plus this sends pacing
6734 			 * time. This gives us a good calculation
6735 			 * as to the rate we have been sending at.
6736 			 */
6737 
6738 			sbw = (uint64_t)(rsm->r_flight_at_send);
6739 			sbw *= (uint64_t)USECS_IN_SECOND;
6740 			sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6741 			sti += rsm->r_pacing_delay;
6742 			sbw /= sti;
6743 			if (sbw < bw) {
6744 				bbr_log_type_bbrupd(bbr, 6, cts,
6745 						    delivered,
6746 						    (uint32_t)sti,
6747 						    (bw >> 32),
6748 						    (uint32_t)bw,
6749 						    rsm->r_first_sent_time, 0, (sbw >> 32),
6750 						    (uint32_t)sbw);
6751 				bw = sbw;
6752 			}
6753 		}
6754 		/* Use the google algorithm for b/w measurements */
6755 		bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6756 		if ((rsm->r_app_limited == 0) ||
6757 		    (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) {
6758 			tcp_bbr_commit_bw(bbr, cts);
6759 			bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6760 					    0, 0, 0, 0,  bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6761 		}
6762 	}
6763 }
6764 
6765 static void
6766 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6767 {
6768 	if (bbr->rc_in_persist == 0) {
6769 		/* We log only when not in persist */
6770 		/* Translate to a Bytes Per Second */
6771 		uint64_t tim, bw;
6772 		uint32_t upper, lower, delivered;
6773 		int no_apply = 0;
6774 
6775 		if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6776 			tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6777 		else
6778 			tim = 1;
6779 		/*
6780 		 * Now that we have processed the tim (skipping the sample
6781 		 * or possibly updating the time, go ahead and
6782 		 * calculate the cdr.
6783 		 */
6784 		delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6785 		bw = (uint64_t)delivered;
6786 		bw *= (uint64_t)USECS_IN_SECOND;
6787 		bw /= tim;
6788 		if (tim < bbr->r_ctl.rc_lowest_rtt) {
6789 			bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6790 					    tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6791 
6792 			no_apply = 1;
6793 		}
6794 		upper = (bw >> 32) & 0x00000000ffffffff;
6795 		lower = bw & 0x00000000ffffffff;
6796 		/*
6797 		 * If we are using this b/w shove it in now so we
6798 		 * can see in the trace viewer if it gets over-ridden.
6799 		 */
6800 		bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6801 		/* Gate by the sending rate */
6802 		if (rsm->r_first_sent_time &&
6803 		    TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6804 			uint64_t sbw, sti;
6805 			/*
6806 			 * We use what was in flight at the time of our
6807 			 * send  and the size of this send to figure
6808 			 * out what we have been sending at (amount).
6809 			 * For the time we take from the time of
6810 			 * the send of the first send outstanding
6811 			 * until this send plus this sends pacing
6812 			 * time. This gives us a good calculation
6813 			 * as to the rate we have been sending at.
6814 			 */
6815 
6816 			sbw = (uint64_t)(rsm->r_flight_at_send);
6817 			sbw *= (uint64_t)USECS_IN_SECOND;
6818 			sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6819 			sti += rsm->r_pacing_delay;
6820 			sbw /= sti;
6821 			if (sbw < bw) {
6822 				bbr_log_type_bbrupd(bbr, 6, cts,
6823 						    delivered,
6824 						    (uint32_t)sti,
6825 						    (bw >> 32),
6826 						    (uint32_t)bw,
6827 						    rsm->r_first_sent_time, 0, (sbw >> 32),
6828 						    (uint32_t)sbw);
6829 				bw = sbw;
6830 			}
6831 			if ((sti > tim) &&
6832 			    (sti < bbr->r_ctl.rc_lowest_rtt)) {
6833 				bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6834 						    (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6835 				no_apply = 1;
6836 			} else
6837 				no_apply = 0;
6838 		}
6839 		bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6840 		if ((no_apply == 0) &&
6841 		    ((rsm->r_app_limited == 0) ||
6842 		     (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) {
6843 			tcp_bbr_commit_bw(bbr, cts);
6844 			bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6845 					    0, 0, 0, 0, bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6846 		}
6847 	}
6848 }
6849 
6850 
6851 static void
6852 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin,
6853     uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to)
6854 {
6855 	uint64_t old_rttprop;
6856 
6857 	/* Update our delivery time and amount */
6858 	bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start);
6859 	bbr->r_ctl.rc_del_time = cts;
6860 	if (rtt == 0) {
6861 		/*
6862 		 * 0 means its a retransmit, for now we don't use these for
6863 		 * the rest of BBR.
6864 		 */
6865 		return;
6866 	}
6867 	if ((bbr->rc_use_google == 0) &&
6868 	    (match != BBR_RTT_BY_EXACTMATCH) &&
6869 	    (match != BBR_RTT_BY_TIMESTAMP)){
6870 		/*
6871 		 * We get a lot of rtt updates, lets not pay attention to
6872 		 * any that are not an exact match. That way we don't have
6873 		 * to worry about timestamps and the whole nonsense of
6874 		 * unsure if its a retransmission etc (if we ever had the
6875 		 * timestamp fixed to always have the last thing sent this
6876 		 * would not be a issue).
6877 		 */
6878 		return;
6879 	}
6880 	if ((bbr_no_retran && bbr->rc_use_google) &&
6881 	    (match != BBR_RTT_BY_EXACTMATCH) &&
6882 	    (match != BBR_RTT_BY_TIMESTAMP)){
6883 		/*
6884 		 * We only do measurements in google mode
6885 		 * with bbr_no_retran on for sure things.
6886 		 */
6887 		return;
6888 	}
6889 	/* Only update srtt if we know by exact match */
6890 	tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin);
6891 	if (ack_type == BBR_CUM_ACKED)
6892 		bbr->rc_ack_is_cumack = 1;
6893 	else
6894 		bbr->rc_ack_is_cumack = 0;
6895 	old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP);
6896         /*
6897 	 * Note the following code differs to the original
6898 	 * BBR spec. It calls for <= not <. However after a
6899 	 * long discussion in email with Neal, he acknowledged
6900 	 * that it should be < than so that we will have flows
6901 	 * going into probe-rtt (we were seeing cases where that
6902 	 * did not happen and caused ugly things to occur). We
6903 	 * have added this agreed upon fix to our code base.
6904 	 */
6905 	if (rtt < old_rttprop) {
6906 		/* Update when we last saw a rtt drop */
6907 		bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0);
6908 		bbr_set_reduced_rtt(bbr, cts, __LINE__);
6909 	}
6910 	bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts,
6911 	    match, rsm->r_start, rsm->r_flags);
6912 	apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6913 	if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) {
6914 		/*
6915 		 * The RTT-prop moved, reset the target (may be a
6916 		 * nop for some states).
6917 		 */
6918 		bbr_set_state_target(bbr, __LINE__);
6919 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)
6920 			bbr_log_rtt_shrinks(bbr, cts, 0, 0,
6921 					    __LINE__, BBR_RTTS_NEW_TARGET, 0);
6922 		else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP))
6923 			/* It went up */
6924 			bbr_check_probe_rtt_limits(bbr, cts);
6925 	}
6926 	if ((bbr->rc_use_google == 0) &&
6927 	    (match == BBR_RTT_BY_TIMESTAMP)) {
6928 		/*
6929 		 * We don't do b/w update with
6930 		 * these since they are not really
6931 		 * reliable.
6932 		 */
6933 		return;
6934 	}
6935 	if (bbr->r_ctl.r_app_limited_until &&
6936 	    (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) {
6937 		/* We are no longer app-limited */
6938 		bbr->r_ctl.r_app_limited_until = 0;
6939 	}
6940 	if (bbr->rc_use_google) {
6941 		bbr_google_measurement(bbr, rsm, rtt, cts);
6942 	} else {
6943 		bbr_nf_measurement(bbr, rsm, rtt, cts);
6944 	}
6945 }
6946 
6947 /*
6948  * Convert a timestamp that the main stack
6949  * uses (milliseconds) into one that bbr uses
6950  * (microseconds). Return that converted timestamp.
6951  */
6952 static uint32_t
6953 bbr_ts_convert(uint32_t cts) {
6954 	uint32_t sec, msec;
6955 
6956 	sec = cts / MS_IN_USEC;
6957 	msec = cts - (MS_IN_USEC * sec);
6958 	return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC));
6959 }
6960 
6961 /*
6962  * Return 0 if we did not update the RTT time, return
6963  * 1 if we did.
6964  */
6965 static int
6966 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr,
6967     struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack)
6968 {
6969 	int32_t i;
6970 	uint32_t t, uts = 0;
6971 
6972 	if ((rsm->r_flags & BBR_ACKED) ||
6973 	    (rsm->r_flags & BBR_WAS_RENEGED) ||
6974 	    (rsm->r_flags & BBR_RXT_CLEARED)) {
6975 		/* Already done */
6976 		return (0);
6977 	}
6978 	if (rsm->r_rtr_cnt == 1) {
6979 		/*
6980 		 * Only one transmit. Hopefully the normal case.
6981 		 */
6982 		if (TSTMP_GT(cts, rsm->r_tim_lastsent[0]))
6983 			t = cts - rsm->r_tim_lastsent[0];
6984 		else
6985 			t = 1;
6986 		if ((int)t <= 0)
6987 			t = 1;
6988 		bbr->r_ctl.rc_last_rtt = t;
6989 		bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6990 				    BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to);
6991 		return (1);
6992 	}
6993 	/* Convert to usecs */
6994 	if ((bbr_can_use_ts_for_rtt == 1) &&
6995 	    (bbr->rc_use_google == 1) &&
6996 	    (ack_type == BBR_CUM_ACKED) &&
6997 	    (to->to_flags & TOF_TS) &&
6998 	    (to->to_tsecr != 0)) {
6999 
7000 		t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr;
7001 		if (t < 1)
7002 			t = 1;
7003 		t *= MS_IN_USEC;
7004 		bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
7005 				    BBR_RTT_BY_TIMESTAMP,
7006 				    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)],
7007 				    ack_type, to);
7008 		return (1);
7009 	}
7010 	uts = bbr_ts_convert(to->to_tsecr);
7011 	if ((to->to_flags & TOF_TS) &&
7012 	    (to->to_tsecr != 0) &&
7013 	    (ack_type == BBR_CUM_ACKED) &&
7014 	    ((rsm->r_flags & BBR_OVERMAX) == 0)) {
7015 		/*
7016 		 * Now which timestamp does it match? In this block the ACK
7017 		 * may be coming from a previous transmission.
7018 		 */
7019 		uint32_t fudge;
7020 
7021 		fudge = BBR_TIMER_FUDGE;
7022 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
7023 			if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) &&
7024 			    (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) {
7025 				if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
7026 					t = cts - rsm->r_tim_lastsent[i];
7027 				else
7028 					t = 1;
7029 				if ((int)t <= 0)
7030 					t = 1;
7031 				bbr->r_ctl.rc_last_rtt = t;
7032 				bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING,
7033 						    rsm->r_tim_lastsent[i], ack_type, to);
7034 				if ((i + 1) < rsm->r_rtr_cnt) {
7035 					/* Likely */
7036 					bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type);
7037 				} else if (rsm->r_flags & BBR_TLP) {
7038 					bbr->rc_tlp_rtx_out = 0;
7039 				}
7040 				return (1);
7041 			}
7042 		}
7043 		/* Fall through if we can't find a matching timestamp */
7044 	}
7045 	/*
7046 	 * Ok its a SACK block that we retransmitted. or a windows
7047 	 * machine without timestamps. We can tell nothing from the
7048 	 * time-stamp since its not there or the time the peer last
7049 	 * recieved a segment that moved forward its cum-ack point.
7050 	 *
7051 	 * Lets look at the last retransmit and see what we can tell
7052 	 * (with BBR for space we only keep 2 note we have to keep
7053 	 * at least 2 so the map can not be condensed more).
7054 	 */
7055 	i = rsm->r_rtr_cnt - 1;
7056 	if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
7057 		t = cts - rsm->r_tim_lastsent[i];
7058 	else
7059 		goto not_sure;
7060 	if (t < bbr->r_ctl.rc_lowest_rtt) {
7061 		/*
7062 		 * We retransmitted and the ack came back in less
7063 		 * than the smallest rtt we have observed in the
7064 		 * windowed rtt. We most likey did an improper
7065 		 * retransmit as outlined in 4.2 Step 3 point 2 in
7066 		 * the rack-draft.
7067 		 *
7068 		 * Use the prior transmission to update all the
7069 		 * information as long as there is only one prior
7070 		 * transmission.
7071 		 */
7072 		if ((rsm->r_flags & BBR_OVERMAX) == 0) {
7073 #ifdef BBR_INVARIANTS
7074 			if (rsm->r_rtr_cnt == 1)
7075 				panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags);
7076 #endif
7077 			i = rsm->r_rtr_cnt - 2;
7078 			if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
7079 				t = cts - rsm->r_tim_lastsent[i];
7080 			else
7081 				t = 1;
7082 			bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET,
7083 					    rsm->r_tim_lastsent[i], ack_type, to);
7084 			bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type);
7085 		} else {
7086 			/*
7087 			 * Too many prior transmissions, just
7088 			 * updated BBR delivered
7089 			 */
7090 not_sure:
7091 			bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
7092 					    BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
7093 		}
7094 	} else {
7095 		/*
7096 		 * We retransmitted it and the retransmit did the
7097 		 * job.
7098 		 */
7099 		if (rsm->r_flags & BBR_TLP)
7100 			bbr->rc_tlp_rtx_out = 0;
7101 		if ((rsm->r_flags & BBR_OVERMAX) == 0)
7102 			bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts,
7103 					    BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to);
7104 		else
7105 			bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
7106 					    BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
7107 		return (1);
7108 	}
7109 	return (0);
7110 }
7111 
7112 /*
7113  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7114  */
7115 static void
7116 bbr_log_sack_passed(struct tcpcb *tp,
7117     struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
7118 {
7119 	struct bbr_sendmap *nrsm;
7120 
7121 	nrsm = rsm;
7122 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap,
7123 	    bbr_head, r_tnext) {
7124 		if (nrsm == rsm) {
7125 			/* Skip orginal segment he is acked */
7126 			continue;
7127 		}
7128 		if (nrsm->r_flags & BBR_ACKED) {
7129 			/* Skip ack'd segments */
7130 			continue;
7131 		}
7132 		if (nrsm->r_flags & BBR_SACK_PASSED) {
7133 			/*
7134 			 * We found one that is already marked
7135 			 * passed, we have been here before and
7136 			 * so all others below this are marked.
7137 			 */
7138 			break;
7139 		}
7140 		BBR_STAT_INC(bbr_sack_passed);
7141 		nrsm->r_flags |= BBR_SACK_PASSED;
7142 		if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) &&
7143 		    bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) {
7144 			bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start;
7145 			bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start;
7146 			nrsm->r_flags |= BBR_MARKED_LOST;
7147 		}
7148 		nrsm->r_flags &= ~BBR_WAS_SACKPASS;
7149 	}
7150 }
7151 
7152 /*
7153  * Returns the number of bytes that were
7154  * newly ack'd by sack blocks.
7155  */
7156 static uint32_t
7157 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack,
7158     struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts)
7159 {
7160 	int32_t times = 0;
7161 	uint32_t start, end, maxseg, changed = 0;
7162 	struct bbr_sendmap *rsm, *nrsm;
7163 	int32_t used_ref = 1;
7164 	uint8_t went_back = 0, went_fwd = 0;
7165 
7166 	maxseg = tp->t_maxseg - bbr->rc_last_options;
7167 	start = sack->start;
7168 	end = sack->end;
7169 	rsm = *prsm;
7170 	if (rsm == NULL)
7171 		used_ref = 0;
7172 
7173 	/* Do we locate the block behind where we last were? */
7174 	if (rsm && SEQ_LT(start, rsm->r_start)) {
7175 		went_back = 1;
7176 		TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
7177 			if (SEQ_GEQ(start, rsm->r_start) &&
7178 			    SEQ_LT(start, rsm->r_end)) {
7179 				goto do_rest_ofb;
7180 			}
7181 		}
7182 	}
7183 start_at_beginning:
7184 	went_fwd = 1;
7185 	/*
7186 	 * Ok lets locate the block where this guy is fwd from rsm (if its
7187 	 * set)
7188 	 */
7189 	TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) {
7190 		if (SEQ_GEQ(start, rsm->r_start) &&
7191 		    SEQ_LT(start, rsm->r_end)) {
7192 			break;
7193 		}
7194 	}
7195 do_rest_ofb:
7196 	if (rsm == NULL) {
7197 		/*
7198 		 * This happens when we get duplicate sack blocks with the
7199 		 * same end. For example SACK 4: 100 SACK 3: 100 The sort
7200 		 * will not change there location so we would just start at
7201 		 * the end of the first one and get lost.
7202 		 */
7203 		if (tp->t_flags & TF_SENTFIN) {
7204 			/*
7205 			 * Check to see if we have not logged the FIN that
7206 			 * went out.
7207 			 */
7208 			nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7209 			if (nrsm && (nrsm->r_end + 1) == tp->snd_max) {
7210 				/*
7211 				 * Ok we did not get the FIN logged.
7212 				 */
7213 				nrsm->r_end++;
7214 				rsm = nrsm;
7215 				goto do_rest_ofb;
7216 			}
7217 		}
7218 		if (times == 1) {
7219 #ifdef BBR_INVARIANTS
7220 			panic("tp:%p bbr:%p sack:%p to:%p prsm:%p",
7221 			    tp, bbr, sack, to, prsm);
7222 #else
7223 			goto out;
7224 #endif
7225 		}
7226 		times++;
7227 		BBR_STAT_INC(bbr_sack_proc_restart);
7228 		rsm = NULL;
7229 		goto start_at_beginning;
7230 	}
7231 	/* Ok we have an ACK for some piece of rsm */
7232 	if (rsm->r_start != start) {
7233 		/*
7234 		 * Need to split this in two pieces the before and after.
7235 		 */
7236 		if (bbr_sack_mergable(rsm, start, end))
7237 			nrsm = bbr_alloc_full_limit(bbr);
7238 		else
7239 			nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7240 		if (nrsm == NULL) {
7241 			/* We could not allocate ignore the sack */
7242 			struct sackblk blk;
7243 
7244 			blk.start = start;
7245 			blk.end = end;
7246 			sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7247 			goto out;
7248 		}
7249 		bbr_clone_rsm(bbr, nrsm, rsm, start);
7250 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7251 		if (rsm->r_in_tmap) {
7252 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7253 			nrsm->r_in_tmap = 1;
7254 		}
7255 		rsm->r_flags &= (~BBR_HAS_FIN);
7256 		rsm = nrsm;
7257 	}
7258 	if (SEQ_GEQ(end, rsm->r_end)) {
7259 		/*
7260 		 * The end of this block is either beyond this guy or right
7261 		 * at this guy.
7262 		 */
7263 		if ((rsm->r_flags & BBR_ACKED) == 0) {
7264 			bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7265 			changed += (rsm->r_end - rsm->r_start);
7266 			bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7267 			bbr_log_sack_passed(tp, bbr, rsm);
7268 			if (rsm->r_flags & BBR_MARKED_LOST) {
7269 				bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7270 			}
7271 			/* Is Reordering occuring? */
7272 			if (rsm->r_flags & BBR_SACK_PASSED) {
7273 				BBR_STAT_INC(bbr_reorder_seen);
7274 				bbr->r_ctl.rc_reorder_ts = cts;
7275 				if (rsm->r_flags & BBR_MARKED_LOST) {
7276 					bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7277 					if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7278 						/* LT sampling also needs adjustment */
7279 						bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7280 				}
7281 			}
7282 			rsm->r_flags |= BBR_ACKED;
7283 			rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7284 			if (rsm->r_in_tmap) {
7285 				TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7286 				rsm->r_in_tmap = 0;
7287 			}
7288 		}
7289 		bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7290 		if (end == rsm->r_end) {
7291 			/* This block only - done */
7292 			goto out;
7293 		}
7294 		/* There is more not coverend by this rsm move on */
7295 		start = rsm->r_end;
7296 		nrsm = TAILQ_NEXT(rsm, r_next);
7297 		rsm = nrsm;
7298 		times = 0;
7299 		goto do_rest_ofb;
7300 	}
7301 	if (rsm->r_flags & BBR_ACKED) {
7302 		/* Been here done that */
7303 		goto out;
7304 	}
7305 	/* Ok we need to split off this one at the tail */
7306 	if (bbr_sack_mergable(rsm, start, end))
7307 		nrsm = bbr_alloc_full_limit(bbr);
7308 	else
7309 		nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7310 	if (nrsm == NULL) {
7311 		/* failed XXXrrs what can we do but loose the sack info? */
7312 		struct sackblk blk;
7313 
7314 		blk.start = start;
7315 		blk.end = end;
7316 		sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7317 		goto out;
7318 	}
7319 	/* Clone it */
7320 	bbr_clone_rsm(bbr, nrsm, rsm, end);
7321 	/* The sack block does not cover this guy fully */
7322 	rsm->r_flags &= (~BBR_HAS_FIN);
7323 	TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7324 	if (rsm->r_in_tmap) {
7325 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7326 		nrsm->r_in_tmap = 1;
7327 	}
7328 	nrsm->r_dupack = 0;
7329 	bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7330 	bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7331 	changed += (rsm->r_end - rsm->r_start);
7332 	bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7333 	bbr_log_sack_passed(tp, bbr, rsm);
7334 	/* Is Reordering occuring? */
7335 	if (rsm->r_flags & BBR_MARKED_LOST) {
7336 		bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7337 	}
7338 	if (rsm->r_flags & BBR_SACK_PASSED) {
7339 		BBR_STAT_INC(bbr_reorder_seen);
7340 		bbr->r_ctl.rc_reorder_ts = cts;
7341 		if (rsm->r_flags & BBR_MARKED_LOST) {
7342 			bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7343 			if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7344 				/* LT sampling also needs adjustment */
7345 				bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7346 		}
7347 	}
7348 	rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7349 	rsm->r_flags |= BBR_ACKED;
7350 	if (rsm->r_in_tmap) {
7351 		TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7352 		rsm->r_in_tmap = 0;
7353 	}
7354 out:
7355 	if (rsm && (rsm->r_flags & BBR_ACKED)) {
7356 		/*
7357 		 * Now can we merge this newly acked
7358 		 * block with either the previous or
7359 		 * next block?
7360 		 */
7361 		nrsm = TAILQ_NEXT(rsm, r_next);
7362 		if (nrsm &&
7363 		    (nrsm->r_flags & BBR_ACKED)) {
7364 			/* yep this and next can be merged */
7365 			rsm = bbr_merge_rsm(bbr, rsm, nrsm);
7366 		}
7367 		/* Now what about the previous? */
7368 		nrsm = TAILQ_PREV(rsm, bbr_head, r_next);
7369 		if (nrsm &&
7370 		    (nrsm->r_flags & BBR_ACKED)) {
7371 			/* yep the previous and this can be merged */
7372 			rsm = bbr_merge_rsm(bbr, nrsm, rsm);
7373 		}
7374 	}
7375 	if (used_ref == 0) {
7376 		BBR_STAT_INC(bbr_sack_proc_all);
7377 	} else {
7378 		BBR_STAT_INC(bbr_sack_proc_short);
7379 	}
7380 	if (went_fwd && went_back) {
7381 		BBR_STAT_INC(bbr_sack_search_both);
7382 	} else if (went_fwd) {
7383 		BBR_STAT_INC(bbr_sack_search_fwd);
7384 	} else if (went_back) {
7385 		BBR_STAT_INC(bbr_sack_search_back);
7386 	}
7387 	/* Save off where the next seq is */
7388 	if (rsm)
7389 		bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next);
7390 	else
7391 		bbr->r_ctl.rc_sacklast = NULL;
7392 	*prsm = rsm;
7393 	return (changed);
7394 }
7395 
7396 
7397 static void inline
7398 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack)
7399 {
7400 	struct bbr_sendmap *tmap;
7401 
7402 	BBR_STAT_INC(bbr_reneges_seen);
7403 	tmap = NULL;
7404 	while (rsm && (rsm->r_flags & BBR_ACKED)) {
7405 		/* Its no longer sacked, mark it so */
7406 		uint32_t oflags;
7407 		bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7408 #ifdef BBR_INVARIANTS
7409 		if (rsm->r_in_tmap) {
7410 			panic("bbr:%p rsm:%p flags:0x%x in tmap?",
7411 			    bbr, rsm, rsm->r_flags);
7412 		}
7413 #endif
7414 		oflags = rsm->r_flags;
7415 		if (rsm->r_flags & BBR_MARKED_LOST) {
7416 			bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7417 			bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7418 			if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7419 				/* LT sampling also needs adjustment */
7420 				bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7421 		}
7422 		rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST);
7423 		rsm->r_flags |= BBR_WAS_RENEGED;
7424 		rsm->r_flags |= BBR_RXT_CLEARED;
7425 		bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__);
7426 		/* Rebuild it into our tmap */
7427 		if (tmap == NULL) {
7428 			TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7429 			tmap = rsm;
7430 		} else {
7431 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7432 			tmap = rsm;
7433 		}
7434 		tmap->r_in_tmap = 1;
7435 		/*
7436 		 * XXXrrs Delivered? Should we do anything here?
7437 		 *
7438 		 * Of course we don't on a rxt timeout so maybe its ok that
7439 		 * we don't?
7440 		 *
7441 		 * For now lets not.
7442 		 */
7443 		rsm = TAILQ_NEXT(rsm, r_next);
7444 	}
7445 	/*
7446 	 * Now lets possibly clear the sack filter so we start recognizing
7447 	 * sacks that cover this area.
7448 	 */
7449 	sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack);
7450 }
7451 
7452 static void
7453 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to)
7454 {
7455 	struct tcp_bbr *bbr;
7456 	struct bbr_sendmap *rsm;
7457 	uint32_t cts;
7458 
7459 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7460 	cts = bbr->r_ctl.rc_rcvtime;
7461 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7462 	if (rsm && (rsm->r_flags & BBR_HAS_SYN)) {
7463 		if ((rsm->r_end - rsm->r_start) <= 1) {
7464 			/* Log out the SYN completely */
7465 			bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7466 			rsm->r_rtr_bytes = 0;
7467 			TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7468 			if (rsm->r_in_tmap) {
7469 				TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7470 				rsm->r_in_tmap = 0;
7471 			}
7472 			if (bbr->r_ctl.rc_next == rsm) {
7473 				/* scoot along the marker */
7474 				bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7475 			}
7476 			if (to != NULL)
7477 				bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0);
7478 			bbr_free(bbr, rsm);
7479 		} else {
7480 			/* There is more (Fast open)? strip out SYN. */
7481 			rsm->r_flags &= ~BBR_HAS_SYN;
7482 			rsm->r_start++;
7483 		}
7484 	}
7485 }
7486 
7487 /*
7488  * Returns the number of bytes that were
7489  * acknowledged by SACK blocks.
7490  */
7491 
7492 static uint32_t
7493 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th,
7494     uint32_t *prev_acked)
7495 {
7496 	uint32_t changed, last_seq, entered_recovery = 0;
7497 	struct tcp_bbr *bbr;
7498 	struct bbr_sendmap *rsm;
7499 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7500 	register uint32_t th_ack;
7501 	int32_t i, j, k, new_sb, num_sack_blks = 0;
7502 	uint32_t cts, acked, ack_point, sack_changed = 0;
7503 	uint32_t p_maxseg, maxseg, p_acked = 0;
7504 
7505 	INP_WLOCK_ASSERT(tp->t_inpcb);
7506 	if (th->th_flags & TH_RST) {
7507 		/* We don't log resets */
7508 		return (0);
7509 	}
7510 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7511 	cts = bbr->r_ctl.rc_rcvtime;
7512 
7513 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7514 	changed = 0;
7515 	maxseg = tp->t_maxseg - bbr->rc_last_options;
7516 	p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg);
7517 	th_ack = th->th_ack;
7518 	if (SEQ_GT(th_ack, tp->snd_una)) {
7519 		acked = th_ack - tp->snd_una;
7520 		bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__);
7521 		bbr->rc_tp->t_acktime = ticks;
7522 	} else
7523 		acked = 0;
7524 	if (SEQ_LEQ(th_ack, tp->snd_una)) {
7525 		/* Only sent here for sack processing */
7526 		goto proc_sack;
7527 	}
7528 	if (rsm && SEQ_GT(th_ack, rsm->r_start)) {
7529 		changed = th_ack - rsm->r_start;
7530 	} else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) {
7531 		/*
7532 		 * For the SYN incoming case we will not have called
7533 		 * tcp_output for the sending of the SYN, so there will be
7534 		 * no map. All other cases should probably be a panic.
7535 		 */
7536 		if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) {
7537 			/*
7538 			 * We have a timestamp that can be used to generate
7539 			 * an initial RTT.
7540 			 */
7541 			uint32_t ts, now, rtt;
7542 
7543 			ts = bbr_ts_convert(to->to_tsecr);
7544 			now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv));
7545 			rtt = now - ts;
7546 			if (rtt < 1)
7547 				rtt = 1;
7548 			bbr_log_type_bbrrttprop(bbr, rtt,
7549 						tp->iss, 0, cts,
7550 						BBR_RTT_BY_TIMESTAMP, tp->iss, 0);
7551 			apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
7552 			changed = 1;
7553 			bbr->r_wanted_output = 1;
7554 			goto out;
7555 		}
7556 		goto proc_sack;
7557 	} else if (rsm == NULL) {
7558 		goto out;
7559 	}
7560 	if (changed) {
7561 		/*
7562 		 * The ACK point is advancing to th_ack, we must drop off
7563 		 * the packets in the rack log and calculate any eligble
7564 		 * RTT's.
7565 		 */
7566 		bbr->r_wanted_output = 1;
7567 more:
7568 		if (rsm == NULL) {
7569 
7570 			if (tp->t_flags & TF_SENTFIN) {
7571 				/* if we send a FIN we will not hav a map */
7572 				goto proc_sack;
7573 			}
7574 #ifdef BBR_INVARIANTS
7575 			panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n",
7576 			    tp,
7577 			    th, tp->t_state, bbr,
7578 			    tp->snd_una, tp->snd_max, changed);
7579 #endif
7580 			goto proc_sack;
7581 		}
7582 	}
7583 	if (SEQ_LT(th_ack, rsm->r_start)) {
7584 		/* Huh map is missing this */
7585 #ifdef BBR_INVARIANTS
7586 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n",
7587 		    rsm->r_start,
7588 		    th_ack, tp->t_state,
7589 		    bbr->r_state, bbr);
7590 		panic("th-ack is bad bbr:%p tp:%p", bbr, tp);
7591 #endif
7592 		goto proc_sack;
7593 	} else if (th_ack == rsm->r_start) {
7594 		/* None here to ack */
7595 		goto proc_sack;
7596 	}
7597 	/*
7598 	 * Clear the dup ack counter, it will
7599 	 * either be freed or if there is some
7600 	 * remaining we need to start it at zero.
7601 	 */
7602 	rsm->r_dupack = 0;
7603 	/* Now do we consume the whole thing? */
7604 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
7605 		/* Its all consumed. */
7606 		uint32_t left;
7607 
7608 		if (rsm->r_flags & BBR_ACKED) {
7609 			/*
7610 			 * It was acked on the scoreboard -- remove it from
7611 			 * total
7612 			 */
7613 			p_acked += (rsm->r_end - rsm->r_start);
7614 			bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7615 			if (bbr->r_ctl.rc_sacked == 0)
7616 				bbr->r_ctl.rc_sacklast = NULL;
7617 		} else {
7618 			bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack);
7619 			if (rsm->r_flags & BBR_MARKED_LOST) {
7620 				bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7621 			}
7622 			if (rsm->r_flags & BBR_SACK_PASSED) {
7623 				/*
7624 				 * There are acked segments ACKED on the
7625 				 * scoreboard further up. We are seeing
7626 				 * reordering.
7627 				 */
7628 				BBR_STAT_INC(bbr_reorder_seen);
7629 				bbr->r_ctl.rc_reorder_ts = cts;
7630 				if (rsm->r_flags & BBR_MARKED_LOST) {
7631 					bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7632 					if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7633 						/* LT sampling also needs adjustment */
7634 						bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7635 				}
7636 			}
7637 			rsm->r_flags &= ~BBR_MARKED_LOST;
7638 		}
7639 		bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7640 		rsm->r_rtr_bytes = 0;
7641 		TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7642 		if (rsm->r_in_tmap) {
7643 			TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7644 			rsm->r_in_tmap = 0;
7645 		}
7646 		if (bbr->r_ctl.rc_next == rsm) {
7647 			/* scoot along the marker */
7648 			bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7649 		}
7650 		bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7651 		/* Adjust the packet counts */
7652 		left = th_ack - rsm->r_end;
7653 		/* Free back to zone */
7654 		bbr_free(bbr, rsm);
7655 		if (left) {
7656 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7657 			goto more;
7658 		}
7659 		goto proc_sack;
7660 	}
7661 	if (rsm->r_flags & BBR_ACKED) {
7662 		/*
7663 		 * It was acked on the scoreboard -- remove it from total
7664 		 * for the part being cum-acked.
7665 		 */
7666 		p_acked += (rsm->r_end - rsm->r_start);
7667 		bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7668 		if (bbr->r_ctl.rc_sacked == 0)
7669 			bbr->r_ctl.rc_sacklast = NULL;
7670 	} else {
7671 		/*
7672 		 * It was acked up to th_ack point for the first time
7673 		 */
7674 		struct bbr_sendmap lrsm;
7675 
7676 		memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap));
7677 		lrsm.r_end = th_ack;
7678 		bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack);
7679 	}
7680 	if ((rsm->r_flags & BBR_MARKED_LOST) &&
7681 	    ((rsm->r_flags & BBR_ACKED) == 0)) {
7682 		/*
7683 		 * It was marked lost and partly ack'd now
7684 		 * for the first time. We lower the rc_lost_bytes
7685 		 * and still leave it MARKED.
7686 		 */
7687 		bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start;
7688 	}
7689 	bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7690 	bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7691 	rsm->r_rtr_bytes = 0;
7692 	/* adjust packet count */
7693 	rsm->r_start = th_ack;
7694 proc_sack:
7695 	/* Check for reneging */
7696 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7697 	if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) {
7698 		/*
7699 		 * The peer has moved snd_una up to the edge of this send,
7700 		 * i.e. one that it had previously acked. The only way that
7701 		 * can be true if the peer threw away data (space issues)
7702 		 * that it had previously sacked (else it would have given
7703 		 * us snd_una up to (rsm->r_end). We need to undo the acked
7704 		 * markings here.
7705 		 *
7706 		 * Note we have to look to make sure th_ack is our
7707 		 * rsm->r_start in case we get an old ack where th_ack is
7708 		 * behind snd_una.
7709 		 */
7710 		bbr_peer_reneges(bbr, rsm, th->th_ack);
7711 	}
7712 	if ((to->to_flags & TOF_SACK) == 0) {
7713 		/* We are done nothing left to log */
7714 		goto out;
7715 	}
7716 	rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7717 	if (rsm) {
7718 		last_seq = rsm->r_end;
7719 	} else {
7720 		last_seq = tp->snd_max;
7721 	}
7722 	/* Sack block processing */
7723 	if (SEQ_GT(th_ack, tp->snd_una))
7724 		ack_point = th_ack;
7725 	else
7726 		ack_point = tp->snd_una;
7727 	for (i = 0; i < to->to_nsacks; i++) {
7728 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
7729 		    &sack, sizeof(sack));
7730 		sack.start = ntohl(sack.start);
7731 		sack.end = ntohl(sack.end);
7732 		if (SEQ_GT(sack.end, sack.start) &&
7733 		    SEQ_GT(sack.start, ack_point) &&
7734 		    SEQ_LT(sack.start, tp->snd_max) &&
7735 		    SEQ_GT(sack.end, ack_point) &&
7736 		    SEQ_LEQ(sack.end, tp->snd_max)) {
7737 			if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) &&
7738 			    (SEQ_LT(sack.end, last_seq)) &&
7739 			    ((sack.end - sack.start) < (p_maxseg / 8))) {
7740 				/*
7741 				 * Not the last piece and its smaller than
7742 				 * 1/8th of a p_maxseg. We ignore this.
7743 				 */
7744 				BBR_STAT_INC(bbr_runt_sacks);
7745 				continue;
7746 			}
7747 			sack_blocks[num_sack_blks] = sack;
7748 			num_sack_blks++;
7749 #ifdef NETFLIX_STATS
7750 		} else if (SEQ_LEQ(sack.start, th_ack) &&
7751 		    SEQ_LEQ(sack.end, th_ack)) {
7752 			/*
7753 			 * Its a D-SACK block.
7754 			 */
7755 			tcp_record_dsack(sack.start, sack.end);
7756 #endif
7757 		}
7758 	}
7759 	if (num_sack_blks == 0)
7760 		goto out;
7761 	/*
7762 	 * Sort the SACK blocks so we can update the rack scoreboard with
7763 	 * just one pass.
7764 	 */
7765 	new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks,
7766 				  num_sack_blks, th->th_ack);
7767 	ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks);
7768 	BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks);
7769 	BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb));
7770 	num_sack_blks = new_sb;
7771 	if (num_sack_blks < 2) {
7772 		goto do_sack_work;
7773 	}
7774 	/* Sort the sacks */
7775 	for (i = 0; i < num_sack_blks; i++) {
7776 		for (j = i + 1; j < num_sack_blks; j++) {
7777 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7778 				sack = sack_blocks[i];
7779 				sack_blocks[i] = sack_blocks[j];
7780 				sack_blocks[j] = sack;
7781 			}
7782 		}
7783 	}
7784 	/*
7785 	 * Now are any of the sack block ends the same (yes some
7786 	 * implememtations send these)?
7787 	 */
7788 again:
7789 	if (num_sack_blks > 1) {
7790 		for (i = 0; i < num_sack_blks; i++) {
7791 			for (j = i + 1; j < num_sack_blks; j++) {
7792 				if (sack_blocks[i].end == sack_blocks[j].end) {
7793 					/*
7794 					 * Ok these two have the same end we
7795 					 * want the smallest end and then
7796 					 * throw away the larger and start
7797 					 * again.
7798 					 */
7799 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7800 						/*
7801 						 * The second block covers
7802 						 * more area use that
7803 						 */
7804 						sack_blocks[i].start = sack_blocks[j].start;
7805 					}
7806 					/*
7807 					 * Now collapse out the dup-sack and
7808 					 * lower the count
7809 					 */
7810 					for (k = (j + 1); k < num_sack_blks; k++) {
7811 						sack_blocks[j].start = sack_blocks[k].start;
7812 						sack_blocks[j].end = sack_blocks[k].end;
7813 						j++;
7814 					}
7815 					num_sack_blks--;
7816 					goto again;
7817 				}
7818 			}
7819 		}
7820 	}
7821 do_sack_work:
7822 	rsm = bbr->r_ctl.rc_sacklast;
7823 	for (i = 0; i < num_sack_blks; i++) {
7824 		acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts);
7825 		if (acked) {
7826 			bbr->r_wanted_output = 1;
7827 			changed += acked;
7828 			sack_changed += acked;
7829 		}
7830 	}
7831 out:
7832 	*prev_acked = p_acked;
7833 	if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) {
7834 		/*
7835 		 * Ok we have a high probability that we need to go in to
7836 		 * recovery since we have data sack'd
7837 		 */
7838 		struct bbr_sendmap *rsm;
7839 
7840 		rsm = bbr_check_recovery_mode(tp, bbr, cts);
7841 		if (rsm) {
7842 			/* Enter recovery */
7843 			entered_recovery = 1;
7844 			bbr->r_wanted_output = 1;
7845 			/*
7846 			 * When we enter recovery we need to assure we send
7847 			 * one packet.
7848 			 */
7849 			if (bbr->r_ctl.rc_resend == NULL) {
7850 				bbr->r_ctl.rc_resend = rsm;
7851 			}
7852 		}
7853 	}
7854 	if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) {
7855 		/*
7856 		 * See if we need to rack-retransmit anything if so set it
7857 		 * up as the thing to resend assuming something else is not
7858 		 * already in that position.
7859 		 */
7860 		if (bbr->r_ctl.rc_resend == NULL) {
7861 			bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
7862 		}
7863 	}
7864 	/*
7865 	 * We return the amount that changed via sack, this is used by the
7866 	 * ack-received code to augment what was changed between th_ack <->
7867 	 * snd_una.
7868 	 */
7869 	return (sack_changed);
7870 }
7871 
7872 static void
7873 bbr_strike_dupack(struct tcp_bbr *bbr)
7874 {
7875 	struct bbr_sendmap *rsm;
7876 
7877 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
7878 	if (rsm && (rsm->r_dupack < 0xff)) {
7879 		rsm->r_dupack++;
7880 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD)
7881 			bbr->r_wanted_output = 1;
7882 	}
7883 }
7884 
7885 /*
7886  * Return value of 1, we do not need to call bbr_process_data().
7887  * return value of 0, bbr_process_data can be called.
7888  * For ret_val if its 0 the TCB is locked and valid, if its non-zero
7889  * its unlocked and probably unsafe to touch the TCB.
7890  */
7891 static int
7892 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
7893     struct tcpcb *tp, struct tcpopt *to,
7894     uint32_t tiwin, int32_t tlen,
7895     int32_t * ofia, int32_t thflags, int32_t * ret_val)
7896 {
7897 	int32_t ourfinisacked = 0;
7898 	int32_t acked_amount;
7899 	uint16_t nsegs;
7900 	int32_t acked;
7901 	uint32_t lost, sack_changed = 0;
7902 	struct mbuf *mfree;
7903 	struct tcp_bbr *bbr;
7904 	uint32_t prev_acked = 0;
7905 
7906 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7907 	lost = bbr->r_ctl.rc_lost;
7908 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
7909 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
7910 		ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
7911 		bbr->r_wanted_output = 1;
7912 		return (1);
7913 	}
7914 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
7915 		/* Process the ack */
7916 		if (bbr->rc_in_persist)
7917 			tp->t_rxtshift = 0;
7918 		if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
7919 		        bbr_strike_dupack(bbr);
7920 		sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
7921 	}
7922 	bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost));
7923 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
7924 		/*
7925 		 * Old ack, behind the last one rcv'd or a duplicate ack
7926 		 * with SACK info.
7927 		 */
7928 		if (th->th_ack == tp->snd_una) {
7929 			bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0);
7930 			if (bbr->r_state == TCPS_SYN_SENT) {
7931 				/*
7932 				 * Special case on where we sent SYN. When
7933 				 * the SYN-ACK is processed in syn_sent
7934 				 * state it bumps the snd_una. This causes
7935 				 * us to hit here even though we did ack 1
7936 				 * byte.
7937 				 *
7938 				 * Go through the nothing left case so we
7939 				 * send data.
7940 				 */
7941 				goto nothing_left;
7942 			}
7943 		}
7944 		return (0);
7945 	}
7946 	/*
7947 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
7948 	 * something we sent.
7949 	 */
7950 	if (tp->t_flags & TF_NEEDSYN) {
7951 		/*
7952 		 * T/TCP: Connection was half-synchronized, and our SYN has
7953 		 * been ACK'd (so connection is now fully synchronized).  Go
7954 		 * to non-starred state, increment snd_una for ACK of SYN,
7955 		 * and check if we can do window scaling.
7956 		 */
7957 		tp->t_flags &= ~TF_NEEDSYN;
7958 		tp->snd_una++;
7959 		/* Do window scaling? */
7960 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
7961 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
7962 			tp->rcv_scale = tp->request_r_scale;
7963 			/* Send window already scaled. */
7964 		}
7965 	}
7966 	INP_WLOCK_ASSERT(tp->t_inpcb);
7967 
7968 	acked = BYTES_THIS_ACK(tp, th);
7969 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
7970 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
7971 
7972 	/*
7973 	 * If we just performed our first retransmit, and the ACK arrives
7974 	 * within our recovery window, then it was a mistake to do the
7975 	 * retransmit in the first place.  Recover our original cwnd and
7976 	 * ssthresh, and proceed to transmit where we left off.
7977 	 */
7978 	if (tp->t_flags & TF_PREVVALID) {
7979 		tp->t_flags &= ~TF_PREVVALID;
7980 		if (tp->t_rxtshift == 1 &&
7981 		    (int)(ticks - tp->t_badrxtwin) < 0)
7982 			bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
7983 	}
7984 	SOCKBUF_LOCK(&so->so_snd);
7985 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
7986 	tp->snd_wnd -= acked_amount;
7987 	mfree = sbcut_locked(&so->so_snd, acked_amount);
7988 	/* NB: sowwakeup_locked() does an implicit unlock. */
7989 	sowwakeup_locked(so);
7990 	m_freem(mfree);
7991 	if (SEQ_GT(th->th_ack, tp->snd_una)) {
7992 		bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
7993 	}
7994 	tp->snd_una = th->th_ack;
7995 	bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost));
7996 	if (IN_RECOVERY(tp->t_flags)) {
7997 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
7998 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
7999 			tcp_bbr_partialack(tp);
8000 		} else {
8001 			bbr_post_recovery(tp);
8002 		}
8003 	}
8004 	if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8005 		tp->snd_recover = tp->snd_una;
8006 	}
8007 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
8008 		tp->snd_nxt = tp->snd_max;
8009 	}
8010 	if (tp->snd_una == tp->snd_max) {
8011 		/* Nothing left outstanding */
8012 nothing_left:
8013 		bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8014 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8015 			bbr->rc_tp->t_acktime = 0;
8016 		if ((sbused(&so->so_snd) == 0) &&
8017 		    (tp->t_flags & TF_SENTFIN)) {
8018 			ourfinisacked = 1;
8019 		}
8020 		bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8021 		if (bbr->rc_in_persist == 0) {
8022 			bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8023 		}
8024 		sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8025 		bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8026 		/*
8027 		 * We invalidate the last ack here since we
8028 		 * don't want to transfer forward the time
8029 		 * for our sum's calculations.
8030 		 */
8031 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8032 		    (sbavail(&so->so_snd) == 0) &&
8033 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
8034 			/*
8035 			 * The socket was gone and the peer sent data, time
8036 			 * to reset him.
8037 			 */
8038 			*ret_val = 1;
8039 			tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
8040 			/* tcp_close will kill the inp pre-log the Reset */
8041 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
8042 			tp = tcp_close(tp);
8043 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
8044 			BBR_STAT_INC(bbr_dropped_af_data);
8045 			return (1);
8046 		}
8047 		/* Set need output so persist might get set */
8048 		bbr->r_wanted_output = 1;
8049 	}
8050 	if (ofia)
8051 		*ofia = ourfinisacked;
8052 	return (0);
8053 }
8054 
8055 static void
8056 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
8057 {
8058 	if (bbr->rc_in_persist == 0) {
8059 		bbr_timer_cancel(bbr, __LINE__, cts);
8060 		bbr->r_ctl.rc_last_delay_val = 0;
8061 		tp->t_rxtshift = 0;
8062 		bbr->rc_in_persist = 1;
8063 		bbr->r_ctl.rc_went_idle_time = cts;
8064 		/* We should be capped when rw went to 0 but just in case */
8065 		bbr_log_type_pesist(bbr, cts, 0, line, 1);
8066 		/* Time freezes for the state, so do the accounting now */
8067 		if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
8068 			uint32_t time_in;
8069 
8070 			time_in = cts - bbr->r_ctl.rc_bbr_state_time;
8071 			if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
8072 				int32_t idx;
8073 
8074 				idx = bbr_state_val(bbr);
8075 				counter_u64_add(bbr_state_time[(idx + 5)], time_in);
8076 			} else {
8077 				counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
8078 			}
8079 		}
8080 		bbr->r_ctl.rc_bbr_state_time = cts;
8081 	}
8082 }
8083 
8084 static void
8085 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time)
8086 {
8087 	/*
8088 	 * Note that if idle time does not exceed our
8089 	 * threshold, we do nothing continuing the state
8090 	 * transitions we were last walking through.
8091 	 */
8092 	if (idle_time >= bbr_idle_restart_threshold) {
8093 		if (bbr->rc_use_idle_restart) {
8094 			bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT;
8095 			/*
8096 			 * Set our target using BBR_UNIT, so
8097 			 * we increase at a dramatic rate but
8098 			 * we stop when we get the pipe
8099 			 * full again for our current b/w estimate.
8100 			 */
8101 			bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
8102 			bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
8103 			bbr_set_state_target(bbr, __LINE__);
8104 			/* Now setup our gains to ramp up */
8105 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
8106 			bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
8107 			bbr_log_type_statechange(bbr, cts, __LINE__);
8108 		} else {
8109 			bbr_substate_change(bbr, cts, __LINE__, 1);
8110 		}
8111 	}
8112 }
8113 
8114 static void
8115 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
8116 {
8117 	uint32_t idle_time;
8118 
8119 	if (bbr->rc_in_persist == 0)
8120 		return;
8121 	idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time);
8122 	bbr->rc_in_persist = 0;
8123 	bbr->rc_hit_state_1 = 0;
8124 	bbr->r_ctl.rc_del_time = cts;
8125 	/*
8126 	 * We invalidate the last ack here since we
8127 	 * don't want to transfer forward the time
8128 	 * for our sum's calculations.
8129 	 */
8130 	if (bbr->rc_inp->inp_in_hpts) {
8131 		tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
8132 		bbr->rc_timer_first = 0;
8133 		bbr->r_ctl.rc_hpts_flags = 0;
8134 		bbr->r_ctl.rc_last_delay_val = 0;
8135 		bbr->r_ctl.rc_hptsi_agg_delay = 0;
8136 		bbr->r_agg_early_set = 0;
8137 		bbr->r_ctl.rc_agg_early = 0;
8138 	}
8139 	bbr_log_type_pesist(bbr, cts, idle_time, line, 0);
8140 	if (idle_time >= bbr_rtt_probe_time) {
8141 		/*
8142 		 * This qualifies as a RTT_PROBE session since we drop the
8143 		 * data outstanding to nothing and waited more than
8144 		 * bbr_rtt_probe_time.
8145 		 */
8146 		bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0);
8147 		bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts;
8148 	}
8149 	tp->t_rxtshift = 0;
8150 	/*
8151 	 * If in probeBW and we have persisted more than an RTT lets do
8152 	 * special handling.
8153 	 */
8154 	/* Force a time based epoch */
8155 	bbr_set_epoch(bbr, cts, __LINE__);
8156 	/*
8157 	 * Setup the lost so we don't count anything against the guy
8158 	 * we have been stuck with during persists.
8159 	 */
8160 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
8161 	/* Time un-freezes for the state */
8162 	bbr->r_ctl.rc_bbr_state_time = cts;
8163 	if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) ||
8164 	    (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) {
8165 		/*
8166 		 * If we are going back to probe-bw
8167 		 * or probe_rtt, we may need to possibly
8168 		 * do a fast restart.
8169 		 */
8170 		bbr_restart_after_idle(bbr, cts, idle_time);
8171 	}
8172 }
8173 
8174 static void
8175 bbr_collapsed_window(struct tcp_bbr *bbr)
8176 {
8177 	/*
8178 	 * Now we must walk the
8179 	 * send map and divide the
8180 	 * ones left stranded. These
8181 	 * guys can't cause us to abort
8182 	 * the connection and are really
8183 	 * "unsent". However if a buggy
8184 	 * client actually did keep some
8185 	 * of the data i.e. collapsed the win
8186 	 * and refused to ack and then opened
8187 	 * the win and acked that data. We would
8188 	 * get into an ack war, the simplier
8189 	 * method then of just pretending we
8190 	 * did not send those segments something
8191 	 * won't work.
8192 	 */
8193 	struct bbr_sendmap *rsm, *nrsm;
8194 	tcp_seq max_seq;
8195 	uint32_t maxseg;
8196 	int can_split = 0;
8197 	int fnd = 0;
8198 
8199 	maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
8200 	max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd;
8201 	bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0);
8202 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
8203 		/* Find the first seq past or at maxseq */
8204 		if (rsm->r_flags & BBR_RWND_COLLAPSED)
8205 			rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8206 		if (SEQ_GEQ(max_seq, rsm->r_start) &&
8207 		    SEQ_GEQ(rsm->r_end, max_seq)) {
8208 			fnd = 1;
8209 			break;
8210 		}
8211 	}
8212 	bbr->rc_has_collapsed = 0;
8213 	if (!fnd) {
8214 		/* Nothing to do strange */
8215 		return;
8216 	}
8217 	/*
8218 	 * Now can we split?
8219 	 *
8220 	 * We don't want to split if splitting
8221 	 * would generate too many small segments
8222 	 * less we let an attacker fragment our
8223 	 * send_map and leave us out of memory.
8224 	 */
8225 	if ((max_seq != rsm->r_start) &&
8226 	    (max_seq != rsm->r_end)){
8227 		/* can we split? */
8228 		int res1, res2;
8229 
8230 		res1 = max_seq - rsm->r_start;
8231 		res2 = rsm->r_end - max_seq;
8232 		if ((res1 >= (maxseg/8)) &&
8233 		    (res2 >= (maxseg/8))) {
8234 			/* No small pieces here */
8235 			can_split = 1;
8236 		} else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) {
8237 			/* We are under the limit */
8238 			can_split = 1;
8239 		}
8240 	}
8241 	/* Ok do we need to split this rsm? */
8242 	if (max_seq == rsm->r_start) {
8243 		/* It's this guy no split required */
8244 		nrsm = rsm;
8245 	} else if (max_seq == rsm->r_end) {
8246 		/* It's the next one no split required. */
8247 		nrsm = TAILQ_NEXT(rsm, r_next);
8248 		if (nrsm == NULL) {
8249 			/* Huh? */
8250 			return;
8251 		}
8252 	} else if (can_split && SEQ_LT(max_seq, rsm->r_end)) {
8253 		/* yep we need to split it */
8254 		nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
8255 		if (nrsm == NULL) {
8256 			/* failed XXXrrs what can we do mark the whole? */
8257 			nrsm = rsm;
8258 			goto no_split;
8259 		}
8260 		/* Clone it */
8261 		bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0);
8262 		bbr_clone_rsm(bbr, nrsm, rsm, max_seq);
8263 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
8264 		if (rsm->r_in_tmap) {
8265 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8266 			nrsm->r_in_tmap = 1;
8267 		}
8268 	} else {
8269 		/*
8270 		 * Split not allowed just start here just
8271 		 * use this guy.
8272 		 */
8273 		nrsm = rsm;
8274 	}
8275 no_split:
8276 	BBR_STAT_INC(bbr_collapsed_win);
8277 	/* reuse fnd as a count */
8278 	fnd = 0;
8279 	TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) {
8280 		nrsm->r_flags |= BBR_RWND_COLLAPSED;
8281 		fnd++;
8282 		bbr->rc_has_collapsed = 1;
8283 	}
8284 	bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd);
8285 }
8286 
8287 static void
8288 bbr_un_collapse_window(struct tcp_bbr *bbr)
8289 {
8290 	struct bbr_sendmap *rsm;
8291 	int cleared = 0;
8292 
8293 	TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
8294 		if (rsm->r_flags & BBR_RWND_COLLAPSED) {
8295 			/* Clear the flag */
8296 			rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8297 			cleared++;
8298 		} else
8299 			break;
8300 	}
8301 	bbr_log_type_rwnd_collapse(bbr,
8302 				   (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared);
8303 	bbr->rc_has_collapsed = 0;
8304 }
8305 
8306 /*
8307  * Return value of 1, the TCB is unlocked and most
8308  * likely gone, return value of 0, the TCB is still
8309  * locked.
8310  */
8311 static int
8312 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8313     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8314     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8315 {
8316 	/*
8317 	 * Update window information. Don't look at window if no ACK: TAC's
8318 	 * send garbage on first SYN.
8319 	 */
8320 	uint16_t nsegs;
8321 	int32_t tfo_syn;
8322 	struct tcp_bbr *bbr;
8323 
8324 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8325 	INP_WLOCK_ASSERT(tp->t_inpcb);
8326 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8327 	if ((thflags & TH_ACK) &&
8328 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8329 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8330 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8331 		/* keep track of pure window updates */
8332 		if (tlen == 0 &&
8333 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8334 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
8335 		tp->snd_wnd = tiwin;
8336 		tp->snd_wl1 = th->th_seq;
8337 		tp->snd_wl2 = th->th_ack;
8338 		if (tp->snd_wnd > tp->max_sndwnd)
8339 			tp->max_sndwnd = tp->snd_wnd;
8340 		bbr->r_wanted_output = 1;
8341 	} else if (thflags & TH_ACK) {
8342 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8343 			tp->snd_wnd = tiwin;
8344 			tp->snd_wl1 = th->th_seq;
8345 			tp->snd_wl2 = th->th_ack;
8346 		}
8347 	}
8348 	if (tp->snd_wnd < ctf_outstanding(tp))
8349 		/* The peer collapsed its window on us */
8350 		bbr_collapsed_window(bbr);
8351  	else if (bbr->rc_has_collapsed)
8352 		bbr_un_collapse_window(bbr);
8353 	/* Was persist timer active and now we have window space? */
8354 	if ((bbr->rc_in_persist != 0) &&
8355 	    (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8356 				bbr_minseg(bbr)))) {
8357 		/*
8358 		 * Make the rate persist at end of persist mode if idle long
8359 		 * enough
8360 		 */
8361 		bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8362 
8363 		/* Make sure we output to start the timer */
8364 		bbr->r_wanted_output = 1;
8365 	}
8366 	/* Do we need to enter persist? */
8367 	if ((bbr->rc_in_persist == 0) &&
8368 	    (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8369 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
8370 	    (tp->snd_max == tp->snd_una) &&
8371 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8372 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8373 		/* No send window.. we must enter persist */
8374 		bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8375 	}
8376 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8377 		m_freem(m);
8378 		return (0);
8379 	}
8380 	/*
8381 	 * We don't support urgent data but
8382 	 * drag along the up just to make sure
8383 	 * if there is a stack switch no one
8384 	 * is surprised.
8385 	 */
8386 	tp->rcv_up = tp->rcv_nxt;
8387 	INP_WLOCK_ASSERT(tp->t_inpcb);
8388 
8389 	/*
8390 	 * Process the segment text, merging it into the TCP sequencing
8391 	 * queue, and arranging for acknowledgment of receipt if necessary.
8392 	 * This process logically involves adjusting tp->rcv_wnd as data is
8393 	 * presented to the user (this happens in tcp_usrreq.c, case
8394 	 * PRU_RCVD).  If a FIN has already been received on this connection
8395 	 * then we just ignore the text.
8396 	 */
8397 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8398 		   IS_FASTOPEN(tp->t_flags));
8399 	if ((tlen || (thflags & TH_FIN) || tfo_syn) &&
8400 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8401 		tcp_seq save_start = th->th_seq;
8402 		tcp_seq save_rnxt  = tp->rcv_nxt;
8403 		int     save_tlen  = tlen;
8404 
8405 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8406 		/*
8407 		 * Insert segment which includes th into TCP reassembly
8408 		 * queue with control block tp.  Set thflags to whether
8409 		 * reassembly now includes a segment with FIN.  This handles
8410 		 * the common case inline (segment is the next to be
8411 		 * received on an established connection, and the queue is
8412 		 * empty), avoiding linkage into and removal from the queue
8413 		 * and repetition of various conversions. Set DELACK for
8414 		 * segments received in order, but ack immediately when
8415 		 * segments are out of order (so fast retransmit can work).
8416 		 */
8417 		if (th->th_seq == tp->rcv_nxt &&
8418 		    SEGQ_EMPTY(tp) &&
8419 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
8420 		    tfo_syn)) {
8421 #ifdef NETFLIX_SB_LIMITS
8422 			u_int mcnt, appended;
8423 
8424 			if (so->so_rcv.sb_shlim) {
8425 				mcnt = m_memcnt(m);
8426 				appended = 0;
8427 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8428 				    CFO_NOSLEEP, NULL) == false) {
8429 					counter_u64_add(tcp_sb_shlim_fails, 1);
8430 					m_freem(m);
8431 					return (0);
8432 				}
8433 			}
8434 
8435 #endif
8436 			if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) {
8437 				bbr->bbr_segs_rcvd += max(1, nsegs);
8438 				tp->t_flags |= TF_DELACK;
8439 				bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8440 			} else {
8441 				bbr->r_wanted_output = 1;
8442 				tp->t_flags |= TF_ACKNOW;
8443 			}
8444 			tp->rcv_nxt += tlen;
8445 			thflags = th->th_flags & TH_FIN;
8446 			KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8447 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8448 			SOCKBUF_LOCK(&so->so_rcv);
8449 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
8450 				m_freem(m);
8451 			else
8452 #ifdef NETFLIX_SB_LIMITS
8453 				appended =
8454 #endif
8455 					sbappendstream_locked(&so->so_rcv, m, 0);
8456 			/* NB: sorwakeup_locked() does an implicit unlock. */
8457 			sorwakeup_locked(so);
8458 #ifdef NETFLIX_SB_LIMITS
8459 			if (so->so_rcv.sb_shlim && appended != mcnt)
8460 				counter_fo_release(so->so_rcv.sb_shlim,
8461 				    mcnt - appended);
8462 #endif
8463 		} else {
8464 			/*
8465 			 * XXX: Due to the header drop above "th" is
8466 			 * theoretically invalid by now.  Fortunately
8467 			 * m_adj() doesn't actually frees any mbufs when
8468 			 * trimming from the head.
8469 			 */
8470 			tcp_seq temp = save_start;
8471 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
8472 			tp->t_flags |= TF_ACKNOW;
8473 		}
8474 		if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0)) {
8475 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8476 				/*
8477 				 * DSACK actually handled in the fastpath
8478 				 * above.
8479 				 */
8480 				tcp_update_sack_list(tp, save_start,
8481 				    save_start + save_tlen);
8482 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8483 				if ((tp->rcv_numsacks >= 1) &&
8484 				    (tp->sackblks[0].end == save_start)) {
8485 					/*
8486 					 * Partial overlap, recorded at todrop
8487 					 * above.
8488 					 */
8489 					tcp_update_sack_list(tp,
8490 					    tp->sackblks[0].start,
8491 					    tp->sackblks[0].end);
8492 				} else {
8493 					tcp_update_dsack_list(tp, save_start,
8494 					    save_start + save_tlen);
8495 				}
8496 			} else if (tlen >= save_tlen) {
8497 				/* Update of sackblks. */
8498 				tcp_update_dsack_list(tp, save_start,
8499 				    save_start + save_tlen);
8500 			} else if (tlen > 0) {
8501 				tcp_update_dsack_list(tp, save_start,
8502 				    save_start + tlen);
8503 			}
8504 		}
8505 	} else {
8506 		m_freem(m);
8507 		thflags &= ~TH_FIN;
8508 	}
8509 
8510 	/*
8511 	 * If FIN is received ACK the FIN and let the user know that the
8512 	 * connection is closing.
8513 	 */
8514 	if (thflags & TH_FIN) {
8515 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8516 			socantrcvmore(so);
8517 			/*
8518 			 * If connection is half-synchronized (ie NEEDSYN
8519 			 * flag on) then delay ACK, so it may be piggybacked
8520 			 * when SYN is sent. Otherwise, since we received a
8521 			 * FIN then no more input can be expected, send ACK
8522 			 * now.
8523 			 */
8524 			if (tp->t_flags & TF_NEEDSYN) {
8525 				tp->t_flags |= TF_DELACK;
8526 				bbr_timer_cancel(bbr,
8527 				    __LINE__, bbr->r_ctl.rc_rcvtime);
8528 			} else {
8529 				tp->t_flags |= TF_ACKNOW;
8530 			}
8531 			tp->rcv_nxt++;
8532 		}
8533 		switch (tp->t_state) {
8534 
8535 			/*
8536 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
8537 			 * CLOSE_WAIT state.
8538 			 */
8539 		case TCPS_SYN_RECEIVED:
8540 			tp->t_starttime = ticks;
8541 			/* FALLTHROUGH */
8542 		case TCPS_ESTABLISHED:
8543 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
8544 			break;
8545 
8546 			/*
8547 			 * If still in FIN_WAIT_1 STATE FIN has not been
8548 			 * acked so enter the CLOSING state.
8549 			 */
8550 		case TCPS_FIN_WAIT_1:
8551 			tcp_state_change(tp, TCPS_CLOSING);
8552 			break;
8553 
8554 			/*
8555 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
8556 			 * starting the time-wait timer, turning off the
8557 			 * other standard timers.
8558 			 */
8559 		case TCPS_FIN_WAIT_2:
8560 			bbr->rc_timer_first = 1;
8561 			bbr_timer_cancel(bbr,
8562 			    __LINE__, bbr->r_ctl.rc_rcvtime);
8563 			INP_WLOCK_ASSERT(tp->t_inpcb);
8564 			tcp_twstart(tp);
8565 			return (1);
8566 		}
8567 	}
8568 	/*
8569 	 * Return any desired output.
8570 	 */
8571 	if ((tp->t_flags & TF_ACKNOW) ||
8572 	    (sbavail(&so->so_snd) > ctf_outstanding(tp))) {
8573 		bbr->r_wanted_output = 1;
8574 	}
8575 	INP_WLOCK_ASSERT(tp->t_inpcb);
8576 	return (0);
8577 }
8578 
8579 /*
8580  * Here nothing is really faster, its just that we
8581  * have broken out the fast-data path also just like
8582  * the fast-ack. Return 1 if we processed the packet
8583  * return 0 if you need to take the "slow-path".
8584  */
8585 static int
8586 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8587     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8588     uint32_t tiwin, int32_t nxt_pkt)
8589 {
8590 	uint16_t nsegs;
8591 	int32_t newsize = 0;	/* automatic sockbuf scaling */
8592 	struct tcp_bbr *bbr;
8593 #ifdef NETFLIX_SB_LIMITS
8594 	u_int mcnt, appended;
8595 #endif
8596 #ifdef TCPDEBUG
8597 	/*
8598 	 * The size of tcp_saveipgen must be the size of the max ip header,
8599 	 * now IPv6.
8600 	 */
8601 	u_char tcp_saveipgen[IP6_HDR_LEN];
8602 	struct tcphdr tcp_savetcp;
8603 	short ostate = 0;
8604 
8605 #endif
8606 	/* On the hpts and we would have called output */
8607 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8608 
8609 	/*
8610 	 * If last ACK falls within this segment's sequence numbers, record
8611 	 * the timestamp. NOTE that the test is modified according to the
8612 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8613 	 */
8614 	if (bbr->r_ctl.rc_resend != NULL) {
8615 		return (0);
8616 	}
8617 	if (tiwin && tiwin != tp->snd_wnd) {
8618 		return (0);
8619 	}
8620 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8621 		return (0);
8622 	}
8623 	if (__predict_false((to->to_flags & TOF_TS) &&
8624 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8625 		return (0);
8626 	}
8627 	if (__predict_false((th->th_ack != tp->snd_una))) {
8628 		return (0);
8629 	}
8630 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8631 		return (0);
8632 	}
8633 	if ((to->to_flags & TOF_TS) != 0 &&
8634 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8635 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
8636 		tp->ts_recent = to->to_tsval;
8637 	}
8638 	/*
8639 	 * This is a pure, in-sequence data packet with nothing on the
8640 	 * reassembly queue and we have enough buffer space to take it.
8641 	 */
8642 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8643 
8644 #ifdef NETFLIX_SB_LIMITS
8645 	if (so->so_rcv.sb_shlim) {
8646 		mcnt = m_memcnt(m);
8647 		appended = 0;
8648 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8649 		    CFO_NOSLEEP, NULL) == false) {
8650 			counter_u64_add(tcp_sb_shlim_fails, 1);
8651 			m_freem(m);
8652 			return (1);
8653 		}
8654 	}
8655 #endif
8656 	/* Clean receiver SACK report if present */
8657 	if (tp->rcv_numsacks)
8658 		tcp_clean_sackreport(tp);
8659 	KMOD_TCPSTAT_INC(tcps_preddat);
8660 	tp->rcv_nxt += tlen;
8661 	/*
8662 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8663 	 */
8664 	tp->snd_wl1 = th->th_seq;
8665 	/*
8666 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8667 	 */
8668 	tp->rcv_up = tp->rcv_nxt;
8669 	KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8670 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8671 #ifdef TCPDEBUG
8672 	if (so->so_options & SO_DEBUG)
8673 		tcp_trace(TA_INPUT, ostate, tp,
8674 		    (void *)tcp_saveipgen, &tcp_savetcp, 0);
8675 #endif
8676 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8677 
8678 	/* Add data to socket buffer. */
8679 	SOCKBUF_LOCK(&so->so_rcv);
8680 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8681 		m_freem(m);
8682 	} else {
8683 		/*
8684 		 * Set new socket buffer size. Give up when limit is
8685 		 * reached.
8686 		 */
8687 		if (newsize)
8688 			if (!sbreserve_locked(&so->so_rcv,
8689 			    newsize, so, NULL))
8690 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8691 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8692 
8693 #ifdef NETFLIX_SB_LIMITS
8694 		appended =
8695 #endif
8696 			sbappendstream_locked(&so->so_rcv, m, 0);
8697 		ctf_calc_rwin(so, tp);
8698 	}
8699 	/* NB: sorwakeup_locked() does an implicit unlock. */
8700 	sorwakeup_locked(so);
8701 #ifdef NETFLIX_SB_LIMITS
8702 	if (so->so_rcv.sb_shlim && mcnt != appended)
8703 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8704 #endif
8705 	if (DELAY_ACK(tp, bbr, nsegs)) {
8706 		bbr->bbr_segs_rcvd += max(1, nsegs);
8707 		tp->t_flags |= TF_DELACK;
8708 		bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8709 	} else {
8710 		bbr->r_wanted_output = 1;
8711 		tp->t_flags |= TF_ACKNOW;
8712 	}
8713 	return (1);
8714 }
8715 
8716 /*
8717  * This subfunction is used to try to highly optimize the
8718  * fast path. We again allow window updates that are
8719  * in sequence to remain in the fast-path. We also add
8720  * in the __predict's to attempt to help the compiler.
8721  * Note that if we return a 0, then we can *not* process
8722  * it and the caller should push the packet into the
8723  * slow-path. If we return 1, then all is well and
8724  * the packet is fully processed.
8725  */
8726 static int
8727 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8728     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8729     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
8730 {
8731 	int32_t acked;
8732 	uint16_t nsegs;
8733 	uint32_t sack_changed;
8734 #ifdef TCPDEBUG
8735 	/*
8736 	 * The size of tcp_saveipgen must be the size of the max ip header,
8737 	 * now IPv6.
8738 	 */
8739 	u_char tcp_saveipgen[IP6_HDR_LEN];
8740 	struct tcphdr tcp_savetcp;
8741 	short ostate = 0;
8742 
8743 #endif
8744 	uint32_t prev_acked = 0;
8745 	struct tcp_bbr *bbr;
8746 
8747 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8748 		/* Old ack, behind (or duplicate to) the last one rcv'd */
8749 		return (0);
8750 	}
8751 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
8752 		/* Above what we have sent? */
8753 		return (0);
8754 	}
8755 	if (__predict_false(tiwin == 0)) {
8756 		/* zero window */
8757 		return (0);
8758 	}
8759 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
8760 		/* We need a SYN or a FIN, unlikely.. */
8761 		return (0);
8762 	}
8763 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
8764 		/* Timestamp is behind .. old ack with seq wrap? */
8765 		return (0);
8766 	}
8767 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
8768 		/* Still recovering */
8769 		return (0);
8770 	}
8771 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8772 	if (__predict_false(bbr->r_ctl.rc_resend != NULL)) {
8773 		/* We are retransmitting */
8774 		return (0);
8775 	}
8776 	if (__predict_false(bbr->rc_in_persist != 0)) {
8777 		/* In persist mode */
8778 		return (0);
8779 	}
8780 	if (bbr->r_ctl.rc_sacked) {
8781 		/* We have sack holes on our scoreboard */
8782 		return (0);
8783 	}
8784 	/* Ok if we reach here, we can process a fast-ack */
8785 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8786 	sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
8787 	/*
8788 	 * We never detect loss in fast ack [we can't
8789 	 * have a sack and can't be in recovery so
8790 	 * we always pass 0 (nothing detected)].
8791 	 */
8792 	bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0);
8793 	/* Did the window get updated? */
8794 	if (tiwin != tp->snd_wnd) {
8795 		tp->snd_wnd = tiwin;
8796 		tp->snd_wl1 = th->th_seq;
8797 		if (tp->snd_wnd > tp->max_sndwnd)
8798 			tp->max_sndwnd = tp->snd_wnd;
8799 	}
8800 	/* Do we need to exit persists? */
8801 	if ((bbr->rc_in_persist != 0) &&
8802 	    (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8803 			       bbr_minseg(bbr)))) {
8804 		bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8805 		bbr->r_wanted_output = 1;
8806 	}
8807 	/* Do we need to enter persists? */
8808 	if ((bbr->rc_in_persist == 0) &&
8809 	    (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8810 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
8811 	    (tp->snd_max == tp->snd_una) &&
8812 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8813 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8814 		/* No send window.. we must enter persist */
8815 		bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8816 	}
8817 	/*
8818 	 * If last ACK falls within this segment's sequence numbers, record
8819 	 * the timestamp. NOTE that the test is modified according to the
8820 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8821 	 */
8822 	if ((to->to_flags & TOF_TS) != 0 &&
8823 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8824 		tp->ts_recent_age = bbr->r_ctl.rc_rcvtime;
8825 		tp->ts_recent = to->to_tsval;
8826 	}
8827 	/*
8828 	 * This is a pure ack for outstanding data.
8829 	 */
8830 	KMOD_TCPSTAT_INC(tcps_predack);
8831 
8832 	/*
8833 	 * "bad retransmit" recovery.
8834 	 */
8835 	if (tp->t_flags & TF_PREVVALID) {
8836 		tp->t_flags &= ~TF_PREVVALID;
8837 		if (tp->t_rxtshift == 1 &&
8838 		    (int)(ticks - tp->t_badrxtwin) < 0)
8839 			bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
8840 	}
8841 	/*
8842 	 * Recalculate the transmit timer / rtt.
8843 	 *
8844 	 * Some boxes send broken timestamp replies during the SYN+ACK
8845 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
8846 	 * and blow up the retransmit timer.
8847 	 */
8848 	acked = BYTES_THIS_ACK(tp, th);
8849 
8850 #ifdef TCP_HHOOK
8851 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
8852 	hhook_run_tcp_est_in(tp, th, to);
8853 #endif
8854 
8855 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
8856 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
8857 	sbdrop(&so->so_snd, acked);
8858 
8859 	if (SEQ_GT(th->th_ack, tp->snd_una))
8860 		bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
8861 	tp->snd_una = th->th_ack;
8862 	if (tp->snd_wnd < ctf_outstanding(tp))
8863 		/* The peer collapsed its window on us */
8864 		bbr_collapsed_window(bbr);
8865 	else if (bbr->rc_has_collapsed)
8866 		bbr_un_collapse_window(bbr);
8867 
8868 	if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8869 		tp->snd_recover = tp->snd_una;
8870 	}
8871 	bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0);
8872 	/*
8873 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
8874 	 */
8875 	tp->snd_wl2 = th->th_ack;
8876 	m_freem(m);
8877 	/*
8878 	 * If all outstanding data are acked, stop retransmit timer,
8879 	 * otherwise restart timer using current (possibly backed-off)
8880 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
8881 	 * If data are ready to send, let tcp_output decide between more
8882 	 * output or persist.
8883 	 */
8884 #ifdef TCPDEBUG
8885 	if (so->so_options & SO_DEBUG)
8886 		tcp_trace(TA_INPUT, ostate, tp,
8887 		    (void *)tcp_saveipgen,
8888 		    &tcp_savetcp, 0);
8889 #endif
8890 	/* Wake up the socket if we have room to write more */
8891 	sowwakeup(so);
8892 	if (tp->snd_una == tp->snd_max) {
8893 		/* Nothing left outstanding */
8894 		bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8895 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8896 			bbr->rc_tp->t_acktime = 0;
8897 		bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8898 		if (bbr->rc_in_persist == 0) {
8899 			bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8900 		}
8901 		sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8902 		bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8903 		/*
8904 		 * We invalidate the last ack here since we
8905 		 * don't want to transfer forward the time
8906 		 * for our sum's calculations.
8907 		 */
8908 		bbr->r_wanted_output = 1;
8909 	}
8910 	if (sbavail(&so->so_snd)) {
8911 		bbr->r_wanted_output = 1;
8912 	}
8913 	return (1);
8914 }
8915 
8916 /*
8917  * Return value of 1, the TCB is unlocked and most
8918  * likely gone, return value of 0, the TCB is still
8919  * locked.
8920  */
8921 static int
8922 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
8923     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8924     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
8925 {
8926 	int32_t todrop;
8927 	int32_t ourfinisacked = 0;
8928 	struct tcp_bbr *bbr;
8929 	int32_t ret_val = 0;
8930 
8931 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8932 	ctf_calc_rwin(so, tp);
8933 	/*
8934 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
8935 	 * SYN, drop the input. if seg contains a RST, then drop the
8936 	 * connection. if seg does not contain SYN, then drop it. Otherwise
8937 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
8938 	 * tp->irs if seg contains ack then advance tp->snd_una. BRR does
8939 	 * not support ECN so we will not say we are capable. if SYN has
8940 	 * been acked change to ESTABLISHED else SYN_RCVD state arrange for
8941 	 * segment to be acked (eventually) continue processing rest of
8942 	 * data/controls, beginning with URG
8943 	 */
8944 	if ((thflags & TH_ACK) &&
8945 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
8946 	    SEQ_GT(th->th_ack, tp->snd_max))) {
8947 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8948 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
8949 		return (1);
8950 	}
8951 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
8952 		TCP_PROBE5(connect__refused, NULL, tp,
8953 		    mtod(m, const char *), tp, th);
8954 		tp = tcp_drop(tp, ECONNREFUSED);
8955 		ctf_do_drop(m, tp);
8956 		return (1);
8957 	}
8958 	if (thflags & TH_RST) {
8959 		ctf_do_drop(m, tp);
8960 		return (1);
8961 	}
8962 	if (!(thflags & TH_SYN)) {
8963 		ctf_do_drop(m, tp);
8964 		return (1);
8965 	}
8966 	tp->irs = th->th_seq;
8967 	tcp_rcvseqinit(tp);
8968 	if (thflags & TH_ACK) {
8969 		int tfo_partial = 0;
8970 
8971 		KMOD_TCPSTAT_INC(tcps_connects);
8972 		soisconnected(so);
8973 #ifdef MAC
8974 		mac_socketpeer_set_from_mbuf(m, so);
8975 #endif
8976 		/* Do window scaling on this connection? */
8977 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
8978 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
8979 			tp->rcv_scale = tp->request_r_scale;
8980 		}
8981 		tp->rcv_adv += min(tp->rcv_wnd,
8982 		    TCP_MAXWIN << tp->rcv_scale);
8983 		/*
8984 		 * If not all the data that was sent in the TFO SYN
8985 		 * has been acked, resend the remainder right away.
8986 		 */
8987 		if (IS_FASTOPEN(tp->t_flags) &&
8988 		    (tp->snd_una != tp->snd_max)) {
8989 			tp->snd_nxt = th->th_ack;
8990 			tfo_partial = 1;
8991 		}
8992 		/*
8993 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
8994 		 * will be turned on later.
8995 		 */
8996 		if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) {
8997 			bbr->bbr_segs_rcvd += 1;
8998 			tp->t_flags |= TF_DELACK;
8999 			bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
9000 		} else {
9001 			bbr->r_wanted_output = 1;
9002 			tp->t_flags |= TF_ACKNOW;
9003 		}
9004 		if (SEQ_GT(th->th_ack, tp->iss)) {
9005 			/*
9006 			 * The SYN is acked
9007 			 * handle it specially.
9008 			 */
9009 			bbr_log_syn(tp, to);
9010 		}
9011 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
9012 			/*
9013 			 * We advance snd_una for the
9014 			 * fast open case. If th_ack is
9015 			 * acknowledging data beyond
9016 			 * snd_una we can't just call
9017 			 * ack-processing since the
9018 			 * data stream in our send-map
9019 			 * will start at snd_una + 1 (one
9020 			 * beyond the SYN). If its just
9021 			 * equal we don't need to do that
9022 			 * and there is no send_map.
9023 			 */
9024 			tp->snd_una++;
9025 		}
9026 		/*
9027 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
9028 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
9029 		 */
9030 		tp->t_starttime = ticks;
9031 		if (tp->t_flags & TF_NEEDFIN) {
9032 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
9033 			tp->t_flags &= ~TF_NEEDFIN;
9034 			thflags &= ~TH_SYN;
9035 		} else {
9036 			tcp_state_change(tp, TCPS_ESTABLISHED);
9037 			TCP_PROBE5(connect__established, NULL, tp,
9038 			    mtod(m, const char *), tp, th);
9039 			cc_conn_init(tp);
9040 		}
9041 	} else {
9042 		/*
9043 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
9044 		 * open.  If segment contains CC option and there is a
9045 		 * cached CC, apply TAO test. If it succeeds, connection is *
9046 		 * half-synchronized. Otherwise, do 3-way handshake:
9047 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
9048 		 * there was no CC option, clear cached CC value.
9049 		 */
9050 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
9051 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
9052 	}
9053 	INP_WLOCK_ASSERT(tp->t_inpcb);
9054 	/*
9055 	 * Advance th->th_seq to correspond to first data byte. If data,
9056 	 * trim to stay within window, dropping FIN if necessary.
9057 	 */
9058 	th->th_seq++;
9059 	if (tlen > tp->rcv_wnd) {
9060 		todrop = tlen - tp->rcv_wnd;
9061 		m_adj(m, -todrop);
9062 		tlen = tp->rcv_wnd;
9063 		thflags &= ~TH_FIN;
9064 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
9065 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
9066 	}
9067 	tp->snd_wl1 = th->th_seq - 1;
9068 	tp->rcv_up = th->th_seq;
9069 	/*
9070 	 * Client side of transaction: already sent SYN and data. If the
9071 	 * remote host used T/TCP to validate the SYN, our data will be
9072 	 * ACK'd; if so, enter normal data segment processing in the middle
9073 	 * of step 5, ack processing. Otherwise, goto step 6.
9074 	 */
9075 	if (thflags & TH_ACK) {
9076 		if ((to->to_flags & TOF_TS) != 0) {
9077 			uint32_t t, rtt;
9078 
9079 			t = tcp_tv_to_mssectick(&bbr->rc_tv);
9080 			if (TSTMP_GEQ(t, to->to_tsecr)) {
9081 				rtt = t - to->to_tsecr;
9082 				if (rtt == 0) {
9083 					rtt = 1;
9084 				}
9085 				rtt *= MS_IN_USEC;
9086 				tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9087 				apply_filter_min_small(&bbr->r_ctl.rc_rttprop,
9088 						       rtt, bbr->r_ctl.rc_rcvtime);
9089 			}
9090 		}
9091 		if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
9092 			return (ret_val);
9093 		/* We may have changed to FIN_WAIT_1 above */
9094 		if (tp->t_state == TCPS_FIN_WAIT_1) {
9095 			/*
9096 			 * In FIN_WAIT_1 STATE in addition to the processing
9097 			 * for the ESTABLISHED state if our FIN is now
9098 			 * acknowledged then enter FIN_WAIT_2.
9099 			 */
9100 			if (ourfinisacked) {
9101 				/*
9102 				 * If we can't receive any more data, then
9103 				 * closing user can proceed. Starting the
9104 				 * timer is contrary to the specification,
9105 				 * but if we don't get a FIN we'll hang
9106 				 * forever.
9107 				 *
9108 				 * XXXjl: we should release the tp also, and
9109 				 * use a compressed state.
9110 				 */
9111 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9112 					soisdisconnected(so);
9113 					tcp_timer_activate(tp, TT_2MSL,
9114 					    (tcp_fast_finwait2_recycle ?
9115 					    tcp_finwait2_timeout :
9116 					    TP_MAXIDLE(tp)));
9117 				}
9118 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
9119 			}
9120 		}
9121 	}
9122 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9123 	    tiwin, thflags, nxt_pkt));
9124 }
9125 
9126 /*
9127  * Return value of 1, the TCB is unlocked and most
9128  * likely gone, return value of 0, the TCB is still
9129  * locked.
9130  */
9131 static int
9132 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
9133 		struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9134 		uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9135 {
9136 	int32_t ourfinisacked = 0;
9137 	int32_t ret_val;
9138 	struct tcp_bbr *bbr;
9139 
9140 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9141 	ctf_calc_rwin(so, tp);
9142 	if ((thflags & TH_ACK) &&
9143 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
9144 	     SEQ_GT(th->th_ack, tp->snd_max))) {
9145 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9146 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9147 		return (1);
9148 	}
9149 	if (IS_FASTOPEN(tp->t_flags)) {
9150 		/*
9151 		 * When a TFO connection is in SYN_RECEIVED, the only valid
9152 		 * packets are the initial SYN, a retransmit/copy of the
9153 		 * initial SYN (possibly with a subset of the original
9154 		 * data), a valid ACK, a FIN, or a RST.
9155 		 */
9156 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
9157 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9158 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9159 			return (1);
9160 		} else if (thflags & TH_SYN) {
9161 			/* non-initial SYN is ignored */
9162 			if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
9163 			    (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
9164 			    (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
9165 				ctf_do_drop(m, NULL);
9166 				return (0);
9167 			}
9168 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
9169 			ctf_do_drop(m, NULL);
9170 			return (0);
9171 		}
9172 	}
9173 	if ((thflags & TH_RST) ||
9174 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9175 		return (ctf_process_rst(m, th, so, tp));
9176 	/*
9177 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9178 	 * it's less than ts_recent, drop it.
9179 	 */
9180 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9181 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9182 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9183 			return (ret_val);
9184 	}
9185 	/*
9186 	 * In the SYN-RECEIVED state, validate that the packet belongs to
9187 	 * this connection before trimming the data to fit the receive
9188 	 * window.  Check the sequence number versus IRS since we know the
9189 	 * sequence numbers haven't wrapped.  This is a partial fix for the
9190 	 * "LAND" DoS attack.
9191 	 */
9192 	if (SEQ_LT(th->th_seq, tp->irs)) {
9193 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9194 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9195 		return (1);
9196 	}
9197 	INP_WLOCK_ASSERT(tp->t_inpcb);
9198 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9199 		return (ret_val);
9200 	}
9201 	/*
9202 	 * If last ACK falls within this segment's sequence numbers, record
9203 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9204 	 * from the latest proposal of the tcplw@cray.com list (Braden
9205 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9206 	 * with our earlier PAWS tests, so this check should be solely
9207 	 * predicated on the sequence space of this segment. 3) That we
9208 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9209 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9210 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9211 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9212 	 * p.869. In such cases, we can still calculate the RTT correctly
9213 	 * when RCV.NXT == Last.ACK.Sent.
9214 	 */
9215 	if ((to->to_flags & TOF_TS) != 0 &&
9216 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9217 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9218 		    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9219 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9220 		tp->ts_recent = to->to_tsval;
9221 	}
9222 	tp->snd_wnd = tiwin;
9223 	/*
9224 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9225 	 * is on (half-synchronized state), then queue data for later
9226 	 * processing; else drop segment and return.
9227 	 */
9228 	if ((thflags & TH_ACK) == 0) {
9229 		if (IS_FASTOPEN(tp->t_flags)) {
9230 			cc_conn_init(tp);
9231 		}
9232 		return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9233 					 tiwin, thflags, nxt_pkt));
9234 	}
9235 	KMOD_TCPSTAT_INC(tcps_connects);
9236 	soisconnected(so);
9237 	/* Do window scaling? */
9238 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9239 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9240 		tp->rcv_scale = tp->request_r_scale;
9241 	}
9242 	/*
9243 	 * ok for the first time in lets see if we can use the ts to figure
9244 	 * out what the initial RTT was.
9245 	 */
9246 	if ((to->to_flags & TOF_TS) != 0) {
9247 		uint32_t t, rtt;
9248 
9249 		t = tcp_tv_to_mssectick(&bbr->rc_tv);
9250 		if (TSTMP_GEQ(t, to->to_tsecr)) {
9251 			rtt = t - to->to_tsecr;
9252 			if (rtt == 0) {
9253 				rtt = 1;
9254 			}
9255 			rtt *= MS_IN_USEC;
9256 			tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9257 			apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime);
9258 		}
9259 	}
9260 	/* Drop off any SYN in the send map (probably not there)  */
9261 	if (thflags & TH_ACK)
9262 		bbr_log_syn(tp, to);
9263 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
9264 
9265 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9266 		tp->t_tfo_pending = NULL;
9267 	}
9268 	/*
9269 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
9270 	 * FIN-WAIT-1
9271 	 */
9272 	tp->t_starttime = ticks;
9273 	if (tp->t_flags & TF_NEEDFIN) {
9274 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
9275 		tp->t_flags &= ~TF_NEEDFIN;
9276 	} else {
9277 		tcp_state_change(tp, TCPS_ESTABLISHED);
9278 		TCP_PROBE5(accept__established, NULL, tp,
9279 			   mtod(m, const char *), tp, th);
9280 		/*
9281 		 * TFO connections call cc_conn_init() during SYN
9282 		 * processing.  Calling it again here for such connections
9283 		 * is not harmless as it would undo the snd_cwnd reduction
9284 		 * that occurs when a TFO SYN|ACK is retransmitted.
9285 		 */
9286 		if (!IS_FASTOPEN(tp->t_flags))
9287 			cc_conn_init(tp);
9288 	}
9289 	/*
9290 	 * Account for the ACK of our SYN prior to
9291 	 * regular ACK processing below, except for
9292 	 * simultaneous SYN, which is handled later.
9293 	 */
9294 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
9295 		tp->snd_una++;
9296 	/*
9297 	 * If segment contains data or ACK, will call tcp_reass() later; if
9298 	 * not, do so now to pass queued data to user.
9299 	 */
9300 	if (tlen == 0 && (thflags & TH_FIN) == 0)
9301 		(void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9302 			(struct mbuf *)0);
9303 	tp->snd_wl1 = th->th_seq - 1;
9304 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9305 		return (ret_val);
9306 	}
9307 	if (tp->t_state == TCPS_FIN_WAIT_1) {
9308 		/* We could have went to FIN_WAIT_1 (or EST) above */
9309 		/*
9310 		 * In FIN_WAIT_1 STATE in addition to the processing for the
9311 		 * ESTABLISHED state if our FIN is now acknowledged then
9312 		 * enter FIN_WAIT_2.
9313 		 */
9314 		if (ourfinisacked) {
9315 			/*
9316 			 * If we can't receive any more data, then closing
9317 			 * user can proceed. Starting the timer is contrary
9318 			 * to the specification, but if we don't get a FIN
9319 			 * we'll hang forever.
9320 			 *
9321 			 * XXXjl: we should release the tp also, and use a
9322 			 * compressed state.
9323 			 */
9324 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9325 				soisdisconnected(so);
9326 				tcp_timer_activate(tp, TT_2MSL,
9327 						   (tcp_fast_finwait2_recycle ?
9328 						    tcp_finwait2_timeout :
9329 						    TP_MAXIDLE(tp)));
9330 			}
9331 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
9332 		}
9333 	}
9334 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9335 				 tiwin, thflags, nxt_pkt));
9336 }
9337 
9338 /*
9339  * Return value of 1, the TCB is unlocked and most
9340  * likely gone, return value of 0, the TCB is still
9341  * locked.
9342  */
9343 static int
9344 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9345     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9346     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9347 {
9348 	struct tcp_bbr *bbr;
9349 	int32_t ret_val;
9350 
9351 	/*
9352 	 * Header prediction: check for the two common cases of a
9353 	 * uni-directional data xfer.  If the packet has no control flags,
9354 	 * is in-sequence, the window didn't change and we're not
9355 	 * retransmitting, it's a candidate.  If the length is zero and the
9356 	 * ack moved forward, we're the sender side of the xfer.  Just free
9357 	 * the data acked & wake any higher level process that was blocked
9358 	 * waiting for space.  If the length is non-zero and the ack didn't
9359 	 * move, we're the receiver side.  If we're getting packets in-order
9360 	 * (the reassembly queue is empty), add the data toc The socket
9361 	 * buffer and note that we need a delayed ack. Make sure that the
9362 	 * hidden state-flags are also off. Since we check for
9363 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9364 	 */
9365 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9366 	if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) {
9367 		/*
9368 		 * If we have delived under 4 segments increase the initial
9369 		 * window if raised by the peer. We use this to determine
9370 		 * dynamic and static rwnd's at the end of a connection.
9371 		 */
9372 		bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd);
9373 	}
9374 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9375 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) &&
9376 	    __predict_true(SEGQ_EMPTY(tp)) &&
9377 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
9378 		if (tlen == 0) {
9379 			if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9380 			    tiwin, nxt_pkt, iptos)) {
9381 				return (0);
9382 			}
9383 		} else {
9384 			if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9385 			    tiwin, nxt_pkt)) {
9386 				return (0);
9387 			}
9388 		}
9389 	}
9390 	ctf_calc_rwin(so, tp);
9391 
9392 	if ((thflags & TH_RST) ||
9393 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9394 		return (ctf_process_rst(m, th, so, tp));
9395 	/*
9396 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9397 	 * synchronized state.
9398 	 */
9399 	if (thflags & TH_SYN) {
9400 		ctf_challenge_ack(m, th, tp, &ret_val);
9401 		return (ret_val);
9402 	}
9403 	/*
9404 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9405 	 * it's less than ts_recent, drop it.
9406 	 */
9407 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9408 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9409 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9410 			return (ret_val);
9411 	}
9412 	INP_WLOCK_ASSERT(tp->t_inpcb);
9413 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9414 		return (ret_val);
9415 	}
9416 	/*
9417 	 * If last ACK falls within this segment's sequence numbers, record
9418 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9419 	 * from the latest proposal of the tcplw@cray.com list (Braden
9420 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9421 	 * with our earlier PAWS tests, so this check should be solely
9422 	 * predicated on the sequence space of this segment. 3) That we
9423 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9424 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9425 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9426 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9427 	 * p.869. In such cases, we can still calculate the RTT correctly
9428 	 * when RCV.NXT == Last.ACK.Sent.
9429 	 */
9430 	if ((to->to_flags & TOF_TS) != 0 &&
9431 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9432 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9433 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9434 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9435 		tp->ts_recent = to->to_tsval;
9436 	}
9437 	/*
9438 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9439 	 * is on (half-synchronized state), then queue data for later
9440 	 * processing; else drop segment and return.
9441 	 */
9442 	if ((thflags & TH_ACK) == 0) {
9443 		if (tp->t_flags & TF_NEEDSYN) {
9444 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9445 			    tiwin, thflags, nxt_pkt));
9446 		} else if (tp->t_flags & TF_ACKNOW) {
9447 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9448 			bbr->r_wanted_output = 1;
9449 			return (ret_val);
9450 		} else {
9451 			ctf_do_drop(m, NULL);
9452 			return (0);
9453 		}
9454 	}
9455 	/*
9456 	 * Ack processing.
9457 	 */
9458 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9459 		return (ret_val);
9460 	}
9461 	if (sbavail(&so->so_snd)) {
9462 		if (ctf_progress_timeout_check(tp, true)) {
9463 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9464 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9465 			return (1);
9466 		}
9467 	}
9468 	/* State changes only happen in bbr_process_data() */
9469 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9470 	    tiwin, thflags, nxt_pkt));
9471 }
9472 
9473 /*
9474  * Return value of 1, the TCB is unlocked and most
9475  * likely gone, return value of 0, the TCB is still
9476  * locked.
9477  */
9478 static int
9479 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9480     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9481     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9482 {
9483 	struct tcp_bbr *bbr;
9484 	int32_t ret_val;
9485 
9486 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9487 	ctf_calc_rwin(so, tp);
9488 	if ((thflags & TH_RST) ||
9489 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9490 		return (ctf_process_rst(m, th, so, tp));
9491 	/*
9492 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9493 	 * synchronized state.
9494 	 */
9495 	if (thflags & TH_SYN) {
9496 		ctf_challenge_ack(m, th, tp, &ret_val);
9497 		return (ret_val);
9498 	}
9499 	/*
9500 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9501 	 * it's less than ts_recent, drop it.
9502 	 */
9503 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9504 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9505 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9506 			return (ret_val);
9507 	}
9508 	INP_WLOCK_ASSERT(tp->t_inpcb);
9509 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9510 		return (ret_val);
9511 	}
9512 	/*
9513 	 * If last ACK falls within this segment's sequence numbers, record
9514 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9515 	 * from the latest proposal of the tcplw@cray.com list (Braden
9516 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9517 	 * with our earlier PAWS tests, so this check should be solely
9518 	 * predicated on the sequence space of this segment. 3) That we
9519 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9520 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9521 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9522 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9523 	 * p.869. In such cases, we can still calculate the RTT correctly
9524 	 * when RCV.NXT == Last.ACK.Sent.
9525 	 */
9526 	if ((to->to_flags & TOF_TS) != 0 &&
9527 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9528 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9529 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9530 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9531 		tp->ts_recent = to->to_tsval;
9532 	}
9533 	/*
9534 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9535 	 * is on (half-synchronized state), then queue data for later
9536 	 * processing; else drop segment and return.
9537 	 */
9538 	if ((thflags & TH_ACK) == 0) {
9539 		if (tp->t_flags & TF_NEEDSYN) {
9540 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9541 			    tiwin, thflags, nxt_pkt));
9542 		} else if (tp->t_flags & TF_ACKNOW) {
9543 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9544 			bbr->r_wanted_output = 1;
9545 			return (ret_val);
9546 		} else {
9547 			ctf_do_drop(m, NULL);
9548 			return (0);
9549 		}
9550 	}
9551 	/*
9552 	 * Ack processing.
9553 	 */
9554 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9555 		return (ret_val);
9556 	}
9557 	if (sbavail(&so->so_snd)) {
9558 		if (ctf_progress_timeout_check(tp, true)) {
9559 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9560 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9561 			return (1);
9562 		}
9563 	}
9564 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9565 	    tiwin, thflags, nxt_pkt));
9566 }
9567 
9568 static int
9569 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr,
9570     struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so)
9571 {
9572 
9573 	if (bbr->rc_allow_data_af_clo == 0) {
9574 close_now:
9575 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9576 		/* tcp_close will kill the inp pre-log the Reset */
9577 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9578 		tp = tcp_close(tp);
9579 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
9580 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
9581 		return (1);
9582 	}
9583 	if (sbavail(&so->so_snd) == 0)
9584 		goto close_now;
9585 	/* Ok we allow data that is ignored and a followup reset */
9586 	tp->rcv_nxt = th->th_seq + *tlen;
9587 	tp->t_flags2 |= TF2_DROP_AF_DATA;
9588 	bbr->r_wanted_output = 1;
9589 	*tlen = 0;
9590 	return (0);
9591 }
9592 
9593 /*
9594  * Return value of 1, the TCB is unlocked and most
9595  * likely gone, return value of 0, the TCB is still
9596  * locked.
9597  */
9598 static int
9599 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9600     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9601     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9602 {
9603 	int32_t ourfinisacked = 0;
9604 	int32_t ret_val;
9605 	struct tcp_bbr *bbr;
9606 
9607 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9608 	ctf_calc_rwin(so, tp);
9609 	if ((thflags & TH_RST) ||
9610 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9611 		return (ctf_process_rst(m, th, so, tp));
9612 	/*
9613 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9614 	 * synchronized state.
9615 	 */
9616 	if (thflags & TH_SYN) {
9617 		ctf_challenge_ack(m, th, tp, &ret_val);
9618 		return (ret_val);
9619 	}
9620 	/*
9621 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9622 	 * it's less than ts_recent, drop it.
9623 	 */
9624 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9625 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9626 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9627 			return (ret_val);
9628 	}
9629 	INP_WLOCK_ASSERT(tp->t_inpcb);
9630 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9631 		return (ret_val);
9632 	}
9633 	/*
9634 	 * If new data are received on a connection after the user processes
9635 	 * are gone, then RST the other end.
9636 	 */
9637 	if ((so->so_state & SS_NOFDREF) && tlen) {
9638 		/*
9639 		 * We call a new function now so we might continue and setup
9640 		 * to reset at all data being ack'd.
9641 		 */
9642 		if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9643 			return (1);
9644 	}
9645 	/*
9646 	 * If last ACK falls within this segment's sequence numbers, record
9647 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9648 	 * from the latest proposal of the tcplw@cray.com list (Braden
9649 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9650 	 * with our earlier PAWS tests, so this check should be solely
9651 	 * predicated on the sequence space of this segment. 3) That we
9652 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9653 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9654 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9655 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9656 	 * p.869. In such cases, we can still calculate the RTT correctly
9657 	 * when RCV.NXT == Last.ACK.Sent.
9658 	 */
9659 	if ((to->to_flags & TOF_TS) != 0 &&
9660 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9661 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9662 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9663 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9664 		tp->ts_recent = to->to_tsval;
9665 	}
9666 	/*
9667 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9668 	 * is on (half-synchronized state), then queue data for later
9669 	 * processing; else drop segment and return.
9670 	 */
9671 	if ((thflags & TH_ACK) == 0) {
9672 		if (tp->t_flags & TF_NEEDSYN) {
9673 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9674 			    tiwin, thflags, nxt_pkt));
9675 		} else if (tp->t_flags & TF_ACKNOW) {
9676 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9677 			bbr->r_wanted_output = 1;
9678 			return (ret_val);
9679 		} else {
9680 			ctf_do_drop(m, NULL);
9681 			return (0);
9682 		}
9683 	}
9684 	/*
9685 	 * Ack processing.
9686 	 */
9687 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9688 		return (ret_val);
9689 	}
9690 	if (ourfinisacked) {
9691 		/*
9692 		 * If we can't receive any more data, then closing user can
9693 		 * proceed. Starting the timer is contrary to the
9694 		 * specification, but if we don't get a FIN we'll hang
9695 		 * forever.
9696 		 *
9697 		 * XXXjl: we should release the tp also, and use a
9698 		 * compressed state.
9699 		 */
9700 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9701 			soisdisconnected(so);
9702 			tcp_timer_activate(tp, TT_2MSL,
9703 			    (tcp_fast_finwait2_recycle ?
9704 			    tcp_finwait2_timeout :
9705 			    TP_MAXIDLE(tp)));
9706 		}
9707 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
9708 	}
9709 	if (sbavail(&so->so_snd)) {
9710 		if (ctf_progress_timeout_check(tp, true)) {
9711 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9712 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9713 			return (1);
9714 		}
9715 	}
9716 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9717 	    tiwin, thflags, nxt_pkt));
9718 }
9719 
9720 /*
9721  * Return value of 1, the TCB is unlocked and most
9722  * likely gone, return value of 0, the TCB is still
9723  * locked.
9724  */
9725 static int
9726 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9727     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9728     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9729 {
9730 	int32_t ourfinisacked = 0;
9731 	int32_t ret_val;
9732 	struct tcp_bbr *bbr;
9733 
9734 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9735 	ctf_calc_rwin(so, tp);
9736 	if ((thflags & TH_RST) ||
9737 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9738 		return (ctf_process_rst(m, th, so, tp));
9739 	/*
9740 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9741 	 * synchronized state.
9742 	 */
9743 	if (thflags & TH_SYN) {
9744 		ctf_challenge_ack(m, th, tp, &ret_val);
9745 		return (ret_val);
9746 	}
9747 	/*
9748 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9749 	 * it's less than ts_recent, drop it.
9750 	 */
9751 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9752 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9753 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9754 			return (ret_val);
9755 	}
9756 	INP_WLOCK_ASSERT(tp->t_inpcb);
9757 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9758 		return (ret_val);
9759 	}
9760 	/*
9761 	 * If new data are received on a connection after the user processes
9762 	 * are gone, then RST the other end.
9763 	 */
9764 	if ((so->so_state & SS_NOFDREF) && tlen) {
9765 		/*
9766 		 * We call a new function now so we might continue and setup
9767 		 * to reset at all data being ack'd.
9768 		 */
9769 		if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9770 			return (1);
9771 	}
9772 	/*
9773 	 * If last ACK falls within this segment's sequence numbers, record
9774 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9775 	 * from the latest proposal of the tcplw@cray.com list (Braden
9776 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9777 	 * with our earlier PAWS tests, so this check should be solely
9778 	 * predicated on the sequence space of this segment. 3) That we
9779 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9780 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9781 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9782 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9783 	 * p.869. In such cases, we can still calculate the RTT correctly
9784 	 * when RCV.NXT == Last.ACK.Sent.
9785 	 */
9786 	if ((to->to_flags & TOF_TS) != 0 &&
9787 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9788 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9789 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9790 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9791 		tp->ts_recent = to->to_tsval;
9792 	}
9793 	/*
9794 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9795 	 * is on (half-synchronized state), then queue data for later
9796 	 * processing; else drop segment and return.
9797 	 */
9798 	if ((thflags & TH_ACK) == 0) {
9799 		if (tp->t_flags & TF_NEEDSYN) {
9800 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9801 			    tiwin, thflags, nxt_pkt));
9802 		} else if (tp->t_flags & TF_ACKNOW) {
9803 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9804 			bbr->r_wanted_output = 1;
9805 			return (ret_val);
9806 		} else {
9807 			ctf_do_drop(m, NULL);
9808 			return (0);
9809 		}
9810 	}
9811 	/*
9812 	 * Ack processing.
9813 	 */
9814 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9815 		return (ret_val);
9816 	}
9817 	if (ourfinisacked) {
9818 		tcp_twstart(tp);
9819 		m_freem(m);
9820 		return (1);
9821 	}
9822 	if (sbavail(&so->so_snd)) {
9823 		if (ctf_progress_timeout_check(tp, true)) {
9824 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9825 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9826 			return (1);
9827 		}
9828 	}
9829 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9830 	    tiwin, thflags, nxt_pkt));
9831 }
9832 
9833 /*
9834  * Return value of 1, the TCB is unlocked and most
9835  * likely gone, return value of 0, the TCB is still
9836  * locked.
9837  */
9838 static int
9839 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9840     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9841     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9842 {
9843 	int32_t ourfinisacked = 0;
9844 	int32_t ret_val;
9845 	struct tcp_bbr *bbr;
9846 
9847 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9848 	ctf_calc_rwin(so, tp);
9849 	if ((thflags & TH_RST) ||
9850 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9851 		return (ctf_process_rst(m, th, so, tp));
9852 	/*
9853 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9854 	 * synchronized state.
9855 	 */
9856 	if (thflags & TH_SYN) {
9857 		ctf_challenge_ack(m, th, tp, &ret_val);
9858 		return (ret_val);
9859 	}
9860 	/*
9861 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9862 	 * it's less than ts_recent, drop it.
9863 	 */
9864 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9865 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9866 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9867 			return (ret_val);
9868 	}
9869 	INP_WLOCK_ASSERT(tp->t_inpcb);
9870 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9871 		return (ret_val);
9872 	}
9873 	/*
9874 	 * If new data are received on a connection after the user processes
9875 	 * are gone, then RST the other end.
9876 	 */
9877 	if ((so->so_state & SS_NOFDREF) && tlen) {
9878 		/*
9879 		 * We call a new function now so we might continue and setup
9880 		 * to reset at all data being ack'd.
9881 		 */
9882 		if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9883 			return (1);
9884 	}
9885 	/*
9886 	 * If last ACK falls within this segment's sequence numbers, record
9887 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9888 	 * from the latest proposal of the tcplw@cray.com list (Braden
9889 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9890 	 * with our earlier PAWS tests, so this check should be solely
9891 	 * predicated on the sequence space of this segment. 3) That we
9892 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9893 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9894 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9895 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9896 	 * p.869. In such cases, we can still calculate the RTT correctly
9897 	 * when RCV.NXT == Last.ACK.Sent.
9898 	 */
9899 	if ((to->to_flags & TOF_TS) != 0 &&
9900 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9901 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9902 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9903 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9904 		tp->ts_recent = to->to_tsval;
9905 	}
9906 	/*
9907 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9908 	 * is on (half-synchronized state), then queue data for later
9909 	 * processing; else drop segment and return.
9910 	 */
9911 	if ((thflags & TH_ACK) == 0) {
9912 		if (tp->t_flags & TF_NEEDSYN) {
9913 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9914 			    tiwin, thflags, nxt_pkt));
9915 		} else if (tp->t_flags & TF_ACKNOW) {
9916 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9917 			bbr->r_wanted_output = 1;
9918 			return (ret_val);
9919 		} else {
9920 			ctf_do_drop(m, NULL);
9921 			return (0);
9922 		}
9923 	}
9924 	/*
9925 	 * case TCPS_LAST_ACK: Ack processing.
9926 	 */
9927 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9928 		return (ret_val);
9929 	}
9930 	if (ourfinisacked) {
9931 		tp = tcp_close(tp);
9932 		ctf_do_drop(m, tp);
9933 		return (1);
9934 	}
9935 	if (sbavail(&so->so_snd)) {
9936 		if (ctf_progress_timeout_check(tp, true)) {
9937 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9938 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9939 			return (1);
9940 		}
9941 	}
9942 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9943 	    tiwin, thflags, nxt_pkt));
9944 }
9945 
9946 
9947 /*
9948  * Return value of 1, the TCB is unlocked and most
9949  * likely gone, return value of 0, the TCB is still
9950  * locked.
9951  */
9952 static int
9953 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
9954     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9955     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9956 {
9957 	int32_t ourfinisacked = 0;
9958 	int32_t ret_val;
9959 	struct tcp_bbr *bbr;
9960 
9961 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9962 	ctf_calc_rwin(so, tp);
9963 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
9964 	if ((thflags & TH_RST) ||
9965 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9966 		return (ctf_process_rst(m, th, so, tp));
9967 
9968 	/*
9969 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9970 	 * synchronized state.
9971 	 */
9972 	if (thflags & TH_SYN) {
9973 		ctf_challenge_ack(m, th, tp, &ret_val);
9974 		return (ret_val);
9975 	}
9976 	INP_WLOCK_ASSERT(tp->t_inpcb);
9977 	/*
9978 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9979 	 * it's less than ts_recent, drop it.
9980 	 */
9981 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9982 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9983 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9984 			return (ret_val);
9985 	}
9986 	INP_WLOCK_ASSERT(tp->t_inpcb);
9987 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9988 		return (ret_val);
9989 	}
9990 	/*
9991 	 * If new data are received on a connection after the user processes
9992 	 * are gone, then we may RST the other end depending on the outcome
9993 	 * of bbr_check_data_after_close.
9994 	 */
9995 	if ((so->so_state & SS_NOFDREF) &&
9996 	    tlen) {
9997 		/*
9998 		 * We call a new function now so we might continue and setup
9999 		 * to reset at all data being ack'd.
10000 		 */
10001 		if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
10002 			return (1);
10003 	}
10004 	INP_WLOCK_ASSERT(tp->t_inpcb);
10005 	/*
10006 	 * If last ACK falls within this segment's sequence numbers, record
10007 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
10008 	 * from the latest proposal of the tcplw@cray.com list (Braden
10009 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
10010 	 * with our earlier PAWS tests, so this check should be solely
10011 	 * predicated on the sequence space of this segment. 3) That we
10012 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10013 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10014 	 * SEG.Len, This modified check allows us to overcome RFC1323's
10015 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10016 	 * p.869. In such cases, we can still calculate the RTT correctly
10017 	 * when RCV.NXT == Last.ACK.Sent.
10018 	 */
10019 	INP_WLOCK_ASSERT(tp->t_inpcb);
10020 	if ((to->to_flags & TOF_TS) != 0 &&
10021 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
10022 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
10023 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
10024 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
10025 		tp->ts_recent = to->to_tsval;
10026 	}
10027 	/*
10028 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
10029 	 * is on (half-synchronized state), then queue data for later
10030 	 * processing; else drop segment and return.
10031 	 */
10032 	if ((thflags & TH_ACK) == 0) {
10033 		if (tp->t_flags & TF_NEEDSYN) {
10034 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
10035 			    tiwin, thflags, nxt_pkt));
10036 		} else if (tp->t_flags & TF_ACKNOW) {
10037 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
10038 			bbr->r_wanted_output = 1;
10039 			return (ret_val);
10040 		} else {
10041 			ctf_do_drop(m, NULL);
10042 			return (0);
10043 		}
10044 	}
10045 	/*
10046 	 * Ack processing.
10047 	 */
10048 	INP_WLOCK_ASSERT(tp->t_inpcb);
10049 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
10050 		return (ret_val);
10051 	}
10052 	if (sbavail(&so->so_snd)) {
10053 		if (ctf_progress_timeout_check(tp, true)) {
10054 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
10055 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10056 			return (1);
10057 		}
10058 	}
10059 	INP_WLOCK_ASSERT(tp->t_inpcb);
10060 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
10061 	    tiwin, thflags, nxt_pkt));
10062 }
10063 
10064 static void
10065 bbr_stop_all_timers(struct tcpcb *tp)
10066 {
10067 	struct tcp_bbr *bbr;
10068 
10069 	/*
10070 	 * Assure no timers are running.
10071 	 */
10072 	if (tcp_timer_active(tp, TT_PERSIST)) {
10073 		/* We enter in persists, set the flag appropriately */
10074 		bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10075 		bbr->rc_in_persist = 1;
10076 	}
10077 	tcp_timer_suspend(tp, TT_PERSIST);
10078 	tcp_timer_suspend(tp, TT_REXMT);
10079 	tcp_timer_suspend(tp, TT_KEEP);
10080 	tcp_timer_suspend(tp, TT_DELACK);
10081 }
10082 
10083 static void
10084 bbr_google_mode_on(struct tcp_bbr *bbr)
10085 {
10086 	bbr->rc_use_google = 1;
10087 	bbr->rc_no_pacing = 0;
10088 	bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10089 	bbr->r_use_policer = bbr_policer_detection_enabled;
10090 	bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10091 	bbr->bbr_use_rack_cheat = 0;
10092 	bbr->r_ctl.rc_incr_tmrs = 0;
10093 	bbr->r_ctl.rc_inc_tcp_oh = 0;
10094 	bbr->r_ctl.rc_inc_ip_oh = 0;
10095 	bbr->r_ctl.rc_inc_enet_oh = 0;
10096 	reset_time(&bbr->r_ctl.rc_delrate,
10097 		   BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10098 	reset_time_small(&bbr->r_ctl.rc_rttprop,
10099 			 (11 * USECS_IN_SECOND));
10100 	tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10101 }
10102 
10103 static void
10104 bbr_google_mode_off(struct tcp_bbr *bbr)
10105 {
10106 	bbr->rc_use_google = 0;
10107 	bbr->r_ctl.bbr_google_discount = 0;
10108 	bbr->no_pacing_until = bbr_no_pacing_until;
10109 	bbr->r_use_policer = 0;
10110 	if (bbr->no_pacing_until)
10111 		bbr->rc_no_pacing = 1;
10112 	else
10113 		bbr->rc_no_pacing = 0;
10114 	if (bbr_use_rack_resend_cheat)
10115 		bbr->bbr_use_rack_cheat = 1;
10116 	else
10117 		bbr->bbr_use_rack_cheat = 0;
10118 	if (bbr_incr_timers)
10119 		bbr->r_ctl.rc_incr_tmrs = 1;
10120 	else
10121 		bbr->r_ctl.rc_incr_tmrs = 0;
10122 	if (bbr_include_tcp_oh)
10123 		bbr->r_ctl.rc_inc_tcp_oh = 1;
10124 	else
10125 		bbr->r_ctl.rc_inc_tcp_oh = 0;
10126 	if (bbr_include_ip_oh)
10127 		bbr->r_ctl.rc_inc_ip_oh = 1;
10128 	else
10129 		bbr->r_ctl.rc_inc_ip_oh = 0;
10130 	if (bbr_include_enet_oh)
10131 		bbr->r_ctl.rc_inc_enet_oh = 1;
10132 	else
10133 		bbr->r_ctl.rc_inc_enet_oh = 0;
10134 	bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10135 	reset_time(&bbr->r_ctl.rc_delrate,
10136 		   bbr_num_pktepo_for_del_limit);
10137 	reset_time_small(&bbr->r_ctl.rc_rttprop,
10138 			 (bbr_filter_len_sec * USECS_IN_SECOND));
10139 	tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10140 }
10141 /*
10142  * Return 0 on success, non-zero on failure
10143  * which indicates the error (usually no memory).
10144  */
10145 static int
10146 bbr_init(struct tcpcb *tp)
10147 {
10148 	struct tcp_bbr *bbr = NULL;
10149 	struct inpcb *inp;
10150 	uint32_t cts;
10151 
10152 	tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO));
10153 	if (tp->t_fb_ptr == NULL) {
10154 		/*
10155 		 * We need to allocate memory but cant. The INP and INP_INFO
10156 		 * locks and they are recusive (happens during setup. So a
10157 		 * scheme to drop the locks fails :(
10158 		 *
10159 		 */
10160 		return (ENOMEM);
10161 	}
10162 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10163 	bbr->rtt_valid = 0;
10164 	inp = tp->t_inpcb;
10165 	inp->inp_flags2 |= INP_CANNOT_DO_ECN;
10166 	inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
10167 	TAILQ_INIT(&bbr->r_ctl.rc_map);
10168 	TAILQ_INIT(&bbr->r_ctl.rc_free);
10169 	TAILQ_INIT(&bbr->r_ctl.rc_tmap);
10170 	bbr->rc_tp = tp;
10171 	if (tp->t_inpcb) {
10172 		bbr->rc_inp = tp->t_inpcb;
10173 	}
10174 	cts = tcp_get_usecs(&bbr->rc_tv);
10175 	tp->t_acktime = 0;
10176 	bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close;
10177 	bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade;
10178 	bbr->rc_tlp_threshold = bbr_tlp_thresh;
10179 	bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh;
10180 	bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay;
10181 	bbr->r_ctl.rc_min_to = bbr_min_to;
10182 	bbr->rc_bbr_state = BBR_STATE_STARTUP;
10183 	bbr->r_ctl.bbr_lost_at_state = 0;
10184 	bbr->r_ctl.rc_lost_at_startup = 0;
10185 	bbr->rc_all_timers_stopped = 0;
10186 	bbr->r_ctl.rc_bbr_lastbtlbw = 0;
10187 	bbr->r_ctl.rc_pkt_epoch_del = 0;
10188 	bbr->r_ctl.rc_pkt_epoch = 0;
10189 	bbr->r_ctl.rc_lowest_rtt = 0xffffffff;
10190 	bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain;
10191 	bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
10192 	bbr->r_ctl.rc_went_idle_time = cts;
10193 	bbr->rc_pacer_started = cts;
10194 	bbr->r_ctl.rc_pkt_epoch_time = cts;
10195 	bbr->r_ctl.rc_rcvtime = cts;
10196 	bbr->r_ctl.rc_bbr_state_time = cts;
10197 	bbr->r_ctl.rc_del_time = cts;
10198 	bbr->r_ctl.rc_tlp_rxt_last_time = cts;
10199 	bbr->r_ctl.last_in_probertt = cts;
10200 	bbr->skip_gain = 0;
10201 	bbr->gain_is_limited = 0;
10202 	bbr->no_pacing_until = bbr_no_pacing_until;
10203 	if (bbr->no_pacing_until)
10204 		bbr->rc_no_pacing = 1;
10205 	if (bbr_use_google_algo) {
10206 		bbr->rc_no_pacing = 0;
10207 		bbr->rc_use_google = 1;
10208 		bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10209 		bbr->r_use_policer = bbr_policer_detection_enabled;
10210 	} else {
10211 		bbr->rc_use_google = 0;
10212 		bbr->r_ctl.bbr_google_discount = 0;
10213 		bbr->r_use_policer = 0;
10214 	}
10215 	if (bbr_ts_limiting)
10216 		bbr->rc_use_ts_limit = 1;
10217 	else
10218 		bbr->rc_use_ts_limit = 0;
10219 	if (bbr_ts_can_raise)
10220 		bbr->ts_can_raise = 1;
10221 	else
10222 		bbr->ts_can_raise = 0;
10223 	if (V_tcp_delack_enabled == 1)
10224 		tp->t_delayed_ack = 2;
10225 	else if (V_tcp_delack_enabled == 0)
10226 		tp->t_delayed_ack = 0;
10227 	else if (V_tcp_delack_enabled < 100)
10228 		tp->t_delayed_ack = V_tcp_delack_enabled;
10229 	else
10230 		tp->t_delayed_ack = 2;
10231 	if (bbr->rc_use_google == 0)
10232 		bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10233 	else
10234 		bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10235 	bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms;
10236 	bbr->rc_max_rto_sec = bbr_rto_max_sec;
10237 	bbr->rc_init_win = bbr_def_init_win;
10238 	if (tp->t_flags & TF_REQ_TSTMP)
10239 		bbr->rc_last_options = TCP_TS_OVERHEAD;
10240 	bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options;
10241 	bbr->r_ctl.rc_high_rwnd = tp->snd_wnd;
10242 	bbr->r_init_rtt = 1;
10243 
10244 	counter_u64_add(bbr_flows_nohdwr_pacing, 1);
10245 	if (bbr_allow_hdwr_pacing)
10246 		bbr->bbr_hdw_pace_ena = 1;
10247 	else
10248 		bbr->bbr_hdw_pace_ena = 0;
10249 	if (bbr_sends_full_iwnd)
10250 		bbr->bbr_init_win_cheat = 1;
10251 	else
10252 		bbr->bbr_init_win_cheat = 0;
10253 	bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max;
10254 	bbr->r_ctl.rc_drain_pg = bbr_drain_gain;
10255 	bbr->r_ctl.rc_startup_pg = bbr_high_gain;
10256 	bbr->rc_loss_exit = bbr_exit_startup_at_loss;
10257 	bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain;
10258 	bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second;
10259 	bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar;
10260 	bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max;
10261 	bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor;
10262 	bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min;
10263 	bbr->r_ctl.bbr_cross_over = bbr_cross_over;
10264 	bbr->r_ctl.rc_rtt_shrinks = cts;
10265 	if (bbr->rc_use_google) {
10266 		setup_time_filter(&bbr->r_ctl.rc_delrate,
10267 				  FILTER_TYPE_MAX,
10268 				  BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10269 		setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10270 					FILTER_TYPE_MIN, (11 * USECS_IN_SECOND));
10271 	} else {
10272 		setup_time_filter(&bbr->r_ctl.rc_delrate,
10273 				  FILTER_TYPE_MAX,
10274 				  bbr_num_pktepo_for_del_limit);
10275 		setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10276 					FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND));
10277 	}
10278 	bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0);
10279 	if (bbr_uses_idle_restart)
10280 		bbr->rc_use_idle_restart = 1;
10281 	else
10282 		bbr->rc_use_idle_restart = 0;
10283 	bbr->r_ctl.rc_bbr_cur_del_rate = 0;
10284 	bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps;
10285 	if (bbr_resends_use_tso)
10286 		bbr->rc_resends_use_tso = 1;
10287 #ifdef NETFLIX_PEAKRATE
10288 	tp->t_peakrate_thr = tp->t_maxpeakrate;
10289 #endif
10290 	if (tp->snd_una != tp->snd_max) {
10291 		/* Create a send map for the current outstanding data */
10292 		struct bbr_sendmap *rsm;
10293 
10294 		rsm = bbr_alloc(bbr);
10295 		if (rsm == NULL) {
10296 			uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10297 			tp->t_fb_ptr = NULL;
10298 			return (ENOMEM);
10299 		}
10300 		rsm->r_flags = BBR_OVERMAX;
10301 		rsm->r_tim_lastsent[0] = cts;
10302 		rsm->r_rtr_cnt = 1;
10303 		rsm->r_rtr_bytes = 0;
10304 		rsm->r_start = tp->snd_una;
10305 		rsm->r_end = tp->snd_max;
10306 		rsm->r_dupack = 0;
10307 		rsm->r_delivered = bbr->r_ctl.rc_delivered;
10308 		rsm->r_ts_valid = 0;
10309 		rsm->r_del_ack_ts = tp->ts_recent;
10310 		rsm->r_del_time = cts;
10311 		if (bbr->r_ctl.r_app_limited_until)
10312 			rsm->r_app_limited = 1;
10313 		else
10314 			rsm->r_app_limited = 0;
10315 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
10316 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
10317 		rsm->r_in_tmap = 1;
10318 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
10319 			rsm->r_bbr_state = bbr_state_val(bbr);
10320 		else
10321 			rsm->r_bbr_state = 8;
10322 	}
10323 	if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0))
10324 		bbr->bbr_use_rack_cheat = 1;
10325 	if (bbr_incr_timers && (bbr->rc_use_google == 0))
10326 		bbr->r_ctl.rc_incr_tmrs = 1;
10327 	if (bbr_include_tcp_oh && (bbr->rc_use_google == 0))
10328 		bbr->r_ctl.rc_inc_tcp_oh = 1;
10329 	if (bbr_include_ip_oh && (bbr->rc_use_google == 0))
10330 		bbr->r_ctl.rc_inc_ip_oh = 1;
10331 	if (bbr_include_enet_oh && (bbr->rc_use_google == 0))
10332 		bbr->r_ctl.rc_inc_enet_oh = 1;
10333 
10334 	bbr_log_type_statechange(bbr, cts, __LINE__);
10335 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
10336 	    (tp->t_srtt)) {
10337 		uint32_t rtt;
10338 
10339 		rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
10340 		apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
10341 	}
10342 	/* announce the settings and state */
10343 	bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT);
10344 	tcp_bbr_tso_size_check(bbr, cts);
10345 	/*
10346 	 * Now call the generic function to start a timer. This will place
10347 	 * the TCB on the hptsi wheel if a timer is needed with appropriate
10348 	 * flags.
10349 	 */
10350 	bbr_stop_all_timers(tp);
10351 	bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0);
10352 	return (0);
10353 }
10354 
10355 /*
10356  * Return 0 if we can accept the connection. Return
10357  * non-zero if we can't handle the connection. A EAGAIN
10358  * means you need to wait until the connection is up.
10359  * a EADDRNOTAVAIL means we can never handle the connection
10360  * (no SACK).
10361  */
10362 static int
10363 bbr_handoff_ok(struct tcpcb *tp)
10364 {
10365 	if ((tp->t_state == TCPS_CLOSED) ||
10366 	    (tp->t_state == TCPS_LISTEN)) {
10367 		/* Sure no problem though it may not stick */
10368 		return (0);
10369 	}
10370 	if ((tp->t_state == TCPS_SYN_SENT) ||
10371 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
10372 		/*
10373 		 * We really don't know you have to get to ESTAB or beyond
10374 		 * to tell.
10375 		 */
10376 		return (EAGAIN);
10377 	}
10378 	if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) {
10379 		return (0);
10380 	}
10381 	/*
10382 	 * If we reach here we don't do SACK on this connection so we can
10383 	 * never do rack.
10384 	 */
10385 	return (EINVAL);
10386 }
10387 
10388 static void
10389 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10390 {
10391 	if (tp->t_fb_ptr) {
10392 		uint32_t calc;
10393 		struct tcp_bbr *bbr;
10394 		struct bbr_sendmap *rsm;
10395 
10396 		bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10397 		if (bbr->r_ctl.crte)
10398 			tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
10399 		bbr_log_flowend(bbr);
10400 		bbr->rc_tp = NULL;
10401 		if (tp->t_inpcb) {
10402 			/* Backout any flags2 we applied */
10403 			tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN;
10404 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10405 			tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
10406 		}
10407 		if (bbr->bbr_hdrw_pacing)
10408 			counter_u64_add(bbr_flows_whdwr_pacing, -1);
10409 		else
10410 			counter_u64_add(bbr_flows_nohdwr_pacing, -1);
10411 		rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10412 		while (rsm) {
10413 			TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
10414 			uma_zfree(bbr_zone, rsm);
10415 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10416 		}
10417 		rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10418 		while (rsm) {
10419 			TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
10420 			uma_zfree(bbr_zone, rsm);
10421 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10422 		}
10423 		calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd;
10424 		if (calc > (bbr->r_ctl.rc_init_rwnd / 10))
10425 			BBR_STAT_INC(bbr_dynamic_rwnd);
10426 		else
10427 			BBR_STAT_INC(bbr_static_rwnd);
10428 		bbr->r_ctl.rc_free_cnt = 0;
10429 		uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10430 		tp->t_fb_ptr = NULL;
10431 	}
10432 	/* Make sure snd_nxt is correctly set */
10433 	tp->snd_nxt = tp->snd_max;
10434 }
10435 
10436 static void
10437 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win)
10438 {
10439 	switch (tp->t_state) {
10440 	case TCPS_SYN_SENT:
10441 		bbr->r_state = TCPS_SYN_SENT;
10442 		bbr->r_substate = bbr_do_syn_sent;
10443 		break;
10444 	case TCPS_SYN_RECEIVED:
10445 		bbr->r_state = TCPS_SYN_RECEIVED;
10446 		bbr->r_substate = bbr_do_syn_recv;
10447 		break;
10448 	case TCPS_ESTABLISHED:
10449 		bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd);
10450 		bbr->r_state = TCPS_ESTABLISHED;
10451 		bbr->r_substate = bbr_do_established;
10452 		break;
10453 	case TCPS_CLOSE_WAIT:
10454 		bbr->r_state = TCPS_CLOSE_WAIT;
10455 		bbr->r_substate = bbr_do_close_wait;
10456 		break;
10457 	case TCPS_FIN_WAIT_1:
10458 		bbr->r_state = TCPS_FIN_WAIT_1;
10459 		bbr->r_substate = bbr_do_fin_wait_1;
10460 		break;
10461 	case TCPS_CLOSING:
10462 		bbr->r_state = TCPS_CLOSING;
10463 		bbr->r_substate = bbr_do_closing;
10464 		break;
10465 	case TCPS_LAST_ACK:
10466 		bbr->r_state = TCPS_LAST_ACK;
10467 		bbr->r_substate = bbr_do_lastack;
10468 		break;
10469 	case TCPS_FIN_WAIT_2:
10470 		bbr->r_state = TCPS_FIN_WAIT_2;
10471 		bbr->r_substate = bbr_do_fin_wait_2;
10472 		break;
10473 	case TCPS_LISTEN:
10474 	case TCPS_CLOSED:
10475 	case TCPS_TIME_WAIT:
10476 	default:
10477 		break;
10478 	};
10479 }
10480 
10481 static void
10482 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog)
10483 {
10484 	/*
10485 	 * Now what state are we going into now? Is there adjustments
10486 	 * needed?
10487 	 */
10488 	int32_t old_state, old_gain;
10489 
10490 
10491 	old_state = bbr_state_val(bbr);
10492 	old_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
10493 	if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) {
10494 		/* Save the lowest srtt we saw in our end of the sub-state */
10495 		bbr->rc_hit_state_1 = 0;
10496 		if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff)
10497 			bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state;
10498 	}
10499 	bbr->rc_bbr_substate++;
10500 	if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) {
10501 		/* Cycle back to first state-> gain */
10502 		bbr->rc_bbr_substate = 0;
10503 	}
10504 	if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10505 		/*
10506 		 * We enter the gain(5/4) cycle (possibly less if
10507 		 * shallow buffer detection is enabled)
10508 		 */
10509 		if (bbr->skip_gain) {
10510 			/*
10511 			 * Hardware pacing has set our rate to
10512 			 * the max and limited our b/w just
10513 			 * do level i.e. no gain.
10514 			 */
10515 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1];
10516 		} else if (bbr->gain_is_limited &&
10517 			   bbr->bbr_hdrw_pacing &&
10518 			   bbr->r_ctl.crte) {
10519 			/*
10520 			 * We can't gain above the hardware pacing
10521 			 * rate which is less than our rate + the gain
10522 			 * calculate the gain needed to reach the hardware
10523 			 * pacing rate..
10524 			 */
10525 			uint64_t bw, rate, gain_calc;
10526 
10527 			bw = bbr_get_bw(bbr);
10528 			rate = bbr->r_ctl.crte->rate;
10529 			if ((rate > bw) &&
10530 			    (((bw *  (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) {
10531 				gain_calc = (rate * BBR_UNIT) / bw;
10532 				if (gain_calc < BBR_UNIT)
10533 					gain_calc = BBR_UNIT;
10534 				bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc;
10535 			} else {
10536 				bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10537 			}
10538 		} else
10539 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10540 		if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) {
10541 			bbr->r_ctl.rc_bbr_state_atflight = cts;
10542 		} else
10543 			bbr->r_ctl.rc_bbr_state_atflight = 0;
10544 	} else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10545 		bbr->rc_hit_state_1 = 1;
10546 		bbr->r_ctl.rc_exta_time_gd = 0;
10547 		bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10548 						     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10549 		if (bbr_state_drain_2_tar) {
10550 			bbr->r_ctl.rc_bbr_state_atflight = 0;
10551 		} else
10552 			bbr->r_ctl.rc_bbr_state_atflight = cts;
10553 		bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN];
10554 	} else {
10555 		/* All other cycles hit here 2-7 */
10556 		if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) {
10557 			if (bbr_sub_drain_slam_cwnd &&
10558 			    (bbr->rc_use_google == 0) &&
10559 			    (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10560 				bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10561 				bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10562 			}
10563 			if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP))
10564 				bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) -
10565 							       bbr_get_rtt(bbr, BBR_RTT_PROP));
10566 			else
10567 				bbr->r_ctl.rc_exta_time_gd = 0;
10568 			if (bbr->r_ctl.rc_exta_time_gd) {
10569 				bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd;
10570 				/* Now chop up the time for each state (div by 7) */
10571 				bbr->r_ctl.rc_level_state_extra /= 7;
10572 				if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) {
10573 					/* Add a randomization */
10574 					bbr_randomize_extra_state_time(bbr);
10575 				}
10576 			}
10577 		}
10578 		bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10579 		bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)];
10580 	}
10581 	if (bbr->rc_use_google) {
10582 		bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10583 	}
10584 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10585 	bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10586 	if (dolog)
10587 		bbr_log_type_statechange(bbr, cts, line);
10588 
10589 	if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10590 		uint32_t time_in;
10591 
10592 		time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10593 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
10594 			counter_u64_add(bbr_state_time[(old_state + 5)], time_in);
10595 		} else {
10596 			counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10597 		}
10598 	}
10599 	bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
10600 	bbr_set_state_target(bbr, __LINE__);
10601 	if (bbr_sub_drain_slam_cwnd &&
10602 	    (bbr->rc_use_google == 0) &&
10603 	    (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10604 		/* Slam down the cwnd */
10605 		bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10606 		bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10607 		if (bbr_sub_drain_app_limit) {
10608 			/* Go app limited if we are on a long drain */
10609 			bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered +
10610 							  ctf_flight_size(bbr->rc_tp,
10611 							      (bbr->r_ctl.rc_sacked +
10612 							       bbr->r_ctl.rc_lost_bytes)));
10613 		}
10614 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10615 	}
10616 	if (bbr->rc_lt_use_bw) {
10617 		/* In policed mode we clamp pacing_gain to BBR_UNIT */
10618 		bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10619 	}
10620 	/* Google changes TSO size every cycle */
10621 	if (bbr->rc_use_google)
10622 		tcp_bbr_tso_size_check(bbr, cts);
10623 	bbr->r_ctl.gain_epoch = cts;
10624 	bbr->r_ctl.rc_bbr_state_time = cts;
10625 	bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch;
10626 }
10627 
10628 static void
10629 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10630 {
10631 	if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) &&
10632 	    (google_allow_early_out == 1) &&
10633 	    (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) {
10634 		/* We have reached out target flight size possibly early */
10635 		goto change_state;
10636 	}
10637 	if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10638 		return;
10639 	}
10640 	if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) {
10641 		/*
10642 		 * Must be a rttProp movement forward before
10643 		 * we can change states.
10644 		 */
10645 		return;
10646 	}
10647 	if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10648 		/*
10649 		 * The needed time has passed but for
10650 		 * the gain cycle extra rules apply:
10651 		 * 1) If we have seen loss, we exit
10652 		 * 2) If we have not reached the target
10653 		 *    we stay in GAIN (gain-to-target).
10654 		 */
10655 		if (google_consider_lost && losses)
10656 			goto change_state;
10657 		if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) {
10658 			return;
10659 		}
10660 	}
10661 change_state:
10662 	/* For gain we must reach our target, all others last 1 rttProp */
10663 	bbr_substate_change(bbr, cts, __LINE__, 1);
10664 }
10665 
10666 static void
10667 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10668 {
10669 	uint32_t flight, bbr_cur_cycle_time;
10670 
10671 	if (bbr->rc_use_google) {
10672 		bbr_set_probebw_google_gains(bbr, cts, losses);
10673 		return;
10674 	}
10675 	if (cts == 0) {
10676 		/*
10677 		 * Never alow cts to be 0 we
10678 		 * do this so we can judge if
10679 		 * we have set a timestamp.
10680 		 */
10681 		cts = 1;
10682 	}
10683 	if (bbr_state_is_pkt_epoch)
10684 		bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
10685 	else
10686 		bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP);
10687 
10688 	if (bbr->r_ctl.rc_bbr_state_atflight == 0) {
10689 		if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10690 			flight = ctf_flight_size(bbr->rc_tp,
10691 				     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10692 			if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) {
10693 				/* Keep it slam down */
10694 				if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) {
10695 					bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10696 					bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10697 				}
10698 				if (bbr_sub_drain_app_limit) {
10699 					/* Go app limited if we are on a long drain */
10700 					bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight);
10701 				}
10702 			}
10703 			if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) &&
10704 			    (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) ||
10705 			     (flight >= bbr->r_ctl.flightsize_at_drain))) {
10706 				/*
10707 				 * Still here after the same time as
10708 				 * the gain. We need to drain harder
10709 				 * for the next srtt. Reduce by a set amount
10710 				 * the gain drop is capped at DRAIN states
10711 				 * value (88).
10712 				 */
10713 				bbr->r_ctl.flightsize_at_drain = flight;
10714 				if (bbr_drain_drop_mul &&
10715 				    bbr_drain_drop_div &&
10716 				    (bbr_drain_drop_mul < bbr_drain_drop_div)) {
10717 					/* Use your specific drop value (def 4/5 = 20%) */
10718 					bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul;
10719 					bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div;
10720 				} else {
10721 					/* You get drop of 20% */
10722 					bbr->r_ctl.rc_bbr_hptsi_gain *= 4;
10723 					bbr->r_ctl.rc_bbr_hptsi_gain /= 5;
10724 				}
10725 				if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) {
10726 					/* Reduce our gain again to the bottom  */
10727 					bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
10728 				}
10729 				bbr_log_exit_gain(bbr, cts, 4);
10730 				/*
10731 				 * Extend out so we wait another
10732 				 * epoch before dropping again.
10733 				 */
10734 				bbr->r_ctl.gain_epoch = cts;
10735 			}
10736 			if (flight <= bbr->r_ctl.rc_target_at_state) {
10737 				if (bbr_sub_drain_slam_cwnd &&
10738 				    (bbr->rc_use_google == 0) &&
10739 				    (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10740 					bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10741 					bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10742 				}
10743 				bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10744 				bbr_log_exit_gain(bbr, cts, 3);
10745 			}
10746 		} else {
10747 			/* Its a gain  */
10748 			if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) {
10749 				bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10750 				goto change_state;
10751 			}
10752 			if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) ||
10753 			    ((ctf_outstanding(bbr->rc_tp) +  bbr->rc_tp->t_maxseg - 1) >=
10754 			     bbr->rc_tp->snd_wnd)) {
10755 				bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10756 				bbr_log_exit_gain(bbr, cts, 2);
10757 			}
10758 		}
10759 		/**
10760 		 * We fall through and return always one of two things has
10761 		 * occured.
10762 		 * 1) We are still not at target
10763 		 *    <or>
10764 		 * 2) We reached the target and set rc_bbr_state_atflight
10765 		 *    which means we no longer hit this block
10766 		 *    next time we are called.
10767 		 */
10768 		return;
10769 	}
10770 change_state:
10771 	if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time))
10772 		return;
10773 	if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) {
10774 		/* Less than a full time-period has passed */
10775 		return;
10776 	}
10777 	if (bbr->r_ctl.rc_level_state_extra &&
10778 	    (bbr_state_val(bbr) > BBR_SUB_DRAIN) &&
10779 	    ((cts - bbr->r_ctl.rc_bbr_state_time) <
10780 	     (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10781 		/* Less than a full time-period + extra has passed */
10782 		return;
10783 	}
10784 	if (bbr_gain_gets_extra_too &&
10785 	    bbr->r_ctl.rc_level_state_extra &&
10786 	    (bbr_state_val(bbr) == BBR_SUB_GAIN) &&
10787 	    ((cts - bbr->r_ctl.rc_bbr_state_time) <
10788 	     (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10789 		/* Less than a full time-period + extra has passed */
10790 		return;
10791 	}
10792 	bbr_substate_change(bbr, cts, __LINE__, 1);
10793 }
10794 
10795 static uint32_t
10796 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain)
10797 {
10798 	uint32_t mss, tar;
10799 
10800 	if (bbr->rc_use_google) {
10801 		/* Google just uses the cwnd target */
10802 		tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain);
10803 	} else {
10804 		mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options),
10805 			  bbr->r_ctl.rc_pace_max_segs);
10806 		/* Get the base cwnd with gain rounded to a mss */
10807 		tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr),
10808 						      gain), mss);
10809 		/* Make sure it is within our min */
10810 		if (tar < get_min_cwnd(bbr))
10811 			return (get_min_cwnd(bbr));
10812 	}
10813 	return (tar);
10814 }
10815 
10816 static void
10817 bbr_set_state_target(struct tcp_bbr *bbr, int line)
10818 {
10819 	uint32_t tar, meth;
10820 
10821 	if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
10822 	    ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
10823 		/* Special case using old probe-rtt method */
10824 		tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10825 		meth = 1;
10826 	} else {
10827 		/* Non-probe-rtt case and reduced probe-rtt  */
10828 		if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
10829 		    (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) {
10830 			/* For gain cycle we use the hptsi gain */
10831 			tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10832 			meth = 2;
10833 		} else if ((bbr_target_is_bbunit) || bbr->rc_use_google) {
10834 			/*
10835 			 * If configured, or for google all other states
10836 			 * get BBR_UNIT.
10837 			 */
10838 			tar = bbr_get_a_state_target(bbr, BBR_UNIT);
10839 			meth = 3;
10840 		} else {
10841 			/*
10842 			 * Or we set a target based on the pacing gain
10843 			 * for non-google mode and default (non-configured).
10844 			 * Note we don't set a target goal below drain (192).
10845 			 */
10846 			if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN])  {
10847 				tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]);
10848 				meth = 4;
10849 			} else {
10850 				tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10851 				meth = 5;
10852 			}
10853 		}
10854 	}
10855 	bbr_log_set_of_state_target(bbr, tar, line, meth);
10856 	bbr->r_ctl.rc_target_at_state = tar;
10857 }
10858 
10859 static void
10860 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
10861 {
10862 	/* Change to probe_rtt */
10863 	uint32_t time_in;
10864 
10865 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10866 	bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10867 					     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10868 	bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain
10869 					  + bbr->r_ctl.rc_delivered);
10870 	/* Setup so we force feed the filter */
10871 	if (bbr->rc_use_google || bbr_probertt_sets_rtt)
10872 		bbr->rc_prtt_set_ts = 1;
10873 	if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10874 		time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10875 		counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10876 	}
10877 	bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0);
10878 	bbr->r_ctl.rc_rtt_shrinks = cts;
10879 	bbr->r_ctl.last_in_probertt = cts;
10880 	bbr->r_ctl.rc_probertt_srttchktim = cts;
10881 	bbr->r_ctl.rc_bbr_state_time = cts;
10882 	bbr->rc_bbr_state = BBR_STATE_PROBE_RTT;
10883 	/* We need to force the filter to update */
10884 
10885 	if ((bbr_sub_drain_slam_cwnd) &&
10886 	    bbr->rc_hit_state_1 &&
10887 	    (bbr->rc_use_google == 0) &&
10888 	    (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10889 		if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd)
10890 			bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10891 	} else
10892 		bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10893 	/* Update the lost */
10894 	bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10895 	if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){
10896 		/* Set to the non-configurable default of 4 (PROBE_RTT_MIN)  */
10897 		bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10898 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10899 		bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10900 		bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10901 		bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6);
10902 		bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd;
10903 	} else {
10904 		/*
10905 		 * We bring it down slowly by using a hptsi gain that is
10906 		 * probably 75%. This will slowly float down our outstanding
10907 		 * without tampering with the cwnd.
10908 		 */
10909 		bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
10910 		bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10911 		bbr_set_state_target(bbr, __LINE__);
10912 		if (bbr_prtt_slam_cwnd &&
10913 		    (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
10914 			bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10915 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10916 		}
10917 	}
10918 	if (ctf_flight_size(bbr->rc_tp,
10919 		(bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
10920 	    bbr->r_ctl.rc_target_at_state) {
10921 		/* We are at target */
10922 		bbr->r_ctl.rc_bbr_enters_probertt = cts;
10923 	} else {
10924 		/* We need to come down to reach target before our time begins */
10925 		bbr->r_ctl.rc_bbr_enters_probertt = 0;
10926 	}
10927 	bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch;
10928 	BBR_STAT_INC(bbr_enter_probertt);
10929 	bbr_log_exit_gain(bbr, cts, 0);
10930 	bbr_log_type_statechange(bbr, cts, line);
10931 }
10932 
10933 static void
10934 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts)
10935 {
10936 	/*
10937 	 * Sanity check on probe-rtt intervals.
10938 	 * In crazy situations where we are competing
10939 	 * against new-reno flows with huge buffers
10940 	 * our rtt-prop interval could come to dominate
10941 	 * things if we can't get through a full set
10942 	 * of cycles, we need to adjust it.
10943 	 */
10944 	if (bbr_can_adjust_probertt &&
10945 	    (bbr->rc_use_google == 0)) {
10946 		uint16_t val = 0;
10947 		uint32_t cur_rttp, fval, newval, baseval;
10948 
10949 		/* Are we to small and go into probe-rtt to often? */
10950 		baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1));
10951 		cur_rttp = roundup(baseval, USECS_IN_SECOND);
10952 		fval = bbr_filter_len_sec * USECS_IN_SECOND;
10953 		if (bbr_is_ratio == 0) {
10954 			if (fval > bbr_rtt_probe_limit)
10955 				newval = cur_rttp + (fval - bbr_rtt_probe_limit);
10956 			else
10957 				newval = cur_rttp;
10958 		} else {
10959 			int mul;
10960 
10961 			mul = fval / bbr_rtt_probe_limit;
10962 			newval = cur_rttp * mul;
10963 		}
10964 		if (cur_rttp > 	bbr->r_ctl.rc_probertt_int) {
10965 			bbr->r_ctl.rc_probertt_int = cur_rttp;
10966 			reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10967 			val = 1;
10968 		} else {
10969 			/*
10970 			 * No adjustments were made
10971 			 * do we need to shrink it?
10972 			 */
10973 			if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) {
10974 				if (cur_rttp <= bbr_rtt_probe_limit) {
10975 					/*
10976 					 * Things have calmed down lets
10977 					 * shrink all the way to default
10978 					 */
10979 					bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10980 					reset_time_small(&bbr->r_ctl.rc_rttprop,
10981 							 (bbr_filter_len_sec * USECS_IN_SECOND));
10982 					cur_rttp = bbr_rtt_probe_limit;
10983 					newval = (bbr_filter_len_sec * USECS_IN_SECOND);
10984 					val = 2;
10985 				} else {
10986 					/*
10987 					 * Well does some adjustment make sense?
10988 					 */
10989 					if (cur_rttp < bbr->r_ctl.rc_probertt_int) {
10990 						/* We can reduce interval time some */
10991 						bbr->r_ctl.rc_probertt_int = cur_rttp;
10992 						reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10993 						val = 3;
10994 					}
10995 				}
10996 			}
10997 		}
10998 		if (val)
10999 			bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val);
11000 	}
11001 }
11002 
11003 static void
11004 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
11005 {
11006 	/* Exit probe-rtt */
11007 
11008 	if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) {
11009 		tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11010 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11011 	}
11012 	bbr_log_exit_gain(bbr, cts, 1);
11013 	bbr->rc_hit_state_1 = 0;
11014 	bbr->r_ctl.rc_rtt_shrinks = cts;
11015 	bbr->r_ctl.last_in_probertt = cts;
11016 	bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0);
11017 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11018 	bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp,
11019 					      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11020 					  bbr->r_ctl.rc_delivered);
11021 	if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11022 		uint32_t time_in;
11023 
11024 		time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11025 		counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11026 	}
11027 	if (bbr->rc_filled_pipe) {
11028 		/* Switch to probe_bw */
11029 		bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11030 		bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11031 		bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
11032 		bbr_substate_change(bbr, cts, __LINE__, 0);
11033 		bbr_log_type_statechange(bbr, cts, __LINE__);
11034 	} else {
11035 		/* Back to startup */
11036 		bbr->rc_bbr_state = BBR_STATE_STARTUP;
11037 		bbr->r_ctl.rc_bbr_state_time = cts;
11038 		/*
11039 		 * We don't want to give a complete free 3
11040 		 * measurements until we exit, so we use
11041 		 * the number of pe's we were in probe-rtt
11042 		 * to add to the startup_epoch. That way
11043 		 * we will still retain the old state.
11044 		 */
11045 		bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt);
11046 		bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11047 		/* Make sure to use the lower pg when shifting back in */
11048 		if (bbr->r_ctl.rc_lost &&
11049 		    bbr_use_lower_gain_in_startup &&
11050 		    (bbr->rc_use_google == 0))
11051 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
11052 		else
11053 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
11054 		bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
11055 		/* Probably not needed but set it anyway */
11056 		bbr_set_state_target(bbr, __LINE__);
11057 		bbr_log_type_statechange(bbr, cts, __LINE__);
11058 		bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11059 		    bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0);
11060 	}
11061 	bbr_check_probe_rtt_limits(bbr, cts);
11062 }
11063 
11064 static int32_t inline
11065 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts)
11066 {
11067 	if ((bbr->rc_past_init_win == 1) &&
11068 	    (bbr->rc_in_persist == 0) &&
11069 	    (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) {
11070 		return (1);
11071 	}
11072 	if (bbr_can_force_probertt &&
11073 	    (bbr->rc_in_persist == 0) &&
11074 	    (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
11075 	    ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
11076 		return (1);
11077 	}
11078 	return (0);
11079 }
11080 
11081 
11082 static int32_t
11083 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t  pkt_epoch)
11084 {
11085 	uint64_t btlbw, gain;
11086 	if (pkt_epoch == 0) {
11087 		/*
11088 		 * Need to be on a pkt-epoch to continue.
11089 		 */
11090 		return (0);
11091 	}
11092 	btlbw = bbr_get_full_bw(bbr);
11093 	gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11094 		 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11095 	if (btlbw >= gain) {
11096 		bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11097 		bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11098 				      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11099 		bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11100 	}
11101 	if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)
11102 		return (1);
11103 	bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11104 			      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11105 	return(0);
11106 }
11107 
11108 static int32_t inline
11109 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch)
11110 {
11111 	/* Have we gained 25% in the last 3 packet based epoch's? */
11112 	uint64_t btlbw, gain;
11113 	int do_exit;
11114 	int delta, rtt_gain;
11115 
11116 	if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11117 	    (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11118 		/*
11119 		 * This qualifies as a RTT_PROBE session since we drop the
11120 		 * data outstanding to nothing and waited more than
11121 		 * bbr_rtt_probe_time.
11122 		 */
11123 		bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11124 		bbr_set_reduced_rtt(bbr, cts, __LINE__);
11125 	}
11126 	if (bbr_should_enter_probe_rtt(bbr, cts)) {
11127 		bbr_enter_probe_rtt(bbr, cts, __LINE__);
11128 		return (0);
11129 	}
11130 	if (bbr->rc_use_google)
11131 		return (bbr_google_startup(bbr, cts,  pkt_epoch));
11132 
11133 	if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11134 	    (bbr_use_lower_gain_in_startup)) {
11135 		/* Drop to a lower gain 1.5 x since we saw loss */
11136 		bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
11137 	}
11138 	if (pkt_epoch == 0) {
11139 		/*
11140 		 * Need to be on a pkt-epoch to continue.
11141 		 */
11142 		return (0);
11143 	}
11144 	if (bbr_rtt_gain_thresh) {
11145 		/*
11146 		 * Do we allow a flow to stay
11147 		 * in startup with no loss and no
11148 		 * gain in rtt over a set threshold?
11149 		 */
11150 		if (bbr->r_ctl.rc_pkt_epoch_rtt &&
11151 		    bbr->r_ctl.startup_last_srtt &&
11152 		    (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) {
11153 			delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt;
11154 			rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt;
11155 		} else
11156 			rtt_gain = 0;
11157 		if ((bbr->r_ctl.startup_last_srtt == 0)  ||
11158 		    (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt))
11159 			/* First time or new lower value */
11160 			bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
11161 
11162 		if ((bbr->r_ctl.rc_lost == 0) &&
11163 		    (rtt_gain < bbr_rtt_gain_thresh)) {
11164 			/*
11165 			 * No loss, and we are under
11166 			 * our gain threhold for
11167 			 * increasing RTT.
11168 			 */
11169 			if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11170 				bbr->r_ctl.rc_bbr_last_startup_epoch++;
11171 			bbr_log_startup_event(bbr, cts, rtt_gain,
11172 					      delta, bbr->r_ctl.startup_last_srtt, 10);
11173 			return (0);
11174 		}
11175 	}
11176 	if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) &&
11177 	    (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) &&
11178 	    (!IN_RECOVERY(bbr->rc_tp->t_flags))) {
11179 		/*
11180 		 * We only assess if we have a new measurment when
11181 		 * we have no loss and are not in recovery.
11182 		 * Drag up by one our last_startup epoch so we will hold
11183 		 * the number of non-gain we have already accumulated.
11184 		 */
11185 		if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11186 			bbr->r_ctl.rc_bbr_last_startup_epoch++;
11187 		bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11188 				      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9);
11189 		return (0);
11190 	}
11191 	/* Case where we reduced the lost (bad retransmit) */
11192 	if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost)
11193 		bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11194 	bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count;
11195 	btlbw = bbr_get_full_bw(bbr);
11196 	if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower)
11197 		gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11198 			 (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11199 	else
11200 		gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11201 			 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11202 	do_exit = 0;
11203 	if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw)
11204 		bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11205 	if (btlbw >= gain) {
11206 		bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11207 		/* Update the lost so we won't exit in next set of tests */
11208 		bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11209 		bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11210 				      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11211 	}
11212 	if ((bbr->rc_loss_exit &&
11213 	     (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11214 	     (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) &&
11215 	    ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) {
11216 		/*
11217 		 * If we had no gain,  we had loss and that loss was above
11218 		 * our threshould, the rwnd is not constrained, and we have
11219 		 * had at least 3 packet epochs exit. Note that this is
11220 		 * switched off by sysctl. Google does not do this by the
11221 		 * way.
11222 		 */
11223 		if ((ctf_flight_size(bbr->rc_tp,
11224 			 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11225 		     (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) {
11226 			do_exit = 1;
11227 			bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11228 					      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4);
11229 		} else {
11230 			/* Just record an updated loss value */
11231 			bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11232 			bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11233 					      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5);
11234 		}
11235 	} else
11236 		bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11237 	if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) ||
11238 	    do_exit) {
11239 		/* Return 1 to exit the startup state. */
11240 		return (1);
11241 	}
11242 	/* Stay in startup */
11243 	bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11244 			      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11245 	return (0);
11246 }
11247 
11248 static void
11249 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses)
11250 {
11251 	/*
11252 	 * A tick occured in the rtt epoch do we need to do anything?
11253 	 */
11254 #ifdef BBR_INVARIANTS
11255 	if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
11256 	    (bbr->rc_bbr_state != BBR_STATE_DRAIN) &&
11257 	    (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) &&
11258 	    (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
11259 	    (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) {
11260 		/* Debug code? */
11261 		panic("Unknown BBR state %d?\n", bbr->rc_bbr_state);
11262 	}
11263 #endif
11264 	if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
11265 		/* Do we exit the startup state? */
11266 		if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) {
11267 			uint32_t time_in;
11268 
11269 			bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11270 					      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6);
11271 			bbr->rc_filled_pipe = 1;
11272 			bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11273 			if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11274 
11275 				time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11276 				counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11277 			} else
11278 				time_in = 0;
11279 			if (bbr->rc_no_pacing)
11280 				bbr->rc_no_pacing = 0;
11281 			bbr->r_ctl.rc_bbr_state_time = cts;
11282 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg;
11283 			bbr->rc_bbr_state = BBR_STATE_DRAIN;
11284 			bbr_set_state_target(bbr, __LINE__);
11285 			if ((bbr->rc_use_google == 0) &&
11286 			    bbr_slam_cwnd_in_main_drain) {
11287 				/* Here we don't have to worry about probe-rtt */
11288 				bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
11289 				bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11290 				bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11291 			}
11292 			bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
11293 			bbr_log_type_statechange(bbr, cts, __LINE__);
11294 			if (ctf_flight_size(bbr->rc_tp,
11295 			        (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
11296 			    bbr->r_ctl.rc_target_at_state) {
11297 				/*
11298 				 * Switch to probe_bw if we are already
11299 				 * there
11300 				 */
11301 				bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11302 				bbr_substate_change(bbr, cts, __LINE__, 0);
11303 				bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11304 				bbr_log_type_statechange(bbr, cts, __LINE__);
11305 			}
11306 		}
11307 	} else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) {
11308 		uint32_t inflight;
11309 		struct tcpcb *tp;
11310 
11311 		tp = bbr->rc_tp;
11312 		inflight = ctf_flight_size(tp,
11313 			      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11314 		if (inflight >= bbr->r_ctl.rc_target_at_state) {
11315 			/* We have reached a flight of the cwnd target */
11316 			bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11317 			bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11318 			bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
11319 			bbr_set_state_target(bbr, __LINE__);
11320 			/*
11321 			 * Rig it so we don't do anything crazy and
11322 			 * start fresh with a new randomization.
11323 			 */
11324 			bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
11325 			bbr->rc_bbr_substate = BBR_SUB_LEVEL6;
11326 			bbr_substate_change(bbr, cts, __LINE__, 1);
11327 		}
11328 	} else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) {
11329 		/* Has in-flight reached the bdp (or less)? */
11330 		uint32_t inflight;
11331 		struct tcpcb *tp;
11332 
11333 		tp = bbr->rc_tp;
11334 		inflight = ctf_flight_size(tp,
11335 			      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11336 		if ((bbr->rc_use_google == 0) &&
11337 		    bbr_slam_cwnd_in_main_drain &&
11338 		    (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11339 			/*
11340 			 * Here we don't have to worry about probe-rtt
11341 			 * re-slam it, but keep it slammed down.
11342 			 */
11343 			bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11344 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11345 		}
11346 		if (inflight <= bbr->r_ctl.rc_target_at_state) {
11347 			/* We have drained */
11348 			bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11349 			bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11350 			if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11351 				uint32_t time_in;
11352 
11353 				time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11354 				counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11355 			}
11356 			if ((bbr->rc_use_google == 0) &&
11357 			    bbr_slam_cwnd_in_main_drain &&
11358 			    (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
11359 				/* Restore the cwnd */
11360 				tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11361 				bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11362 			}
11363 			/* Setup probe-rtt has being done now RRS-HERE */
11364 			bbr->r_ctl.rc_rtt_shrinks = cts;
11365 			bbr->r_ctl.last_in_probertt = cts;
11366 			bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0);
11367 			/* Randomly pick a sub-state */
11368 			bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11369 			bbr_substate_change(bbr, cts, __LINE__, 0);
11370 			bbr_log_type_statechange(bbr, cts, __LINE__);
11371 		}
11372 	} else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) {
11373 		uint32_t flight;
11374 
11375 		flight = ctf_flight_size(bbr->rc_tp,
11376 			     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11377 		bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered);
11378 		if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) &&
11379 		    (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11380 			/*
11381 			 * We must keep cwnd at the desired MSS.
11382 			 */
11383 			bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
11384 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11385 		} else if ((bbr_prtt_slam_cwnd) &&
11386 			   (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11387 			/* Re-slam it */
11388 			bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11389 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11390 		}
11391 		if (bbr->r_ctl.rc_bbr_enters_probertt == 0) {
11392 			/* Has outstanding reached our target? */
11393 			if (flight <= bbr->r_ctl.rc_target_at_state) {
11394 				bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0);
11395 				bbr->r_ctl.rc_bbr_enters_probertt = cts;
11396 				/* If time is exactly 0, be 1usec off */
11397 				if (bbr->r_ctl.rc_bbr_enters_probertt == 0)
11398 					bbr->r_ctl.rc_bbr_enters_probertt = 1;
11399 				if (bbr->rc_use_google == 0) {
11400 					/*
11401 					 * Restore any lowering that as occured to
11402 					 * reach here
11403 					 */
11404 					if (bbr->r_ctl.bbr_rttprobe_gain_val)
11405 						bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
11406 					else
11407 						bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11408 				}
11409 			}
11410 			if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) &&
11411 			    (bbr->rc_use_google == 0) &&
11412 			    bbr->r_ctl.bbr_rttprobe_gain_val &&
11413 			    (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) ||
11414 			     (flight >= bbr->r_ctl.flightsize_at_drain))) {
11415 				/*
11416 				 * We have doddled with our current hptsi
11417 				 * gain an srtt and have still not made it
11418 				 * to target, or we have increased our flight.
11419 				 * Lets reduce the gain by xx%
11420 				 * flooring the reduce at DRAIN (based on
11421 				 * mul/div)
11422 				 */
11423 				int red;
11424 
11425 				bbr->r_ctl.flightsize_at_drain = flight;
11426 				bbr->r_ctl.rc_probertt_srttchktim = cts;
11427 				red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1);
11428 				if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) {
11429 					/* Reduce our gain again */
11430 					bbr->r_ctl.rc_bbr_hptsi_gain -= red;
11431 					bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0);
11432 				} else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) {
11433 					/* one more chance before we give up */
11434 					bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
11435 					bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0);
11436 				} else {
11437 					/* At the very bottom */
11438 					bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1);
11439 				}
11440 			}
11441 		}
11442 		if (bbr->r_ctl.rc_bbr_enters_probertt &&
11443 		    (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) &&
11444 		    ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) {
11445 			/* Time to exit probe RTT normally */
11446 			bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts);
11447 		}
11448 	} else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
11449 		if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11450 		    (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11451 			/*
11452 			 * This qualifies as a RTT_PROBE session since we
11453 			 * drop the data outstanding to nothing and waited
11454 			 * more than bbr_rtt_probe_time.
11455 			 */
11456 			bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11457 			bbr_set_reduced_rtt(bbr, cts, __LINE__);
11458 		}
11459 		if (bbr_should_enter_probe_rtt(bbr, cts)) {
11460 			bbr_enter_probe_rtt(bbr, cts, __LINE__);
11461 		} else {
11462 			bbr_set_probebw_gains(bbr, cts, losses);
11463 		}
11464 	}
11465 }
11466 
11467 static void
11468 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses)
11469 {
11470 	int32_t epoch = 0;
11471 
11472 	if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
11473 		bbr_set_epoch(bbr, cts, line);
11474 		/* At each epoch doe lt bw sampling */
11475 		epoch = 1;
11476 	}
11477 	bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses);
11478 }
11479 
11480 static int
11481 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
11482     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
11483     int32_t nxt_pkt, struct timeval *tv)
11484 {
11485 	int32_t thflags, retval;
11486 	uint32_t cts, lcts;
11487 	uint32_t tiwin;
11488 	struct tcpopt to;
11489 	struct tcp_bbr *bbr;
11490 	struct bbr_sendmap *rsm;
11491 	struct timeval ltv;
11492 	int32_t did_out = 0;
11493 	int32_t in_recovery;
11494 	uint16_t nsegs;
11495 	int32_t prev_state;
11496 	uint32_t lost;
11497 
11498 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
11499 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11500 	/* add in our stats */
11501 	kern_prefetch(bbr, &prev_state);
11502 	prev_state = 0;
11503 	thflags = th->th_flags;
11504 	/*
11505 	 * If this is either a state-changing packet or current state isn't
11506 	 * established, we require a write lock on tcbinfo.  Otherwise, we
11507 	 * allow the tcbinfo to be in either alocked or unlocked, as the
11508 	 * caller may have unnecessarily acquired a write lock due to a
11509 	 * race.
11510 	 */
11511 	INP_WLOCK_ASSERT(tp->t_inpcb);
11512 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
11513 	    __func__));
11514 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
11515 	    __func__));
11516 
11517 	tp->t_rcvtime = ticks;
11518 	/*
11519 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
11520 	 * the scale is zero.
11521 	 */
11522 	tiwin = th->th_win << tp->snd_scale;
11523 #ifdef STATS
11524 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
11525 #endif
11526 	/*
11527 	 * Parse options on any incoming segment.
11528 	 */
11529 	tcp_dooptions(&to, (u_char *)(th + 1),
11530 	    (th->th_off << 2) - sizeof(struct tcphdr),
11531 	    (thflags & TH_SYN) ? TO_SYN : 0);
11532 
11533 	if (m->m_flags & M_TSTMP) {
11534 		/* Prefer the hardware timestamp if present */
11535 		struct timespec ts;
11536 
11537 		mbuf_tstmp2timespec(m, &ts);
11538 		bbr->rc_tv.tv_sec = ts.tv_sec;
11539 		bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11540 		bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11541 	} else if (m->m_flags & M_TSTMP_LRO) {
11542 		/* Next the arrival timestamp */
11543 		struct timespec ts;
11544 
11545 		mbuf_tstmp2timespec(m, &ts);
11546 		bbr->rc_tv.tv_sec = ts.tv_sec;
11547 		bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11548 		bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11549 	} else {
11550 		/*
11551 		 * Ok just get the current time.
11552 		 */
11553 		bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv);
11554 	}
11555 	/*
11556 	 * If echoed timestamp is later than the current time, fall back to
11557 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
11558 	 * were used when this connection was established.
11559 	 */
11560 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
11561 		to.to_tsecr -= tp->ts_offset;
11562 		if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv)))
11563 			to.to_tsecr = 0;
11564 	}
11565 	/*
11566 	 * If its the first time in we need to take care of options and
11567 	 * verify we can do SACK for rack!
11568 	 */
11569 	if (bbr->r_state == 0) {
11570 		/*
11571 		 * Process options only when we get SYN/ACK back. The SYN
11572 		 * case for incoming connections is handled in tcp_syncache.
11573 		 * According to RFC1323 the window field in a SYN (i.e., a
11574 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
11575 		 * this is traditional behavior, may need to be cleaned up.
11576 		 */
11577 		if (bbr->rc_inp == NULL) {
11578 			bbr->rc_inp = tp->t_inpcb;
11579 		}
11580 		/*
11581 		 * We need to init rc_inp here since its not init'd when
11582 		 * bbr_init is called
11583 		 */
11584 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
11585 			if ((to.to_flags & TOF_SCALE) &&
11586 			    (tp->t_flags & TF_REQ_SCALE)) {
11587 				tp->t_flags |= TF_RCVD_SCALE;
11588 				tp->snd_scale = to.to_wscale;
11589 			}
11590 			/*
11591 			 * Initial send window.  It will be updated with the
11592 			 * next incoming segment to the scaled value.
11593 			 */
11594 			tp->snd_wnd = th->th_win;
11595 			if (to.to_flags & TOF_TS) {
11596 				tp->t_flags |= TF_RCVD_TSTMP;
11597 				tp->ts_recent = to.to_tsval;
11598 				tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
11599 			}
11600 			if (to.to_flags & TOF_MSS)
11601 				tcp_mss(tp, to.to_mss);
11602 			if ((tp->t_flags & TF_SACK_PERMIT) &&
11603 			    (to.to_flags & TOF_SACKPERM) == 0)
11604 				tp->t_flags &= ~TF_SACK_PERMIT;
11605 			if (IS_FASTOPEN(tp->t_flags)) {
11606 				if (to.to_flags & TOF_FASTOPEN) {
11607 					uint16_t mss;
11608 
11609 					if (to.to_flags & TOF_MSS)
11610 						mss = to.to_mss;
11611 					else
11612 						if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
11613 							mss = TCP6_MSS;
11614 						else
11615 							mss = TCP_MSS;
11616 					tcp_fastopen_update_cache(tp, mss,
11617 					    to.to_tfo_len, to.to_tfo_cookie);
11618 				} else
11619 					tcp_fastopen_disable_path(tp);
11620 			}
11621 		}
11622 		/*
11623 		 * At this point we are at the initial call. Here we decide
11624 		 * if we are doing RACK or not. We do this by seeing if
11625 		 * TF_SACK_PERMIT is set, if not rack is *not* possible and
11626 		 * we switch to the default code.
11627 		 */
11628 		if ((tp->t_flags & TF_SACK_PERMIT) == 0) {
11629 			/* Bail */
11630 			tcp_switch_back_to_default(tp);
11631 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
11632 			    tlen, iptos);
11633 			return (1);
11634 		}
11635 		/* Set the flag */
11636 		bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
11637 		tcp_set_hpts(tp->t_inpcb);
11638 		sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack);
11639 	}
11640 	if (thflags & TH_ACK) {
11641 		/* Track ack types */
11642 		if (to.to_flags & TOF_SACK)
11643 			BBR_STAT_INC(bbr_acks_with_sacks);
11644 		else
11645 			BBR_STAT_INC(bbr_plain_acks);
11646 	}
11647 	/*
11648 	 * This is the one exception case where we set the rack state
11649 	 * always. All other times (timers etc) we must have a rack-state
11650 	 * set (so we assure we have done the checks above for SACK).
11651 	 */
11652 	if (thflags & TH_FIN)
11653 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
11654 	if (bbr->r_state != tp->t_state)
11655 		bbr_set_state(tp, bbr, tiwin);
11656 
11657 	if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL)
11658 		kern_prefetch(rsm, &prev_state);
11659 	prev_state = bbr->r_state;
11660 	bbr->rc_ack_was_delayed = 0;
11661 	lost = bbr->r_ctl.rc_lost;
11662 	bbr->rc_is_pkt_epoch_now = 0;
11663 	if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) {
11664 		/* Get the real time into lcts and figure the real delay */
11665 		lcts = tcp_get_usecs(&ltv);
11666 		if (TSTMP_GT(lcts, cts)) {
11667 			bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts;
11668 			bbr->rc_ack_was_delayed = 1;
11669 			if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay,
11670 				     bbr->r_ctl.highest_hdwr_delay))
11671 				bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay;
11672 		} else {
11673 			bbr->r_ctl.rc_ack_hdwr_delay = 0;
11674 			bbr->rc_ack_was_delayed = 0;
11675 		}
11676 	} else {
11677 		bbr->r_ctl.rc_ack_hdwr_delay = 0;
11678 		bbr->rc_ack_was_delayed = 0;
11679 	}
11680 	bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m);
11681 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
11682 		retval = 0;
11683 		m_freem(m);
11684                 goto done_with_input;
11685         }
11686         /*
11687          * If a segment with the ACK-bit set arrives in the SYN-SENT state
11688          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
11689          */
11690         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
11691             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
11692 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11693 		ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11694                 return (1);
11695         }
11696 	in_recovery = IN_RECOVERY(tp->t_flags);
11697 	if (tiwin > bbr->r_ctl.rc_high_rwnd)
11698 		bbr->r_ctl.rc_high_rwnd = tiwin;
11699 #ifdef BBR_INVARIANTS
11700 	if ((tp->t_inpcb->inp_flags & INP_DROPPED) ||
11701 	    (tp->t_inpcb->inp_flags2 & INP_FREED)) {
11702 		panic("tp:%p bbr:%p given a dropped inp:%p",
11703 		    tp, bbr, tp->t_inpcb);
11704 	}
11705 #endif
11706 	bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp,
11707 					    (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11708 	bbr->rtt_valid = 0;
11709 	if (to.to_flags & TOF_TS) {
11710 		bbr->rc_ts_valid = 1;
11711 		bbr->r_ctl.last_inbound_ts = to.to_tsval;
11712 	} else {
11713 		bbr->rc_ts_valid = 0;
11714 		bbr->r_ctl.last_inbound_ts = 0;
11715 	}
11716 	retval = (*bbr->r_substate) (m, th, so,
11717 	    tp, &to, drop_hdrlen,
11718 	    tlen, tiwin, thflags, nxt_pkt, iptos);
11719 #ifdef BBR_INVARIANTS
11720 	if ((retval == 0) &&
11721 	    (tp->t_inpcb == NULL)) {
11722 		panic("retval:%d tp:%p t_inpcb:NULL state:%d",
11723 		    retval, tp, prev_state);
11724 	}
11725 #endif
11726 	if (nxt_pkt == 0)
11727 		BBR_STAT_INC(bbr_rlock_left_ret0);
11728 	else
11729 		BBR_STAT_INC(bbr_rlock_left_ret1);
11730 	if (retval == 0) {
11731 		/*
11732 		 * If retval is 1 the tcb is unlocked and most likely the tp
11733 		 * is gone.
11734 		 */
11735 		INP_WLOCK_ASSERT(tp->t_inpcb);
11736 		tcp_bbr_xmit_timer_commit(bbr, tp, cts);
11737 		if (bbr->rc_is_pkt_epoch_now)
11738 			bbr_set_pktepoch(bbr, cts, __LINE__);
11739 		bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost));
11740 		if (nxt_pkt == 0) {
11741 			if (bbr->r_wanted_output != 0) {
11742 				bbr->rc_output_starts_timer = 0;
11743 				did_out = 1;
11744 				(void)tp->t_fb->tfb_tcp_output(tp);
11745 			} else
11746 				bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0);
11747 		}
11748 		if ((nxt_pkt == 0) &&
11749 		    ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11750 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
11751 		     (tp->t_flags & TF_DELACK) ||
11752 		     ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11753 		      (tp->t_state <= TCPS_CLOSING)))) {
11754 			/*
11755 			 * We could not send (probably in the hpts but
11756 			 * stopped the timer)?
11757 			 */
11758 			if ((tp->snd_max == tp->snd_una) &&
11759 			    ((tp->t_flags & TF_DELACK) == 0) &&
11760 			    (bbr->rc_inp->inp_in_hpts) &&
11761 			    (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11762 				/*
11763 				 * keep alive not needed if we are hptsi
11764 				 * output yet
11765 				 */
11766 				;
11767 			} else {
11768 				if (bbr->rc_inp->inp_in_hpts) {
11769 					tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
11770 					if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11771 					    (TSTMP_GT(lcts, bbr->rc_pacer_started))) {
11772 						uint32_t del;
11773 
11774 						del = lcts - bbr->rc_pacer_started;
11775 						if (bbr->r_ctl.rc_last_delay_val > del) {
11776 							BBR_STAT_INC(bbr_force_timer_start);
11777 							bbr->r_ctl.rc_last_delay_val -= del;
11778 							bbr->rc_pacer_started = lcts;
11779 						} else {
11780 							/* We are late */
11781 							bbr->r_ctl.rc_last_delay_val = 0;
11782 							BBR_STAT_INC(bbr_force_output);
11783 							(void)tp->t_fb->tfb_tcp_output(tp);
11784 						}
11785 					}
11786 				}
11787 				bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val,
11788 				    0);
11789 			}
11790 		} else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) {
11791 			/* Do we have the correct timer running? */
11792 			bbr_timer_audit(tp, bbr, lcts, &so->so_snd);
11793 		}
11794 		/* Do we have a new state */
11795 		if (bbr->r_state != tp->t_state)
11796 			bbr_set_state(tp, bbr, tiwin);
11797 done_with_input:
11798 		bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out);
11799 		if (did_out)
11800 			bbr->r_wanted_output = 0;
11801 #ifdef BBR_INVARIANTS
11802 		if (tp->t_inpcb == NULL) {
11803 			panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
11804 			    did_out,
11805 			    retval, tp, prev_state);
11806 		}
11807 #endif
11808 	}
11809 	return (retval);
11810 }
11811 
11812 static void
11813 bbr_log_type_hrdwtso(struct tcpcb *tp, struct tcp_bbr *bbr, int len, int mod, int what_we_can_send)
11814 {
11815 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
11816 		union tcp_log_stackspecific log;
11817 		struct timeval tv;
11818 		uint32_t cts;
11819 
11820 		cts = tcp_get_usecs(&tv);
11821 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
11822 		log.u_bbr.flex1 = bbr->r_ctl.rc_pace_min_segs;
11823 		log.u_bbr.flex2 = what_we_can_send;
11824 		log.u_bbr.flex3 = bbr->r_ctl.rc_pace_max_segs;
11825 		log.u_bbr.flex4 = len;
11826 		log.u_bbr.flex5 = 0;
11827 		log.u_bbr.flex7 = mod;
11828 		log.u_bbr.flex8 = 1;
11829 		TCP_LOG_EVENTP(tp, NULL,
11830 		    &tp->t_inpcb->inp_socket->so_rcv,
11831 		    &tp->t_inpcb->inp_socket->so_snd,
11832 		    TCP_HDWR_TLS, 0,
11833 		    0, &log, false, &tv);
11834 	}
11835 }
11836 
11837 static void
11838 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
11839     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11840 {
11841 	struct timeval tv;
11842 	int retval;
11843 
11844 	/* First lets see if we have old packets */
11845 	if (tp->t_in_pkt) {
11846 		if (ctf_do_queued_segments(so, tp, 1)) {
11847 			m_freem(m);
11848 			return;
11849 		}
11850 	}
11851 	if (m->m_flags & M_TSTMP_LRO) {
11852 		tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
11853 		tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
11854 	} else {
11855 		/* Should not be should we kassert instead? */
11856 		tcp_get_usecs(&tv);
11857 	}
11858 	retval = bbr_do_segment_nounlock(m, th, so, tp,
11859 					 drop_hdrlen, tlen, iptos, 0, &tv);
11860 	if (retval == 0)
11861 		INP_WUNLOCK(tp->t_inpcb);
11862 }
11863 
11864 /*
11865  * Return how much data can be sent without violating the
11866  * cwnd or rwnd.
11867  */
11868 
11869 static inline uint32_t
11870 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin,
11871     uint32_t avail, int32_t sb_offset, uint32_t cts)
11872 {
11873 	uint32_t len;
11874 
11875 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
11876 		/* We never want to go over our peers rcv-window */
11877 		len = 0;
11878 	} else {
11879 		uint32_t flight;
11880 
11881 		flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11882 		if (flight >= sendwin) {
11883 			/*
11884 			 * We have in flight what we are allowed by cwnd (if
11885 			 * it was rwnd blocking it would have hit above out
11886 			 * >= tp->snd_wnd).
11887 			 */
11888 			return (0);
11889 		}
11890 		len = sendwin - flight;
11891 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11892 			/* We would send too much (beyond the rwnd) */
11893 			len = tp->snd_wnd - ctf_outstanding(tp);
11894 		}
11895 		if ((len + sb_offset) > avail) {
11896 			/*
11897 			 * We don't have that much in the SB, how much is
11898 			 * there?
11899 			 */
11900 			len = avail - sb_offset;
11901 		}
11902 	}
11903 	return (len);
11904 }
11905 
11906 static inline void
11907 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11908 {
11909 #ifdef NETFLIX_STATS
11910 	KMOD_TCPSTAT_INC(tcps_sndpack_error);
11911 	KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len);
11912 #endif
11913 }
11914 
11915 static inline void
11916 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11917 {
11918 	if (error) {
11919 		bbr_do_error_accounting(tp, bbr, rsm, len, error);
11920 		return;
11921 	}
11922 	if (rsm) {
11923 		if (rsm->r_flags & BBR_TLP) {
11924 			/*
11925 			 * TLP should not count in retran count, but in its
11926 			 * own bin
11927 			 */
11928 #ifdef NETFLIX_STATS
11929 			tp->t_sndtlppack++;
11930 			tp->t_sndtlpbyte += len;
11931 			KMOD_TCPSTAT_INC(tcps_tlpresends);
11932 			KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len);
11933 #endif
11934 		} else {
11935 			/* Retransmit */
11936 			tp->t_sndrexmitpack++;
11937 			KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
11938 			KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
11939 #ifdef STATS
11940 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
11941 			    len);
11942 #endif
11943 		}
11944 		/*
11945 		 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is
11946 		 * sub-state
11947 		 */
11948 		counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len);
11949 		if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) {
11950 			/* Non probe_bw log in 1, 2, or 4. */
11951 			counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len);
11952 		} else {
11953 			/*
11954 			 * Log our probe state 3, and log also 5-13 to show
11955 			 * us the recovery sub-state for the send. This
11956 			 * means that 3 == (5+6+7+8+9+10+11+12+13)
11957 			 */
11958 			counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len);
11959 			counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len);
11960 		}
11961 		/* Place in both 16's the totals of retransmitted */
11962 		counter_u64_add(bbr_state_lost[16], len);
11963 		counter_u64_add(bbr_state_resend[16], len);
11964 		/* Place in 17's the total sent */
11965 		counter_u64_add(bbr_state_resend[17], len);
11966 		counter_u64_add(bbr_state_lost[17], len);
11967 
11968 	} else {
11969 		/* New sends */
11970 		KMOD_TCPSTAT_INC(tcps_sndpack);
11971 		KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
11972 		/* Place in 17's the total sent */
11973 		counter_u64_add(bbr_state_resend[17], len);
11974 		counter_u64_add(bbr_state_lost[17], len);
11975 #ifdef STATS
11976 		stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
11977 		    len);
11978 #endif
11979 	}
11980 }
11981 
11982 static void
11983 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level)
11984 {
11985 	if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) {
11986 		/*
11987 		 * Limit the cwnd to not be above N x the target plus whats
11988 		 * is outstanding. The target is based on the current b/w
11989 		 * estimate.
11990 		 */
11991 		uint32_t target;
11992 
11993 		target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT);
11994 		target += ctf_outstanding(tp);
11995 		target *= bbr_target_cwnd_mult_limit;
11996 		if (tp->snd_cwnd > target)
11997 			tp->snd_cwnd = target;
11998 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__);
11999 	}
12000 }
12001 
12002 static int
12003 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg)
12004 {
12005 	/*
12006 	 * "adv" is the amount we could increase the window, taking into
12007 	 * account that we are limited by TCP_MAXWIN << tp->rcv_scale.
12008 	 */
12009 	uint32_t adv;
12010 	int32_t oldwin;
12011 
12012 	adv = min(recwin, TCP_MAXWIN << tp->rcv_scale);
12013 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
12014 		oldwin = (tp->rcv_adv - tp->rcv_nxt);
12015 		adv -= oldwin;
12016 	} else
12017 		oldwin = 0;
12018 
12019 	/*
12020 	 * If the new window size ends up being the same as the old size
12021 	 * when it is scaled, then don't force a window update.
12022 	 */
12023 	if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
12024 		return (0);
12025 
12026 	if (adv >= (2 * maxseg) &&
12027 	    (adv >= (so->so_rcv.sb_hiwat / 4) ||
12028 	    recwin <= (so->so_rcv.sb_hiwat / 8) ||
12029 	    so->so_rcv.sb_hiwat <= 8 * maxseg)) {
12030 		return (1);
12031 	}
12032 	if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat)
12033 		return (1);
12034 	return (0);
12035 }
12036 
12037 /*
12038  * Return 0 on success and a errno on failure to send.
12039  * Note that a 0 return may not mean we sent anything
12040  * if the TCB was on the hpts. A non-zero return
12041  * does indicate the error we got from ip[6]_output.
12042  */
12043 static int
12044 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv)
12045 {
12046 	struct socket *so;
12047 	int32_t len;
12048 	uint32_t cts;
12049 	uint32_t recwin, sendwin;
12050 	int32_t sb_offset;
12051 	int32_t flags, abandon, error = 0;
12052 	struct tcp_log_buffer *lgb = NULL;
12053 	struct mbuf *m;
12054 	struct mbuf *mb;
12055 	uint32_t if_hw_tsomaxsegcount = 0;
12056 	uint32_t if_hw_tsomaxsegsize = 0;
12057 	uint32_t if_hw_tsomax = 0;
12058 	struct ip *ip = NULL;
12059 #ifdef TCPDEBUG
12060 	struct ipovly *ipov = NULL;
12061 #endif
12062 	struct tcp_bbr *bbr;
12063 	struct tcphdr *th;
12064 #ifdef NETFLIX_TCPOUDP
12065 	struct udphdr *udp = NULL;
12066 #endif
12067 	u_char opt[TCP_MAXOLEN];
12068 	unsigned ipoptlen, optlen, hdrlen;
12069 #ifdef NETFLIX_TCPOUDP
12070 	unsigned ulen;
12071 #endif
12072 	uint32_t bbr_seq;
12073 	uint32_t delay_calc=0;
12074 	uint8_t doing_tlp = 0;
12075 	uint8_t local_options;
12076 #ifdef BBR_INVARIANTS
12077 	uint8_t doing_retran_from = 0;
12078 	uint8_t picked_up_retran = 0;
12079 #endif
12080 	uint8_t wanted_cookie = 0;
12081 	uint8_t more_to_rxt=0;
12082 	int32_t prefetch_so_done = 0;
12083 	int32_t prefetch_rsm = 0;
12084  	uint32_t what_we_can = 0;
12085 	uint32_t tot_len = 0;
12086 	uint32_t rtr_cnt = 0;
12087 	uint32_t maxseg, pace_max_segs, p_maxseg;
12088 	int32_t csum_flags;
12089  	int32_t hw_tls;
12090 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12091 	unsigned ipsec_optlen = 0;
12092 
12093 #endif
12094 	volatile int32_t sack_rxmit;
12095 	struct bbr_sendmap *rsm = NULL;
12096 	int32_t tso, mtu;
12097 	int force_tso = 0;
12098 	struct tcpopt to;
12099 	int32_t slot = 0;
12100 	struct inpcb *inp;
12101 	struct sockbuf *sb;
12102 	uint32_t hpts_calling;
12103 #ifdef INET6
12104 	struct ip6_hdr *ip6 = NULL;
12105 	int32_t isipv6;
12106 #endif
12107 	uint8_t app_limited = BBR_JR_SENT_DATA;
12108 	uint8_t filled_all = 0;
12109 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
12110 	/* We take a cache hit here */
12111 	memcpy(&bbr->rc_tv, tv, sizeof(struct timeval));
12112 	cts = tcp_tv_to_usectick(&bbr->rc_tv);
12113 	inp = bbr->rc_inp;
12114 	so = inp->inp_socket;
12115 	sb = &so->so_snd;
12116 #ifdef KERN_TLS
12117  	if (sb->sb_flags & SB_TLS_IFNET)
12118  		hw_tls = 1;
12119  	else
12120 #endif
12121  		hw_tls = 0;
12122 	kern_prefetch(sb, &maxseg);
12123 	maxseg = tp->t_maxseg - bbr->rc_last_options;
12124 	if (bbr_minseg(bbr) < maxseg) {
12125 		tcp_bbr_tso_size_check(bbr, cts);
12126 	}
12127 	/* Remove any flags that indicate we are pacing on the inp  */
12128 	pace_max_segs = bbr->r_ctl.rc_pace_max_segs;
12129 	p_maxseg = min(maxseg, pace_max_segs);
12130 	INP_WLOCK_ASSERT(inp);
12131 #ifdef TCP_OFFLOAD
12132 	if (tp->t_flags & TF_TOE)
12133 		return (tcp_offload_output(tp));
12134 #endif
12135 
12136 #ifdef INET6
12137 	if (bbr->r_state) {
12138 		/* Use the cache line loaded if possible */
12139 		isipv6 = bbr->r_is_v6;
12140 	} else {
12141 		isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
12142 	}
12143 #endif
12144 	if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
12145 	    inp->inp_in_hpts) {
12146 		/*
12147 		 * We are on the hpts for some timer but not hptsi output.
12148 		 * Possibly remove from the hpts so we can send/recv etc.
12149 		 */
12150 		if ((tp->t_flags & TF_ACKNOW) == 0) {
12151 			/*
12152 			 * No immediate demand right now to send an ack, but
12153 			 * the user may have read, making room for new data
12154 			 * (a window update). If so we may want to cancel
12155 			 * whatever timer is running (KEEP/DEL-ACK?) and
12156 			 * continue to send out a window update. Or we may
12157 			 * have gotten more data into the socket buffer to
12158 			 * send.
12159 			 */
12160 			recwin = min(max(sbspace(&so->so_rcv), 0),
12161 			    TCP_MAXWIN << tp->rcv_scale);
12162 			if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) &&
12163 			    ((tcp_outflags[tp->t_state] & TH_RST) == 0) &&
12164 			    ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <=
12165 			    (tp->snd_max - tp->snd_una))) {
12166 				/*
12167 				 * Nothing new to send and no window update
12168 				 * is needed to send. Lets just return and
12169 				 * let the timer-run off.
12170 				 */
12171 				return (0);
12172 			}
12173 		}
12174 		tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12175 		bbr_timer_cancel(bbr, __LINE__, cts);
12176 	}
12177 	if (bbr->r_ctl.rc_last_delay_val) {
12178 		/* Calculate a rough delay for early escape to sending  */
12179 		if (SEQ_GT(cts, bbr->rc_pacer_started))
12180 			delay_calc = cts - bbr->rc_pacer_started;
12181 		if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12182 			delay_calc -= bbr->r_ctl.rc_last_delay_val;
12183 		else
12184 			delay_calc = 0;
12185 	}
12186 	/* Mark that we have called bbr_output(). */
12187 	if ((bbr->r_timer_override) ||
12188 	    (tp->t_state < TCPS_ESTABLISHED)) {
12189 		/* Timeouts or early states are exempt */
12190 		if (inp->inp_in_hpts)
12191 			tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12192 	} else if (inp->inp_in_hpts) {
12193 		if ((bbr->r_ctl.rc_last_delay_val) &&
12194 		    (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
12195 		    delay_calc) {
12196 			/*
12197 			 * We were being paced for output and the delay has
12198 			 * already exceeded when we were supposed to be
12199 			 * called, lets go ahead and pull out of the hpts
12200 			 * and call output.
12201 			 */
12202 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1);
12203 			bbr->r_ctl.rc_last_delay_val = 0;
12204 			tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12205 		} else if (tp->t_state == TCPS_CLOSED) {
12206 			bbr->r_ctl.rc_last_delay_val = 0;
12207 			tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12208 		} else {
12209 			/*
12210 			 * On the hpts, you shall not pass! even if ACKNOW
12211 			 * is on, we will when the hpts fires, unless of
12212 			 * course we are overdue.
12213 			 */
12214 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1);
12215 			return (0);
12216 		}
12217 	}
12218 	bbr->rc_cwnd_limited = 0;
12219 	if (bbr->r_ctl.rc_last_delay_val) {
12220 		/* recalculate the real delay and deal with over/under  */
12221 		if (SEQ_GT(cts, bbr->rc_pacer_started))
12222 			delay_calc = cts - bbr->rc_pacer_started;
12223 		else
12224 			delay_calc = 0;
12225 		if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12226 			/* Setup the delay which will be added in */
12227 			delay_calc -= bbr->r_ctl.rc_last_delay_val;
12228 		else {
12229 			/*
12230 			 * We are early setup to adjust
12231 			 * our slot time.
12232 			 */
12233 			uint64_t merged_val;
12234 
12235 			bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc);
12236 			bbr->r_agg_early_set = 1;
12237 			if (bbr->r_ctl.rc_hptsi_agg_delay) {
12238 				if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) {
12239 					/* Nope our previous late cancels out the early */
12240 					bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early;
12241 					bbr->r_agg_early_set = 0;
12242 					bbr->r_ctl.rc_agg_early = 0;
12243 				} else {
12244 					bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay;
12245 					bbr->r_ctl.rc_hptsi_agg_delay = 0;
12246 				}
12247 			}
12248 			merged_val = bbr->rc_pacer_started;
12249 			merged_val <<= 32;
12250 			merged_val |= bbr->r_ctl.rc_last_delay_val;
12251 			bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls,
12252 						 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val,
12253 						 bbr->r_agg_early_set, 3);
12254 			bbr->r_ctl.rc_last_delay_val = 0;
12255 			BBR_STAT_INC(bbr_early);
12256 			delay_calc = 0;
12257 		}
12258 	} else {
12259 		/* We were not delayed due to hptsi */
12260 		if (bbr->r_agg_early_set)
12261 			bbr->r_ctl.rc_agg_early = 0;
12262 		bbr->r_agg_early_set = 0;
12263 		delay_calc = 0;
12264 	}
12265 	if (delay_calc) {
12266 		/*
12267 		 * We had a hptsi delay which means we are falling behind on
12268 		 * sending at the expected rate. Calculate an extra amount
12269 		 * of data we can send, if any, to put us back on track.
12270 		 */
12271 		if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay)
12272 			bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff;
12273 		else
12274 			bbr->r_ctl.rc_hptsi_agg_delay += delay_calc;
12275 	}
12276 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12277 	if ((tp->snd_una == tp->snd_max) &&
12278 	    (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
12279 	    (sbavail(sb))) {
12280 		/*
12281 		 * Ok we have been idle with nothing outstanding
12282 		 * we possibly need to start fresh with either a new
12283 		 * suite of states or a fast-ramp up.
12284 		 */
12285 		bbr_restart_after_idle(bbr,
12286 				       cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time));
12287 	}
12288 	/*
12289 	 * Now was there a hptsi delay where we are behind? We only count
12290 	 * being behind if: a) We are not in recovery. b) There was a delay.
12291 	 * <and> c) We had room to send something.
12292 	 *
12293 	 */
12294 	hpts_calling = inp->inp_hpts_calls;
12295 	inp->inp_hpts_calls = 0;
12296 	if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
12297 		if (bbr_process_timers(tp, bbr, cts, hpts_calling)) {
12298 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1);
12299 			return (0);
12300 		}
12301 	}
12302 	bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12303 	if (hpts_calling &&
12304 	    (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12305 		bbr->r_ctl.rc_last_delay_val = 0;
12306 	}
12307 	bbr->r_timer_override = 0;
12308 	bbr->r_wanted_output = 0;
12309 	/*
12310 	 * For TFO connections in SYN_RECEIVED, only allow the initial
12311 	 * SYN|ACK and those sent by the retransmit timer.
12312 	 */
12313 	if (IS_FASTOPEN(tp->t_flags) &&
12314 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
12315 	     (tp->t_state == TCPS_SYN_SENT)) &&
12316 	    SEQ_GT(tp->snd_max, tp->snd_una) &&	/* initial SYN or SYN|ACK sent */
12317 	    (tp->t_rxtshift == 0)) {	/* not a retransmit */
12318 		return (0);
12319 	}
12320 	/*
12321 	 * Before sending anything check for a state update. For hpts
12322 	 * calling without input this is important. If its input calling
12323 	 * then this was already done.
12324 	 */
12325 	if (bbr->rc_use_google == 0)
12326 		bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12327 again:
12328 	/*
12329 	 * If we've recently taken a timeout, snd_max will be greater than
12330 	 * snd_max. BBR in general does not pay much attention to snd_nxt
12331 	 * for historic reasons the persist timer still uses it. This means
12332 	 * we have to look at it. All retransmissions that are not persits
12333 	 * use the rsm that needs to be sent so snd_nxt is ignored. At the
12334 	 * end of this routine we pull snd_nxt always up to snd_max.
12335 	 */
12336 	doing_tlp = 0;
12337 #ifdef BBR_INVARIANTS
12338 	doing_retran_from = picked_up_retran = 0;
12339 #endif
12340 	error = 0;
12341 	tso = 0;
12342 	slot = 0;
12343 	mtu = 0;
12344 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12345 	sb_offset = tp->snd_max - tp->snd_una;
12346 	flags = tcp_outflags[tp->t_state];
12347 	sack_rxmit = 0;
12348 	len = 0;
12349 	rsm = NULL;
12350 	if (flags & TH_RST) {
12351 		SOCKBUF_LOCK(sb);
12352 		goto send;
12353 	}
12354 recheck_resend:
12355 	while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
12356 		/* We need to always have one in reserve */
12357 		rsm = bbr_alloc(bbr);
12358 		if (rsm == NULL) {
12359 			error = ENOMEM;
12360 			/* Lie to get on the hpts */
12361 			tot_len = tp->t_maxseg;
12362 			if (hpts_calling)
12363 				/* Retry in a ms */
12364 				slot = 1001;
12365 			goto just_return_nolock;
12366 		}
12367 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
12368 		bbr->r_ctl.rc_free_cnt++;
12369 		rsm = NULL;
12370 	}
12371 	/* What do we send, a resend? */
12372 	if (bbr->r_ctl.rc_resend == NULL) {
12373 		/* Check for rack timeout */
12374 		bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
12375 		if (bbr->r_ctl.rc_resend) {
12376 #ifdef BBR_INVARIANTS
12377 			picked_up_retran = 1;
12378 #endif
12379 			bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend);
12380 		}
12381 	}
12382 	if (bbr->r_ctl.rc_resend) {
12383 		rsm = bbr->r_ctl.rc_resend;
12384 #ifdef BBR_INVARIANTS
12385 		doing_retran_from = 1;
12386 #endif
12387 		/* Remove any TLP flags its a RACK or T-O */
12388 		rsm->r_flags &= ~BBR_TLP;
12389 		bbr->r_ctl.rc_resend = NULL;
12390 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12391 #ifdef BBR_INVARIANTS
12392 			panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n",
12393 			    tp, bbr, rsm, rsm->r_start, tp->snd_una);
12394 			goto recheck_resend;
12395 #else
12396 			/* TSNH */
12397 			rsm = NULL;
12398 			goto recheck_resend;
12399 #endif
12400 		}
12401 		rtr_cnt++;
12402 		if (rsm->r_flags & BBR_HAS_SYN) {
12403 			/* Only retransmit a SYN by itself */
12404 			len = 0;
12405 			if ((flags & TH_SYN) == 0) {
12406 				/* Huh something is wrong */
12407 				rsm->r_start++;
12408 				if (rsm->r_start == rsm->r_end) {
12409 					/* Clean it up, somehow we missed the ack? */
12410 					bbr_log_syn(tp, NULL);
12411 				} else {
12412 					/* TFO with data? */
12413 					rsm->r_flags &= ~BBR_HAS_SYN;
12414 					len = rsm->r_end - rsm->r_start;
12415 				}
12416 			} else {
12417 				/* Retransmitting SYN */
12418 				rsm = NULL;
12419 				SOCKBUF_LOCK(sb);
12420 				goto send;
12421 			}
12422 		} else
12423 			len = rsm->r_end - rsm->r_start;
12424 		if ((bbr->rc_resends_use_tso == 0) &&
12425 #ifdef KERN_TLS
12426 		    ((sb->sb_flags & SB_TLS_IFNET) == 0) &&
12427 #endif
12428 		    (len > maxseg)) {
12429 			len = maxseg;
12430 			more_to_rxt = 1;
12431 		}
12432 		sb_offset = rsm->r_start - tp->snd_una;
12433 		if (len > 0) {
12434 			sack_rxmit = 1;
12435 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
12436 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12437 			    min(len, maxseg));
12438 		} else {
12439 			/* I dont think this can happen */
12440 			rsm = NULL;
12441 			goto recheck_resend;
12442 		}
12443 		BBR_STAT_INC(bbr_resends_set);
12444 	} else if (bbr->r_ctl.rc_tlp_send) {
12445 		/*
12446 		 * Tail loss probe
12447 		 */
12448 		doing_tlp = 1;
12449 		rsm = bbr->r_ctl.rc_tlp_send;
12450 		bbr->r_ctl.rc_tlp_send = NULL;
12451 		sack_rxmit = 1;
12452 		len = rsm->r_end - rsm->r_start;
12453 		rtr_cnt++;
12454 		if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12455 			len = maxseg;
12456 
12457 		if (SEQ_GT(tp->snd_una, rsm->r_start)) {
12458 #ifdef BBR_INVARIANTS
12459 			panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u",
12460 			    tp, bbr, tp->snd_una, rsm, rsm->r_start);
12461 #else
12462 			/* TSNH */
12463 			rsm = NULL;
12464 			goto recheck_resend;
12465 #endif
12466 		}
12467 		sb_offset = rsm->r_start - tp->snd_una;
12468 		BBR_STAT_INC(bbr_tlp_set);
12469 	}
12470 	/*
12471 	 * Enforce a connection sendmap count limit if set
12472 	 * as long as we are not retransmiting.
12473 	 */
12474 	if ((rsm == NULL) &&
12475 	    (V_tcp_map_entries_limit > 0) &&
12476 	    (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
12477 		BBR_STAT_INC(bbr_alloc_limited);
12478 		if (!bbr->alloc_limit_reported) {
12479 			bbr->alloc_limit_reported = 1;
12480 			BBR_STAT_INC(bbr_alloc_limited_conns);
12481 		}
12482 		goto just_return_nolock;
12483 	}
12484 #ifdef BBR_INVARIANTS
12485 	if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) {
12486 		panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u",
12487 		    tp, bbr, rsm, sb_offset, len);
12488 	}
12489 #endif
12490 	/*
12491 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
12492 	 * state flags.
12493 	 */
12494 	if (tp->t_flags & TF_NEEDFIN && (rsm == NULL))
12495 		flags |= TH_FIN;
12496 	if (tp->t_flags & TF_NEEDSYN)
12497 		flags |= TH_SYN;
12498 
12499 	if (rsm && (rsm->r_flags & BBR_HAS_FIN)) {
12500 		/* we are retransmitting the fin */
12501 		len--;
12502 		if (len) {
12503 			/*
12504 			 * When retransmitting data do *not* include the
12505 			 * FIN. This could happen from a TLP probe if we
12506 			 * allowed data with a FIN.
12507 			 */
12508 			flags &= ~TH_FIN;
12509 		}
12510 	} else if (rsm) {
12511 		if (flags & TH_FIN)
12512 			flags &= ~TH_FIN;
12513 	}
12514 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12515 		void *end_rsm;
12516 
12517 		end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
12518 		if (end_rsm)
12519 			kern_prefetch(end_rsm, &prefetch_rsm);
12520 		prefetch_rsm = 1;
12521 	}
12522 	SOCKBUF_LOCK(sb);
12523 	/*
12524 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
12525 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12526 	 * negative length.  This can also occur when TCP opens up its
12527 	 * congestion window while receiving additional duplicate acks after
12528 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
12529 	 * the fast-retransmit.
12530 	 *
12531 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
12532 	 * set to snd_una, the sb_offset will be 0, and the length may wind
12533 	 * up 0.
12534 	 *
12535 	 * If sack_rxmit is true we are retransmitting from the scoreboard
12536 	 * in which case len is already set.
12537 	 */
12538 	if (sack_rxmit == 0) {
12539 		uint32_t avail;
12540 
12541 		avail = sbavail(sb);
12542 		if (SEQ_GT(tp->snd_max, tp->snd_una))
12543 			sb_offset = tp->snd_max - tp->snd_una;
12544 		else
12545 			sb_offset = 0;
12546 		if (bbr->rc_tlp_new_data) {
12547 			/* TLP is forcing out new data */
12548 			uint32_t tlplen;
12549 
12550 			doing_tlp = 1;
12551 			tlplen = maxseg;
12552 
12553 			if (tlplen > (uint32_t)(avail - sb_offset)) {
12554 				tlplen = (uint32_t)(avail - sb_offset);
12555 			}
12556 			if (tlplen > tp->snd_wnd) {
12557 				len = tp->snd_wnd;
12558 			} else {
12559 				len = tlplen;
12560 			}
12561 			bbr->rc_tlp_new_data = 0;
12562 		} else {
12563 			what_we_can = len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts);
12564 			if ((len < p_maxseg) &&
12565 			    (bbr->rc_in_persist == 0) &&
12566 			    (ctf_outstanding(tp) >= (2 * p_maxseg)) &&
12567 			    ((avail - sb_offset) >= p_maxseg)) {
12568 				/*
12569 				 * We are not completing whats in the socket
12570 				 * buffer (i.e. there is at least a segment
12571 				 * waiting to send) and we have 2 or more
12572 				 * segments outstanding. There is no sense
12573 				 * of sending a little piece. Lets defer and
12574 				 * and wait until we can send a whole
12575 				 * segment.
12576 				 */
12577 				len = 0;
12578 			}
12579 			if (bbr->rc_in_persist) {
12580 				/*
12581 				 * We are in persists, figure out if
12582 				 * a retransmit is available (maybe the previous
12583 				 * persists we sent) or if we have to send new
12584 				 * data.
12585 				 */
12586 				rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
12587 				if (rsm) {
12588 					len = rsm->r_end - rsm->r_start;
12589 					if (rsm->r_flags & BBR_HAS_FIN)
12590 						len--;
12591 					if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12592 						len = maxseg;
12593 					if (len > 1)
12594 						BBR_STAT_INC(bbr_persist_reneg);
12595 					/*
12596 					 * XXXrrs we could force the len to
12597 					 * 1 byte here to cause the chunk to
12598 					 * split apart.. but that would then
12599 					 * mean we always retransmit it as
12600 					 * one byte even after the window
12601 					 * opens.
12602 					 */
12603 					sack_rxmit = 1;
12604 					sb_offset = rsm->r_start - tp->snd_una;
12605 				} else {
12606 					/*
12607 					 * First time through in persists or peer
12608 					 * acked our one byte. Though we do have
12609 					 * to have something in the sb.
12610 					 */
12611 					len = 1;
12612 					sb_offset = 0;
12613 					if (avail == 0)
12614 					    len = 0;
12615 				}
12616 			}
12617 		}
12618 	}
12619 	if (prefetch_so_done == 0) {
12620 		kern_prefetch(so, &prefetch_so_done);
12621 		prefetch_so_done = 1;
12622 	}
12623 	/*
12624 	 * Lop off SYN bit if it has already been sent.  However, if this is
12625 	 * SYN-SENT state and if segment contains data and if we don't know
12626 	 * that foreign host supports TAO, suppress sending segment.
12627 	 */
12628 	if ((flags & TH_SYN) && (rsm == NULL) &&
12629 	    SEQ_GT(tp->snd_max, tp->snd_una)) {
12630 		if (tp->t_state != TCPS_SYN_RECEIVED)
12631 			flags &= ~TH_SYN;
12632 		/*
12633 		 * When sending additional segments following a TFO SYN|ACK,
12634 		 * do not include the SYN bit.
12635 		 */
12636 		if (IS_FASTOPEN(tp->t_flags) &&
12637 		    (tp->t_state == TCPS_SYN_RECEIVED))
12638 			flags &= ~TH_SYN;
12639 		sb_offset--, len++;
12640 		if (sbavail(sb) == 0)
12641 			len = 0;
12642 	} else if ((flags & TH_SYN) && rsm) {
12643 		/*
12644 		 * Subtract one from the len for the SYN being
12645 		 * retransmitted.
12646 		 */
12647 		len--;
12648 	}
12649 	/*
12650 	 * Be careful not to send data and/or FIN on SYN segments. This
12651 	 * measure is needed to prevent interoperability problems with not
12652 	 * fully conformant TCP implementations.
12653 	 */
12654 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12655 		len = 0;
12656 		flags &= ~TH_FIN;
12657 	}
12658 	/*
12659 	 * On TFO sockets, ensure no data is sent in the following cases:
12660 	 *
12661 	 *  - When retransmitting SYN|ACK on a passively-created socket
12662 	 *  - When retransmitting SYN on an actively created socket
12663 	 *  - When sending a zero-length cookie (cookie request) on an
12664 	 *    actively created socket
12665 	 *  - When the socket is in the CLOSED state (RST is being sent)
12666 	 */
12667 	if (IS_FASTOPEN(tp->t_flags) &&
12668 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12669 	     ((tp->t_state == TCPS_SYN_SENT) &&
12670 	      (tp->t_tfo_client_cookie_len == 0)) ||
12671 	     (flags & TH_RST))) {
12672 		len = 0;
12673 		sack_rxmit = 0;
12674 		rsm = NULL;
12675 	}
12676 	/* Without fast-open there should never be data sent on a SYN */
12677 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags)))
12678 		len = 0;
12679 	if (len <= 0) {
12680 		/*
12681 		 * If FIN has been sent but not acked, but we haven't been
12682 		 * called to retransmit, len will be < 0.  Otherwise, window
12683 		 * shrank after we sent into it.  If window shrank to 0,
12684 		 * cancel pending retransmit, pull snd_nxt back to (closed)
12685 		 * window, and set the persist timer if it isn't already
12686 		 * going.  If the window didn't close completely, just wait
12687 		 * for an ACK.
12688 		 *
12689 		 * We also do a general check here to ensure that we will
12690 		 * set the persist timer when we have data to send, but a
12691 		 * 0-byte window. This makes sure the persist timer is set
12692 		 * even if the packet hits one of the "goto send" lines
12693 		 * below.
12694 		 */
12695 		len = 0;
12696 		if ((tp->snd_wnd == 0) &&
12697 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12698 		    (tp->snd_una == tp->snd_max) &&
12699 		    (sb_offset < (int)sbavail(sb))) {
12700 			/*
12701 			 * Not enough room in the rwnd to send
12702 			 * a paced segment out.
12703 			 */
12704 			bbr_enter_persist(tp, bbr, cts, __LINE__);
12705 		}
12706 	} else if ((rsm == NULL) &&
12707 		   (doing_tlp == 0) &&
12708 		   (len < bbr->r_ctl.rc_pace_max_segs)) {
12709 		/*
12710 		 * We are not sending a full segment for
12711 		 * some reason. Should we not send anything (think
12712 		 * sws or persists)?
12713 		 */
12714 		if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12715 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12716 		    (len < (int)(sbavail(sb) - sb_offset))) {
12717 			/*
12718 			 * Here the rwnd is less than
12719 			 * the pacing size, this is not a retransmit,
12720 			 * we are established and
12721 			 * the send is not the last in the socket buffer
12722 			 * lets not send, and possibly enter persists.
12723 			 */
12724 			len = 0;
12725 			if (tp->snd_max == tp->snd_una)
12726 				bbr_enter_persist(tp, bbr, cts, __LINE__);
12727 		} else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) &&
12728 			   (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12729 						 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12730 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12731 			   (len < bbr_minseg(bbr))) {
12732 			/*
12733 			 * Here we are not retransmitting, and
12734 			 * the cwnd is not so small that we could
12735 			 * not send at least a min size (rxt timer
12736 			 * not having gone off), We have 2 segments or
12737 			 * more already in flight, its not the tail end
12738 			 * of the socket buffer  and the cwnd is blocking
12739 			 * us from sending out minimum pacing segment size.
12740 			 * Lets not send anything.
12741 			 */
12742 			bbr->rc_cwnd_limited = 1;
12743 			len = 0;
12744 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12745 			    min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12746 			   (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12747 						 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12748 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12749 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
12750 			/*
12751 			 * Here we have a send window but we have
12752 			 * filled it up and we can't send another pacing segment.
12753 			 * We also have in flight more than 2 segments
12754 			 * and we are not completing the sb i.e. we allow
12755 			 * the last bytes of the sb to go out even if
12756 			 * its not a full pacing segment.
12757 			 */
12758 			len = 0;
12759 		}
12760 	}
12761 	/* len will be >= 0 after this point. */
12762 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12763 	tcp_sndbuf_autoscale(tp, so, sendwin);
12764 	/*
12765 	 *
12766 	 */
12767 	if (bbr->rc_in_persist &&
12768 	    len &&
12769 	    (rsm == NULL) &&
12770 	    (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) {
12771 		/*
12772 		 * We are in persist, not doing a retransmit and don't have enough space
12773 		 * yet to send a full TSO. So is it at the end of the sb
12774 		 * if so we need to send else nuke to 0 and don't send.
12775 		 */
12776 		int sbleft;
12777 		if (sbavail(sb) > sb_offset)
12778 			sbleft = sbavail(sb) - sb_offset;
12779 		else
12780 			sbleft = 0;
12781 		if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) {
12782 			/* not at end of sb lets not send */
12783 			len = 0;
12784 		}
12785 	}
12786 	/*
12787 	 * Decide if we can use TCP Segmentation Offloading (if supported by
12788 	 * hardware).
12789 	 *
12790 	 * TSO may only be used if we are in a pure bulk sending state.  The
12791 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12792 	 * options prevent using TSO.  With TSO the TCP header is the same
12793 	 * (except for the sequence number) for all generated packets.  This
12794 	 * makes it impossible to transmit any options which vary per
12795 	 * generated segment or packet.
12796 	 *
12797 	 * IPv4 handling has a clear separation of ip options and ip header
12798 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen()
12799 	 * does the right thing below to provide length of just ip options
12800 	 * and thus checking for ipoptlen is enough to decide if ip options
12801 	 * are present.
12802 	 */
12803 #ifdef INET6
12804 	if (isipv6)
12805 		ipoptlen = ip6_optlen(inp);
12806 	else
12807 #endif
12808 	if (inp->inp_options)
12809 		ipoptlen = inp->inp_options->m_len -
12810 		    offsetof(struct ipoption, ipopt_list);
12811 	else
12812 		ipoptlen = 0;
12813 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12814 	/*
12815 	 * Pre-calculate here as we save another lookup into the darknesses
12816 	 * of IPsec that way and can actually decide if TSO is ok.
12817 	 */
12818 #ifdef INET6
12819 	if (isipv6 && IPSEC_ENABLED(ipv6))
12820 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
12821 #ifdef INET
12822 	else
12823 #endif
12824 #endif				/* INET6 */
12825 #ifdef INET
12826 	if (IPSEC_ENABLED(ipv4))
12827 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
12828 #endif				/* INET */
12829 #endif				/* IPSEC */
12830 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12831 	ipoptlen += ipsec_optlen;
12832 #endif
12833 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
12834 	    (len > maxseg) &&
12835 	    (tp->t_port == 0) &&
12836 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
12837 	    tp->rcv_numsacks == 0 &&
12838 	    ipoptlen == 0)
12839 		tso = 1;
12840 
12841 	recwin = min(max(sbspace(&so->so_rcv), 0),
12842 	    TCP_MAXWIN << tp->rcv_scale);
12843 	/*
12844 	 * Sender silly window avoidance.   We transmit under the following
12845 	 * conditions when len is non-zero:
12846 	 *
12847 	 * - We have a full segment (or more with TSO) - This is the last
12848 	 * buffer in a write()/send() and we are either idle or running
12849 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
12850 	 * then 1/2 the maximum send window's worth of data (receiver may be
12851 	 * limited the window size) - we need to retransmit
12852 	 */
12853 	if (rsm)
12854 		goto send;
12855 	if (len) {
12856 		if (sack_rxmit)
12857 			goto send;
12858 		if (len >= p_maxseg)
12859 			goto send;
12860 		/*
12861 		 * NOTE! on localhost connections an 'ack' from the remote
12862 		 * end may occur synchronously with the output and cause us
12863 		 * to flush a buffer queued with moretocome.  XXX
12864 		 *
12865 		 */
12866 		if (((tp->t_flags & TF_MORETOCOME) == 0) &&	/* normal case */
12867 		    ((tp->t_flags & TF_NODELAY) ||
12868 		    ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) &&
12869 		    (tp->t_flags & TF_NOPUSH) == 0) {
12870 			goto send;
12871 		}
12872 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
12873 			goto send;
12874 		}
12875 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
12876 			goto send;
12877 		}
12878 	}
12879 	/*
12880 	 * Sending of standalone window updates.
12881 	 *
12882 	 * Window updates are important when we close our window due to a
12883 	 * full socket buffer and are opening it again after the application
12884 	 * reads data from it.  Once the window has opened again and the
12885 	 * remote end starts to send again the ACK clock takes over and
12886 	 * provides the most current window information.
12887 	 *
12888 	 * We must avoid the silly window syndrome whereas every read from
12889 	 * the receive buffer, no matter how small, causes a window update
12890 	 * to be sent.  We also should avoid sending a flurry of window
12891 	 * updates when the socket buffer had queued a lot of data and the
12892 	 * application is doing small reads.
12893 	 *
12894 	 * Prevent a flurry of pointless window updates by only sending an
12895 	 * update when we can increase the advertized window by more than
12896 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
12897 	 * full or is very small be more aggressive and send an update
12898 	 * whenever we can increase by two mss sized segments. In all other
12899 	 * situations the ACK's to new incoming data will carry further
12900 	 * window increases.
12901 	 *
12902 	 * Don't send an independent window update if a delayed ACK is
12903 	 * pending (it will get piggy-backed on it) or the remote side
12904 	 * already has done a half-close and won't send more data.  Skip
12905 	 * this if the connection is in T/TCP half-open state.
12906 	 */
12907 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
12908 	    !(tp->t_flags & TF_DELACK) &&
12909 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
12910 		/* Check to see if we should do a window update */
12911 		if (bbr_window_update_needed(tp, so, recwin, maxseg))
12912 			goto send;
12913 	}
12914 	/*
12915 	 * Send if we owe the peer an ACK, RST, SYN.  ACKNOW
12916 	 * is also a catch-all for the retransmit timer timeout case.
12917 	 */
12918 	if (tp->t_flags & TF_ACKNOW) {
12919 		goto send;
12920 	}
12921 	if (flags & TH_RST) {
12922 		/* Always send a RST if one is due */
12923 		goto send;
12924 	}
12925 	if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) {
12926 		goto send;
12927 	}
12928 	/*
12929 	 * If our state indicates that FIN should be sent and we have not
12930 	 * yet done so, then we need to send.
12931 	 */
12932 	if (flags & TH_FIN &&
12933 	    ((tp->t_flags & TF_SENTFIN) == 0)) {
12934 		goto send;
12935 	}
12936 	/*
12937 	 * No reason to send a segment, just return.
12938 	 */
12939 just_return:
12940 	SOCKBUF_UNLOCK(sb);
12941 just_return_nolock:
12942 	if (tot_len)
12943 		slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
12944 	if (bbr->rc_no_pacing)
12945 		slot = 0;
12946 	if (tot_len == 0) {
12947 		if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >=
12948 		    tp->snd_wnd) {
12949 			BBR_STAT_INC(bbr_rwnd_limited);
12950 			app_limited = BBR_JR_RWND_LIMITED;
12951 			bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12952 			if ((bbr->rc_in_persist == 0) &&
12953 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
12954 			    (tp->snd_max == tp->snd_una) &&
12955 			    sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
12956 				/* No send window.. we must enter persist */
12957 				bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
12958 			}
12959 		} else if (ctf_outstanding(tp) >= sbavail(sb)) {
12960 			BBR_STAT_INC(bbr_app_limited);
12961 			app_limited = BBR_JR_APP_LIMITED;
12962 			bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12963 		} else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12964 						 bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) {
12965 			BBR_STAT_INC(bbr_cwnd_limited);
12966  			app_limited = BBR_JR_CWND_LIMITED;
12967 			bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12968 									bbr->r_ctl.rc_lost_bytes)));
12969 			bbr->rc_cwnd_limited = 1;
12970 		} else {
12971 			BBR_STAT_INC(bbr_app_limited);
12972 			app_limited = BBR_JR_APP_LIMITED;
12973 			bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12974 		}
12975 		bbr->r_ctl.rc_hptsi_agg_delay = 0;
12976 		bbr->r_agg_early_set = 0;
12977 		bbr->r_ctl.rc_agg_early = 0;
12978 		bbr->r_ctl.rc_last_delay_val = 0;
12979 	} else if (bbr->rc_use_google == 0)
12980 		bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12981 	/* Are we app limited? */
12982 	if ((app_limited == BBR_JR_APP_LIMITED) ||
12983 	    (app_limited == BBR_JR_RWND_LIMITED)) {
12984 		/**
12985 		 * We are application limited.
12986 		 */
12987 		bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12988 								       bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered);
12989 	}
12990 	if (tot_len == 0)
12991 		counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1);
12992 	/* Dont update the time if we did not send */
12993 	bbr->r_ctl.rc_last_delay_val = 0;
12994 	bbr->rc_output_starts_timer = 1;
12995 	bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len);
12996 	bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len);
12997 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12998 		/* Make sure snd_nxt is drug up */
12999 		tp->snd_nxt = tp->snd_max;
13000 	}
13001 	return (error);
13002 
13003 send:
13004 	if (doing_tlp == 0) {
13005 		/*
13006 		 * Data not a TLP, and its not the rxt firing. If it is the
13007 		 * rxt firing, we want to leave the tlp_in_progress flag on
13008 		 * so we don't send another TLP. It has to be a rack timer
13009 		 * or normal send (response to acked data) to clear the tlp
13010 		 * in progress flag.
13011 		 */
13012 		bbr->rc_tlp_in_progress = 0;
13013 		bbr->rc_tlp_rtx_out = 0;
13014 	} else {
13015 		/*
13016 		 * Its a TLP.
13017 		 */
13018 		bbr->rc_tlp_in_progress = 1;
13019 	}
13020 	bbr_timer_cancel(bbr, __LINE__, cts);
13021 	if (rsm == NULL) {
13022 		if (sbused(sb) > 0) {
13023 			/*
13024 			 * This is sub-optimal. We only send a stand alone
13025 			 * FIN on its own segment.
13026 			 */
13027 			if (flags & TH_FIN) {
13028 				flags &= ~TH_FIN;
13029 				if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) {
13030 					/* Lets not send this */
13031 					slot = 0;
13032 					goto just_return;
13033 				}
13034 			}
13035 		}
13036 	} else {
13037 		/*
13038 		 * We do *not* send a FIN on a retransmit if it has data.
13039 		 * The if clause here where len > 1 should never come true.
13040 		 */
13041 		if ((len > 0) &&
13042 		    (((rsm->r_flags & BBR_HAS_FIN) == 0) &&
13043 		    (flags & TH_FIN))) {
13044 			flags &= ~TH_FIN;
13045 			len--;
13046 		}
13047 	}
13048 	SOCKBUF_LOCK_ASSERT(sb);
13049 	if (len > 0) {
13050 		if ((tp->snd_una == tp->snd_max) &&
13051 		    (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
13052 			/*
13053 			 * This qualifies as a RTT_PROBE session since we
13054 			 * drop the data outstanding to nothing and waited
13055 			 * more than bbr_rtt_probe_time.
13056 			 */
13057 			bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
13058 			bbr_set_reduced_rtt(bbr, cts, __LINE__);
13059 		}
13060 		if (len >= maxseg)
13061 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
13062 		else
13063 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
13064 	}
13065 	/*
13066 	 * Before ESTABLISHED, force sending of initial options unless TCP
13067 	 * set not to do any options. NOTE: we assume that the IP/TCP header
13068 	 * plus TCP options always fit in a single mbuf, leaving room for a
13069 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
13070 	 * + optlen <= MCLBYTES
13071 	 */
13072 	optlen = 0;
13073 #ifdef INET6
13074 	if (isipv6)
13075 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
13076 	else
13077 #endif
13078 		hdrlen = sizeof(struct tcpiphdr);
13079 
13080 	/*
13081 	 * Compute options for segment. We only have to care about SYN and
13082 	 * established connection segments.  Options for SYN-ACK segments
13083 	 * are handled in TCP syncache.
13084 	 */
13085 	to.to_flags = 0;
13086 	local_options = 0;
13087 	if ((tp->t_flags & TF_NOOPT) == 0) {
13088 		/* Maximum segment size. */
13089 		if (flags & TH_SYN) {
13090 			to.to_mss = tcp_mssopt(&inp->inp_inc);
13091 #ifdef NETFLIX_TCPOUDP
13092 			if (tp->t_port)
13093 				to.to_mss -= V_tcp_udp_tunneling_overhead;
13094 #endif
13095 			to.to_flags |= TOF_MSS;
13096 			/*
13097 			 * On SYN or SYN|ACK transmits on TFO connections,
13098 			 * only include the TFO option if it is not a
13099 			 * retransmit, as the presence of the TFO option may
13100 			 * have caused the original SYN or SYN|ACK to have
13101 			 * been dropped by a middlebox.
13102 			 */
13103 			if (IS_FASTOPEN(tp->t_flags) &&
13104 			    (tp->t_rxtshift == 0)) {
13105 				if (tp->t_state == TCPS_SYN_RECEIVED) {
13106 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
13107 					to.to_tfo_cookie =
13108 					    (u_int8_t *)&tp->t_tfo_cookie.server;
13109 					to.to_flags |= TOF_FASTOPEN;
13110 					wanted_cookie = 1;
13111 				} else if (tp->t_state == TCPS_SYN_SENT) {
13112 					to.to_tfo_len =
13113 					    tp->t_tfo_client_cookie_len;
13114 					to.to_tfo_cookie =
13115 					    tp->t_tfo_cookie.client;
13116 					to.to_flags |= TOF_FASTOPEN;
13117 					wanted_cookie = 1;
13118 				}
13119 			}
13120 		}
13121 		/* Window scaling. */
13122 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
13123 			to.to_wscale = tp->request_r_scale;
13124 			to.to_flags |= TOF_SCALE;
13125 		}
13126 		/* Timestamps. */
13127 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
13128 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
13129 			to.to_tsval = 	tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset;
13130 			to.to_tsecr = tp->ts_recent;
13131 			to.to_flags |= TOF_TS;
13132 			local_options += TCPOLEN_TIMESTAMP + 2;
13133 		}
13134 		/* Set receive buffer autosizing timestamp. */
13135 		if (tp->rfbuf_ts == 0 &&
13136 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
13137 			tp->rfbuf_ts = 	tcp_tv_to_mssectick(&bbr->rc_tv);
13138 		/* Selective ACK's. */
13139 		if (flags & TH_SYN)
13140 			to.to_flags |= TOF_SACKPERM;
13141 		else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13142 		    tp->rcv_numsacks > 0) {
13143 			to.to_flags |= TOF_SACK;
13144 			to.to_nsacks = tp->rcv_numsacks;
13145 			to.to_sacks = (u_char *)tp->sackblks;
13146 		}
13147 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13148 		/* TCP-MD5 (RFC2385). */
13149 		if (tp->t_flags & TF_SIGNATURE)
13150 			to.to_flags |= TOF_SIGNATURE;
13151 #endif				/* TCP_SIGNATURE */
13152 
13153 		/* Processing the options. */
13154 		hdrlen += (optlen = tcp_addoptions(&to, opt));
13155 		/*
13156 		 * If we wanted a TFO option to be added, but it was unable
13157 		 * to fit, ensure no data is sent.
13158 		 */
13159 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
13160 		    !(to.to_flags & TOF_FASTOPEN))
13161 			len = 0;
13162 	}
13163 #ifdef NETFLIX_TCPOUDP
13164 	if (tp->t_port) {
13165 		if (V_tcp_udp_tunneling_port == 0) {
13166 			/* The port was removed?? */
13167 			SOCKBUF_UNLOCK(&so->so_snd);
13168 			return (EHOSTUNREACH);
13169 		}
13170 		hdrlen += sizeof(struct udphdr);
13171 	}
13172 #endif
13173 #ifdef INET6
13174 	if (isipv6)
13175 		ipoptlen = ip6_optlen(tp->t_inpcb);
13176 	else
13177 #endif
13178 	if (tp->t_inpcb->inp_options)
13179 		ipoptlen = tp->t_inpcb->inp_options->m_len -
13180 		    offsetof(struct ipoption, ipopt_list);
13181 	else
13182 		ipoptlen = 0;
13183 	ipoptlen = 0;
13184 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
13185 	ipoptlen += ipsec_optlen;
13186 #endif
13187 	if (bbr->rc_last_options != local_options) {
13188 		/*
13189 		 * Cache the options length this generally does not change
13190 		 * on a connection. We use this to calculate TSO.
13191 		 */
13192 		bbr->rc_last_options = local_options;
13193 	}
13194 	maxseg = tp->t_maxseg - (ipoptlen + optlen);
13195 	p_maxseg = min(maxseg, pace_max_segs);
13196 	/*
13197 	 * Adjust data length if insertion of options will bump the packet
13198 	 * length beyond the t_maxseg length. Clear the FIN bit because we
13199 	 * cut off the tail of the segment.
13200 	 */
13201 #ifdef KERN_TLS
13202  	/* force TSO for so TLS offload can get mss */
13203  	if (sb->sb_flags & SB_TLS_IFNET) {
13204  		force_tso = 1;
13205  	}
13206 #endif
13207 
13208 	if (len > maxseg) {
13209 		if (len != 0 && (flags & TH_FIN)) {
13210 			flags &= ~TH_FIN;
13211 		}
13212 		if (tso) {
13213 			uint32_t moff;
13214 			int32_t max_len;
13215 
13216 			/* extract TSO information */
13217 			if_hw_tsomax = tp->t_tsomax;
13218 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
13219 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
13220 			KASSERT(ipoptlen == 0,
13221 			    ("%s: TSO can't do IP options", __func__));
13222 
13223 			/*
13224 			 * Check if we should limit by maximum payload
13225 			 * length:
13226 			 */
13227 			if (if_hw_tsomax != 0) {
13228 				/* compute maximum TSO length */
13229 				max_len = (if_hw_tsomax - hdrlen -
13230 				    max_linkhdr);
13231 				if (max_len <= 0) {
13232 					len = 0;
13233 				} else if (len > max_len) {
13234 					len = max_len;
13235 				}
13236 			}
13237 			/*
13238 			 * Prevent the last segment from being fractional
13239 			 * unless the send sockbuf can be emptied:
13240 			 */
13241 			if (((sb_offset + len) < sbavail(sb)) &&
13242 			    (hw_tls == 0)) {
13243 				moff = len % (uint32_t)maxseg;
13244 				if (moff != 0) {
13245 					len -= moff;
13246 				}
13247 			}
13248 			/*
13249 			 * In case there are too many small fragments don't
13250 			 * use TSO:
13251 			 */
13252 			if (len <= maxseg) {
13253 				len = maxseg;
13254 				tso = 0;
13255 			}
13256 		} else {
13257 			/* Not doing TSO */
13258 			if (optlen + ipoptlen >= tp->t_maxseg) {
13259 				/*
13260 				 * Since we don't have enough space to put
13261 				 * the IP header chain and the TCP header in
13262 				 * one packet as required by RFC 7112, don't
13263 				 * send it. Also ensure that at least one
13264 				 * byte of the payload can be put into the
13265 				 * TCP segment.
13266 				 */
13267 				SOCKBUF_UNLOCK(&so->so_snd);
13268 				error = EMSGSIZE;
13269 				sack_rxmit = 0;
13270 				goto out;
13271 			}
13272 			len = maxseg;
13273 		}
13274 	} else {
13275 		/* Not doing TSO */
13276 		if_hw_tsomaxsegcount = 0;
13277 		tso = 0;
13278 	}
13279 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
13280 	    ("%s: len > IP_MAXPACKET", __func__));
13281 #ifdef DIAGNOSTIC
13282 #ifdef INET6
13283 	if (max_linkhdr + hdrlen > MCLBYTES)
13284 #else
13285 	if (max_linkhdr + hdrlen > MHLEN)
13286 #endif
13287 		panic("tcphdr too big");
13288 #endif
13289 	/*
13290 	 * This KASSERT is here to catch edge cases at a well defined place.
13291 	 * Before, those had triggered (random) panic conditions further
13292 	 * down.
13293 	 */
13294 #ifdef BBR_INVARIANTS
13295 	if (sack_rxmit) {
13296 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
13297 			panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u",
13298 			    rsm, tp, bbr, rsm->r_start, tp->snd_una);
13299 		}
13300 	}
13301 #endif
13302 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
13303 	if ((len == 0) &&
13304 	    (flags & TH_FIN) &&
13305 	    (sbused(sb))) {
13306 		/*
13307 		 * We have outstanding data, don't send a fin by itself!.
13308 		 */
13309 		slot = 0;
13310 		goto just_return;
13311 	}
13312 	/*
13313 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
13314 	 * and initialize the header from the template for sends on this
13315 	 * connection.
13316 	 */
13317 	if (len) {
13318 		uint32_t moff;
13319 		uint32_t orig_len;
13320 
13321 		/*
13322 		 * We place a limit on sending with hptsi.
13323 		 */
13324 		if ((rsm == NULL) && len > pace_max_segs)
13325 			len = pace_max_segs;
13326 		if (len <= maxseg)
13327 			tso = 0;
13328 #ifdef INET6
13329 		if (MHLEN < hdrlen + max_linkhdr)
13330 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
13331 		else
13332 #endif
13333 			m = m_gethdr(M_NOWAIT, MT_DATA);
13334 
13335 		if (m == NULL) {
13336 			BBR_STAT_INC(bbr_failed_mbuf_aloc);
13337 			bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13338 			SOCKBUF_UNLOCK(sb);
13339 			error = ENOBUFS;
13340 			sack_rxmit = 0;
13341 			goto out;
13342 		}
13343 		m->m_data += max_linkhdr;
13344 		m->m_len = hdrlen;
13345 		/*
13346 		 * Start the m_copy functions from the closest mbuf to the
13347 		 * sb_offset in the socket buffer chain.
13348 		 */
13349 		if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) {
13350 #ifdef BBR_INVARIANTS
13351 			if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0)))
13352 				panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u",
13353 				    tp, bbr, len, sb_offset, sbavail(sb), rsm,
13354 				    doing_retran_from,
13355 				    picked_up_retran,
13356 				    doing_tlp);
13357 
13358 #endif
13359 			/*
13360 			 * In this messed up situation we have two choices,
13361 			 * a) pretend the send worked, and just start timers
13362 			 * and what not (not good since that may lead us
13363 			 * back here a lot). <or> b) Send the lowest segment
13364 			 * in the map. <or> c) Drop the connection. Lets do
13365 			 * <b> which if it continues to happen will lead to
13366 			 * <c> via timeouts.
13367 			 */
13368 			BBR_STAT_INC(bbr_offset_recovery);
13369 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
13370 			sb_offset = 0;
13371 			if (rsm == NULL) {
13372 				sack_rxmit = 0;
13373 				len = sbavail(sb);
13374 			} else {
13375 				sack_rxmit = 1;
13376 				if (rsm->r_start != tp->snd_una) {
13377 					/*
13378 					 * Things are really messed up, <c>
13379 					 * is the only thing to do.
13380 					 */
13381 					BBR_STAT_INC(bbr_offset_drop);
13382 					tcp_set_inp_to_drop(inp, EFAULT);
13383 					return (0);
13384 				}
13385 				len = rsm->r_end - rsm->r_start;
13386 			}
13387 			if (len > sbavail(sb))
13388 				len = sbavail(sb);
13389 			if (len > maxseg)
13390 				len = maxseg;
13391 		}
13392 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
13393 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
13394 			m_copydata(mb, moff, (int)len,
13395 			    mtod(m, caddr_t)+hdrlen);
13396 			if (rsm == NULL)
13397 				sbsndptr_adv(sb, mb, len);
13398 			m->m_len += len;
13399 		} else {
13400 			struct sockbuf *msb;
13401 
13402 			if (rsm)
13403 				msb = NULL;
13404 			else
13405 				msb = sb;
13406 #ifdef BBR_INVARIANTS
13407 			if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) {
13408 				if (rsm) {
13409 					panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u rsm:%p snd_una:%u rsm_start:%u flg:%x %u:%u:%u sr:%d ",
13410 					    tp, bbr, len, moff,
13411 					    sbavail(sb), rsm,
13412 					    tp->snd_una, rsm->r_flags, rsm->r_start,
13413 					    doing_retran_from,
13414 					    picked_up_retran,
13415 					    doing_tlp, sack_rxmit);
13416 				} else {
13417 					panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u",
13418 					    tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una);
13419 				}
13420 			}
13421 #endif
13422 			orig_len = len;
13423 			m->m_next = tcp_m_copym(
13424 				mb, moff, &len,
13425 				if_hw_tsomaxsegcount,
13426 				if_hw_tsomaxsegsize, msb,
13427 				((rsm == NULL) ? hw_tls : 0)
13428 #ifdef NETFLIX_COPY_ARGS
13429 				, &filled_all
13430 #endif
13431 				);
13432 			if (len <= maxseg && !force_tso) {
13433 				/*
13434 				 * Must have ran out of mbufs for the copy
13435 				 * shorten it to no longer need tso. Lets
13436 				 * not put on sendalot since we are low on
13437 				 * mbufs.
13438 				 */
13439 				tso = 0;
13440 			}
13441 			if (m->m_next == NULL) {
13442 				SOCKBUF_UNLOCK(sb);
13443 				(void)m_free(m);
13444 				error = ENOBUFS;
13445 				sack_rxmit = 0;
13446 				goto out;
13447 			}
13448 		}
13449 #ifdef BBR_INVARIANTS
13450 		if (tso && len < maxseg) {
13451 			panic("tp:%p tso on, but len:%d < maxseg:%d",
13452 			    tp, len, maxseg);
13453 		}
13454 		if (tso && if_hw_tsomaxsegcount) {
13455 			int32_t seg_cnt = 0;
13456 			struct mbuf *foo;
13457 
13458 			foo = m;
13459 			while (foo) {
13460 				seg_cnt++;
13461 				foo = foo->m_next;
13462 			}
13463 			if (seg_cnt > if_hw_tsomaxsegcount) {
13464 				panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount);
13465 			}
13466 		}
13467 #endif
13468 		/*
13469 		 * If we're sending everything we've got, set PUSH. (This
13470 		 * will keep happy those implementations which only give
13471 		 * data to the user when a buffer fills or a PUSH comes in.)
13472 		 */
13473 		if (sb_offset + len == sbused(sb) &&
13474 		    sbused(sb) &&
13475 		    !(flags & TH_SYN)) {
13476 			flags |= TH_PUSH;
13477 		}
13478 		SOCKBUF_UNLOCK(sb);
13479 	} else {
13480 		SOCKBUF_UNLOCK(sb);
13481 		if (tp->t_flags & TF_ACKNOW)
13482 			KMOD_TCPSTAT_INC(tcps_sndacks);
13483 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
13484 			KMOD_TCPSTAT_INC(tcps_sndctrl);
13485 		else
13486 			KMOD_TCPSTAT_INC(tcps_sndwinup);
13487 
13488 		m = m_gethdr(M_NOWAIT, MT_DATA);
13489 		if (m == NULL) {
13490 			BBR_STAT_INC(bbr_failed_mbuf_aloc);
13491 			bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13492 			error = ENOBUFS;
13493 			/* Fudge the send time since we could not send */
13494 			sack_rxmit = 0;
13495 			goto out;
13496 		}
13497 #ifdef INET6
13498 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13499 		    MHLEN >= hdrlen) {
13500 			M_ALIGN(m, hdrlen);
13501 		} else
13502 #endif
13503 			m->m_data += max_linkhdr;
13504 		m->m_len = hdrlen;
13505 	}
13506 	SOCKBUF_UNLOCK_ASSERT(sb);
13507 	m->m_pkthdr.rcvif = (struct ifnet *)0;
13508 #ifdef MAC
13509 	mac_inpcb_create_mbuf(inp, m);
13510 #endif
13511 #ifdef INET6
13512 	if (isipv6) {
13513 		ip6 = mtod(m, struct ip6_hdr *);
13514 #ifdef NETFLIX_TCPOUDP
13515 		if (tp->t_port) {
13516 			udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr));
13517 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13518 			udp->uh_dport = tp->t_port;
13519 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
13520 			udp->uh_ulen = htons(ulen);
13521 			th = (struct tcphdr *)(udp + 1);
13522 		} else {
13523 #endif
13524 			th = (struct tcphdr *)(ip6 + 1);
13525 
13526 #ifdef NETFLIX_TCPOUDP
13527 		}
13528 #endif
13529 		tcpip_fillheaders(inp,
13530 #ifdef NETFLIX_TCPOUDP
13531 				  tp->t_port,
13532 #endif
13533 				  ip6, th);
13534 	} else
13535 #endif				/* INET6 */
13536 	{
13537 		ip = mtod(m, struct ip *);
13538 #ifdef TCPDEBUG
13539 		ipov = (struct ipovly *)ip;
13540 #endif
13541 #ifdef NETFLIX_TCPOUDP
13542 		if (tp->t_port) {
13543 			udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip));
13544 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13545 			udp->uh_dport = tp->t_port;
13546 			ulen = hdrlen + len - sizeof(struct ip);
13547 			udp->uh_ulen = htons(ulen);
13548 			th = (struct tcphdr *)(udp + 1);
13549 		} else
13550 #endif
13551 			th = (struct tcphdr *)(ip + 1);
13552 		tcpip_fillheaders(inp,
13553 #ifdef NETFLIX_TCPOUDP
13554 				  tp->t_port,
13555 #endif
13556 				  ip, th);
13557 	}
13558 	/*
13559 	 * If we are doing retransmissions, then snd_nxt will not reflect
13560 	 * the first unsent octet.  For ACK only packets, we do not want the
13561 	 * sequence number of the retransmitted packet, we want the sequence
13562 	 * number of the next unsent octet.  So, if there is no data (and no
13563 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
13564 	 * ti_seq.  But if we are in persist state, snd_max might reflect
13565 	 * one byte beyond the right edge of the window, so use snd_nxt in
13566 	 * that case, since we know we aren't doing a retransmission.
13567 	 * (retransmit and persist are mutually exclusive...)
13568 	 */
13569 	if (sack_rxmit == 0) {
13570 		if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) {
13571 			/* New data (including new persists) */
13572 			th->th_seq = htonl(tp->snd_max);
13573 			bbr_seq = tp->snd_max;
13574 		} else if (flags & TH_SYN) {
13575 			/* Syn's always send from iss */
13576 			th->th_seq = htonl(tp->iss);
13577 			bbr_seq = tp->iss;
13578 		} else if (flags & TH_FIN) {
13579 			if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) {
13580 				/*
13581 				 * If we sent the fin already its 1 minus
13582 				 * snd_max
13583 				 */
13584 				th->th_seq = (htonl(tp->snd_max - 1));
13585 				bbr_seq = (tp->snd_max - 1);
13586 			} else {
13587 				/* First time FIN use snd_max */
13588 				th->th_seq = htonl(tp->snd_max);
13589 				bbr_seq = tp->snd_max;
13590 			}
13591 		} else if (flags & TH_RST) {
13592 			/*
13593 			 * For a Reset send the last cum ack in sequence
13594 			 * (this like any other choice may still generate a
13595 			 * challenge ack, if a ack-update packet is in
13596 			 * flight).
13597 			 */
13598 			th->th_seq = htonl(tp->snd_una);
13599 			bbr_seq = tp->snd_una;
13600 		} else {
13601 			/*
13602 			 * len == 0 and not persist we use snd_max, sending
13603 			 * an ack unless we have sent the fin then its 1
13604 			 * minus.
13605 			 */
13606 			/*
13607 			 * XXXRRS Question if we are in persists and we have
13608 			 * nothing outstanding to send and we have not sent
13609 			 * a FIN, we will send an ACK. In such a case it
13610 			 * might be better to send (tp->snd_una - 1) which
13611 			 * would force the peer to ack.
13612 			 */
13613 			if (tp->t_flags & TF_SENTFIN) {
13614 				th->th_seq = htonl(tp->snd_max - 1);
13615 				bbr_seq = (tp->snd_max - 1);
13616 			} else {
13617 				th->th_seq = htonl(tp->snd_max);
13618 				bbr_seq = tp->snd_max;
13619 			}
13620 		}
13621 	} else {
13622 		/* All retransmits use the rsm to guide the send */
13623 		th->th_seq = htonl(rsm->r_start);
13624 		bbr_seq = rsm->r_start;
13625 	}
13626 	th->th_ack = htonl(tp->rcv_nxt);
13627 	if (optlen) {
13628 		bcopy(opt, th + 1, optlen);
13629 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13630 	}
13631 	th->th_flags = flags;
13632 	/*
13633 	 * Calculate receive window.  Don't shrink window, but avoid silly
13634 	 * window syndrome.
13635 	 */
13636 	if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) &&
13637 				  recwin < maxseg)))
13638 		recwin = 0;
13639 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13640 	    recwin < (tp->rcv_adv - tp->rcv_nxt))
13641 		recwin = (tp->rcv_adv - tp->rcv_nxt);
13642 	if (recwin > TCP_MAXWIN << tp->rcv_scale)
13643 		recwin = TCP_MAXWIN << tp->rcv_scale;
13644 
13645 	/*
13646 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13647 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
13648 	 * handled in syncache.
13649 	 */
13650 	if (flags & TH_SYN)
13651 		th->th_win = htons((u_short)
13652 		    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13653 	else {
13654 		/* Avoid shrinking window with window scaling. */
13655 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
13656 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13657 	}
13658 	/*
13659 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13660 	 * window.  This may cause the remote transmitter to stall.  This
13661 	 * flag tells soreceive() to disable delayed acknowledgements when
13662 	 * draining the buffer.  This can occur if the receiver is
13663 	 * attempting to read more data than can be buffered prior to
13664 	 * transmitting on the connection.
13665 	 */
13666 	if (th->th_win == 0) {
13667 		tp->t_sndzerowin++;
13668 		tp->t_flags |= TF_RXWIN0SENT;
13669 	} else
13670 		tp->t_flags &= ~TF_RXWIN0SENT;
13671 	/*
13672 	 * We don't support urgent data, but drag along
13673 	 * the pointer in case of a stack switch.
13674 	 */
13675 	tp->snd_up = tp->snd_una;
13676 
13677 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13678 	if (to.to_flags & TOF_SIGNATURE) {
13679 		/*
13680 		 * Calculate MD5 signature and put it into the place
13681 		 * determined before. NOTE: since TCP options buffer doesn't
13682 		 * point into mbuf's data, calculate offset and use it.
13683 		 */
13684 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13685 		    (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13686 			/*
13687 			 * Do not send segment if the calculation of MD5
13688 			 * digest has failed.
13689 			 */
13690 			goto out;
13691 		}
13692 	}
13693 #endif
13694 
13695 	/*
13696 	 * Put TCP length in extended header, and then checksum extended
13697 	 * header and data.
13698 	 */
13699 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
13700 #ifdef INET6
13701 	if (isipv6) {
13702 		/*
13703 		 * ip6_plen is not need to be filled now, and will be filled
13704 		 * in ip6_output.
13705 		 */
13706 #ifdef NETFLIX_TCPOUDP
13707 		if (tp->t_port) {
13708 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13709 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13710 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13711 			th->th_sum = htons(0);
13712 			UDPSTAT_INC(udps_opackets);
13713 		} else {
13714 #endif
13715 			csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13716 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13717 			th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
13718 			    optlen + len, IPPROTO_TCP, 0);
13719 #ifdef NETFLIX_TCPOUDP
13720 		}
13721 #endif
13722 	}
13723 #endif
13724 #if defined(INET6) && defined(INET)
13725 	else
13726 #endif
13727 #ifdef INET
13728 	{
13729 #ifdef NETFLIX_TCPOUDP
13730 		if (tp->t_port) {
13731 			m->m_pkthdr.csum_flags = CSUM_UDP;
13732 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13733 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13734 			    ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13735 			th->th_sum = htons(0);
13736 			UDPSTAT_INC(udps_opackets);
13737 		} else {
13738 #endif
13739 			csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP;
13740 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13741 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
13742 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13743 			    IPPROTO_TCP + len + optlen));
13744 #ifdef NETFLIX_TCPOUDP
13745 		}
13746 #endif
13747 		/* IP version must be set here for ipv4/ipv6 checking later */
13748 		KASSERT(ip->ip_v == IPVERSION,
13749 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
13750 	}
13751 #endif
13752 
13753 	/*
13754 	 * Enable TSO and specify the size of the segments. The TCP pseudo
13755 	 * header checksum is always provided. XXX: Fixme: This is currently
13756 	 * not the case for IPv6.
13757 	 */
13758 	if (tso || force_tso) {
13759 		KASSERT(force_tso || len > maxseg,
13760 		    ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg));
13761 		m->m_pkthdr.csum_flags |= CSUM_TSO;
13762 		csum_flags |= CSUM_TSO;
13763 		m->m_pkthdr.tso_segsz = maxseg;
13764 	}
13765 	KASSERT(len + hdrlen == m_length(m, NULL),
13766 	    ("%s: mbuf chain different than expected: %d + %u != %u",
13767 	    __func__, len, hdrlen, m_length(m, NULL)));
13768 
13769 #ifdef TCP_HHOOK
13770 	/* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */
13771 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
13772 #endif
13773 #ifdef TCPDEBUG
13774 	/*
13775 	 * Trace.
13776 	 */
13777 	if (so->so_options & SO_DEBUG) {
13778 		u_short save = 0;
13779 
13780 #ifdef INET6
13781 		if (!isipv6)
13782 #endif
13783 		{
13784 			save = ipov->ih_len;
13785 			ipov->ih_len = htons(m->m_pkthdr.len	/* - hdrlen +
13786 			      * (th->th_off << 2) */ );
13787 		}
13788 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
13789 #ifdef INET6
13790 		if (!isipv6)
13791 #endif
13792 			ipov->ih_len = save;
13793 	}
13794 #endif				/* TCPDEBUG */
13795 
13796 	/* Log to the black box */
13797 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13798 		union tcp_log_stackspecific log;
13799 
13800 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
13801 		/* Record info on type of transmission */
13802 		log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay;
13803 		log.u_bbr.flex2 = (bbr->r_recovery_bw << 3);
13804 		log.u_bbr.flex3 = maxseg;
13805 		log.u_bbr.flex4 = delay_calc;
13806 		/* Encode filled_all into the upper flex5 bit */
13807 		log.u_bbr.flex5 = bbr->rc_past_init_win;
13808 		log.u_bbr.flex5 <<= 1;
13809 		log.u_bbr.flex5 |= bbr->rc_no_pacing;
13810 		log.u_bbr.flex5 <<= 29;
13811 		if (filled_all)
13812 			log.u_bbr.flex5 |= 0x80000000;
13813 		log.u_bbr.flex5 |= tp->t_maxseg;
13814 		log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs;
13815 		log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr);
13816 		/* lets poke in the low and the high here for debugging */
13817 		log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
13818 		if (rsm || sack_rxmit) {
13819 			if (doing_tlp)
13820 				log.u_bbr.flex8 = 2;
13821 			else
13822 				log.u_bbr.flex8 = 1;
13823 		} else {
13824 			log.u_bbr.flex8 = 0;
13825 		}
13826 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13827 		    len, &log, false, NULL, NULL, 0, tv);
13828 	} else {
13829 		lgb = NULL;
13830 	}
13831 	/*
13832 	 * Fill in IP length and desired time to live and send to IP level.
13833 	 * There should be a better way to handle ttl and tos; we could keep
13834 	 * them in the template, but need a way to checksum without them.
13835 	 */
13836 	/*
13837 	 * m->m_pkthdr.len should have been set before cksum calcuration,
13838 	 * because in6_cksum() need it.
13839 	 */
13840 #ifdef INET6
13841 	if (isipv6) {
13842 		/*
13843 		 * we separately set hoplimit for every segment, since the
13844 		 * user might want to change the value via setsockopt. Also,
13845 		 * desired default hop limit might be changed via Neighbor
13846 		 * Discovery.
13847 		 */
13848 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13849 
13850 		/*
13851 		 * Set the packet size here for the benefit of DTrace
13852 		 * probes. ip6_output() will set it properly; it's supposed
13853 		 * to include the option header lengths as well.
13854 		 */
13855 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13856 
13857 		if (V_path_mtu_discovery && maxseg > V_tcp_minmss)
13858 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13859 		else
13860 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13861 
13862 		if (tp->t_state == TCPS_SYN_SENT)
13863 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13864 
13865 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13866 		/* TODO: IPv6 IP6TOS_ECT bit on */
13867 		error = ip6_output(m, inp->in6p_outputopts,
13868 		    &inp->inp_route6,
13869 		    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13870 		    NULL, NULL, inp);
13871 
13872 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
13873 			mtu = inp->inp_route6.ro_nh->nh_mtu;
13874 	}
13875 #endif				/* INET6 */
13876 #if defined(INET) && defined(INET6)
13877 	else
13878 #endif
13879 #ifdef INET
13880 	{
13881 		ip->ip_len = htons(m->m_pkthdr.len);
13882 #ifdef INET6
13883 		if (isipv6)
13884 			ip->ip_ttl = in6_selecthlim(inp, NULL);
13885 #endif				/* INET6 */
13886 		/*
13887 		 * If we do path MTU discovery, then we set DF on every
13888 		 * packet. This might not be the best thing to do according
13889 		 * to RFC3390 Section 2. However the tcp hostcache migitates
13890 		 * the problem so it affects only the first tcp connection
13891 		 * with a host.
13892 		 *
13893 		 * NB: Don't set DF on small MTU/MSS to have a safe
13894 		 * fallback.
13895 		 */
13896 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13897 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13898 			if (tp->t_port == 0 || len < V_tcp_minmss) {
13899 				ip->ip_off |= htons(IP_DF);
13900 			}
13901 		} else {
13902 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13903 		}
13904 
13905 		if (tp->t_state == TCPS_SYN_SENT)
13906 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
13907 
13908 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
13909 
13910 		error = ip_output(m, inp->inp_options, &inp->inp_route,
13911 		    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
13912 		    inp);
13913 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
13914 			mtu = inp->inp_route.ro_nh->nh_mtu;
13915 	}
13916 #endif				/* INET */
13917 out:
13918 
13919 	if (lgb) {
13920 		lgb->tlb_errno = error;
13921 		lgb = NULL;
13922 	}
13923 	/*
13924 	 * In transmit state, time the transmission and arrange for the
13925 	 * retransmit.  In persist state, just set snd_max.
13926 	 */
13927 	if (error == 0) {
13928 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13929 		    (tp->t_flags & TF_SACK_PERMIT) &&
13930 		    tp->rcv_numsacks > 0)
13931 			tcp_clean_dsack_blocks(tp);
13932 		/* We sent an ack clear the bbr_segs_rcvd count */
13933 		bbr->output_error_seen = 0;
13934 		bbr->oerror_cnt = 0;
13935 		bbr->bbr_segs_rcvd = 0;
13936 		if (len == 0)
13937 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1);
13938 		else if (hw_tls) {
13939 			if (filled_all ||
13940 			    (len >= bbr->r_ctl.rc_pace_max_segs))
13941 				BBR_STAT_INC(bbr_meets_tso_thresh);
13942 			else {
13943 				if (doing_tlp) {
13944 					BBR_STAT_INC(bbr_miss_tlp);
13945 					bbr_log_type_hrdwtso(tp, bbr, len, 1, what_we_can);
13946 
13947 
13948 				} else if (rsm) {
13949 					BBR_STAT_INC(bbr_miss_retran);
13950 					bbr_log_type_hrdwtso(tp, bbr, len, 2, what_we_can);
13951 				} else if ((ctf_outstanding(tp) + bbr->r_ctl.rc_pace_max_segs) > sbavail(sb)) {
13952 					BBR_STAT_INC(bbr_miss_tso_app);
13953 					bbr_log_type_hrdwtso(tp, bbr, len, 3, what_we_can);
13954 				} else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13955 								 bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_pace_max_segs) > tp->snd_cwnd) {
13956 					BBR_STAT_INC(bbr_miss_tso_cwnd);
13957 					bbr_log_type_hrdwtso(tp, bbr, len, 4, what_we_can);
13958 				} else if ((ctf_outstanding(tp) + bbr->r_ctl.rc_pace_max_segs) > tp->snd_wnd) {
13959 					BBR_STAT_INC(bbr_miss_tso_rwnd);
13960 					bbr_log_type_hrdwtso(tp, bbr, len, 5, what_we_can);
13961 				} else {
13962 					BBR_STAT_INC(bbr_miss_unknown);
13963 					bbr_log_type_hrdwtso(tp, bbr, len, 6, what_we_can);
13964 				}
13965 			}
13966 		}
13967 		/* Do accounting for new sends */
13968 		if ((len > 0) && (rsm == NULL)) {
13969 			int idx;
13970 			if (tp->snd_una == tp->snd_max) {
13971 				/*
13972 				 * Special case to match google, when
13973 				 * nothing is in flight the delivered
13974 				 * time does get updated to the current
13975 				 * time (see tcp_rate_bsd.c).
13976 				 */
13977 				bbr->r_ctl.rc_del_time = cts;
13978 			}
13979 			if (len >= maxseg) {
13980 				idx = (len / maxseg) + 3;
13981 				if (idx >= TCP_MSS_ACCT_ATIMER)
13982 					counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1);
13983 				else
13984 					counter_u64_add(bbr_out_size[idx], 1);
13985 			} else {
13986 				/* smaller than a MSS */
13987 				idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options);
13988 				if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV)
13989 					idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1);
13990 				counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1);
13991 			}
13992 		}
13993 	}
13994 	abandon = 0;
13995 	/*
13996 	 * We must do the send accounting before we log the output,
13997 	 * otherwise the state of the rsm could change and we account to the
13998 	 * wrong bucket.
13999 	 */
14000 	if (len > 0) {
14001 		bbr_do_send_accounting(tp, bbr, rsm, len, error);
14002 		if (error == 0) {
14003 			if (tp->snd_una == tp->snd_max)
14004 				bbr->r_ctl.rc_tlp_rxt_last_time = cts;
14005 		}
14006 	}
14007 	bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error,
14008 	    cts, mb, &abandon, rsm, 0, sb);
14009 	if (abandon) {
14010 		/*
14011 		 * If bbr_log_output destroys the TCB or sees a TH_RST being
14012 		 * sent we should hit this condition.
14013 		 */
14014 		return (0);
14015 	}
14016 	if (bbr->rc_in_persist == 0) {
14017 		/*
14018 		 * Advance snd_nxt over sequence space of this segment.
14019 		 */
14020 		if (error)
14021 			/* We don't log or do anything with errors */
14022 			goto skip_upd;
14023 
14024 		if (tp->snd_una == tp->snd_max &&
14025 		    (len || (flags & (TH_SYN | TH_FIN)))) {
14026 			/*
14027 			 * Update the time we just added data since none was
14028 			 * outstanding.
14029 			 */
14030 			bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
14031 			bbr->rc_tp->t_acktime  = ticks;
14032 		}
14033 		if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) {
14034 			if (flags & TH_SYN) {
14035 				/*
14036 				 * Smack the snd_max to iss + 1
14037 				 * if its a FO we will add len below.
14038 				 */
14039 				tp->snd_max = tp->iss + 1;
14040 			}
14041 			if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
14042 				tp->snd_max++;
14043 				tp->t_flags |= TF_SENTFIN;
14044 			}
14045 		}
14046 		if (sack_rxmit == 0)
14047 			tp->snd_max += len;
14048 skip_upd:
14049 		if ((error == 0) && len)
14050 			tot_len += len;
14051 	} else {
14052 		/* Persists case */
14053 		int32_t xlen = len;
14054 
14055 		if (error)
14056 			goto nomore;
14057 
14058 		if (flags & TH_SYN)
14059 			++xlen;
14060 		if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
14061 			++xlen;
14062 			tp->t_flags |= TF_SENTFIN;
14063 		}
14064 		if (xlen && (tp->snd_una == tp->snd_max)) {
14065 			/*
14066 			 * Update the time we just added data since none was
14067 			 * outstanding.
14068 			 */
14069 			bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
14070 			bbr->rc_tp->t_acktime = ticks;
14071 		}
14072 		if (sack_rxmit == 0)
14073 			tp->snd_max += xlen;
14074 		tot_len += (len + optlen + ipoptlen);
14075 	}
14076 nomore:
14077 	if (error) {
14078 		/*
14079 		 * Failures do not advance the seq counter above. For the
14080 		 * case of ENOBUFS we will fall out and become ack-clocked.
14081 		 * capping the cwnd at the current flight.
14082 		 * Everything else will just have to retransmit with the timer
14083 		 * (no pacer).
14084 		 */
14085 		SOCKBUF_UNLOCK_ASSERT(sb);
14086 		BBR_STAT_INC(bbr_saw_oerr);
14087 		/* Clear all delay/early tracks */
14088 		bbr->r_ctl.rc_hptsi_agg_delay = 0;
14089 		bbr->r_ctl.rc_agg_early = 0;
14090 		bbr->r_agg_early_set = 0;
14091 		bbr->output_error_seen = 1;
14092 		if (bbr->oerror_cnt < 0xf)
14093 			bbr->oerror_cnt++;
14094 		if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) {
14095 			/* drop the session */
14096 			tcp_set_inp_to_drop(inp, ENETDOWN);
14097 		}
14098 		switch (error) {
14099 		case ENOBUFS:
14100 			/*
14101 			 * Make this guy have to get ack's to send
14102 			 * more but lets make sure we don't
14103 			 * slam him below a T-O (1MSS).
14104 			 */
14105 			if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
14106 				tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14107 								    bbr->r_ctl.rc_lost_bytes)) - maxseg;
14108 				if (tp->snd_cwnd < maxseg)
14109 					tp->snd_cwnd = maxseg;
14110 			}
14111 			slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt;
14112 			BBR_STAT_INC(bbr_saw_enobuf);
14113 			if (bbr->bbr_hdrw_pacing)
14114 				counter_u64_add(bbr_hdwr_pacing_enobuf, 1);
14115 			else
14116 				counter_u64_add(bbr_nohdwr_pacing_enobuf, 1);
14117 			/*
14118 			 * Here even in the enobuf's case we want to do our
14119 			 * state update. The reason being we may have been
14120 			 * called by the input function. If so we have had
14121 			 * things change.
14122 			 */
14123 			error = 0;
14124 			goto enobufs;
14125 		case EMSGSIZE:
14126 			/*
14127 			 * For some reason the interface we used initially
14128 			 * to send segments changed to another or lowered
14129 			 * its MTU. If TSO was active we either got an
14130 			 * interface without TSO capabilits or TSO was
14131 			 * turned off. If we obtained mtu from ip_output()
14132 			 * then update it and try again.
14133 			 */
14134 			/* Turn on tracing (or try to) */
14135 			{
14136 				int old_maxseg;
14137 
14138 				old_maxseg = tp->t_maxseg;
14139 				BBR_STAT_INC(bbr_saw_emsgsiz);
14140 				bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts);
14141 				if (mtu != 0)
14142 					tcp_mss_update(tp, -1, mtu, NULL, NULL);
14143 				if (old_maxseg <= tp->t_maxseg) {
14144 					/* Huh it did not shrink? */
14145 					tp->t_maxseg = old_maxseg - 40;
14146 					bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts);
14147 				}
14148 				/*
14149 				 * Nuke all other things that can interfere
14150 				 * with slot
14151 				 */
14152 				if ((tot_len + len) && (len >= tp->t_maxseg)) {
14153 					slot = bbr_get_pacing_delay(bbr,
14154 					    bbr->r_ctl.rc_bbr_hptsi_gain,
14155 					    (tot_len + len), cts, 0);
14156 					if (slot < bbr_error_base_paceout)
14157 						slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
14158 				} else
14159 					slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
14160 				bbr->rc_output_starts_timer = 1;
14161 				bbr_start_hpts_timer(bbr, tp, cts, 10, slot,
14162 				    tot_len);
14163 				return (error);
14164 			}
14165 		case EPERM:
14166 			tp->t_softerror = error;
14167 			/* Fall through */
14168 		case EHOSTDOWN:
14169 		case EHOSTUNREACH:
14170 		case ENETDOWN:
14171 		case ENETUNREACH:
14172 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
14173 				tp->t_softerror = error;
14174 			}
14175 			/* FALLTHROUGH */
14176 		default:
14177 			slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt;
14178 			bbr->rc_output_starts_timer = 1;
14179 			bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0);
14180 			return (error);
14181 		}
14182 #ifdef STATS
14183 	} else if (((tp->t_flags & TF_GPUTINPROG) == 0) &&
14184 		    len &&
14185 		    (rsm == NULL) &&
14186 	    (bbr->rc_in_persist == 0)) {
14187 		tp->gput_seq = bbr_seq;
14188 		tp->gput_ack = bbr_seq +
14189 		    min(sbavail(&so->so_snd) - sb_offset, sendwin);
14190 		tp->gput_ts = cts;
14191 		tp->t_flags |= TF_GPUTINPROG;
14192 #endif
14193 	}
14194 	KMOD_TCPSTAT_INC(tcps_sndtotal);
14195 	if ((bbr->bbr_hdw_pace_ena) &&
14196 	    (bbr->bbr_attempt_hdwr_pace == 0) &&
14197 	    (bbr->rc_past_init_win) &&
14198 	    (bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
14199 	    (get_filter_value(&bbr->r_ctl.rc_delrate)) &&
14200 	    (inp->inp_route.ro_nh &&
14201 	     inp->inp_route.ro_nh->nh_ifp)) {
14202 		/*
14203 		 * We are past the initial window and
14204 		 * have at least one measurement so we
14205 		 * could use hardware pacing if its available.
14206 		 * We have an interface and we have not attempted
14207 		 * to setup hardware pacing, lets try to now.
14208 		 */
14209 		uint64_t rate_wanted;
14210 		int err = 0;
14211 
14212 		rate_wanted = bbr_get_hardware_rate(bbr);
14213 		bbr->bbr_attempt_hdwr_pace = 1;
14214 		bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp,
14215 						      inp->inp_route.ro_nh->nh_ifp,
14216 						      rate_wanted,
14217 						      (RS_PACING_GEQ|RS_PACING_SUB_OK),
14218 						      &err);
14219 		if (bbr->r_ctl.crte) {
14220 			bbr_type_log_hdwr_pacing(bbr,
14221 						 bbr->r_ctl.crte->ptbl->rs_ifp,
14222 						 rate_wanted,
14223 						 bbr->r_ctl.crte->rate,
14224 						 __LINE__, cts, err);
14225 			BBR_STAT_INC(bbr_hdwr_rl_add_ok);
14226 			counter_u64_add(bbr_flows_nohdwr_pacing, -1);
14227 			counter_u64_add(bbr_flows_whdwr_pacing, 1);
14228 			bbr->bbr_hdrw_pacing = 1;
14229 			/* Now what is our gain status? */
14230 			if (bbr->r_ctl.crte->rate < rate_wanted) {
14231 				/* We have a problem */
14232 				bbr_setup_less_of_rate(bbr, cts,
14233 						       bbr->r_ctl.crte->rate, rate_wanted);
14234 			} else {
14235 				/* We are good */
14236 				bbr->gain_is_limited = 0;
14237 				bbr->skip_gain = 0;
14238 			}
14239 			tcp_bbr_tso_size_check(bbr, cts);
14240 		} else {
14241 			bbr_type_log_hdwr_pacing(bbr,
14242 						 inp->inp_route.ro_nh->nh_ifp,
14243 						 rate_wanted,
14244 						 0,
14245 						 __LINE__, cts, err);
14246 			BBR_STAT_INC(bbr_hdwr_rl_add_fail);
14247 		}
14248 	}
14249 	if (bbr->bbr_hdrw_pacing) {
14250 		/*
14251 		 * Worry about cases where the route
14252 		 * changes or something happened that we
14253 		 * lost our hardware pacing possibly during
14254 		 * the last ip_output call.
14255 		 */
14256 		if (inp->inp_snd_tag == NULL) {
14257 			/* A change during ip output disabled hw pacing? */
14258 			bbr->bbr_hdrw_pacing = 0;
14259 		} else if ((inp->inp_route.ro_nh == NULL) ||
14260 		    (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) {
14261 			/*
14262 			 * We had an interface or route change,
14263 			 * detach from the current hdwr pacing
14264 			 * and setup to re-attempt next go
14265 			 * round.
14266 			 */
14267 			bbr->bbr_hdrw_pacing = 0;
14268 			bbr->bbr_attempt_hdwr_pace = 0;
14269 			tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
14270 			tcp_bbr_tso_size_check(bbr, cts);
14271 		}
14272 	}
14273 	/*
14274 	 * Data sent (as far as we can tell). If this advertises a larger
14275 	 * window than any other segment, then remember the size of the
14276 	 * advertised window. Any pending ACK has now been sent.
14277 	 */
14278 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
14279 		tp->rcv_adv = tp->rcv_nxt + recwin;
14280 
14281 	tp->last_ack_sent = tp->rcv_nxt;
14282 	if ((error == 0) &&
14283 	    (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) &&
14284 	    (doing_tlp == 0) &&
14285 	    (tso == 0) &&
14286 	    (hw_tls == 0) &&
14287 	    (len > 0) &&
14288 	    ((flags & TH_RST) == 0) &&
14289 	    (IN_RECOVERY(tp->t_flags) == 0) &&
14290 	    (bbr->rc_in_persist == 0) &&
14291 	    (tot_len < bbr->r_ctl.rc_pace_max_segs)) {
14292 		/*
14293 		 * For non-tso we need to goto again until we have sent out
14294 		 * enough data to match what we are hptsi out every hptsi
14295 		 * interval.
14296 		 */
14297 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14298 			/* Make sure snd_nxt is drug up */
14299 			tp->snd_nxt = tp->snd_max;
14300 		}
14301 		if (rsm != NULL) {
14302 			rsm = NULL;
14303 			goto skip_again;
14304 		}
14305 		rsm = NULL;
14306 		sack_rxmit = 0;
14307 		tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14308 		goto again;
14309 	}
14310 skip_again:
14311 	if ((error == 0) && (flags & TH_FIN))
14312 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
14313 	if ((error == 0) && (flags & TH_RST))
14314 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
14315 	if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) {
14316 		/*
14317 		 * Calculate/Re-Calculate the hptsi slot in usecs based on
14318 		 * what we have sent so far
14319 		 */
14320 		slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
14321 		if (bbr->rc_no_pacing)
14322 			slot = 0;
14323 	}
14324 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14325 enobufs:
14326 	if (bbr->rc_use_google == 0)
14327 		bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
14328 	bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14329 							bbr->r_ctl.rc_lost_bytes)));
14330 	bbr->rc_output_starts_timer = 1;
14331 	if (bbr->bbr_use_rack_cheat &&
14332 	    (more_to_rxt ||
14333 	     ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) {
14334 		/* Rack cheats and shotguns out all rxt's 1ms apart */
14335 		if (slot > 1000)
14336 			slot = 1000;
14337 	}
14338 	if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) {
14339 		/*
14340 		 * We don't change the tso size until some number of sends
14341 		 * to give the hardware commands time to get down
14342 		 * to the interface.
14343 		 */
14344 		bbr->r_ctl.bbr_hdwr_cnt_noset_snt++;
14345 		if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) {
14346 			bbr->hw_pacing_set = 1;
14347 			tcp_bbr_tso_size_check(bbr, cts);
14348 		}
14349 	}
14350 	bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len);
14351 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14352 		/* Make sure snd_nxt is drug up */
14353 		tp->snd_nxt = tp->snd_max;
14354 	}
14355 	return (error);
14356 
14357 }
14358 
14359 /*
14360  * See bbr_output_wtime() for return values.
14361  */
14362 static int
14363 bbr_output(struct tcpcb *tp)
14364 {
14365 	int32_t ret;
14366 	struct timeval tv;
14367 	struct tcp_bbr *bbr;
14368 
14369 	NET_EPOCH_ASSERT();
14370 
14371 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14372 	INP_WLOCK_ASSERT(tp->t_inpcb);
14373 	(void)tcp_get_usecs(&tv);
14374 	ret = bbr_output_wtime(tp, &tv);
14375 	return (ret);
14376 }
14377 
14378 static void
14379 bbr_mtu_chg(struct tcpcb *tp)
14380 {
14381 	struct tcp_bbr *bbr;
14382 	struct bbr_sendmap *rsm, *frsm = NULL;
14383 	uint32_t maxseg;
14384 
14385 	/*
14386 	 * The MTU has changed. a) Clear the sack filter. b) Mark everything
14387 	 * over the current size as SACK_PASS so a retransmit will occur.
14388 	 */
14389 
14390 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14391 	maxseg = tp->t_maxseg - bbr->rc_last_options;
14392 	sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
14393 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
14394 		/* Don't mess with ones acked (by sack?) */
14395 		if (rsm->r_flags & BBR_ACKED)
14396 			continue;
14397 		if ((rsm->r_end - rsm->r_start) > maxseg) {
14398 			/*
14399 			 * We mark sack-passed on all the previous large
14400 			 * sends we did. This will force them to retransmit.
14401 			 */
14402 			rsm->r_flags |= BBR_SACK_PASSED;
14403 			if (((rsm->r_flags & BBR_MARKED_LOST) == 0) &&
14404 			    bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) {
14405 				bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
14406 				bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
14407 				rsm->r_flags |= BBR_MARKED_LOST;
14408 			}
14409 			if (frsm == NULL)
14410 				frsm = rsm;
14411 		}
14412 	}
14413 	if (frsm) {
14414 		bbr->r_ctl.rc_resend = frsm;
14415 	}
14416 }
14417 
14418 /*
14419  * bbr_ctloutput() must drop the inpcb lock before performing copyin on
14420  * socket option arguments.  When it re-acquires the lock after the copy, it
14421  * has to revalidate that the connection is still valid for the socket
14422  * option.
14423  */
14424 static int
14425 bbr_set_sockopt(struct socket *so, struct sockopt *sopt,
14426 		struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14427 {
14428 	struct epoch_tracker et;
14429 	int32_t error = 0, optval;
14430 
14431 	switch (sopt->sopt_name) {
14432 	case TCP_RACK_PACE_MAX_SEG:
14433 	case TCP_RACK_MIN_TO:
14434 	case TCP_RACK_REORD_THRESH:
14435 	case TCP_RACK_REORD_FADE:
14436 	case TCP_RACK_TLP_THRESH:
14437 	case TCP_RACK_PKT_DELAY:
14438 	case TCP_BBR_ALGORITHM:
14439 	case TCP_BBR_TSLIMITS:
14440 	case TCP_BBR_IWINTSO:
14441 	case TCP_BBR_RECFORCE:
14442 	case TCP_BBR_STARTUP_PG:
14443 	case TCP_BBR_DRAIN_PG:
14444 	case TCP_BBR_RWND_IS_APP:
14445 	case TCP_BBR_PROBE_RTT_INT:
14446 	case TCP_BBR_PROBE_RTT_GAIN:
14447 	case TCP_BBR_PROBE_RTT_LEN:
14448 	case TCP_BBR_STARTUP_LOSS_EXIT:
14449 	case TCP_BBR_USEDEL_RATE:
14450 	case TCP_BBR_MIN_RTO:
14451 	case TCP_BBR_MAX_RTO:
14452 	case TCP_BBR_PACE_PER_SEC:
14453 	case TCP_DELACK:
14454 	case TCP_BBR_PACE_DEL_TAR:
14455 	case TCP_BBR_SEND_IWND_IN_TSO:
14456 	case TCP_BBR_EXTRA_STATE:
14457 	case TCP_BBR_UTTER_MAX_TSO:
14458 	case TCP_BBR_MIN_TOPACEOUT:
14459 	case TCP_BBR_FLOOR_MIN_TSO:
14460 	case TCP_BBR_TSTMP_RAISES:
14461 	case TCP_BBR_POLICER_DETECT:
14462 	case TCP_BBR_USE_RACK_CHEAT:
14463 	case TCP_DATA_AFTER_CLOSE:
14464 	case TCP_BBR_HDWR_PACE:
14465 	case TCP_BBR_PACE_SEG_MAX:
14466 	case TCP_BBR_PACE_SEG_MIN:
14467 	case TCP_BBR_PACE_CROSS:
14468 	case TCP_BBR_PACE_OH:
14469 #ifdef NETFLIX_PEAKRATE
14470 	case TCP_MAXPEAKRATE:
14471 #endif
14472 	case TCP_BBR_TMR_PACE_OH:
14473 	case TCP_BBR_RACK_RTT_USE:
14474 	case TCP_BBR_RETRAN_WTSO:
14475 		break;
14476 	default:
14477 		return (tcp_default_ctloutput(so, sopt, inp, tp));
14478 		break;
14479 	}
14480 	INP_WUNLOCK(inp);
14481 	error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14482 	if (error)
14483 		return (error);
14484 	INP_WLOCK(inp);
14485 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
14486 		INP_WUNLOCK(inp);
14487 		return (ECONNRESET);
14488 	}
14489 	tp = intotcpcb(inp);
14490 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14491 	switch (sopt->sopt_name) {
14492 	case TCP_BBR_PACE_PER_SEC:
14493 		BBR_OPTS_INC(tcp_bbr_pace_per_sec);
14494 		bbr->r_ctl.bbr_hptsi_per_second = optval;
14495 		break;
14496 	case TCP_BBR_PACE_DEL_TAR:
14497 		BBR_OPTS_INC(tcp_bbr_pace_del_tar);
14498 		bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval;
14499 		break;
14500 	case TCP_BBR_PACE_SEG_MAX:
14501 		BBR_OPTS_INC(tcp_bbr_pace_seg_max);
14502 		bbr->r_ctl.bbr_hptsi_segments_max = optval;
14503 		break;
14504 	case TCP_BBR_PACE_SEG_MIN:
14505 		BBR_OPTS_INC(tcp_bbr_pace_seg_min);
14506 		bbr->r_ctl.bbr_hptsi_bytes_min = optval;
14507 		break;
14508 	case TCP_BBR_PACE_CROSS:
14509 		BBR_OPTS_INC(tcp_bbr_pace_cross);
14510 		bbr->r_ctl.bbr_cross_over = optval;
14511 		break;
14512 	case TCP_BBR_ALGORITHM:
14513 		BBR_OPTS_INC(tcp_bbr_algorithm);
14514 		if (optval && (bbr->rc_use_google == 0)) {
14515 			/* Turn on the google mode */
14516 			bbr_google_mode_on(bbr);
14517 			if ((optval > 3) && (optval < 500)) {
14518 				/*
14519 				 * Must be at least greater than .3%
14520 				 * and must be less than 50.0%.
14521 				 */
14522 				bbr->r_ctl.bbr_google_discount = optval;
14523 			}
14524 		} else if ((optval == 0) && (bbr->rc_use_google == 1)) {
14525 			/* Turn off the google mode */
14526 			bbr_google_mode_off(bbr);
14527 		}
14528 		break;
14529 	case TCP_BBR_TSLIMITS:
14530 		BBR_OPTS_INC(tcp_bbr_tslimits);
14531 		if (optval == 1)
14532 			bbr->rc_use_ts_limit = 1;
14533 		else if (optval == 0)
14534 			bbr->rc_use_ts_limit = 0;
14535 		else
14536 			error = EINVAL;
14537 		break;
14538 
14539 	case TCP_BBR_IWINTSO:
14540 		BBR_OPTS_INC(tcp_bbr_iwintso);
14541 		if ((optval >= 0) && (optval < 128)) {
14542 			uint32_t twin;
14543 
14544 			bbr->rc_init_win = optval;
14545 			twin = bbr_initial_cwnd(bbr, tp);
14546 			if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd))
14547 				tp->snd_cwnd = twin;
14548 			else
14549 				error = EBUSY;
14550 		} else
14551 			error = EINVAL;
14552 		break;
14553 	case TCP_BBR_STARTUP_PG:
14554 		BBR_OPTS_INC(tcp_bbr_startup_pg);
14555 		if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) {
14556 			bbr->r_ctl.rc_startup_pg = optval;
14557 			if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
14558 				bbr->r_ctl.rc_bbr_hptsi_gain = optval;
14559 			}
14560 		} else
14561 			error = EINVAL;
14562 		break;
14563 	case TCP_BBR_DRAIN_PG:
14564 		BBR_OPTS_INC(tcp_bbr_drain_pg);
14565 		if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE))
14566 			bbr->r_ctl.rc_drain_pg = optval;
14567 		else
14568 			error = EINVAL;
14569 		break;
14570 	case TCP_BBR_PROBE_RTT_LEN:
14571 		BBR_OPTS_INC(tcp_bbr_probertt_len);
14572 		if (optval <= 1)
14573 			reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND));
14574 		else
14575 			error = EINVAL;
14576 		break;
14577 	case TCP_BBR_PROBE_RTT_GAIN:
14578 		BBR_OPTS_INC(tcp_bbr_probertt_gain);
14579 		if (optval <= BBR_UNIT)
14580 			bbr->r_ctl.bbr_rttprobe_gain_val = optval;
14581 		else
14582 			error = EINVAL;
14583 		break;
14584 	case TCP_BBR_PROBE_RTT_INT:
14585 		BBR_OPTS_INC(tcp_bbr_probe_rtt_int);
14586 		if (optval > 1000)
14587 			bbr->r_ctl.rc_probertt_int = optval;
14588 		else
14589 			error = EINVAL;
14590 		break;
14591 	case TCP_BBR_MIN_TOPACEOUT:
14592 		BBR_OPTS_INC(tcp_bbr_topaceout);
14593 		if (optval == 0) {
14594 			bbr->no_pacing_until = 0;
14595 			bbr->rc_no_pacing = 0;
14596 		} else if (optval <= 0x00ff) {
14597 			bbr->no_pacing_until = optval;
14598 			if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) &&
14599 			    (bbr->rc_bbr_state == BBR_STATE_STARTUP)){
14600 				/* Turn on no pacing */
14601 				bbr->rc_no_pacing = 1;
14602 			}
14603 		} else
14604 			error = EINVAL;
14605 		break;
14606 	case TCP_BBR_STARTUP_LOSS_EXIT:
14607 		BBR_OPTS_INC(tcp_bbr_startup_loss_exit);
14608 		bbr->rc_loss_exit = optval;
14609 		break;
14610 	case TCP_BBR_USEDEL_RATE:
14611 		error = EINVAL;
14612 		break;
14613 	case TCP_BBR_MIN_RTO:
14614 		BBR_OPTS_INC(tcp_bbr_min_rto);
14615 		bbr->r_ctl.rc_min_rto_ms = optval;
14616 		break;
14617 	case TCP_BBR_MAX_RTO:
14618 		BBR_OPTS_INC(tcp_bbr_max_rto);
14619 		bbr->rc_max_rto_sec = optval;
14620 		break;
14621 	case TCP_RACK_MIN_TO:
14622 		/* Minimum time between rack t-o's in ms */
14623 		BBR_OPTS_INC(tcp_rack_min_to);
14624 		bbr->r_ctl.rc_min_to = optval;
14625 		break;
14626 	case TCP_RACK_REORD_THRESH:
14627 		/* RACK reorder threshold (shift amount) */
14628 		BBR_OPTS_INC(tcp_rack_reord_thresh);
14629 		if ((optval > 0) && (optval < 31))
14630 			bbr->r_ctl.rc_reorder_shift = optval;
14631 		else
14632 			error = EINVAL;
14633 		break;
14634 	case TCP_RACK_REORD_FADE:
14635 		/* Does reordering fade after ms time */
14636 		BBR_OPTS_INC(tcp_rack_reord_fade);
14637 		bbr->r_ctl.rc_reorder_fade = optval;
14638 		break;
14639 	case TCP_RACK_TLP_THRESH:
14640 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14641 		BBR_OPTS_INC(tcp_rack_tlp_thresh);
14642 		if (optval)
14643 			bbr->rc_tlp_threshold = optval;
14644 		else
14645 			error = EINVAL;
14646 		break;
14647 	case TCP_BBR_USE_RACK_CHEAT:
14648 		BBR_OPTS_INC(tcp_use_rackcheat);
14649 		if (bbr->rc_use_google) {
14650 			error = EINVAL;
14651 			break;
14652 		}
14653 		BBR_OPTS_INC(tcp_rack_cheat);
14654 		if (optval)
14655 			bbr->bbr_use_rack_cheat = 1;
14656 		else
14657 			bbr->bbr_use_rack_cheat = 0;
14658 		break;
14659 	case TCP_BBR_FLOOR_MIN_TSO:
14660 		BBR_OPTS_INC(tcp_utter_max_tso);
14661 		if ((optval >= 0) && (optval < 40))
14662 			bbr->r_ctl.bbr_hptsi_segments_floor = optval;
14663 		else
14664 			error = EINVAL;
14665 		break;
14666 	case TCP_BBR_UTTER_MAX_TSO:
14667 		BBR_OPTS_INC(tcp_utter_max_tso);
14668 		if ((optval >= 0) && (optval < 0xffff))
14669 			bbr->r_ctl.bbr_utter_max = optval;
14670 		else
14671 			error = EINVAL;
14672 		break;
14673 
14674 	case TCP_BBR_EXTRA_STATE:
14675 		BBR_OPTS_INC(tcp_extra_state);
14676 		if (optval)
14677 			bbr->rc_use_idle_restart = 1;
14678 		else
14679 			bbr->rc_use_idle_restart = 0;
14680 		break;
14681 	case TCP_BBR_SEND_IWND_IN_TSO:
14682 		BBR_OPTS_INC(tcp_iwnd_tso);
14683 		if (optval) {
14684 			bbr->bbr_init_win_cheat = 1;
14685 			if (bbr->rc_past_init_win == 0) {
14686 				uint32_t cts;
14687 				cts = tcp_get_usecs(&bbr->rc_tv);
14688 				tcp_bbr_tso_size_check(bbr, cts);
14689 			}
14690 		} else
14691 			bbr->bbr_init_win_cheat = 0;
14692 		break;
14693 	case TCP_BBR_HDWR_PACE:
14694 		BBR_OPTS_INC(tcp_hdwr_pacing);
14695 		if (optval){
14696 			bbr->bbr_hdw_pace_ena = 1;
14697 			bbr->bbr_attempt_hdwr_pace = 0;
14698 		} else {
14699 			bbr->bbr_hdw_pace_ena = 0;
14700 #ifdef RATELIMIT
14701 			if (bbr->bbr_hdrw_pacing) {
14702 				bbr->bbr_hdrw_pacing = 0;
14703 				in_pcbdetach_txrtlmt(bbr->rc_inp);
14704 			}
14705 #endif
14706 		}
14707 		break;
14708 
14709 	case TCP_DELACK:
14710 		BBR_OPTS_INC(tcp_delack);
14711 		if (optval < 100) {
14712 			if (optval == 0) /* off */
14713 				tp->t_delayed_ack = 0;
14714 			else if (optval == 1) /* on which is 2 */
14715 				tp->t_delayed_ack = 2;
14716 			else /* higher than 2 and less than 100 */
14717 				tp->t_delayed_ack = optval;
14718 			if (tp->t_flags & TF_DELACK) {
14719 				tp->t_flags &= ~TF_DELACK;
14720 				tp->t_flags |= TF_ACKNOW;
14721 				NET_EPOCH_ENTER(et);
14722 				bbr_output(tp);
14723 				NET_EPOCH_EXIT(et);
14724 			}
14725 		} else
14726 			error = EINVAL;
14727 		break;
14728 	case TCP_RACK_PKT_DELAY:
14729 		/* RACK added ms i.e. rack-rtt + reord + N */
14730 		BBR_OPTS_INC(tcp_rack_pkt_delay);
14731 		bbr->r_ctl.rc_pkt_delay = optval;
14732 		break;
14733 #ifdef NETFLIX_PEAKRATE
14734 	case TCP_MAXPEAKRATE:
14735 		BBR_OPTS_INC(tcp_maxpeak);
14736 		error = tcp_set_maxpeakrate(tp, optval);
14737 		if (!error)
14738 			tp->t_peakrate_thr = tp->t_maxpeakrate;
14739 		break;
14740 #endif
14741 	case TCP_BBR_RETRAN_WTSO:
14742 		BBR_OPTS_INC(tcp_retran_wtso);
14743 		if (optval)
14744 			bbr->rc_resends_use_tso = 1;
14745 		else
14746 			bbr->rc_resends_use_tso = 0;
14747 		break;
14748 	case TCP_DATA_AFTER_CLOSE:
14749 		BBR_OPTS_INC(tcp_data_ac);
14750 		if (optval)
14751 			bbr->rc_allow_data_af_clo = 1;
14752 		else
14753 			bbr->rc_allow_data_af_clo = 0;
14754 		break;
14755 	case TCP_BBR_POLICER_DETECT:
14756 		BBR_OPTS_INC(tcp_policer_det);
14757 		if (bbr->rc_use_google == 0)
14758 			error = EINVAL;
14759 		else if (optval)
14760 			bbr->r_use_policer = 1;
14761 		else
14762 			bbr->r_use_policer = 0;
14763 		break;
14764 
14765 	case TCP_BBR_TSTMP_RAISES:
14766 		BBR_OPTS_INC(tcp_ts_raises);
14767 		if (optval)
14768 			bbr->ts_can_raise = 1;
14769 		else
14770 			bbr->ts_can_raise = 0;
14771 		break;
14772 	case TCP_BBR_TMR_PACE_OH:
14773 		BBR_OPTS_INC(tcp_pacing_oh_tmr);
14774 		if (bbr->rc_use_google) {
14775 			error = EINVAL;
14776 		} else {
14777 			if (optval)
14778 				bbr->r_ctl.rc_incr_tmrs = 1;
14779 			else
14780 				bbr->r_ctl.rc_incr_tmrs = 0;
14781 		}
14782 		break;
14783 	case TCP_BBR_PACE_OH:
14784 		BBR_OPTS_INC(tcp_pacing_oh);
14785 		if (bbr->rc_use_google) {
14786 			error = EINVAL;
14787 		} else {
14788 			if (optval > (BBR_INCL_TCP_OH|
14789 				      BBR_INCL_IP_OH|
14790 				      BBR_INCL_ENET_OH)) {
14791 				error = EINVAL;
14792 				break;
14793 			}
14794 			if (optval & BBR_INCL_TCP_OH)
14795 				bbr->r_ctl.rc_inc_tcp_oh = 1;
14796 			else
14797 				bbr->r_ctl.rc_inc_tcp_oh = 0;
14798 			if (optval & BBR_INCL_IP_OH)
14799 				bbr->r_ctl.rc_inc_ip_oh = 1;
14800 			else
14801 				bbr->r_ctl.rc_inc_ip_oh = 0;
14802 			if (optval & BBR_INCL_ENET_OH)
14803 				bbr->r_ctl.rc_inc_enet_oh = 1;
14804 			else
14805 				bbr->r_ctl.rc_inc_enet_oh = 0;
14806 		}
14807 		break;
14808 	default:
14809 		return (tcp_default_ctloutput(so, sopt, inp, tp));
14810 		break;
14811 	}
14812 #ifdef NETFLIX_STATS
14813 	tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14814 #endif
14815 	INP_WUNLOCK(inp);
14816 	return (error);
14817 }
14818 
14819 /*
14820  * return 0 on success, error-num on failure
14821  */
14822 static int
14823 bbr_get_sockopt(struct socket *so, struct sockopt *sopt,
14824     struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14825 {
14826 	int32_t error, optval;
14827 
14828 	/*
14829 	 * Because all our options are either boolean or an int, we can just
14830 	 * pull everything into optval and then unlock and copy. If we ever
14831 	 * add a option that is not a int, then this will have quite an
14832 	 * impact to this routine.
14833 	 */
14834 	switch (sopt->sopt_name) {
14835 	case TCP_BBR_PACE_PER_SEC:
14836 		optval = bbr->r_ctl.bbr_hptsi_per_second;
14837 		break;
14838 	case TCP_BBR_PACE_DEL_TAR:
14839 		optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar;
14840 		break;
14841 	case TCP_BBR_PACE_SEG_MAX:
14842 		optval = bbr->r_ctl.bbr_hptsi_segments_max;
14843 		break;
14844 	case TCP_BBR_MIN_TOPACEOUT:
14845 		optval = bbr->no_pacing_until;
14846 		break;
14847 	case TCP_BBR_PACE_SEG_MIN:
14848 		optval = bbr->r_ctl.bbr_hptsi_bytes_min;
14849 		break;
14850 	case TCP_BBR_PACE_CROSS:
14851 		optval = bbr->r_ctl.bbr_cross_over;
14852 		break;
14853 	case TCP_BBR_ALGORITHM:
14854 		optval = bbr->rc_use_google;
14855 		break;
14856 	case TCP_BBR_TSLIMITS:
14857 		optval = bbr->rc_use_ts_limit;
14858 		break;
14859 	case TCP_BBR_IWINTSO:
14860 		optval = bbr->rc_init_win;
14861 		break;
14862 	case TCP_BBR_STARTUP_PG:
14863 		optval = bbr->r_ctl.rc_startup_pg;
14864 		break;
14865 	case TCP_BBR_DRAIN_PG:
14866 		optval = bbr->r_ctl.rc_drain_pg;
14867 		break;
14868 	case TCP_BBR_PROBE_RTT_INT:
14869 		optval = bbr->r_ctl.rc_probertt_int;
14870 		break;
14871 	case TCP_BBR_PROBE_RTT_LEN:
14872 		optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND);
14873 		break;
14874 	case TCP_BBR_PROBE_RTT_GAIN:
14875 		optval = bbr->r_ctl.bbr_rttprobe_gain_val;
14876 		break;
14877 	case TCP_BBR_STARTUP_LOSS_EXIT:
14878 		optval = bbr->rc_loss_exit;
14879 		break;
14880 	case TCP_BBR_USEDEL_RATE:
14881 		error = EINVAL;
14882 		break;
14883 	case TCP_BBR_MIN_RTO:
14884 		optval = bbr->r_ctl.rc_min_rto_ms;
14885 		break;
14886 	case TCP_BBR_MAX_RTO:
14887 		optval = bbr->rc_max_rto_sec;
14888 		break;
14889 	case TCP_RACK_PACE_MAX_SEG:
14890 		/* Max segments in a pace */
14891 		optval = bbr->r_ctl.rc_pace_max_segs;
14892 		break;
14893 	case TCP_RACK_MIN_TO:
14894 		/* Minimum time between rack t-o's in ms */
14895 		optval = bbr->r_ctl.rc_min_to;
14896 		break;
14897 	case TCP_RACK_REORD_THRESH:
14898 		/* RACK reorder threshold (shift amount) */
14899 		optval = bbr->r_ctl.rc_reorder_shift;
14900 		break;
14901 	case TCP_RACK_REORD_FADE:
14902 		/* Does reordering fade after ms time */
14903 		optval = bbr->r_ctl.rc_reorder_fade;
14904 		break;
14905 	case TCP_BBR_USE_RACK_CHEAT:
14906 		/* Do we use the rack cheat for rxt */
14907 		optval = bbr->bbr_use_rack_cheat;
14908 		break;
14909 	case TCP_BBR_FLOOR_MIN_TSO:
14910 		optval = bbr->r_ctl.bbr_hptsi_segments_floor;
14911 		break;
14912 	case TCP_BBR_UTTER_MAX_TSO:
14913 		optval = bbr->r_ctl.bbr_utter_max;
14914 		break;
14915 	case TCP_BBR_SEND_IWND_IN_TSO:
14916 		/* Do we send TSO size segments initially */
14917 		optval = bbr->bbr_init_win_cheat;
14918 		break;
14919 	case TCP_BBR_EXTRA_STATE:
14920 		optval = bbr->rc_use_idle_restart;
14921 		break;
14922 	case TCP_RACK_TLP_THRESH:
14923 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14924 		optval = bbr->rc_tlp_threshold;
14925 		break;
14926 	case TCP_RACK_PKT_DELAY:
14927 		/* RACK added ms i.e. rack-rtt + reord + N */
14928 		optval = bbr->r_ctl.rc_pkt_delay;
14929 		break;
14930 	case TCP_BBR_RETRAN_WTSO:
14931 		optval = bbr->rc_resends_use_tso;
14932 		break;
14933 	case TCP_DATA_AFTER_CLOSE:
14934 		optval = bbr->rc_allow_data_af_clo;
14935 		break;
14936 	case TCP_DELACK:
14937 		optval = tp->t_delayed_ack;
14938 		break;
14939 	case TCP_BBR_HDWR_PACE:
14940 		optval = bbr->bbr_hdw_pace_ena;
14941 		break;
14942 	case TCP_BBR_POLICER_DETECT:
14943 		optval = bbr->r_use_policer;
14944 		break;
14945 	case TCP_BBR_TSTMP_RAISES:
14946 		optval = bbr->ts_can_raise;
14947 		break;
14948 	case TCP_BBR_TMR_PACE_OH:
14949 		optval = bbr->r_ctl.rc_incr_tmrs;
14950 		break;
14951 	case TCP_BBR_PACE_OH:
14952 		optval = 0;
14953 		if (bbr->r_ctl.rc_inc_tcp_oh)
14954 			optval |= BBR_INCL_TCP_OH;
14955 		if (bbr->r_ctl.rc_inc_ip_oh)
14956 			optval |= BBR_INCL_IP_OH;
14957 		if (bbr->r_ctl.rc_inc_enet_oh)
14958 			optval |= BBR_INCL_ENET_OH;
14959 		break;
14960 	default:
14961 		return (tcp_default_ctloutput(so, sopt, inp, tp));
14962 		break;
14963 	}
14964 	INP_WUNLOCK(inp);
14965 	error = sooptcopyout(sopt, &optval, sizeof optval);
14966 	return (error);
14967 }
14968 
14969 /*
14970  * return 0 on success, error-num on failure
14971  */
14972 static int
14973 bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
14974 {
14975 	int32_t error = EINVAL;
14976 	struct tcp_bbr *bbr;
14977 
14978 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14979 	if (bbr == NULL) {
14980 		/* Huh? */
14981 		goto out;
14982 	}
14983 	if (sopt->sopt_dir == SOPT_SET) {
14984 		return (bbr_set_sockopt(so, sopt, inp, tp, bbr));
14985 	} else if (sopt->sopt_dir == SOPT_GET) {
14986 		return (bbr_get_sockopt(so, sopt, inp, tp, bbr));
14987 	}
14988 out:
14989 	INP_WUNLOCK(inp);
14990 	return (error);
14991 }
14992 
14993 static int
14994 bbr_pru_options(struct tcpcb *tp, int flags)
14995 {
14996 	if (flags & PRUS_OOB)
14997 		return (EOPNOTSUPP);
14998 	return (0);
14999 }
15000 
15001 struct tcp_function_block __tcp_bbr = {
15002 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
15003 	.tfb_tcp_output = bbr_output,
15004 	.tfb_do_queued_segments = ctf_do_queued_segments,
15005 	.tfb_do_segment_nounlock = bbr_do_segment_nounlock,
15006 	.tfb_tcp_do_segment = bbr_do_segment,
15007 	.tfb_tcp_ctloutput = bbr_ctloutput,
15008 	.tfb_tcp_fb_init = bbr_init,
15009 	.tfb_tcp_fb_fini = bbr_fini,
15010 	.tfb_tcp_timer_stop_all = bbr_stopall,
15011 	.tfb_tcp_timer_activate = bbr_timer_activate,
15012 	.tfb_tcp_timer_active = bbr_timer_active,
15013 	.tfb_tcp_timer_stop = bbr_timer_stop,
15014 	.tfb_tcp_rexmit_tmr = bbr_remxt_tmr,
15015 	.tfb_tcp_handoff_ok = bbr_handoff_ok,
15016 	.tfb_tcp_mtu_chg = bbr_mtu_chg,
15017 	.tfb_pru_options = bbr_pru_options,
15018 };
15019 
15020 static const char *bbr_stack_names[] = {
15021 	__XSTRING(STACKNAME),
15022 #ifdef STACKALIAS
15023 	__XSTRING(STACKALIAS),
15024 #endif
15025 };
15026 
15027 static bool bbr_mod_inited = false;
15028 
15029 static int
15030 tcp_addbbr(module_t mod, int32_t type, void *data)
15031 {
15032 	int32_t err = 0;
15033 	int num_stacks;
15034 
15035 	switch (type) {
15036 	case MOD_LOAD:
15037 		printf("Attempting to load " __XSTRING(MODNAME) "\n");
15038 		bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
15039 		    sizeof(struct bbr_sendmap),
15040 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
15041 		bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
15042 		    sizeof(struct tcp_bbr),
15043 		    NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
15044 		sysctl_ctx_init(&bbr_sysctl_ctx);
15045 		bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
15046 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
15047 		    OID_AUTO,
15048 #ifdef STACKALIAS
15049 		    __XSTRING(STACKALIAS),
15050 #else
15051 		    __XSTRING(STACKNAME),
15052 #endif
15053 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
15054 		    "");
15055 		if (bbr_sysctl_root == NULL) {
15056 			printf("Failed to add sysctl node\n");
15057 			err = EFAULT;
15058 			goto free_uma;
15059 		}
15060 		bbr_init_sysctls();
15061 		num_stacks = nitems(bbr_stack_names);
15062 		err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK,
15063 		    bbr_stack_names, &num_stacks);
15064 		if (err) {
15065 			printf("Failed to register %s stack name for "
15066 			    "%s module\n", bbr_stack_names[num_stacks],
15067 			    __XSTRING(MODNAME));
15068 			sysctl_ctx_free(&bbr_sysctl_ctx);
15069 	free_uma:
15070 			uma_zdestroy(bbr_zone);
15071 			uma_zdestroy(bbr_pcb_zone);
15072 			bbr_counter_destroy();
15073 			printf("Failed to register " __XSTRING(MODNAME)
15074 			    " module err:%d\n", err);
15075 			return (err);
15076 		}
15077 		tcp_lro_reg_mbufq();
15078 		bbr_mod_inited = true;
15079 		printf(__XSTRING(MODNAME) " is now available\n");
15080 		break;
15081 	case MOD_QUIESCE:
15082 		err = deregister_tcp_functions(&__tcp_bbr, true, false);
15083 		break;
15084 	case MOD_UNLOAD:
15085 		err = deregister_tcp_functions(&__tcp_bbr, false, true);
15086 		if (err == EBUSY)
15087 			break;
15088 		if (bbr_mod_inited) {
15089 			uma_zdestroy(bbr_zone);
15090 			uma_zdestroy(bbr_pcb_zone);
15091 			sysctl_ctx_free(&bbr_sysctl_ctx);
15092 			bbr_counter_destroy();
15093 			printf(__XSTRING(MODNAME)
15094 			    " is now no longer available\n");
15095 			bbr_mod_inited = false;
15096 		}
15097 		tcp_lro_dereg_mbufq();
15098 		err = 0;
15099 		break;
15100 	default:
15101 		return (EOPNOTSUPP);
15102 	}
15103 	return (err);
15104 }
15105 
15106 static moduledata_t tcp_bbr = {
15107 	.name = __XSTRING(MODNAME),
15108 	    .evhand = tcp_addbbr,
15109 	    .priv = 0
15110 };
15111 
15112 MODULE_VERSION(MODNAME, 1);
15113 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
15114 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
15115