xref: /freebsd/sys/netinet/tcp_stacks/bbr.c (revision 91ef6f14f234a12eb6cc961be01a483e9a0a5955) 
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_ratelimit.h"
40 #include <sys/param.h>
41 #include <sys/arb.h>
42 #include <sys/module.h>
43 #include <sys/kernel.h>
44 #include <sys/libkern.h>
45 #ifdef TCP_HHOOK
46 #include <sys/hhook.h>
47 #endif
48 #include <sys/malloc.h>
49 #include <sys/mbuf.h>
50 #include <sys/proc.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/sysctl.h>
54 #include <sys/systm.h>
55 #ifdef STATS
56 #include <sys/qmath.h>
57 #include <sys/tree.h>
58 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
59 #endif
60 #include <sys/refcount.h>
61 #include <sys/queue.h>
62 #include <sys/eventhandler.h>
63 #include <sys/smp.h>
64 #include <sys/kthread.h>
65 #include <sys/lock.h>
66 #include <sys/mutex.h>
67 #include <sys/tim_filter.h>
68 #include <sys/time.h>
69 #include <sys/protosw.h>
70 #include <vm/uma.h>
71 #include <sys/kern_prefetch.h>
72 
73 #include <net/route.h>
74 #include <net/route/nhop.h>
75 #include <net/vnet.h>
76 
77 #define TCPSTATES		/* for logging */
78 
79 #include <netinet/in.h>
80 #include <netinet/in_kdtrace.h>
81 #include <netinet/in_pcb.h>
82 #include <netinet/ip.h>
83 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
84 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
85 #include <netinet/ip_var.h>
86 #include <netinet/ip6.h>
87 #include <netinet6/in6_pcb.h>
88 #include <netinet6/ip6_var.h>
89 #define	TCPOUTFLAGS
90 #include <netinet/tcp.h>
91 #include <netinet/tcp_fsm.h>
92 #include <netinet/tcp_seq.h>
93 #include <netinet/tcp_timer.h>
94 #include <netinet/tcp_var.h>
95 #include <netinet/tcpip.h>
96 #include <netinet/tcp_hpts.h>
97 #include <netinet/cc/cc.h>
98 #include <netinet/tcp_log_buf.h>
99 #include <netinet/tcp_ratelimit.h>
100 #include <netinet/tcp_lro.h>
101 #ifdef TCP_OFFLOAD
102 #include <netinet/tcp_offload.h>
103 #endif
104 #ifdef INET6
105 #include <netinet6/tcp6_var.h>
106 #endif
107 #include <netinet/tcp_fastopen.h>
108 
109 #include <netipsec/ipsec_support.h>
110 #include <net/if.h>
111 #include <net/if_var.h>
112 #include <net/ethernet.h>
113 
114 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
115 #include <netipsec/ipsec.h>
116 #include <netipsec/ipsec6.h>
117 #endif				/* IPSEC */
118 
119 #include <netinet/udp.h>
120 #include <netinet/udp_var.h>
121 #include <machine/in_cksum.h>
122 
123 #ifdef MAC
124 #include <security/mac/mac_framework.h>
125 #endif
126 
127 #include "sack_filter.h"
128 #include "tcp_bbr.h"
129 #include "rack_bbr_common.h"
130 uma_zone_t bbr_zone;
131 uma_zone_t bbr_pcb_zone;
132 
133 struct sysctl_ctx_list bbr_sysctl_ctx;
134 struct sysctl_oid *bbr_sysctl_root;
135 
136 #define	TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \
137 	(tv) = (value); \
138 	if ((u_long)(tv) < (u_long)(tvmin)) \
139 		(tv) = (tvmin); \
140 	if ((u_long)(tv) > (u_long)(tvmax)) \
141 		(tv) = (tvmax); \
142 } while(0)
143 
144 /*#define BBR_INVARIANT 1*/
145 
146 /*
147  * initial window
148  */
149 static uint32_t bbr_def_init_win = 10;
150 static int32_t bbr_persist_min = 250000;	/* 250ms */
151 static int32_t bbr_persist_max = 1000000;	/* 1 Second */
152 static int32_t bbr_cwnd_may_shrink = 0;
153 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP;
154 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT;
155 static int32_t bbr_hardware_pacing_limit = 8000;
156 static int32_t bbr_quanta = 3;	/* How much extra quanta do we get? */
157 static int32_t bbr_no_retran = 0;
158 
159 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */
160 static int32_t bbr_max_net_error_cnt = 10;
161 /* Should the following be dynamic too -- loss wise */
162 static int32_t bbr_rtt_gain_thresh = 0;
163 /* Measurement controls */
164 static int32_t bbr_use_google_algo = 1;
165 static int32_t bbr_ts_limiting = 1;
166 static int32_t bbr_ts_can_raise = 0;
167 static int32_t bbr_do_red = 600;
168 static int32_t bbr_red_scale = 20000;
169 static int32_t bbr_red_mul = 1;
170 static int32_t bbr_red_div = 2;
171 static int32_t bbr_red_growth_restrict = 1;
172 static int32_t  bbr_target_is_bbunit = 0;
173 static int32_t bbr_drop_limit = 0;
174 /*
175  * How much gain do we need to see to
176  * stay in startup?
177  */
178 static int32_t bbr_marks_rxt_sack_passed = 0;
179 static int32_t bbr_start_exit = 25;
180 static int32_t bbr_low_start_exit = 25;	/* When we are in reduced gain */
181 static int32_t bbr_startup_loss_thresh = 2000;	/* 20.00% loss */
182 static int32_t bbr_hptsi_max_mul = 1;	/* These two mul/div assure a min pacing */
183 static int32_t bbr_hptsi_max_div = 2;	/* time, 0 means turned off. We need this
184 					 * if we go back ever to where the pacer
185 					 * has priority over timers.
186 					 */
187 static int32_t bbr_policer_call_from_rack_to = 0;
188 static int32_t bbr_policer_detection_enabled = 1;
189 static int32_t bbr_min_measurements_req = 1;	/* We need at least 2
190 						 * measurements before we are
191 						 * "good" note that 2 == 1.
192 						 * This is because we use a >
193 						 * comparison. This means if
194 						 * min_measure was 0, it takes
195 						 * num-measures > min(0) and
196 						 * you get 1 measurement and
197 						 * you are good. Set to 1, you
198 						 * have to have two
199 						 * measurements (this is done
200 						 * to prevent it from being ok
201 						 * to have no measurements). */
202 static int32_t bbr_no_pacing_until = 4;
203 
204 static int32_t bbr_min_usec_delta = 20000;	/* 20,000 usecs */
205 static int32_t bbr_min_peer_delta = 20;		/* 20 units */
206 static int32_t bbr_delta_percent = 150;		/* 15.0 % */
207 
208 static int32_t bbr_target_cwnd_mult_limit = 8;
209 /*
210  * bbr_cwnd_min_val is the number of
211  * segments we hold to in the RTT probe
212  * state typically 4.
213  */
214 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS;
215 
216 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS;
217 
218 static int32_t bbr_gain_to_target = 1;
219 static int32_t bbr_gain_gets_extra_too = 1;
220 /*
221  * bbr_high_gain is the 2/ln(2) value we need
222  * to double the sending rate in startup. This
223  * is used for both cwnd and hptsi gain's.
224  */
225 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1;
226 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1;
227 static int32_t bbr_use_lower_gain_in_startup = 1;
228 
229 /* thresholds for reduction on drain in sub-states/drain */
230 static int32_t bbr_drain_rtt = BBR_SRTT;
231 static int32_t bbr_drain_floor = 88;
232 static int32_t google_allow_early_out = 1;
233 static int32_t google_consider_lost = 1;
234 static int32_t bbr_drain_drop_mul = 4;
235 static int32_t bbr_drain_drop_div = 5;
236 static int32_t bbr_rand_ot = 50;
237 static int32_t bbr_can_force_probertt = 0;
238 static int32_t bbr_can_adjust_probertt = 1;
239 static int32_t bbr_probertt_sets_rtt = 0;
240 static int32_t bbr_can_use_ts_for_rtt = 1;
241 static int32_t bbr_is_ratio = 0;
242 static int32_t bbr_sub_drain_app_limit = 1;
243 static int32_t bbr_prtt_slam_cwnd = 1;
244 static int32_t bbr_sub_drain_slam_cwnd = 1;
245 static int32_t bbr_slam_cwnd_in_main_drain = 1;
246 static int32_t bbr_filter_len_sec = 6;	/* How long does the rttProp filter
247 					 * hold */
248 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4);
249 /*
250  * bbr_drain_gain is the reverse of the high_gain
251  * designed to drain back out the standing queue
252  * that is formed in startup by causing a larger
253  * hptsi gain and thus drainging the packets
254  * in flight.
255  */
256 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885;
257 static int32_t bbr_rttprobe_gain = 192;
258 
259 /*
260  * The cwnd_gain is the default cwnd gain applied when
261  * calculating a target cwnd. Note that the cwnd is
262  * a secondary factor in the way BBR works (see the
263  * paper and think about it, it will take some time).
264  * Basically the hptsi_gain spreads the packets out
265  * so you never get more than BDP to the peer even
266  * if the cwnd is high. In our implemenation that
267  * means in non-recovery/retransmission scenarios
268  * cwnd will never be reached by the flight-size.
269  */
270 static int32_t bbr_cwnd_gain = BBR_UNIT * 2;
271 static int32_t bbr_tlp_type_to_use = BBR_SRTT;
272 static int32_t bbr_delack_time = 100000;	/* 100ms in useconds */
273 static int32_t bbr_sack_not_required = 0;	/* set to one to allow non-sack to use bbr */
274 static int32_t bbr_initial_bw_bps = 62500;	/* 500kbps in bytes ps */
275 static int32_t bbr_ignore_data_after_close = 1;
276 static int16_t bbr_hptsi_gain[] = {
277 	(BBR_UNIT *5 / 4),
278 	(BBR_UNIT * 3 / 4),
279 	BBR_UNIT,
280 	BBR_UNIT,
281 	BBR_UNIT,
282 	BBR_UNIT,
283 	BBR_UNIT,
284 	BBR_UNIT
285 };
286 int32_t bbr_use_rack_resend_cheat = 1;
287 int32_t bbr_sends_full_iwnd = 1;
288 
289 #define BBR_HPTSI_GAIN_MAX 8
290 /*
291  * The BBR module incorporates a number of
292  * TCP ideas that have been put out into the IETF
293  * over the last few years:
294  * - Yuchung Cheng's RACK TCP (for which its named) that
295  *    will stop us using the number of dup acks and instead
296  *    use time as the gage of when we retransmit.
297  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
298  *    of Dukkipati et.al.
299  * - Van Jacobson's et.al BBR.
300  *
301  * RACK depends on SACK, so if an endpoint arrives that
302  * cannot do SACK the state machine below will shuttle the
303  * connection back to using the "default" TCP stack that is
304  * in FreeBSD.
305  *
306  * To implement BBR and RACK the original TCP stack was first decomposed
307  * into a functional state machine with individual states
308  * for each of the possible TCP connection states. The do_segment
309  * functions role in life is to mandate the connection supports SACK
310  * initially and then assure that the RACK state matches the conenction
311  * state before calling the states do_segment function. Data processing
312  * of inbound segments also now happens in the hpts_do_segment in general
313  * with only one exception. This is so we can keep the connection on
314  * a single CPU.
315  *
316  * Each state is simplified due to the fact that the original do_segment
317  * has been decomposed and we *know* what state we are in (no
318  * switches on the state) and all tests for SACK are gone. This
319  * greatly simplifies what each state does.
320  *
321  * TCP output is also over-written with a new version since it
322  * must maintain the new rack scoreboard and has had hptsi
323  * integrated as a requirment. Still todo is to eliminate the
324  * use of the callout_() system and use the hpts for all
325  * timers as well.
326  */
327 static uint32_t bbr_rtt_probe_time = 200000;	/* 200ms in micro seconds */
328 static uint32_t bbr_rtt_probe_cwndtarg = 4;	/* How many mss's outstanding */
329 static const int32_t bbr_min_req_free = 2;	/* The min we must have on the
330 						 * free list */
331 static int32_t bbr_tlp_thresh = 1;
332 static int32_t bbr_reorder_thresh = 2;
333 static int32_t bbr_reorder_fade = 60000000;	/* 0 - never fade, def
334 						 * 60,000,000 - 60 seconds */
335 static int32_t bbr_pkt_delay = 1000;
336 static int32_t bbr_min_to = 1000;	/* Number of usec's minimum timeout */
337 static int32_t bbr_incr_timers = 1;
338 
339 static int32_t bbr_tlp_min = 10000;	/* 10ms in usecs */
340 static int32_t bbr_delayed_ack_time = 200000;	/* 200ms in usecs */
341 static int32_t bbr_exit_startup_at_loss = 1;
342 
343 /*
344  * bbr_lt_bw_ratio is 1/8th
345  * bbr_lt_bw_diff is  < 4 Kbit/sec
346  */
347 static uint64_t bbr_lt_bw_diff = 4000 / 8;	/* In bytes per second */
348 static uint64_t bbr_lt_bw_ratio = 8;	/* For 1/8th */
349 static uint32_t bbr_lt_bw_max_rtts = 48;	/* How many rtt's do we use
350 						 * the lt_bw for */
351 static uint32_t bbr_lt_intvl_min_rtts = 4;	/* Min num of RTT's to measure
352 						 * lt_bw */
353 static int32_t bbr_lt_intvl_fp = 0;		/* False positive epoch diff */
354 static int32_t bbr_lt_loss_thresh = 196;	/* Lost vs delivered % */
355 static int32_t bbr_lt_fd_thresh = 100;		/* false detection % */
356 
357 static int32_t bbr_verbose_logging = 0;
358 /*
359  * Currently regular tcp has a rto_min of 30ms
360  * the backoff goes 12 times so that ends up
361  * being a total of 122.850 seconds before a
362  * connection is killed.
363  */
364 static int32_t bbr_rto_min_ms = 30;	/* 30ms same as main freebsd */
365 static int32_t bbr_rto_max_sec = 4;	/* 4 seconds */
366 
367 /****************************************************/
368 /* DEFAULT TSO SIZING  (cpu performance impacting)  */
369 /****************************************************/
370 /* What amount is our formula using to get TSO size */
371 static int32_t bbr_hptsi_per_second = 1000;
372 
373 /*
374  * For hptsi under bbr_cross_over connections what is delay
375  * target 7ms (in usec) combined with a seg_max of 2
376  * gets us close to identical google behavior in
377  * TSO size selection (possibly more 1MSS sends).
378  */
379 static int32_t bbr_hptsi_segments_delay_tar = 7000;
380 
381 /* Does pacing delay include overhead's in its time calculations? */
382 static int32_t bbr_include_enet_oh = 0;
383 static int32_t bbr_include_ip_oh = 1;
384 static int32_t bbr_include_tcp_oh = 1;
385 static int32_t bbr_google_discount = 10;
386 
387 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */
388 static int32_t bbr_state_is_pkt_epoch = 0;
389 static int32_t bbr_state_drain_2_tar = 1;
390 /* What is the max the 0 - bbr_cross_over MBPS TSO target
391  * can reach using our delay target. Note that this
392  * value becomes the floor for the cross over
393  * algorithm.
394  */
395 static int32_t bbr_hptsi_segments_max = 2;
396 static int32_t bbr_hptsi_segments_floor = 1;
397 static int32_t bbr_hptsi_utter_max = 0;
398 
399 /* What is the min the 0 - bbr_cross-over MBPS  TSO target can be */
400 static int32_t bbr_hptsi_bytes_min = 1460;
401 static int32_t bbr_all_get_min = 0;
402 
403 /* Cross over point from algo-a to algo-b */
404 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS;
405 
406 /* Do we deal with our restart state? */
407 static int32_t bbr_uses_idle_restart = 0;
408 static int32_t bbr_idle_restart_threshold = 100000;	/* 100ms in useconds */
409 
410 /* Do we allow hardware pacing? */
411 static int32_t bbr_allow_hdwr_pacing = 0;
412 static int32_t bbr_hdwr_pace_adjust = 2;	/* multipler when we calc the tso size */
413 static int32_t bbr_hdwr_pace_floor = 1;
414 static int32_t bbr_hdwr_pacing_delay_cnt = 10;
415 
416 /****************************************************/
417 static int32_t bbr_resends_use_tso = 0;
418 static int32_t bbr_tlp_max_resend = 2;
419 static int32_t bbr_sack_block_limit = 128;
420 
421 #define  BBR_MAX_STAT 19
422 counter_u64_t bbr_state_time[BBR_MAX_STAT];
423 counter_u64_t bbr_state_lost[BBR_MAX_STAT];
424 counter_u64_t bbr_state_resend[BBR_MAX_STAT];
425 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE];
426 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE];
427 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE];
428 counter_u64_t bbr_flows_whdwr_pacing;
429 counter_u64_t bbr_flows_nohdwr_pacing;
430 
431 counter_u64_t bbr_nohdwr_pacing_enobuf;
432 counter_u64_t bbr_hdwr_pacing_enobuf;
433 
434 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr);
435 
436 /*
437  * Static defintions we need for forward declarations.
438  */
439 static uint32_t
440 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain,
441 		      uint32_t useconds_time, uint64_t bw);
442 static uint32_t
443 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain);
444 static void
445 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win);
446 static void
447 bbr_set_probebw_gains(struct tcp_bbr *bbr,  uint32_t cts, uint32_t losses);
448 static void
449 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line,
450 		    int dolog);
451 static uint32_t
452 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain);
453 static void
454 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch,
455 		 int32_t pkt_epoch, uint32_t losses);
456 static uint32_t
457 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts,
458 		     struct bbr_sendmap *rsm);
459 static uint32_t
460 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp);
461 static uint32_t
462 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
463 		    struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts);
464 static void
465 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts,
466 		 int32_t line);
467 static void
468 bbr_set_state_target(struct tcp_bbr *bbr, int line);
469 static void
470 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line);
471 static void
472 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick,
473 		       int event, int line);
474 static void
475 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts);
476 static void
477 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts);
478 static void
479 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
480 		    uint32_t rtt, uint32_t line, uint8_t is_start,
481 		    uint16_t set);
482 static struct bbr_sendmap *
483 bbr_find_lowest_rsm(struct tcp_bbr *bbr);
484 static __inline uint32_t
485 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type);
486 static void
487 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot,
488 		 uint8_t which);
489 static void
490 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts,
491 		  uint32_t time_since_sent, uint32_t srtt,
492 		  uint32_t thresh, uint32_t to);
493 static void
494 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag);
495 static void
496 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot,
497 		    uint32_t del_by, uint32_t cts, uint32_t sloton,
498 		    uint32_t prev_delay);
499 static void
500 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts,
501 		  int32_t line);
502 static void
503 bbr_stop_all_timers(struct tcpcb *tp, struct tcp_bbr *bbr);
504 static void
505 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts);
506 static void
507 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts);
508 static void
509 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts);
510 static void
511 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
512 			  uint32_t cts, uint32_t usecs, uint64_t bw,
513 			  uint32_t override, int mod);
514 static int
515 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt);
516 
517 static inline uint8_t
518 bbr_state_val(struct tcp_bbr *bbr)
519 {
520 	return(bbr->rc_bbr_substate);
521 }
522 
523 static inline uint32_t
524 get_min_cwnd(struct tcp_bbr *bbr)
525 {
526 	int mss;
527 
528 	mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options),
529 		  bbr->r_ctl.rc_pace_max_segs);
530 	if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED)
531 		return (bbr_cwnd_min_val_hs * mss);
532 	else
533 		return (bbr_cwnd_min_val * mss);
534 }
535 
536 static uint32_t
537 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr)
538 {
539 	uint64_t srtt, var;
540 	uint64_t ret_val;
541 
542 	bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
543 	if (tp->t_srtt == 0) {
544 		srtt = (uint64_t)BBR_INITIAL_RTO;
545 		var = 0;
546 	} else {
547 		srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
548 		var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT);
549 	}
550 	TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]),
551 	    bbr_persist_min, bbr_persist_max);
552 	return ((uint32_t)ret_val);
553 }
554 
555 static uint32_t
556 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
557 {
558 	/*
559 	 * Start the FR timer, we do this based on getting the first one in
560 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
561 	 * events we need to stop the running timer (if its running) before
562 	 * starting the new one.
563 	 */
564 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
565 	int32_t idx;
566 	int32_t is_tlp_timer = 0;
567 	struct bbr_sendmap *rsm;
568 
569 	if (bbr->rc_all_timers_stopped) {
570 		/* All timers have been stopped none are to run */
571 		return (0);
572 	}
573 	if (bbr->rc_in_persist) {
574 		/* We can't start any timer in persists */
575 		return (bbr_get_persists_timer_val(tp, bbr));
576 	}
577 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
578 	if ((rsm == NULL) ||
579 	    ((tp->t_flags & TF_SACK_PERMIT) == 0) ||
580 	    (tp->t_state < TCPS_ESTABLISHED)) {
581 		/* Nothing on the send map */
582 activate_rxt:
583 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
584 		    sbavail(&tptosocket(tp)->so_snd)) {
585 			uint64_t tov;
586 
587 			time_since_sent = 0;
588 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
589 			if (rsm) {
590 				idx = rsm->r_rtr_cnt - 1;
591 				if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
592 					tstmp_touse = rsm->r_tim_lastsent[idx];
593 				else
594 					tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
595 				if (TSTMP_GT(tstmp_touse, cts))
596 				    time_since_sent = cts - tstmp_touse;
597 			}
598 			bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
599 			if (tp->t_srtt == 0)
600 				tov = BBR_INITIAL_RTO;
601 			else
602 				tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) +
603 				    ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT);
604 			if (tp->t_rxtshift)
605 				tov *= tcp_backoff[tp->t_rxtshift];
606 			if (tov > time_since_sent)
607 				tov -= time_since_sent;
608 			else
609 				tov = bbr->r_ctl.rc_min_to;
610 			TCPT_RANGESET_NOSLOP(to, tov,
611 			    (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC),
612 			    (bbr->rc_max_rto_sec * USECS_IN_SECOND));
613 			bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to);
614 			return (to);
615 		}
616 		return (0);
617 	}
618 	if (rsm->r_flags & BBR_ACKED) {
619 		rsm = bbr_find_lowest_rsm(bbr);
620 		if (rsm == NULL) {
621 			/* No lowest? */
622 			goto activate_rxt;
623 		}
624 	}
625 	/* Convert from ms to usecs */
626 	if (rsm->r_flags & BBR_SACK_PASSED) {
627 		if ((tp->t_flags & TF_SENTFIN) &&
628 		    ((tp->snd_max - tp->snd_una) == 1) &&
629 		    (rsm->r_flags & BBR_HAS_FIN)) {
630 			/*
631 			 * We don't start a bbr rack timer if all we have is
632 			 * a FIN outstanding.
633 			 */
634 			goto activate_rxt;
635 		}
636 		srtt = bbr_get_rtt(bbr, BBR_RTT_RACK);
637 		thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm);
638 		idx = rsm->r_rtr_cnt - 1;
639 		exp = rsm->r_tim_lastsent[idx] + thresh;
640 		if (SEQ_GEQ(exp, cts)) {
641 			to = exp - cts;
642 			if (to < bbr->r_ctl.rc_min_to) {
643 				to = bbr->r_ctl.rc_min_to;
644 			}
645 		} else {
646 			to = bbr->r_ctl.rc_min_to;
647 		}
648 	} else {
649 		/* Ok we need to do a TLP not RACK */
650 		if (bbr->rc_tlp_in_progress != 0) {
651 			/*
652 			 * The previous send was a TLP.
653 			 */
654 			goto activate_rxt;
655 		}
656 		rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
657 		if (rsm == NULL) {
658 			/* We found no rsm to TLP with. */
659 			goto activate_rxt;
660 		}
661 		if (rsm->r_flags & BBR_HAS_FIN) {
662 			/* If its a FIN we don't do TLP */
663 			rsm = NULL;
664 			goto activate_rxt;
665 		}
666 		time_since_sent = 0;
667 		idx = rsm->r_rtr_cnt - 1;
668 		if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
669 			tstmp_touse = rsm->r_tim_lastsent[idx];
670 		else
671 			tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
672 		if (TSTMP_GT(tstmp_touse, cts))
673 		    time_since_sent = cts - tstmp_touse;
674 		is_tlp_timer = 1;
675 		srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use);
676 		thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts);
677 		if (thresh > time_since_sent)
678 			to = thresh - time_since_sent;
679 		else
680 			to = bbr->r_ctl.rc_min_to;
681 		if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
682 			/*
683 			 * If the TLP time works out to larger than the max
684 			 * RTO lets not do TLP.. just RTO.
685 			 */
686 			goto activate_rxt;
687 		}
688 		if ((bbr->rc_tlp_rtx_out == 1) &&
689 		    (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) {
690 			/*
691 			 * Second retransmit of the same TLP
692 			 * lets not.
693 			 */
694 			bbr->rc_tlp_rtx_out = 0;
695 			goto activate_rxt;
696 		}
697 		if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) {
698 			/*
699 			 * The tail is no longer the last one I did a probe
700 			 * on
701 			 */
702 			bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
703 			bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
704 		}
705 	}
706 	if (is_tlp_timer == 0) {
707 		BBR_STAT_INC(bbr_to_arm_rack);
708 		bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
709 	} else {
710 		bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to);
711 		if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
712 			/*
713 			 * We have exceeded how many times we can retran the
714 			 * current TLP timer, switch to the RTO timer.
715 			 */
716 			goto activate_rxt;
717 		} else {
718 			BBR_STAT_INC(bbr_to_arm_tlp);
719 			bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
720 		}
721 	}
722 	return (to);
723 }
724 
725 static inline int32_t
726 bbr_minseg(struct tcp_bbr *bbr)
727 {
728 	return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options);
729 }
730 
731 static void
732 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len)
733 {
734 	struct inpcb *inp = tptoinpcb(tp);
735 	struct hpts_diag diag;
736 	uint32_t delayed_ack = 0;
737 	uint32_t left = 0;
738 	uint32_t hpts_timeout;
739 	uint8_t stopped;
740 	int32_t delay_calc = 0;
741 	uint32_t prev_delay = 0;
742 
743 	if (tcp_in_hpts(inp)) {
744 		/* A previous call is already set up */
745 		return;
746 	}
747 	if ((tp->t_state == TCPS_CLOSED) ||
748 	    (tp->t_state == TCPS_LISTEN)) {
749 		return;
750 	}
751 	stopped = bbr->rc_tmr_stopped;
752 	if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
753 		left = bbr->r_ctl.rc_timer_exp - cts;
754 	}
755 	bbr->r_ctl.rc_hpts_flags = 0;
756 	bbr->r_ctl.rc_timer_exp = 0;
757 	prev_delay = bbr->r_ctl.rc_last_delay_val;
758 	if (bbr->r_ctl.rc_last_delay_val &&
759 	    (slot == 0)) {
760 		/*
761 		 * If a previous pacer delay was in place we
762 		 * are not coming from the output side (where
763 		 * we calculate a delay, more likely a timer).
764 		 */
765 		slot = bbr->r_ctl.rc_last_delay_val;
766 		if (TSTMP_GT(cts, bbr->rc_pacer_started)) {
767 			/* Compensate for time passed  */
768 			delay_calc = cts - bbr->rc_pacer_started;
769 			if (delay_calc <= slot)
770 				slot -= delay_calc;
771 		}
772 	}
773 	/* Do we have early to make up for by pushing out the pacing time? */
774 	if (bbr->r_agg_early_set) {
775 		bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2);
776 		slot += bbr->r_ctl.rc_agg_early;
777 		bbr->r_ctl.rc_agg_early = 0;
778 		bbr->r_agg_early_set = 0;
779 	}
780 	/* Are we running a total debt that needs to be compensated for? */
781 	if (bbr->r_ctl.rc_hptsi_agg_delay) {
782 		if (slot > bbr->r_ctl.rc_hptsi_agg_delay) {
783 			/* We nuke the delay */
784 			slot -= bbr->r_ctl.rc_hptsi_agg_delay;
785 			bbr->r_ctl.rc_hptsi_agg_delay = 0;
786 		} else {
787 			/* We nuke some of the delay, put in a minimal 100usecs  */
788 			bbr->r_ctl.rc_hptsi_agg_delay -= slot;
789 			bbr->r_ctl.rc_last_delay_val = slot = 100;
790 		}
791 	}
792 	bbr->r_ctl.rc_last_delay_val = slot;
793 	hpts_timeout = bbr_timer_start(tp, bbr, cts);
794 	if (tp->t_flags & TF_DELACK) {
795 		if (bbr->rc_in_persist == 0) {
796 			delayed_ack = bbr_delack_time;
797 		} else {
798 			/*
799 			 * We are in persists and have
800 			 * gotten a new data element.
801 			 */
802 			if (hpts_timeout > bbr_delack_time) {
803 				/*
804 				 * Lets make the persists timer (which acks)
805 				 * be the smaller of hpts_timeout and bbr_delack_time.
806 				 */
807 				hpts_timeout = bbr_delack_time;
808 			}
809 		}
810 	}
811 	if (delayed_ack &&
812 	    ((hpts_timeout == 0) ||
813 	     (delayed_ack < hpts_timeout))) {
814 		/* We need a Delayed ack timer */
815 		bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
816 		hpts_timeout = delayed_ack;
817 	}
818 	if (slot) {
819 		/* Mark that we have a pacing timer up */
820 		BBR_STAT_INC(bbr_paced_segments);
821 		bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
822 	}
823 	/*
824 	 * If no timers are going to run and we will fall off thfe hptsi
825 	 * wheel, we resort to a keep-alive timer if its configured.
826 	 */
827 	if ((hpts_timeout == 0) &&
828 	    (slot == 0)) {
829 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
830 		    (tp->t_state <= TCPS_CLOSING)) {
831 			/*
832 			 * Ok we have no timer (persists, rack, tlp, rxt  or
833 			 * del-ack), we don't have segments being paced. So
834 			 * all that is left is the keepalive timer.
835 			 */
836 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
837 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
838 			} else {
839 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
840 			}
841 			bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
842 		}
843 	}
844 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
845 	    (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
846 		/*
847 		 * RACK, TLP, persists and RXT timers all are restartable
848 		 * based on actions input .. i.e we received a packet (ack
849 		 * or sack) and that changes things (rw, or snd_una etc).
850 		 * Thus we can restart them with a new value. For
851 		 * keep-alive, delayed_ack we keep track of what was left
852 		 * and restart the timer with a smaller value.
853 		 */
854 		if (left < hpts_timeout)
855 			hpts_timeout = left;
856 	}
857 	if (bbr->r_ctl.rc_incr_tmrs && slot &&
858 	    (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
859 		/*
860 		 * If configured to do so, and the timer is either
861 		 * the TLP or RXT timer, we need to increase the timeout
862 		 * by the pacing time. Consider the bottleneck at my
863 		 * machine as an example, we are sending something
864 		 * to start a TLP on. The last packet won't be emitted
865 		 * fully until the pacing time (the bottleneck will hold
866 		 * the data in place). Once the packet is emitted that
867 		 * is when we want to start waiting for the TLP. This
868 		 * is most evident with hardware pacing (where the nic
869 		 * is holding the packet(s) before emitting). But it
870 		 * can also show up in the network so we do it for all
871 		 * cases. Technically we would take off one packet from
872 		 * this extra delay but this is easier and being more
873 		 * conservative is probably better.
874 		 */
875 		hpts_timeout += slot;
876 	}
877 	if (hpts_timeout) {
878 		/*
879 		 * Hack alert for now we can't time-out over 2147 seconds (a
880 		 * bit more than 35min)
881 		 */
882 		if (hpts_timeout > 0x7ffffffe)
883 			hpts_timeout = 0x7ffffffe;
884 		bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
885 	} else
886 		bbr->r_ctl.rc_timer_exp = 0;
887 	if ((slot) &&
888 	    (bbr->rc_use_google ||
889 	     bbr->output_error_seen ||
890 	     (slot <= hpts_timeout))  ) {
891 		/*
892 		 * Tell LRO that it can queue packets while
893 		 * we pace.
894 		 */
895 		bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
896 		if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
897 		    (bbr->rc_cwnd_limited == 0)) {
898 			/*
899 			 * If we are not cwnd limited and we
900 			 * are running a rack timer we put on
901 			 * the do not disturbe even for sack.
902 			 */
903 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
904 		} else
905 			inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
906 		bbr->rc_pacer_started = cts;
907 
908 		(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot),
909 					   __LINE__, &diag);
910 		bbr->rc_timer_first = 0;
911 		bbr->bbr_timer_src = frm;
912 		bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1);
913 		bbr_log_hpts_diag(bbr, cts, &diag);
914 	} else if (hpts_timeout) {
915 		(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
916 					   __LINE__, &diag);
917 		/*
918 		 * We add the flag here as well if the slot is set,
919 		 * since hpts will call in to clear the queue first before
920 		 * calling the output routine (which does our timers).
921 		 * We don't want to set the flag if its just a timer
922 		 * else the arrival of data might (that causes us
923 		 * to send more) might get delayed. Imagine being
924 		 * on a keep-alive timer and a request comes in for
925 		 * more data.
926 		 */
927 		if (slot)
928 			bbr->rc_pacer_started = cts;
929 		if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
930 		    (bbr->rc_cwnd_limited == 0)) {
931 			/*
932 			 * For a rack timer, don't wake us even
933 			 * if a sack arrives as long as we are
934 			 * not cwnd limited.
935 			 */
936 			bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
937 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
938 		} else {
939 			/* All other timers wake us up */
940 			bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
941 			inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
942 		}
943 		bbr->bbr_timer_src = frm;
944 		bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0);
945 		bbr_log_hpts_diag(bbr, cts, &diag);
946 		bbr->rc_timer_first = 1;
947 	}
948 	bbr->rc_tmr_stopped = 0;
949 	bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay);
950 }
951 
952 static void
953 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb)
954 {
955 	/*
956 	 * We received an ack, and then did not call send or were bounced
957 	 * out due to the hpts was running. Now a timer is up as well, is it
958 	 * the right timer?
959 	 */
960 	struct inpcb *inp;
961 	struct bbr_sendmap *rsm;
962 	uint32_t hpts_timeout;
963 	int tmr_up;
964 
965 	tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
966 	if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
967 		return;
968 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
969 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
970 	    (tmr_up == PACE_TMR_RXT)) {
971 		/* Should be an RXT */
972 		return;
973 	}
974 	inp = bbr->rc_inp;
975 	if (rsm == NULL) {
976 		/* Nothing outstanding? */
977 		if (tp->t_flags & TF_DELACK) {
978 			if (tmr_up == PACE_TMR_DELACK)
979 				/*
980 				 * We are supposed to have delayed ack up
981 				 * and we do
982 				 */
983 				return;
984 		} else if (sbavail(&inp->inp_socket->so_snd) &&
985 		    (tmr_up == PACE_TMR_RXT)) {
986 			/*
987 			 * if we hit enobufs then we would expect the
988 			 * possibility of nothing outstanding and the RXT up
989 			 * (and the hptsi timer).
990 			 */
991 			return;
992 		} else if (((V_tcp_always_keepalive ||
993 			    inp->inp_socket->so_options & SO_KEEPALIVE) &&
994 			    (tp->t_state <= TCPS_CLOSING)) &&
995 			    (tmr_up == PACE_TMR_KEEP) &&
996 		    (tp->snd_max == tp->snd_una)) {
997 			/* We should have keep alive up and we do */
998 			return;
999 		}
1000 	}
1001 	if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) {
1002 		if ((tp->t_flags & TF_SENTFIN) &&
1003 		    ((tp->snd_max - tp->snd_una) == 1) &&
1004 		    (rsm->r_flags & BBR_HAS_FIN)) {
1005 			/* needs to be a RXT */
1006 			if (tmr_up == PACE_TMR_RXT)
1007 				return;
1008 			else
1009 				goto wrong_timer;
1010 		} else if (tmr_up == PACE_TMR_RACK)
1011 			return;
1012 		else
1013 			goto wrong_timer;
1014 	} else if (rsm && (tmr_up == PACE_TMR_RACK)) {
1015 		/* Rack timer has priority if we have data out */
1016 		return;
1017 	} else if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1018 		    ((tmr_up == PACE_TMR_TLP) ||
1019 	    (tmr_up == PACE_TMR_RXT))) {
1020 		/*
1021 		 * Either a TLP or RXT is fine if no sack-passed is in place
1022 		 * and data is outstanding.
1023 		 */
1024 		return;
1025 	} else if (tmr_up == PACE_TMR_DELACK) {
1026 		/*
1027 		 * If the delayed ack was going to go off before the
1028 		 * rtx/tlp/rack timer were going to expire, then that would
1029 		 * be the timer in control. Note we don't check the time
1030 		 * here trusting the code is correct.
1031 		 */
1032 		return;
1033 	}
1034 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1035 	    ((tmr_up == PACE_TMR_RXT) ||
1036 	     (tmr_up == PACE_TMR_TLP) ||
1037 	     (tmr_up == PACE_TMR_RACK))) {
1038 		/*
1039 		 * We have outstanding data and
1040 		 * we *do* have a RACK, TLP or RXT
1041 		 * timer running. We won't restart
1042 		 * anything here since thats probably ok we
1043 		 * will get called with some timer here shortly.
1044 		 */
1045 		return;
1046 	}
1047 	/*
1048 	 * Ok the timer originally started is not what we want now. We will
1049 	 * force the hpts to be stopped if any, and restart with the slot
1050 	 * set to what was in the saved slot.
1051 	 */
1052 wrong_timer:
1053 	if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) {
1054 		if (tcp_in_hpts(inp))
1055 			tcp_hpts_remove(inp);
1056 		bbr_timer_cancel(bbr, __LINE__, cts);
1057 		bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val,
1058 		    0);
1059 	} else {
1060 		/*
1061 		 * Output is hptsi so we just need to switch the type of
1062 		 * timer. We don't bother with keep-alive, since when we
1063 		 * jump through the output, it will start the keep-alive if
1064 		 * nothing is sent.
1065 		 *
1066 		 * We only need a delayed-ack added and or the hpts_timeout.
1067 		 */
1068 		hpts_timeout = bbr_timer_start(tp, bbr, cts);
1069 		if (tp->t_flags & TF_DELACK) {
1070 			if (hpts_timeout == 0) {
1071 				hpts_timeout = bbr_delack_time;
1072 				bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1073 			}
1074 			else if (hpts_timeout > bbr_delack_time) {
1075 				hpts_timeout = bbr_delack_time;
1076 				bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1077 			}
1078 		}
1079 		if (hpts_timeout) {
1080 			if (hpts_timeout > 0x7ffffffe)
1081 				hpts_timeout = 0x7ffffffe;
1082 			bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
1083 		}
1084 	}
1085 }
1086 
1087 int32_t bbr_clear_lost = 0;
1088 
1089 /*
1090  * Considers the two time values now (cts) and earlier.
1091  * If cts is smaller than earlier, we could have
1092  * had a sequence wrap (our counter wraps every
1093  * 70 min or so) or it could be just clock skew
1094  * getting us two different time values. Clock skew
1095  * will show up within 10ms or so. So in such
1096  * a case (where cts is behind earlier time by
1097  * less than 10ms) we return 0. Otherwise we
1098  * return the true difference between them.
1099  */
1100 static inline uint32_t
1101 bbr_calc_time(uint32_t cts, uint32_t earlier_time) {
1102 	/*
1103 	 * Given two timestamps, the current time stamp cts, and some other
1104 	 * time-stamp taken in theory earlier return the difference. The
1105 	 * trick is here sometimes locking will get the other timestamp
1106 	 * after the cts. If this occurs we need to return 0.
1107 	 */
1108 	if (TSTMP_GEQ(cts, earlier_time))
1109 		return (cts - earlier_time);
1110 	/*
1111 	 * cts is behind earlier_time if its less than 10ms consider it 0.
1112 	 * If its more than 10ms difference then we had a time wrap. Else
1113 	 * its just the normal locking foo. I wonder if we should not go to
1114 	 * 64bit TS and get rid of this issue.
1115 	 */
1116 	if (TSTMP_GEQ((cts + 10000), earlier_time))
1117 		return (0);
1118 	/*
1119 	 * Ok the time must have wrapped. So we need to answer a large
1120 	 * amount of time, which the normal subtraction should do.
1121 	 */
1122 	return (cts - earlier_time);
1123 }
1124 
1125 static int
1126 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS)
1127 {
1128 	uint32_t stat;
1129 	int32_t error;
1130 
1131 	error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t));
1132 	if (error || req->newptr == NULL)
1133 		return error;
1134 
1135 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
1136 	if (error)
1137 		return (error);
1138 	if (stat == 1) {
1139 #ifdef BBR_INVARIANTS
1140 		printf("Clearing BBR lost counters\n");
1141 #endif
1142 		COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT);
1143 		COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT);
1144 		COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT);
1145 	} else if (stat == 2) {
1146 #ifdef BBR_INVARIANTS
1147 		printf("Clearing BBR option counters\n");
1148 #endif
1149 		COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE);
1150 	} else if (stat == 3) {
1151 #ifdef BBR_INVARIANTS
1152 		printf("Clearing BBR stats counters\n");
1153 #endif
1154 		COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE);
1155 	} else if (stat == 4) {
1156 #ifdef BBR_INVARIANTS
1157 		printf("Clearing BBR out-size counters\n");
1158 #endif
1159 		COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE);
1160 	}
1161 	bbr_clear_lost = 0;
1162 	return (0);
1163 }
1164 
1165 static void
1166 bbr_init_sysctls(void)
1167 {
1168 	struct sysctl_oid *bbr_probertt;
1169 	struct sysctl_oid *bbr_hptsi;
1170 	struct sysctl_oid *bbr_measure;
1171 	struct sysctl_oid *bbr_cwnd;
1172 	struct sysctl_oid *bbr_timeout;
1173 	struct sysctl_oid *bbr_states;
1174 	struct sysctl_oid *bbr_startup;
1175 	struct sysctl_oid *bbr_policer;
1176 
1177 	/* Probe rtt controls */
1178 	bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1179 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1180 	    OID_AUTO,
1181 	    "probertt",
1182 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1183 	    "");
1184 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1185 	    SYSCTL_CHILDREN(bbr_probertt),
1186 	    OID_AUTO, "gain", CTLFLAG_RW,
1187 	    &bbr_rttprobe_gain, 192,
1188 	    "What is the filter gain drop in probe_rtt (0=disable)?");
1189 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1190 	    SYSCTL_CHILDREN(bbr_probertt),
1191 	    OID_AUTO, "cwnd", CTLFLAG_RW,
1192 	    &bbr_rtt_probe_cwndtarg, 4,
1193 	    "How many mss's are outstanding during probe-rtt");
1194 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1195 	    SYSCTL_CHILDREN(bbr_probertt),
1196 	    OID_AUTO, "int", CTLFLAG_RW,
1197 	    &bbr_rtt_probe_limit, 4000000,
1198 	    "If RTT has not shrank in this many micro-seconds enter probe-rtt");
1199 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1200 	    SYSCTL_CHILDREN(bbr_probertt),
1201 	    OID_AUTO, "mintime", CTLFLAG_RW,
1202 	    &bbr_rtt_probe_time, 200000,
1203 	    "How many microseconds in probe-rtt");
1204 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1205 	    SYSCTL_CHILDREN(bbr_probertt),
1206 	    OID_AUTO, "filter_len_sec", CTLFLAG_RW,
1207 	    &bbr_filter_len_sec, 6,
1208 	    "How long in seconds does the rttProp filter run?");
1209 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1210 	    SYSCTL_CHILDREN(bbr_probertt),
1211 	    OID_AUTO, "drain_rtt", CTLFLAG_RW,
1212 	    &bbr_drain_rtt, BBR_SRTT,
1213 	    "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?");
1214 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1215 	    SYSCTL_CHILDREN(bbr_probertt),
1216 	    OID_AUTO, "can_force", CTLFLAG_RW,
1217 	    &bbr_can_force_probertt, 0,
1218 	    "If we keep setting new low rtt's but delay going in probe-rtt can we force in??");
1219 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1220 	    SYSCTL_CHILDREN(bbr_probertt),
1221 	    OID_AUTO, "enter_sets_force", CTLFLAG_RW,
1222 	    &bbr_probertt_sets_rtt, 0,
1223 	    "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?");
1224 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1225 	    SYSCTL_CHILDREN(bbr_probertt),
1226 	    OID_AUTO, "can_adjust", CTLFLAG_RW,
1227 	    &bbr_can_adjust_probertt, 1,
1228 	    "Can we dynamically adjust the probe-rtt limits and times?");
1229 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1230 	    SYSCTL_CHILDREN(bbr_probertt),
1231 	    OID_AUTO, "is_ratio", CTLFLAG_RW,
1232 	    &bbr_is_ratio, 0,
1233 	    "is the limit to filter a ratio?");
1234 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1235 	    SYSCTL_CHILDREN(bbr_probertt),
1236 	    OID_AUTO, "use_cwnd", CTLFLAG_RW,
1237 	    &bbr_prtt_slam_cwnd, 0,
1238 	    "Should we set/recover cwnd?");
1239 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1240 	    SYSCTL_CHILDREN(bbr_probertt),
1241 	    OID_AUTO, "can_use_ts", CTLFLAG_RW,
1242 	    &bbr_can_use_ts_for_rtt, 1,
1243 	    "Can we use the ms timestamp if available for retransmistted rtt calculations?");
1244 
1245 	/* Pacing controls */
1246 	bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1247 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1248 	    OID_AUTO,
1249 	    "pacing",
1250 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1251 	    "");
1252 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1253 	    SYSCTL_CHILDREN(bbr_hptsi),
1254 	    OID_AUTO, "hw_pacing", CTLFLAG_RW,
1255 	    &bbr_allow_hdwr_pacing, 1,
1256 	    "Do we allow hardware pacing?");
1257 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1258 	    SYSCTL_CHILDREN(bbr_hptsi),
1259 	    OID_AUTO, "hw_pacing_limit", CTLFLAG_RW,
1260 	    &bbr_hardware_pacing_limit, 4000,
1261 	    "Do we have a limited number of connections for pacing chelsio (0=no limit)?");
1262 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1263 	    SYSCTL_CHILDREN(bbr_hptsi),
1264 	    OID_AUTO, "hw_pacing_adj", CTLFLAG_RW,
1265 	    &bbr_hdwr_pace_adjust, 2,
1266 	    "Multiplier to calculated tso size?");
1267 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1268 	    SYSCTL_CHILDREN(bbr_hptsi),
1269 	    OID_AUTO, "hw_pacing_floor", CTLFLAG_RW,
1270 	    &bbr_hdwr_pace_floor, 1,
1271 	    "Do we invoke the hardware pacing floor?");
1272 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1273 	    SYSCTL_CHILDREN(bbr_hptsi),
1274 	    OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW,
1275 	    &bbr_hdwr_pacing_delay_cnt, 10,
1276 	    "How many packets must be sent after hdwr pacing is enabled");
1277 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1278 	    SYSCTL_CHILDREN(bbr_hptsi),
1279 	    OID_AUTO, "bw_cross", CTLFLAG_RW,
1280 	    &bbr_cross_over, 3000000,
1281 	    "What is the point where we cross over to linux like TSO size set");
1282 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1283 	    SYSCTL_CHILDREN(bbr_hptsi),
1284 	    OID_AUTO, "seg_deltarg", CTLFLAG_RW,
1285 	    &bbr_hptsi_segments_delay_tar, 7000,
1286 	    "What is the worse case delay target for hptsi < 48Mbp connections");
1287 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1288 	    SYSCTL_CHILDREN(bbr_hptsi),
1289 	    OID_AUTO, "enet_oh", CTLFLAG_RW,
1290 	    &bbr_include_enet_oh, 0,
1291 	    "Do we include the ethernet overhead in calculating pacing delay?");
1292 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1293 	    SYSCTL_CHILDREN(bbr_hptsi),
1294 	    OID_AUTO, "ip_oh", CTLFLAG_RW,
1295 	    &bbr_include_ip_oh, 1,
1296 	    "Do we include the IP overhead in calculating pacing delay?");
1297 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1298 	    SYSCTL_CHILDREN(bbr_hptsi),
1299 	    OID_AUTO, "tcp_oh", CTLFLAG_RW,
1300 	    &bbr_include_tcp_oh, 0,
1301 	    "Do we include the TCP overhead in calculating pacing delay?");
1302 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1303 	    SYSCTL_CHILDREN(bbr_hptsi),
1304 	    OID_AUTO, "google_discount", CTLFLAG_RW,
1305 	    &bbr_google_discount, 10,
1306 	    "What is the default google discount percentage wise for pacing (11 = 1.1%%)?");
1307 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1308 	    SYSCTL_CHILDREN(bbr_hptsi),
1309 	    OID_AUTO, "all_get_min", CTLFLAG_RW,
1310 	    &bbr_all_get_min, 0,
1311 	    "If you are less than a MSS do you just get the min?");
1312 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1313 	    SYSCTL_CHILDREN(bbr_hptsi),
1314 	    OID_AUTO, "tso_min", CTLFLAG_RW,
1315 	    &bbr_hptsi_bytes_min, 1460,
1316 	    "For 0 -> 24Mbps what is floor number of segments for TSO");
1317 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1318 	    SYSCTL_CHILDREN(bbr_hptsi),
1319 	    OID_AUTO, "seg_tso_max", CTLFLAG_RW,
1320 	    &bbr_hptsi_segments_max, 6,
1321 	    "For 0 -> 24Mbps what is top number of segments for TSO");
1322 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1323 	    SYSCTL_CHILDREN(bbr_hptsi),
1324 	    OID_AUTO, "seg_floor", CTLFLAG_RW,
1325 	    &bbr_hptsi_segments_floor, 1,
1326 	    "Minimum TSO size we will fall too in segments");
1327 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1328 	    SYSCTL_CHILDREN(bbr_hptsi),
1329 	    OID_AUTO, "utter_max", CTLFLAG_RW,
1330 	    &bbr_hptsi_utter_max, 0,
1331 	    "The absolute maximum that any pacing (outside of hardware) can be");
1332 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1333 	    SYSCTL_CHILDREN(bbr_hptsi),
1334 	    OID_AUTO, "seg_divisor", CTLFLAG_RW,
1335 	    &bbr_hptsi_per_second, 100,
1336 	    "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps ");
1337 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1338 	    SYSCTL_CHILDREN(bbr_hptsi),
1339 	    OID_AUTO, "srtt_mul", CTLFLAG_RW,
1340 	    &bbr_hptsi_max_mul, 1,
1341 	    "The multiplier for pace len max");
1342 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1343 	    SYSCTL_CHILDREN(bbr_hptsi),
1344 	    OID_AUTO, "srtt_div", CTLFLAG_RW,
1345 	    &bbr_hptsi_max_div, 2,
1346 	    "The divisor for pace len max");
1347 	/* Measurement controls */
1348 	bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1349 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1350 	    OID_AUTO,
1351 	    "measure",
1352 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1353 	    "Measurement controls");
1354 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1355 	    SYSCTL_CHILDREN(bbr_measure),
1356 	    OID_AUTO, "min_i_bw", CTLFLAG_RW,
1357 	    &bbr_initial_bw_bps, 62500,
1358 	    "Minimum initial b/w in bytes per second");
1359 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1360 	    SYSCTL_CHILDREN(bbr_measure),
1361 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1362 	    &bbr_sack_not_required, 0,
1363 	    "Do we allow bbr to run on connections not supporting SACK?");
1364 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1365 	    SYSCTL_CHILDREN(bbr_measure),
1366 	    OID_AUTO, "use_google", CTLFLAG_RW,
1367 	    &bbr_use_google_algo, 0,
1368 	    "Use has close to google V1.0 has possible?");
1369 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1370 	    SYSCTL_CHILDREN(bbr_measure),
1371 	    OID_AUTO, "ts_limiting", CTLFLAG_RW,
1372 	    &bbr_ts_limiting, 1,
1373 	    "Do we attempt to use the peers timestamp to limit b/w caculations?");
1374 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1375 	    SYSCTL_CHILDREN(bbr_measure),
1376 	    OID_AUTO, "ts_can_raise", CTLFLAG_RW,
1377 	    &bbr_ts_can_raise, 0,
1378 	    "Can we raise the b/w via timestamp b/w calculation?");
1379 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1380 	    SYSCTL_CHILDREN(bbr_measure),
1381 	    OID_AUTO, "ts_delta", CTLFLAG_RW,
1382 	    &bbr_min_usec_delta, 20000,
1383 	    "How long in usec between ts of our sends in ts validation code?");
1384 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1385 	    SYSCTL_CHILDREN(bbr_measure),
1386 	    OID_AUTO, "ts_peer_delta", CTLFLAG_RW,
1387 	    &bbr_min_peer_delta, 20,
1388 	    "What min numerical value should be between the peer deltas?");
1389 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1390 	    SYSCTL_CHILDREN(bbr_measure),
1391 	    OID_AUTO, "ts_delta_percent", CTLFLAG_RW,
1392 	    &bbr_delta_percent, 150,
1393 	    "What percentage (150 = 15.0) do we allow variance for?");
1394 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1395 	    SYSCTL_CHILDREN(bbr_measure),
1396 	    OID_AUTO, "min_measure_good_bw", CTLFLAG_RW,
1397 	    &bbr_min_measurements_req, 1,
1398 	    "What is the minimum measurement count we need before we switch to our b/w estimate");
1399 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1400 	    SYSCTL_CHILDREN(bbr_measure),
1401 	    OID_AUTO, "min_measure_before_pace", CTLFLAG_RW,
1402 	    &bbr_no_pacing_until, 4,
1403 	    "How many pkt-epoch's (0 is off) do we need before pacing is on?");
1404 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1405 	    SYSCTL_CHILDREN(bbr_measure),
1406 	    OID_AUTO, "quanta", CTLFLAG_RW,
1407 	    &bbr_quanta, 2,
1408 	    "Extra quanta to add when calculating the target (ID section 4.2.3.2).");
1409 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1410 	    SYSCTL_CHILDREN(bbr_measure),
1411 	    OID_AUTO, "noretran", CTLFLAG_RW,
1412 	    &bbr_no_retran, 0,
1413 	    "Should google mode not use retransmission measurements for the b/w estimation?");
1414 	/* State controls */
1415 	bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1416 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1417 	    OID_AUTO,
1418 	    "states",
1419 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1420 	    "State controls");
1421 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1422 	    SYSCTL_CHILDREN(bbr_states),
1423 	    OID_AUTO, "idle_restart", CTLFLAG_RW,
1424 	    &bbr_uses_idle_restart, 0,
1425 	    "Do we use a new special idle_restart state to ramp back up quickly?");
1426 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1427 	    SYSCTL_CHILDREN(bbr_states),
1428 	    OID_AUTO, "idle_restart_threshold", CTLFLAG_RW,
1429 	    &bbr_idle_restart_threshold, 100000,
1430 	    "How long must we be idle before we restart??");
1431 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1432 	    SYSCTL_CHILDREN(bbr_states),
1433 	    OID_AUTO, "use_pkt_epoch", CTLFLAG_RW,
1434 	    &bbr_state_is_pkt_epoch, 0,
1435 	    "Do we use a pkt-epoch for substate if 0 rttProp?");
1436 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1437 	    SYSCTL_CHILDREN(bbr_states),
1438 	    OID_AUTO, "startup_rtt_gain", CTLFLAG_RW,
1439 	    &bbr_rtt_gain_thresh, 0,
1440 	    "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?");
1441 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1442 	    SYSCTL_CHILDREN(bbr_states),
1443 	    OID_AUTO, "drain_floor", CTLFLAG_RW,
1444 	    &bbr_drain_floor, 88,
1445 	    "What is the lowest we can drain (pg) too?");
1446 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1447 	    SYSCTL_CHILDREN(bbr_states),
1448 	    OID_AUTO, "drain_2_target", CTLFLAG_RW,
1449 	    &bbr_state_drain_2_tar, 1,
1450 	    "Do we drain to target in drain substate?");
1451 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1452 	    SYSCTL_CHILDREN(bbr_states),
1453 	    OID_AUTO, "gain_2_target", CTLFLAG_RW,
1454 	    &bbr_gain_to_target, 1,
1455 	    "Does probe bw gain to target??");
1456 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1457 	    SYSCTL_CHILDREN(bbr_states),
1458 	    OID_AUTO, "gain_extra_time", CTLFLAG_RW,
1459 	    &bbr_gain_gets_extra_too, 1,
1460 	    "Does probe bw gain get the extra time too?");
1461 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1462 	    SYSCTL_CHILDREN(bbr_states),
1463 	    OID_AUTO, "ld_div", CTLFLAG_RW,
1464 	    &bbr_drain_drop_div, 5,
1465 	    "Long drain drop divider?");
1466 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1467 	    SYSCTL_CHILDREN(bbr_states),
1468 	    OID_AUTO, "ld_mul", CTLFLAG_RW,
1469 	    &bbr_drain_drop_mul, 4,
1470 	    "Long drain drop multiplier?");
1471 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1472 	    SYSCTL_CHILDREN(bbr_states),
1473 	    OID_AUTO, "rand_ot_disc", CTLFLAG_RW,
1474 	    &bbr_rand_ot, 50,
1475 	    "Random discount of the ot?");
1476 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1477 	    SYSCTL_CHILDREN(bbr_states),
1478 	    OID_AUTO, "dr_filter_life", CTLFLAG_RW,
1479 	    &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT,
1480 	    "How many packet-epochs does the b/w delivery rate last?");
1481 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1482 	    SYSCTL_CHILDREN(bbr_states),
1483 	    OID_AUTO, "subdrain_applimited", CTLFLAG_RW,
1484 	    &bbr_sub_drain_app_limit, 0,
1485 	    "Does our sub-state drain invoke app limited if its long?");
1486 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1487 	    SYSCTL_CHILDREN(bbr_states),
1488 	    OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW,
1489 	    &bbr_sub_drain_slam_cwnd, 0,
1490 	    "Should we set/recover cwnd for sub-state drain?");
1491 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1492 	    SYSCTL_CHILDREN(bbr_states),
1493 	    OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW,
1494 	    &bbr_slam_cwnd_in_main_drain, 0,
1495 	    "Should we set/recover cwnd for main-state drain?");
1496 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1497 	    SYSCTL_CHILDREN(bbr_states),
1498 	    OID_AUTO, "google_gets_earlyout", CTLFLAG_RW,
1499 	    &google_allow_early_out, 1,
1500 	    "Should we allow google probe-bw/drain to exit early at flight target?");
1501 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1502 	    SYSCTL_CHILDREN(bbr_states),
1503 	    OID_AUTO, "google_exit_loss", CTLFLAG_RW,
1504 	    &google_consider_lost, 1,
1505 	    "Should we have losses exit gain of probebw in google mode??");
1506 	/* Startup controls */
1507 	bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1508 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1509 	    OID_AUTO,
1510 	    "startup",
1511 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1512 	    "Startup controls");
1513 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1514 	    SYSCTL_CHILDREN(bbr_startup),
1515 	    OID_AUTO, "cheat_iwnd", CTLFLAG_RW,
1516 	    &bbr_sends_full_iwnd, 1,
1517 	    "Do we not pace but burst out initial windows has our TSO size?");
1518 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1519 	    SYSCTL_CHILDREN(bbr_startup),
1520 	    OID_AUTO, "loss_threshold", CTLFLAG_RW,
1521 	    &bbr_startup_loss_thresh, 2000,
1522 	    "In startup what is the loss threshold in a pe that will exit us from startup?");
1523 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1524 	    SYSCTL_CHILDREN(bbr_startup),
1525 	    OID_AUTO, "use_lowerpg", CTLFLAG_RW,
1526 	    &bbr_use_lower_gain_in_startup, 1,
1527 	    "Should we use a lower hptsi gain if we see loss in startup?");
1528 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1529 	    SYSCTL_CHILDREN(bbr_startup),
1530 	    OID_AUTO, "gain", CTLFLAG_RW,
1531 	    &bbr_start_exit, 25,
1532 	    "What gain percent do we need to see to stay in startup??");
1533 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1534 	    SYSCTL_CHILDREN(bbr_startup),
1535 	    OID_AUTO, "low_gain", CTLFLAG_RW,
1536 	    &bbr_low_start_exit, 15,
1537 	    "What gain percent do we need to see to stay in the lower gain startup??");
1538 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1539 	    SYSCTL_CHILDREN(bbr_startup),
1540 	    OID_AUTO, "loss_exit", CTLFLAG_RW,
1541 	    &bbr_exit_startup_at_loss, 1,
1542 	    "Should we exit startup at loss in an epoch if we are not gaining?");
1543 	/* CWND controls */
1544 	bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1545 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1546 	    OID_AUTO,
1547 	    "cwnd",
1548 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1549 	    "Cwnd controls");
1550 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1551 	    SYSCTL_CHILDREN(bbr_cwnd),
1552 	    OID_AUTO, "tar_rtt", CTLFLAG_RW,
1553 	    &bbr_cwndtarget_rtt_touse, 0,
1554 	    "Target cwnd rtt measurement to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?");
1555 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1556 	    SYSCTL_CHILDREN(bbr_cwnd),
1557 	    OID_AUTO, "may_shrink", CTLFLAG_RW,
1558 	    &bbr_cwnd_may_shrink, 0,
1559 	    "Can the cwnd shrink if it would grow to more than the target?");
1560 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1561 	    SYSCTL_CHILDREN(bbr_cwnd),
1562 	    OID_AUTO, "max_target_limit", CTLFLAG_RW,
1563 	    &bbr_target_cwnd_mult_limit, 8,
1564 	    "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?");
1565 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1566 	    SYSCTL_CHILDREN(bbr_cwnd),
1567 	    OID_AUTO, "highspeed_min", CTLFLAG_RW,
1568 	    &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS,
1569 	    "What is the high-speed min cwnd (rttProp under 1ms)");
1570 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1571 	    SYSCTL_CHILDREN(bbr_cwnd),
1572 	    OID_AUTO, "lowspeed_min", CTLFLAG_RW,
1573 	    &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS,
1574 	    "What is the min cwnd (rttProp > 1ms)");
1575 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1576 	    SYSCTL_CHILDREN(bbr_cwnd),
1577 	    OID_AUTO, "initwin", CTLFLAG_RW,
1578 	    &bbr_def_init_win, 10,
1579 	    "What is the BBR initial window, if 0 use tcp version");
1580 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1581 	    SYSCTL_CHILDREN(bbr_cwnd),
1582 	    OID_AUTO, "do_loss_red", CTLFLAG_RW,
1583 	    &bbr_do_red, 600,
1584 	    "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?");
1585 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1586 	    SYSCTL_CHILDREN(bbr_cwnd),
1587 	    OID_AUTO, "red_scale", CTLFLAG_RW,
1588 	    &bbr_red_scale, 20000,
1589 	    "What RTT do we scale with?");
1590 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1591 	    SYSCTL_CHILDREN(bbr_cwnd),
1592 	    OID_AUTO, "red_growslow", CTLFLAG_RW,
1593 	    &bbr_red_growth_restrict, 1,
1594 	    "Do we restrict cwnd growth for whats in flight?");
1595 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1596 	    SYSCTL_CHILDREN(bbr_cwnd),
1597 	    OID_AUTO, "red_div", CTLFLAG_RW,
1598 	    &bbr_red_div, 2,
1599 	    "If we reduce whats the divisor?");
1600 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1601 	    SYSCTL_CHILDREN(bbr_cwnd),
1602 	    OID_AUTO, "red_mul", CTLFLAG_RW,
1603 	    &bbr_red_mul, 1,
1604 	    "If we reduce whats the mulitiplier?");
1605 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1606 	    SYSCTL_CHILDREN(bbr_cwnd),
1607 	    OID_AUTO, "target_is_unit", CTLFLAG_RW,
1608 	    &bbr_target_is_bbunit, 0,
1609 	    "Is the state target the pacing_gain or BBR_UNIT?");
1610 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1611 	    SYSCTL_CHILDREN(bbr_cwnd),
1612 	    OID_AUTO, "drop_limit", CTLFLAG_RW,
1613 	    &bbr_drop_limit, 0,
1614 	    "Number of segments limit for drop (0=use min_cwnd w/flight)?");
1615 
1616 	/* Timeout controls */
1617 	bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1618 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1619 	    OID_AUTO,
1620 	    "timeout",
1621 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1622 	    "Time out controls");
1623 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1624 	    SYSCTL_CHILDREN(bbr_timeout),
1625 	    OID_AUTO, "delack", CTLFLAG_RW,
1626 	    &bbr_delack_time, 100000,
1627 	    "BBR's delayed ack time");
1628 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1629 	    SYSCTL_CHILDREN(bbr_timeout),
1630 	    OID_AUTO, "tlp_uses", CTLFLAG_RW,
1631 	    &bbr_tlp_type_to_use, 3,
1632 	    "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt");
1633 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1634 	    SYSCTL_CHILDREN(bbr_timeout),
1635 	    OID_AUTO, "persmin", CTLFLAG_RW,
1636 	    &bbr_persist_min, 250000,
1637 	    "What is the minimum time in microseconds between persists");
1638 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1639 	    SYSCTL_CHILDREN(bbr_timeout),
1640 	    OID_AUTO, "persmax", CTLFLAG_RW,
1641 	    &bbr_persist_max, 1000000,
1642 	    "What is the largest delay in microseconds between persists");
1643 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1644 	    SYSCTL_CHILDREN(bbr_timeout),
1645 	    OID_AUTO, "tlp_minto", CTLFLAG_RW,
1646 	    &bbr_tlp_min, 10000,
1647 	    "TLP Min timeout in usecs");
1648 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1649 	    SYSCTL_CHILDREN(bbr_timeout),
1650 	    OID_AUTO, "tlp_dack_time", CTLFLAG_RW,
1651 	    &bbr_delayed_ack_time, 200000,
1652 	    "TLP delayed ack compensation value");
1653 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1654 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1655 	    OID_AUTO, "minrto", CTLFLAG_RW,
1656 	    &bbr_rto_min_ms, 30,
1657 	    "Minimum RTO in ms");
1658 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1659 	    SYSCTL_CHILDREN(bbr_timeout),
1660 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1661 	    &bbr_rto_max_sec, 4,
1662 	    "Maximum RTO in seconds -- should be at least as large as min_rto");
1663 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1664 	    SYSCTL_CHILDREN(bbr_timeout),
1665 	    OID_AUTO, "tlp_retry", CTLFLAG_RW,
1666 	    &bbr_tlp_max_resend, 2,
1667 	    "How many times does TLP retry a single segment or multiple with no ACK");
1668 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1669 	    SYSCTL_CHILDREN(bbr_timeout),
1670 	    OID_AUTO, "minto", CTLFLAG_RW,
1671 	    &bbr_min_to, 1000,
1672 	    "Minimum rack timeout in useconds");
1673 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1674 	    SYSCTL_CHILDREN(bbr_timeout),
1675 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1676 	    &bbr_pkt_delay, 1000,
1677 	    "Extra RACK time (in useconds) besides reordering thresh");
1678 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1679 	    SYSCTL_CHILDREN(bbr_timeout),
1680 	    OID_AUTO, "incr_tmrs", CTLFLAG_RW,
1681 	    &bbr_incr_timers, 1,
1682 	    "Increase the RXT/TLP timer by the pacing time used?");
1683 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1684 	    SYSCTL_CHILDREN(bbr_timeout),
1685 	    OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW,
1686 	    &bbr_marks_rxt_sack_passed, 0,
1687 	    "Mark sack passed on all those not ack'd when a RXT hits?");
1688 	/* Policer controls */
1689 	bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1690 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1691 	    OID_AUTO,
1692 	    "policer",
1693 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1694 	    "Policer controls");
1695 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1696 	    SYSCTL_CHILDREN(bbr_policer),
1697 	    OID_AUTO, "detect_enable", CTLFLAG_RW,
1698 	    &bbr_policer_detection_enabled, 1,
1699 	    "Is policer detection enabled??");
1700 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1701 	    SYSCTL_CHILDREN(bbr_policer),
1702 	    OID_AUTO, "min_pes", CTLFLAG_RW,
1703 	    &bbr_lt_intvl_min_rtts, 4,
1704 	    "Minimum number of PE's?");
1705 	SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1706 	    SYSCTL_CHILDREN(bbr_policer),
1707 	    OID_AUTO, "bwdiff", CTLFLAG_RW,
1708 	    &bbr_lt_bw_diff, (4000/8),
1709 	    "Minimal bw diff?");
1710 	SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1711 	    SYSCTL_CHILDREN(bbr_policer),
1712 	    OID_AUTO, "bwratio", CTLFLAG_RW,
1713 	    &bbr_lt_bw_ratio, 8,
1714 	    "Minimal bw diff?");
1715 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1716 	    SYSCTL_CHILDREN(bbr_policer),
1717 	    OID_AUTO, "from_rack_rxt", CTLFLAG_RW,
1718 	    &bbr_policer_call_from_rack_to, 0,
1719 	    "Do we call the policer detection code from a rack-timeout?");
1720 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1721 	    SYSCTL_CHILDREN(bbr_policer),
1722 	    OID_AUTO, "false_postive", CTLFLAG_RW,
1723 	    &bbr_lt_intvl_fp, 0,
1724 	    "What packet epoch do we do false-positive detection at (0=no)?");
1725 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1726 	    SYSCTL_CHILDREN(bbr_policer),
1727 	    OID_AUTO, "loss_thresh", CTLFLAG_RW,
1728 	    &bbr_lt_loss_thresh, 196,
1729 	    "Loss threshold 196 = 19.6%?");
1730 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1731 	    SYSCTL_CHILDREN(bbr_policer),
1732 	    OID_AUTO, "false_postive_thresh", CTLFLAG_RW,
1733 	    &bbr_lt_fd_thresh, 100,
1734 	    "What percentage is the false detection threshold (150=15.0)?");
1735 	/* All the rest */
1736 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1737 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1738 	    OID_AUTO, "cheat_rxt", CTLFLAG_RW,
1739 	    &bbr_use_rack_resend_cheat, 0,
1740 	    "Do we burst 1ms between sends on retransmissions (like rack)?");
1741 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1742 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1743 	    OID_AUTO, "error_paceout", CTLFLAG_RW,
1744 	    &bbr_error_base_paceout, 10000,
1745 	    "When we hit an error what is the min to pace out in usec's?");
1746 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1747 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1748 	    OID_AUTO, "kill_paceout", CTLFLAG_RW,
1749 	    &bbr_max_net_error_cnt, 10,
1750 	    "When we hit this many errors in a row, kill the session?");
1751 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1752 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1753 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1754 	    &bbr_ignore_data_after_close, 1,
1755 	    "Do we hold off sending a RST until all pending data is ack'd");
1756 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1757 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1758 	    OID_AUTO, "resend_use_tso", CTLFLAG_RW,
1759 	    &bbr_resends_use_tso, 0,
1760 	    "Can resends use TSO?");
1761 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1762 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1763 	    OID_AUTO, "sblklimit", CTLFLAG_RW,
1764 	    &bbr_sack_block_limit, 128,
1765 	    "When do we start ignoring small sack blocks");
1766 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1767 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1768 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1769 	    &bbr_verbose_logging, 0,
1770 	    "Should BBR black box logging be verbose");
1771 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1772 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1773 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1774 	    &bbr_reorder_thresh, 2,
1775 	    "What factor for rack will be added when seeing reordering (shift right)");
1776 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1777 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1778 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1779 	    &bbr_reorder_fade, 0,
1780 	    "Does reorder detection fade, if so how many ms (0 means never)");
1781 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1782 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1783 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1784 	    &bbr_tlp_thresh, 1,
1785 	    "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1786 	/* Stats and counters */
1787 	/* The pacing counters for hdwr/software can't be in the array */
1788 	bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1789 	bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1790 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1791 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1792 	    OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD,
1793 	    &bbr_hdwr_pacing_enobuf,
1794 	    "Total number of enobufs for hardware paced flows");
1795 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1796 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1797 	    OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD,
1798 	    &bbr_nohdwr_pacing_enobuf,
1799 	    "Total number of enobufs for non-hardware paced flows");
1800 
1801 	bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK);
1802 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1803 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1804 	    OID_AUTO, "hdwr_pacing", CTLFLAG_RD,
1805 	    &bbr_flows_whdwr_pacing,
1806 	    "Total number of hardware paced flows");
1807 	bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK);
1808 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1809 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1810 	    OID_AUTO, "software_pacing", CTLFLAG_RD,
1811 	    &bbr_flows_nohdwr_pacing,
1812 	    "Total number of software paced flows");
1813 	COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK);
1814 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1815 	    OID_AUTO, "stats", CTLFLAG_RD,
1816 	    bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats");
1817 	COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK);
1818 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1819 	    OID_AUTO, "opts", CTLFLAG_RD,
1820 	    bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats");
1821 	COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK);
1822 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1823 	    OID_AUTO, "lost", CTLFLAG_RD,
1824 	    bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur");
1825 	COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK);
1826 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1827 	    OID_AUTO, "stateresend", CTLFLAG_RD,
1828 	    bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend");
1829 	COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK);
1830 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1831 	    OID_AUTO, "statetime", CTLFLAG_RD,
1832 	    bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states");
1833 	COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1834 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1835 	    OID_AUTO, "outsize", CTLFLAG_RD,
1836 	    bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls");
1837 	SYSCTL_ADD_PROC(&bbr_sysctl_ctx,
1838 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1839 	    OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1840 	    &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters");
1841 }
1842 
1843 static void
1844 bbr_counter_destroy(void)
1845 {
1846 	COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE);
1847 	COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE);
1848 	COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE);
1849 	COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT);
1850 	COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT);
1851 	COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT);
1852 	counter_u64_free(bbr_nohdwr_pacing_enobuf);
1853 	counter_u64_free(bbr_hdwr_pacing_enobuf);
1854 	counter_u64_free(bbr_flows_whdwr_pacing);
1855 	counter_u64_free(bbr_flows_nohdwr_pacing);
1856 
1857 }
1858 
1859 static __inline void
1860 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts)
1861 {
1862 	memset(l, 0, sizeof(union tcp_log_stackspecific));
1863 	l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate;
1864 	l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate);
1865 	l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
1866 	l->bw_inuse = bbr_get_bw(bbr);
1867 	l->inflight = ctf_flight_size(bbr->rc_tp,
1868 			  (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
1869 	l->applimited = bbr->r_ctl.r_app_limited_until;
1870 	l->delivered = bbr->r_ctl.rc_delivered;
1871 	l->timeStamp = cts;
1872 	l->lost = bbr->r_ctl.rc_lost;
1873 	l->bbr_state = bbr->rc_bbr_state;
1874 	l->bbr_substate = bbr_state_val(bbr);
1875 	l->epoch = bbr->r_ctl.rc_rtt_epoch;
1876 	l->lt_epoch = bbr->r_ctl.rc_lt_epoch;
1877 	l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
1878 	l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain;
1879 	l->inhpts = tcp_in_hpts(bbr->rc_inp);
1880 	l->use_lt_bw = bbr->rc_lt_use_bw;
1881 	l->pkts_out = bbr->r_ctl.rc_flight_at_input;
1882 	l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch;
1883 }
1884 
1885 static void
1886 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason)
1887 {
1888 	if (tcp_bblogging_on(bbr->rc_tp)) {
1889 		union tcp_log_stackspecific log;
1890 
1891 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1892 		log.u_bbr.flex1 = 0;
1893 		log.u_bbr.flex2 = 0;
1894 		log.u_bbr.flex5 = 0;
1895 		log.u_bbr.flex3 = 0;
1896 		log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate;
1897 		log.u_bbr.flex7 = reason;
1898 		log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt;
1899 		log.u_bbr.flex8 = 0;
1900 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1901 		    &bbr->rc_inp->inp_socket->so_rcv,
1902 		    &bbr->rc_inp->inp_socket->so_snd,
1903 		    BBR_LOG_BW_RED_EV, 0,
1904 		    0, &log, false, &bbr->rc_tv);
1905 	}
1906 }
1907 
1908 static void
1909 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count)
1910 {
1911 	if (tcp_bblogging_on(bbr->rc_tp)) {
1912 		union tcp_log_stackspecific log;
1913 
1914 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1915 		log.u_bbr.flex1 = seq;
1916 		log.u_bbr.flex2 = count;
1917 		log.u_bbr.flex8 = mode;
1918 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1919 		    &bbr->rc_inp->inp_socket->so_rcv,
1920 		    &bbr->rc_inp->inp_socket->so_snd,
1921 		    BBR_LOG_LOWGAIN, 0,
1922 		    0, &log, false, &bbr->rc_tv);
1923 	}
1924 }
1925 
1926 static void
1927 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling,
1928     uint8_t reason, uint32_t p_maxseg, int len)
1929 {
1930 	if (tcp_bblogging_on(bbr->rc_tp)) {
1931 		union tcp_log_stackspecific log;
1932 
1933 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1934 		log.u_bbr.flex1 = p_maxseg;
1935 		log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags;
1936 		log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
1937 		log.u_bbr.flex4 = reason;
1938 		log.u_bbr.flex5 = bbr->rc_in_persist;
1939 		log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val;
1940 		log.u_bbr.flex7 = p_maxseg;
1941 		log.u_bbr.flex8 = bbr->rc_in_persist;
1942 		log.u_bbr.pkts_out = 0;
1943 		log.u_bbr.applimited = len;
1944 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1945 		    &bbr->rc_inp->inp_socket->so_rcv,
1946 		    &bbr->rc_inp->inp_socket->so_snd,
1947 		    BBR_LOG_JUSTRET, 0,
1948 		    tlen, &log, false, &bbr->rc_tv);
1949 	}
1950 }
1951 
1952 static void
1953 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq)
1954 {
1955 	if (tcp_bblogging_on(bbr->rc_tp)) {
1956 		union tcp_log_stackspecific log;
1957 
1958 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1959 		log.u_bbr.flex1 = seq;
1960 		log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
1961 		log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start;
1962 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1963 		    &bbr->rc_inp->inp_socket->so_rcv,
1964 		    &bbr->rc_inp->inp_socket->so_snd,
1965 		    BBR_LOG_ENTREC, 0,
1966 		    0, &log, false, &bbr->rc_tv);
1967 	}
1968 }
1969 
1970 static void
1971 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)
1972 {
1973 	if (tcp_bblogging_on(tp)) {
1974 		union tcp_log_stackspecific log;
1975 
1976 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1977 		log.u_bbr.flex1 = tso;
1978 		log.u_bbr.flex2 = maxseg;
1979 		log.u_bbr.flex3 = mtu;
1980 		log.u_bbr.flex4 = csum_flags;
1981 		TCP_LOG_EVENTP(tp, NULL,
1982 		    &bbr->rc_inp->inp_socket->so_rcv,
1983 		    &bbr->rc_inp->inp_socket->so_snd,
1984 		    BBR_LOG_MSGSIZE, 0,
1985 		    0, &log, false, &bbr->rc_tv);
1986 	}
1987 }
1988 
1989 static void
1990 bbr_log_flowend(struct tcp_bbr *bbr)
1991 {
1992 	if (tcp_bblogging_on(bbr->rc_tp)) {
1993 		union tcp_log_stackspecific log;
1994 		struct sockbuf *r, *s;
1995 		struct timeval tv;
1996 
1997 		if (bbr->rc_inp->inp_socket) {
1998 			r = &bbr->rc_inp->inp_socket->so_rcv;
1999 			s = &bbr->rc_inp->inp_socket->so_snd;
2000 		} else {
2001 			r = s = NULL;
2002 		}
2003 		bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv));
2004 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2005 		    r, s,
2006 		    TCP_LOG_FLOWEND, 0,
2007 		    0, &log, false, &tv);
2008 	}
2009 }
2010 
2011 static void
2012 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line,
2013     uint32_t lost, uint32_t del)
2014 {
2015 	if (tcp_bblogging_on(bbr->rc_tp)) {
2016 		union tcp_log_stackspecific log;
2017 
2018 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2019 		log.u_bbr.flex1 = lost;
2020 		log.u_bbr.flex2 = del;
2021 		log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw;
2022 		log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt;
2023 		log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2024 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2025 		log.u_bbr.flex7 = line;
2026 		log.u_bbr.flex8 = 0;
2027 		log.u_bbr.inflight = bbr->r_ctl.r_measurement_count;
2028 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2029 		    &bbr->rc_inp->inp_socket->so_rcv,
2030 		    &bbr->rc_inp->inp_socket->so_snd,
2031 		    BBR_LOG_PKT_EPOCH, 0,
2032 		    0, &log, false, &bbr->rc_tv);
2033 	}
2034 }
2035 
2036 static void
2037 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time)
2038 {
2039 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2040 		union tcp_log_stackspecific log;
2041 
2042 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2043 		log.u_bbr.flex1 = bbr->r_ctl.rc_lost;
2044 		log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat;
2045 		log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat;
2046 		log.u_bbr.flex7 = line;
2047 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2048 		    &bbr->rc_inp->inp_socket->so_rcv,
2049 		    &bbr->rc_inp->inp_socket->so_snd,
2050 		    BBR_LOG_TIME_EPOCH, 0,
2051 		    0, &log, false, &bbr->rc_tv);
2052 	}
2053 }
2054 
2055 static void
2056 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth)
2057 {
2058 	if (tcp_bblogging_on(bbr->rc_tp)) {
2059 		union tcp_log_stackspecific log;
2060 
2061 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2062 		log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2063 		log.u_bbr.flex2 = new_tar;
2064 		log.u_bbr.flex3 = line;
2065 		log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2066 		log.u_bbr.flex5 = bbr_quanta;
2067 		log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs;
2068 		log.u_bbr.flex7 = bbr->rc_last_options;
2069 		log.u_bbr.flex8 = meth;
2070 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2071 		    &bbr->rc_inp->inp_socket->so_rcv,
2072 		    &bbr->rc_inp->inp_socket->so_snd,
2073 		    BBR_LOG_STATE_TARGET, 0,
2074 		    0, &log, false, &bbr->rc_tv);
2075 	}
2076 
2077 }
2078 
2079 static void
2080 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2081 {
2082 	if (tcp_bblogging_on(bbr->rc_tp)) {
2083 		union tcp_log_stackspecific log;
2084 
2085 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2086 		log.u_bbr.flex1 = line;
2087 		log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2088 		log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2089 		if (bbr_state_is_pkt_epoch)
2090 			log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
2091 		else
2092 			log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP);
2093 		log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2094 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2095 		log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000);
2096 		log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra;
2097 		log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state;
2098 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2099 		    &bbr->rc_inp->inp_socket->so_rcv,
2100 		    &bbr->rc_inp->inp_socket->so_snd,
2101 		    BBR_LOG_STATE, 0,
2102 		    0, &log, false, &bbr->rc_tv);
2103 	}
2104 }
2105 
2106 static void
2107 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
2108 		    uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond)
2109 {
2110 	if (tcp_bblogging_on(bbr->rc_tp)) {
2111 		union tcp_log_stackspecific log;
2112 
2113 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2114 		log.u_bbr.flex1 = line;
2115 		log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2116 		log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt;
2117 		log.u_bbr.flex4 = applied;
2118 		log.u_bbr.flex5 = rtt;
2119 		log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2120 		log.u_bbr.flex7 = cond;
2121 		log.u_bbr.flex8 = reas;
2122 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2123 		    &bbr->rc_inp->inp_socket->so_rcv,
2124 		    &bbr->rc_inp->inp_socket->so_snd,
2125 		    BBR_LOG_RTT_SHRINKS, 0,
2126 		    0, &log, false, &bbr->rc_tv);
2127 	}
2128 }
2129 
2130 static void
2131 bbr_log_type_exit_rec(struct tcp_bbr *bbr)
2132 {
2133 	if (tcp_bblogging_on(bbr->rc_tp)) {
2134 		union tcp_log_stackspecific log;
2135 
2136 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2137 		log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start;
2138 		log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
2139 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2140 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2141 		    &bbr->rc_inp->inp_socket->so_rcv,
2142 		    &bbr->rc_inp->inp_socket->so_snd,
2143 		    BBR_LOG_EXITREC, 0,
2144 		    0, &log, false, &bbr->rc_tv);
2145 	}
2146 }
2147 
2148 static void
2149 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg,
2150     uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line)
2151 {
2152 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2153 		union tcp_log_stackspecific log;
2154 
2155 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2156 		log.u_bbr.flex1 = line;
2157 		log.u_bbr.flex2 = prev_acked;
2158 		log.u_bbr.flex3 = bytes_this_ack;
2159 		log.u_bbr.flex4 = chg;
2160 		log.u_bbr.flex5 = th_ack;
2161 		log.u_bbr.flex6 = target;
2162 		log.u_bbr.flex8 = meth;
2163 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2164 		    &bbr->rc_inp->inp_socket->so_rcv,
2165 		    &bbr->rc_inp->inp_socket->so_snd,
2166 		    BBR_LOG_CWND, 0,
2167 		    0, &log, false, &bbr->rc_tv);
2168 	}
2169 }
2170 
2171 static void
2172 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin)
2173 {
2174 	/*
2175 	 * Log the rtt sample we are applying to the srtt algorithm in
2176 	 * useconds.
2177 	 */
2178 	if (tcp_bblogging_on(bbr->rc_tp)) {
2179 		union tcp_log_stackspecific log;
2180 
2181 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2182 		log.u_bbr.flex1 = rtt;
2183 		log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time;
2184 		log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay;
2185 		log.u_bbr.flex4 = bbr->rc_tp->ts_offset;
2186 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2187 		log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv);
2188 		log.u_bbr.flex6 = tsin;
2189 		log.u_bbr.flex7 = 0;
2190 		log.u_bbr.flex8 = bbr->rc_ack_was_delayed;
2191 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2192 		    &bbr->rc_inp->inp_socket->so_rcv,
2193 		    &bbr->rc_inp->inp_socket->so_snd,
2194 		    TCP_LOG_RTT, 0,
2195 		    0, &log, false, &bbr->rc_tv);
2196 	}
2197 }
2198 
2199 static void
2200 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit)
2201 {
2202 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2203 		union tcp_log_stackspecific log;
2204 
2205 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2206 		log.u_bbr.flex1 = time_in;
2207 		log.u_bbr.flex2 = line;
2208 		log.u_bbr.flex8 = enter_exit;
2209 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2210 		    &bbr->rc_inp->inp_socket->so_rcv,
2211 		    &bbr->rc_inp->inp_socket->so_snd,
2212 		    BBR_LOG_PERSIST, 0,
2213 		    0, &log, false, &bbr->rc_tv);
2214 	}
2215 }
2216 static void
2217 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts)
2218 {
2219 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2220 		union tcp_log_stackspecific log;
2221 
2222 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2223 		log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age;
2224 		log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2225 		log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2226 		log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time;
2227 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2228 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2229 		    &bbr->rc_inp->inp_socket->so_rcv,
2230 		    &bbr->rc_inp->inp_socket->so_snd,
2231 		    BBR_LOG_ACKCLEAR, 0,
2232 		    0, &log, false, &bbr->rc_tv);
2233 	}
2234 }
2235 
2236 static void
2237 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen,
2238 		  uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m)
2239 {
2240 	if (tcp_bblogging_on(bbr->rc_tp)) {
2241 		union tcp_log_stackspecific log;
2242 		struct timeval tv;
2243 
2244 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2245 		log.u_bbr.flex1 = nsegs;
2246 		log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes;
2247 		if (m) {
2248 			struct timespec ts;
2249 
2250 			log.u_bbr.flex3 = m->m_flags;
2251 			if (m->m_flags & M_TSTMP) {
2252 				mbuf_tstmp2timespec(m, &ts);
2253 				tv.tv_sec = ts.tv_sec;
2254 				tv.tv_usec = ts.tv_nsec / 1000;
2255 				log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv);
2256 			} else {
2257 				log.u_bbr.lt_epoch = 0;
2258 			}
2259 			if (m->m_flags & M_TSTMP_LRO) {
2260 				mbuf_tstmp2timeval(m, &tv);
2261 				log.u_bbr.flex5 = tcp_tv_to_usectick(&tv);
2262 			} else {
2263 				/* No arrival timestamp */
2264 				log.u_bbr.flex5 = 0;
2265 			}
2266 
2267 			log.u_bbr.pkts_out = tcp_get_usecs(&tv);
2268 		} else {
2269 			log.u_bbr.flex3 = 0;
2270 			log.u_bbr.flex5 = 0;
2271 			log.u_bbr.flex6 = 0;
2272 			log.u_bbr.pkts_out = 0;
2273 		}
2274 		log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state;
2275 		log.u_bbr.flex7 = bbr->r_wanted_output;
2276 		log.u_bbr.flex8 = bbr->rc_in_persist;
2277 		TCP_LOG_EVENTP(bbr->rc_tp, th,
2278 		    &bbr->rc_inp->inp_socket->so_rcv,
2279 		    &bbr->rc_inp->inp_socket->so_snd,
2280 		    TCP_LOG_IN, 0,
2281 		    tlen, &log, true, &bbr->rc_tv);
2282 	}
2283 }
2284 
2285 static void
2286 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out)
2287 {
2288 	if (tcp_bblogging_on(bbr->rc_tp)) {
2289 		union tcp_log_stackspecific log;
2290 
2291 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2292 		log.u_bbr.flex1 = did_out;
2293 		log.u_bbr.flex2 = nxt_pkt;
2294 		log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val;
2295 		log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2296 		log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp;
2297 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes;
2298 		log.u_bbr.flex7 = bbr->r_wanted_output;
2299 		log.u_bbr.flex8 = bbr->rc_in_persist;
2300 		log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay;
2301 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2302 		    &bbr->rc_inp->inp_socket->so_rcv,
2303 		    &bbr->rc_inp->inp_socket->so_snd,
2304 		    BBR_LOG_DOSEG_DONE, 0,
2305 		    0, &log, true, &bbr->rc_tv);
2306 	}
2307 }
2308 
2309 static void
2310 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts,
2311     int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz)
2312 {
2313 	if (tcp_bblogging_on(bbr->rc_tp)) {
2314 		union tcp_log_stackspecific log;
2315 
2316 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2317 		log.u_bbr.flex1 = line;
2318 		log.u_bbr.flex2 = o_len;
2319 		log.u_bbr.flex3 = segcnt;
2320 		log.u_bbr.flex4 = segsiz;
2321 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2322 		    &bbr->rc_inp->inp_socket->so_rcv,
2323 		    &bbr->rc_inp->inp_socket->so_snd,
2324 		    BBR_LOG_ENOBUF_JMP, ENOBUFS,
2325 		    len, &log, true, &bbr->rc_tv);
2326 	}
2327 }
2328 
2329 static void
2330 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling)
2331 {
2332 	if (tcp_bblogging_on(bbr->rc_tp)) {
2333 		union tcp_log_stackspecific log;
2334 
2335 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2336 		log.u_bbr.flex1 = timers;
2337 		log.u_bbr.flex2 = ret;
2338 		log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
2339 		log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2340 		log.u_bbr.flex5 = cts;
2341 		log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2342 		log.u_bbr.flex8 = hpts_calling;
2343 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2344 		    &bbr->rc_inp->inp_socket->so_rcv,
2345 		    &bbr->rc_inp->inp_socket->so_snd,
2346 		    BBR_LOG_TO_PROCESS, 0,
2347 		    0, &log, false, &bbr->rc_tv);
2348 	}
2349 }
2350 
2351 static void
2352 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num)
2353 {
2354 	if (tcp_bblogging_on(bbr->rc_tp)) {
2355 		union tcp_log_stackspecific log;
2356 		uint64_t ar;
2357 
2358 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2359 		log.u_bbr.flex1 = bbr->bbr_timer_src;
2360 		log.u_bbr.flex2 = 0;
2361 		log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2362 		ar = (uint64_t)(bbr->r_ctl.rc_resend);
2363 		ar >>= 32;
2364 		ar &= 0x00000000ffffffff;
2365 		log.u_bbr.flex4 = (uint32_t)ar;
2366 		ar = (uint64_t)bbr->r_ctl.rc_resend;
2367 		ar &= 0x00000000ffffffff;
2368 		log.u_bbr.flex5 = (uint32_t)ar;
2369 		log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2370 		log.u_bbr.flex8 = to_num;
2371 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2372 		    &bbr->rc_inp->inp_socket->so_rcv,
2373 		    &bbr->rc_inp->inp_socket->so_snd,
2374 		    BBR_LOG_RTO, 0,
2375 		    0, &log, false, &bbr->rc_tv);
2376 	}
2377 }
2378 
2379 static void
2380 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason)
2381 {
2382 	if (tcp_bblogging_on(bbr->rc_tp)) {
2383 		union tcp_log_stackspecific log;
2384 
2385 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2386 		log.u_bbr.flex1 = flex1;
2387 		log.u_bbr.flex2 = flex2;
2388 		log.u_bbr.flex3 = flex3;
2389 		log.u_bbr.flex4 = 0;
2390 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2391 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2392 		log.u_bbr.flex8 = reason;
2393 		log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw;
2394 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2395 		    &bbr->rc_inp->inp_socket->so_rcv,
2396 		    &bbr->rc_inp->inp_socket->so_snd,
2397 		    BBR_LOG_REDUCE, 0,
2398 		    0, &log, false, &bbr->rc_tv);
2399 	}
2400 }
2401 
2402 static void
2403 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag)
2404 {
2405 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2406 		union tcp_log_stackspecific log;
2407 
2408 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2409 		log.u_bbr.flex1 = diag->p_nxt_slot;
2410 		log.u_bbr.flex2 = diag->p_cur_slot;
2411 		log.u_bbr.flex3 = diag->slot_req;
2412 		log.u_bbr.flex4 = diag->inp_hptsslot;
2413 		log.u_bbr.flex5 = diag->slot_remaining;
2414 		log.u_bbr.flex6 = diag->need_new_to;
2415 		log.u_bbr.flex7 = diag->p_hpts_active;
2416 		log.u_bbr.flex8 = diag->p_on_min_sleep;
2417 		/* Hijack other fields as needed  */
2418 		log.u_bbr.epoch = diag->have_slept;
2419 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
2420 		log.u_bbr.pkts_out = diag->co_ret;
2421 		log.u_bbr.applimited = diag->hpts_sleep_time;
2422 		log.u_bbr.delivered = diag->p_prev_slot;
2423 		log.u_bbr.inflight = diag->p_runningslot;
2424 		log.u_bbr.bw_inuse = diag->wheel_slot;
2425 		log.u_bbr.rttProp = diag->wheel_cts;
2426 		log.u_bbr.delRate = diag->maxslots;
2427 		log.u_bbr.cur_del_rate = diag->p_curtick;
2428 		log.u_bbr.cur_del_rate <<= 32;
2429 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
2430 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2431 		    &bbr->rc_inp->inp_socket->so_rcv,
2432 		    &bbr->rc_inp->inp_socket->so_snd,
2433 		    BBR_LOG_HPTSDIAG, 0,
2434 		    0, &log, false, &bbr->rc_tv);
2435 	}
2436 }
2437 
2438 static void
2439 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
2440     uint32_t thresh, uint32_t to)
2441 {
2442 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2443 		union tcp_log_stackspecific log;
2444 
2445 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2446 		log.u_bbr.flex1 = bbr->rc_tp->t_rttvar;
2447 		log.u_bbr.flex2 = time_since_sent;
2448 		log.u_bbr.flex3 = srtt;
2449 		log.u_bbr.flex4 = thresh;
2450 		log.u_bbr.flex5 = to;
2451 		log.u_bbr.flex6 = bbr->rc_tp->t_srtt;
2452 		log.u_bbr.flex8 = mode;
2453 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2454 		    &bbr->rc_inp->inp_socket->so_rcv,
2455 		    &bbr->rc_inp->inp_socket->so_snd,
2456 		    BBR_LOG_TIMERPREP, 0,
2457 		    0, &log, false, &bbr->rc_tv);
2458 	}
2459 }
2460 
2461 static void
2462 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
2463     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod)
2464 {
2465 	if (tcp_bblogging_on(bbr->rc_tp)) {
2466 		union tcp_log_stackspecific log;
2467 
2468 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2469 		log.u_bbr.flex1 = usecs;
2470 		log.u_bbr.flex2 = len;
2471 		log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff);
2472 		log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff);
2473 		if (override)
2474 			log.u_bbr.flex5 = (1 << 2);
2475 		else
2476 			log.u_bbr.flex5 = 0;
2477 		log.u_bbr.flex6 = override;
2478 		log.u_bbr.flex7 = gain;
2479 		log.u_bbr.flex8 = mod;
2480 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2481 		    &bbr->rc_inp->inp_socket->so_rcv,
2482 		    &bbr->rc_inp->inp_socket->so_snd,
2483 		    BBR_LOG_HPTSI_CALC, 0,
2484 		    len, &log, false, &bbr->rc_tv);
2485 	}
2486 }
2487 
2488 static void
2489 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2490 {
2491 	if (tcp_bblogging_on(bbr->rc_tp)) {
2492 		union tcp_log_stackspecific log;
2493 
2494 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2495 
2496 		log.u_bbr.flex1 = bbr->bbr_timer_src;
2497 		log.u_bbr.flex2 = to;
2498 		log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2499 		log.u_bbr.flex4 = slot;
2500 		log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot;
2501 		log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2502 		log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2;
2503 		log.u_bbr.flex8 = which;
2504 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2505 		    &bbr->rc_inp->inp_socket->so_rcv,
2506 		    &bbr->rc_inp->inp_socket->so_snd,
2507 		    BBR_LOG_TIMERSTAR, 0,
2508 		    0, &log, false, &bbr->rc_tv);
2509 	}
2510 }
2511 
2512 static void
2513 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)
2514 {
2515 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2516 		union tcp_log_stackspecific log;
2517 
2518 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2519 		log.u_bbr.flex1 = thresh;
2520 		log.u_bbr.flex2 = lro;
2521 		log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts;
2522 		log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
2523 		log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2524 		log.u_bbr.flex6 = srtt;
2525 		log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift;
2526 		log.u_bbr.flex8 = frm;
2527 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2528 		    &bbr->rc_inp->inp_socket->so_rcv,
2529 		    &bbr->rc_inp->inp_socket->so_snd,
2530 		    BBR_LOG_THRESH_CALC, 0,
2531 		    0, &log, false, &bbr->rc_tv);
2532 	}
2533 }
2534 
2535 static void
2536 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed)
2537 {
2538 	if (tcp_bblogging_on(bbr->rc_tp)) {
2539 		union tcp_log_stackspecific log;
2540 
2541 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2542 		log.u_bbr.flex1 = line;
2543 		log.u_bbr.flex2 = bbr->bbr_timer_src;
2544 		log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2545 		log.u_bbr.flex4 = bbr->rc_in_persist;
2546 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2547 		log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2548 		log.u_bbr.flex8 = hpts_removed;
2549 		log.u_bbr.pkts_out = bbr->rc_pacer_started;
2550 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2551 		    &bbr->rc_inp->inp_socket->so_rcv,
2552 		    &bbr->rc_inp->inp_socket->so_snd,
2553 		    BBR_LOG_TIMERCANC, 0,
2554 		    0, &log, false, &bbr->rc_tv);
2555 	}
2556 }
2557 
2558 static void
2559 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta)
2560 {
2561 	if (tcp_bblogging_on(bbr->rc_tp)) {
2562 		union tcp_log_stackspecific log;
2563 
2564 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2565 		log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio;
2566 		log.u_bbr.flex2 = (peer_delta >> 32);
2567 		log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff);
2568 		log.u_bbr.flex4 = (delta >> 32);
2569 		log.u_bbr.flex5 = (delta & 0x00000000ffffffff);
2570 		log.u_bbr.flex7 = bbr->rc_ts_clock_set;
2571 		log.u_bbr.flex8 = bbr->rc_ts_cant_be_used;
2572 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2573 		    &bbr->rc_inp->inp_socket->so_rcv,
2574 		    &bbr->rc_inp->inp_socket->so_snd,
2575 		    BBR_LOG_TSTMP_VAL, 0,
2576 		    0, &log, false, &bbr->rc_tv);
2577 	}
2578 }
2579 
2580 static void
2581 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)
2582 {
2583 	if (tcp_bblogging_on(bbr->rc_tp)) {
2584 		union tcp_log_stackspecific log;
2585 
2586 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2587 		log.u_bbr.flex1 = tsosz;
2588 		log.u_bbr.flex2 = tls;
2589 		log.u_bbr.flex3 = tcp_min_hptsi_time;
2590 		log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min;
2591 		log.u_bbr.flex5 = old_val;
2592 		log.u_bbr.flex6 = maxseg;
2593 		log.u_bbr.flex7 = bbr->rc_no_pacing;
2594 		log.u_bbr.flex7 <<= 1;
2595 		log.u_bbr.flex7 |= bbr->rc_past_init_win;
2596 		if (hdwr)
2597 			log.u_bbr.flex8 = 0x80 | bbr->rc_use_google;
2598 		else
2599 			log.u_bbr.flex8 = bbr->rc_use_google;
2600 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2601 		    &bbr->rc_inp->inp_socket->so_rcv,
2602 		    &bbr->rc_inp->inp_socket->so_snd,
2603 		    BBR_LOG_BBRTSO, 0,
2604 		    0, &log, false, &bbr->rc_tv);
2605 	}
2606 }
2607 
2608 static void
2609 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm,
2610 		      uint32_t flags, uint32_t line)
2611 {
2612 	if (tcp_bblogging_on(bbr->rc_tp)) {
2613 		union tcp_log_stackspecific log;
2614 
2615 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2616 		log.u_bbr.flex1 = line;
2617 		log.u_bbr.flex2 = rsm->r_start;
2618 		log.u_bbr.flex3 = rsm->r_end;
2619 		log.u_bbr.flex4 = rsm->r_delivered;
2620 		log.u_bbr.flex5 = rsm->r_rtr_cnt;
2621 		log.u_bbr.flex6 = rsm->r_dupack;
2622 		log.u_bbr.flex7 = rsm->r_tim_lastsent[0];
2623 		log.u_bbr.flex8 = rsm->r_flags;
2624 		/* Hijack the pkts_out fids */
2625 		log.u_bbr.applimited = flags;
2626 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2627 		    &bbr->rc_inp->inp_socket->so_rcv,
2628 		    &bbr->rc_inp->inp_socket->so_snd,
2629 		    BBR_RSM_CLEARED, 0,
2630 		    0, &log, false, &bbr->rc_tv);
2631 	}
2632 }
2633 
2634 static void
2635 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts,
2636     uint32_t flex3, uint32_t flex2, uint32_t flex5,
2637     uint32_t flex6, uint32_t pkts_out, int flex7,
2638     uint32_t flex4, uint32_t flex1)
2639 {
2640 
2641 	if (tcp_bblogging_on(bbr->rc_tp)) {
2642 		union tcp_log_stackspecific log;
2643 
2644 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2645 		log.u_bbr.flex1 = flex1;
2646 		log.u_bbr.flex2 = flex2;
2647 		log.u_bbr.flex3 = flex3;
2648 		log.u_bbr.flex4 = flex4;
2649 		log.u_bbr.flex5 = flex5;
2650 		log.u_bbr.flex6 = flex6;
2651 		log.u_bbr.flex7 = flex7;
2652 		/* Hijack the pkts_out fids */
2653 		log.u_bbr.pkts_out = pkts_out;
2654 		log.u_bbr.flex8 = flex8;
2655 		if (bbr->rc_ack_was_delayed)
2656 			log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay;
2657 		else
2658 			log.u_bbr.epoch = 0;
2659 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2660 		    &bbr->rc_inp->inp_socket->so_rcv,
2661 		    &bbr->rc_inp->inp_socket->so_snd,
2662 		    BBR_LOG_BBRUPD, 0,
2663 		    flex2, &log, false, &bbr->rc_tv);
2664 	}
2665 }
2666 
2667 static void
2668 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason,
2669 	uint32_t newbw, uint32_t obw, uint32_t diff,
2670 	uint32_t tim)
2671 {
2672 	if (/*bbr_verbose_logging && */tcp_bblogging_on(bbr->rc_tp)) {
2673 		union tcp_log_stackspecific log;
2674 
2675 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2676 		log.u_bbr.flex1 = reason;
2677 		log.u_bbr.flex2 = newbw;
2678 		log.u_bbr.flex3 = obw;
2679 		log.u_bbr.flex4 = diff;
2680 		log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost;
2681 		log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del;
2682 		log.u_bbr.flex7 = bbr->rc_lt_is_sampling;
2683 		log.u_bbr.pkts_out = tim;
2684 		log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw;
2685 		if (bbr->rc_lt_use_bw == 0)
2686 			log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
2687 		else
2688 			log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
2689 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2690 		    &bbr->rc_inp->inp_socket->so_rcv,
2691 		    &bbr->rc_inp->inp_socket->so_snd,
2692 		    BBR_LOG_BWSAMP, 0,
2693 		    0, &log, false, &bbr->rc_tv);
2694 	}
2695 }
2696 
2697 static inline void
2698 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line)
2699 {
2700 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2701 		union tcp_log_stackspecific log;
2702 
2703 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2704 		log.u_bbr.flex1 = line;
2705 		log.u_bbr.flex2 = tick;
2706 		log.u_bbr.flex3 = tp->t_maxunacktime;
2707 		log.u_bbr.flex4 = tp->t_acktime;
2708 		log.u_bbr.flex8 = event;
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_PROGRESS, 0,
2713 		    0, &log, false, &bbr->rc_tv);
2714 	}
2715 }
2716 
2717 static void
2718 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp,
2719 			 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts,
2720 			 int error)
2721 {
2722 	if (tcp_bblogging_on(bbr->rc_tp)) {
2723 		union tcp_log_stackspecific log;
2724 
2725 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2726 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2727 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2728 		log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2729 		log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2730 		log.u_bbr.bw_inuse = rate;
2731 		log.u_bbr.flex5 = line;
2732 		log.u_bbr.flex6 = error;
2733 		log.u_bbr.flex8 = bbr->skip_gain;
2734 		log.u_bbr.flex8 <<= 1;
2735 		log.u_bbr.flex8 |= bbr->gain_is_limited;
2736 		log.u_bbr.flex8 <<= 1;
2737 		log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing;
2738 		log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
2739 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2740 		    &bbr->rc_inp->inp_socket->so_rcv,
2741 		    &bbr->rc_inp->inp_socket->so_snd,
2742 		    BBR_LOG_HDWR_PACE, 0,
2743 		    0, &log, false, &bbr->rc_tv);
2744 	}
2745 }
2746 
2747 static void
2748 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)
2749 {
2750 	if (tcp_bblogging_on(bbr->rc_tp)) {
2751 		union tcp_log_stackspecific log;
2752 
2753 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2754 		log.u_bbr.flex1 = slot;
2755 		log.u_bbr.flex2 = del_by;
2756 		log.u_bbr.flex3 = prev_delay;
2757 		log.u_bbr.flex4 = line;
2758 		log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val;
2759 		log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay;
2760 		log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags);
2761 		log.u_bbr.flex8 = bbr->rc_in_persist;
2762 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2763 		    &bbr->rc_inp->inp_socket->so_rcv,
2764 		    &bbr->rc_inp->inp_socket->so_snd,
2765 		    BBR_LOG_BBRSND, 0,
2766 		    len, &log, false, &bbr->rc_tv);
2767 	}
2768 }
2769 
2770 static void
2771 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)
2772 {
2773 	if (tcp_bblogging_on(bbr->rc_tp)) {
2774 		union tcp_log_stackspecific log;
2775 
2776 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2777 		log.u_bbr.flex1 = bbr->r_ctl.rc_delivered;
2778 		log.u_bbr.flex2 = 0;
2779 		log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt;
2780 		log.u_bbr.flex4 = end;
2781 		log.u_bbr.flex5 = seq;
2782 		log.u_bbr.flex6 = t;
2783 		log.u_bbr.flex7 = match;
2784 		log.u_bbr.flex8 = flags;
2785 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2786 		    &bbr->rc_inp->inp_socket->so_rcv,
2787 		    &bbr->rc_inp->inp_socket->so_snd,
2788 		    BBR_LOG_BBRRTT, 0,
2789 		    0, &log, false, &bbr->rc_tv);
2790 	}
2791 }
2792 
2793 static void
2794 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method)
2795 {
2796 	if (tcp_bblogging_on(bbr->rc_tp)) {
2797 		union tcp_log_stackspecific log;
2798 
2799 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2800 		log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2801 		log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
2802 		log.u_bbr.flex3 = bbr->r_ctl.gain_epoch;
2803 		log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2804 		log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs;
2805 		log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight;
2806 		log.u_bbr.flex7 = 0;
2807 		log.u_bbr.flex8 = entry_method;
2808 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2809 		    &bbr->rc_inp->inp_socket->so_rcv,
2810 		    &bbr->rc_inp->inp_socket->so_snd,
2811 		    BBR_LOG_EXIT_GAIN, 0,
2812 		    0, &log, false, &bbr->rc_tv);
2813 	}
2814 }
2815 
2816 static void
2817 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired)
2818 {
2819 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2820 		union tcp_log_stackspecific log;
2821 
2822 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2823 		/* R-HU */
2824 		log.u_bbr.flex1 = 0;
2825 		log.u_bbr.flex2 = 0;
2826 		log.u_bbr.flex3 = 0;
2827 		log.u_bbr.flex4 = 0;
2828 		log.u_bbr.flex7 = 0;
2829 		log.u_bbr.flex8 = settings_desired;
2830 
2831 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2832 		    &bbr->rc_inp->inp_socket->so_rcv,
2833 		    &bbr->rc_inp->inp_socket->so_snd,
2834 		    BBR_LOG_SETTINGS_CHG, 0,
2835 		    0, &log, false, &bbr->rc_tv);
2836 	}
2837 }
2838 
2839 /*
2840  * Returns the bw from the our filter.
2841  */
2842 static inline uint64_t
2843 bbr_get_full_bw(struct tcp_bbr *bbr)
2844 {
2845 	uint64_t bw;
2846 
2847 	bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2848 
2849 	return (bw);
2850 }
2851 
2852 static inline void
2853 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2854 {
2855 	uint64_t calclr;
2856 	uint32_t lost, del;
2857 
2858 	if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch)
2859 		lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch;
2860 	else
2861 		lost = 0;
2862 	del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del;
2863 	if (lost == 0)  {
2864 		calclr = 0;
2865 	} else if (del) {
2866 		calclr = lost;
2867 		calclr *= (uint64_t)1000;
2868 		calclr /= (uint64_t)del;
2869 	} else {
2870 		/* Nothing delivered? 100.0% loss */
2871 		calclr = 1000;
2872 	}
2873 	bbr->r_ctl.rc_pkt_epoch_loss_rate =  (uint32_t)calclr;
2874 	if (IN_RECOVERY(bbr->rc_tp->t_flags))
2875 		bbr->r_ctl.recovery_lr += (uint32_t)calclr;
2876 	bbr->r_ctl.rc_pkt_epoch++;
2877 	if (bbr->rc_no_pacing &&
2878 	    (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) {
2879 		bbr->rc_no_pacing = 0;
2880 		tcp_bbr_tso_size_check(bbr, cts);
2881 	}
2882 	bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time);
2883 	bbr->r_ctl.rc_pkt_epoch_time = cts;
2884 	/* What was our loss rate */
2885 	bbr_log_pkt_epoch(bbr, cts, line, lost, del);
2886 	bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered;
2887 	bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost;
2888 }
2889 
2890 static inline void
2891 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2892 {
2893 	uint32_t epoch_time;
2894 
2895 	/* Tick the RTT clock */
2896 	bbr->r_ctl.rc_rtt_epoch++;
2897 	epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start;
2898 	bbr_log_time_epoch(bbr, cts, line, epoch_time);
2899 	bbr->r_ctl.rc_rcv_epoch_start = cts;
2900 }
2901 
2902 static inline void
2903 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked)
2904 {
2905 	if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) {
2906 		bbr->rc_is_pkt_epoch_now = 1;
2907 	}
2908 }
2909 
2910 /*
2911  * Returns the bw from either the b/w filter
2912  * or from the lt_bw (if the connection is being
2913  * policed).
2914  */
2915 static inline uint64_t
2916 __bbr_get_bw(struct tcp_bbr *bbr)
2917 {
2918 	uint64_t bw, min_bw;
2919 	uint64_t rtt;
2920 	int gm_measure_cnt = 1;
2921 
2922 	/*
2923 	 * For startup we make, like google, a
2924 	 * minimum b/w. This is generated from the
2925 	 * IW and the rttProp. We do fall back to srtt
2926 	 * if for some reason (initial handshake) we don't
2927 	 * have a rttProp. We, in the worst case, fall back
2928 	 * to the configured min_bw (rc_initial_hptsi_bw).
2929 	 */
2930 	if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
2931 		/* Attempt first to use rttProp */
2932 		rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2933 		if (rtt && (rtt < 0xffffffff)) {
2934 measure:
2935 			min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2936 				((uint64_t)1000000);
2937 			min_bw /= rtt;
2938 			if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2939 				min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2940 			}
2941 
2942 		} else if (bbr->rc_tp->t_srtt != 0) {
2943 			/* No rttProp, use srtt? */
2944 			rtt = bbr_get_rtt(bbr, BBR_SRTT);
2945 			goto measure;
2946 		} else {
2947 			min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2948 		}
2949 	} else
2950 		min_bw = 0;
2951 
2952 	if ((bbr->rc_past_init_win == 0) &&
2953 	    (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp)))
2954 		bbr->rc_past_init_win = 1;
2955 	if ((bbr->rc_use_google)  && (bbr->r_ctl.r_measurement_count >= 1))
2956 		gm_measure_cnt = 0;
2957 	if (gm_measure_cnt &&
2958 	    ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) ||
2959 	     (bbr->rc_past_init_win == 0))) {
2960 		/* For google we use our guess rate until we get 1 measurement */
2961 
2962 use_initial_window:
2963 		rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2964 		if (rtt && (rtt < 0xffffffff)) {
2965 			/*
2966 			 * We have an RTT measurement. Use that in
2967 			 * combination with our initial window to calculate
2968 			 * a b/w.
2969 			 */
2970 			bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2971 				((uint64_t)1000000);
2972 			bw /= rtt;
2973 			if (bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2974 				bw = bbr->r_ctl.rc_initial_hptsi_bw;
2975 			}
2976 		} else {
2977 			/* Drop back to the 40 and punt to a default */
2978 			bw = bbr->r_ctl.rc_initial_hptsi_bw;
2979 		}
2980 		if (bw < 1)
2981 			/* Probably should panic */
2982 			bw = 1;
2983 		if (bw > min_bw)
2984 			return (bw);
2985 		else
2986 			return (min_bw);
2987 	}
2988 	if (bbr->rc_lt_use_bw)
2989 		bw = bbr->r_ctl.rc_lt_bw;
2990 	else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0))
2991 		bw = bbr->r_ctl.red_bw;
2992 	else
2993 		bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2994 	if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) {
2995 		/*
2996 		 * Enforce user set rate limit, keep in mind that
2997 		 * t_peakrate_thr is in B/s already
2998 		 */
2999 		bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw);
3000 	}
3001 	if (bw == 0) {
3002 		/* We should not be at 0, go to the initial window then  */
3003 		goto use_initial_window;
3004 	}
3005 	if (bw < 1)
3006 		/* Probably should panic */
3007 		bw = 1;
3008 	if (bw < min_bw)
3009 		bw = min_bw;
3010 	return (bw);
3011 }
3012 
3013 static inline uint64_t
3014 bbr_get_bw(struct tcp_bbr *bbr)
3015 {
3016 	uint64_t bw;
3017 
3018 	bw = __bbr_get_bw(bbr);
3019 	return (bw);
3020 }
3021 
3022 static inline void
3023 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts)
3024 {
3025 	bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch;
3026 	bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time;
3027 	bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3028 	bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3029 }
3030 
3031 static inline void
3032 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts)
3033 {
3034 	bbr->rc_lt_is_sampling = 0;
3035 	bbr->rc_lt_use_bw = 0;
3036 	bbr->r_ctl.rc_lt_bw = 0;
3037 	bbr_reset_lt_bw_interval(bbr, cts);
3038 }
3039 
3040 static inline void
3041 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin)
3042 {
3043 	uint64_t diff;
3044 
3045 	/* Do we have a previous sample? */
3046 	if (bbr->r_ctl.rc_lt_bw) {
3047 		/* Get the diff in bytes per second */
3048 		if (bbr->r_ctl.rc_lt_bw > bw)
3049 			diff = bbr->r_ctl.rc_lt_bw - bw;
3050 		else
3051 			diff = bw - bbr->r_ctl.rc_lt_bw;
3052 		if ((diff <= bbr_lt_bw_diff) ||
3053 		    (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) {
3054 			/* Consider us policed */
3055 			uint32_t saved_bw;
3056 
3057 			saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw;
3058 			bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2;	/* average of two */
3059 			bbr->rc_lt_use_bw = 1;
3060 			bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
3061 			/*
3062 			 * Use pkt based epoch for measuring length of
3063 			 * policer up
3064 			 */
3065 			bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch;
3066 			/*
3067 			 * reason 4 is we need to start consider being
3068 			 * policed
3069 			 */
3070 			bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin);
3071 			return;
3072 		}
3073 	}
3074 	bbr->r_ctl.rc_lt_bw = bw;
3075 	bbr_reset_lt_bw_interval(bbr, cts);
3076 	bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin);
3077 }
3078 
3079 static void
3080 bbr_randomize_extra_state_time(struct tcp_bbr *bbr)
3081 {
3082 	uint32_t ran, deduct;
3083 
3084 	ran = arc4random_uniform(bbr_rand_ot);
3085 	if (ran) {
3086 		deduct = bbr->r_ctl.rc_level_state_extra / ran;
3087 		bbr->r_ctl.rc_level_state_extra -= deduct;
3088 	}
3089 }
3090 /*
3091  * Return randomly the starting state
3092  * to use in probebw.
3093  */
3094 static uint8_t
3095 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts)
3096 {
3097 	uint32_t ran;
3098 	uint8_t ret_val;
3099 
3100 	/* Initialize the offset to 0 */
3101 	bbr->r_ctl.rc_exta_time_gd = 0;
3102 	bbr->rc_hit_state_1 = 0;
3103 	bbr->r_ctl.rc_level_state_extra = 0;
3104 	ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1));
3105 	/*
3106 	 * The math works funny here :) the return value is used to set the
3107 	 * substate and then the state change is called which increments by
3108 	 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when
3109 	 * we fully enter the state. Note that the (8 - 1 - ran) assures that
3110 	 * we return 1 - 7, so we dont return 0 and end up starting in
3111 	 * state 1 (DRAIN).
3112 	 */
3113 	ret_val = BBR_SUBSTATE_COUNT - 1 - ran;
3114 	/* Set an epoch */
3115 	if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP))
3116 		bbr_set_epoch(bbr, cts, __LINE__);
3117 
3118 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
3119 	return (ret_val);
3120 }
3121 
3122 static void
3123 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected)
3124 {
3125 	uint32_t diff, d_time;
3126 	uint64_t del_time, bw, lost, delivered;
3127 
3128 	if (bbr->r_use_policer == 0)
3129 		return;
3130 	if (bbr->rc_lt_use_bw) {
3131 		/* We are using lt bw do we stop yet? */
3132 		diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
3133 		if (diff > bbr_lt_bw_max_rtts) {
3134 			/* Reset it all */
3135 reset_all:
3136 			bbr_reset_lt_bw_sampling(bbr, cts);
3137 			if (bbr->rc_filled_pipe) {
3138 				bbr_set_epoch(bbr, cts, __LINE__);
3139 				bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
3140 				bbr_substate_change(bbr, cts, __LINE__, 0);
3141 				bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
3142 				bbr_log_type_statechange(bbr, cts, __LINE__);
3143 			} else {
3144 				/*
3145 				 * This should not happen really
3146 				 * unless we remove the startup/drain
3147 				 * restrictions above.
3148 				 */
3149 				bbr->rc_bbr_state = BBR_STATE_STARTUP;
3150 				bbr_set_epoch(bbr, cts, __LINE__);
3151 				bbr->r_ctl.rc_bbr_state_time = cts;
3152 				bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
3153 				bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
3154 				bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
3155 				bbr_set_state_target(bbr, __LINE__);
3156 				bbr_log_type_statechange(bbr, cts, __LINE__);
3157 			}
3158 			/* reason 0 is to stop using lt-bw */
3159 			bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0);
3160 			return;
3161 		}
3162 		if (bbr_lt_intvl_fp == 0) {
3163 			/* Not doing false-positive detection */
3164 			return;
3165 		}
3166 		/* False positive detection */
3167 		if (diff == bbr_lt_intvl_fp) {
3168 			/* At bbr_lt_intvl_fp we record the lost */
3169 			bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3170 			bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3171 		} else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) {
3172 			/* Now is our loss rate still high? */
3173 			lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3174 			delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3175 			if ((delivered == 0) ||
3176 			    (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) {
3177 				/* No still below our threshold */
3178 				bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0);
3179 			} else {
3180 				/* Yikes its still high, it must be a false positive */
3181 				bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0);
3182 				goto reset_all;
3183 			}
3184 		}
3185 		return;
3186 	}
3187 	/*
3188 	 * Wait for the first loss before sampling, to let the policer
3189 	 * exhaust its tokens and estimate the steady-state rate allowed by
3190 	 * the policer. Starting samples earlier includes bursts that
3191 	 * over-estimate the bw.
3192 	 */
3193 	if (bbr->rc_lt_is_sampling == 0) {
3194 		/* reason 1 is to begin doing the sampling  */
3195 		if (loss_detected == 0)
3196 			return;
3197 		bbr_reset_lt_bw_interval(bbr, cts);
3198 		bbr->rc_lt_is_sampling = 1;
3199 		bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0);
3200 		return;
3201 	}
3202 	/* Now how long were we delivering long term last> */
3203 	if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time))
3204 		d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time;
3205 	else
3206 		d_time = 0;
3207 
3208 	/* To avoid underestimates, reset sampling if we run out of data. */
3209 	if (bbr->r_ctl.r_app_limited_until) {
3210 		/* Can not measure in app-limited state */
3211 		bbr_reset_lt_bw_sampling(bbr, cts);
3212 		/* reason 2 is to reset sampling due to app limits  */
3213 		bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time);
3214 		return;
3215 	}
3216 	diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
3217 	if (diff < bbr_lt_intvl_min_rtts) {
3218 		/*
3219 		 * need more samples (we don't
3220 		 * start on a round like linux so
3221 		 * we need 1 more).
3222 		 */
3223 		/* 6 is not_enough time or no-loss */
3224 		bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3225 		return;
3226 	}
3227 	if (diff > (4 * bbr_lt_intvl_min_rtts)) {
3228 		/*
3229 		 * For now if we wait too long, reset all sampling. We need
3230 		 * to do some research here, its possible that we should
3231 		 * base this on how much loss as occurred.. something like
3232 		 * if its under 10% (or some thresh) reset all otherwise
3233 		 * don't.  Thats for phase II I guess.
3234 		 */
3235 		bbr_reset_lt_bw_sampling(bbr, cts);
3236  		/* reason 3 is to reset sampling due too long of sampling */
3237 		bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3238 		return;
3239 	}
3240 	/*
3241 	 * End sampling interval when a packet is lost, so we estimate the
3242 	 * policer tokens were exhausted. Stopping the sampling before the
3243 	 * tokens are exhausted under-estimates the policed rate.
3244 	 */
3245 	if (loss_detected == 0) {
3246 		/* 6 is not_enough time or no-loss */
3247 		bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3248 		return;
3249 	}
3250 	/* Calculate packets lost and delivered in sampling interval. */
3251 	lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3252 	delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3253 	if ((delivered == 0) ||
3254 	    (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) {
3255 		bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time);
3256 		return;
3257 	}
3258 	if (d_time < 1000) {
3259 		/* Not enough time. wait */
3260 		/* 6 is not_enough time or no-loss */
3261 		bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3262 		return;
3263 	}
3264 	if (d_time >= (0xffffffff / USECS_IN_MSEC)) {
3265 		/* Too long */
3266 		bbr_reset_lt_bw_sampling(bbr, cts);
3267  		/* reason 3 is to reset sampling due too long of sampling */
3268 		bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3269 		return;
3270 	}
3271 	del_time = d_time;
3272 	bw = delivered;
3273 	bw *= (uint64_t)USECS_IN_SECOND;
3274 	bw /= del_time;
3275 	bbr_lt_bw_samp_done(bbr, bw, cts, d_time);
3276 }
3277 
3278 /*
3279  * Allocate a sendmap from our zone.
3280  */
3281 static struct bbr_sendmap *
3282 bbr_alloc(struct tcp_bbr *bbr)
3283 {
3284 	struct bbr_sendmap *rsm;
3285 
3286 	BBR_STAT_INC(bbr_to_alloc);
3287 	rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO));
3288 	if (rsm) {
3289 		bbr->r_ctl.rc_num_maps_alloced++;
3290 		return (rsm);
3291 	}
3292 	if (bbr->r_ctl.rc_free_cnt) {
3293 		BBR_STAT_INC(bbr_to_alloc_emerg);
3294 		rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
3295 		TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
3296 		bbr->r_ctl.rc_free_cnt--;
3297 		return (rsm);
3298 	}
3299 	BBR_STAT_INC(bbr_to_alloc_failed);
3300 	return (NULL);
3301 }
3302 
3303 static struct bbr_sendmap *
3304 bbr_alloc_full_limit(struct tcp_bbr *bbr)
3305 {
3306 	if ((V_tcp_map_entries_limit > 0) &&
3307 	    (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3308 		BBR_STAT_INC(bbr_alloc_limited);
3309 		if (!bbr->alloc_limit_reported) {
3310 			bbr->alloc_limit_reported = 1;
3311 			BBR_STAT_INC(bbr_alloc_limited_conns);
3312 		}
3313 		return (NULL);
3314 	}
3315 	return (bbr_alloc(bbr));
3316 }
3317 
3318 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3319 static struct bbr_sendmap *
3320 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type)
3321 {
3322 	struct bbr_sendmap *rsm;
3323 
3324 	if (limit_type) {
3325 		/* currently there is only one limit type */
3326 		if (V_tcp_map_split_limit > 0 &&
3327 		    bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
3328 			BBR_STAT_INC(bbr_split_limited);
3329 			if (!bbr->alloc_limit_reported) {
3330 				bbr->alloc_limit_reported = 1;
3331 				BBR_STAT_INC(bbr_alloc_limited_conns);
3332 			}
3333 			return (NULL);
3334 		}
3335 	}
3336 
3337 	/* allocate and mark in the limit type, if set */
3338 	rsm = bbr_alloc(bbr);
3339 	if (rsm != NULL && limit_type) {
3340 		rsm->r_limit_type = limit_type;
3341 		bbr->r_ctl.rc_num_split_allocs++;
3342 	}
3343 	return (rsm);
3344 }
3345 
3346 static void
3347 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
3348 {
3349 	if (rsm->r_limit_type) {
3350 		/* currently there is only one limit type */
3351 		bbr->r_ctl.rc_num_split_allocs--;
3352 	}
3353 	if (rsm->r_is_smallmap)
3354 		bbr->r_ctl.rc_num_small_maps_alloced--;
3355 	if (bbr->r_ctl.rc_tlp_send == rsm)
3356 		bbr->r_ctl.rc_tlp_send = NULL;
3357 	if (bbr->r_ctl.rc_resend == rsm) {
3358 		bbr->r_ctl.rc_resend = NULL;
3359 	}
3360 	if (bbr->r_ctl.rc_next == rsm)
3361 		bbr->r_ctl.rc_next = NULL;
3362 	if (bbr->r_ctl.rc_sacklast == rsm)
3363 		bbr->r_ctl.rc_sacklast = NULL;
3364 	if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
3365 		memset(rsm, 0, sizeof(struct bbr_sendmap));
3366 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
3367 		rsm->r_limit_type = 0;
3368 		bbr->r_ctl.rc_free_cnt++;
3369 		return;
3370 	}
3371 	bbr->r_ctl.rc_num_maps_alloced--;
3372 	uma_zfree(bbr_zone, rsm);
3373 }
3374 
3375 /*
3376  * Returns the BDP.
3377  */
3378 static uint64_t
3379 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) {
3380 	/*
3381 	 * Calculate the bytes in flight needed given the bw (in bytes per
3382 	 * second) and the specifyed rtt in useconds. We need to put out the
3383 	 * returned value per RTT to match that rate. Gain will normally
3384 	 * raise it up from there.
3385 	 *
3386 	 * This should not overflow as long as the bandwidth is below 1
3387 	 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller
3388 	 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30).
3389 	 */
3390 	uint64_t usec_per_sec;
3391 
3392 	usec_per_sec = USECS_IN_SECOND;
3393 	return ((rtt * bw) / usec_per_sec);
3394 }
3395 
3396 /*
3397  * Return the initial cwnd.
3398  */
3399 static uint32_t
3400 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp)
3401 {
3402 	uint32_t i_cwnd;
3403 
3404 	if (bbr->rc_init_win) {
3405 		i_cwnd = bbr->rc_init_win * tp->t_maxseg;
3406 	} else if (V_tcp_initcwnd_segments)
3407 		i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg),
3408 		    max(2 * tp->t_maxseg, 14600));
3409 	else if (V_tcp_do_rfc3390)
3410 		i_cwnd = min(4 * tp->t_maxseg,
3411 		    max(2 * tp->t_maxseg, 4380));
3412 	else {
3413 		/* Per RFC5681 Section 3.1 */
3414 		if (tp->t_maxseg > 2190)
3415 			i_cwnd = 2 * tp->t_maxseg;
3416 		else if (tp->t_maxseg > 1095)
3417 			i_cwnd = 3 * tp->t_maxseg;
3418 		else
3419 			i_cwnd = 4 * tp->t_maxseg;
3420 	}
3421 	return (i_cwnd);
3422 }
3423 
3424 /*
3425  * Given a specified gain, return the target
3426  * cwnd based on that gain.
3427  */
3428 static uint32_t
3429 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw)
3430 {
3431 	uint64_t bdp, rtt;
3432 	uint32_t cwnd;
3433 
3434 	if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) ||
3435 	    (bbr_get_full_bw(bbr) == 0)) {
3436 		/* No measurements yet */
3437 		return (bbr_initial_cwnd(bbr, bbr->rc_tp));
3438 	}
3439 	/*
3440 	 * Get bytes per RTT needed (rttProp is normally in
3441 	 * bbr_cwndtarget_rtt_touse)
3442 	 */
3443 	rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse);
3444 	/* Get the bdp from the two values */
3445 	bdp = bbr_get_bw_delay_prod(rtt, bw);
3446 	/* Now apply the gain */
3447 	cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT));
3448 
3449 	return (cwnd);
3450 }
3451 
3452 static uint32_t
3453 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain)
3454 {
3455 	uint32_t cwnd, mss;
3456 
3457 	mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
3458 	/* Get the base cwnd with gain rounded to a mss */
3459 	cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss);
3460 	/*
3461 	 * Add in N (2 default since we do not have a
3462 	 * fq layer to trap packets in) quanta's per the I-D
3463 	 * section 4.2.3.2 quanta adjust.
3464 	 */
3465 	cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs);
3466 	if (bbr->rc_use_google) {
3467 		if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3468 		   (bbr_state_val(bbr) == BBR_SUB_GAIN)) {
3469 			/*
3470 			 * The linux implementation adds
3471 			 * an extra 2 x mss in gain cycle which
3472 			 * is documented no-where except in the code.
3473 			 * so we add more for Neal undocumented feature
3474 			 */
3475 			cwnd += 2 * mss;
3476 		}
3477  		if ((cwnd / mss) & 0x1) {
3478 			/* Round up for odd num mss */
3479 			cwnd += mss;
3480 		}
3481 	}
3482 	/* Are we below the min cwnd? */
3483 	if (cwnd < get_min_cwnd(bbr))
3484 		return (get_min_cwnd(bbr));
3485 	return (cwnd);
3486 }
3487 
3488 static uint16_t
3489 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain)
3490 {
3491 	if (gain < 1)
3492 		gain = 1;
3493 	return (gain);
3494 }
3495 
3496 static uint32_t
3497 bbr_get_header_oh(struct tcp_bbr *bbr)
3498 {
3499 	int seg_oh;
3500 
3501 	seg_oh = 0;
3502 	if (bbr->r_ctl.rc_inc_tcp_oh) {
3503 		/* Do we include TCP overhead? */
3504 		seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr));
3505 	}
3506 	if (bbr->r_ctl.rc_inc_ip_oh) {
3507 		/* Do we include IP overhead? */
3508 #ifdef INET6
3509 		if (bbr->r_is_v6) {
3510 			seg_oh += sizeof(struct ip6_hdr);
3511 		} else
3512 #endif
3513 		{
3514 
3515 #ifdef INET
3516 			seg_oh += sizeof(struct ip);
3517 #endif
3518 		}
3519 	}
3520 	if (bbr->r_ctl.rc_inc_enet_oh) {
3521 		/* Do we include the ethernet overhead?  */
3522 		seg_oh += sizeof(struct ether_header);
3523 	}
3524 	return(seg_oh);
3525 }
3526 
3527 static uint32_t
3528 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw)
3529 {
3530 	uint64_t divor, res, tim;
3531 
3532 	if (useconds_time == 0)
3533 		return (0);
3534 	gain = bbr_gain_adjust(bbr, gain);
3535 	divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT;
3536 	tim = useconds_time;
3537 	res = (tim * bw * gain) / divor;
3538 	if (res == 0)
3539 		res = 1;
3540 	return ((uint32_t)res);
3541 }
3542 
3543 /*
3544  * Given a gain and a length return the delay in useconds that
3545  * should be used to evenly space out packets
3546  * on the connection (based on the gain factor).
3547  */
3548 static uint32_t
3549 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog)
3550 {
3551 	uint64_t bw, lentim, res;
3552 	uint32_t usecs, srtt, over = 0;
3553 	uint32_t seg_oh, num_segs, maxseg;
3554 
3555 	if (len == 0)
3556 		return (0);
3557 
3558 	maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
3559 	num_segs = (len + maxseg - 1) / maxseg;
3560 	if (bbr->rc_use_google == 0) {
3561 		seg_oh = bbr_get_header_oh(bbr);
3562 		len += (num_segs * seg_oh);
3563 	}
3564 	gain = bbr_gain_adjust(bbr, gain);
3565 	bw = bbr_get_bw(bbr);
3566 	if (bbr->rc_use_google) {
3567 		uint64_t cbw;
3568 
3569 		/*
3570 		 * Reduce the b/w by the google discount
3571 		 * factor 10 = 1%.
3572 		 */
3573 		cbw = bw *  (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount);
3574 		cbw /= (uint64_t)1000;
3575 		/* We don't apply a discount if it results in 0 */
3576 		if (cbw > 0)
3577 			bw = cbw;
3578 	}
3579 	lentim = ((uint64_t)len *
3580 		  (uint64_t)USECS_IN_SECOND *
3581 		  (uint64_t)BBR_UNIT);
3582 	res = lentim / ((uint64_t)gain * bw);
3583 	if (res == 0)
3584 		res = 1;
3585 	usecs = (uint32_t)res;
3586 	srtt = bbr_get_rtt(bbr, BBR_SRTT);
3587 	if (bbr_hptsi_max_mul && bbr_hptsi_max_div &&
3588 	    (bbr->rc_use_google == 0) &&
3589 	    (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) {
3590 		/*
3591 		 * We cannot let the delay be more than 1/2 the srtt time.
3592 		 * Otherwise we cannot pace out or send properly.
3593 		 */
3594 		over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div;
3595 		BBR_STAT_INC(bbr_hpts_min_time);
3596 	}
3597 	if (!nolog)
3598 		bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1);
3599 	return (usecs);
3600 }
3601 
3602 static void
3603 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack,
3604 		 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses)
3605 {
3606 	uint64_t bw;
3607 	uint32_t cwnd, target_cwnd, saved_bytes, maxseg;
3608 	int32_t meth;
3609 
3610 	INP_WLOCK_ASSERT(tptoinpcb(tp));
3611 
3612 #ifdef STATS
3613 	if ((tp->t_flags & TF_GPUTINPROG) &&
3614 	    SEQ_GEQ(th->th_ack, tp->gput_ack)) {
3615 		/*
3616 		 * Strech acks and compressed acks will cause this to
3617 		 * oscillate but we are doing it the same way as the main
3618 		 * stack so it will be compariable (though possibly not
3619 		 * ideal).
3620 		 */
3621 		int32_t cgput;
3622 		int64_t gput, time_stamp;
3623 
3624 		gput = (int64_t) (th->th_ack - tp->gput_seq) * 8;
3625 		time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000));
3626 		cgput = gput / time_stamp;
3627 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3628 					 cgput);
3629 		if (tp->t_stats_gput_prev > 0)
3630 			stats_voi_update_abs_s32(tp->t_stats,
3631 						 VOI_TCP_GPUT_ND,
3632 						 ((gput - tp->t_stats_gput_prev) * 100) /
3633 						 tp->t_stats_gput_prev);
3634 		tp->t_flags &= ~TF_GPUTINPROG;
3635 		tp->t_stats_gput_prev = cgput;
3636 	}
3637 #endif
3638 	if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3639 	    ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
3640 		/* We don't change anything in probe-rtt */
3641 		return;
3642 	}
3643 	maxseg = tp->t_maxseg - bbr->rc_last_options;
3644 	saved_bytes = bytes_this_ack;
3645 	bytes_this_ack += sack_changed;
3646 	if (bytes_this_ack > prev_acked) {
3647 		bytes_this_ack -= prev_acked;
3648 		/*
3649 		 * A byte ack'd gives us a full mss
3650 		 * to be like linux i.e. they count packets.
3651 		 */
3652 		if ((bytes_this_ack < maxseg) && bbr->rc_use_google)
3653 			bytes_this_ack = maxseg;
3654 	} else {
3655 		/* Unlikely */
3656 		bytes_this_ack = 0;
3657 	}
3658 	cwnd = tp->snd_cwnd;
3659 	bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3660 	if (bw)
3661 		target_cwnd = bbr_get_target_cwnd(bbr,
3662 						  bw,
3663 						  (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain);
3664 	else
3665 		target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp);
3666 	if (IN_RECOVERY(tp->t_flags) &&
3667 	    (bbr->bbr_prev_in_rec == 0)) {
3668 		/*
3669 		 * We are entering recovery and
3670 		 * thus packet conservation.
3671 		 */
3672 		bbr->pkt_conservation = 1;
3673 		bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime;
3674 		cwnd = ctf_flight_size(tp,
3675 				       (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3676 			bytes_this_ack;
3677 	}
3678 	if (IN_RECOVERY(tp->t_flags)) {
3679 		uint32_t flight;
3680 
3681 		bbr->bbr_prev_in_rec = 1;
3682 		if (cwnd > losses) {
3683 			cwnd -= losses;
3684 			if (cwnd < maxseg)
3685 				cwnd = maxseg;
3686 		} else
3687 			cwnd = maxseg;
3688 		flight = ctf_flight_size(tp,
3689 					 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3690 		bbr_log_type_cwndupd(bbr, flight, 0,
3691 				     losses, 10, 0, 0, line);
3692 		if (bbr->pkt_conservation) {
3693 			uint32_t time_in;
3694 
3695 			if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start))
3696 				time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start;
3697 			else
3698 				time_in = 0;
3699 
3700 			if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
3701 				/* Clear packet conservation after an rttProp */
3702 				bbr->pkt_conservation = 0;
3703 			} else {
3704 				if ((flight + bytes_this_ack) > cwnd)
3705 					cwnd = flight + bytes_this_ack;
3706 				if (cwnd < get_min_cwnd(bbr))
3707 					cwnd = get_min_cwnd(bbr);
3708 				tp->snd_cwnd = cwnd;
3709 				bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed,
3710 						     prev_acked, 1, target_cwnd, th->th_ack, line);
3711 				return;
3712 			}
3713 		}
3714 	} else
3715 		bbr->bbr_prev_in_rec = 0;
3716 	if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) {
3717 		bbr->r_ctl.restrict_growth--;
3718 		if (bytes_this_ack > maxseg)
3719 			bytes_this_ack = maxseg;
3720 	}
3721 	if (bbr->rc_filled_pipe) {
3722 		/*
3723 		 * Here we have exited startup and filled the pipe. We will
3724 		 * thus allow the cwnd to shrink to the target. We hit here
3725 		 * mostly.
3726 		 */
3727 		uint32_t s_cwnd;
3728 
3729 		meth = 2;
3730 		s_cwnd = min((cwnd + bytes_this_ack), target_cwnd);
3731 		if (s_cwnd > cwnd)
3732 			cwnd = s_cwnd;
3733 		else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing)
3734 			cwnd = s_cwnd;
3735 	} else {
3736 		/*
3737 		 * Here we are still in startup, we increase cwnd by what
3738 		 * has been acked.
3739 		 */
3740 		if ((cwnd < target_cwnd) ||
3741 		    (bbr->rc_past_init_win == 0)) {
3742 			meth = 3;
3743 			cwnd += bytes_this_ack;
3744 		} else {
3745 			/*
3746 			 * Method 4 means we are at target so no gain in
3747 			 * startup and past the initial window.
3748 			 */
3749 			meth = 4;
3750 		}
3751 	}
3752 	tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr));
3753 	bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line);
3754 }
3755 
3756 static void
3757 tcp_bbr_partialack(struct tcpcb *tp)
3758 {
3759 	struct tcp_bbr *bbr;
3760 
3761 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3762 	INP_WLOCK_ASSERT(tptoinpcb(tp));
3763 	if (ctf_flight_size(tp,
3764 		(bbr->r_ctl.rc_sacked  + bbr->r_ctl.rc_lost_bytes)) <=
3765 	    tp->snd_cwnd) {
3766 		bbr->r_wanted_output = 1;
3767 	}
3768 }
3769 
3770 static void
3771 bbr_post_recovery(struct tcpcb *tp)
3772 {
3773 	struct tcp_bbr *bbr;
3774 	uint32_t  flight;
3775 
3776 	INP_WLOCK_ASSERT(tptoinpcb(tp));
3777 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3778 	/*
3779 	 * Here we just exit recovery.
3780 	 */
3781 	EXIT_RECOVERY(tp->t_flags);
3782 	/* Lock in our b/w reduction for the specified number of pkt-epochs */
3783 	bbr->r_recovery_bw = 0;
3784 	tp->snd_recover = tp->snd_una;
3785 	tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3786 	bbr->pkt_conservation = 0;
3787 	if (bbr->rc_use_google == 0) {
3788 		/*
3789 		 * For non-google mode lets
3790 		 * go ahead and make sure we clear
3791 		 * the recovery state so if we
3792 		 * bounce back in to recovery we
3793 		 * will do PC.
3794 		 */
3795 		bbr->bbr_prev_in_rec = 0;
3796 	}
3797 	bbr_log_type_exit_rec(bbr);
3798 	if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3799 		tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
3800 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__);
3801 	} else {
3802 		/* For probe-rtt case lets fix up its saved_cwnd */
3803 		if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) {
3804 			bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent;
3805 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__);
3806 		}
3807 	}
3808 	flight = ctf_flight_size(tp,
3809 		     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3810 	if ((bbr->rc_use_google == 0) &&
3811 	    bbr_do_red) {
3812 		uint64_t val, lr2use;
3813 		uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd;
3814 		uint32_t *cwnd_p;
3815 
3816 		if (bbr_get_rtt(bbr, BBR_SRTT)) {
3817 			val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000);
3818 			val /= bbr_get_rtt(bbr, BBR_SRTT);
3819 			ratio = (uint32_t)val;
3820 		} else
3821 			ratio = 1000;
3822 
3823 		bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div,
3824 				     bbr->r_ctl.recovery_lr, 21,
3825 				     ratio,
3826 				     bbr->r_ctl.rc_red_cwnd_pe,
3827 				     __LINE__);
3828 		if ((ratio < bbr_do_red) || (bbr_do_red == 0))
3829 			goto done;
3830 		if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3831 		     bbr_prtt_slam_cwnd) ||
3832 		    (bbr_sub_drain_slam_cwnd &&
3833 		     (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3834 		     bbr->rc_hit_state_1 &&
3835 		     (bbr_state_val(bbr) == BBR_SUB_DRAIN)) ||
3836 		    ((bbr->rc_bbr_state == BBR_STATE_DRAIN) &&
3837 		     bbr_slam_cwnd_in_main_drain)) {
3838 			/*
3839 			 * Here we must poke at the saved cwnd
3840 			 * as well as the cwnd.
3841 			 */
3842 			cwnd = bbr->r_ctl.rc_saved_cwnd;
3843 			cwnd_p = &bbr->r_ctl.rc_saved_cwnd;
3844 		} else {
3845  			cwnd = tp->snd_cwnd;
3846 			cwnd_p = &tp->snd_cwnd;
3847 		}
3848 		maxseg = tp->t_maxseg - bbr->rc_last_options;
3849 		/* Add the overall lr with the recovery lr */
3850 		if (bbr->r_ctl.rc_lost == 0)
3851 			lr2use = 0;
3852 		else if (bbr->r_ctl.rc_delivered == 0)
3853 			lr2use = 1000;
3854 		else {
3855 			lr2use = bbr->r_ctl.rc_lost * 1000;
3856 			lr2use /= bbr->r_ctl.rc_delivered;
3857 		}
3858 		lr2use += bbr->r_ctl.recovery_lr;
3859 		acks_inflight = (flight / (maxseg * 2));
3860 		if (bbr_red_scale) {
3861 			lr2use *= bbr_get_rtt(bbr, BBR_SRTT);
3862 			lr2use /= bbr_red_scale;
3863 			if ((bbr_red_growth_restrict) &&
3864 			    ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1))
3865 			    bbr->r_ctl.restrict_growth += acks_inflight;
3866 		}
3867 		if (lr2use) {
3868 			val = (uint64_t)cwnd * lr2use;
3869 			val /= 1000;
3870 			if (cwnd > val)
3871 				newcwnd = roundup((cwnd - val), maxseg);
3872 			else
3873 				newcwnd = maxseg;
3874 		} else {
3875 			val = (uint64_t)cwnd * (uint64_t)bbr_red_mul;
3876 			val /= (uint64_t)bbr_red_div;
3877 			newcwnd = roundup((uint32_t)val, maxseg);
3878 		}
3879 		/* with standard delayed acks how many acks can I expect? */
3880 		if (bbr_drop_limit == 0) {
3881 			/*
3882 			 * Anticpate how much we will
3883 			 * raise the cwnd based on the acks.
3884 			 */
3885 			if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) {
3886 				/* We do enforce the min (with the acks) */
3887 				newcwnd = (get_min_cwnd(bbr) - acks_inflight);
3888 			}
3889 		} else {
3890 			/*
3891 			 * A strict drop limit of N is inplace
3892 			 */
3893 			if (newcwnd < (bbr_drop_limit * maxseg)) {
3894 				newcwnd = bbr_drop_limit * maxseg;
3895 			}
3896 		}
3897 		/* For the next N acks do we restrict the growth */
3898 		*cwnd_p = newcwnd;
3899 		if (tp->snd_cwnd > newcwnd)
3900 			tp->snd_cwnd = newcwnd;
3901 		bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22,
3902 				     (uint32_t)lr2use,
3903 				     bbr_get_rtt(bbr, BBR_SRTT), __LINE__);
3904 		bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch;
3905 	}
3906 done:
3907 	bbr->r_ctl.recovery_lr = 0;
3908 	if (flight <= tp->snd_cwnd) {
3909 		bbr->r_wanted_output = 1;
3910 	}
3911 	tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3912 }
3913 
3914 static void
3915 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts)
3916 {
3917 	bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3918 	/* Limit the drop in b/w to 1/2 our current filter. */
3919 	if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate)
3920 		bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate;
3921 	if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2))
3922 		bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2;
3923 	tcp_bbr_tso_size_check(bbr, cts);
3924 }
3925 
3926 static void
3927 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm)
3928 {
3929 	struct tcp_bbr *bbr;
3930 
3931 	INP_WLOCK_ASSERT(tptoinpcb(tp));
3932 #ifdef STATS
3933 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
3934 #endif
3935 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3936 	switch (type) {
3937 	case CC_NDUPACK:
3938 		if (!IN_RECOVERY(tp->t_flags)) {
3939 			tp->snd_recover = tp->snd_max;
3940 			/* Start a new epoch */
3941 			bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
3942 			if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) {
3943 				/*
3944 				 * Move forward the lt epoch
3945 				 * so it won't count the truncated
3946 				 * epoch.
3947 				 */
3948 				bbr->r_ctl.rc_lt_epoch++;
3949 			}
3950 			if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
3951 				/*
3952 				 * Just like the policer detection code
3953 				 * if we are in startup we must push
3954 				 * forward the last startup epoch
3955 				 * to hide the truncated PE.
3956 				 */
3957 				bbr->r_ctl.rc_bbr_last_startup_epoch++;
3958 			}
3959 			bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd;
3960 			ENTER_RECOVERY(tp->t_flags);
3961 			bbr->rc_tlp_rtx_out = 0;
3962 			bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate;
3963 			tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3964 			if (tcp_in_hpts(bbr->rc_inp) &&
3965 			    ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) {
3966 				/*
3967 				 * When we enter recovery, we need to restart
3968 				 * any timers. This may mean we gain an agg
3969 				 * early, which will be made up for at the last
3970 				 * rxt out.
3971 				 */
3972 				bbr->rc_timer_first = 1;
3973 				bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
3974 			}
3975 			/*
3976 			 * Calculate a new cwnd based on to the current
3977 			 * delivery rate with no gain. We get the bdp
3978 			 * without gaining it up like we normally would and
3979 			 * we use the last cur_del_rate.
3980 			 */
3981 			if ((bbr->rc_use_google == 0) &&
3982 			    (bbr->r_ctl.bbr_rttprobe_gain_val ||
3983 			     (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) {
3984 				tp->snd_cwnd = ctf_flight_size(tp,
3985 					           (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3986 					(tp->t_maxseg - bbr->rc_last_options);
3987 				if (tp->snd_cwnd < get_min_cwnd(bbr)) {
3988 					/* We always gate to min cwnd */
3989 					tp->snd_cwnd = get_min_cwnd(bbr);
3990 				}
3991 				bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__);
3992 			}
3993 			bbr_log_type_enter_rec(bbr, rsm->r_start);
3994 		}
3995 		break;
3996 	case CC_RTO_ERR:
3997 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
3998 		/* RTO was unnecessary, so reset everything. */
3999 		bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime);
4000 		if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
4001 			tp->snd_cwnd = tp->snd_cwnd_prev;
4002 			tp->snd_ssthresh = tp->snd_ssthresh_prev;
4003 			tp->snd_recover = tp->snd_recover_prev;
4004 			tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
4005 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__);
4006 		}
4007 		tp->t_badrxtwin = 0;
4008 		break;
4009 	}
4010 }
4011 
4012 /*
4013  * Indicate whether this ack should be delayed.  We can delay the ack if
4014  * following conditions are met:
4015  *	- There is no delayed ack timer in progress.
4016  *	- Our last ack wasn't a 0-sized window. We never want to delay
4017  *	  the ack that opens up a 0-sized window.
4018  *	- LRO wasn't used for this segment. We make sure by checking that the
4019  *	  segment size is not larger than the MSS.
4020  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
4021  *	  connection.
4022  *	- The data being acked is less than a full segment (a stretch ack
4023  *        of more than a segment we should ack.
4024  *      - nsegs is 1 (if its more than that we received more than 1 ack).
4025  */
4026 #define DELAY_ACK(tp, bbr, nsegs)				\
4027 	(((tp->t_flags & TF_RXWIN0SENT) == 0) &&		\
4028 	 ((tp->t_flags & TF_DELACK) == 0) && 		 	\
4029 	 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) &&	\
4030 	 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4031 
4032 /*
4033  * Return the lowest RSM in the map of
4034  * packets still in flight that is not acked.
4035  * This should normally find on the first one
4036  * since we remove packets from the send
4037  * map after they are marked ACKED.
4038  */
4039 static struct bbr_sendmap *
4040 bbr_find_lowest_rsm(struct tcp_bbr *bbr)
4041 {
4042 	struct bbr_sendmap *rsm;
4043 
4044 	/*
4045 	 * Walk the time-order transmitted list looking for an rsm that is
4046 	 * not acked. This will be the one that was sent the longest time
4047 	 * ago that is still outstanding.
4048 	 */
4049 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) {
4050 		if (rsm->r_flags & BBR_ACKED) {
4051 			continue;
4052 		}
4053 		goto finish;
4054 	}
4055 finish:
4056 	return (rsm);
4057 }
4058 
4059 static struct bbr_sendmap *
4060 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
4061 {
4062 	struct bbr_sendmap *prsm;
4063 
4064 	/*
4065 	 * Walk the sequence order list backward until we hit and arrive at
4066 	 * the highest seq not acked. In theory when this is called it
4067 	 * should be the last segment (which it was not).
4068 	 */
4069 	prsm = rsm;
4070 	TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4071 		if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) {
4072 			continue;
4073 		}
4074 		return (prsm);
4075 	}
4076 	return (NULL);
4077 }
4078 
4079 /*
4080  * Returns to the caller the number of microseconds that
4081  * the packet can be outstanding before we think we
4082  * should have had an ack returned.
4083  */
4084 static uint32_t
4085 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm)
4086 {
4087 	/*
4088 	 * lro is the flag we use to determine if we have seen reordering.
4089 	 * If it gets set we have seen reordering. The reorder logic either
4090 	 * works in one of two ways:
4091 	 *
4092 	 * If reorder-fade is configured, then we track the last time we saw
4093 	 * re-ordering occur. If we reach the point where enough time as
4094 	 * passed we no longer consider reordering has occuring.
4095 	 *
4096 	 * Or if reorder-face is 0, then once we see reordering we consider
4097 	 * the connection to alway be subject to reordering and just set lro
4098 	 * to 1.
4099 	 *
4100 	 * In the end if lro is non-zero we add the extra time for
4101 	 * reordering in.
4102 	 */
4103 	int32_t lro;
4104 	uint32_t thresh, t_rxtcur;
4105 
4106 	if (srtt == 0)
4107 		srtt = 1;
4108 	if (bbr->r_ctl.rc_reorder_ts) {
4109 		if (bbr->r_ctl.rc_reorder_fade) {
4110 			if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) {
4111 				lro = cts - bbr->r_ctl.rc_reorder_ts;
4112 				if (lro == 0) {
4113 					/*
4114 					 * No time as passed since the last
4115 					 * reorder, mark it as reordering.
4116 					 */
4117 					lro = 1;
4118 				}
4119 			} else {
4120 				/* Negative time? */
4121 				lro = 0;
4122 			}
4123 			if (lro > bbr->r_ctl.rc_reorder_fade) {
4124 				/* Turn off reordering seen too */
4125 				bbr->r_ctl.rc_reorder_ts = 0;
4126 				lro = 0;
4127 			}
4128 		} else {
4129 			/* Reodering does not fade */
4130 			lro = 1;
4131 		}
4132 	} else {
4133 		lro = 0;
4134 	}
4135 	thresh = srtt + bbr->r_ctl.rc_pkt_delay;
4136 	if (lro) {
4137 		/* It must be set, if not you get 1/4 rtt */
4138 		if (bbr->r_ctl.rc_reorder_shift)
4139 			thresh += (srtt >> bbr->r_ctl.rc_reorder_shift);
4140 		else
4141 			thresh += (srtt >> 2);
4142 	} else {
4143 		thresh += 1000;
4144 	}
4145 	/* We don't let the rack timeout be above a RTO */
4146 	if ((bbr->rc_tp)->t_srtt == 0)
4147 		t_rxtcur = BBR_INITIAL_RTO;
4148 	else
4149 		t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
4150 	if (thresh > t_rxtcur) {
4151 		thresh = t_rxtcur;
4152 	}
4153 	/* And we don't want it above the RTO max either */
4154 	if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4155 		thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4156 	}
4157 	bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK);
4158 	return (thresh);
4159 }
4160 
4161 /*
4162  * Return to the caller the amount of time in mico-seconds
4163  * that should be used for the TLP timer from the last
4164  * send time of this packet.
4165  */
4166 static uint32_t
4167 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
4168     struct bbr_sendmap *rsm, uint32_t srtt,
4169     uint32_t cts)
4170 {
4171 	uint32_t thresh, len, maxseg, t_rxtcur;
4172 	struct bbr_sendmap *prsm;
4173 
4174 	if (srtt == 0)
4175 		srtt = 1;
4176 	if (bbr->rc_tlp_threshold)
4177 		thresh = srtt + (srtt / bbr->rc_tlp_threshold);
4178 	else
4179 		thresh = (srtt * 2);
4180 	maxseg = tp->t_maxseg - bbr->rc_last_options;
4181 	/* Get the previous sent packet, if any  */
4182 	len = rsm->r_end - rsm->r_start;
4183 
4184 	/* 2.1 behavior */
4185 	prsm = TAILQ_PREV(rsm, bbr_head, r_tnext);
4186 	if (prsm && (len <= maxseg)) {
4187 		/*
4188 		 * Two packets outstanding, thresh should be (2*srtt) +
4189 		 * possible inter-packet delay (if any).
4190 		 */
4191 		uint32_t inter_gap = 0;
4192 		int idx, nidx;
4193 
4194 		idx = rsm->r_rtr_cnt - 1;
4195 		nidx = prsm->r_rtr_cnt - 1;
4196 		if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4197 			/* Yes it was sent later (or at the same time) */
4198 			inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4199 		}
4200 		thresh += inter_gap;
4201 	} else if (len <= maxseg) {
4202 		/*
4203 		 * Possibly compensate for delayed-ack.
4204 		 */
4205 		uint32_t alt_thresh;
4206 
4207 		alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time;
4208 		if (alt_thresh > thresh)
4209 			thresh = alt_thresh;
4210 	}
4211 	/* Not above the current  RTO */
4212 	if (tp->t_srtt == 0)
4213 		t_rxtcur = BBR_INITIAL_RTO;
4214 	else
4215 		t_rxtcur = TICKS_2_USEC(tp->t_rxtcur);
4216 
4217 	bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP);
4218 	/* Not above an RTO */
4219 	if (thresh > t_rxtcur) {
4220 		thresh = t_rxtcur;
4221 	}
4222 	/* Not above a RTO max */
4223 	if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4224 		thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4225 	}
4226 	/* And now apply the user TLP min */
4227 	if (thresh < bbr_tlp_min) {
4228 		thresh = bbr_tlp_min;
4229 	}
4230 	return (thresh);
4231 }
4232 
4233 /*
4234  * Return one of three RTTs to use (in microseconds).
4235  */
4236 static __inline uint32_t
4237 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type)
4238 {
4239 	uint32_t f_rtt;
4240 	uint32_t srtt;
4241 
4242 	f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
4243 	if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) {
4244 		/* We have no rtt at all */
4245 		if (bbr->rc_tp->t_srtt == 0)
4246 			f_rtt = BBR_INITIAL_RTO;
4247 		else
4248 			f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4249 		/*
4250 		 * Since we don't know how good the rtt is apply a
4251 		 * delayed-ack min
4252 		 */
4253 		if (f_rtt < bbr_delayed_ack_time) {
4254 			f_rtt = bbr_delayed_ack_time;
4255 		}
4256 	}
4257 	/* Take the filter version or last measured pkt-rtt */
4258 	if (rtt_type == BBR_RTT_PROP) {
4259 		srtt = f_rtt;
4260 	} else if (rtt_type == BBR_RTT_PKTRTT) {
4261 		if (bbr->r_ctl.rc_pkt_epoch_rtt) {
4262 			srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
4263 		} else {
4264 			/* No pkt rtt yet */
4265 			srtt = f_rtt;
4266 		}
4267 	} else if (rtt_type == BBR_RTT_RACK) {
4268 		srtt = bbr->r_ctl.rc_last_rtt;
4269 		/* We need to add in any internal delay for our timer */
4270 		if (bbr->rc_ack_was_delayed)
4271 			srtt += bbr->r_ctl.rc_ack_hdwr_delay;
4272 	} else if (rtt_type == BBR_SRTT) {
4273 		srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4274 	} else {
4275 		/* TSNH */
4276 		srtt = f_rtt;
4277 #ifdef BBR_INVARIANTS
4278 		panic("Unknown rtt request type %d", rtt_type);
4279 #endif
4280 	}
4281 	return (srtt);
4282 }
4283 
4284 static int
4285 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts)
4286 {
4287 	uint32_t thresh;
4288 
4289 	thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK),
4290 				      cts, rsm);
4291 	if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) {
4292 		/* It is lost (past time) */
4293 		return (1);
4294 	}
4295 	return (0);
4296 }
4297 
4298 /*
4299  * Return a sendmap if we need to retransmit something.
4300  */
4301 static struct bbr_sendmap *
4302 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4303 {
4304 	/*
4305 	 * Check to see that we don't need to fall into recovery. We will
4306 	 * need to do so if our oldest transmit is past the time we should
4307 	 * have had an ack.
4308 	 */
4309 
4310 	struct bbr_sendmap *rsm;
4311 	int32_t idx;
4312 
4313 	if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) {
4314 		/* Nothing outstanding that we know of */
4315 		return (NULL);
4316 	}
4317 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
4318 	if (rsm == NULL) {
4319 		/* Nothing in the transmit map */
4320 		return (NULL);
4321 	}
4322 	if (tp->t_flags & TF_SENTFIN) {
4323 		/* Fin restricted, don't find anything once a fin is sent */
4324 		return (NULL);
4325 	}
4326 	if (rsm->r_flags & BBR_ACKED) {
4327 		/*
4328 		 * Ok the first one is acked (this really should not happen
4329 		 * since we remove the from the tmap once they are acked)
4330 		 */
4331 		rsm = bbr_find_lowest_rsm(bbr);
4332 		if (rsm == NULL)
4333 			return (NULL);
4334 	}
4335 	idx = rsm->r_rtr_cnt - 1;
4336 	if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) {
4337 		/* Send timestamp is the same or less? can't be ready */
4338 		return (NULL);
4339 	}
4340 	/* Get our RTT time */
4341 	if (bbr_is_lost(bbr, rsm, cts) &&
4342 	    ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
4343 	     (rsm->r_flags & BBR_SACK_PASSED))) {
4344 		if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4345 			rsm->r_flags |= BBR_MARKED_LOST;
4346 			bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4347 			bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4348 		}
4349 		bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm);
4350 #ifdef BBR_INVARIANTS
4351 		if ((rsm->r_end - rsm->r_start) == 0)
4352 			panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm);
4353 #endif
4354 		return (rsm);
4355 	}
4356 	return (NULL);
4357 }
4358 
4359 /*
4360  * RACK Timer, here we simply do logging and house keeping.
4361  * the normal bbr_output_wtime() function will call the
4362  * appropriate thing to check if we need to do a RACK retransmit.
4363  * We return 1, saying don't proceed with bbr_output_wtime only
4364  * when all timers have been stopped (destroyed PCB?).
4365  */
4366 static int
4367 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4368 {
4369 	/*
4370 	 * This timer simply provides an internal trigger to send out data.
4371 	 * The check_recovery_mode call will see if there are needed
4372 	 * retransmissions, if so we will enter fast-recovery. The output
4373 	 * call may or may not do the same thing depending on sysctl
4374 	 * settings.
4375 	 */
4376 	uint32_t lost;
4377 
4378 	if (bbr->rc_all_timers_stopped) {
4379 		return (1);
4380 	}
4381 	if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4382 		/* Its not time yet */
4383 		return (0);
4384 	}
4385 	BBR_STAT_INC(bbr_to_tot);
4386 	lost = bbr->r_ctl.rc_lost;
4387 	if (bbr->r_state && (bbr->r_state != tp->t_state))
4388 		bbr_set_state(tp, bbr, 0);
4389 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK);
4390 	if (bbr->r_ctl.rc_resend == NULL) {
4391 		/* Lets do the check here */
4392 		bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
4393 	}
4394 	if (bbr_policer_call_from_rack_to)
4395 		bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4396 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
4397 	return (0);
4398 }
4399 
4400 static __inline void
4401 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start)
4402 {
4403 	int idx;
4404 
4405 	nrsm->r_start = start;
4406 	nrsm->r_end = rsm->r_end;
4407 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
4408 	nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed;
4409 	nrsm->r_flags = rsm->r_flags;
4410 	/* We don't transfer forward the SYN flag */
4411 	nrsm->r_flags &= ~BBR_HAS_SYN;
4412 	/* We move forward the FIN flag, not that this should happen */
4413 	rsm->r_flags &= ~BBR_HAS_FIN;
4414 	nrsm->r_dupack = rsm->r_dupack;
4415 	nrsm->r_rtr_bytes = 0;
4416 	nrsm->r_is_gain = rsm->r_is_gain;
4417 	nrsm->r_is_drain = rsm->r_is_drain;
4418 	nrsm->r_delivered = rsm->r_delivered;
4419 	nrsm->r_ts_valid = rsm->r_ts_valid;
4420 	nrsm->r_del_ack_ts = rsm->r_del_ack_ts;
4421 	nrsm->r_del_time = rsm->r_del_time;
4422 	nrsm->r_app_limited = rsm->r_app_limited;
4423 	nrsm->r_first_sent_time = rsm->r_first_sent_time;
4424 	nrsm->r_flight_at_send = rsm->r_flight_at_send;
4425 	/* We split a piece the lower section looses any just_ret flag. */
4426 	nrsm->r_bbr_state = rsm->r_bbr_state;
4427 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
4428 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
4429 	}
4430 	rsm->r_end = nrsm->r_start;
4431 	idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
4432 	idx /= 8;
4433 	/* Check if we got too small */
4434 	if ((rsm->r_is_smallmap == 0) &&
4435 	    ((rsm->r_end - rsm->r_start) <= idx)) {
4436 		bbr->r_ctl.rc_num_small_maps_alloced++;
4437 		rsm->r_is_smallmap = 1;
4438 	}
4439 	/* Check the new one as well */
4440 	if ((nrsm->r_end - nrsm->r_start) <= idx) {
4441 		bbr->r_ctl.rc_num_small_maps_alloced++;
4442 		nrsm->r_is_smallmap = 1;
4443 	}
4444 }
4445 
4446 static int
4447 bbr_sack_mergable(struct bbr_sendmap *at,
4448 		  uint32_t start, uint32_t end)
4449 {
4450 	/*
4451 	 * Given a sack block defined by
4452 	 * start and end, and a current position
4453 	 * at. Return 1 if either side of at
4454 	 * would show that the block is mergable
4455 	 * to that side. A block to be mergable
4456 	 * must have overlap with the start/end
4457 	 * and be in the SACK'd state.
4458 	 */
4459 	struct bbr_sendmap *l_rsm;
4460 	struct bbr_sendmap *r_rsm;
4461 
4462 	/* first get the either side blocks */
4463 	l_rsm = TAILQ_PREV(at, bbr_head, r_next);
4464 	r_rsm = TAILQ_NEXT(at, r_next);
4465 	if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) {
4466 		/* Potentially mergeable */
4467 		if ((l_rsm->r_end == start) ||
4468 		    (SEQ_LT(start, l_rsm->r_end) &&
4469 		     SEQ_GT(end, l_rsm->r_end))) {
4470 			    /*
4471 			     * map blk   |------|
4472 			     * sack blk         |------|
4473 			     * <or>
4474 			     * map blk   |------|
4475 			     * sack blk      |------|
4476 			     */
4477 			    return (1);
4478 		    }
4479 	}
4480 	if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) {
4481 		/* Potentially mergeable */
4482 		if ((r_rsm->r_start == end) ||
4483 		    (SEQ_LT(start, r_rsm->r_start) &&
4484 		     SEQ_GT(end, r_rsm->r_start))) {
4485 			/*
4486 			 * map blk          |---------|
4487 			 * sack blk    |----|
4488 			 * <or>
4489 			 * map blk          |---------|
4490 			 * sack blk    |-------|
4491 			 */
4492 			return (1);
4493 		}
4494 	}
4495 	return (0);
4496 }
4497 
4498 static struct bbr_sendmap *
4499 bbr_merge_rsm(struct tcp_bbr *bbr,
4500 	      struct bbr_sendmap *l_rsm,
4501 	      struct bbr_sendmap *r_rsm)
4502 {
4503 	/*
4504 	 * We are merging two ack'd RSM's,
4505 	 * the l_rsm is on the left (lower seq
4506 	 * values) and the r_rsm is on the right
4507 	 * (higher seq value). The simplest way
4508 	 * to merge these is to move the right
4509 	 * one into the left. I don't think there
4510 	 * is any reason we need to try to find
4511 	 * the oldest (or last oldest retransmitted).
4512 	 */
4513 	l_rsm->r_end = r_rsm->r_end;
4514 	if (l_rsm->r_dupack < r_rsm->r_dupack)
4515 		l_rsm->r_dupack = r_rsm->r_dupack;
4516 	if (r_rsm->r_rtr_bytes)
4517 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
4518 	if (r_rsm->r_in_tmap) {
4519 		/* This really should not happen */
4520 		TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext);
4521 	}
4522 	if (r_rsm->r_app_limited)
4523 		l_rsm->r_app_limited = r_rsm->r_app_limited;
4524 	/* Now the flags */
4525 	if (r_rsm->r_flags & BBR_HAS_FIN)
4526 		l_rsm->r_flags |= BBR_HAS_FIN;
4527 	if (r_rsm->r_flags & BBR_TLP)
4528 		l_rsm->r_flags |= BBR_TLP;
4529 	if (r_rsm->r_flags & BBR_RWND_COLLAPSED)
4530 		l_rsm->r_flags |= BBR_RWND_COLLAPSED;
4531 	if (r_rsm->r_flags & BBR_MARKED_LOST) {
4532 		/* This really should not happen */
4533 		bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start;
4534 	}
4535 	TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next);
4536 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
4537 		/* Transfer the split limit to the map we free */
4538 		r_rsm->r_limit_type = l_rsm->r_limit_type;
4539 		l_rsm->r_limit_type = 0;
4540 	}
4541 	bbr_free(bbr, r_rsm);
4542 	return(l_rsm);
4543 }
4544 
4545 /*
4546  * TLP Timer, here we simply setup what segment we want to
4547  * have the TLP expire on, the normal bbr_output_wtime() will then
4548  * send it out.
4549  *
4550  * We return 1, saying don't proceed with bbr_output_wtime only
4551  * when all timers have been stopped (destroyed PCB?).
4552  */
4553 static int
4554 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4555 {
4556 	/*
4557 	 * Tail Loss Probe.
4558 	 */
4559 	struct bbr_sendmap *rsm = NULL;
4560 	struct socket *so;
4561 	uint32_t amm;
4562 	uint32_t out, avail;
4563 	uint32_t maxseg;
4564 	int collapsed_win = 0;
4565 
4566 	if (bbr->rc_all_timers_stopped) {
4567 		return (1);
4568 	}
4569 	if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4570 		/* Its not time yet */
4571 		return (0);
4572 	}
4573 	if (ctf_progress_timeout_check(tp, true)) {
4574 		bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4575 		return (-ETIMEDOUT);	/* tcp_drop() */
4576 	}
4577 	/* Did we somehow get into persists? */
4578 	if (bbr->rc_in_persist) {
4579 		return (0);
4580 	}
4581 	if (bbr->r_state && (bbr->r_state != tp->t_state))
4582 		bbr_set_state(tp, bbr, 0);
4583 	BBR_STAT_INC(bbr_tlp_tot);
4584 	maxseg = tp->t_maxseg - bbr->rc_last_options;
4585 	/*
4586 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
4587 	 * need to figure out how to force a full MSS segment out.
4588 	 */
4589 	so = tptosocket(tp);
4590 	avail = sbavail(&so->so_snd);
4591 	out = ctf_outstanding(tp);
4592 	if (out > tp->snd_wnd) {
4593 		/* special case, we need a retransmission */
4594 		collapsed_win = 1;
4595 		goto need_retran;
4596 	}
4597 	if (avail > out) {
4598 		/* New data is available */
4599 		amm = avail - out;
4600 		if (amm > maxseg) {
4601 			amm = maxseg;
4602 		} else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) {
4603 			/* not enough to fill a MTU and no-delay is off */
4604 			goto need_retran;
4605 		}
4606 		/* Set the send-new override */
4607 		if ((out + amm) <= tp->snd_wnd) {
4608 			bbr->rc_tlp_new_data = 1;
4609 		} else {
4610 			goto need_retran;
4611 		}
4612 		bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4613 		bbr->r_ctl.rc_last_tlp_seq = tp->snd_max;
4614 		bbr->r_ctl.rc_tlp_send = NULL;
4615 		/* cap any slots */
4616 		BBR_STAT_INC(bbr_tlp_newdata);
4617 		goto send;
4618 	}
4619 need_retran:
4620 	/*
4621 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
4622 	 * optionally the first un-acked segment.
4623 	 */
4624 	if (collapsed_win == 0) {
4625 		rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
4626 		if (rsm && (BBR_ACKED | BBR_HAS_FIN)) {
4627 			rsm = bbr_find_high_nonack(bbr, rsm);
4628 		}
4629 		if (rsm == NULL) {
4630 			goto restore;
4631 		}
4632 	} else {
4633 		/*
4634 		 * We must find the last segment
4635 		 * that was acceptable by the client.
4636 		 */
4637 		TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4638 			if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) {
4639 				/* Found one */
4640 				break;
4641 			}
4642 		}
4643 		if (rsm == NULL) {
4644 			/* None? if so send the first */
4645 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4646 			if (rsm == NULL)
4647 				goto restore;
4648 		}
4649 	}
4650 	if ((rsm->r_end - rsm->r_start) > maxseg) {
4651 		/*
4652 		 * We need to split this the last segment in two.
4653 		 */
4654 		struct bbr_sendmap *nrsm;
4655 
4656 		nrsm = bbr_alloc_full_limit(bbr);
4657 		if (nrsm == NULL) {
4658 			/*
4659 			 * We can't get memory to split, we can either just
4660 			 * not split it. Or retransmit the whole piece, lets
4661 			 * do the large send (BTLP :-) ).
4662 			 */
4663 			goto go_for_it;
4664 		}
4665 		bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg));
4666 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
4667 		if (rsm->r_in_tmap) {
4668 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
4669 			nrsm->r_in_tmap = 1;
4670 		}
4671 		rsm->r_flags &= (~BBR_HAS_FIN);
4672 		rsm = nrsm;
4673 	}
4674 go_for_it:
4675 	bbr->r_ctl.rc_tlp_send = rsm;
4676 	bbr->rc_tlp_rtx_out = 1;
4677 	if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) {
4678 		bbr->r_ctl.rc_tlp_seg_send_cnt++;
4679 		tp->t_rxtshift++;
4680 	} else {
4681 		bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
4682 		bbr->r_ctl.rc_tlp_seg_send_cnt = 1;
4683 	}
4684 send:
4685 	if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
4686 		/*
4687 		 * Can't [re]/transmit a segment we have retransmitted the
4688 		 * max times. We need the retransmit timer to take over.
4689 		 */
4690 restore:
4691 		bbr->rc_tlp_new_data = 0;
4692 		bbr->r_ctl.rc_tlp_send = NULL;
4693 		if (rsm)
4694 			rsm->r_flags &= ~BBR_TLP;
4695 		BBR_STAT_INC(bbr_tlp_retran_fail);
4696 		return (0);
4697 	} else if (rsm) {
4698 		rsm->r_flags |= BBR_TLP;
4699 	}
4700 	if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) &&
4701 	    (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) {
4702 		/*
4703 		 * We have retransmitted to many times for TLP. Switch to
4704 		 * the regular RTO timer
4705 		 */
4706 		goto restore;
4707 	}
4708 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP);
4709 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
4710 	return (0);
4711 }
4712 
4713 /*
4714  * Delayed ack Timer, here we simply need to setup the
4715  * ACK_NOW flag and remove the DELACK flag. From there
4716  * the output routine will send the ack out.
4717  *
4718  * We only return 1, saying don't proceed, if all timers
4719  * are stopped (destroyed PCB?).
4720  */
4721 static int
4722 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4723 {
4724 	if (bbr->rc_all_timers_stopped) {
4725 		return (1);
4726 	}
4727 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK);
4728 	tp->t_flags &= ~TF_DELACK;
4729 	tp->t_flags |= TF_ACKNOW;
4730 	KMOD_TCPSTAT_INC(tcps_delack);
4731 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4732 	return (0);
4733 }
4734 
4735 /*
4736  * Here we send a KEEP-ALIVE like probe to the
4737  * peer, we do not send data.
4738  *
4739  * We only return 1, saying don't proceed, if all timers
4740  * are stopped (destroyed PCB?).
4741  */
4742 static int
4743 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4744 {
4745 	struct tcptemp *t_template;
4746 	int32_t retval = 1;
4747 
4748 	if (bbr->rc_all_timers_stopped) {
4749 		return (1);
4750 	}
4751 	if (bbr->rc_in_persist == 0)
4752 		return (0);
4753 
4754 	/*
4755 	 * Persistence timer into zero window. Force a byte to be output, if
4756 	 * possible.
4757 	 */
4758 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST);
4759 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
4760 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
4761 	/*
4762 	 * Have we exceeded the user specified progress time?
4763 	 */
4764 	if (ctf_progress_timeout_check(tp, true)) {
4765 		bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4766 		return (-ETIMEDOUT);	/* tcp_drop() */
4767 	}
4768 	/*
4769 	 * Hack: if the peer is dead/unreachable, we do not time out if the
4770 	 * window is closed.  After a full backoff, drop the connection if
4771 	 * the idle time (no responses to probes) reaches the maximum
4772 	 * backoff that we would use if retransmitting.
4773 	 */
4774 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
4775 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
4776 	    ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
4777 		KMOD_TCPSTAT_INC(tcps_persistdrop);
4778 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4779 		return (-ETIMEDOUT);	/* tcp_drop() */
4780 	}
4781 	if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) &&
4782 	    tp->snd_una == tp->snd_max) {
4783 		bbr_exit_persist(tp, bbr, cts, __LINE__);
4784 		retval = 0;
4785 		goto out;
4786 	}
4787 	/*
4788 	 * If the user has closed the socket then drop a persisting
4789 	 * connection after a much reduced timeout.
4790 	 */
4791 	if (tp->t_state > TCPS_CLOSE_WAIT &&
4792 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
4793 		KMOD_TCPSTAT_INC(tcps_persistdrop);
4794 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4795 		return (-ETIMEDOUT);	/* tcp_drop() */
4796 	}
4797 	t_template = tcpip_maketemplate(bbr->rc_inp);
4798 	if (t_template) {
4799 		tcp_respond(tp, t_template->tt_ipgen,
4800 			    &t_template->tt_t, (struct mbuf *)NULL,
4801 			    tp->rcv_nxt, tp->snd_una - 1, 0);
4802 		/* This sends an ack */
4803 		if (tp->t_flags & TF_DELACK)
4804 			tp->t_flags &= ~TF_DELACK;
4805 		free(t_template, M_TEMP);
4806 	}
4807 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
4808 		tp->t_rxtshift++;
4809 	bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0);
4810 out:
4811 	return (retval);
4812 }
4813 
4814 /*
4815  * If a keepalive goes off, we had no other timers
4816  * happening. We always return 1 here since this
4817  * routine either drops the connection or sends
4818  * out a segment with respond.
4819  */
4820 static int
4821 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4822 {
4823 	struct tcptemp *t_template;
4824 	struct inpcb *inp = tptoinpcb(tp);
4825 
4826 	if (bbr->rc_all_timers_stopped) {
4827 		return (1);
4828 	}
4829 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
4830 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP);
4831 	/*
4832 	 * Keep-alive timer went off; send something or drop connection if
4833 	 * idle for too long.
4834 	 */
4835 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
4836 	if (tp->t_state < TCPS_ESTABLISHED)
4837 		goto dropit;
4838 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4839 	    tp->t_state <= TCPS_CLOSING) {
4840 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
4841 			goto dropit;
4842 		/*
4843 		 * Send a packet designed to force a response if the peer is
4844 		 * up and reachable: either an ACK if the connection is
4845 		 * still alive, or an RST if the peer has closed the
4846 		 * connection due to timeout or reboot. Using sequence
4847 		 * number tp->snd_una-1 causes the transmitted zero-length
4848 		 * segment to lie outside the receive window; by the
4849 		 * protocol spec, this requires the correspondent TCP to
4850 		 * respond.
4851 		 */
4852 		KMOD_TCPSTAT_INC(tcps_keepprobe);
4853 		t_template = tcpip_maketemplate(inp);
4854 		if (t_template) {
4855 			tcp_respond(tp, t_template->tt_ipgen,
4856 			    &t_template->tt_t, (struct mbuf *)NULL,
4857 			    tp->rcv_nxt, tp->snd_una - 1, 0);
4858 			free(t_template, M_TEMP);
4859 		}
4860 	}
4861 	bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0);
4862 	return (1);
4863 dropit:
4864 	KMOD_TCPSTAT_INC(tcps_keepdrops);
4865 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
4866 	return (-ETIMEDOUT);	/* tcp_drop() */
4867 }
4868 
4869 /*
4870  * Retransmit helper function, clear up all the ack
4871  * flags and take care of important book keeping.
4872  */
4873 static void
4874 bbr_remxt_tmr(struct tcpcb *tp)
4875 {
4876 	/*
4877 	 * The retransmit timer went off, all sack'd blocks must be
4878 	 * un-acked.
4879 	 */
4880 	struct bbr_sendmap *rsm, *trsm = NULL;
4881 	struct tcp_bbr *bbr;
4882 	uint32_t cts, lost;
4883 
4884 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
4885 	cts = tcp_get_usecs(&bbr->rc_tv);
4886 	lost = bbr->r_ctl.rc_lost;
4887 	if (bbr->r_state && (bbr->r_state != tp->t_state))
4888 		bbr_set_state(tp, bbr, 0);
4889 
4890 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
4891 		if (rsm->r_flags & BBR_ACKED) {
4892 			uint32_t old_flags;
4893 
4894 			rsm->r_dupack = 0;
4895 			if (rsm->r_in_tmap == 0) {
4896 				/* We must re-add it back to the tlist */
4897 				if (trsm == NULL) {
4898 					TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
4899 				} else {
4900 					TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext);
4901 				}
4902 				rsm->r_in_tmap = 1;
4903 			}
4904 			old_flags = rsm->r_flags;
4905 			rsm->r_flags |= BBR_RXT_CLEARED;
4906 			rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS);
4907 			bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
4908 		} else {
4909 			if ((tp->t_state < TCPS_ESTABLISHED) &&
4910 			    (rsm->r_start == tp->snd_una)) {
4911 				/*
4912 				 * Special case for TCP FO. Where
4913 				 * we sent more data beyond the snd_max.
4914 				 * We don't mark that as lost and stop here.
4915 				 */
4916 				break;
4917 			}
4918 			if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4919 				bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4920 				bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4921 			}
4922 			if (bbr_marks_rxt_sack_passed) {
4923 				/*
4924 				 * With this option, we will rack out
4925 				 * in 1ms increments the rest of the packets.
4926 				 */
4927 				rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST;
4928 				rsm->r_flags &= ~BBR_WAS_SACKPASS;
4929 			} else {
4930 				/*
4931 				 * With this option we only mark them lost
4932 				 * and remove all sack'd markings. We will run
4933 				 * another RXT or a TLP. This will cause
4934 				 * us to eventually send more based on what
4935 				 * ack's come in.
4936 				 */
4937 				rsm->r_flags |= BBR_MARKED_LOST;
4938 				rsm->r_flags &= ~BBR_WAS_SACKPASS;
4939 				rsm->r_flags &= ~BBR_SACK_PASSED;
4940 			}
4941 		}
4942 		trsm = rsm;
4943 	}
4944 	bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4945 	/* Clear the count (we just un-acked them) */
4946 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR);
4947 	bbr->rc_tlp_new_data = 0;
4948 	bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4949 	/* zap the behindness on a rxt */
4950 	bbr->r_ctl.rc_hptsi_agg_delay = 0;
4951 	bbr->r_agg_early_set = 0;
4952 	bbr->r_ctl.rc_agg_early = 0;
4953 	bbr->rc_tlp_rtx_out = 0;
4954 	bbr->r_ctl.rc_sacked = 0;
4955 	bbr->r_ctl.rc_sacklast = NULL;
4956 	bbr->r_timer_override = 1;
4957 	bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4958 }
4959 
4960 /*
4961  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
4962  * we will setup to retransmit the lowest seq number outstanding.
4963  */
4964 static int
4965 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4966 {
4967 	struct inpcb *inp = tptoinpcb(tp);
4968 	int32_t rexmt;
4969 	int32_t retval = 0;
4970 	bool isipv6;
4971 
4972 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
4973 	if (bbr->rc_all_timers_stopped) {
4974 		return (1);
4975 	}
4976 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
4977 	    (tp->snd_una == tp->snd_max)) {
4978 		/* Nothing outstanding .. nothing to do */
4979 		return (0);
4980 	}
4981 	/*
4982 	 * Retransmission timer went off.  Message has not been acked within
4983 	 * retransmit interval.  Back off to a longer retransmit interval
4984 	 * and retransmit one segment.
4985 	 */
4986 	if (ctf_progress_timeout_check(tp, true)) {
4987 		bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4988 		return (-ETIMEDOUT);	/* tcp_drop() */
4989 	}
4990 	bbr_remxt_tmr(tp);
4991 	if ((bbr->r_ctl.rc_resend == NULL) ||
4992 	    ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) {
4993 		/*
4994 		 * If the rwnd collapsed on
4995 		 * the one we are retransmitting
4996 		 * it does not count against the
4997 		 * rxt count.
4998 		 */
4999 		tp->t_rxtshift++;
5000 	}
5001 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
5002 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
5003 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
5004 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5005 		/* XXXGL: previously t_softerror was casted to uint16_t */
5006 		MPASS(tp->t_softerror >= 0);
5007 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
5008 		return (retval);	/* tcp_drop() */
5009 	}
5010 	if (tp->t_state == TCPS_SYN_SENT) {
5011 		/*
5012 		 * If the SYN was retransmitted, indicate CWND to be limited
5013 		 * to 1 segment in cc_conn_init().
5014 		 */
5015 		tp->snd_cwnd = 1;
5016 	} else if (tp->t_rxtshift == 1) {
5017 		/*
5018 		 * first retransmit; record ssthresh and cwnd so they can be
5019 		 * recovered if this turns out to be a "bad" retransmit. A
5020 		 * retransmit is considered "bad" if an ACK for this segment
5021 		 * is received within RTT/2 interval; the assumption here is
5022 		 * that the ACK was already in flight.  See "On Estimating
5023 		 * End-to-End Network Path Properties" by Allman and Paxson
5024 		 * for more details.
5025 		 */
5026 		tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5027 		if (!IN_RECOVERY(tp->t_flags)) {
5028 			tp->snd_cwnd_prev = tp->snd_cwnd;
5029 			tp->snd_ssthresh_prev = tp->snd_ssthresh;
5030 			tp->snd_recover_prev = tp->snd_recover;
5031 			tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5032 			tp->t_flags |= TF_PREVVALID;
5033 		} else {
5034 			tp->t_flags &= ~TF_PREVVALID;
5035 		}
5036 		tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5037 	} else {
5038 		tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5039 		tp->t_flags &= ~TF_PREVVALID;
5040 	}
5041 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
5042 	if ((tp->t_state == TCPS_SYN_SENT) ||
5043 	    (tp->t_state == TCPS_SYN_RECEIVED))
5044 		rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift];
5045 	else
5046 		rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5047 	TCPT_RANGESET(tp->t_rxtcur, rexmt,
5048 	    MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms),
5049 	    MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
5050 	/*
5051 	 * We enter the path for PLMTUD if connection is established or, if
5052 	 * connection is FIN_WAIT_1 status, reason for the last is that if
5053 	 * amount of data we send is very small, we could send it in couple
5054 	 * of packets and process straight to FIN. In that case we won't
5055 	 * catch ESTABLISHED state.
5056 	 */
5057 #ifdef INET6
5058 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
5059 #else
5060 	isipv6 = false;
5061 #endif
5062 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
5063 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
5064 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
5065 	    ((tp->t_state == TCPS_ESTABLISHED) ||
5066 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
5067 		/*
5068 		 * Idea here is that at each stage of mtu probe (usually,
5069 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
5070 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5071 		 * should take care of that.
5072 		 */
5073 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5074 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5075 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5076 		    tp->t_rxtshift % 2 == 0)) {
5077 			/*
5078 			 * Enter Path MTU Black-hole Detection mechanism: -
5079 			 * Disable Path MTU Discovery (IP "DF" bit). -
5080 			 * Reduce MTU to lower value than what we negotiated
5081 			 * with peer.
5082 			 */
5083 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5084 				/*
5085 				 * Record that we may have found a black
5086 				 * hole.
5087 				 */
5088 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5089 				/* Keep track of previous MSS. */
5090 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5091 			}
5092 			/*
5093 			 * Reduce the MSS to blackhole value or to the
5094 			 * default in an attempt to retransmit.
5095 			 */
5096 #ifdef INET6
5097 			isipv6 = bbr->r_is_v6;
5098 			if (isipv6 &&
5099 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5100 				/* Use the sysctl tuneable blackhole MSS. */
5101 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5102 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5103 			} else if (isipv6) {
5104 				/* Use the default MSS. */
5105 				tp->t_maxseg = V_tcp_v6mssdflt;
5106 				/*
5107 				 * Disable Path MTU Discovery when we switch
5108 				 * to minmss.
5109 				 */
5110 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5111 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5112 			}
5113 #endif
5114 #if defined(INET6) && defined(INET)
5115 			else
5116 #endif
5117 #ifdef INET
5118 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5119 				/* Use the sysctl tuneable blackhole MSS. */
5120 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5121 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5122 			} else {
5123 				/* Use the default MSS. */
5124 				tp->t_maxseg = V_tcp_mssdflt;
5125 				/*
5126 				 * Disable Path MTU Discovery when we switch
5127 				 * to minmss.
5128 				 */
5129 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5130 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5131 			}
5132 #endif
5133 		} else {
5134 			/*
5135 			 * If further retransmissions are still unsuccessful
5136 			 * with a lowered MTU, maybe this isn't a blackhole
5137 			 * and we restore the previous MSS and blackhole
5138 			 * detection flags. The limit '6' is determined by
5139 			 * giving each probe stage (1448, 1188, 524) 2
5140 			 * chances to recover.
5141 			 */
5142 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5143 			    (tp->t_rxtshift >= 6)) {
5144 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5145 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5146 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5147 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5148 			}
5149 		}
5150 	}
5151 	/*
5152 	 * Disable RFC1323 and SACK if we haven't got any response to our
5153 	 * third SYN to work-around some broken terminal servers (most of
5154 	 * which have hopefully been retired) that have bad VJ header
5155 	 * compression code which trashes TCP segments containing
5156 	 * unknown-to-them TCP options.
5157 	 */
5158 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
5159 	    (tp->t_rxtshift == 3))
5160 		tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT);
5161 	/*
5162 	 * If we backed off this far, our srtt estimate is probably bogus.
5163 	 * Clobber it so we'll take the next rtt measurement as our srtt;
5164 	 * move the current srtt into rttvar to keep the current retransmit
5165 	 * times until then.
5166 	 */
5167 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5168 #ifdef INET6
5169 		if (bbr->r_is_v6)
5170 			in6_losing(inp);
5171 		else
5172 #endif
5173 			in_losing(inp);
5174 		tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5175 		tp->t_srtt = 0;
5176 	}
5177 	sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
5178 	tp->snd_recover = tp->snd_max;
5179 	tp->t_flags |= TF_ACKNOW;
5180 	tp->t_rtttime = 0;
5181 
5182 	return (retval);
5183 }
5184 
5185 static int
5186 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling)
5187 {
5188 	int32_t ret = 0;
5189 	int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5190 
5191 	if (timers == 0) {
5192 		return (0);
5193 	}
5194 	if (tp->t_state == TCPS_LISTEN) {
5195 		/* no timers on listen sockets */
5196 		if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5197 			return (0);
5198 		return (1);
5199 	}
5200 	if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
5201 		uint32_t left;
5202 
5203 		if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5204 			ret = -1;
5205 			bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5206 			return (0);
5207 		}
5208 		if (hpts_calling == 0) {
5209 			ret = -2;
5210 			bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5211 			return (0);
5212 		}
5213 		/*
5214 		 * Ok our timer went off early and we are not paced false
5215 		 * alarm, go back to sleep.
5216 		 */
5217 		left = bbr->r_ctl.rc_timer_exp - cts;
5218 		ret = -3;
5219 		bbr_log_to_processing(bbr, cts, ret, left, hpts_calling);
5220 		tcp_hpts_insert(tptoinpcb(tp), HPTS_USEC_TO_SLOTS(left));
5221 		return (1);
5222 	}
5223 	bbr->rc_tmr_stopped = 0;
5224 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5225 	if (timers & PACE_TMR_DELACK) {
5226 		ret = bbr_timeout_delack(tp, bbr, cts);
5227 	} else if (timers & PACE_TMR_PERSIT) {
5228 		ret = bbr_timeout_persist(tp, bbr, cts);
5229 	} else if (timers & PACE_TMR_RACK) {
5230 		bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5231 		ret = bbr_timeout_rack(tp, bbr, cts);
5232 	} else if (timers & PACE_TMR_TLP) {
5233 		bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5234 		ret = bbr_timeout_tlp(tp, bbr, cts);
5235 	} else if (timers & PACE_TMR_RXT) {
5236 		bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5237 		ret = bbr_timeout_rxt(tp, bbr, cts);
5238 	} else if (timers & PACE_TMR_KEEP) {
5239 		ret = bbr_timeout_keepalive(tp, bbr, cts);
5240 	}
5241 	bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling);
5242 	return (ret);
5243 }
5244 
5245 static void
5246 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts)
5247 {
5248 	if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5249 		uint8_t hpts_removed = 0;
5250 
5251 		if (tcp_in_hpts(bbr->rc_inp) &&
5252 		    (bbr->rc_timer_first == 1)) {
5253 			/*
5254 			 * If we are canceling timer's when we have the
5255 			 * timer ahead of the output being paced. We also
5256 			 * must remove ourselves from the hpts.
5257 			 */
5258 			hpts_removed = 1;
5259 			tcp_hpts_remove(bbr->rc_inp);
5260 			if (bbr->r_ctl.rc_last_delay_val) {
5261 				/* Update the last hptsi delay too */
5262 				uint32_t time_since_send;
5263 
5264 				if (TSTMP_GT(cts, bbr->rc_pacer_started))
5265 					time_since_send = cts - bbr->rc_pacer_started;
5266 				else
5267 					time_since_send = 0;
5268 				if (bbr->r_ctl.rc_last_delay_val > time_since_send) {
5269 					/* Cut down our slot time */
5270 					bbr->r_ctl.rc_last_delay_val -= time_since_send;
5271 				} else {
5272 					bbr->r_ctl.rc_last_delay_val = 0;
5273 				}
5274 				bbr->rc_pacer_started = cts;
5275 			}
5276 		}
5277 		bbr->rc_timer_first = 0;
5278 		bbr_log_to_cancel(bbr, line, cts, hpts_removed);
5279 		bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5280 		bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5281 	}
5282 }
5283 
5284 static int
5285 bbr_stopall(struct tcpcb *tp)
5286 {
5287 	struct tcp_bbr *bbr;
5288 
5289 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
5290 	bbr->rc_all_timers_stopped = 1;
5291 	return (0);
5292 }
5293 
5294 static uint32_t
5295 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts)
5296 {
5297 	struct bbr_sendmap *rsm;
5298 
5299 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
5300 	if ((rsm == NULL) || (u_rsm == rsm))
5301 		return (cts);
5302 	return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
5303 }
5304 
5305 static void
5306 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr,
5307      struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time)
5308 {
5309 	int32_t idx;
5310 
5311 	rsm->r_rtr_cnt++;
5312 	rsm->r_dupack = 0;
5313 	if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) {
5314 		rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS;
5315 		rsm->r_flags |= BBR_OVERMAX;
5316 	}
5317 	if (rsm->r_flags & BBR_RWND_COLLAPSED) {
5318 		/* Take off the collapsed flag at rxt */
5319 		rsm->r_flags &= ~BBR_RWND_COLLAPSED;
5320 	}
5321 	if (rsm->r_flags & BBR_MARKED_LOST) {
5322 		/* We have retransmitted, its no longer lost */
5323 		rsm->r_flags &= ~BBR_MARKED_LOST;
5324 		bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
5325 	}
5326 	if (rsm->r_flags & BBR_RXT_CLEARED) {
5327 		/*
5328 		 * We hit a RXT timer on it and
5329 		 * we cleared the "acked" flag.
5330 		 * We now have it going back into
5331 		 * flight, we can remove the cleared
5332 		 * flag and possibly do accounting on
5333 		 * this piece.
5334 		 */
5335 		rsm->r_flags &= ~BBR_RXT_CLEARED;
5336 	}
5337 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) {
5338 		bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5339 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5340 	}
5341 	idx = rsm->r_rtr_cnt - 1;
5342 	rsm->r_tim_lastsent[idx] = cts;
5343 	rsm->r_pacing_delay = pacing_time;
5344 	rsm->r_delivered = bbr->r_ctl.rc_delivered;
5345 	rsm->r_ts_valid = bbr->rc_ts_valid;
5346 	if (bbr->rc_ts_valid)
5347 		rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5348 	if (bbr->r_ctl.r_app_limited_until)
5349 		rsm->r_app_limited = 1;
5350 	else
5351 		rsm->r_app_limited = 0;
5352 	if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5353 		rsm->r_bbr_state = bbr_state_val(bbr);
5354 	else
5355 		rsm->r_bbr_state = 8;
5356 	if (rsm->r_flags & BBR_ACKED) {
5357 		/* Problably MTU discovery messing with us */
5358 		uint32_t old_flags;
5359 
5360 		old_flags = rsm->r_flags;
5361 		rsm->r_flags &= ~BBR_ACKED;
5362 		bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
5363 		bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5364 		if (bbr->r_ctl.rc_sacked == 0)
5365 			bbr->r_ctl.rc_sacklast = NULL;
5366 	}
5367 	if (rsm->r_in_tmap) {
5368 		TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5369 	}
5370 	TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5371 	rsm->r_in_tmap = 1;
5372 	if (rsm->r_flags & BBR_SACK_PASSED) {
5373 		/* We have retransmitted due to the SACK pass */
5374 		rsm->r_flags &= ~BBR_SACK_PASSED;
5375 		rsm->r_flags |= BBR_WAS_SACKPASS;
5376 	}
5377 	rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5378 	rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5379 						(bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5380 	bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
5381 	if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5382 		rsm->r_is_gain = 1;
5383 		rsm->r_is_drain = 0;
5384 	} else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5385 		rsm->r_is_drain = 1;
5386 		rsm->r_is_gain = 0;
5387 	} else {
5388 		rsm->r_is_drain = 0;
5389 		rsm->r_is_gain = 0;
5390 	}
5391 	rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */
5392 }
5393 
5394 /*
5395  * Returns 0, or the sequence where we stopped
5396  * updating. We also update the lenp to be the amount
5397  * of data left.
5398  */
5399 
5400 static uint32_t
5401 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr,
5402     struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time)
5403 {
5404 	/*
5405 	 * We (re-)transmitted starting at rsm->r_start for some length
5406 	 * (possibly less than r_end.
5407 	 */
5408 	struct bbr_sendmap *nrsm;
5409 	uint32_t c_end;
5410 	int32_t len;
5411 
5412 	len = *lenp;
5413 	c_end = rsm->r_start + len;
5414 	if (SEQ_GEQ(c_end, rsm->r_end)) {
5415 		/*
5416 		 * We retransmitted the whole piece or more than the whole
5417 		 * slopping into the next rsm.
5418 		 */
5419 		bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5420 		if (c_end == rsm->r_end) {
5421 			*lenp = 0;
5422 			return (0);
5423 		} else {
5424 			int32_t act_len;
5425 
5426 			/* Hangs over the end return whats left */
5427 			act_len = rsm->r_end - rsm->r_start;
5428 			*lenp = (len - act_len);
5429 			return (rsm->r_end);
5430 		}
5431 		/* We don't get out of this block. */
5432 	}
5433 	/*
5434 	 * Here we retransmitted less than the whole thing which means we
5435 	 * have to split this into what was transmitted and what was not.
5436 	 */
5437 	nrsm = bbr_alloc_full_limit(bbr);
5438 	if (nrsm == NULL) {
5439 		*lenp = 0;
5440 		return (0);
5441 	}
5442 	/*
5443 	 * So here we are going to take the original rsm and make it what we
5444 	 * retransmitted. nrsm will be the tail portion we did not
5445 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
5446 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
5447 	 * 1, 6 and the new piece will be 6, 11.
5448 	 */
5449 	bbr_clone_rsm(bbr, nrsm, rsm, c_end);
5450 	TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
5451 	nrsm->r_dupack = 0;
5452 	if (rsm->r_in_tmap) {
5453 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5454 		nrsm->r_in_tmap = 1;
5455 	}
5456 	rsm->r_flags &= (~BBR_HAS_FIN);
5457 	bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5458 	*lenp = 0;
5459 	return (0);
5460 }
5461 
5462 static uint64_t
5463 bbr_get_hardware_rate(struct tcp_bbr *bbr)
5464 {
5465 	uint64_t bw;
5466 
5467 	bw = bbr_get_bw(bbr);
5468 	bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN];
5469 	bw /= (uint64_t)BBR_UNIT;
5470 	return(bw);
5471 }
5472 
5473 static void
5474 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts,
5475 		       uint64_t act_rate, uint64_t rate_wanted)
5476 {
5477 	/*
5478 	 * We could not get a full gains worth
5479 	 * of rate.
5480 	 */
5481 	if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) {
5482 		/* we can't even get the real rate */
5483 		uint64_t red;
5484 
5485 		bbr->skip_gain = 1;
5486 		bbr->gain_is_limited = 0;
5487 		red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate;
5488 		if (red)
5489 			filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts);
5490 	} else {
5491 		/* We can use a lower gain */
5492 		bbr->skip_gain = 0;
5493 		bbr->gain_is_limited = 1;
5494 	}
5495 }
5496 
5497 static void
5498 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts)
5499 {
5500 	const struct tcp_hwrate_limit_table *nrte;
5501 	int error, rate = -1;
5502 
5503 	if (bbr->r_ctl.crte == NULL)
5504 		return;
5505 	if ((bbr->rc_inp->inp_route.ro_nh == NULL) ||
5506 	    (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) {
5507 		/* Lost our routes? */
5508 		/* Clear the way for a re-attempt */
5509 		bbr->bbr_attempt_hdwr_pace = 0;
5510 lost_rate:
5511 		bbr->gain_is_limited = 0;
5512 		bbr->skip_gain = 0;
5513 		bbr->bbr_hdrw_pacing = 0;
5514 		counter_u64_add(bbr_flows_whdwr_pacing, -1);
5515 		counter_u64_add(bbr_flows_nohdwr_pacing, 1);
5516 		tcp_bbr_tso_size_check(bbr, cts);
5517 		return;
5518 	}
5519 	rate = bbr_get_hardware_rate(bbr);
5520 	nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte,
5521 				   bbr->rc_tp,
5522 				   bbr->rc_inp->inp_route.ro_nh->nh_ifp,
5523 				   rate,
5524 				   (RS_PACING_GEQ|RS_PACING_SUB_OK),
5525 				   &error, NULL);
5526 	if (nrte == NULL) {
5527 		goto lost_rate;
5528 	}
5529 	if (nrte != bbr->r_ctl.crte) {
5530 		bbr->r_ctl.crte = nrte;
5531 		if (error == 0)  {
5532 			BBR_STAT_INC(bbr_hdwr_rl_mod_ok);
5533 			if (bbr->r_ctl.crte->rate < rate) {
5534 				/* We have a problem */
5535 				bbr_setup_less_of_rate(bbr, cts,
5536 						       bbr->r_ctl.crte->rate, rate);
5537 			} else {
5538 				/* We are good */
5539 				bbr->gain_is_limited = 0;
5540 				bbr->skip_gain = 0;
5541 			}
5542 		} else {
5543 			/* A failure should release the tag */
5544 			BBR_STAT_INC(bbr_hdwr_rl_mod_fail);
5545 			bbr->gain_is_limited = 0;
5546 			bbr->skip_gain = 0;
5547 			bbr->bbr_hdrw_pacing = 0;
5548 		}
5549 		bbr_type_log_hdwr_pacing(bbr,
5550 					 bbr->r_ctl.crte->ptbl->rs_ifp,
5551 					 rate,
5552 					 ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate),
5553 					 __LINE__,
5554 					 cts,
5555 					 error);
5556 	}
5557 }
5558 
5559 static void
5560 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts)
5561 {
5562 	/*
5563 	 * If we have hardware pacing support
5564 	 * we need to factor that in for our
5565 	 * TSO size.
5566 	 */
5567 	const struct tcp_hwrate_limit_table *rlp;
5568 	uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay;
5569 
5570 	if ((bbr->bbr_hdrw_pacing == 0) ||
5571 	    (IN_RECOVERY(bbr->rc_tp->t_flags)) ||
5572 	    (bbr->r_ctl.crte == NULL))
5573 		return;
5574 	if (bbr->hw_pacing_set == 0) {
5575 		/* Not yet by the hdwr pacing count delay */
5576 		return;
5577 	}
5578 	if (bbr_hdwr_pace_adjust == 0) {
5579 		/* No adjustment */
5580 		return;
5581 	}
5582 	rlp = bbr->r_ctl.crte;
5583 	if (bbr->rc_tp->t_maxseg > bbr->rc_last_options)
5584 		maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5585 	else
5586 		maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5587 	/*
5588 	 * So lets first get the
5589 	 * time we will take between
5590 	 * TSO sized sends currently without
5591 	 * hardware help.
5592 	 */
5593 	cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT,
5594 		        bbr->r_ctl.rc_pace_max_segs, cts, 1);
5595 	hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg;
5596 	hdwr_delay *= rlp->time_between;
5597 	if (cur_delay > hdwr_delay)
5598 		delta = cur_delay - hdwr_delay;
5599 	else
5600 		delta = 0;
5601 	bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay,
5602 			     (bbr->r_ctl.rc_pace_max_segs / maxseg),
5603 			     1);
5604 	if (delta &&
5605 	    (delta < (max(rlp->time_between,
5606 			  bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) {
5607 		/*
5608 		 * Now lets divide by the pacing
5609 		 * time between each segment the
5610 		 * hardware sends rounding up and
5611 		 * derive a bytes from that. We multiply
5612 		 * that by bbr_hdwr_pace_adjust to get
5613 		 * more bang for our buck.
5614 		 *
5615 		 * The goal is to have the software pacer
5616 		 * waiting no more than an additional
5617 		 * pacing delay if we can (without the
5618 		 * compensation i.e. x bbr_hdwr_pace_adjust).
5619 		 */
5620 		seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between),
5621 			     (bbr->r_ctl.rc_pace_max_segs/maxseg));
5622 		seg_sz *= bbr_hdwr_pace_adjust;
5623 		if (bbr_hdwr_pace_floor &&
5624 		    (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5625 			/* Currently hardware paces
5626 			 * out rs_min_seg segments at a time.
5627 			 * We need to make sure we always send at least
5628 			 * a full burst of bbr_hdwr_pace_floor down.
5629 			 */
5630 			seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5631 		}
5632 		seg_sz *= maxseg;
5633 	} else if (delta == 0) {
5634 		/*
5635 		 * The highest pacing rate is
5636 		 * above our b/w gained. This means
5637 		 * we probably are going quite fast at
5638 		 * the hardware highest rate. Lets just multiply
5639 		 * the calculated TSO size by the
5640 		 * multiplier factor (its probably
5641 		 * 4 segments in the default config for
5642 		 * mlx).
5643 		 */
5644 		seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust;
5645 		if (bbr_hdwr_pace_floor &&
5646 		    (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5647 			/* Currently hardware paces
5648 			 * out rs_min_seg segments at a time.
5649 			 * We need to make sure we always send at least
5650 			 * a full burst of bbr_hdwr_pace_floor down.
5651 			 */
5652 			seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5653 		}
5654 	} else {
5655 		/*
5656 		 * The pacing time difference is so
5657 		 * big that the hardware will
5658 		 * pace out more rapidly then we
5659 		 * really want and then we
5660 		 * will have a long delay. Lets just keep
5661 		 * the same TSO size so its as if
5662 		 * we were not using hdwr pacing (we
5663 		 * just gain a bit of spacing from the
5664 		 * hardware if seg_sz > 1).
5665 		 */
5666 		seg_sz = bbr->r_ctl.rc_pace_max_segs;
5667 	}
5668 	if (seg_sz > bbr->r_ctl.rc_pace_max_segs)
5669 		new_tso = seg_sz;
5670 	else
5671 		new_tso = bbr->r_ctl.rc_pace_max_segs;
5672 	if (new_tso >= (PACE_MAX_IP_BYTES-maxseg))
5673 		new_tso = PACE_MAX_IP_BYTES - maxseg;
5674 
5675 	if (new_tso != bbr->r_ctl.rc_pace_max_segs) {
5676 		bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0);
5677 		bbr->r_ctl.rc_pace_max_segs = new_tso;
5678 	}
5679 }
5680 
5681 static void
5682 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts)
5683 {
5684 	uint64_t bw;
5685 	uint32_t old_tso = 0, new_tso;
5686 	uint32_t maxseg, bytes;
5687 	uint32_t tls_seg=0;
5688 	/*
5689 	 * Google/linux uses the following algorithm to determine
5690 	 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18):
5691 	 *
5692 	 *  bytes = bw_in_bytes_per_second / 1000
5693 	 *  bytes = min(bytes, 64k)
5694 	 *  tso_segs = bytes / MSS
5695 	 *  if (bw < 1.2Mbs)
5696 	 *      min_tso_segs = 1
5697 	 *  else
5698 	 *	min_tso_segs = 2
5699 	 * tso_segs = max(tso_segs, min_tso_segs)
5700 	 *
5701 	 * * Note apply a device specific limit (we apply this in the
5702 	 *   tcp_m_copym).
5703 	 * Note that before the initial measurement is made google bursts out
5704 	 * a full iwnd just like new-reno/cubic.
5705 	 *
5706 	 * We do not use this algorithm. Instead we
5707 	 * use a two phased approach:
5708 	 *
5709 	 *  if ( bw <= per-tcb-cross-over)
5710 	 *     goal_tso =  calculate how much with this bw we
5711 	 *                 can send in goal-time seconds.
5712 	 *     if (goal_tso > mss)
5713 	 *         seg = goal_tso / mss
5714 	 *         tso = seg * mss
5715 	 *     else
5716 	 *         tso = mss
5717 	 *     if (tso > per-tcb-max)
5718 	 *         tso = per-tcb-max
5719 	 *  else if ( bw > 512Mbps)
5720 	 *     tso = max-tso (64k/mss)
5721 	 *  else
5722 	 *     goal_tso = bw / per-tcb-divsor
5723 	 *     seg = (goal_tso + mss-1)/mss
5724 	 *     tso = seg * mss
5725 	 *
5726 	 * if (tso < per-tcb-floor)
5727 	 *    tso = per-tcb-floor
5728 	 * if (tso > per-tcb-utter_max)
5729 	 *    tso = per-tcb-utter_max
5730 	 *
5731 	 * Note the default per-tcb-divisor is 1000 (same as google).
5732 	 * the goal cross over is 30Mbps however. To recreate googles
5733 	 * algorithm you need to set:
5734 	 *
5735 	 * cross-over = 23,168,000 bps
5736 	 * goal-time = 18000
5737 	 * per-tcb-max = 2
5738 	 * per-tcb-divisor = 1000
5739 	 * per-tcb-floor = 1
5740 	 *
5741 	 * This will get you "google bbr" behavior with respect to tso size.
5742 	 *
5743 	 * Note we do set anything TSO size until we are past the initial
5744 	 * window. Before that we gnerally use either a single MSS
5745 	 * or we use the full IW size (so we burst a IW at a time)
5746 	 */
5747 
5748 	if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) {
5749 		maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5750 	} else {
5751 		maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5752 	}
5753 	old_tso = bbr->r_ctl.rc_pace_max_segs;
5754 	if (bbr->rc_past_init_win == 0) {
5755 		/*
5756 		 * Not enough data has been acknowledged to make a
5757 		 * judgement. Set up the initial TSO based on if we
5758 		 * are sending a full IW at once or not.
5759 		 */
5760 		if (bbr->rc_use_google)
5761 			bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2);
5762 		else if (bbr->bbr_init_win_cheat)
5763 			bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp);
5764 		else
5765 			bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5766 		if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg)
5767 			bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg;
5768 		if (bbr->r_ctl.rc_pace_max_segs == 0) {
5769 			bbr->r_ctl.rc_pace_max_segs = maxseg;
5770 		}
5771 		bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0);
5772 			bbr_adjust_for_hw_pacing(bbr, cts);
5773 		return;
5774 	}
5775 	/**
5776 	 * Now lets set the TSO goal based on our delivery rate in
5777 	 * bytes per second. Note we only do this if
5778 	 * we have acked at least the initial cwnd worth of data.
5779 	 */
5780 	bw = bbr_get_bw(bbr);
5781 	if (IN_RECOVERY(bbr->rc_tp->t_flags) &&
5782 	     (bbr->rc_use_google == 0)) {
5783 		/* We clamp to one MSS in recovery */
5784 		new_tso = maxseg;
5785 	} else if (bbr->rc_use_google) {
5786 		int min_tso_segs;
5787 
5788 		/* Google considers the gain too */
5789 		if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) {
5790 			bw *= bbr->r_ctl.rc_bbr_hptsi_gain;
5791 			bw /= BBR_UNIT;
5792 		}
5793 		bytes = bw / 1024;
5794 		if (bytes > (64 * 1024))
5795 			bytes = 64 * 1024;
5796 		new_tso = bytes / maxseg;
5797 		if (bw < ONE_POINT_TWO_MEG)
5798 			min_tso_segs = 1;
5799 		else
5800 			min_tso_segs = 2;
5801 		if (new_tso < min_tso_segs)
5802 			new_tso = min_tso_segs;
5803 		new_tso *= maxseg;
5804 	} else if (bbr->rc_no_pacing) {
5805 		new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg;
5806 	} else if (bw <= bbr->r_ctl.bbr_cross_over) {
5807 		/*
5808 		 * Calculate the worse case b/w TSO if we are inserting no
5809 		 * more than a delay_target number of TSO's.
5810 		 */
5811 		uint32_t tso_len, min_tso;
5812 
5813 		tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw);
5814 		if (tso_len > maxseg) {
5815 			new_tso = tso_len / maxseg;
5816 			if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max)
5817 				new_tso = bbr->r_ctl.bbr_hptsi_segments_max;
5818 			new_tso *= maxseg;
5819 		} else {
5820 			/*
5821 			 * less than a full sized frame yikes.. long rtt or
5822 			 * low bw?
5823 			 */
5824 			min_tso = bbr_minseg(bbr);
5825 			if ((tso_len > min_tso) && (bbr_all_get_min == 0))
5826 				new_tso = rounddown(tso_len, min_tso);
5827 			else
5828 				new_tso = min_tso;
5829 		}
5830 	} else if (bw > FIVETWELVE_MBPS) {
5831 		/*
5832 		 * This guy is so fast b/w wise that we can TSO as large as
5833 		 * possible of segments that the NIC will allow.
5834 		 */
5835 		new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5836 	} else {
5837 		/*
5838 		 * This formula is based on attempting to send a segment or
5839 		 * more every bbr_hptsi_per_second. The default is 1000
5840 		 * which means you are targeting what you can send every 1ms
5841 		 * based on the peers bw.
5842 		 *
5843 		 * If the number drops to say 500, then you are looking more
5844 		 * at 2ms and you will raise how much we send in a single
5845 		 * TSO thus saving CPU (less bbr_output_wtime() calls). The
5846 		 * trade off of course is you will send more at once and
5847 		 * thus tend to clump up the sends into larger "bursts"
5848 		 * building a queue.
5849 		 */
5850 		bw /= bbr->r_ctl.bbr_hptsi_per_second;
5851 		new_tso = roundup(bw, (uint64_t)maxseg);
5852 		/*
5853 		 * Gate the floor to match what our lower than 48Mbps
5854 		 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus
5855 		 * becomes the floor for this calculation.
5856 		 */
5857 		if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg))
5858 			new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg);
5859 	}
5860 	if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor)))
5861 		new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor;
5862 	if (new_tso > PACE_MAX_IP_BYTES)
5863 		new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5864 	/* Enforce an utter maximum. */
5865 	if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) {
5866 		new_tso = bbr->r_ctl.bbr_utter_max * maxseg;
5867 	}
5868 	if (old_tso != new_tso) {
5869 		/* Only log changes */
5870 		bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0);
5871 		bbr->r_ctl.rc_pace_max_segs = new_tso;
5872 	}
5873 	/* We have hardware pacing! */
5874 	bbr_adjust_for_hw_pacing(bbr, cts);
5875 }
5876 
5877 static void
5878 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len,
5879     uint32_t seq_out, uint16_t th_flags, int32_t err, uint32_t cts,
5880     struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc,
5881     struct sockbuf *sb)
5882 {
5883 
5884 	struct bbr_sendmap *rsm, *nrsm;
5885 	register uint32_t snd_max, snd_una;
5886 	uint32_t pacing_time;
5887 	/*
5888 	 * Add to the RACK log of packets in flight or retransmitted. If
5889 	 * there is a TS option we will use the TS echoed, if not we will
5890 	 * grab a TS.
5891 	 *
5892 	 * Retransmissions will increment the count and move the ts to its
5893 	 * proper place. Note that if options do not include TS's then we
5894 	 * won't be able to effectively use the ACK for an RTT on a retran.
5895 	 *
5896 	 * Notes about r_start and r_end. Lets consider a send starting at
5897 	 * sequence 1 for 10 bytes. In such an example the r_start would be
5898 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
5899 	 * This means that r_end is actually the first sequence for the next
5900 	 * slot (11).
5901 	 *
5902 	 */
5903 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5904 	if (err) {
5905 		/*
5906 		 * We don't log errors -- we could but snd_max does not
5907 		 * advance in this case either.
5908 		 */
5909 		return;
5910 	}
5911 	if (th_flags & TH_RST) {
5912 		/*
5913 		 * We don't log resets and we return immediately from
5914 		 * sending
5915 		 */
5916 		*abandon = 1;
5917 		return;
5918 	}
5919 	snd_una = tp->snd_una;
5920 	if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) {
5921 		/*
5922 		 * The call to bbr_log_output is made before bumping
5923 		 * snd_max. This means we can record one extra byte on a SYN
5924 		 * or FIN if seq_out is adding more on and a FIN is present
5925 		 * (and we are not resending).
5926 		 */
5927 		if ((th_flags & TH_SYN) && (tp->iss == seq_out))
5928 			len++;
5929 		if (th_flags & TH_FIN)
5930 			len++;
5931 	}
5932 	if (SEQ_LEQ((seq_out + len), snd_una)) {
5933 		/* Are sending an old segment to induce an ack (keep-alive)? */
5934 		return;
5935 	}
5936 	if (SEQ_LT(seq_out, snd_una)) {
5937 		/* huh? should we panic? */
5938 		uint32_t end;
5939 
5940 		end = seq_out + len;
5941 		seq_out = snd_una;
5942 		len = end - seq_out;
5943 	}
5944 	snd_max = tp->snd_max;
5945 	if (len == 0) {
5946 		/* We don't log zero window probes */
5947 		return;
5948 	}
5949 	pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1);
5950 	/* First question is it a retransmission? */
5951 	if (seq_out == snd_max) {
5952 again:
5953 		rsm = bbr_alloc(bbr);
5954 		if (rsm == NULL) {
5955 			return;
5956 		}
5957 		rsm->r_flags = 0;
5958 		if (th_flags & TH_SYN)
5959 			rsm->r_flags |= BBR_HAS_SYN;
5960 		if (th_flags & TH_FIN)
5961 			rsm->r_flags |= BBR_HAS_FIN;
5962 		rsm->r_tim_lastsent[0] = cts;
5963 		rsm->r_rtr_cnt = 1;
5964 		rsm->r_rtr_bytes = 0;
5965 		rsm->r_start = seq_out;
5966 		rsm->r_end = rsm->r_start + len;
5967 		rsm->r_dupack = 0;
5968 		rsm->r_delivered = bbr->r_ctl.rc_delivered;
5969 		rsm->r_pacing_delay = pacing_time;
5970 		rsm->r_ts_valid = bbr->rc_ts_valid;
5971 		if (bbr->rc_ts_valid)
5972 			rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5973 		rsm->r_del_time = bbr->r_ctl.rc_del_time;
5974 		if (bbr->r_ctl.r_app_limited_until)
5975 			rsm->r_app_limited = 1;
5976 		else
5977 			rsm->r_app_limited = 0;
5978 		rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5979 		rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5980 						(bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5981 		/*
5982 		 * Here we must also add in this rsm since snd_max
5983 		 * is updated after we return from a new send.
5984 		 */
5985 		rsm->r_flight_at_send += len;
5986 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
5987 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5988 		rsm->r_in_tmap = 1;
5989 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5990 			rsm->r_bbr_state = bbr_state_val(bbr);
5991 		else
5992 			rsm->r_bbr_state = 8;
5993 		if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5994 			rsm->r_is_gain = 1;
5995 			rsm->r_is_drain = 0;
5996 		} else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5997 			rsm->r_is_drain = 1;
5998 			rsm->r_is_gain = 0;
5999 		} else {
6000 			rsm->r_is_drain = 0;
6001 			rsm->r_is_gain = 0;
6002 		}
6003 		return;
6004 	}
6005 	/*
6006 	 * If we reach here its a retransmission and we need to find it.
6007 	 */
6008 more:
6009 	if (hintrsm && (hintrsm->r_start == seq_out)) {
6010 		rsm = hintrsm;
6011 		hintrsm = NULL;
6012 	} else if (bbr->r_ctl.rc_next) {
6013 		/* We have a hint from a previous run */
6014 		rsm = bbr->r_ctl.rc_next;
6015 	} else {
6016 		/* No hints sorry */
6017 		rsm = NULL;
6018 	}
6019 	if ((rsm) && (rsm->r_start == seq_out)) {
6020 		/*
6021 		 * We used rc_next or hintrsm  to retransmit, hopefully the
6022 		 * likely case.
6023 		 */
6024 		seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6025 		if (len == 0) {
6026 			return;
6027 		} else {
6028 			goto more;
6029 		}
6030 	}
6031 	/* Ok it was not the last pointer go through it the hard way. */
6032 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6033 		if (rsm->r_start == seq_out) {
6034 			seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6035 			bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
6036 			if (len == 0) {
6037 				return;
6038 			} else {
6039 				continue;
6040 			}
6041 		}
6042 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6043 			/* Transmitted within this piece */
6044 			/*
6045 			 * Ok we must split off the front and then let the
6046 			 * update do the rest
6047 			 */
6048 			nrsm = bbr_alloc_full_limit(bbr);
6049 			if (nrsm == NULL) {
6050 				bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
6051 				return;
6052 			}
6053 			/*
6054 			 * copy rsm to nrsm and then trim the front of rsm
6055 			 * to not include this part.
6056 			 */
6057 			bbr_clone_rsm(bbr, nrsm, rsm, seq_out);
6058 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
6059 			if (rsm->r_in_tmap) {
6060 				TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6061 				nrsm->r_in_tmap = 1;
6062 			}
6063 			rsm->r_flags &= (~BBR_HAS_FIN);
6064 			seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time);
6065 			if (len == 0) {
6066 				return;
6067 			}
6068 		}
6069 	}
6070 	/*
6071 	 * Hmm not found in map did they retransmit both old and on into the
6072 	 * new?
6073 	 */
6074 	if (seq_out == tp->snd_max) {
6075 		goto again;
6076 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
6077 #ifdef BBR_INVARIANTS
6078 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6079 		    seq_out, len, tp->snd_una, tp->snd_max);
6080 		printf("Starting Dump of all rack entries\n");
6081 		TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6082 			printf("rsm:%p start:%u end:%u\n",
6083 			    rsm, rsm->r_start, rsm->r_end);
6084 		}
6085 		printf("Dump complete\n");
6086 		panic("seq_out not found rack:%p tp:%p",
6087 		    bbr, tp);
6088 #endif
6089 	} else {
6090 #ifdef BBR_INVARIANTS
6091 		/*
6092 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
6093 		 * flag)
6094 		 */
6095 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6096 		    seq_out, len, tp->snd_max, tp);
6097 #endif
6098 	}
6099 }
6100 
6101 static void
6102 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt)
6103 {
6104 	/*
6105 	 * Collapse timeout back the cum-ack moved.
6106 	 */
6107 	tp->t_rxtshift = 0;
6108 	tp->t_softerror = 0;
6109 }
6110 
6111 static void
6112 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin)
6113 {
6114 	bbr->rtt_valid = 1;
6115 	bbr->r_ctl.cur_rtt = rtt_usecs;
6116 	bbr->r_ctl.ts_in = tsin;
6117 	if (rsm_send_time)
6118 		bbr->r_ctl.cur_rtt_send_time = rsm_send_time;
6119 }
6120 
6121 static void
6122 bbr_make_timestamp_determination(struct tcp_bbr *bbr)
6123 {
6124 	/**
6125 	 * We have in our bbr control:
6126 	 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp).
6127 	 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts).
6128 	 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts)
6129 	 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time)
6130 	 *
6131 	 * Now we can calculate the time between the sends by doing:
6132 	 *
6133 	 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts
6134 	 *
6135 	 * And the peer's time between receiving them by doing:
6136 	 *
6137 	 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp
6138 	 *
6139 	 * We want to figure out if the timestamp values are in msec, 10msec or usec.
6140 	 * We also may find that we can't use the timestamps if say we see
6141 	 * that the peer_delta indicates that though we may have taken 10ms to
6142 	 * pace out the data, it only saw 1ms between the two packets. This would
6143 	 * indicate that somewhere on the path is a batching entity that is giving
6144 	 * out time-slices of the actual b/w. This would mean we could not use
6145 	 * reliably the peers timestamps.
6146 	 *
6147 	 * We expect delta > peer_delta initially. Until we figure out the
6148 	 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio.
6149 	 * If we place 1000 there then its a ms vs our usec. If we place 10000 there
6150 	 * then its 10ms vs our usec. If the peer is running a usec clock we would
6151 	 * put a 1 there. If the value is faster then ours, we will disable the
6152 	 * use of timestamps (though we could revist this later if we find it to be not
6153 	 * just an isolated one or two flows)).
6154 	 *
6155 	 * To detect the batching middle boxes we will come up with our compensation and
6156 	 * if with it in place, we find the peer is drastically off (by some margin) in
6157 	 * the smaller direction, then we will assume the worst case and disable use of timestamps.
6158 	 *
6159 	 */
6160 	uint64_t delta, peer_delta, delta_up;
6161 
6162 	delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts;
6163 	if (delta < bbr_min_usec_delta) {
6164 		/*
6165 		 * Have not seen a min amount of time
6166 		 * between our send times so we can
6167 		 * make a determination of the timestamp
6168 		 * yet.
6169 		 */
6170 		return;
6171 	}
6172 	peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp;
6173 	if (peer_delta < bbr_min_peer_delta) {
6174 		/*
6175 		 * We may have enough in the form of
6176 		 * our delta but the peers number
6177 		 * has not changed that much. It could
6178 		 * be its clock ratio is such that
6179 		 * we need more data (10ms tick) or
6180 		 * there may be other compression scenarios
6181 		 * going on. In any event we need the
6182 		 * spread to be larger.
6183 		 */
6184 		return;
6185 	}
6186 	/* Ok lets first see which way our delta is going */
6187 	if (peer_delta > delta) {
6188 		/* Very unlikely, the peer without
6189 		 * compensation shows that it saw
6190 		 * the two sends arrive further apart
6191 		 * then we saw then in micro-seconds.
6192 		 */
6193 		if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) {
6194 			/* well it looks like the peer is a micro-second clock. */
6195 			bbr->rc_ts_clock_set = 1;
6196 			bbr->r_ctl.bbr_peer_tsratio = 1;
6197 		} else {
6198 			bbr->rc_ts_cant_be_used = 1;
6199 			bbr->rc_ts_clock_set = 1;
6200 		}
6201 		return;
6202 	}
6203 	/* Ok we know that the peer_delta is smaller than our send distance */
6204 	bbr->rc_ts_clock_set = 1;
6205 	/* First question is it within the percentage that they are using usec time? */
6206 	delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent;
6207 	if ((peer_delta + delta_up) >= delta) {
6208 		/* Its a usec clock */
6209 		bbr->r_ctl.bbr_peer_tsratio = 1;
6210 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6211 		return;
6212 	}
6213 	/* Ok if not usec, what about 10usec (though unlikely)? */
6214 	delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent;
6215 	if (((peer_delta * 10) + delta_up) >= delta) {
6216 		bbr->r_ctl.bbr_peer_tsratio = 10;
6217 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6218 		return;
6219 	}
6220 	/* And what about 100usec (though again unlikely)? */
6221 	delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent;
6222 	if (((peer_delta * 100) + delta_up) >= delta) {
6223 		bbr->r_ctl.bbr_peer_tsratio = 100;
6224 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6225 		return;
6226 	}
6227 	/* And how about 1 msec (the most likely one)? */
6228 	delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent;
6229 	if (((peer_delta * 1000) + delta_up) >= delta) {
6230 		bbr->r_ctl.bbr_peer_tsratio = 1000;
6231 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6232 		return;
6233 	}
6234 	/* Ok if not msec could it be 10 msec? */
6235 	delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent;
6236 	if (((peer_delta * 10000) + delta_up) >= delta) {
6237 		bbr->r_ctl.bbr_peer_tsratio = 10000;
6238 		return;
6239 	}
6240 	/* If we fall down here the clock tick so slowly we can't use it */
6241 	bbr->rc_ts_cant_be_used = 1;
6242 	bbr->r_ctl.bbr_peer_tsratio = 0;
6243 	bbr_log_tstmp_validation(bbr, peer_delta, delta);
6244 }
6245 
6246 /*
6247  * Collect new round-trip time estimate
6248  * and update averages and current timeout.
6249  */
6250 static void
6251 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts)
6252 {
6253 	int32_t delta;
6254 	uint32_t rtt, tsin;
6255 	int32_t rtt_ticks;
6256 
6257 	if (bbr->rtt_valid == 0)
6258 		/* No valid sample */
6259 		return;
6260 
6261 	rtt = bbr->r_ctl.cur_rtt;
6262 	tsin = bbr->r_ctl.ts_in;
6263 	if (bbr->rc_prtt_set_ts) {
6264 		/*
6265 		 * We are to force feed the rttProp filter due
6266 		 * to an entry into PROBE_RTT. This assures
6267 		 * that the times are sync'd between when we
6268 		 * go into PROBE_RTT and the filter expiration.
6269 		 *
6270 		 * Google does not use a true filter, so they do
6271 		 * this implicitly since they only keep one value
6272 		 * and when they enter probe-rtt they update the
6273 		 * value to the newest rtt.
6274 		 */
6275 		uint32_t rtt_prop;
6276 
6277 		bbr->rc_prtt_set_ts = 0;
6278 		rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
6279 		if (rtt > rtt_prop)
6280 			filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts);
6281 		else
6282 			apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6283 	}
6284 #ifdef STATS
6285 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rtt));
6286 #endif
6287 	if (bbr->rc_ack_was_delayed)
6288 		rtt += bbr->r_ctl.rc_ack_hdwr_delay;
6289 
6290 	if (rtt < bbr->r_ctl.rc_lowest_rtt)
6291 		bbr->r_ctl.rc_lowest_rtt = rtt;
6292 	bbr_log_rtt_sample(bbr, rtt, tsin);
6293 	if (bbr->r_init_rtt) {
6294 		/*
6295 		 * The initial rtt is not-trusted, nuke it and lets get
6296 		 * our first valid measurement in.
6297 		 */
6298 		bbr->r_init_rtt = 0;
6299 		tp->t_srtt = 0;
6300 	}
6301 	if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) {
6302 		/*
6303 		 * So we have not yet figured out
6304 		 * what the peers TSTMP value is
6305 		 * in (most likely ms). We need a
6306 		 * series of cum-ack's to determine
6307 		 * this reliably.
6308 		 */
6309 		if (bbr->rc_ack_is_cumack) {
6310 			if (bbr->rc_ts_data_set) {
6311 				/* Lets attempt to determine the timestamp granularity. */
6312 				bbr_make_timestamp_determination(bbr);
6313 			} else {
6314 				bbr->rc_ts_data_set = 1;
6315 				bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts;
6316 				bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time;
6317 			}
6318 		} else {
6319 			/*
6320 			 * We have to have consecutive acks
6321 			 * reset any "filled" state to none.
6322 			 */
6323 			bbr->rc_ts_data_set = 0;
6324 		}
6325 	}
6326 	/* Round it up */
6327 	rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1)));
6328 	if (rtt_ticks == 0)
6329 		rtt_ticks = 1;
6330 	if (tp->t_srtt != 0) {
6331 		/*
6332 		 * srtt is stored as fixed point with 5 bits after the
6333 		 * binary point (i.e., scaled by 8).  The following magic is
6334 		 * equivalent to the smoothing algorithm in rfc793 with an
6335 		 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6336 		 * Adjust rtt to origin 0.
6337 		 */
6338 
6339 		delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT)
6340 		    - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6341 
6342 		tp->t_srtt += delta;
6343 		if (tp->t_srtt <= 0)
6344 			tp->t_srtt = 1;
6345 
6346 		/*
6347 		 * We accumulate a smoothed rtt variance (actually, a
6348 		 * smoothed mean difference), then set the retransmit timer
6349 		 * to smoothed rtt + 4 times the smoothed variance. rttvar
6350 		 * is stored as fixed point with 4 bits after the binary
6351 		 * point (scaled by 16).  The following is equivalent to
6352 		 * rfc793 smoothing with an alpha of .75 (rttvar =
6353 		 * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
6354 		 * wired-in beta.
6355 		 */
6356 		if (delta < 0)
6357 			delta = -delta;
6358 		delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6359 		tp->t_rttvar += delta;
6360 		if (tp->t_rttvar <= 0)
6361 			tp->t_rttvar = 1;
6362 	} else {
6363 		/*
6364 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
6365 		 * variance to half the rtt (so our first retransmit happens
6366 		 * at 3*rtt).
6367 		 */
6368 		tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT;
6369 		tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1);
6370 	}
6371 	KMOD_TCPSTAT_INC(tcps_rttupdated);
6372 	if (tp->t_rttupdated < UCHAR_MAX)
6373 		tp->t_rttupdated++;
6374 #ifdef STATS
6375 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks));
6376 #endif
6377 	/*
6378 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
6379 	 * way we do the smoothing, srtt and rttvar will each average +1/2
6380 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
6381 	 * tick of rounding and 1 extra tick because of +-1/2 tick
6382 	 * uncertainty in the firing of the timer.  The bias will give us
6383 	 * exactly the 1.5 tick we need.  But, because the bias is
6384 	 * statistical, we have to test that we don't drop below the minimum
6385 	 * feasible timer (which is 2 ticks).
6386 	 */
6387 	TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6388 	    max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2),
6389 	    MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
6390 
6391 	/*
6392 	 * We received an ack for a packet that wasn't retransmitted; it is
6393 	 * probably safe to discard any error indications we've received
6394 	 * recently.  This isn't quite right, but close enough for now (a
6395 	 * route might have failed after we sent a segment, and the return
6396 	 * path might not be symmetrical).
6397 	 */
6398 	tp->t_softerror = 0;
6399 	rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
6400 	if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt)
6401 		bbr->r_ctl.bbr_smallest_srtt_this_state = rtt;
6402 }
6403 
6404 static void
6405 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line)
6406 {
6407 	bbr->r_ctl.rc_rtt_shrinks = cts;
6408 	if (bbr_can_force_probertt &&
6409 	    (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
6410 	    ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
6411 		/*
6412 		 * We should enter probe-rtt its been too long
6413 		 * since we have been there.
6414 		 */
6415 		bbr_enter_probe_rtt(bbr, cts, __LINE__);
6416 	} else
6417 		bbr_check_probe_rtt_limits(bbr, cts);
6418 }
6419 
6420 static void
6421 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts)
6422 {
6423 	uint64_t orig_bw;
6424 
6425 	if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) {
6426 		/* We never apply a zero measurement */
6427 		bbr_log_type_bbrupd(bbr, 20, cts, 0, 0,
6428 				    0, 0, 0, 0, 0, 0);
6429 		return;
6430 	}
6431 	if (bbr->r_ctl.r_measurement_count < 0xffffffff)
6432 		bbr->r_ctl.r_measurement_count++;
6433 	orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
6434 	apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch);
6435 	bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw,
6436 			    (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate),
6437 			    0, 0, 0, 0, 0, 0);
6438 	if (orig_bw &&
6439 	    (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) {
6440 		if (bbr->bbr_hdrw_pacing) {
6441 			/*
6442 			 * Apply a new rate to the hardware
6443 			 * possibly.
6444 			 */
6445 			bbr_update_hardware_pacing_rate(bbr, cts);
6446 		}
6447 		bbr_set_state_target(bbr, __LINE__);
6448 		tcp_bbr_tso_size_check(bbr, cts);
6449 		if (bbr->r_recovery_bw)  {
6450 			bbr_setup_red_bw(bbr, cts);
6451 			bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW);
6452 		}
6453 	} else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate))
6454 		tcp_bbr_tso_size_check(bbr, cts);
6455 }
6456 
6457 static void
6458 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6459 {
6460 	if (bbr->rc_in_persist == 0) {
6461 		/* We log only when not in persist */
6462 		/* Translate to a Bytes Per Second */
6463 		uint64_t tim, bw, ts_diff, ts_bw;
6464 		uint32_t delivered;
6465 
6466 		if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6467 			tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6468 		else
6469 			tim = 1;
6470 		/*
6471 		 * Now that we have processed the tim (skipping the sample
6472 		 * or possibly updating the time, go ahead and
6473 		 * calculate the cdr.
6474 		 */
6475 		delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6476 		bw = (uint64_t)delivered;
6477 		bw *= (uint64_t)USECS_IN_SECOND;
6478 		bw /= tim;
6479 		if (bw == 0) {
6480 			/* We must have a calculatable amount */
6481 			return;
6482 		}
6483 		/*
6484 		 * If we are using this b/w shove it in now so we
6485 		 * can see in the trace viewer if it gets over-ridden.
6486 		 */
6487 		if (rsm->r_ts_valid &&
6488 		    bbr->rc_ts_valid &&
6489 		    bbr->rc_ts_clock_set &&
6490 		    (bbr->rc_ts_cant_be_used == 0) &&
6491 		    bbr->rc_use_ts_limit) {
6492 			ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1);
6493 			ts_diff *= bbr->r_ctl.bbr_peer_tsratio;
6494 			if ((delivered == 0) ||
6495 			    (rtt < 1000)) {
6496 				/* Can't use the ts */
6497 				bbr_log_type_bbrupd(bbr, 61, cts,
6498 						    ts_diff,
6499 						    bbr->r_ctl.last_inbound_ts,
6500 						    rsm->r_del_ack_ts, 0,
6501 						    0, 0, 0, delivered);
6502 			} else {
6503 				ts_bw = (uint64_t)delivered;
6504 				ts_bw *= (uint64_t)USECS_IN_SECOND;
6505 				ts_bw /= ts_diff;
6506 				bbr_log_type_bbrupd(bbr, 62, cts,
6507 						    (ts_bw >> 32),
6508 						    (ts_bw & 0xffffffff), 0, 0,
6509 						    0, 0, ts_diff, delivered);
6510 				if ((bbr->ts_can_raise) &&
6511 				    (ts_bw > bw)) {
6512 					bbr_log_type_bbrupd(bbr, 8, cts,
6513 							    delivered,
6514 							    ts_diff,
6515 							    (bw >> 32),
6516 							    (bw & 0x00000000ffffffff),
6517 							    0, 0, 0, 0);
6518 					bw = ts_bw;
6519 				} else if (ts_bw && (ts_bw < bw)) {
6520 					bbr_log_type_bbrupd(bbr, 7, cts,
6521 							    delivered,
6522 							    ts_diff,
6523 							    (bw >> 32),
6524 							    (bw & 0x00000000ffffffff),
6525 							    0, 0, 0, 0);
6526 					bw = ts_bw;
6527 				}
6528 			}
6529 		}
6530 		if (rsm->r_first_sent_time &&
6531 		    TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6532 			uint64_t sbw, sti;
6533 			/*
6534 			 * We use what was in flight at the time of our
6535 			 * send  and the size of this send to figure
6536 			 * out what we have been sending at (amount).
6537 			 * For the time we take from the time of
6538 			 * the send of the first send outstanding
6539 			 * until this send plus this sends pacing
6540 			 * time. This gives us a good calculation
6541 			 * as to the rate we have been sending at.
6542 			 */
6543 
6544 			sbw = (uint64_t)(rsm->r_flight_at_send);
6545 			sbw *= (uint64_t)USECS_IN_SECOND;
6546 			sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6547 			sti += rsm->r_pacing_delay;
6548 			sbw /= sti;
6549 			if (sbw < bw) {
6550 				bbr_log_type_bbrupd(bbr, 6, cts,
6551 						    delivered,
6552 						    (uint32_t)sti,
6553 						    (bw >> 32),
6554 						    (uint32_t)bw,
6555 						    rsm->r_first_sent_time, 0, (sbw >> 32),
6556 						    (uint32_t)sbw);
6557 				bw = sbw;
6558 			}
6559 		}
6560 		/* Use the google algorithm for b/w measurements */
6561 		bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6562 		if ((rsm->r_app_limited == 0) ||
6563 		    (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) {
6564 			tcp_bbr_commit_bw(bbr, cts);
6565 			bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6566 					    0, 0, 0, 0,  bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6567 		}
6568 	}
6569 }
6570 
6571 static void
6572 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6573 {
6574 	if (bbr->rc_in_persist == 0) {
6575 		/* We log only when not in persist */
6576 		/* Translate to a Bytes Per Second */
6577 		uint64_t tim, bw;
6578 		uint32_t delivered;
6579 		int no_apply = 0;
6580 
6581 		if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6582 			tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6583 		else
6584 			tim = 1;
6585 		/*
6586 		 * Now that we have processed the tim (skipping the sample
6587 		 * or possibly updating the time, go ahead and
6588 		 * calculate the cdr.
6589 		 */
6590 		delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6591 		bw = (uint64_t)delivered;
6592 		bw *= (uint64_t)USECS_IN_SECOND;
6593 		bw /= tim;
6594 		if (tim < bbr->r_ctl.rc_lowest_rtt) {
6595 			bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6596 					    tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6597 
6598 			no_apply = 1;
6599 		}
6600 		/*
6601 		 * If we are using this b/w shove it in now so we
6602 		 * can see in the trace viewer if it gets over-ridden.
6603 		 */
6604 		bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6605 		/* Gate by the sending rate */
6606 		if (rsm->r_first_sent_time &&
6607 		    TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6608 			uint64_t sbw, sti;
6609 			/*
6610 			 * We use what was in flight at the time of our
6611 			 * send  and the size of this send to figure
6612 			 * out what we have been sending at (amount).
6613 			 * For the time we take from the time of
6614 			 * the send of the first send outstanding
6615 			 * until this send plus this sends pacing
6616 			 * time. This gives us a good calculation
6617 			 * as to the rate we have been sending at.
6618 			 */
6619 
6620 			sbw = (uint64_t)(rsm->r_flight_at_send);
6621 			sbw *= (uint64_t)USECS_IN_SECOND;
6622 			sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6623 			sti += rsm->r_pacing_delay;
6624 			sbw /= sti;
6625 			if (sbw < bw) {
6626 				bbr_log_type_bbrupd(bbr, 6, cts,
6627 						    delivered,
6628 						    (uint32_t)sti,
6629 						    (bw >> 32),
6630 						    (uint32_t)bw,
6631 						    rsm->r_first_sent_time, 0, (sbw >> 32),
6632 						    (uint32_t)sbw);
6633 				bw = sbw;
6634 			}
6635 			if ((sti > tim) &&
6636 			    (sti < bbr->r_ctl.rc_lowest_rtt)) {
6637 				bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6638 						    (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6639 				no_apply = 1;
6640 			} else
6641 				no_apply = 0;
6642 		}
6643 		bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6644 		if ((no_apply == 0) &&
6645 		    ((rsm->r_app_limited == 0) ||
6646 		     (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) {
6647 			tcp_bbr_commit_bw(bbr, cts);
6648 			bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6649 					    0, 0, 0, 0, bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6650 		}
6651 	}
6652 }
6653 
6654 static void
6655 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin,
6656     uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to)
6657 {
6658 	uint64_t old_rttprop;
6659 
6660 	/* Update our delivery time and amount */
6661 	bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start);
6662 	bbr->r_ctl.rc_del_time = cts;
6663 	if (rtt == 0) {
6664 		/*
6665 		 * 0 means its a retransmit, for now we don't use these for
6666 		 * the rest of BBR.
6667 		 */
6668 		return;
6669 	}
6670 	if ((bbr->rc_use_google == 0) &&
6671 	    (match != BBR_RTT_BY_EXACTMATCH) &&
6672 	    (match != BBR_RTT_BY_TIMESTAMP)){
6673 		/*
6674 		 * We get a lot of rtt updates, lets not pay attention to
6675 		 * any that are not an exact match. That way we don't have
6676 		 * to worry about timestamps and the whole nonsense of
6677 		 * unsure if its a retransmission etc (if we ever had the
6678 		 * timestamp fixed to always have the last thing sent this
6679 		 * would not be a issue).
6680 		 */
6681 		return;
6682 	}
6683 	if ((bbr_no_retran && bbr->rc_use_google) &&
6684 	    (match != BBR_RTT_BY_EXACTMATCH) &&
6685 	    (match != BBR_RTT_BY_TIMESTAMP)){
6686 		/*
6687 		 * We only do measurements in google mode
6688 		 * with bbr_no_retran on for sure things.
6689 		 */
6690 		return;
6691 	}
6692 	/* Only update srtt if we know by exact match */
6693 	tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin);
6694 	if (ack_type == BBR_CUM_ACKED)
6695 		bbr->rc_ack_is_cumack = 1;
6696 	else
6697 		bbr->rc_ack_is_cumack = 0;
6698 	old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP);
6699 	/*
6700 	 * Note the following code differs to the original
6701 	 * BBR spec. It calls for <= not <. However after a
6702 	 * long discussion in email with Neal, he acknowledged
6703 	 * that it should be < than so that we will have flows
6704 	 * going into probe-rtt (we were seeing cases where that
6705 	 * did not happen and caused ugly things to occur). We
6706 	 * have added this agreed upon fix to our code base.
6707 	 */
6708 	if (rtt < old_rttprop) {
6709 		/* Update when we last saw a rtt drop */
6710 		bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0);
6711 		bbr_set_reduced_rtt(bbr, cts, __LINE__);
6712 	}
6713 	bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts,
6714 	    match, rsm->r_start, rsm->r_flags);
6715 	apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6716 	if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) {
6717 		/*
6718 		 * The RTT-prop moved, reset the target (may be a
6719 		 * nop for some states).
6720 		 */
6721 		bbr_set_state_target(bbr, __LINE__);
6722 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)
6723 			bbr_log_rtt_shrinks(bbr, cts, 0, 0,
6724 					    __LINE__, BBR_RTTS_NEW_TARGET, 0);
6725 		else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP))
6726 			/* It went up */
6727 			bbr_check_probe_rtt_limits(bbr, cts);
6728 	}
6729 	if ((bbr->rc_use_google == 0) &&
6730 	    (match == BBR_RTT_BY_TIMESTAMP)) {
6731 		/*
6732 		 * We don't do b/w update with
6733 		 * these since they are not really
6734 		 * reliable.
6735 		 */
6736 		return;
6737 	}
6738 	if (bbr->r_ctl.r_app_limited_until &&
6739 	    (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) {
6740 		/* We are no longer app-limited */
6741 		bbr->r_ctl.r_app_limited_until = 0;
6742 	}
6743 	if (bbr->rc_use_google) {
6744 		bbr_google_measurement(bbr, rsm, rtt, cts);
6745 	} else {
6746 		bbr_nf_measurement(bbr, rsm, rtt, cts);
6747 	}
6748 }
6749 
6750 /*
6751  * Convert a timestamp that the main stack
6752  * uses (milliseconds) into one that bbr uses
6753  * (microseconds). Return that converted timestamp.
6754  */
6755 static uint32_t
6756 bbr_ts_convert(uint32_t cts) {
6757 	uint32_t sec, msec;
6758 
6759 	sec = cts / MS_IN_USEC;
6760 	msec = cts - (MS_IN_USEC * sec);
6761 	return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC));
6762 }
6763 
6764 /*
6765  * Return 0 if we did not update the RTT time, return
6766  * 1 if we did.
6767  */
6768 static int
6769 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr,
6770     struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack)
6771 {
6772 	int32_t i;
6773 	uint32_t t, uts = 0;
6774 
6775 	if ((rsm->r_flags & BBR_ACKED) ||
6776 	    (rsm->r_flags & BBR_WAS_RENEGED) ||
6777 	    (rsm->r_flags & BBR_RXT_CLEARED)) {
6778 		/* Already done */
6779 		return (0);
6780 	}
6781 	if (rsm->r_rtt_not_allowed) {
6782 		/* Not allowed */
6783 		return (0);
6784 	}
6785 	if (rsm->r_rtr_cnt == 1) {
6786 		/*
6787 		 * Only one transmit. Hopefully the normal case.
6788 		 */
6789 		if (TSTMP_GT(cts, rsm->r_tim_lastsent[0]))
6790 			t = cts - rsm->r_tim_lastsent[0];
6791 		else
6792 			t = 1;
6793 		if ((int)t <= 0)
6794 			t = 1;
6795 		bbr->r_ctl.rc_last_rtt = t;
6796 		bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6797 				    BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to);
6798 		return (1);
6799 	}
6800 	/* Convert to usecs */
6801 	if ((bbr_can_use_ts_for_rtt == 1) &&
6802 	    (bbr->rc_use_google == 1) &&
6803 	    (ack_type == BBR_CUM_ACKED) &&
6804 	    (to->to_flags & TOF_TS) &&
6805 	    (to->to_tsecr != 0)) {
6806 		t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr;
6807 		if (t < 1)
6808 			t = 1;
6809 		t *= MS_IN_USEC;
6810 		bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6811 				    BBR_RTT_BY_TIMESTAMP,
6812 				    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)],
6813 				    ack_type, to);
6814 		return (1);
6815 	}
6816 	uts = bbr_ts_convert(to->to_tsecr);
6817 	if ((to->to_flags & TOF_TS) &&
6818 	    (to->to_tsecr != 0) &&
6819 	    (ack_type == BBR_CUM_ACKED) &&
6820 	    ((rsm->r_flags & BBR_OVERMAX) == 0)) {
6821 		/*
6822 		 * Now which timestamp does it match? In this block the ACK
6823 		 * may be coming from a previous transmission.
6824 		 */
6825 		uint32_t fudge;
6826 
6827 		fudge = BBR_TIMER_FUDGE;
6828 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
6829 			if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) &&
6830 			    (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) {
6831 				if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6832 					t = cts - rsm->r_tim_lastsent[i];
6833 				else
6834 					t = 1;
6835 				if ((int)t <= 0)
6836 					t = 1;
6837 				bbr->r_ctl.rc_last_rtt = t;
6838 				bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING,
6839 						    rsm->r_tim_lastsent[i], ack_type, to);
6840 				if ((i + 1) < rsm->r_rtr_cnt) {
6841 					/* Likely */
6842 					return (0);
6843 				} else if (rsm->r_flags & BBR_TLP) {
6844 					bbr->rc_tlp_rtx_out = 0;
6845 				}
6846 				return (1);
6847 			}
6848 		}
6849 		/* Fall through if we can't find a matching timestamp */
6850 	}
6851 	/*
6852 	 * Ok its a SACK block that we retransmitted. or a windows
6853 	 * machine without timestamps. We can tell nothing from the
6854 	 * time-stamp since its not there or the time the peer last
6855 	 * recieved a segment that moved forward its cum-ack point.
6856 	 *
6857 	 * Lets look at the last retransmit and see what we can tell
6858 	 * (with BBR for space we only keep 2 note we have to keep
6859 	 * at least 2 so the map can not be condensed more).
6860 	 */
6861 	i = rsm->r_rtr_cnt - 1;
6862 	if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6863 		t = cts - rsm->r_tim_lastsent[i];
6864 	else
6865 		goto not_sure;
6866 	if (t < bbr->r_ctl.rc_lowest_rtt) {
6867 		/*
6868 		 * We retransmitted and the ack came back in less
6869 		 * than the smallest rtt we have observed in the
6870 		 * windowed rtt. We most likey did an improper
6871 		 * retransmit as outlined in 4.2 Step 3 point 2 in
6872 		 * the rack-draft.
6873 		 *
6874 		 * Use the prior transmission to update all the
6875 		 * information as long as there is only one prior
6876 		 * transmission.
6877 		 */
6878 		if ((rsm->r_flags & BBR_OVERMAX) == 0) {
6879 #ifdef BBR_INVARIANTS
6880 			if (rsm->r_rtr_cnt == 1)
6881 				panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags);
6882 #endif
6883 			i = rsm->r_rtr_cnt - 2;
6884 			if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6885 				t = cts - rsm->r_tim_lastsent[i];
6886 			else
6887 				t = 1;
6888 			bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET,
6889 					    rsm->r_tim_lastsent[i], ack_type, to);
6890 			return (0);
6891 		} else {
6892 			/*
6893 			 * Too many prior transmissions, just
6894 			 * updated BBR delivered
6895 			 */
6896 not_sure:
6897 			bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6898 					    BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
6899 		}
6900 	} else {
6901 		/*
6902 		 * We retransmitted it and the retransmit did the
6903 		 * job.
6904 		 */
6905 		if (rsm->r_flags & BBR_TLP)
6906 			bbr->rc_tlp_rtx_out = 0;
6907 		if ((rsm->r_flags & BBR_OVERMAX) == 0)
6908 			bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts,
6909 					    BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to);
6910 		else
6911 			bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6912 					    BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
6913 		return (1);
6914 	}
6915 	return (0);
6916 }
6917 
6918 /*
6919  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
6920  */
6921 static void
6922 bbr_log_sack_passed(struct tcpcb *tp,
6923     struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
6924 {
6925 	struct bbr_sendmap *nrsm;
6926 
6927 	nrsm = rsm;
6928 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap,
6929 	    bbr_head, r_tnext) {
6930 		if (nrsm == rsm) {
6931 			/* Skip original segment he is acked */
6932 			continue;
6933 		}
6934 		if (nrsm->r_flags & BBR_ACKED) {
6935 			/* Skip ack'd segments */
6936 			continue;
6937 		}
6938 		if (nrsm->r_flags & BBR_SACK_PASSED) {
6939 			/*
6940 			 * We found one that is already marked
6941 			 * passed, we have been here before and
6942 			 * so all others below this are marked.
6943 			 */
6944 			break;
6945 		}
6946 		BBR_STAT_INC(bbr_sack_passed);
6947 		nrsm->r_flags |= BBR_SACK_PASSED;
6948 		if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) &&
6949 		    bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) {
6950 			bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start;
6951 			bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start;
6952 			nrsm->r_flags |= BBR_MARKED_LOST;
6953 		}
6954 		nrsm->r_flags &= ~BBR_WAS_SACKPASS;
6955 	}
6956 }
6957 
6958 /*
6959  * Returns the number of bytes that were
6960  * newly ack'd by sack blocks.
6961  */
6962 static uint32_t
6963 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack,
6964     struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts)
6965 {
6966 	int32_t times = 0;
6967 	uint32_t start, end, changed = 0;
6968 	struct bbr_sendmap *rsm, *nrsm;
6969 	int32_t used_ref = 1;
6970 	uint8_t went_back = 0, went_fwd = 0;
6971 
6972 	start = sack->start;
6973 	end = sack->end;
6974 	rsm = *prsm;
6975 	if (rsm == NULL)
6976 		used_ref = 0;
6977 
6978 	/* Do we locate the block behind where we last were? */
6979 	if (rsm && SEQ_LT(start, rsm->r_start)) {
6980 		went_back = 1;
6981 		TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
6982 			if (SEQ_GEQ(start, rsm->r_start) &&
6983 			    SEQ_LT(start, rsm->r_end)) {
6984 				goto do_rest_ofb;
6985 			}
6986 		}
6987 	}
6988 start_at_beginning:
6989 	went_fwd = 1;
6990 	/*
6991 	 * Ok lets locate the block where this guy is fwd from rsm (if its
6992 	 * set)
6993 	 */
6994 	TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) {
6995 		if (SEQ_GEQ(start, rsm->r_start) &&
6996 		    SEQ_LT(start, rsm->r_end)) {
6997 			break;
6998 		}
6999 	}
7000 do_rest_ofb:
7001 	if (rsm == NULL) {
7002 		/*
7003 		 * This happens when we get duplicate sack blocks with the
7004 		 * same end. For example SACK 4: 100 SACK 3: 100 The sort
7005 		 * will not change there location so we would just start at
7006 		 * the end of the first one and get lost.
7007 		 */
7008 		if (tp->t_flags & TF_SENTFIN) {
7009 			/*
7010 			 * Check to see if we have not logged the FIN that
7011 			 * went out.
7012 			 */
7013 			nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7014 			if (nrsm && (nrsm->r_end + 1) == tp->snd_max) {
7015 				/*
7016 				 * Ok we did not get the FIN logged.
7017 				 */
7018 				nrsm->r_end++;
7019 				rsm = nrsm;
7020 				goto do_rest_ofb;
7021 			}
7022 		}
7023 		if (times == 1) {
7024 #ifdef BBR_INVARIANTS
7025 			panic("tp:%p bbr:%p sack:%p to:%p prsm:%p",
7026 			    tp, bbr, sack, to, prsm);
7027 #else
7028 			goto out;
7029 #endif
7030 		}
7031 		times++;
7032 		BBR_STAT_INC(bbr_sack_proc_restart);
7033 		rsm = NULL;
7034 		goto start_at_beginning;
7035 	}
7036 	/* Ok we have an ACK for some piece of rsm */
7037 	if (rsm->r_start != start) {
7038 		/*
7039 		 * Need to split this in two pieces the before and after.
7040 		 */
7041 		if (bbr_sack_mergable(rsm, start, end))
7042 			nrsm = bbr_alloc_full_limit(bbr);
7043 		else
7044 			nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7045 		if (nrsm == NULL) {
7046 			/* We could not allocate ignore the sack */
7047 			struct sackblk blk;
7048 
7049 			blk.start = start;
7050 			blk.end = end;
7051 			sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7052 			goto out;
7053 		}
7054 		bbr_clone_rsm(bbr, nrsm, rsm, start);
7055 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7056 		if (rsm->r_in_tmap) {
7057 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7058 			nrsm->r_in_tmap = 1;
7059 		}
7060 		rsm->r_flags &= (~BBR_HAS_FIN);
7061 		rsm = nrsm;
7062 	}
7063 	if (SEQ_GEQ(end, rsm->r_end)) {
7064 		/*
7065 		 * The end of this block is either beyond this guy or right
7066 		 * at this guy.
7067 		 */
7068 		if ((rsm->r_flags & BBR_ACKED) == 0) {
7069 			bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7070 			changed += (rsm->r_end - rsm->r_start);
7071 			bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7072 			bbr_log_sack_passed(tp, bbr, rsm);
7073 			if (rsm->r_flags & BBR_MARKED_LOST) {
7074 				bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7075 			}
7076 			/* Is Reordering occuring? */
7077 			if (rsm->r_flags & BBR_SACK_PASSED) {
7078 				BBR_STAT_INC(bbr_reorder_seen);
7079 				bbr->r_ctl.rc_reorder_ts = cts;
7080 				if (rsm->r_flags & BBR_MARKED_LOST) {
7081 					bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7082 					if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7083 						/* LT sampling also needs adjustment */
7084 						bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7085 				}
7086 			}
7087 			rsm->r_flags |= BBR_ACKED;
7088 			rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7089 			if (rsm->r_in_tmap) {
7090 				TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7091 				rsm->r_in_tmap = 0;
7092 			}
7093 		}
7094 		bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7095 		if (end == rsm->r_end) {
7096 			/* This block only - done */
7097 			goto out;
7098 		}
7099 		/* There is more not coverend by this rsm move on */
7100 		start = rsm->r_end;
7101 		nrsm = TAILQ_NEXT(rsm, r_next);
7102 		rsm = nrsm;
7103 		times = 0;
7104 		goto do_rest_ofb;
7105 	}
7106 	if (rsm->r_flags & BBR_ACKED) {
7107 		/* Been here done that */
7108 		goto out;
7109 	}
7110 	/* Ok we need to split off this one at the tail */
7111 	if (bbr_sack_mergable(rsm, start, end))
7112 		nrsm = bbr_alloc_full_limit(bbr);
7113 	else
7114 		nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7115 	if (nrsm == NULL) {
7116 		/* failed XXXrrs what can we do but loose the sack info? */
7117 		struct sackblk blk;
7118 
7119 		blk.start = start;
7120 		blk.end = end;
7121 		sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7122 		goto out;
7123 	}
7124 	/* Clone it */
7125 	bbr_clone_rsm(bbr, nrsm, rsm, end);
7126 	/* The sack block does not cover this guy fully */
7127 	rsm->r_flags &= (~BBR_HAS_FIN);
7128 	TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7129 	if (rsm->r_in_tmap) {
7130 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7131 		nrsm->r_in_tmap = 1;
7132 	}
7133 	nrsm->r_dupack = 0;
7134 	bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7135 	bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7136 	changed += (rsm->r_end - rsm->r_start);
7137 	bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7138 	bbr_log_sack_passed(tp, bbr, rsm);
7139 	/* Is Reordering occuring? */
7140 	if (rsm->r_flags & BBR_MARKED_LOST) {
7141 		bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7142 	}
7143 	if (rsm->r_flags & BBR_SACK_PASSED) {
7144 		BBR_STAT_INC(bbr_reorder_seen);
7145 		bbr->r_ctl.rc_reorder_ts = cts;
7146 		if (rsm->r_flags & BBR_MARKED_LOST) {
7147 			bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7148 			if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7149 				/* LT sampling also needs adjustment */
7150 				bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7151 		}
7152 	}
7153 	rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7154 	rsm->r_flags |= BBR_ACKED;
7155 	if (rsm->r_in_tmap) {
7156 		TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7157 		rsm->r_in_tmap = 0;
7158 	}
7159 out:
7160 	if (rsm && (rsm->r_flags & BBR_ACKED)) {
7161 		/*
7162 		 * Now can we merge this newly acked
7163 		 * block with either the previous or
7164 		 * next block?
7165 		 */
7166 		nrsm = TAILQ_NEXT(rsm, r_next);
7167 		if (nrsm &&
7168 		    (nrsm->r_flags & BBR_ACKED)) {
7169 			/* yep this and next can be merged */
7170 			rsm = bbr_merge_rsm(bbr, rsm, nrsm);
7171 		}
7172 		/* Now what about the previous? */
7173 		nrsm = TAILQ_PREV(rsm, bbr_head, r_next);
7174 		if (nrsm &&
7175 		    (nrsm->r_flags & BBR_ACKED)) {
7176 			/* yep the previous and this can be merged */
7177 			rsm = bbr_merge_rsm(bbr, nrsm, rsm);
7178 		}
7179 	}
7180 	if (used_ref == 0) {
7181 		BBR_STAT_INC(bbr_sack_proc_all);
7182 	} else {
7183 		BBR_STAT_INC(bbr_sack_proc_short);
7184 	}
7185 	if (went_fwd && went_back) {
7186 		BBR_STAT_INC(bbr_sack_search_both);
7187 	} else if (went_fwd) {
7188 		BBR_STAT_INC(bbr_sack_search_fwd);
7189 	} else if (went_back) {
7190 		BBR_STAT_INC(bbr_sack_search_back);
7191 	}
7192 	/* Save off where the next seq is */
7193 	if (rsm)
7194 		bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next);
7195 	else
7196 		bbr->r_ctl.rc_sacklast = NULL;
7197 	*prsm = rsm;
7198 	return (changed);
7199 }
7200 
7201 static void inline
7202 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack)
7203 {
7204 	struct bbr_sendmap *tmap;
7205 
7206 	BBR_STAT_INC(bbr_reneges_seen);
7207 	tmap = NULL;
7208 	while (rsm && (rsm->r_flags & BBR_ACKED)) {
7209 		/* Its no longer sacked, mark it so */
7210 		uint32_t oflags;
7211 		bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7212 #ifdef BBR_INVARIANTS
7213 		if (rsm->r_in_tmap) {
7214 			panic("bbr:%p rsm:%p flags:0x%x in tmap?",
7215 			    bbr, rsm, rsm->r_flags);
7216 		}
7217 #endif
7218 		oflags = rsm->r_flags;
7219 		if (rsm->r_flags & BBR_MARKED_LOST) {
7220 			bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7221 			bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7222 			if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7223 				/* LT sampling also needs adjustment */
7224 				bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7225 		}
7226 		rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST);
7227 		rsm->r_flags |= BBR_WAS_RENEGED;
7228 		rsm->r_flags |= BBR_RXT_CLEARED;
7229 		bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__);
7230 		/* Rebuild it into our tmap */
7231 		if (tmap == NULL) {
7232 			TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7233 			tmap = rsm;
7234 		} else {
7235 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7236 			tmap = rsm;
7237 		}
7238 		tmap->r_in_tmap = 1;
7239 		/*
7240 		 * XXXrrs Delivered? Should we do anything here?
7241 		 *
7242 		 * Of course we don't on a rxt timeout so maybe its ok that
7243 		 * we don't?
7244 		 *
7245 		 * For now lets not.
7246 		 */
7247 		rsm = TAILQ_NEXT(rsm, r_next);
7248 	}
7249 	/*
7250 	 * Now lets possibly clear the sack filter so we start recognizing
7251 	 * sacks that cover this area.
7252 	 */
7253 	sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack);
7254 }
7255 
7256 static void
7257 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to)
7258 {
7259 	struct tcp_bbr *bbr;
7260 	struct bbr_sendmap *rsm;
7261 	uint32_t cts;
7262 
7263 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7264 	cts = bbr->r_ctl.rc_rcvtime;
7265 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7266 	if (rsm && (rsm->r_flags & BBR_HAS_SYN)) {
7267 		if ((rsm->r_end - rsm->r_start) <= 1) {
7268 			/* Log out the SYN completely */
7269 			bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7270 			rsm->r_rtr_bytes = 0;
7271 			TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7272 			if (rsm->r_in_tmap) {
7273 				TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7274 				rsm->r_in_tmap = 0;
7275 			}
7276 			if (bbr->r_ctl.rc_next == rsm) {
7277 				/* scoot along the marker */
7278 				bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7279 			}
7280 			if (to != NULL)
7281 				bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0);
7282 			bbr_free(bbr, rsm);
7283 		} else {
7284 			/* There is more (Fast open)? strip out SYN. */
7285 			rsm->r_flags &= ~BBR_HAS_SYN;
7286 			rsm->r_start++;
7287 		}
7288 	}
7289 }
7290 
7291 /*
7292  * Returns the number of bytes that were
7293  * acknowledged by SACK blocks.
7294  */
7295 
7296 static uint32_t
7297 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th,
7298     uint32_t *prev_acked)
7299 {
7300 	uint32_t changed, last_seq, entered_recovery = 0;
7301 	struct tcp_bbr *bbr;
7302 	struct bbr_sendmap *rsm;
7303 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7304 	register uint32_t th_ack;
7305 	int32_t i, j, k, new_sb, num_sack_blks = 0;
7306 	uint32_t cts, acked, ack_point, sack_changed = 0;
7307 	uint32_t p_maxseg, maxseg, p_acked = 0;
7308 
7309 	INP_WLOCK_ASSERT(tptoinpcb(tp));
7310 	if (tcp_get_flags(th) & TH_RST) {
7311 		/* We don't log resets */
7312 		return (0);
7313 	}
7314 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7315 	cts = bbr->r_ctl.rc_rcvtime;
7316 
7317 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7318 	changed = 0;
7319 	maxseg = tp->t_maxseg - bbr->rc_last_options;
7320 	p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg);
7321 	th_ack = th->th_ack;
7322 	if (SEQ_GT(th_ack, tp->snd_una)) {
7323 		acked = th_ack - tp->snd_una;
7324 		bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__);
7325 		bbr->rc_tp->t_acktime = ticks;
7326 	} else
7327 		acked = 0;
7328 	if (SEQ_LEQ(th_ack, tp->snd_una)) {
7329 		/* Only sent here for sack processing */
7330 		goto proc_sack;
7331 	}
7332 	if (rsm && SEQ_GT(th_ack, rsm->r_start)) {
7333 		changed = th_ack - rsm->r_start;
7334 	} else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) {
7335 		/*
7336 		 * For the SYN incoming case we will not have called
7337 		 * tcp_output for the sending of the SYN, so there will be
7338 		 * no map. All other cases should probably be a panic.
7339 		 */
7340 		if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) {
7341 			/*
7342 			 * We have a timestamp that can be used to generate
7343 			 * an initial RTT.
7344 			 */
7345 			uint32_t ts, now, rtt;
7346 
7347 			ts = bbr_ts_convert(to->to_tsecr);
7348 			now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv));
7349 			rtt = now - ts;
7350 			if (rtt < 1)
7351 				rtt = 1;
7352 			bbr_log_type_bbrrttprop(bbr, rtt,
7353 						tp->iss, 0, cts,
7354 						BBR_RTT_BY_TIMESTAMP, tp->iss, 0);
7355 			apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
7356 			changed = 1;
7357 			bbr->r_wanted_output = 1;
7358 			goto out;
7359 		}
7360 		goto proc_sack;
7361 	} else if (rsm == NULL) {
7362 		goto out;
7363 	}
7364 	if (changed) {
7365 		/*
7366 		 * The ACK point is advancing to th_ack, we must drop off
7367 		 * the packets in the rack log and calculate any eligble
7368 		 * RTT's.
7369 		 */
7370 		bbr->r_wanted_output = 1;
7371 more:
7372 		if (rsm == NULL) {
7373 			if (tp->t_flags & TF_SENTFIN) {
7374 				/* if we send a FIN we will not hav a map */
7375 				goto proc_sack;
7376 			}
7377 #ifdef BBR_INVARIANTS
7378 			panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n",
7379 			    tp,
7380 			    th, tp->t_state, bbr,
7381 			    tp->snd_una, tp->snd_max, changed);
7382 #endif
7383 			goto proc_sack;
7384 		}
7385 	}
7386 	if (SEQ_LT(th_ack, rsm->r_start)) {
7387 		/* Huh map is missing this */
7388 #ifdef BBR_INVARIANTS
7389 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n",
7390 		    rsm->r_start,
7391 		    th_ack, tp->t_state,
7392 		    bbr->r_state, bbr);
7393 		panic("th-ack is bad bbr:%p tp:%p", bbr, tp);
7394 #endif
7395 		goto proc_sack;
7396 	} else if (th_ack == rsm->r_start) {
7397 		/* None here to ack */
7398 		goto proc_sack;
7399 	}
7400 	/*
7401 	 * Clear the dup ack counter, it will
7402 	 * either be freed or if there is some
7403 	 * remaining we need to start it at zero.
7404 	 */
7405 	rsm->r_dupack = 0;
7406 	/* Now do we consume the whole thing? */
7407 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
7408 		/* Its all consumed. */
7409 		uint32_t left;
7410 
7411 		if (rsm->r_flags & BBR_ACKED) {
7412 			/*
7413 			 * It was acked on the scoreboard -- remove it from
7414 			 * total
7415 			 */
7416 			p_acked += (rsm->r_end - rsm->r_start);
7417 			bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7418 			if (bbr->r_ctl.rc_sacked == 0)
7419 				bbr->r_ctl.rc_sacklast = NULL;
7420 		} else {
7421 			bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack);
7422 			if (rsm->r_flags & BBR_MARKED_LOST) {
7423 				bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7424 			}
7425 			if (rsm->r_flags & BBR_SACK_PASSED) {
7426 				/*
7427 				 * There are acked segments ACKED on the
7428 				 * scoreboard further up. We are seeing
7429 				 * reordering.
7430 				 */
7431 				BBR_STAT_INC(bbr_reorder_seen);
7432 				bbr->r_ctl.rc_reorder_ts = cts;
7433 				if (rsm->r_flags & BBR_MARKED_LOST) {
7434 					bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7435 					if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7436 						/* LT sampling also needs adjustment */
7437 						bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7438 				}
7439 			}
7440 			rsm->r_flags &= ~BBR_MARKED_LOST;
7441 		}
7442 		bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7443 		rsm->r_rtr_bytes = 0;
7444 		TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7445 		if (rsm->r_in_tmap) {
7446 			TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7447 			rsm->r_in_tmap = 0;
7448 		}
7449 		if (bbr->r_ctl.rc_next == rsm) {
7450 			/* scoot along the marker */
7451 			bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7452 		}
7453 		bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7454 		/* Adjust the packet counts */
7455 		left = th_ack - rsm->r_end;
7456 		/* Free back to zone */
7457 		bbr_free(bbr, rsm);
7458 		if (left) {
7459 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7460 			goto more;
7461 		}
7462 		goto proc_sack;
7463 	}
7464 	if (rsm->r_flags & BBR_ACKED) {
7465 		/*
7466 		 * It was acked on the scoreboard -- remove it from total
7467 		 * for the part being cum-acked.
7468 		 */
7469 		p_acked += (rsm->r_end - rsm->r_start);
7470 		bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7471 		if (bbr->r_ctl.rc_sacked == 0)
7472 			bbr->r_ctl.rc_sacklast = NULL;
7473 	} else {
7474 		/*
7475 		 * It was acked up to th_ack point for the first time
7476 		 */
7477 		struct bbr_sendmap lrsm;
7478 
7479 		memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap));
7480 		lrsm.r_end = th_ack;
7481 		bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack);
7482 	}
7483 	if ((rsm->r_flags & BBR_MARKED_LOST) &&
7484 	    ((rsm->r_flags & BBR_ACKED) == 0)) {
7485 		/*
7486 		 * It was marked lost and partly ack'd now
7487 		 * for the first time. We lower the rc_lost_bytes
7488 		 * and still leave it MARKED.
7489 		 */
7490 		bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start;
7491 	}
7492 	bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7493 	bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7494 	rsm->r_rtr_bytes = 0;
7495 	/* adjust packet count */
7496 	rsm->r_start = th_ack;
7497 proc_sack:
7498 	/* Check for reneging */
7499 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7500 	if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) {
7501 		/*
7502 		 * The peer has moved snd_una up to the edge of this send,
7503 		 * i.e. one that it had previously acked. The only way that
7504 		 * can be true if the peer threw away data (space issues)
7505 		 * that it had previously sacked (else it would have given
7506 		 * us snd_una up to (rsm->r_end). We need to undo the acked
7507 		 * markings here.
7508 		 *
7509 		 * Note we have to look to make sure th_ack is our
7510 		 * rsm->r_start in case we get an old ack where th_ack is
7511 		 * behind snd_una.
7512 		 */
7513 		bbr_peer_reneges(bbr, rsm, th->th_ack);
7514 	}
7515 	if ((to->to_flags & TOF_SACK) == 0) {
7516 		/* We are done nothing left to log */
7517 		goto out;
7518 	}
7519 	rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7520 	if (rsm) {
7521 		last_seq = rsm->r_end;
7522 	} else {
7523 		last_seq = tp->snd_max;
7524 	}
7525 	/* Sack block processing */
7526 	if (SEQ_GT(th_ack, tp->snd_una))
7527 		ack_point = th_ack;
7528 	else
7529 		ack_point = tp->snd_una;
7530 	for (i = 0; i < to->to_nsacks; i++) {
7531 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
7532 		    &sack, sizeof(sack));
7533 		sack.start = ntohl(sack.start);
7534 		sack.end = ntohl(sack.end);
7535 		if (SEQ_GT(sack.end, sack.start) &&
7536 		    SEQ_GT(sack.start, ack_point) &&
7537 		    SEQ_LT(sack.start, tp->snd_max) &&
7538 		    SEQ_GT(sack.end, ack_point) &&
7539 		    SEQ_LEQ(sack.end, tp->snd_max)) {
7540 			if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) &&
7541 			    (SEQ_LT(sack.end, last_seq)) &&
7542 			    ((sack.end - sack.start) < (p_maxseg / 8))) {
7543 				/*
7544 				 * Not the last piece and its smaller than
7545 				 * 1/8th of a p_maxseg. We ignore this.
7546 				 */
7547 				BBR_STAT_INC(bbr_runt_sacks);
7548 				continue;
7549 			}
7550 			sack_blocks[num_sack_blks] = sack;
7551 			num_sack_blks++;
7552 		} else if (SEQ_LEQ(sack.start, th_ack) &&
7553 		    SEQ_LEQ(sack.end, th_ack)) {
7554 			/*
7555 			 * Its a D-SACK block.
7556 			 */
7557 			tcp_record_dsack(tp, sack.start, sack.end, 0);
7558 		}
7559 	}
7560 	if (num_sack_blks == 0)
7561 		goto out;
7562 	/*
7563 	 * Sort the SACK blocks so we can update the rack scoreboard with
7564 	 * just one pass.
7565 	 */
7566 	new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks,
7567 				  num_sack_blks, th->th_ack);
7568 	ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks);
7569 	BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks);
7570 	BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb));
7571 	num_sack_blks = new_sb;
7572 	if (num_sack_blks < 2) {
7573 		goto do_sack_work;
7574 	}
7575 	/* Sort the sacks */
7576 	for (i = 0; i < num_sack_blks; i++) {
7577 		for (j = i + 1; j < num_sack_blks; j++) {
7578 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7579 				sack = sack_blocks[i];
7580 				sack_blocks[i] = sack_blocks[j];
7581 				sack_blocks[j] = sack;
7582 			}
7583 		}
7584 	}
7585 	/*
7586 	 * Now are any of the sack block ends the same (yes some
7587 	 * implememtations send these)?
7588 	 */
7589 again:
7590 	if (num_sack_blks > 1) {
7591 		for (i = 0; i < num_sack_blks; i++) {
7592 			for (j = i + 1; j < num_sack_blks; j++) {
7593 				if (sack_blocks[i].end == sack_blocks[j].end) {
7594 					/*
7595 					 * Ok these two have the same end we
7596 					 * want the smallest end and then
7597 					 * throw away the larger and start
7598 					 * again.
7599 					 */
7600 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7601 						/*
7602 						 * The second block covers
7603 						 * more area use that
7604 						 */
7605 						sack_blocks[i].start = sack_blocks[j].start;
7606 					}
7607 					/*
7608 					 * Now collapse out the dup-sack and
7609 					 * lower the count
7610 					 */
7611 					for (k = (j + 1); k < num_sack_blks; k++) {
7612 						sack_blocks[j].start = sack_blocks[k].start;
7613 						sack_blocks[j].end = sack_blocks[k].end;
7614 						j++;
7615 					}
7616 					num_sack_blks--;
7617 					goto again;
7618 				}
7619 			}
7620 		}
7621 	}
7622 do_sack_work:
7623 	rsm = bbr->r_ctl.rc_sacklast;
7624 	for (i = 0; i < num_sack_blks; i++) {
7625 		acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts);
7626 		if (acked) {
7627 			bbr->r_wanted_output = 1;
7628 			changed += acked;
7629 			sack_changed += acked;
7630 		}
7631 	}
7632 out:
7633 	*prev_acked = p_acked;
7634 	if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) {
7635 		/*
7636 		 * Ok we have a high probability that we need to go in to
7637 		 * recovery since we have data sack'd
7638 		 */
7639 		struct bbr_sendmap *rsm;
7640 
7641 		rsm = bbr_check_recovery_mode(tp, bbr, cts);
7642 		if (rsm) {
7643 			/* Enter recovery */
7644 			entered_recovery = 1;
7645 			bbr->r_wanted_output = 1;
7646 			/*
7647 			 * When we enter recovery we need to assure we send
7648 			 * one packet.
7649 			 */
7650 			if (bbr->r_ctl.rc_resend == NULL) {
7651 				bbr->r_ctl.rc_resend = rsm;
7652 			}
7653 		}
7654 	}
7655 	if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) {
7656 		/*
7657 		 * See if we need to rack-retransmit anything if so set it
7658 		 * up as the thing to resend assuming something else is not
7659 		 * already in that position.
7660 		 */
7661 		if (bbr->r_ctl.rc_resend == NULL) {
7662 			bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
7663 		}
7664 	}
7665 	/*
7666 	 * We return the amount that changed via sack, this is used by the
7667 	 * ack-received code to augment what was changed between th_ack <->
7668 	 * snd_una.
7669 	 */
7670 	return (sack_changed);
7671 }
7672 
7673 static void
7674 bbr_strike_dupack(struct tcp_bbr *bbr)
7675 {
7676 	struct bbr_sendmap *rsm;
7677 
7678 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
7679 	if (rsm && (rsm->r_dupack < 0xff)) {
7680 		rsm->r_dupack++;
7681 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD)
7682 			bbr->r_wanted_output = 1;
7683 	}
7684 }
7685 
7686 /*
7687  * Return value of 1, we do not need to call bbr_process_data().
7688  * return value of 0, bbr_process_data can be called.
7689  * For ret_val if its 0 the TCB is locked and valid, if its non-zero
7690  * its unlocked and probably unsafe to touch the TCB.
7691  */
7692 static int
7693 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
7694     struct tcpcb *tp, struct tcpopt *to,
7695     uint32_t tiwin, int32_t tlen,
7696     int32_t * ofia, int32_t thflags, int32_t * ret_val)
7697 {
7698 	int32_t ourfinisacked = 0;
7699 	int32_t acked_amount;
7700 	uint16_t nsegs;
7701 	int32_t acked;
7702 	uint32_t lost, sack_changed = 0;
7703 	struct mbuf *mfree;
7704 	struct tcp_bbr *bbr;
7705 	uint32_t prev_acked = 0;
7706 
7707 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7708 	lost = bbr->r_ctl.rc_lost;
7709 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
7710 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
7711 		ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
7712 		bbr->r_wanted_output = 1;
7713 		return (1);
7714 	}
7715 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
7716 		/* Process the ack */
7717 		if (bbr->rc_in_persist)
7718 			tp->t_rxtshift = 0;
7719 		if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
7720 			bbr_strike_dupack(bbr);
7721 		sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
7722 	}
7723 	bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost));
7724 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
7725 		/*
7726 		 * Old ack, behind the last one rcv'd or a duplicate ack
7727 		 * with SACK info.
7728 		 */
7729 		if (th->th_ack == tp->snd_una) {
7730 			bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0);
7731 			if (bbr->r_state == TCPS_SYN_SENT) {
7732 				/*
7733 				 * Special case on where we sent SYN. When
7734 				 * the SYN-ACK is processed in syn_sent
7735 				 * state it bumps the snd_una. This causes
7736 				 * us to hit here even though we did ack 1
7737 				 * byte.
7738 				 *
7739 				 * Go through the nothing left case so we
7740 				 * send data.
7741 				 */
7742 				goto nothing_left;
7743 			}
7744 		}
7745 		return (0);
7746 	}
7747 	/*
7748 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
7749 	 * something we sent.
7750 	 */
7751 	if (tp->t_flags & TF_NEEDSYN) {
7752 		/*
7753 		 * T/TCP: Connection was half-synchronized, and our SYN has
7754 		 * been ACK'd (so connection is now fully synchronized).  Go
7755 		 * to non-starred state, increment snd_una for ACK of SYN,
7756 		 * and check if we can do window scaling.
7757 		 */
7758 		tp->t_flags &= ~TF_NEEDSYN;
7759 		tp->snd_una++;
7760 		/* Do window scaling? */
7761 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
7762 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
7763 			tp->rcv_scale = tp->request_r_scale;
7764 			/* Send window already scaled. */
7765 		}
7766 	}
7767 	INP_WLOCK_ASSERT(tptoinpcb(tp));
7768 
7769 	acked = BYTES_THIS_ACK(tp, th);
7770 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
7771 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
7772 
7773 	/*
7774 	 * If we just performed our first retransmit, and the ACK arrives
7775 	 * within our recovery window, then it was a mistake to do the
7776 	 * retransmit in the first place.  Recover our original cwnd and
7777 	 * ssthresh, and proceed to transmit where we left off.
7778 	 */
7779 	if (tp->t_flags & TF_PREVVALID) {
7780 		tp->t_flags &= ~TF_PREVVALID;
7781 		if (tp->t_rxtshift == 1 &&
7782 		    (int)(ticks - tp->t_badrxtwin) < 0)
7783 			bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
7784 	}
7785 	SOCKBUF_LOCK(&so->so_snd);
7786 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
7787 	tp->snd_wnd -= acked_amount;
7788 	mfree = sbcut_locked(&so->so_snd, acked_amount);
7789 	/* NB: sowwakeup_locked() does an implicit unlock. */
7790 	sowwakeup_locked(so);
7791 	m_freem(mfree);
7792 	if (SEQ_GT(th->th_ack, tp->snd_una)) {
7793 		bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
7794 	}
7795 	tp->snd_una = th->th_ack;
7796 	bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost));
7797 	if (IN_RECOVERY(tp->t_flags)) {
7798 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
7799 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
7800 			tcp_bbr_partialack(tp);
7801 		} else {
7802 			bbr_post_recovery(tp);
7803 		}
7804 	}
7805 	if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
7806 		tp->snd_recover = tp->snd_una;
7807 	}
7808 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
7809 		tp->snd_nxt = tp->snd_max;
7810 	}
7811 	if (tp->snd_una == tp->snd_max) {
7812 		/* Nothing left outstanding */
7813 nothing_left:
7814 		bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
7815 		if (sbavail(&so->so_snd) == 0)
7816 			bbr->rc_tp->t_acktime = 0;
7817 		if ((sbused(&so->so_snd) == 0) &&
7818 		    (tp->t_flags & TF_SENTFIN)) {
7819 			ourfinisacked = 1;
7820 		}
7821 		bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
7822 		if (bbr->rc_in_persist == 0) {
7823 			bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
7824 		}
7825 		sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
7826 		bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
7827 		/*
7828 		 * We invalidate the last ack here since we
7829 		 * don't want to transfer forward the time
7830 		 * for our sum's calculations.
7831 		 */
7832 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7833 		    (sbavail(&so->so_snd) == 0) &&
7834 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
7835 			/*
7836 			 * The socket was gone and the peer sent data, time
7837 			 * to reset him.
7838 			 */
7839 			*ret_val = 1;
7840 			tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
7841 			/* tcp_close will kill the inp pre-log the Reset */
7842 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
7843 			tp = tcp_close(tp);
7844 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
7845 			BBR_STAT_INC(bbr_dropped_af_data);
7846 			return (1);
7847 		}
7848 		/* Set need output so persist might get set */
7849 		bbr->r_wanted_output = 1;
7850 	}
7851 	if (ofia)
7852 		*ofia = ourfinisacked;
7853 	return (0);
7854 }
7855 
7856 static void
7857 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
7858 {
7859 	if (bbr->rc_in_persist == 0) {
7860 		bbr_timer_cancel(bbr, __LINE__, cts);
7861 		bbr->r_ctl.rc_last_delay_val = 0;
7862 		tp->t_rxtshift = 0;
7863 		bbr->rc_in_persist = 1;
7864 		bbr->r_ctl.rc_went_idle_time = cts;
7865 		/* We should be capped when rw went to 0 but just in case */
7866 		bbr_log_type_pesist(bbr, cts, 0, line, 1);
7867 		/* Time freezes for the state, so do the accounting now */
7868 		if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
7869 			uint32_t time_in;
7870 
7871 			time_in = cts - bbr->r_ctl.rc_bbr_state_time;
7872 			if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
7873 				int32_t idx;
7874 
7875 				idx = bbr_state_val(bbr);
7876 				counter_u64_add(bbr_state_time[(idx + 5)], time_in);
7877 			} else {
7878 				counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
7879 			}
7880 		}
7881 		bbr->r_ctl.rc_bbr_state_time = cts;
7882 	}
7883 }
7884 
7885 static void
7886 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time)
7887 {
7888 	/*
7889 	 * Note that if idle time does not exceed our
7890 	 * threshold, we do nothing continuing the state
7891 	 * transitions we were last walking through.
7892 	 */
7893 	if (idle_time >= bbr_idle_restart_threshold) {
7894 		if (bbr->rc_use_idle_restart) {
7895 			bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT;
7896 			/*
7897 			 * Set our target using BBR_UNIT, so
7898 			 * we increase at a dramatic rate but
7899 			 * we stop when we get the pipe
7900 			 * full again for our current b/w estimate.
7901 			 */
7902 			bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
7903 			bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
7904 			bbr_set_state_target(bbr, __LINE__);
7905 			/* Now setup our gains to ramp up */
7906 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
7907 			bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
7908 			bbr_log_type_statechange(bbr, cts, __LINE__);
7909 		} else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
7910 			bbr_substate_change(bbr, cts, __LINE__, 1);
7911 		}
7912 	}
7913 }
7914 
7915 static void
7916 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
7917 {
7918 	uint32_t idle_time;
7919 
7920 	if (bbr->rc_in_persist == 0)
7921 		return;
7922 	idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time);
7923 	bbr->rc_in_persist = 0;
7924 	bbr->rc_hit_state_1 = 0;
7925 	bbr->r_ctl.rc_del_time = cts;
7926 	/*
7927 	 * We invalidate the last ack here since we
7928 	 * don't want to transfer forward the time
7929 	 * for our sum's calculations.
7930 	 */
7931 	if (tcp_in_hpts(bbr->rc_inp)) {
7932 		tcp_hpts_remove(bbr->rc_inp);
7933 		bbr->rc_timer_first = 0;
7934 		bbr->r_ctl.rc_hpts_flags = 0;
7935 		bbr->r_ctl.rc_last_delay_val = 0;
7936 		bbr->r_ctl.rc_hptsi_agg_delay = 0;
7937 		bbr->r_agg_early_set = 0;
7938 		bbr->r_ctl.rc_agg_early = 0;
7939 	}
7940 	bbr_log_type_pesist(bbr, cts, idle_time, line, 0);
7941 	if (idle_time >= bbr_rtt_probe_time) {
7942 		/*
7943 		 * This qualifies as a RTT_PROBE session since we drop the
7944 		 * data outstanding to nothing and waited more than
7945 		 * bbr_rtt_probe_time.
7946 		 */
7947 		bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0);
7948 		bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts;
7949 	}
7950 	tp->t_rxtshift = 0;
7951 	/*
7952 	 * If in probeBW and we have persisted more than an RTT lets do
7953 	 * special handling.
7954 	 */
7955 	/* Force a time based epoch */
7956 	bbr_set_epoch(bbr, cts, __LINE__);
7957 	/*
7958 	 * Setup the lost so we don't count anything against the guy
7959 	 * we have been stuck with during persists.
7960 	 */
7961 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
7962 	/* Time un-freezes for the state */
7963 	bbr->r_ctl.rc_bbr_state_time = cts;
7964 	if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) ||
7965 	    (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) {
7966 		/*
7967 		 * If we are going back to probe-bw
7968 		 * or probe_rtt, we may need to possibly
7969 		 * do a fast restart.
7970 		 */
7971 		bbr_restart_after_idle(bbr, cts, idle_time);
7972 	}
7973 }
7974 
7975 static void
7976 bbr_collapsed_window(struct tcp_bbr *bbr)
7977 {
7978 	/*
7979 	 * Now we must walk the
7980 	 * send map and divide the
7981 	 * ones left stranded. These
7982 	 * guys can't cause us to abort
7983 	 * the connection and are really
7984 	 * "unsent". However if a buggy
7985 	 * client actually did keep some
7986 	 * of the data i.e. collapsed the win
7987 	 * and refused to ack and then opened
7988 	 * the win and acked that data. We would
7989 	 * get into an ack war, the simplier
7990 	 * method then of just pretending we
7991 	 * did not send those segments something
7992 	 * won't work.
7993 	 */
7994 	struct bbr_sendmap *rsm, *nrsm;
7995 	tcp_seq max_seq;
7996 	uint32_t maxseg;
7997 	int can_split = 0;
7998 	int fnd = 0;
7999 
8000 	maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
8001 	max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd;
8002 	bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0);
8003 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
8004 		/* Find the first seq past or at maxseq */
8005 		if (rsm->r_flags & BBR_RWND_COLLAPSED)
8006 			rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8007 		if (SEQ_GEQ(max_seq, rsm->r_start) &&
8008 		    SEQ_GEQ(rsm->r_end, max_seq)) {
8009 			fnd = 1;
8010 			break;
8011 		}
8012 	}
8013 	bbr->rc_has_collapsed = 0;
8014 	if (!fnd) {
8015 		/* Nothing to do strange */
8016 		return;
8017 	}
8018 	/*
8019 	 * Now can we split?
8020 	 *
8021 	 * We don't want to split if splitting
8022 	 * would generate too many small segments
8023 	 * less we let an attacker fragment our
8024 	 * send_map and leave us out of memory.
8025 	 */
8026 	if ((max_seq != rsm->r_start) &&
8027 	    (max_seq != rsm->r_end)){
8028 		/* can we split? */
8029 		int res1, res2;
8030 
8031 		res1 = max_seq - rsm->r_start;
8032 		res2 = rsm->r_end - max_seq;
8033 		if ((res1 >= (maxseg/8)) &&
8034 		    (res2 >= (maxseg/8))) {
8035 			/* No small pieces here */
8036 			can_split = 1;
8037 		} else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) {
8038 			/* We are under the limit */
8039 			can_split = 1;
8040 		}
8041 	}
8042 	/* Ok do we need to split this rsm? */
8043 	if (max_seq == rsm->r_start) {
8044 		/* It's this guy no split required */
8045 		nrsm = rsm;
8046 	} else if (max_seq == rsm->r_end) {
8047 		/* It's the next one no split required. */
8048 		nrsm = TAILQ_NEXT(rsm, r_next);
8049 		if (nrsm == NULL) {
8050 			/* Huh? */
8051 			return;
8052 		}
8053 	} else if (can_split && SEQ_LT(max_seq, rsm->r_end)) {
8054 		/* yep we need to split it */
8055 		nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
8056 		if (nrsm == NULL) {
8057 			/* failed XXXrrs what can we do mark the whole? */
8058 			nrsm = rsm;
8059 			goto no_split;
8060 		}
8061 		/* Clone it */
8062 		bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0);
8063 		bbr_clone_rsm(bbr, nrsm, rsm, max_seq);
8064 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
8065 		if (rsm->r_in_tmap) {
8066 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8067 			nrsm->r_in_tmap = 1;
8068 		}
8069 	} else {
8070 		/*
8071 		 * Split not allowed just start here just
8072 		 * use this guy.
8073 		 */
8074 		nrsm = rsm;
8075 	}
8076 no_split:
8077 	BBR_STAT_INC(bbr_collapsed_win);
8078 	/* reuse fnd as a count */
8079 	fnd = 0;
8080 	TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) {
8081 		nrsm->r_flags |= BBR_RWND_COLLAPSED;
8082 		fnd++;
8083 		bbr->rc_has_collapsed = 1;
8084 	}
8085 	bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd);
8086 }
8087 
8088 static void
8089 bbr_un_collapse_window(struct tcp_bbr *bbr)
8090 {
8091 	struct bbr_sendmap *rsm;
8092 	int cleared = 0;
8093 
8094 	TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
8095 		if (rsm->r_flags & BBR_RWND_COLLAPSED) {
8096 			/* Clear the flag */
8097 			rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8098 			cleared++;
8099 		} else
8100 			break;
8101 	}
8102 	bbr_log_type_rwnd_collapse(bbr,
8103 				   (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared);
8104 	bbr->rc_has_collapsed = 0;
8105 }
8106 
8107 /*
8108  * Return value of 1, the TCB is unlocked and most
8109  * likely gone, return value of 0, the TCB is still
8110  * locked.
8111  */
8112 static int
8113 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8114     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8115     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8116 {
8117 	/*
8118 	 * Update window information. Don't look at window if no ACK: TAC's
8119 	 * send garbage on first SYN.
8120 	 */
8121 	uint16_t nsegs;
8122 	int32_t tfo_syn;
8123 	struct tcp_bbr *bbr;
8124 
8125 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8126 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8127 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8128 	if ((thflags & TH_ACK) &&
8129 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8130 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8131 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8132 		/* keep track of pure window updates */
8133 		if (tlen == 0 &&
8134 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8135 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
8136 		tp->snd_wnd = tiwin;
8137 		tp->snd_wl1 = th->th_seq;
8138 		tp->snd_wl2 = th->th_ack;
8139 		if (tp->snd_wnd > tp->max_sndwnd)
8140 			tp->max_sndwnd = tp->snd_wnd;
8141 		bbr->r_wanted_output = 1;
8142 	} else if (thflags & TH_ACK) {
8143 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8144 			tp->snd_wnd = tiwin;
8145 			tp->snd_wl1 = th->th_seq;
8146 			tp->snd_wl2 = th->th_ack;
8147 		}
8148 	}
8149 	if (tp->snd_wnd < ctf_outstanding(tp))
8150 		/* The peer collapsed its window on us */
8151 		bbr_collapsed_window(bbr);
8152  	else if (bbr->rc_has_collapsed)
8153 		bbr_un_collapse_window(bbr);
8154 	/* Was persist timer active and now we have window space? */
8155 	if ((bbr->rc_in_persist != 0) &&
8156 	    (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8157 				bbr_minseg(bbr)))) {
8158 		/*
8159 		 * Make the rate persist at end of persist mode if idle long
8160 		 * enough
8161 		 */
8162 		bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8163 
8164 		/* Make sure we output to start the timer */
8165 		bbr->r_wanted_output = 1;
8166 	}
8167 	/* Do we need to enter persist? */
8168 	if ((bbr->rc_in_persist == 0) &&
8169 	    (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8170 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
8171 	    (tp->snd_max == tp->snd_una) &&
8172 	    sbavail(&so->so_snd) &&
8173 	    (sbavail(&so->so_snd) > tp->snd_wnd)) {
8174 		/* No send window.. we must enter persist */
8175 		bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8176 	}
8177 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8178 		m_freem(m);
8179 		return (0);
8180 	}
8181 	/*
8182 	 * We don't support urgent data but
8183 	 * drag along the up just to make sure
8184 	 * if there is a stack switch no one
8185 	 * is surprised.
8186 	 */
8187 	tp->rcv_up = tp->rcv_nxt;
8188 
8189 	/*
8190 	 * Process the segment text, merging it into the TCP sequencing
8191 	 * queue, and arranging for acknowledgment of receipt if necessary.
8192 	 * This process logically involves adjusting tp->rcv_wnd as data is
8193 	 * presented to the user (this happens in tcp_usrreq.c, case
8194 	 * PRU_RCVD).  If a FIN has already been received on this connection
8195 	 * then we just ignore the text.
8196 	 */
8197 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8198 		   IS_FASTOPEN(tp->t_flags));
8199 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
8200 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8201 		tcp_seq save_start = th->th_seq;
8202 		tcp_seq save_rnxt  = tp->rcv_nxt;
8203 		int     save_tlen  = tlen;
8204 
8205 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8206 		/*
8207 		 * Insert segment which includes th into TCP reassembly
8208 		 * queue with control block tp.  Set thflags to whether
8209 		 * reassembly now includes a segment with FIN.  This handles
8210 		 * the common case inline (segment is the next to be
8211 		 * received on an established connection, and the queue is
8212 		 * empty), avoiding linkage into and removal from the queue
8213 		 * and repetition of various conversions. Set DELACK for
8214 		 * segments received in order, but ack immediately when
8215 		 * segments are out of order (so fast retransmit can work).
8216 		 */
8217 		if (th->th_seq == tp->rcv_nxt &&
8218 		    SEGQ_EMPTY(tp) &&
8219 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
8220 		    tfo_syn)) {
8221 #ifdef NETFLIX_SB_LIMITS
8222 			u_int mcnt, appended;
8223 
8224 			if (so->so_rcv.sb_shlim) {
8225 				mcnt = m_memcnt(m);
8226 				appended = 0;
8227 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8228 				    CFO_NOSLEEP, NULL) == false) {
8229 					counter_u64_add(tcp_sb_shlim_fails, 1);
8230 					m_freem(m);
8231 					return (0);
8232 				}
8233 			}
8234 
8235 #endif
8236 			if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) {
8237 				bbr->bbr_segs_rcvd += max(1, nsegs);
8238 				tp->t_flags |= TF_DELACK;
8239 				bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8240 			} else {
8241 				bbr->r_wanted_output = 1;
8242 				tp->t_flags |= TF_ACKNOW;
8243 			}
8244 			tp->rcv_nxt += tlen;
8245 			if (tlen &&
8246 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8247 			    (tp->t_fbyte_in == 0)) {
8248 				tp->t_fbyte_in = ticks;
8249 				if (tp->t_fbyte_in == 0)
8250 					tp->t_fbyte_in = 1;
8251 				if (tp->t_fbyte_out && tp->t_fbyte_in)
8252 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8253 			}
8254 			thflags = tcp_get_flags(th) & TH_FIN;
8255 			KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8256 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8257 			SOCKBUF_LOCK(&so->so_rcv);
8258 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
8259 				m_freem(m);
8260 			else
8261 #ifdef NETFLIX_SB_LIMITS
8262 				appended =
8263 #endif
8264 					sbappendstream_locked(&so->so_rcv, m, 0);
8265 			/* NB: sorwakeup_locked() does an implicit unlock. */
8266 			sorwakeup_locked(so);
8267 #ifdef NETFLIX_SB_LIMITS
8268 			if (so->so_rcv.sb_shlim && appended != mcnt)
8269 				counter_fo_release(so->so_rcv.sb_shlim,
8270 				    mcnt - appended);
8271 #endif
8272 
8273 		} else {
8274 			/*
8275 			 * XXX: Due to the header drop above "th" is
8276 			 * theoretically invalid by now.  Fortunately
8277 			 * m_adj() doesn't actually frees any mbufs when
8278 			 * trimming from the head.
8279 			 */
8280 			tcp_seq temp = save_start;
8281 
8282 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
8283 			tp->t_flags |= TF_ACKNOW;
8284 			if (tp->t_flags & TF_WAKESOR) {
8285 				tp->t_flags &= ~TF_WAKESOR;
8286 				/* NB: sorwakeup_locked() does an implicit unlock. */
8287 				sorwakeup_locked(so);
8288 			}
8289 		}
8290 		if ((tp->t_flags & TF_SACK_PERMIT) &&
8291 		    (save_tlen > 0) &&
8292 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
8293 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8294 				/*
8295 				 * DSACK actually handled in the fastpath
8296 				 * above.
8297 				 */
8298 				tcp_update_sack_list(tp, save_start,
8299 				    save_start + save_tlen);
8300 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8301 				if ((tp->rcv_numsacks >= 1) &&
8302 				    (tp->sackblks[0].end == save_start)) {
8303 					/*
8304 					 * Partial overlap, recorded at todrop
8305 					 * above.
8306 					 */
8307 					tcp_update_sack_list(tp,
8308 					    tp->sackblks[0].start,
8309 					    tp->sackblks[0].end);
8310 				} else {
8311 					tcp_update_dsack_list(tp, save_start,
8312 					    save_start + save_tlen);
8313 				}
8314 			} else if (tlen >= save_tlen) {
8315 				/* Update of sackblks. */
8316 				tcp_update_dsack_list(tp, save_start,
8317 				    save_start + save_tlen);
8318 			} else if (tlen > 0) {
8319 				tcp_update_dsack_list(tp, save_start,
8320 				    save_start + tlen);
8321 			}
8322 		}
8323 	} else {
8324 		m_freem(m);
8325 		thflags &= ~TH_FIN;
8326 	}
8327 
8328 	/*
8329 	 * If FIN is received ACK the FIN and let the user know that the
8330 	 * connection is closing.
8331 	 */
8332 	if (thflags & TH_FIN) {
8333 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8334 			/* The socket upcall is handled by socantrcvmore. */
8335 			socantrcvmore(so);
8336 			/*
8337 			 * If connection is half-synchronized (ie NEEDSYN
8338 			 * flag on) then delay ACK, so it may be piggybacked
8339 			 * when SYN is sent. Otherwise, since we received a
8340 			 * FIN then no more input can be expected, send ACK
8341 			 * now.
8342 			 */
8343 			if (tp->t_flags & TF_NEEDSYN) {
8344 				tp->t_flags |= TF_DELACK;
8345 				bbr_timer_cancel(bbr,
8346 				    __LINE__, bbr->r_ctl.rc_rcvtime);
8347 			} else {
8348 				tp->t_flags |= TF_ACKNOW;
8349 			}
8350 			tp->rcv_nxt++;
8351 		}
8352 		switch (tp->t_state) {
8353 			/*
8354 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
8355 			 * CLOSE_WAIT state.
8356 			 */
8357 		case TCPS_SYN_RECEIVED:
8358 			tp->t_starttime = ticks;
8359 			/* FALLTHROUGH */
8360 		case TCPS_ESTABLISHED:
8361 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
8362 			break;
8363 
8364 			/*
8365 			 * If still in FIN_WAIT_1 STATE FIN has not been
8366 			 * acked so enter the CLOSING state.
8367 			 */
8368 		case TCPS_FIN_WAIT_1:
8369 			tcp_state_change(tp, TCPS_CLOSING);
8370 			break;
8371 
8372 			/*
8373 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
8374 			 * starting the time-wait timer, turning off the
8375 			 * other standard timers.
8376 			 */
8377 		case TCPS_FIN_WAIT_2:
8378 			bbr->rc_timer_first = 1;
8379 			bbr_timer_cancel(bbr,
8380 			    __LINE__, bbr->r_ctl.rc_rcvtime);
8381 			tcp_twstart(tp);
8382 			return (1);
8383 		}
8384 	}
8385 	/*
8386 	 * Return any desired output.
8387 	 */
8388 	if ((tp->t_flags & TF_ACKNOW) ||
8389 	    (sbavail(&so->so_snd) > ctf_outstanding(tp))) {
8390 		bbr->r_wanted_output = 1;
8391 	}
8392 	return (0);
8393 }
8394 
8395 /*
8396  * Here nothing is really faster, its just that we
8397  * have broken out the fast-data path also just like
8398  * the fast-ack. Return 1 if we processed the packet
8399  * return 0 if you need to take the "slow-path".
8400  */
8401 static int
8402 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8403     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8404     uint32_t tiwin, int32_t nxt_pkt)
8405 {
8406 	uint16_t nsegs;
8407 	int32_t newsize = 0;	/* automatic sockbuf scaling */
8408 	struct tcp_bbr *bbr;
8409 #ifdef NETFLIX_SB_LIMITS
8410 	u_int mcnt, appended;
8411 #endif
8412 
8413 	/* On the hpts and we would have called output */
8414 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8415 
8416 	/*
8417 	 * If last ACK falls within this segment's sequence numbers, record
8418 	 * the timestamp. NOTE that the test is modified according to the
8419 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8420 	 */
8421 	if (bbr->r_ctl.rc_resend != NULL) {
8422 		return (0);
8423 	}
8424 	if (tiwin && tiwin != tp->snd_wnd) {
8425 		return (0);
8426 	}
8427 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8428 		return (0);
8429 	}
8430 	if (__predict_false((to->to_flags & TOF_TS) &&
8431 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8432 		return (0);
8433 	}
8434 	if (__predict_false((th->th_ack != tp->snd_una))) {
8435 		return (0);
8436 	}
8437 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8438 		return (0);
8439 	}
8440 	if ((to->to_flags & TOF_TS) != 0 &&
8441 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8442 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
8443 		tp->ts_recent = to->to_tsval;
8444 	}
8445 	/*
8446 	 * This is a pure, in-sequence data packet with nothing on the
8447 	 * reassembly queue and we have enough buffer space to take it.
8448 	 */
8449 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8450 
8451 #ifdef NETFLIX_SB_LIMITS
8452 	if (so->so_rcv.sb_shlim) {
8453 		mcnt = m_memcnt(m);
8454 		appended = 0;
8455 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8456 		    CFO_NOSLEEP, NULL) == false) {
8457 			counter_u64_add(tcp_sb_shlim_fails, 1);
8458 			m_freem(m);
8459 			return (1);
8460 		}
8461 	}
8462 #endif
8463 	/* Clean receiver SACK report if present */
8464 	if (tp->rcv_numsacks)
8465 		tcp_clean_sackreport(tp);
8466 	KMOD_TCPSTAT_INC(tcps_preddat);
8467 	tp->rcv_nxt += tlen;
8468 	if (tlen &&
8469 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8470 	    (tp->t_fbyte_in == 0)) {
8471 		tp->t_fbyte_in = ticks;
8472 		if (tp->t_fbyte_in == 0)
8473 			tp->t_fbyte_in = 1;
8474 		if (tp->t_fbyte_out && tp->t_fbyte_in)
8475 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8476 	}
8477 	/*
8478 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8479 	 */
8480 	tp->snd_wl1 = th->th_seq;
8481 	/*
8482 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8483 	 */
8484 	tp->rcv_up = tp->rcv_nxt;
8485 	KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8486 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8487 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8488 
8489 	/* Add data to socket buffer. */
8490 	SOCKBUF_LOCK(&so->so_rcv);
8491 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8492 		m_freem(m);
8493 	} else {
8494 		/*
8495 		 * Set new socket buffer size. Give up when limit is
8496 		 * reached.
8497 		 */
8498 		if (newsize)
8499 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
8500 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8501 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8502 
8503 #ifdef NETFLIX_SB_LIMITS
8504 		appended =
8505 #endif
8506 			sbappendstream_locked(&so->so_rcv, m, 0);
8507 		ctf_calc_rwin(so, tp);
8508 	}
8509 	/* NB: sorwakeup_locked() does an implicit unlock. */
8510 	sorwakeup_locked(so);
8511 #ifdef NETFLIX_SB_LIMITS
8512 	if (so->so_rcv.sb_shlim && mcnt != appended)
8513 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8514 #endif
8515 	if (DELAY_ACK(tp, bbr, nsegs)) {
8516 		bbr->bbr_segs_rcvd += max(1, nsegs);
8517 		tp->t_flags |= TF_DELACK;
8518 		bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8519 	} else {
8520 		bbr->r_wanted_output = 1;
8521 		tp->t_flags |= TF_ACKNOW;
8522 	}
8523 	return (1);
8524 }
8525 
8526 /*
8527  * This subfunction is used to try to highly optimize the
8528  * fast path. We again allow window updates that are
8529  * in sequence to remain in the fast-path. We also add
8530  * in the __predict's to attempt to help the compiler.
8531  * Note that if we return a 0, then we can *not* process
8532  * it and the caller should push the packet into the
8533  * slow-path. If we return 1, then all is well and
8534  * the packet is fully processed.
8535  */
8536 static int
8537 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8538     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8539     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
8540 {
8541 	int32_t acked;
8542 	uint16_t nsegs;
8543 	uint32_t sack_changed;
8544 	uint32_t prev_acked = 0;
8545 	struct tcp_bbr *bbr;
8546 
8547 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8548 		/* Old ack, behind (or duplicate to) the last one rcv'd */
8549 		return (0);
8550 	}
8551 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
8552 		/* Above what we have sent? */
8553 		return (0);
8554 	}
8555 	if (__predict_false(tiwin == 0)) {
8556 		/* zero window */
8557 		return (0);
8558 	}
8559 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
8560 		/* We need a SYN or a FIN, unlikely.. */
8561 		return (0);
8562 	}
8563 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
8564 		/* Timestamp is behind .. old ack with seq wrap? */
8565 		return (0);
8566 	}
8567 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
8568 		/* Still recovering */
8569 		return (0);
8570 	}
8571 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8572 	if (__predict_false(bbr->r_ctl.rc_resend != NULL)) {
8573 		/* We are retransmitting */
8574 		return (0);
8575 	}
8576 	if (__predict_false(bbr->rc_in_persist != 0)) {
8577 		/* In persist mode */
8578 		return (0);
8579 	}
8580 	if (bbr->r_ctl.rc_sacked) {
8581 		/* We have sack holes on our scoreboard */
8582 		return (0);
8583 	}
8584 	/* Ok if we reach here, we can process a fast-ack */
8585 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8586 	sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
8587 	/*
8588 	 * We never detect loss in fast ack [we can't
8589 	 * have a sack and can't be in recovery so
8590 	 * we always pass 0 (nothing detected)].
8591 	 */
8592 	bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0);
8593 	/* Did the window get updated? */
8594 	if (tiwin != tp->snd_wnd) {
8595 		tp->snd_wnd = tiwin;
8596 		tp->snd_wl1 = th->th_seq;
8597 		if (tp->snd_wnd > tp->max_sndwnd)
8598 			tp->max_sndwnd = tp->snd_wnd;
8599 	}
8600 	/* Do we need to exit persists? */
8601 	if ((bbr->rc_in_persist != 0) &&
8602 	    (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8603 			       bbr_minseg(bbr)))) {
8604 		bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8605 		bbr->r_wanted_output = 1;
8606 	}
8607 	/* Do we need to enter persists? */
8608 	if ((bbr->rc_in_persist == 0) &&
8609 	    (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8610 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
8611 	    (tp->snd_max == tp->snd_una) &&
8612 	    sbavail(&so->so_snd) &&
8613 	    (sbavail(&so->so_snd) > tp->snd_wnd)) {
8614 		/* No send window.. we must enter persist */
8615 		bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8616 	}
8617 	/*
8618 	 * If last ACK falls within this segment's sequence numbers, record
8619 	 * the timestamp. NOTE that the test is modified according to the
8620 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8621 	 */
8622 	if ((to->to_flags & TOF_TS) != 0 &&
8623 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8624 		tp->ts_recent_age = bbr->r_ctl.rc_rcvtime;
8625 		tp->ts_recent = to->to_tsval;
8626 	}
8627 	/*
8628 	 * This is a pure ack for outstanding data.
8629 	 */
8630 	KMOD_TCPSTAT_INC(tcps_predack);
8631 
8632 	/*
8633 	 * "bad retransmit" recovery.
8634 	 */
8635 	if (tp->t_flags & TF_PREVVALID) {
8636 		tp->t_flags &= ~TF_PREVVALID;
8637 		if (tp->t_rxtshift == 1 &&
8638 		    (int)(ticks - tp->t_badrxtwin) < 0)
8639 			bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
8640 	}
8641 	/*
8642 	 * Recalculate the transmit timer / rtt.
8643 	 *
8644 	 * Some boxes send broken timestamp replies during the SYN+ACK
8645 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
8646 	 * and blow up the retransmit timer.
8647 	 */
8648 	acked = BYTES_THIS_ACK(tp, th);
8649 
8650 #ifdef TCP_HHOOK
8651 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
8652 	hhook_run_tcp_est_in(tp, th, to);
8653 #endif
8654 
8655 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
8656 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
8657 	sbdrop(&so->so_snd, acked);
8658 
8659 	if (SEQ_GT(th->th_ack, tp->snd_una))
8660 		bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
8661 	tp->snd_una = th->th_ack;
8662 	if (tp->snd_wnd < ctf_outstanding(tp))
8663 		/* The peer collapsed its window on us */
8664 		bbr_collapsed_window(bbr);
8665 	else if (bbr->rc_has_collapsed)
8666 		bbr_un_collapse_window(bbr);
8667 
8668 	if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8669 		tp->snd_recover = tp->snd_una;
8670 	}
8671 	bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0);
8672 	/*
8673 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
8674 	 */
8675 	tp->snd_wl2 = th->th_ack;
8676 	m_freem(m);
8677 	/*
8678 	 * If all outstanding data are acked, stop retransmit timer,
8679 	 * otherwise restart timer using current (possibly backed-off)
8680 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
8681 	 * If data are ready to send, let tcp_output decide between more
8682 	 * output or persist.
8683 	 * Wake up the socket if we have room to write more.
8684 	 */
8685 	sowwakeup(so);
8686 	if (tp->snd_una == tp->snd_max) {
8687 		/* Nothing left outstanding */
8688 		bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8689 		if (sbavail(&so->so_snd) == 0)
8690 			bbr->rc_tp->t_acktime = 0;
8691 		bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8692 		if (bbr->rc_in_persist == 0) {
8693 			bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8694 		}
8695 		sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8696 		bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8697 		/*
8698 		 * We invalidate the last ack here since we
8699 		 * don't want to transfer forward the time
8700 		 * for our sum's calculations.
8701 		 */
8702 		bbr->r_wanted_output = 1;
8703 	}
8704 	if (sbavail(&so->so_snd)) {
8705 		bbr->r_wanted_output = 1;
8706 	}
8707 	return (1);
8708 }
8709 
8710 /*
8711  * Return value of 1, the TCB is unlocked and most
8712  * likely gone, return value of 0, the TCB is still
8713  * locked.
8714  */
8715 static int
8716 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
8717     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8718     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
8719 {
8720 	int32_t todrop;
8721 	int32_t ourfinisacked = 0;
8722 	struct tcp_bbr *bbr;
8723 	int32_t ret_val = 0;
8724 
8725 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8726 
8727 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8728 	ctf_calc_rwin(so, tp);
8729 	/*
8730 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
8731 	 * SYN, drop the input. if seg contains a RST, then drop the
8732 	 * connection. if seg does not contain SYN, then drop it. Otherwise
8733 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
8734 	 * tp->irs if seg contains ack then advance tp->snd_una. BRR does
8735 	 * not support ECN so we will not say we are capable. if SYN has
8736 	 * been acked change to ESTABLISHED else SYN_RCVD state arrange for
8737 	 * segment to be acked (eventually) continue processing rest of
8738 	 * data/controls, beginning with URG
8739 	 */
8740 	if ((thflags & TH_ACK) &&
8741 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
8742 	    SEQ_GT(th->th_ack, tp->snd_max))) {
8743 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8744 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
8745 		return (1);
8746 	}
8747 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
8748 		TCP_PROBE5(connect__refused, NULL, tp,
8749 		    mtod(m, const char *), tp, th);
8750 		tp = tcp_drop(tp, ECONNREFUSED);
8751 		ctf_do_drop(m, tp);
8752 		return (1);
8753 	}
8754 	if (thflags & TH_RST) {
8755 		ctf_do_drop(m, tp);
8756 		return (1);
8757 	}
8758 	if (!(thflags & TH_SYN)) {
8759 		ctf_do_drop(m, tp);
8760 		return (1);
8761 	}
8762 	tp->irs = th->th_seq;
8763 	tcp_rcvseqinit(tp);
8764 	if (thflags & TH_ACK) {
8765 		int tfo_partial = 0;
8766 
8767 		KMOD_TCPSTAT_INC(tcps_connects);
8768 		soisconnected(so);
8769 #ifdef MAC
8770 		mac_socketpeer_set_from_mbuf(m, so);
8771 #endif
8772 		/* Do window scaling on this connection? */
8773 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
8774 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
8775 			tp->rcv_scale = tp->request_r_scale;
8776 		}
8777 		tp->rcv_adv += min(tp->rcv_wnd,
8778 		    TCP_MAXWIN << tp->rcv_scale);
8779 		/*
8780 		 * If not all the data that was sent in the TFO SYN
8781 		 * has been acked, resend the remainder right away.
8782 		 */
8783 		if (IS_FASTOPEN(tp->t_flags) &&
8784 		    (tp->snd_una != tp->snd_max)) {
8785 			tp->snd_nxt = th->th_ack;
8786 			tfo_partial = 1;
8787 		}
8788 		/*
8789 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
8790 		 * will be turned on later.
8791 		 */
8792 		if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) {
8793 			bbr->bbr_segs_rcvd += 1;
8794 			tp->t_flags |= TF_DELACK;
8795 			bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8796 		} else {
8797 			bbr->r_wanted_output = 1;
8798 			tp->t_flags |= TF_ACKNOW;
8799 		}
8800 		if (SEQ_GT(th->th_ack, tp->iss)) {
8801 			/*
8802 			 * The SYN is acked
8803 			 * handle it specially.
8804 			 */
8805 			bbr_log_syn(tp, to);
8806 		}
8807 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
8808 			/*
8809 			 * We advance snd_una for the
8810 			 * fast open case. If th_ack is
8811 			 * acknowledging data beyond
8812 			 * snd_una we can't just call
8813 			 * ack-processing since the
8814 			 * data stream in our send-map
8815 			 * will start at snd_una + 1 (one
8816 			 * beyond the SYN). If its just
8817 			 * equal we don't need to do that
8818 			 * and there is no send_map.
8819 			 */
8820 			tp->snd_una++;
8821 		}
8822 		/*
8823 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
8824 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
8825 		 */
8826 		tp->t_starttime = ticks;
8827 		if (tp->t_flags & TF_NEEDFIN) {
8828 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
8829 			tp->t_flags &= ~TF_NEEDFIN;
8830 			thflags &= ~TH_SYN;
8831 		} else {
8832 			tcp_state_change(tp, TCPS_ESTABLISHED);
8833 			TCP_PROBE5(connect__established, NULL, tp,
8834 			    mtod(m, const char *), tp, th);
8835 			cc_conn_init(tp);
8836 		}
8837 	} else {
8838 		/*
8839 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
8840 		 * open.  If segment contains CC option and there is a
8841 		 * cached CC, apply TAO test. If it succeeds, connection is *
8842 		 * half-synchronized. Otherwise, do 3-way handshake:
8843 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
8844 		 * there was no CC option, clear cached CC value.
8845 		 */
8846 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
8847 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
8848 	}
8849 	/*
8850 	 * Advance th->th_seq to correspond to first data byte. If data,
8851 	 * trim to stay within window, dropping FIN if necessary.
8852 	 */
8853 	th->th_seq++;
8854 	if (tlen > tp->rcv_wnd) {
8855 		todrop = tlen - tp->rcv_wnd;
8856 		m_adj(m, -todrop);
8857 		tlen = tp->rcv_wnd;
8858 		thflags &= ~TH_FIN;
8859 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
8860 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
8861 	}
8862 	tp->snd_wl1 = th->th_seq - 1;
8863 	tp->rcv_up = th->th_seq;
8864 	/*
8865 	 * Client side of transaction: already sent SYN and data. If the
8866 	 * remote host used T/TCP to validate the SYN, our data will be
8867 	 * ACK'd; if so, enter normal data segment processing in the middle
8868 	 * of step 5, ack processing. Otherwise, goto step 6.
8869 	 */
8870 	if (thflags & TH_ACK) {
8871 		if ((to->to_flags & TOF_TS) != 0) {
8872 			uint32_t t, rtt;
8873 
8874 			t = tcp_tv_to_mssectick(&bbr->rc_tv);
8875 			if (TSTMP_GEQ(t, to->to_tsecr)) {
8876 				rtt = t - to->to_tsecr;
8877 				if (rtt == 0) {
8878 					rtt = 1;
8879 				}
8880 				rtt *= MS_IN_USEC;
8881 				tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
8882 				apply_filter_min_small(&bbr->r_ctl.rc_rttprop,
8883 						       rtt, bbr->r_ctl.rc_rcvtime);
8884 			}
8885 		}
8886 		if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
8887 			return (ret_val);
8888 		/* We may have changed to FIN_WAIT_1 above */
8889 		if (tp->t_state == TCPS_FIN_WAIT_1) {
8890 			/*
8891 			 * In FIN_WAIT_1 STATE in addition to the processing
8892 			 * for the ESTABLISHED state if our FIN is now
8893 			 * acknowledged then enter FIN_WAIT_2.
8894 			 */
8895 			if (ourfinisacked) {
8896 				/*
8897 				 * If we can't receive any more data, then
8898 				 * closing user can proceed. Starting the
8899 				 * timer is contrary to the specification,
8900 				 * but if we don't get a FIN we'll hang
8901 				 * forever.
8902 				 *
8903 				 * XXXjl: we should release the tp also, and
8904 				 * use a compressed state.
8905 				 */
8906 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8907 					soisdisconnected(so);
8908 					tcp_timer_activate(tp, TT_2MSL,
8909 					    (tcp_fast_finwait2_recycle ?
8910 					    tcp_finwait2_timeout :
8911 					    TP_MAXIDLE(tp)));
8912 				}
8913 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
8914 			}
8915 		}
8916 	}
8917 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
8918 	    tiwin, thflags, nxt_pkt));
8919 }
8920 
8921 /*
8922  * Return value of 1, the TCB is unlocked and most
8923  * likely gone, return value of 0, the TCB is still
8924  * locked.
8925  */
8926 static int
8927 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
8928 		struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8929 		uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
8930 {
8931 	int32_t ourfinisacked = 0;
8932 	int32_t ret_val;
8933 	struct tcp_bbr *bbr;
8934 
8935 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8936 
8937 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8938 	ctf_calc_rwin(so, tp);
8939 	if ((thflags & TH_ACK) &&
8940 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
8941 	     SEQ_GT(th->th_ack, tp->snd_max))) {
8942 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8943 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
8944 		return (1);
8945 	}
8946 	if (IS_FASTOPEN(tp->t_flags)) {
8947 		/*
8948 		 * When a TFO connection is in SYN_RECEIVED, the only valid
8949 		 * packets are the initial SYN, a retransmit/copy of the
8950 		 * initial SYN (possibly with a subset of the original
8951 		 * data), a valid ACK, a FIN, or a RST.
8952 		 */
8953 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
8954 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8955 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
8956 			return (1);
8957 		} else if (thflags & TH_SYN) {
8958 			/* non-initial SYN is ignored */
8959 			if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
8960 			    (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
8961 			    (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
8962 				ctf_do_drop(m, NULL);
8963 				return (0);
8964 			}
8965 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
8966 			ctf_do_drop(m, NULL);
8967 			return (0);
8968 		}
8969 	}
8970 	if ((thflags & TH_RST) ||
8971 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
8972 		return (ctf_process_rst(m, th, so, tp));
8973 	/*
8974 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
8975 	 * it's less than ts_recent, drop it.
8976 	 */
8977 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
8978 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
8979 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
8980 			return (ret_val);
8981 	}
8982 	/*
8983 	 * In the SYN-RECEIVED state, validate that the packet belongs to
8984 	 * this connection before trimming the data to fit the receive
8985 	 * window.  Check the sequence number versus IRS since we know the
8986 	 * sequence numbers haven't wrapped.  This is a partial fix for the
8987 	 * "LAND" DoS attack.
8988 	 */
8989 	if (SEQ_LT(th->th_seq, tp->irs)) {
8990 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8991 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
8992 		return (1);
8993 	}
8994 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
8995 		return (ret_val);
8996 	}
8997 	/*
8998 	 * If last ACK falls within this segment's sequence numbers, record
8999 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9000 	 * from the latest proposal of the tcplw@cray.com list (Braden
9001 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9002 	 * with our earlier PAWS tests, so this check should be solely
9003 	 * predicated on the sequence space of this segment. 3) That we
9004 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9005 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9006 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9007 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9008 	 * p.869. In such cases, we can still calculate the RTT correctly
9009 	 * when RCV.NXT == Last.ACK.Sent.
9010 	 */
9011 	if ((to->to_flags & TOF_TS) != 0 &&
9012 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9013 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9014 		    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9015 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9016 		tp->ts_recent = to->to_tsval;
9017 	}
9018 	tp->snd_wnd = tiwin;
9019 	/*
9020 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9021 	 * is on (half-synchronized state), then queue data for later
9022 	 * processing; else drop segment and return.
9023 	 */
9024 	if ((thflags & TH_ACK) == 0) {
9025 		if (IS_FASTOPEN(tp->t_flags)) {
9026 			cc_conn_init(tp);
9027 		}
9028 		return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9029 					 tiwin, thflags, nxt_pkt));
9030 	}
9031 	KMOD_TCPSTAT_INC(tcps_connects);
9032 	if (tp->t_flags & TF_SONOTCONN) {
9033 		tp->t_flags &= ~TF_SONOTCONN;
9034 		soisconnected(so);
9035 	}
9036 	/* Do window scaling? */
9037 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9038 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9039 		tp->rcv_scale = tp->request_r_scale;
9040 	}
9041 	/*
9042 	 * ok for the first time in lets see if we can use the ts to figure
9043 	 * out what the initial RTT was.
9044 	 */
9045 	if ((to->to_flags & TOF_TS) != 0) {
9046 		uint32_t t, rtt;
9047 
9048 		t = tcp_tv_to_mssectick(&bbr->rc_tv);
9049 		if (TSTMP_GEQ(t, to->to_tsecr)) {
9050 			rtt = t - to->to_tsecr;
9051 			if (rtt == 0) {
9052 				rtt = 1;
9053 			}
9054 			rtt *= MS_IN_USEC;
9055 			tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9056 			apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime);
9057 		}
9058 	}
9059 	/* Drop off any SYN in the send map (probably not there)  */
9060 	if (thflags & TH_ACK)
9061 		bbr_log_syn(tp, to);
9062 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
9063 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9064 		tp->t_tfo_pending = NULL;
9065 	}
9066 	/*
9067 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
9068 	 * FIN-WAIT-1
9069 	 */
9070 	tp->t_starttime = ticks;
9071 	if (tp->t_flags & TF_NEEDFIN) {
9072 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
9073 		tp->t_flags &= ~TF_NEEDFIN;
9074 	} else {
9075 		tcp_state_change(tp, TCPS_ESTABLISHED);
9076 		TCP_PROBE5(accept__established, NULL, tp,
9077 			   mtod(m, const char *), tp, th);
9078 		/*
9079 		 * TFO connections call cc_conn_init() during SYN
9080 		 * processing.  Calling it again here for such connections
9081 		 * is not harmless as it would undo the snd_cwnd reduction
9082 		 * that occurs when a TFO SYN|ACK is retransmitted.
9083 		 */
9084 		if (!IS_FASTOPEN(tp->t_flags))
9085 			cc_conn_init(tp);
9086 	}
9087 	/*
9088 	 * Account for the ACK of our SYN prior to
9089 	 * regular ACK processing below, except for
9090 	 * simultaneous SYN, which is handled later.
9091 	 */
9092 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
9093 		tp->snd_una++;
9094 	/*
9095 	 * If segment contains data or ACK, will call tcp_reass() later; if
9096 	 * not, do so now to pass queued data to user.
9097 	 */
9098 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
9099 		(void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9100 			(struct mbuf *)0);
9101 		if (tp->t_flags & TF_WAKESOR) {
9102 			tp->t_flags &= ~TF_WAKESOR;
9103 			/* NB: sorwakeup_locked() does an implicit unlock. */
9104 			sorwakeup_locked(so);
9105 		}
9106 	}
9107 	tp->snd_wl1 = th->th_seq - 1;
9108 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9109 		return (ret_val);
9110 	}
9111 	if (tp->t_state == TCPS_FIN_WAIT_1) {
9112 		/* We could have went to FIN_WAIT_1 (or EST) above */
9113 		/*
9114 		 * In FIN_WAIT_1 STATE in addition to the processing for the
9115 		 * ESTABLISHED state if our FIN is now acknowledged then
9116 		 * enter FIN_WAIT_2.
9117 		 */
9118 		if (ourfinisacked) {
9119 			/*
9120 			 * If we can't receive any more data, then closing
9121 			 * user can proceed. Starting the timer is contrary
9122 			 * to the specification, but if we don't get a FIN
9123 			 * we'll hang forever.
9124 			 *
9125 			 * XXXjl: we should release the tp also, and use a
9126 			 * compressed state.
9127 			 */
9128 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9129 				soisdisconnected(so);
9130 				tcp_timer_activate(tp, TT_2MSL,
9131 						   (tcp_fast_finwait2_recycle ?
9132 						    tcp_finwait2_timeout :
9133 						    TP_MAXIDLE(tp)));
9134 			}
9135 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
9136 		}
9137 	}
9138 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9139 				 tiwin, thflags, nxt_pkt));
9140 }
9141 
9142 /*
9143  * Return value of 1, the TCB is unlocked and most
9144  * likely gone, return value of 0, the TCB is still
9145  * locked.
9146  */
9147 static int
9148 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9149     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9150     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9151 {
9152 	struct tcp_bbr *bbr;
9153 	int32_t ret_val;
9154 
9155 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9156 
9157 	/*
9158 	 * Header prediction: check for the two common cases of a
9159 	 * uni-directional data xfer.  If the packet has no control flags,
9160 	 * is in-sequence, the window didn't change and we're not
9161 	 * retransmitting, it's a candidate.  If the length is zero and the
9162 	 * ack moved forward, we're the sender side of the xfer.  Just free
9163 	 * the data acked & wake any higher level process that was blocked
9164 	 * waiting for space.  If the length is non-zero and the ack didn't
9165 	 * move, we're the receiver side.  If we're getting packets in-order
9166 	 * (the reassembly queue is empty), add the data toc The socket
9167 	 * buffer and note that we need a delayed ack. Make sure that the
9168 	 * hidden state-flags are also off. Since we check for
9169 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9170 	 */
9171 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9172 	if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) {
9173 		/*
9174 		 * If we have delived under 4 segments increase the initial
9175 		 * window if raised by the peer. We use this to determine
9176 		 * dynamic and static rwnd's at the end of a connection.
9177 		 */
9178 		bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd);
9179 	}
9180 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9181 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) &&
9182 	    __predict_true(SEGQ_EMPTY(tp)) &&
9183 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
9184 		if (tlen == 0) {
9185 			if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9186 			    tiwin, nxt_pkt, iptos)) {
9187 				return (0);
9188 			}
9189 		} else {
9190 			if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9191 			    tiwin, nxt_pkt)) {
9192 				return (0);
9193 			}
9194 		}
9195 	}
9196 	ctf_calc_rwin(so, tp);
9197 
9198 	if ((thflags & TH_RST) ||
9199 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9200 		return (ctf_process_rst(m, th, so, tp));
9201 	/*
9202 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9203 	 * synchronized state.
9204 	 */
9205 	if (thflags & TH_SYN) {
9206 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9207 		return (ret_val);
9208 	}
9209 	/*
9210 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9211 	 * it's less than ts_recent, drop it.
9212 	 */
9213 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9214 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9215 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9216 			return (ret_val);
9217 	}
9218 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9219 		return (ret_val);
9220 	}
9221 	/*
9222 	 * If last ACK falls within this segment's sequence numbers, record
9223 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9224 	 * from the latest proposal of the tcplw@cray.com list (Braden
9225 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9226 	 * with our earlier PAWS tests, so this check should be solely
9227 	 * predicated on the sequence space of this segment. 3) That we
9228 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9229 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9230 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9231 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9232 	 * p.869. In such cases, we can still calculate the RTT correctly
9233 	 * when RCV.NXT == Last.ACK.Sent.
9234 	 */
9235 	if ((to->to_flags & TOF_TS) != 0 &&
9236 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9237 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9238 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9239 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9240 		tp->ts_recent = to->to_tsval;
9241 	}
9242 	/*
9243 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9244 	 * is on (half-synchronized state), then queue data for later
9245 	 * processing; else drop segment and return.
9246 	 */
9247 	if ((thflags & TH_ACK) == 0) {
9248 		if (tp->t_flags & TF_NEEDSYN) {
9249 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9250 			    tiwin, thflags, nxt_pkt));
9251 		} else if (tp->t_flags & TF_ACKNOW) {
9252 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9253 			bbr->r_wanted_output = 1;
9254 			return (ret_val);
9255 		} else {
9256 			ctf_do_drop(m, NULL);
9257 			return (0);
9258 		}
9259 	}
9260 	/*
9261 	 * Ack processing.
9262 	 */
9263 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9264 		return (ret_val);
9265 	}
9266 	if (sbavail(&so->so_snd)) {
9267 		if (ctf_progress_timeout_check(tp, true)) {
9268 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9269 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9270 			return (1);
9271 		}
9272 	}
9273 	/* State changes only happen in bbr_process_data() */
9274 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9275 	    tiwin, thflags, nxt_pkt));
9276 }
9277 
9278 /*
9279  * Return value of 1, the TCB is unlocked and most
9280  * likely gone, return value of 0, the TCB is still
9281  * locked.
9282  */
9283 static int
9284 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9285     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9286     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9287 {
9288 	struct tcp_bbr *bbr;
9289 	int32_t ret_val;
9290 
9291 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9292 
9293 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9294 	ctf_calc_rwin(so, tp);
9295 	if ((thflags & TH_RST) ||
9296 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9297 		return (ctf_process_rst(m, th, so, tp));
9298 	/*
9299 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9300 	 * synchronized state.
9301 	 */
9302 	if (thflags & TH_SYN) {
9303 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9304 		return (ret_val);
9305 	}
9306 	/*
9307 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9308 	 * it's less than ts_recent, drop it.
9309 	 */
9310 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9311 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9312 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9313 			return (ret_val);
9314 	}
9315 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9316 		return (ret_val);
9317 	}
9318 	/*
9319 	 * If last ACK falls within this segment's sequence numbers, record
9320 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9321 	 * from the latest proposal of the tcplw@cray.com list (Braden
9322 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9323 	 * with our earlier PAWS tests, so this check should be solely
9324 	 * predicated on the sequence space of this segment. 3) That we
9325 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9326 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9327 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9328 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9329 	 * p.869. In such cases, we can still calculate the RTT correctly
9330 	 * when RCV.NXT == Last.ACK.Sent.
9331 	 */
9332 	if ((to->to_flags & TOF_TS) != 0 &&
9333 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9334 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9335 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9336 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9337 		tp->ts_recent = to->to_tsval;
9338 	}
9339 	/*
9340 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9341 	 * is on (half-synchronized state), then queue data for later
9342 	 * processing; else drop segment and return.
9343 	 */
9344 	if ((thflags & TH_ACK) == 0) {
9345 		if (tp->t_flags & TF_NEEDSYN) {
9346 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9347 			    tiwin, thflags, nxt_pkt));
9348 		} else if (tp->t_flags & TF_ACKNOW) {
9349 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9350 			bbr->r_wanted_output = 1;
9351 			return (ret_val);
9352 		} else {
9353 			ctf_do_drop(m, NULL);
9354 			return (0);
9355 		}
9356 	}
9357 	/*
9358 	 * Ack processing.
9359 	 */
9360 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9361 		return (ret_val);
9362 	}
9363 	if (sbavail(&so->so_snd)) {
9364 		if (ctf_progress_timeout_check(tp, true)) {
9365 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9366 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9367 			return (1);
9368 		}
9369 	}
9370 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9371 	    tiwin, thflags, nxt_pkt));
9372 }
9373 
9374 static int
9375 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr,
9376     struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so)
9377 {
9378 
9379 	if (bbr->rc_allow_data_af_clo == 0) {
9380 close_now:
9381 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9382 		/* tcp_close will kill the inp pre-log the Reset */
9383 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9384 		tp = tcp_close(tp);
9385 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
9386 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
9387 		return (1);
9388 	}
9389 	if (sbavail(&so->so_snd) == 0)
9390 		goto close_now;
9391 	/* Ok we allow data that is ignored and a followup reset */
9392 	tp->rcv_nxt = th->th_seq + *tlen;
9393 	tp->t_flags2 |= TF2_DROP_AF_DATA;
9394 	bbr->r_wanted_output = 1;
9395 	*tlen = 0;
9396 	return (0);
9397 }
9398 
9399 /*
9400  * Return value of 1, the TCB is unlocked and most
9401  * likely gone, return value of 0, the TCB is still
9402  * locked.
9403  */
9404 static int
9405 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9406     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9407     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9408 {
9409 	int32_t ourfinisacked = 0;
9410 	int32_t ret_val;
9411 	struct tcp_bbr *bbr;
9412 
9413 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9414 
9415 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9416 	ctf_calc_rwin(so, tp);
9417 	if ((thflags & TH_RST) ||
9418 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9419 		return (ctf_process_rst(m, th, so, tp));
9420 	/*
9421 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9422 	 * synchronized state.
9423 	 */
9424 	if (thflags & TH_SYN) {
9425 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9426 		return (ret_val);
9427 	}
9428 	/*
9429 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9430 	 * it's less than ts_recent, drop it.
9431 	 */
9432 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9433 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9434 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9435 			return (ret_val);
9436 	}
9437 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9438 		return (ret_val);
9439 	}
9440 	/*
9441 	 * If new data are received on a connection after the user processes
9442 	 * are gone, then RST the other end.
9443 	 * We call a new function now so we might continue and setup
9444 	 * to reset at all data being ack'd.
9445 	 */
9446 	if ((tp->t_flags & TF_CLOSED) && tlen &&
9447 	    bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9448 		return (1);
9449 	/*
9450 	 * If last ACK falls within this segment's sequence numbers, record
9451 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9452 	 * from the latest proposal of the tcplw@cray.com list (Braden
9453 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9454 	 * with our earlier PAWS tests, so this check should be solely
9455 	 * predicated on the sequence space of this segment. 3) That we
9456 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9457 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9458 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9459 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9460 	 * p.869. In such cases, we can still calculate the RTT correctly
9461 	 * when RCV.NXT == Last.ACK.Sent.
9462 	 */
9463 	if ((to->to_flags & TOF_TS) != 0 &&
9464 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9465 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9466 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9467 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9468 		tp->ts_recent = to->to_tsval;
9469 	}
9470 	/*
9471 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9472 	 * is on (half-synchronized state), then queue data for later
9473 	 * processing; else drop segment and return.
9474 	 */
9475 	if ((thflags & TH_ACK) == 0) {
9476 		if (tp->t_flags & TF_NEEDSYN) {
9477 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9478 			    tiwin, thflags, nxt_pkt));
9479 		} else if (tp->t_flags & TF_ACKNOW) {
9480 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9481 			bbr->r_wanted_output = 1;
9482 			return (ret_val);
9483 		} else {
9484 			ctf_do_drop(m, NULL);
9485 			return (0);
9486 		}
9487 	}
9488 	/*
9489 	 * Ack processing.
9490 	 */
9491 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9492 		return (ret_val);
9493 	}
9494 	if (ourfinisacked) {
9495 		/*
9496 		 * If we can't receive any more data, then closing user can
9497 		 * proceed. Starting the timer is contrary to the
9498 		 * specification, but if we don't get a FIN we'll hang
9499 		 * forever.
9500 		 *
9501 		 * XXXjl: we should release the tp also, and use a
9502 		 * compressed state.
9503 		 */
9504 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9505 			soisdisconnected(so);
9506 			tcp_timer_activate(tp, TT_2MSL,
9507 			    (tcp_fast_finwait2_recycle ?
9508 			    tcp_finwait2_timeout :
9509 			    TP_MAXIDLE(tp)));
9510 		}
9511 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
9512 	}
9513 	if (sbavail(&so->so_snd)) {
9514 		if (ctf_progress_timeout_check(tp, true)) {
9515 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9516 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9517 			return (1);
9518 		}
9519 	}
9520 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9521 	    tiwin, thflags, nxt_pkt));
9522 }
9523 
9524 /*
9525  * Return value of 1, the TCB is unlocked and most
9526  * likely gone, return value of 0, the TCB is still
9527  * locked.
9528  */
9529 static int
9530 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9531     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9532     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9533 {
9534 	int32_t ourfinisacked = 0;
9535 	int32_t ret_val;
9536 	struct tcp_bbr *bbr;
9537 
9538 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9539 
9540 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9541 	ctf_calc_rwin(so, tp);
9542 	if ((thflags & TH_RST) ||
9543 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9544 		return (ctf_process_rst(m, th, so, tp));
9545 	/*
9546 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9547 	 * synchronized state.
9548 	 */
9549 	if (thflags & TH_SYN) {
9550 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9551 		return (ret_val);
9552 	}
9553 	/*
9554 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9555 	 * it's less than ts_recent, drop it.
9556 	 */
9557 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9558 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9559 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9560 			return (ret_val);
9561 	}
9562 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9563 		return (ret_val);
9564 	}
9565 	/*
9566 	 * If new data are received on a connection after the user processes
9567 	 * are gone, then RST the other end.
9568 	 * We call a new function now so we might continue and setup
9569 	 * to reset at all data being ack'd.
9570 	 */
9571 	if ((tp->t_flags & TF_CLOSED) && tlen &&
9572 	    bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9573 		return (1);
9574 	/*
9575 	 * If last ACK falls within this segment's sequence numbers, record
9576 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9577 	 * from the latest proposal of the tcplw@cray.com list (Braden
9578 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9579 	 * with our earlier PAWS tests, so this check should be solely
9580 	 * predicated on the sequence space of this segment. 3) That we
9581 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9582 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9583 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9584 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9585 	 * p.869. In such cases, we can still calculate the RTT correctly
9586 	 * when RCV.NXT == Last.ACK.Sent.
9587 	 */
9588 	if ((to->to_flags & TOF_TS) != 0 &&
9589 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9590 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9591 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9592 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9593 		tp->ts_recent = to->to_tsval;
9594 	}
9595 	/*
9596 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9597 	 * is on (half-synchronized state), then queue data for later
9598 	 * processing; else drop segment and return.
9599 	 */
9600 	if ((thflags & TH_ACK) == 0) {
9601 		if (tp->t_flags & TF_NEEDSYN) {
9602 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9603 			    tiwin, thflags, nxt_pkt));
9604 		} else if (tp->t_flags & TF_ACKNOW) {
9605 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9606 			bbr->r_wanted_output = 1;
9607 			return (ret_val);
9608 		} else {
9609 			ctf_do_drop(m, NULL);
9610 			return (0);
9611 		}
9612 	}
9613 	/*
9614 	 * Ack processing.
9615 	 */
9616 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9617 		return (ret_val);
9618 	}
9619 	if (ourfinisacked) {
9620 		tcp_twstart(tp);
9621 		m_freem(m);
9622 		return (1);
9623 	}
9624 	if (sbavail(&so->so_snd)) {
9625 		if (ctf_progress_timeout_check(tp, true)) {
9626 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9627 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9628 			return (1);
9629 		}
9630 	}
9631 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9632 	    tiwin, thflags, nxt_pkt));
9633 }
9634 
9635 /*
9636  * Return value of 1, the TCB is unlocked and most
9637  * likely gone, return value of 0, the TCB is still
9638  * locked.
9639  */
9640 static int
9641 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9642     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9643     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9644 {
9645 	int32_t ourfinisacked = 0;
9646 	int32_t ret_val;
9647 	struct tcp_bbr *bbr;
9648 
9649 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9650 
9651 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9652 	ctf_calc_rwin(so, tp);
9653 	if ((thflags & TH_RST) ||
9654 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9655 		return (ctf_process_rst(m, th, so, tp));
9656 	/*
9657 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9658 	 * synchronized state.
9659 	 */
9660 	if (thflags & TH_SYN) {
9661 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9662 		return (ret_val);
9663 	}
9664 	/*
9665 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9666 	 * it's less than ts_recent, drop it.
9667 	 */
9668 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9669 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9670 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9671 			return (ret_val);
9672 	}
9673 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9674 		return (ret_val);
9675 	}
9676 	/*
9677 	 * If new data are received on a connection after the user processes
9678 	 * are gone, then RST the other end.
9679 	 * We call a new function now so we might continue and setup
9680 	 * to reset at all data being ack'd.
9681 	 */
9682 	if ((tp->t_flags & TF_CLOSED) && tlen &&
9683 	    bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9684 		return (1);
9685 	/*
9686 	 * If last ACK falls within this segment's sequence numbers, record
9687 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9688 	 * from the latest proposal of the tcplw@cray.com list (Braden
9689 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9690 	 * with our earlier PAWS tests, so this check should be solely
9691 	 * predicated on the sequence space of this segment. 3) That we
9692 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9693 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9694 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9695 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9696 	 * p.869. In such cases, we can still calculate the RTT correctly
9697 	 * when RCV.NXT == Last.ACK.Sent.
9698 	 */
9699 	if ((to->to_flags & TOF_TS) != 0 &&
9700 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9701 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9702 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9703 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9704 		tp->ts_recent = to->to_tsval;
9705 	}
9706 	/*
9707 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9708 	 * is on (half-synchronized state), then queue data for later
9709 	 * processing; else drop segment and return.
9710 	 */
9711 	if ((thflags & TH_ACK) == 0) {
9712 		if (tp->t_flags & TF_NEEDSYN) {
9713 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9714 			    tiwin, thflags, nxt_pkt));
9715 		} else if (tp->t_flags & TF_ACKNOW) {
9716 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9717 			bbr->r_wanted_output = 1;
9718 			return (ret_val);
9719 		} else {
9720 			ctf_do_drop(m, NULL);
9721 			return (0);
9722 		}
9723 	}
9724 	/*
9725 	 * case TCPS_LAST_ACK: Ack processing.
9726 	 */
9727 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9728 		return (ret_val);
9729 	}
9730 	if (ourfinisacked) {
9731 		tp = tcp_close(tp);
9732 		ctf_do_drop(m, tp);
9733 		return (1);
9734 	}
9735 	if (sbavail(&so->so_snd)) {
9736 		if (ctf_progress_timeout_check(tp, true)) {
9737 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9738 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9739 			return (1);
9740 		}
9741 	}
9742 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9743 	    tiwin, thflags, nxt_pkt));
9744 }
9745 
9746 /*
9747  * Return value of 1, the TCB is unlocked and most
9748  * likely gone, return value of 0, the TCB is still
9749  * locked.
9750  */
9751 static int
9752 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
9753     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9754     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9755 {
9756 	int32_t ourfinisacked = 0;
9757 	int32_t ret_val;
9758 	struct tcp_bbr *bbr;
9759 
9760 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9761 
9762 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9763 	ctf_calc_rwin(so, tp);
9764 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
9765 	if ((thflags & TH_RST) ||
9766 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9767 		return (ctf_process_rst(m, th, so, tp));
9768 
9769 	/*
9770 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9771 	 * synchronized state.
9772 	 */
9773 	if (thflags & TH_SYN) {
9774 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9775 		return (ret_val);
9776 	}
9777 	/*
9778 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9779 	 * it's less than ts_recent, drop it.
9780 	 */
9781 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9782 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9783 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9784 			return (ret_val);
9785 	}
9786 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9787 		return (ret_val);
9788 	}
9789 	/*
9790 	 * If new data are received on a connection after the user processes
9791 	 * are gone, then we may RST the other end depending on the outcome
9792 	 * of bbr_check_data_after_close.
9793 	 * We call a new function now so we might continue and setup
9794 	 * to reset at all data being ack'd.
9795 	 */
9796 	if ((tp->t_flags & TF_CLOSED) && tlen &&
9797 	    bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9798 		return (1);
9799 	/*
9800 	 * If last ACK falls within this segment's sequence numbers, record
9801 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9802 	 * from the latest proposal of the tcplw@cray.com list (Braden
9803 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9804 	 * with our earlier PAWS tests, so this check should be solely
9805 	 * predicated on the sequence space of this segment. 3) That we
9806 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9807 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9808 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9809 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9810 	 * p.869. In such cases, we can still calculate the RTT correctly
9811 	 * when RCV.NXT == Last.ACK.Sent.
9812 	 */
9813 	if ((to->to_flags & TOF_TS) != 0 &&
9814 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9815 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9816 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9817 		tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9818 		tp->ts_recent = to->to_tsval;
9819 	}
9820 	/*
9821 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9822 	 * is on (half-synchronized state), then queue data for later
9823 	 * processing; else drop segment and return.
9824 	 */
9825 	if ((thflags & TH_ACK) == 0) {
9826 		if (tp->t_flags & TF_NEEDSYN) {
9827 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9828 			    tiwin, thflags, nxt_pkt));
9829 		} else if (tp->t_flags & TF_ACKNOW) {
9830 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9831 			bbr->r_wanted_output = 1;
9832 			return (ret_val);
9833 		} else {
9834 			ctf_do_drop(m, NULL);
9835 			return (0);
9836 		}
9837 	}
9838 	/*
9839 	 * Ack processing.
9840 	 */
9841 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9842 		return (ret_val);
9843 	}
9844 	if (sbavail(&so->so_snd)) {
9845 		if (ctf_progress_timeout_check(tp, true)) {
9846 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9847 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9848 			return (1);
9849 		}
9850 	}
9851 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9852 	    tiwin, thflags, nxt_pkt));
9853 }
9854 
9855 static void
9856 bbr_stop_all_timers(struct tcpcb *tp, struct tcp_bbr *bbr)
9857 {
9858 	/*
9859 	 * Assure no timers are running.
9860 	 */
9861 	if (tcp_timer_active(tp, TT_PERSIST)) {
9862 		/* We enter in persists, set the flag appropriately */
9863 		bbr->rc_in_persist = 1;
9864 	}
9865 }
9866 
9867 static void
9868 bbr_google_mode_on(struct tcp_bbr *bbr)
9869 {
9870 	bbr->rc_use_google = 1;
9871 	bbr->rc_no_pacing = 0;
9872 	bbr->r_ctl.bbr_google_discount = bbr_google_discount;
9873 	bbr->r_use_policer = bbr_policer_detection_enabled;
9874 	bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
9875 	bbr->bbr_use_rack_cheat = 0;
9876 	bbr->r_ctl.rc_incr_tmrs = 0;
9877 	bbr->r_ctl.rc_inc_tcp_oh = 0;
9878 	bbr->r_ctl.rc_inc_ip_oh = 0;
9879 	bbr->r_ctl.rc_inc_enet_oh = 0;
9880 	reset_time(&bbr->r_ctl.rc_delrate,
9881 		   BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
9882 	reset_time_small(&bbr->r_ctl.rc_rttprop,
9883 			 (11 * USECS_IN_SECOND));
9884 	tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
9885 }
9886 
9887 static void
9888 bbr_google_mode_off(struct tcp_bbr *bbr)
9889 {
9890 	bbr->rc_use_google = 0;
9891 	bbr->r_ctl.bbr_google_discount = 0;
9892 	bbr->no_pacing_until = bbr_no_pacing_until;
9893 	bbr->r_use_policer = 0;
9894 	if (bbr->no_pacing_until)
9895 		bbr->rc_no_pacing = 1;
9896 	else
9897 		bbr->rc_no_pacing = 0;
9898 	if (bbr_use_rack_resend_cheat)
9899 		bbr->bbr_use_rack_cheat = 1;
9900 	else
9901 		bbr->bbr_use_rack_cheat = 0;
9902 	if (bbr_incr_timers)
9903 		bbr->r_ctl.rc_incr_tmrs = 1;
9904 	else
9905 		bbr->r_ctl.rc_incr_tmrs = 0;
9906 	if (bbr_include_tcp_oh)
9907 		bbr->r_ctl.rc_inc_tcp_oh = 1;
9908 	else
9909 		bbr->r_ctl.rc_inc_tcp_oh = 0;
9910 	if (bbr_include_ip_oh)
9911 		bbr->r_ctl.rc_inc_ip_oh = 1;
9912 	else
9913 		bbr->r_ctl.rc_inc_ip_oh = 0;
9914 	if (bbr_include_enet_oh)
9915 		bbr->r_ctl.rc_inc_enet_oh = 1;
9916 	else
9917 		bbr->r_ctl.rc_inc_enet_oh = 0;
9918 	bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
9919 	reset_time(&bbr->r_ctl.rc_delrate,
9920 		   bbr_num_pktepo_for_del_limit);
9921 	reset_time_small(&bbr->r_ctl.rc_rttprop,
9922 			 (bbr_filter_len_sec * USECS_IN_SECOND));
9923 	tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
9924 }
9925 /*
9926  * Return 0 on success, non-zero on failure
9927  * which indicates the error (usually no memory).
9928  */
9929 static int
9930 bbr_init(struct tcpcb *tp, void **ptr)
9931 {
9932 	struct inpcb *inp = tptoinpcb(tp);
9933 	struct tcp_bbr *bbr = NULL;
9934 	uint32_t cts;
9935 
9936 	*ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO));
9937 	if (*ptr == NULL) {
9938 		/*
9939 		 * We need to allocate memory but cant. The INP and INP_INFO
9940 		 * locks and they are recursive (happens during setup. So a
9941 		 * scheme to drop the locks fails :(
9942 		 *
9943 		 */
9944 		return (ENOMEM);
9945 	}
9946 	bbr = (struct tcp_bbr *)*ptr;
9947 	bbr->rtt_valid = 0;
9948 	inp->inp_flags2 |= INP_CANNOT_DO_ECN;
9949 	inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
9950 	/* Take off any undesired flags */
9951 	inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
9952 	inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
9953 	inp->inp_flags2 &= ~INP_MBUF_ACKCMP;
9954 	inp->inp_flags2 &= ~INP_MBUF_L_ACKS;
9955 
9956 	TAILQ_INIT(&bbr->r_ctl.rc_map);
9957 	TAILQ_INIT(&bbr->r_ctl.rc_free);
9958 	TAILQ_INIT(&bbr->r_ctl.rc_tmap);
9959 	bbr->rc_tp = tp;
9960 	bbr->rc_inp = inp;
9961 	cts = tcp_get_usecs(&bbr->rc_tv);
9962 	tp->t_acktime = 0;
9963 	bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close;
9964 	bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade;
9965 	bbr->rc_tlp_threshold = bbr_tlp_thresh;
9966 	bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh;
9967 	bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay;
9968 	bbr->r_ctl.rc_min_to = bbr_min_to;
9969 	bbr->rc_bbr_state = BBR_STATE_STARTUP;
9970 	bbr->r_ctl.bbr_lost_at_state = 0;
9971 	bbr->r_ctl.rc_lost_at_startup = 0;
9972 	bbr->rc_all_timers_stopped = 0;
9973 	bbr->r_ctl.rc_bbr_lastbtlbw = 0;
9974 	bbr->r_ctl.rc_pkt_epoch_del = 0;
9975 	bbr->r_ctl.rc_pkt_epoch = 0;
9976 	bbr->r_ctl.rc_lowest_rtt = 0xffffffff;
9977 	bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain;
9978 	bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
9979 	bbr->r_ctl.rc_went_idle_time = cts;
9980 	bbr->rc_pacer_started = cts;
9981 	bbr->r_ctl.rc_pkt_epoch_time = cts;
9982 	bbr->r_ctl.rc_rcvtime = cts;
9983 	bbr->r_ctl.rc_bbr_state_time = cts;
9984 	bbr->r_ctl.rc_del_time = cts;
9985 	bbr->r_ctl.rc_tlp_rxt_last_time = cts;
9986 	bbr->r_ctl.last_in_probertt = cts;
9987 	bbr->skip_gain = 0;
9988 	bbr->gain_is_limited = 0;
9989 	bbr->no_pacing_until = bbr_no_pacing_until;
9990 	if (bbr->no_pacing_until)
9991 		bbr->rc_no_pacing = 1;
9992 	if (bbr_use_google_algo) {
9993 		bbr->rc_no_pacing = 0;
9994 		bbr->rc_use_google = 1;
9995 		bbr->r_ctl.bbr_google_discount = bbr_google_discount;
9996 		bbr->r_use_policer = bbr_policer_detection_enabled;
9997 	} else {
9998 		bbr->rc_use_google = 0;
9999 		bbr->r_ctl.bbr_google_discount = 0;
10000 		bbr->r_use_policer = 0;
10001 	}
10002 	if (bbr_ts_limiting)
10003 		bbr->rc_use_ts_limit = 1;
10004 	else
10005 		bbr->rc_use_ts_limit = 0;
10006 	if (bbr_ts_can_raise)
10007 		bbr->ts_can_raise = 1;
10008 	else
10009 		bbr->ts_can_raise = 0;
10010 	if (V_tcp_delack_enabled == 1)
10011 		tp->t_delayed_ack = 2;
10012 	else if (V_tcp_delack_enabled == 0)
10013 		tp->t_delayed_ack = 0;
10014 	else if (V_tcp_delack_enabled < 100)
10015 		tp->t_delayed_ack = V_tcp_delack_enabled;
10016 	else
10017 		tp->t_delayed_ack = 2;
10018 	if (bbr->rc_use_google == 0)
10019 		bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10020 	else
10021 		bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10022 	bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms;
10023 	bbr->rc_max_rto_sec = bbr_rto_max_sec;
10024 	bbr->rc_init_win = bbr_def_init_win;
10025 	if (tp->t_flags & TF_REQ_TSTMP)
10026 		bbr->rc_last_options = TCP_TS_OVERHEAD;
10027 	bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options;
10028 	bbr->r_ctl.rc_high_rwnd = tp->snd_wnd;
10029 	bbr->r_init_rtt = 1;
10030 
10031 	counter_u64_add(bbr_flows_nohdwr_pacing, 1);
10032 	if (bbr_allow_hdwr_pacing)
10033 		bbr->bbr_hdw_pace_ena = 1;
10034 	else
10035 		bbr->bbr_hdw_pace_ena = 0;
10036 	if (bbr_sends_full_iwnd)
10037 		bbr->bbr_init_win_cheat = 1;
10038 	else
10039 		bbr->bbr_init_win_cheat = 0;
10040 	bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max;
10041 	bbr->r_ctl.rc_drain_pg = bbr_drain_gain;
10042 	bbr->r_ctl.rc_startup_pg = bbr_high_gain;
10043 	bbr->rc_loss_exit = bbr_exit_startup_at_loss;
10044 	bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain;
10045 	bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second;
10046 	bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar;
10047 	bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max;
10048 	bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor;
10049 	bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min;
10050 	bbr->r_ctl.bbr_cross_over = bbr_cross_over;
10051 	bbr->r_ctl.rc_rtt_shrinks = cts;
10052 	if (bbr->rc_use_google) {
10053 		setup_time_filter(&bbr->r_ctl.rc_delrate,
10054 				  FILTER_TYPE_MAX,
10055 				  BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10056 		setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10057 					FILTER_TYPE_MIN, (11 * USECS_IN_SECOND));
10058 	} else {
10059 		setup_time_filter(&bbr->r_ctl.rc_delrate,
10060 				  FILTER_TYPE_MAX,
10061 				  bbr_num_pktepo_for_del_limit);
10062 		setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10063 					FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND));
10064 	}
10065 	bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0);
10066 	if (bbr_uses_idle_restart)
10067 		bbr->rc_use_idle_restart = 1;
10068 	else
10069 		bbr->rc_use_idle_restart = 0;
10070 	bbr->r_ctl.rc_bbr_cur_del_rate = 0;
10071 	bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps;
10072 	if (bbr_resends_use_tso)
10073 		bbr->rc_resends_use_tso = 1;
10074 #ifdef NETFLIX_PEAKRATE
10075 	tp->t_peakrate_thr = tp->t_maxpeakrate;
10076 #endif
10077 	if (tp->snd_una != tp->snd_max) {
10078 		/* Create a send map for the current outstanding data */
10079 		struct bbr_sendmap *rsm;
10080 
10081 		rsm = bbr_alloc(bbr);
10082 		if (rsm == NULL) {
10083 			uma_zfree(bbr_pcb_zone, *ptr);
10084 			*ptr = NULL;
10085 			return (ENOMEM);
10086 		}
10087 		rsm->r_rtt_not_allowed = 1;
10088 		rsm->r_tim_lastsent[0] = cts;
10089 		rsm->r_rtr_cnt = 1;
10090 		rsm->r_rtr_bytes = 0;
10091 		rsm->r_start = tp->snd_una;
10092 		rsm->r_end = tp->snd_max;
10093 		rsm->r_dupack = 0;
10094 		rsm->r_delivered = bbr->r_ctl.rc_delivered;
10095 		rsm->r_ts_valid = 0;
10096 		rsm->r_del_ack_ts = tp->ts_recent;
10097 		rsm->r_del_time = cts;
10098 		if (bbr->r_ctl.r_app_limited_until)
10099 			rsm->r_app_limited = 1;
10100 		else
10101 			rsm->r_app_limited = 0;
10102 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
10103 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
10104 		rsm->r_in_tmap = 1;
10105 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
10106 			rsm->r_bbr_state = bbr_state_val(bbr);
10107 		else
10108 			rsm->r_bbr_state = 8;
10109 	}
10110 	if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0))
10111 		bbr->bbr_use_rack_cheat = 1;
10112 	if (bbr_incr_timers && (bbr->rc_use_google == 0))
10113 		bbr->r_ctl.rc_incr_tmrs = 1;
10114 	if (bbr_include_tcp_oh && (bbr->rc_use_google == 0))
10115 		bbr->r_ctl.rc_inc_tcp_oh = 1;
10116 	if (bbr_include_ip_oh && (bbr->rc_use_google == 0))
10117 		bbr->r_ctl.rc_inc_ip_oh = 1;
10118 	if (bbr_include_enet_oh && (bbr->rc_use_google == 0))
10119 		bbr->r_ctl.rc_inc_enet_oh = 1;
10120 
10121 	bbr_log_type_statechange(bbr, cts, __LINE__);
10122 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
10123 	    (tp->t_srtt)) {
10124 		uint32_t rtt;
10125 
10126 		rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
10127 		apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
10128 	}
10129 	/* announce the settings and state */
10130 	bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT);
10131 	tcp_bbr_tso_size_check(bbr, cts);
10132 	/*
10133 	 * Now call the generic function to start a timer. This will place
10134 	 * the TCB on the hptsi wheel if a timer is needed with appropriate
10135 	 * flags.
10136 	 */
10137 	bbr_stop_all_timers(tp, bbr);
10138 	/*
10139 	 * Validate the timers are not in usec, if they are convert.
10140 	 * BBR should in theory move to USEC and get rid of a
10141 	 * lot of the TICKS_2 calls.. but for now we stay
10142 	 * with tick timers.
10143 	 */
10144 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_TICKS);
10145 	TCPT_RANGESET(tp->t_rxtcur,
10146 	    ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1,
10147 	    tp->t_rttmin, TCPTV_REXMTMAX);
10148 	bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0);
10149 	return (0);
10150 }
10151 
10152 /*
10153  * Return 0 if we can accept the connection. Return
10154  * non-zero if we can't handle the connection. A EAGAIN
10155  * means you need to wait until the connection is up.
10156  * a EADDRNOTAVAIL means we can never handle the connection
10157  * (no SACK).
10158  */
10159 static int
10160 bbr_handoff_ok(struct tcpcb *tp)
10161 {
10162 	if ((tp->t_state == TCPS_CLOSED) ||
10163 	    (tp->t_state == TCPS_LISTEN)) {
10164 		/* Sure no problem though it may not stick */
10165 		return (0);
10166 	}
10167 	if ((tp->t_state == TCPS_SYN_SENT) ||
10168 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
10169 		/*
10170 		 * We really don't know you have to get to ESTAB or beyond
10171 		 * to tell.
10172 		 */
10173 		return (EAGAIN);
10174 	}
10175 	if (tp->t_flags & TF_SENTFIN)
10176 		return (EINVAL);
10177 	if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) {
10178 		return (0);
10179 	}
10180 	/*
10181 	 * If we reach here we don't do SACK on this connection so we can
10182 	 * never do rack.
10183 	 */
10184 	return (EINVAL);
10185 }
10186 
10187 static void
10188 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10189 {
10190 	if (tp->t_fb_ptr) {
10191 		uint32_t calc;
10192 		struct tcp_bbr *bbr;
10193 		struct bbr_sendmap *rsm;
10194 
10195 		bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10196 		if (bbr->r_ctl.crte)
10197 			tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
10198 		bbr_log_flowend(bbr);
10199 		bbr->rc_tp = NULL;
10200 		if (bbr->bbr_hdrw_pacing)
10201 			counter_u64_add(bbr_flows_whdwr_pacing, -1);
10202 		else
10203 			counter_u64_add(bbr_flows_nohdwr_pacing, -1);
10204 		if (bbr->r_ctl.crte != NULL) {
10205 			tcp_rel_pacing_rate(bbr->r_ctl.crte, tp);
10206 			bbr->r_ctl.crte = NULL;
10207 		}
10208 		rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10209 		while (rsm) {
10210 			TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
10211 			uma_zfree(bbr_zone, rsm);
10212 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10213 		}
10214 		rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10215 		while (rsm) {
10216 			TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
10217 			uma_zfree(bbr_zone, rsm);
10218 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10219 		}
10220 		calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd;
10221 		if (calc > (bbr->r_ctl.rc_init_rwnd / 10))
10222 			BBR_STAT_INC(bbr_dynamic_rwnd);
10223 		else
10224 			BBR_STAT_INC(bbr_static_rwnd);
10225 		bbr->r_ctl.rc_free_cnt = 0;
10226 		uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10227 		tp->t_fb_ptr = NULL;
10228 	}
10229 	/* Make sure snd_nxt is correctly set */
10230 	tp->snd_nxt = tp->snd_max;
10231 }
10232 
10233 static void
10234 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win)
10235 {
10236 	switch (tp->t_state) {
10237 	case TCPS_SYN_SENT:
10238 		bbr->r_state = TCPS_SYN_SENT;
10239 		bbr->r_substate = bbr_do_syn_sent;
10240 		break;
10241 	case TCPS_SYN_RECEIVED:
10242 		bbr->r_state = TCPS_SYN_RECEIVED;
10243 		bbr->r_substate = bbr_do_syn_recv;
10244 		break;
10245 	case TCPS_ESTABLISHED:
10246 		bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd);
10247 		bbr->r_state = TCPS_ESTABLISHED;
10248 		bbr->r_substate = bbr_do_established;
10249 		break;
10250 	case TCPS_CLOSE_WAIT:
10251 		bbr->r_state = TCPS_CLOSE_WAIT;
10252 		bbr->r_substate = bbr_do_close_wait;
10253 		break;
10254 	case TCPS_FIN_WAIT_1:
10255 		bbr->r_state = TCPS_FIN_WAIT_1;
10256 		bbr->r_substate = bbr_do_fin_wait_1;
10257 		break;
10258 	case TCPS_CLOSING:
10259 		bbr->r_state = TCPS_CLOSING;
10260 		bbr->r_substate = bbr_do_closing;
10261 		break;
10262 	case TCPS_LAST_ACK:
10263 		bbr->r_state = TCPS_LAST_ACK;
10264 		bbr->r_substate = bbr_do_lastack;
10265 		break;
10266 	case TCPS_FIN_WAIT_2:
10267 		bbr->r_state = TCPS_FIN_WAIT_2;
10268 		bbr->r_substate = bbr_do_fin_wait_2;
10269 		break;
10270 	case TCPS_LISTEN:
10271 	case TCPS_CLOSED:
10272 	case TCPS_TIME_WAIT:
10273 	default:
10274 		break;
10275 	};
10276 }
10277 
10278 static void
10279 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog)
10280 {
10281 	/*
10282 	 * Now what state are we going into now? Is there adjustments
10283 	 * needed?
10284 	 */
10285 	int32_t old_state;
10286 
10287 	old_state = bbr_state_val(bbr);
10288 	if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) {
10289 		/* Save the lowest srtt we saw in our end of the sub-state */
10290 		bbr->rc_hit_state_1 = 0;
10291 		if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff)
10292 			bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state;
10293 	}
10294 	bbr->rc_bbr_substate++;
10295 	if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) {
10296 		/* Cycle back to first state-> gain */
10297 		bbr->rc_bbr_substate = 0;
10298 	}
10299 	if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10300 		/*
10301 		 * We enter the gain(5/4) cycle (possibly less if
10302 		 * shallow buffer detection is enabled)
10303 		 */
10304 		if (bbr->skip_gain) {
10305 			/*
10306 			 * Hardware pacing has set our rate to
10307 			 * the max and limited our b/w just
10308 			 * do level i.e. no gain.
10309 			 */
10310 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1];
10311 		} else if (bbr->gain_is_limited &&
10312 			   bbr->bbr_hdrw_pacing &&
10313 			   bbr->r_ctl.crte) {
10314 			/*
10315 			 * We can't gain above the hardware pacing
10316 			 * rate which is less than our rate + the gain
10317 			 * calculate the gain needed to reach the hardware
10318 			 * pacing rate..
10319 			 */
10320 			uint64_t bw, rate, gain_calc;
10321 
10322 			bw = bbr_get_bw(bbr);
10323 			rate = bbr->r_ctl.crte->rate;
10324 			if ((rate > bw) &&
10325 			    (((bw *  (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) {
10326 				gain_calc = (rate * BBR_UNIT) / bw;
10327 				if (gain_calc < BBR_UNIT)
10328 					gain_calc = BBR_UNIT;
10329 				bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc;
10330 			} else {
10331 				bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10332 			}
10333 		} else
10334 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10335 		if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) {
10336 			bbr->r_ctl.rc_bbr_state_atflight = cts;
10337 		} else
10338 			bbr->r_ctl.rc_bbr_state_atflight = 0;
10339 	} else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10340 		bbr->rc_hit_state_1 = 1;
10341 		bbr->r_ctl.rc_exta_time_gd = 0;
10342 		bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10343 						     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10344 		if (bbr_state_drain_2_tar) {
10345 			bbr->r_ctl.rc_bbr_state_atflight = 0;
10346 		} else
10347 			bbr->r_ctl.rc_bbr_state_atflight = cts;
10348 		bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN];
10349 	} else {
10350 		/* All other cycles hit here 2-7 */
10351 		if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) {
10352 			if (bbr_sub_drain_slam_cwnd &&
10353 			    (bbr->rc_use_google == 0) &&
10354 			    (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10355 				bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10356 				bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10357 			}
10358 			if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP))
10359 				bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) -
10360 							       bbr_get_rtt(bbr, BBR_RTT_PROP));
10361 			else
10362 				bbr->r_ctl.rc_exta_time_gd = 0;
10363 			if (bbr->r_ctl.rc_exta_time_gd) {
10364 				bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd;
10365 				/* Now chop up the time for each state (div by 7) */
10366 				bbr->r_ctl.rc_level_state_extra /= 7;
10367 				if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) {
10368 					/* Add a randomization */
10369 					bbr_randomize_extra_state_time(bbr);
10370 				}
10371 			}
10372 		}
10373 		bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10374 		bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)];
10375 	}
10376 	if (bbr->rc_use_google) {
10377 		bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10378 	}
10379 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10380 	bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10381 	if (dolog)
10382 		bbr_log_type_statechange(bbr, cts, line);
10383 
10384 	if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10385 		uint32_t time_in;
10386 
10387 		time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10388 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
10389 			counter_u64_add(bbr_state_time[(old_state + 5)], time_in);
10390 		} else {
10391 			counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10392 		}
10393 	}
10394 	bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
10395 	bbr_set_state_target(bbr, __LINE__);
10396 	if (bbr_sub_drain_slam_cwnd &&
10397 	    (bbr->rc_use_google == 0) &&
10398 	    (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10399 		/* Slam down the cwnd */
10400 		bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10401 		bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10402 		if (bbr_sub_drain_app_limit) {
10403 			/* Go app limited if we are on a long drain */
10404 			bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered +
10405 							  ctf_flight_size(bbr->rc_tp,
10406 							      (bbr->r_ctl.rc_sacked +
10407 							       bbr->r_ctl.rc_lost_bytes)));
10408 		}
10409 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10410 	}
10411 	if (bbr->rc_lt_use_bw) {
10412 		/* In policed mode we clamp pacing_gain to BBR_UNIT */
10413 		bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10414 	}
10415 	/* Google changes TSO size every cycle */
10416 	if (bbr->rc_use_google)
10417 		tcp_bbr_tso_size_check(bbr, cts);
10418 	bbr->r_ctl.gain_epoch = cts;
10419 	bbr->r_ctl.rc_bbr_state_time = cts;
10420 	bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch;
10421 }
10422 
10423 static void
10424 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10425 {
10426 	if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) &&
10427 	    (google_allow_early_out == 1) &&
10428 	    (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) {
10429 		/* We have reached out target flight size possibly early */
10430 		goto change_state;
10431 	}
10432 	if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10433 		return;
10434 	}
10435 	if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) {
10436 		/*
10437 		 * Must be a rttProp movement forward before
10438 		 * we can change states.
10439 		 */
10440 		return;
10441 	}
10442 	if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10443 		/*
10444 		 * The needed time has passed but for
10445 		 * the gain cycle extra rules apply:
10446 		 * 1) If we have seen loss, we exit
10447 		 * 2) If we have not reached the target
10448 		 *    we stay in GAIN (gain-to-target).
10449 		 */
10450 		if (google_consider_lost && losses)
10451 			goto change_state;
10452 		if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) {
10453 			return;
10454 		}
10455 	}
10456 change_state:
10457 	/* For gain we must reach our target, all others last 1 rttProp */
10458 	bbr_substate_change(bbr, cts, __LINE__, 1);
10459 }
10460 
10461 static void
10462 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10463 {
10464 	uint32_t flight, bbr_cur_cycle_time;
10465 
10466 	if (bbr->rc_use_google) {
10467 		bbr_set_probebw_google_gains(bbr, cts, losses);
10468 		return;
10469 	}
10470 	if (cts == 0) {
10471 		/*
10472 		 * Never alow cts to be 0 we
10473 		 * do this so we can judge if
10474 		 * we have set a timestamp.
10475 		 */
10476 		cts = 1;
10477 	}
10478 	if (bbr_state_is_pkt_epoch)
10479 		bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
10480 	else
10481 		bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP);
10482 
10483 	if (bbr->r_ctl.rc_bbr_state_atflight == 0) {
10484 		if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10485 			flight = ctf_flight_size(bbr->rc_tp,
10486 				     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10487 			if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) {
10488 				/* Keep it slam down */
10489 				if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) {
10490 					bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10491 					bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10492 				}
10493 				if (bbr_sub_drain_app_limit) {
10494 					/* Go app limited if we are on a long drain */
10495 					bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight);
10496 				}
10497 			}
10498 			if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) &&
10499 			    (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) ||
10500 			     (flight >= bbr->r_ctl.flightsize_at_drain))) {
10501 				/*
10502 				 * Still here after the same time as
10503 				 * the gain. We need to drain harder
10504 				 * for the next srtt. Reduce by a set amount
10505 				 * the gain drop is capped at DRAIN states
10506 				 * value (88).
10507 				 */
10508 				bbr->r_ctl.flightsize_at_drain = flight;
10509 				if (bbr_drain_drop_mul &&
10510 				    bbr_drain_drop_div &&
10511 				    (bbr_drain_drop_mul < bbr_drain_drop_div)) {
10512 					/* Use your specific drop value (def 4/5 = 20%) */
10513 					bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul;
10514 					bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div;
10515 				} else {
10516 					/* You get drop of 20% */
10517 					bbr->r_ctl.rc_bbr_hptsi_gain *= 4;
10518 					bbr->r_ctl.rc_bbr_hptsi_gain /= 5;
10519 				}
10520 				if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) {
10521 					/* Reduce our gain again to the bottom  */
10522 					bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
10523 				}
10524 				bbr_log_exit_gain(bbr, cts, 4);
10525 				/*
10526 				 * Extend out so we wait another
10527 				 * epoch before dropping again.
10528 				 */
10529 				bbr->r_ctl.gain_epoch = cts;
10530 			}
10531 			if (flight <= bbr->r_ctl.rc_target_at_state) {
10532 				if (bbr_sub_drain_slam_cwnd &&
10533 				    (bbr->rc_use_google == 0) &&
10534 				    (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10535 					bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10536 					bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10537 				}
10538 				bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10539 				bbr_log_exit_gain(bbr, cts, 3);
10540 			}
10541 		} else {
10542 			/* Its a gain  */
10543 			if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) {
10544 				bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10545 				goto change_state;
10546 			}
10547 			if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) ||
10548 			    ((ctf_outstanding(bbr->rc_tp) +  bbr->rc_tp->t_maxseg - 1) >=
10549 			     bbr->rc_tp->snd_wnd)) {
10550 				bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10551 				bbr_log_exit_gain(bbr, cts, 2);
10552 			}
10553 		}
10554 		/**
10555 		 * We fall through and return always one of two things has
10556 		 * occurred.
10557 		 * 1) We are still not at target
10558 		 *    <or>
10559 		 * 2) We reached the target and set rc_bbr_state_atflight
10560 		 *    which means we no longer hit this block
10561 		 *    next time we are called.
10562 		 */
10563 		return;
10564 	}
10565 change_state:
10566 	if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time))
10567 		return;
10568 	if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) {
10569 		/* Less than a full time-period has passed */
10570 		return;
10571 	}
10572 	if (bbr->r_ctl.rc_level_state_extra &&
10573 	    (bbr_state_val(bbr) > BBR_SUB_DRAIN) &&
10574 	    ((cts - bbr->r_ctl.rc_bbr_state_time) <
10575 	     (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10576 		/* Less than a full time-period + extra has passed */
10577 		return;
10578 	}
10579 	if (bbr_gain_gets_extra_too &&
10580 	    bbr->r_ctl.rc_level_state_extra &&
10581 	    (bbr_state_val(bbr) == BBR_SUB_GAIN) &&
10582 	    ((cts - bbr->r_ctl.rc_bbr_state_time) <
10583 	     (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10584 		/* Less than a full time-period + extra has passed */
10585 		return;
10586 	}
10587 	bbr_substate_change(bbr, cts, __LINE__, 1);
10588 }
10589 
10590 static uint32_t
10591 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain)
10592 {
10593 	uint32_t mss, tar;
10594 
10595 	if (bbr->rc_use_google) {
10596 		/* Google just uses the cwnd target */
10597 		tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain);
10598 	} else {
10599 		mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options),
10600 			  bbr->r_ctl.rc_pace_max_segs);
10601 		/* Get the base cwnd with gain rounded to a mss */
10602 		tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr),
10603 						      gain), mss);
10604 		/* Make sure it is within our min */
10605 		if (tar < get_min_cwnd(bbr))
10606 			return (get_min_cwnd(bbr));
10607 	}
10608 	return (tar);
10609 }
10610 
10611 static void
10612 bbr_set_state_target(struct tcp_bbr *bbr, int line)
10613 {
10614 	uint32_t tar, meth;
10615 
10616 	if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
10617 	    ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
10618 		/* Special case using old probe-rtt method */
10619 		tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10620 		meth = 1;
10621 	} else {
10622 		/* Non-probe-rtt case and reduced probe-rtt  */
10623 		if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
10624 		    (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) {
10625 			/* For gain cycle we use the hptsi gain */
10626 			tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10627 			meth = 2;
10628 		} else if ((bbr_target_is_bbunit) || bbr->rc_use_google) {
10629 			/*
10630 			 * If configured, or for google all other states
10631 			 * get BBR_UNIT.
10632 			 */
10633 			tar = bbr_get_a_state_target(bbr, BBR_UNIT);
10634 			meth = 3;
10635 		} else {
10636 			/*
10637 			 * Or we set a target based on the pacing gain
10638 			 * for non-google mode and default (non-configured).
10639 			 * Note we don't set a target goal below drain (192).
10640 			 */
10641 			if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN])  {
10642 				tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]);
10643 				meth = 4;
10644 			} else {
10645 				tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10646 				meth = 5;
10647 			}
10648 		}
10649 	}
10650 	bbr_log_set_of_state_target(bbr, tar, line, meth);
10651 	bbr->r_ctl.rc_target_at_state = tar;
10652 }
10653 
10654 static void
10655 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
10656 {
10657 	/* Change to probe_rtt */
10658 	uint32_t time_in;
10659 
10660 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10661 	bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10662 					     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10663 	bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain
10664 					  + bbr->r_ctl.rc_delivered);
10665 	/* Setup so we force feed the filter */
10666 	if (bbr->rc_use_google || bbr_probertt_sets_rtt)
10667 		bbr->rc_prtt_set_ts = 1;
10668 	if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10669 		time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10670 		counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10671 	}
10672 	bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0);
10673 	bbr->r_ctl.rc_rtt_shrinks = cts;
10674 	bbr->r_ctl.last_in_probertt = cts;
10675 	bbr->r_ctl.rc_probertt_srttchktim = cts;
10676 	bbr->r_ctl.rc_bbr_state_time = cts;
10677 	bbr->rc_bbr_state = BBR_STATE_PROBE_RTT;
10678 	/* We need to force the filter to update */
10679 
10680 	if ((bbr_sub_drain_slam_cwnd) &&
10681 	    bbr->rc_hit_state_1 &&
10682 	    (bbr->rc_use_google == 0) &&
10683 	    (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10684 		if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd)
10685 			bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10686 	} else
10687 		bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10688 	/* Update the lost */
10689 	bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10690 	if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){
10691 		/* Set to the non-configurable default of 4 (PROBE_RTT_MIN)  */
10692 		bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10693 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10694 		bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10695 		bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10696 		bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6);
10697 		bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd;
10698 	} else {
10699 		/*
10700 		 * We bring it down slowly by using a hptsi gain that is
10701 		 * probably 75%. This will slowly float down our outstanding
10702 		 * without tampering with the cwnd.
10703 		 */
10704 		bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
10705 		bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10706 		bbr_set_state_target(bbr, __LINE__);
10707 		if (bbr_prtt_slam_cwnd &&
10708 		    (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
10709 			bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10710 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10711 		}
10712 	}
10713 	if (ctf_flight_size(bbr->rc_tp,
10714 		(bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
10715 	    bbr->r_ctl.rc_target_at_state) {
10716 		/* We are at target */
10717 		bbr->r_ctl.rc_bbr_enters_probertt = cts;
10718 	} else {
10719 		/* We need to come down to reach target before our time begins */
10720 		bbr->r_ctl.rc_bbr_enters_probertt = 0;
10721 	}
10722 	bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch;
10723 	BBR_STAT_INC(bbr_enter_probertt);
10724 	bbr_log_exit_gain(bbr, cts, 0);
10725 	bbr_log_type_statechange(bbr, cts, line);
10726 }
10727 
10728 static void
10729 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts)
10730 {
10731 	/*
10732 	 * Sanity check on probe-rtt intervals.
10733 	 * In crazy situations where we are competing
10734 	 * against new-reno flows with huge buffers
10735 	 * our rtt-prop interval could come to dominate
10736 	 * things if we can't get through a full set
10737 	 * of cycles, we need to adjust it.
10738 	 */
10739 	if (bbr_can_adjust_probertt &&
10740 	    (bbr->rc_use_google == 0)) {
10741 		uint16_t val = 0;
10742 		uint32_t cur_rttp, fval, newval, baseval;
10743 
10744 		/* Are we to small and go into probe-rtt to often? */
10745 		baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1));
10746 		cur_rttp = roundup(baseval, USECS_IN_SECOND);
10747 		fval = bbr_filter_len_sec * USECS_IN_SECOND;
10748 		if (bbr_is_ratio == 0) {
10749 			if (fval > bbr_rtt_probe_limit)
10750 				newval = cur_rttp + (fval - bbr_rtt_probe_limit);
10751 			else
10752 				newval = cur_rttp;
10753 		} else {
10754 			int mul;
10755 
10756 			mul = fval / bbr_rtt_probe_limit;
10757 			newval = cur_rttp * mul;
10758 		}
10759 		if (cur_rttp > 	bbr->r_ctl.rc_probertt_int) {
10760 			bbr->r_ctl.rc_probertt_int = cur_rttp;
10761 			reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10762 			val = 1;
10763 		} else {
10764 			/*
10765 			 * No adjustments were made
10766 			 * do we need to shrink it?
10767 			 */
10768 			if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) {
10769 				if (cur_rttp <= bbr_rtt_probe_limit) {
10770 					/*
10771 					 * Things have calmed down lets
10772 					 * shrink all the way to default
10773 					 */
10774 					bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10775 					reset_time_small(&bbr->r_ctl.rc_rttprop,
10776 							 (bbr_filter_len_sec * USECS_IN_SECOND));
10777 					cur_rttp = bbr_rtt_probe_limit;
10778 					newval = (bbr_filter_len_sec * USECS_IN_SECOND);
10779 					val = 2;
10780 				} else {
10781 					/*
10782 					 * Well does some adjustment make sense?
10783 					 */
10784 					if (cur_rttp < bbr->r_ctl.rc_probertt_int) {
10785 						/* We can reduce interval time some */
10786 						bbr->r_ctl.rc_probertt_int = cur_rttp;
10787 						reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10788 						val = 3;
10789 					}
10790 				}
10791 			}
10792 		}
10793 		if (val)
10794 			bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val);
10795 	}
10796 }
10797 
10798 static void
10799 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
10800 {
10801 	/* Exit probe-rtt */
10802 
10803 	if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) {
10804 		tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10805 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10806 	}
10807 	bbr_log_exit_gain(bbr, cts, 1);
10808 	bbr->rc_hit_state_1 = 0;
10809 	bbr->r_ctl.rc_rtt_shrinks = cts;
10810 	bbr->r_ctl.last_in_probertt = cts;
10811 	bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0);
10812 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10813 	bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp,
10814 					      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
10815 					  bbr->r_ctl.rc_delivered);
10816 	if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10817 		uint32_t time_in;
10818 
10819 		time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10820 		counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10821 	}
10822 	if (bbr->rc_filled_pipe) {
10823 		/* Switch to probe_bw */
10824 		bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
10825 		bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
10826 		bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10827 		bbr_substate_change(bbr, cts, __LINE__, 0);
10828 		bbr_log_type_statechange(bbr, cts, __LINE__);
10829 	} else {
10830 		/* Back to startup */
10831 		bbr->rc_bbr_state = BBR_STATE_STARTUP;
10832 		bbr->r_ctl.rc_bbr_state_time = cts;
10833 		/*
10834 		 * We don't want to give a complete free 3
10835 		 * measurements until we exit, so we use
10836 		 * the number of pe's we were in probe-rtt
10837 		 * to add to the startup_epoch. That way
10838 		 * we will still retain the old state.
10839 		 */
10840 		bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt);
10841 		bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10842 		/* Make sure to use the lower pg when shifting back in */
10843 		if (bbr->r_ctl.rc_lost &&
10844 		    bbr_use_lower_gain_in_startup &&
10845 		    (bbr->rc_use_google == 0))
10846 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
10847 		else
10848 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
10849 		bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
10850 		/* Probably not needed but set it anyway */
10851 		bbr_set_state_target(bbr, __LINE__);
10852 		bbr_log_type_statechange(bbr, cts, __LINE__);
10853 		bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10854 		    bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0);
10855 	}
10856 	bbr_check_probe_rtt_limits(bbr, cts);
10857 }
10858 
10859 static int32_t inline
10860 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts)
10861 {
10862 	if ((bbr->rc_past_init_win == 1) &&
10863 	    (bbr->rc_in_persist == 0) &&
10864 	    (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) {
10865 		return (1);
10866 	}
10867 	if (bbr_can_force_probertt &&
10868 	    (bbr->rc_in_persist == 0) &&
10869 	    (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
10870 	    ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
10871 		return (1);
10872 	}
10873 	return (0);
10874 }
10875 
10876 static int32_t
10877 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t  pkt_epoch)
10878 {
10879 	uint64_t btlbw, gain;
10880 	if (pkt_epoch == 0) {
10881 		/*
10882 		 * Need to be on a pkt-epoch to continue.
10883 		 */
10884 		return (0);
10885 	}
10886 	btlbw = bbr_get_full_bw(bbr);
10887 	gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
10888 		 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
10889 	if (btlbw >= gain) {
10890 		bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
10891 		bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10892 				      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
10893 		bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
10894 	}
10895 	if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)
10896 		return (1);
10897 	bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10898 			      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
10899 	return(0);
10900 }
10901 
10902 static int32_t inline
10903 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch)
10904 {
10905 	/* Have we gained 25% in the last 3 packet based epoch's? */
10906 	uint64_t btlbw, gain;
10907 	int do_exit;
10908 	int delta, rtt_gain;
10909 
10910 	if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
10911 	    (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
10912 		/*
10913 		 * This qualifies as a RTT_PROBE session since we drop the
10914 		 * data outstanding to nothing and waited more than
10915 		 * bbr_rtt_probe_time.
10916 		 */
10917 		bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
10918 		bbr_set_reduced_rtt(bbr, cts, __LINE__);
10919 	}
10920 	if (bbr_should_enter_probe_rtt(bbr, cts)) {
10921 		bbr_enter_probe_rtt(bbr, cts, __LINE__);
10922 		return (0);
10923 	}
10924 	if (bbr->rc_use_google)
10925 		return (bbr_google_startup(bbr, cts,  pkt_epoch));
10926 
10927 	if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
10928 	    (bbr_use_lower_gain_in_startup)) {
10929 		/* Drop to a lower gain 1.5 x since we saw loss */
10930 		bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
10931 	}
10932 	if (pkt_epoch == 0) {
10933 		/*
10934 		 * Need to be on a pkt-epoch to continue.
10935 		 */
10936 		return (0);
10937 	}
10938 	if (bbr_rtt_gain_thresh) {
10939 		/*
10940 		 * Do we allow a flow to stay
10941 		 * in startup with no loss and no
10942 		 * gain in rtt over a set threshold?
10943 		 */
10944 		if (bbr->r_ctl.rc_pkt_epoch_rtt &&
10945 		    bbr->r_ctl.startup_last_srtt &&
10946 		    (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) {
10947 			delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt;
10948 			rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt;
10949 		} else
10950 			rtt_gain = 0;
10951 		if ((bbr->r_ctl.startup_last_srtt == 0)  ||
10952 		    (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt))
10953 			/* First time or new lower value */
10954 			bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
10955 
10956 		if ((bbr->r_ctl.rc_lost == 0) &&
10957 		    (rtt_gain < bbr_rtt_gain_thresh)) {
10958 			/*
10959 			 * No loss, and we are under
10960 			 * our gain threhold for
10961 			 * increasing RTT.
10962 			 */
10963 			if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
10964 				bbr->r_ctl.rc_bbr_last_startup_epoch++;
10965 			bbr_log_startup_event(bbr, cts, rtt_gain,
10966 					      delta, bbr->r_ctl.startup_last_srtt, 10);
10967 			return (0);
10968 		}
10969 	}
10970 	if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) &&
10971 	    (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) &&
10972 	    (!IN_RECOVERY(bbr->rc_tp->t_flags))) {
10973 		/*
10974 		 * We only assess if we have a new measurement when
10975 		 * we have no loss and are not in recovery.
10976 		 * Drag up by one our last_startup epoch so we will hold
10977 		 * the number of non-gain we have already accumulated.
10978 		 */
10979 		if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
10980 			bbr->r_ctl.rc_bbr_last_startup_epoch++;
10981 		bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10982 				      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9);
10983 		return (0);
10984 	}
10985 	/* Case where we reduced the lost (bad retransmit) */
10986 	if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost)
10987 		bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10988 	bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count;
10989 	btlbw = bbr_get_full_bw(bbr);
10990 	if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower)
10991 		gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
10992 			 (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
10993 	else
10994 		gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
10995 			 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
10996 	do_exit = 0;
10997 	if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw)
10998 		bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
10999 	if (btlbw >= gain) {
11000 		bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11001 		/* Update the lost so we won't exit in next set of tests */
11002 		bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11003 		bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11004 				      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11005 	}
11006 	if ((bbr->rc_loss_exit &&
11007 	     (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11008 	     (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) &&
11009 	    ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) {
11010 		/*
11011 		 * If we had no gain,  we had loss and that loss was above
11012 		 * our threshould, the rwnd is not constrained, and we have
11013 		 * had at least 3 packet epochs exit. Note that this is
11014 		 * switched off by sysctl. Google does not do this by the
11015 		 * way.
11016 		 */
11017 		if ((ctf_flight_size(bbr->rc_tp,
11018 			 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11019 		     (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) {
11020 			do_exit = 1;
11021 			bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11022 					      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4);
11023 		} else {
11024 			/* Just record an updated loss value */
11025 			bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11026 			bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11027 					      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5);
11028 		}
11029 	} else
11030 		bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11031 	if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) ||
11032 	    do_exit) {
11033 		/* Return 1 to exit the startup state. */
11034 		return (1);
11035 	}
11036 	/* Stay in startup */
11037 	bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11038 			      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11039 	return (0);
11040 }
11041 
11042 static void
11043 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses)
11044 {
11045 	/*
11046 	 * A tick occurred in the rtt epoch do we need to do anything?
11047 	 */
11048 #ifdef BBR_INVARIANTS
11049 	if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
11050 	    (bbr->rc_bbr_state != BBR_STATE_DRAIN) &&
11051 	    (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) &&
11052 	    (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
11053 	    (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) {
11054 		/* Debug code? */
11055 		panic("Unknown BBR state %d?\n", bbr->rc_bbr_state);
11056 	}
11057 #endif
11058 	if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
11059 		/* Do we exit the startup state? */
11060 		if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) {
11061 			uint32_t time_in;
11062 
11063 			bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11064 					      bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6);
11065 			bbr->rc_filled_pipe = 1;
11066 			bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11067 			if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11068 				time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11069 				counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11070 			} else
11071 				time_in = 0;
11072 			if (bbr->rc_no_pacing)
11073 				bbr->rc_no_pacing = 0;
11074 			bbr->r_ctl.rc_bbr_state_time = cts;
11075 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg;
11076 			bbr->rc_bbr_state = BBR_STATE_DRAIN;
11077 			bbr_set_state_target(bbr, __LINE__);
11078 			if ((bbr->rc_use_google == 0) &&
11079 			    bbr_slam_cwnd_in_main_drain) {
11080 				/* Here we don't have to worry about probe-rtt */
11081 				bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
11082 				bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11083 				bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11084 			}
11085 			bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
11086 			bbr_log_type_statechange(bbr, cts, __LINE__);
11087 			if (ctf_flight_size(bbr->rc_tp,
11088 			        (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
11089 			    bbr->r_ctl.rc_target_at_state) {
11090 				/*
11091 				 * Switch to probe_bw if we are already
11092 				 * there
11093 				 */
11094 				bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11095 				bbr_substate_change(bbr, cts, __LINE__, 0);
11096 				bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11097 				bbr_log_type_statechange(bbr, cts, __LINE__);
11098 			}
11099 		}
11100 	} else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) {
11101 		uint32_t inflight;
11102 		struct tcpcb *tp;
11103 
11104 		tp = bbr->rc_tp;
11105 		inflight = ctf_flight_size(tp,
11106 			      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11107 		if (inflight >= bbr->r_ctl.rc_target_at_state) {
11108 			/* We have reached a flight of the cwnd target */
11109 			bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11110 			bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11111 			bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
11112 			bbr_set_state_target(bbr, __LINE__);
11113 			/*
11114 			 * Rig it so we don't do anything crazy and
11115 			 * start fresh with a new randomization.
11116 			 */
11117 			bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
11118 			bbr->rc_bbr_substate = BBR_SUB_LEVEL6;
11119 			bbr_substate_change(bbr, cts, __LINE__, 1);
11120 		}
11121 	} else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) {
11122 		/* Has in-flight reached the bdp (or less)? */
11123 		uint32_t inflight;
11124 		struct tcpcb *tp;
11125 
11126 		tp = bbr->rc_tp;
11127 		inflight = ctf_flight_size(tp,
11128 			      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11129 		if ((bbr->rc_use_google == 0) &&
11130 		    bbr_slam_cwnd_in_main_drain &&
11131 		    (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11132 			/*
11133 			 * Here we don't have to worry about probe-rtt
11134 			 * re-slam it, but keep it slammed down.
11135 			 */
11136 			bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11137 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11138 		}
11139 		if (inflight <= bbr->r_ctl.rc_target_at_state) {
11140 			/* We have drained */
11141 			bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11142 			bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11143 			if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11144 				uint32_t time_in;
11145 
11146 				time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11147 				counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11148 			}
11149 			if ((bbr->rc_use_google == 0) &&
11150 			    bbr_slam_cwnd_in_main_drain &&
11151 			    (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
11152 				/* Restore the cwnd */
11153 				tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11154 				bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11155 			}
11156 			/* Setup probe-rtt has being done now RRS-HERE */
11157 			bbr->r_ctl.rc_rtt_shrinks = cts;
11158 			bbr->r_ctl.last_in_probertt = cts;
11159 			bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0);
11160 			/* Randomly pick a sub-state */
11161 			bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11162 			bbr_substate_change(bbr, cts, __LINE__, 0);
11163 			bbr_log_type_statechange(bbr, cts, __LINE__);
11164 		}
11165 	} else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) {
11166 		uint32_t flight;
11167 
11168 		flight = ctf_flight_size(bbr->rc_tp,
11169 			     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11170 		bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered);
11171 		if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) &&
11172 		    (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11173 			/*
11174 			 * We must keep cwnd at the desired MSS.
11175 			 */
11176 			bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
11177 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11178 		} else if ((bbr_prtt_slam_cwnd) &&
11179 			   (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11180 			/* Re-slam it */
11181 			bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11182 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11183 		}
11184 		if (bbr->r_ctl.rc_bbr_enters_probertt == 0) {
11185 			/* Has outstanding reached our target? */
11186 			if (flight <= bbr->r_ctl.rc_target_at_state) {
11187 				bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0);
11188 				bbr->r_ctl.rc_bbr_enters_probertt = cts;
11189 				/* If time is exactly 0, be 1usec off */
11190 				if (bbr->r_ctl.rc_bbr_enters_probertt == 0)
11191 					bbr->r_ctl.rc_bbr_enters_probertt = 1;
11192 				if (bbr->rc_use_google == 0) {
11193 					/*
11194 					 * Restore any lowering that as occurred to
11195 					 * reach here
11196 					 */
11197 					if (bbr->r_ctl.bbr_rttprobe_gain_val)
11198 						bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
11199 					else
11200 						bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11201 				}
11202 			}
11203 			if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) &&
11204 			    (bbr->rc_use_google == 0) &&
11205 			    bbr->r_ctl.bbr_rttprobe_gain_val &&
11206 			    (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) ||
11207 			     (flight >= bbr->r_ctl.flightsize_at_drain))) {
11208 				/*
11209 				 * We have doddled with our current hptsi
11210 				 * gain an srtt and have still not made it
11211 				 * to target, or we have increased our flight.
11212 				 * Lets reduce the gain by xx%
11213 				 * flooring the reduce at DRAIN (based on
11214 				 * mul/div)
11215 				 */
11216 				int red;
11217 
11218 				bbr->r_ctl.flightsize_at_drain = flight;
11219 				bbr->r_ctl.rc_probertt_srttchktim = cts;
11220 				red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1);
11221 				if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) {
11222 					/* Reduce our gain again */
11223 					bbr->r_ctl.rc_bbr_hptsi_gain -= red;
11224 					bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0);
11225 				} else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) {
11226 					/* one more chance before we give up */
11227 					bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
11228 					bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0);
11229 				} else {
11230 					/* At the very bottom */
11231 					bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1);
11232 				}
11233 			}
11234 		}
11235 		if (bbr->r_ctl.rc_bbr_enters_probertt &&
11236 		    (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) &&
11237 		    ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) {
11238 			/* Time to exit probe RTT normally */
11239 			bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts);
11240 		}
11241 	} else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
11242 		if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11243 		    (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11244 			/*
11245 			 * This qualifies as a RTT_PROBE session since we
11246 			 * drop the data outstanding to nothing and waited
11247 			 * more than bbr_rtt_probe_time.
11248 			 */
11249 			bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11250 			bbr_set_reduced_rtt(bbr, cts, __LINE__);
11251 		}
11252 		if (bbr_should_enter_probe_rtt(bbr, cts)) {
11253 			bbr_enter_probe_rtt(bbr, cts, __LINE__);
11254 		} else {
11255 			bbr_set_probebw_gains(bbr, cts, losses);
11256 		}
11257 	}
11258 }
11259 
11260 static void
11261 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses)
11262 {
11263 	int32_t epoch = 0;
11264 
11265 	if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
11266 		bbr_set_epoch(bbr, cts, line);
11267 		/* At each epoch doe lt bw sampling */
11268 		epoch = 1;
11269 	}
11270 	bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses);
11271 }
11272 
11273 static int
11274 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
11275     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
11276     int32_t nxt_pkt, struct timeval *tv)
11277 {
11278 	struct inpcb *inp = tptoinpcb(tp);
11279 	int32_t thflags, retval;
11280 	uint32_t cts, lcts;
11281 	uint32_t tiwin;
11282 	struct tcpopt to;
11283 	struct tcp_bbr *bbr;
11284 	struct bbr_sendmap *rsm;
11285 	struct timeval ltv;
11286 	int32_t did_out = 0;
11287 	uint16_t nsegs;
11288 	int32_t prev_state;
11289 	uint32_t lost;
11290 
11291 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
11292 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11293 	/* add in our stats */
11294 	kern_prefetch(bbr, &prev_state);
11295 	prev_state = 0;
11296 	thflags = tcp_get_flags(th);
11297 	/*
11298 	 * If this is either a state-changing packet or current state isn't
11299 	 * established, we require a write lock on tcbinfo.  Otherwise, we
11300 	 * allow the tcbinfo to be in either alocked or unlocked, as the
11301 	 * caller may have unnecessarily acquired a write lock due to a
11302 	 * race.
11303 	 */
11304 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11305 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
11306 	    __func__));
11307 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
11308 	    __func__));
11309 
11310 	tp->t_rcvtime = ticks;
11311 	/*
11312 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
11313 	 * the scale is zero.
11314 	 */
11315 	tiwin = th->th_win << tp->snd_scale;
11316 #ifdef STATS
11317 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
11318 #endif
11319 
11320 	if (m->m_flags & M_TSTMP) {
11321 		/* Prefer the hardware timestamp if present */
11322 		struct timespec ts;
11323 
11324 		mbuf_tstmp2timespec(m, &ts);
11325 		bbr->rc_tv.tv_sec = ts.tv_sec;
11326 		bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11327 		bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11328 	} else if (m->m_flags & M_TSTMP_LRO) {
11329 		/* Next the arrival timestamp */
11330 		struct timespec ts;
11331 
11332 		mbuf_tstmp2timespec(m, &ts);
11333 		bbr->rc_tv.tv_sec = ts.tv_sec;
11334 		bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11335 		bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11336 	} else {
11337 		/*
11338 		 * Ok just get the current time.
11339 		 */
11340 		bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv);
11341 	}
11342 	/*
11343 	 * Parse options on any incoming segment.
11344 	 */
11345 	tcp_dooptions(&to, (u_char *)(th + 1),
11346 	    (th->th_off << 2) - sizeof(struct tcphdr),
11347 	    (thflags & TH_SYN) ? TO_SYN : 0);
11348 
11349 	/*
11350 	 * If timestamps were negotiated during SYN/ACK and a
11351 	 * segment without a timestamp is received, silently drop
11352 	 * the segment, unless it is a RST segment or missing timestamps are
11353 	 * tolerated.
11354 	 * See section 3.2 of RFC 7323.
11355 	 */
11356 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
11357 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
11358 		retval = 0;
11359 		m_freem(m);
11360 		goto done_with_input;
11361 	}
11362 	/*
11363 	 * If echoed timestamp is later than the current time, fall back to
11364 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
11365 	 * were used when this connection was established.
11366 	 */
11367 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
11368 		to.to_tsecr -= tp->ts_offset;
11369 		if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv)))
11370 			to.to_tsecr = 0;
11371 	}
11372 	/*
11373 	 * If its the first time in we need to take care of options and
11374 	 * verify we can do SACK for rack!
11375 	 */
11376 	if (bbr->r_state == 0) {
11377 		/*
11378 		 * Process options only when we get SYN/ACK back. The SYN
11379 		 * case for incoming connections is handled in tcp_syncache.
11380 		 * According to RFC1323 the window field in a SYN (i.e., a
11381 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
11382 		 * this is traditional behavior, may need to be cleaned up.
11383 		 */
11384 		if (bbr->rc_inp == NULL) {
11385 			bbr->rc_inp = inp;
11386 		}
11387 		/*
11388 		 * We need to init rc_inp here since its not init'd when
11389 		 * bbr_init is called
11390 		 */
11391 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
11392 			if ((to.to_flags & TOF_SCALE) &&
11393 			    (tp->t_flags & TF_REQ_SCALE)) {
11394 				tp->t_flags |= TF_RCVD_SCALE;
11395 				tp->snd_scale = to.to_wscale;
11396 			} else
11397 				tp->t_flags &= ~TF_REQ_SCALE;
11398 			/*
11399 			 * Initial send window.  It will be updated with the
11400 			 * next incoming segment to the scaled value.
11401 			 */
11402 			tp->snd_wnd = th->th_win;
11403 			if ((to.to_flags & TOF_TS) &&
11404 			    (tp->t_flags & TF_REQ_TSTMP)) {
11405 				tp->t_flags |= TF_RCVD_TSTMP;
11406 				tp->ts_recent = to.to_tsval;
11407 				tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
11408 			} else
11409 			    tp->t_flags &= ~TF_REQ_TSTMP;
11410 			if (to.to_flags & TOF_MSS)
11411 				tcp_mss(tp, to.to_mss);
11412 			if ((tp->t_flags & TF_SACK_PERMIT) &&
11413 			    (to.to_flags & TOF_SACKPERM) == 0)
11414 				tp->t_flags &= ~TF_SACK_PERMIT;
11415 			if (IS_FASTOPEN(tp->t_flags)) {
11416 				if (to.to_flags & TOF_FASTOPEN) {
11417 					uint16_t mss;
11418 
11419 					if (to.to_flags & TOF_MSS)
11420 						mss = to.to_mss;
11421 					else
11422 						if ((inp->inp_vflag & INP_IPV6) != 0)
11423 							mss = TCP6_MSS;
11424 						else
11425 							mss = TCP_MSS;
11426 					tcp_fastopen_update_cache(tp, mss,
11427 					    to.to_tfo_len, to.to_tfo_cookie);
11428 				} else
11429 					tcp_fastopen_disable_path(tp);
11430 			}
11431 		}
11432 		/*
11433 		 * At this point we are at the initial call. Here we decide
11434 		 * if we are doing RACK or not. We do this by seeing if
11435 		 * TF_SACK_PERMIT is set, if not rack is *not* possible and
11436 		 * we switch to the default code.
11437 		 */
11438 		if ((tp->t_flags & TF_SACK_PERMIT) == 0) {
11439 			/* Bail */
11440 			tcp_switch_back_to_default(tp);
11441 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
11442 			    tlen, iptos);
11443 			return (1);
11444 		}
11445 		/* Set the flag */
11446 		bbr->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
11447 		tcp_set_hpts(inp);
11448 		sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack);
11449 	}
11450 	if (thflags & TH_ACK) {
11451 		/* Track ack types */
11452 		if (to.to_flags & TOF_SACK)
11453 			BBR_STAT_INC(bbr_acks_with_sacks);
11454 		else
11455 			BBR_STAT_INC(bbr_plain_acks);
11456 	}
11457 	/*
11458 	 * This is the one exception case where we set the rack state
11459 	 * always. All other times (timers etc) we must have a rack-state
11460 	 * set (so we assure we have done the checks above for SACK).
11461 	 */
11462 	if (thflags & TH_FIN)
11463 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
11464 	if (bbr->r_state != tp->t_state)
11465 		bbr_set_state(tp, bbr, tiwin);
11466 
11467 	if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL)
11468 		kern_prefetch(rsm, &prev_state);
11469 	prev_state = bbr->r_state;
11470 	bbr->rc_ack_was_delayed = 0;
11471 	lost = bbr->r_ctl.rc_lost;
11472 	bbr->rc_is_pkt_epoch_now = 0;
11473 	if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) {
11474 		/* Get the real time into lcts and figure the real delay */
11475 		lcts = tcp_get_usecs(&ltv);
11476 		if (TSTMP_GT(lcts, cts)) {
11477 			bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts;
11478 			bbr->rc_ack_was_delayed = 1;
11479 			if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay,
11480 				     bbr->r_ctl.highest_hdwr_delay))
11481 				bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay;
11482 		} else {
11483 			bbr->r_ctl.rc_ack_hdwr_delay = 0;
11484 			bbr->rc_ack_was_delayed = 0;
11485 		}
11486 	} else {
11487 		bbr->r_ctl.rc_ack_hdwr_delay = 0;
11488 		bbr->rc_ack_was_delayed = 0;
11489 	}
11490 	bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m);
11491 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
11492 		retval = 0;
11493 		m_freem(m);
11494 		goto done_with_input;
11495 	}
11496 	/*
11497 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
11498 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
11499 	 */
11500 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
11501 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
11502 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11503 		ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11504 		return (1);
11505 	}
11506 	if (tiwin > bbr->r_ctl.rc_high_rwnd)
11507 		bbr->r_ctl.rc_high_rwnd = tiwin;
11508 	bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp,
11509 					    (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11510 	bbr->rtt_valid = 0;
11511 	if (to.to_flags & TOF_TS) {
11512 		bbr->rc_ts_valid = 1;
11513 		bbr->r_ctl.last_inbound_ts = to.to_tsval;
11514 	} else {
11515 		bbr->rc_ts_valid = 0;
11516 		bbr->r_ctl.last_inbound_ts = 0;
11517 	}
11518 	retval = (*bbr->r_substate) (m, th, so,
11519 	    tp, &to, drop_hdrlen,
11520 	    tlen, tiwin, thflags, nxt_pkt, iptos);
11521 	if (nxt_pkt == 0)
11522 		BBR_STAT_INC(bbr_rlock_left_ret0);
11523 	else
11524 		BBR_STAT_INC(bbr_rlock_left_ret1);
11525 	if (retval == 0) {
11526 		/*
11527 		 * If retval is 1 the tcb is unlocked and most likely the tp
11528 		 * is gone.
11529 		 */
11530 		INP_WLOCK_ASSERT(inp);
11531 		tcp_bbr_xmit_timer_commit(bbr, tp, cts);
11532 		if (bbr->rc_is_pkt_epoch_now)
11533 			bbr_set_pktepoch(bbr, cts, __LINE__);
11534 		bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost));
11535 		if (nxt_pkt == 0) {
11536 			if (bbr->r_wanted_output != 0) {
11537 				bbr->rc_output_starts_timer = 0;
11538 				did_out = 1;
11539 				if (tcp_output(tp) < 0)
11540 					return (1);
11541 			} else
11542 				bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0);
11543 		}
11544 		if ((nxt_pkt == 0) &&
11545 		    ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11546 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
11547 		     (tp->t_flags & TF_DELACK) ||
11548 		     ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11549 		      (tp->t_state <= TCPS_CLOSING)))) {
11550 			/*
11551 			 * We could not send (probably in the hpts but
11552 			 * stopped the timer)?
11553 			 */
11554 			if ((tp->snd_max == tp->snd_una) &&
11555 			    ((tp->t_flags & TF_DELACK) == 0) &&
11556 			    (tcp_in_hpts(bbr->rc_inp)) &&
11557 			    (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11558 				/*
11559 				 * keep alive not needed if we are hptsi
11560 				 * output yet
11561 				 */
11562 				;
11563 			} else {
11564 				if (tcp_in_hpts(bbr->rc_inp)) {
11565 					tcp_hpts_remove(bbr->rc_inp);
11566 					if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11567 					    (TSTMP_GT(lcts, bbr->rc_pacer_started))) {
11568 						uint32_t del;
11569 
11570 						del = lcts - bbr->rc_pacer_started;
11571 						if (bbr->r_ctl.rc_last_delay_val > del) {
11572 							BBR_STAT_INC(bbr_force_timer_start);
11573 							bbr->r_ctl.rc_last_delay_val -= del;
11574 							bbr->rc_pacer_started = lcts;
11575 						} else {
11576 							/* We are late */
11577 							bbr->r_ctl.rc_last_delay_val = 0;
11578 							BBR_STAT_INC(bbr_force_output);
11579 							if (tcp_output(tp) < 0)
11580 								return (1);
11581 						}
11582 					}
11583 				}
11584 				bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val,
11585 				    0);
11586 			}
11587 		} else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) {
11588 			/* Do we have the correct timer running? */
11589 			bbr_timer_audit(tp, bbr, lcts, &so->so_snd);
11590 		}
11591 		/* Do we have a new state */
11592 		if (bbr->r_state != tp->t_state)
11593 			bbr_set_state(tp, bbr, tiwin);
11594 done_with_input:
11595 		bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out);
11596 		if (did_out)
11597 			bbr->r_wanted_output = 0;
11598 	}
11599 	return (retval);
11600 }
11601 
11602 static void
11603 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
11604     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11605 {
11606 	struct timeval tv;
11607 	int retval;
11608 
11609 	/* First lets see if we have old packets */
11610 	if (tp->t_in_pkt) {
11611 		if (ctf_do_queued_segments(so, tp, 1)) {
11612 			m_freem(m);
11613 			return;
11614 		}
11615 	}
11616 	if (m->m_flags & M_TSTMP_LRO) {
11617 		mbuf_tstmp2timeval(m, &tv);
11618 	} else {
11619 		/* Should not be should we kassert instead? */
11620 		tcp_get_usecs(&tv);
11621 	}
11622 	retval = bbr_do_segment_nounlock(m, th, so, tp,
11623 					 drop_hdrlen, tlen, iptos, 0, &tv);
11624 	if (retval == 0) {
11625 		INP_WUNLOCK(tptoinpcb(tp));
11626 	}
11627 }
11628 
11629 /*
11630  * Return how much data can be sent without violating the
11631  * cwnd or rwnd.
11632  */
11633 
11634 static inline uint32_t
11635 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin,
11636     uint32_t avail, int32_t sb_offset, uint32_t cts)
11637 {
11638 	uint32_t len;
11639 
11640 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
11641 		/* We never want to go over our peers rcv-window */
11642 		len = 0;
11643 	} else {
11644 		uint32_t flight;
11645 
11646 		flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11647 		if (flight >= sendwin) {
11648 			/*
11649 			 * We have in flight what we are allowed by cwnd (if
11650 			 * it was rwnd blocking it would have hit above out
11651 			 * >= tp->snd_wnd).
11652 			 */
11653 			return (0);
11654 		}
11655 		len = sendwin - flight;
11656 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11657 			/* We would send too much (beyond the rwnd) */
11658 			len = tp->snd_wnd - ctf_outstanding(tp);
11659 		}
11660 		if ((len + sb_offset) > avail) {
11661 			/*
11662 			 * We don't have that much in the SB, how much is
11663 			 * there?
11664 			 */
11665 			len = avail - sb_offset;
11666 		}
11667 	}
11668 	return (len);
11669 }
11670 
11671 static inline void
11672 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11673 {
11674 #ifdef NETFLIX_STATS
11675 	KMOD_TCPSTAT_INC(tcps_sndpack_error);
11676 	KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len);
11677 #endif
11678 }
11679 
11680 static inline void
11681 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11682 {
11683 	if (error) {
11684 		bbr_do_error_accounting(tp, bbr, rsm, len, error);
11685 		return;
11686 	}
11687 	if (rsm) {
11688 		if (rsm->r_flags & BBR_TLP) {
11689 			/*
11690 			 * TLP should not count in retran count, but in its
11691 			 * own bin
11692 			 */
11693 #ifdef NETFLIX_STATS
11694 			KMOD_TCPSTAT_INC(tcps_tlpresends);
11695 			KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len);
11696 #endif
11697 		} else {
11698 			/* Retransmit */
11699 			tp->t_sndrexmitpack++;
11700 			KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
11701 			KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
11702 #ifdef STATS
11703 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
11704 			    len);
11705 #endif
11706 		}
11707 		/*
11708 		 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is
11709 		 * sub-state
11710 		 */
11711 		counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len);
11712 		if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) {
11713 			/* Non probe_bw log in 1, 2, or 4. */
11714 			counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len);
11715 		} else {
11716 			/*
11717 			 * Log our probe state 3, and log also 5-13 to show
11718 			 * us the recovery sub-state for the send. This
11719 			 * means that 3 == (5+6+7+8+9+10+11+12+13)
11720 			 */
11721 			counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len);
11722 			counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len);
11723 		}
11724 		/* Place in both 16's the totals of retransmitted */
11725 		counter_u64_add(bbr_state_lost[16], len);
11726 		counter_u64_add(bbr_state_resend[16], len);
11727 		/* Place in 17's the total sent */
11728 		counter_u64_add(bbr_state_resend[17], len);
11729 		counter_u64_add(bbr_state_lost[17], len);
11730 
11731 	} else {
11732 		/* New sends */
11733 		KMOD_TCPSTAT_INC(tcps_sndpack);
11734 		KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
11735 		/* Place in 17's the total sent */
11736 		counter_u64_add(bbr_state_resend[17], len);
11737 		counter_u64_add(bbr_state_lost[17], len);
11738 #ifdef STATS
11739 		stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
11740 		    len);
11741 #endif
11742 	}
11743 }
11744 
11745 static void
11746 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level)
11747 {
11748 	if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) {
11749 		/*
11750 		 * Limit the cwnd to not be above N x the target plus whats
11751 		 * is outstanding. The target is based on the current b/w
11752 		 * estimate.
11753 		 */
11754 		uint32_t target;
11755 
11756 		target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT);
11757 		target += ctf_outstanding(tp);
11758 		target *= bbr_target_cwnd_mult_limit;
11759 		if (tp->snd_cwnd > target)
11760 			tp->snd_cwnd = target;
11761 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__);
11762 	}
11763 }
11764 
11765 static int
11766 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg)
11767 {
11768 	/*
11769 	 * "adv" is the amount we could increase the window, taking into
11770 	 * account that we are limited by TCP_MAXWIN << tp->rcv_scale.
11771 	 */
11772 	int32_t adv;
11773 	int32_t oldwin;
11774 
11775 	adv = recwin;
11776 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
11777 		oldwin = (tp->rcv_adv - tp->rcv_nxt);
11778 		if (adv > oldwin)
11779 			adv -= oldwin;
11780 		else {
11781 			/* We can't increase the window */
11782 			adv = 0;
11783 		}
11784 	} else
11785 		oldwin = 0;
11786 
11787 	/*
11788 	 * If the new window size ends up being the same as or less
11789 	 * than the old size when it is scaled, then don't force
11790 	 * a window update.
11791 	 */
11792 	if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
11793 		return (0);
11794 
11795 	if (adv >= (2 * maxseg) &&
11796 	    (adv >= (so->so_rcv.sb_hiwat / 4) ||
11797 	    recwin <= (so->so_rcv.sb_hiwat / 8) ||
11798 	    so->so_rcv.sb_hiwat <= 8 * maxseg)) {
11799 		return (1);
11800 	}
11801 	if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat)
11802 		return (1);
11803 	return (0);
11804 }
11805 
11806 /*
11807  * Return 0 on success and a errno on failure to send.
11808  * Note that a 0 return may not mean we sent anything
11809  * if the TCB was on the hpts. A non-zero return
11810  * does indicate the error we got from ip[6]_output.
11811  */
11812 static int
11813 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv)
11814 {
11815 	struct socket *so;
11816 	int32_t len;
11817 	uint32_t cts;
11818 	uint32_t recwin, sendwin;
11819 	int32_t sb_offset;
11820 	int32_t flags, abandon, error = 0;
11821 	struct tcp_log_buffer *lgb = NULL;
11822 	struct mbuf *m;
11823 	struct mbuf *mb;
11824 	uint32_t if_hw_tsomaxsegcount = 0;
11825 	uint32_t if_hw_tsomaxsegsize = 0;
11826 	uint32_t if_hw_tsomax = 0;
11827 	struct ip *ip = NULL;
11828 	struct tcp_bbr *bbr;
11829 	struct tcphdr *th;
11830 	struct udphdr *udp = NULL;
11831 	u_char opt[TCP_MAXOLEN];
11832 	unsigned ipoptlen, optlen, hdrlen;
11833 	unsigned ulen;
11834 	uint32_t bbr_seq;
11835 	uint32_t delay_calc=0;
11836 	uint8_t doing_tlp = 0;
11837 	uint8_t local_options;
11838 #ifdef BBR_INVARIANTS
11839 	uint8_t doing_retran_from = 0;
11840 	uint8_t picked_up_retran = 0;
11841 #endif
11842 	uint8_t wanted_cookie = 0;
11843 	uint8_t more_to_rxt=0;
11844 	int32_t prefetch_so_done = 0;
11845 	int32_t prefetch_rsm = 0;
11846 	uint32_t tot_len = 0;
11847 	uint32_t maxseg, pace_max_segs, p_maxseg;
11848 	int32_t csum_flags = 0;
11849  	int32_t hw_tls;
11850 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
11851 	unsigned ipsec_optlen = 0;
11852 
11853 #endif
11854 	volatile int32_t sack_rxmit;
11855 	struct bbr_sendmap *rsm = NULL;
11856 	int32_t tso, mtu;
11857 	struct tcpopt to;
11858 	int32_t slot = 0;
11859 	struct inpcb *inp;
11860 	struct sockbuf *sb;
11861 	uint32_t hpts_calling;
11862 #ifdef INET6
11863 	struct ip6_hdr *ip6 = NULL;
11864 	int32_t isipv6;
11865 #endif
11866 	uint8_t app_limited = BBR_JR_SENT_DATA;
11867 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11868 	/* We take a cache hit here */
11869 	memcpy(&bbr->rc_tv, tv, sizeof(struct timeval));
11870 	cts = tcp_tv_to_usectick(&bbr->rc_tv);
11871 	inp = bbr->rc_inp;
11872 	so = inp->inp_socket;
11873 	sb = &so->so_snd;
11874 	if (tp->t_nic_ktls_xmit)
11875  		hw_tls = 1;
11876  	else
11877  		hw_tls = 0;
11878 	kern_prefetch(sb, &maxseg);
11879 	maxseg = tp->t_maxseg - bbr->rc_last_options;
11880 	if (bbr_minseg(bbr) < maxseg) {
11881 		tcp_bbr_tso_size_check(bbr, cts);
11882 	}
11883 	/* Remove any flags that indicate we are pacing on the inp  */
11884 	pace_max_segs = bbr->r_ctl.rc_pace_max_segs;
11885 	p_maxseg = min(maxseg, pace_max_segs);
11886 	INP_WLOCK_ASSERT(inp);
11887 #ifdef TCP_OFFLOAD
11888 	if (tp->t_flags & TF_TOE)
11889 		return (tcp_offload_output(tp));
11890 #endif
11891 
11892 #ifdef INET6
11893 	if (bbr->r_state) {
11894 		/* Use the cache line loaded if possible */
11895 		isipv6 = bbr->r_is_v6;
11896 	} else {
11897 		isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
11898 	}
11899 #endif
11900 	if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
11901 	    tcp_in_hpts(inp)) {
11902 		/*
11903 		 * We are on the hpts for some timer but not hptsi output.
11904 		 * Possibly remove from the hpts so we can send/recv etc.
11905 		 */
11906 		if ((tp->t_flags & TF_ACKNOW) == 0) {
11907 			/*
11908 			 * No immediate demand right now to send an ack, but
11909 			 * the user may have read, making room for new data
11910 			 * (a window update). If so we may want to cancel
11911 			 * whatever timer is running (KEEP/DEL-ACK?) and
11912 			 * continue to send out a window update. Or we may
11913 			 * have gotten more data into the socket buffer to
11914 			 * send.
11915 			 */
11916 			recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
11917 				      (long)TCP_MAXWIN << tp->rcv_scale);
11918 			if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) &&
11919 			    ((tcp_outflags[tp->t_state] & TH_RST) == 0) &&
11920 			    ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <=
11921 			    (tp->snd_max - tp->snd_una))) {
11922 				/*
11923 				 * Nothing new to send and no window update
11924 				 * is needed to send. Lets just return and
11925 				 * let the timer-run off.
11926 				 */
11927 				return (0);
11928 			}
11929 		}
11930 		tcp_hpts_remove(inp);
11931 		bbr_timer_cancel(bbr, __LINE__, cts);
11932 	}
11933 	if (bbr->r_ctl.rc_last_delay_val) {
11934 		/* Calculate a rough delay for early escape to sending  */
11935 		if (SEQ_GT(cts, bbr->rc_pacer_started))
11936 			delay_calc = cts - bbr->rc_pacer_started;
11937 		if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
11938 			delay_calc -= bbr->r_ctl.rc_last_delay_val;
11939 		else
11940 			delay_calc = 0;
11941 	}
11942 	/* Mark that we have called bbr_output(). */
11943 	if ((bbr->r_timer_override) ||
11944 	    (tp->t_state < TCPS_ESTABLISHED)) {
11945 		/* Timeouts or early states are exempt */
11946 		if (tcp_in_hpts(inp))
11947 			tcp_hpts_remove(inp);
11948 	} else if (tcp_in_hpts(inp)) {
11949 		if ((bbr->r_ctl.rc_last_delay_val) &&
11950 		    (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11951 		    delay_calc) {
11952 			/*
11953 			 * We were being paced for output and the delay has
11954 			 * already exceeded when we were supposed to be
11955 			 * called, lets go ahead and pull out of the hpts
11956 			 * and call output.
11957 			 */
11958 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1);
11959 			bbr->r_ctl.rc_last_delay_val = 0;
11960 			tcp_hpts_remove(inp);
11961 		} else if (tp->t_state == TCPS_CLOSED) {
11962 			bbr->r_ctl.rc_last_delay_val = 0;
11963 			tcp_hpts_remove(inp);
11964 		} else {
11965 			/*
11966 			 * On the hpts, you shall not pass! even if ACKNOW
11967 			 * is on, we will when the hpts fires, unless of
11968 			 * course we are overdue.
11969 			 */
11970 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1);
11971 			return (0);
11972 		}
11973 	}
11974 	bbr->rc_cwnd_limited = 0;
11975 	if (bbr->r_ctl.rc_last_delay_val) {
11976 		/* recalculate the real delay and deal with over/under  */
11977 		if (SEQ_GT(cts, bbr->rc_pacer_started))
11978 			delay_calc = cts - bbr->rc_pacer_started;
11979 		else
11980 			delay_calc = 0;
11981 		if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
11982 			/* Setup the delay which will be added in */
11983 			delay_calc -= bbr->r_ctl.rc_last_delay_val;
11984 		else {
11985 			/*
11986 			 * We are early setup to adjust
11987 			 * our slot time.
11988 			 */
11989 			uint64_t merged_val;
11990 
11991 			bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc);
11992 			bbr->r_agg_early_set = 1;
11993 			if (bbr->r_ctl.rc_hptsi_agg_delay) {
11994 				if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) {
11995 					/* Nope our previous late cancels out the early */
11996 					bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early;
11997 					bbr->r_agg_early_set = 0;
11998 					bbr->r_ctl.rc_agg_early = 0;
11999 				} else {
12000 					bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay;
12001 					bbr->r_ctl.rc_hptsi_agg_delay = 0;
12002 				}
12003 			}
12004 			merged_val = bbr->rc_pacer_started;
12005 			merged_val <<= 32;
12006 			merged_val |= bbr->r_ctl.rc_last_delay_val;
12007 			bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls,
12008 						 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val,
12009 						 bbr->r_agg_early_set, 3);
12010 			bbr->r_ctl.rc_last_delay_val = 0;
12011 			BBR_STAT_INC(bbr_early);
12012 			delay_calc = 0;
12013 		}
12014 	} else {
12015 		/* We were not delayed due to hptsi */
12016 		if (bbr->r_agg_early_set)
12017 			bbr->r_ctl.rc_agg_early = 0;
12018 		bbr->r_agg_early_set = 0;
12019 		delay_calc = 0;
12020 	}
12021 	if (delay_calc) {
12022 		/*
12023 		 * We had a hptsi delay which means we are falling behind on
12024 		 * sending at the expected rate. Calculate an extra amount
12025 		 * of data we can send, if any, to put us back on track.
12026 		 */
12027 		if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay)
12028 			bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff;
12029 		else
12030 			bbr->r_ctl.rc_hptsi_agg_delay += delay_calc;
12031 	}
12032 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12033 	if ((tp->snd_una == tp->snd_max) &&
12034 	    (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
12035 	    (sbavail(sb))) {
12036 		/*
12037 		 * Ok we have been idle with nothing outstanding
12038 		 * we possibly need to start fresh with either a new
12039 		 * suite of states or a fast-ramp up.
12040 		 */
12041 		bbr_restart_after_idle(bbr,
12042 				       cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time));
12043 	}
12044 	/*
12045 	 * Now was there a hptsi delay where we are behind? We only count
12046 	 * being behind if: a) We are not in recovery. b) There was a delay.
12047 	 * <and> c) We had room to send something.
12048 	 *
12049 	 */
12050 	hpts_calling = inp->inp_hpts_calls;
12051 	inp->inp_hpts_calls = 0;
12052 	if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
12053 		int retval;
12054 
12055 		retval = bbr_process_timers(tp, bbr, cts, hpts_calling);
12056 		if (retval != 0) {
12057 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1);
12058 			/*
12059 			 * If timers want tcp_drop(), then pass error out,
12060 			 * otherwise suppress it.
12061 			 */
12062 			return (retval < 0 ? retval : 0);
12063 		}
12064 	}
12065 	bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12066 	if (hpts_calling &&
12067 	    (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12068 		bbr->r_ctl.rc_last_delay_val = 0;
12069 	}
12070 	bbr->r_timer_override = 0;
12071 	bbr->r_wanted_output = 0;
12072 	/*
12073 	 * For TFO connections in SYN_RECEIVED, only allow the initial
12074 	 * SYN|ACK and those sent by the retransmit timer.
12075 	 */
12076 	if (IS_FASTOPEN(tp->t_flags) &&
12077 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
12078 	     (tp->t_state == TCPS_SYN_SENT)) &&
12079 	    SEQ_GT(tp->snd_max, tp->snd_una) &&	/* initial SYN or SYN|ACK sent */
12080 	    (tp->t_rxtshift == 0)) {	/* not a retransmit */
12081 		len = 0;
12082 		goto just_return_nolock;
12083 	}
12084 	/*
12085 	 * Before sending anything check for a state update. For hpts
12086 	 * calling without input this is important. If its input calling
12087 	 * then this was already done.
12088 	 */
12089 	if (bbr->rc_use_google == 0)
12090 		bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12091 again:
12092 	/*
12093 	 * If we've recently taken a timeout, snd_max will be greater than
12094 	 * snd_max. BBR in general does not pay much attention to snd_nxt
12095 	 * for historic reasons the persist timer still uses it. This means
12096 	 * we have to look at it. All retransmissions that are not persits
12097 	 * use the rsm that needs to be sent so snd_nxt is ignored. At the
12098 	 * end of this routine we pull snd_nxt always up to snd_max.
12099 	 */
12100 	doing_tlp = 0;
12101 #ifdef BBR_INVARIANTS
12102 	doing_retran_from = picked_up_retran = 0;
12103 #endif
12104 	error = 0;
12105 	tso = 0;
12106 	slot = 0;
12107 	mtu = 0;
12108 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12109 	sb_offset = tp->snd_max - tp->snd_una;
12110 	flags = tcp_outflags[tp->t_state];
12111 	sack_rxmit = 0;
12112 	len = 0;
12113 	rsm = NULL;
12114 	if (flags & TH_RST) {
12115 		SOCKBUF_LOCK(sb);
12116 		goto send;
12117 	}
12118 recheck_resend:
12119 	while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
12120 		/* We need to always have one in reserve */
12121 		rsm = bbr_alloc(bbr);
12122 		if (rsm == NULL) {
12123 			error = ENOMEM;
12124 			/* Lie to get on the hpts */
12125 			tot_len = tp->t_maxseg;
12126 			if (hpts_calling)
12127 				/* Retry in a ms */
12128 				slot = 1001;
12129 			goto just_return_nolock;
12130 		}
12131 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
12132 		bbr->r_ctl.rc_free_cnt++;
12133 		rsm = NULL;
12134 	}
12135 	/* What do we send, a resend? */
12136 	if (bbr->r_ctl.rc_resend == NULL) {
12137 		/* Check for rack timeout */
12138 		bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
12139 		if (bbr->r_ctl.rc_resend) {
12140 #ifdef BBR_INVARIANTS
12141 			picked_up_retran = 1;
12142 #endif
12143 			bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend);
12144 		}
12145 	}
12146 	if (bbr->r_ctl.rc_resend) {
12147 		rsm = bbr->r_ctl.rc_resend;
12148 #ifdef BBR_INVARIANTS
12149 		doing_retran_from = 1;
12150 #endif
12151 		/* Remove any TLP flags its a RACK or T-O */
12152 		rsm->r_flags &= ~BBR_TLP;
12153 		bbr->r_ctl.rc_resend = NULL;
12154 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12155 #ifdef BBR_INVARIANTS
12156 			panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n",
12157 			    tp, bbr, rsm, rsm->r_start, tp->snd_una);
12158 			goto recheck_resend;
12159 #else
12160 			/* TSNH */
12161 			rsm = NULL;
12162 			goto recheck_resend;
12163 #endif
12164 		}
12165 		if (rsm->r_flags & BBR_HAS_SYN) {
12166 			/* Only retransmit a SYN by itself */
12167 			len = 0;
12168 			if ((flags & TH_SYN) == 0) {
12169 				/* Huh something is wrong */
12170 				rsm->r_start++;
12171 				if (rsm->r_start == rsm->r_end) {
12172 					/* Clean it up, somehow we missed the ack? */
12173 					bbr_log_syn(tp, NULL);
12174 				} else {
12175 					/* TFO with data? */
12176 					rsm->r_flags &= ~BBR_HAS_SYN;
12177 					len = rsm->r_end - rsm->r_start;
12178 				}
12179 			} else {
12180 				/* Retransmitting SYN */
12181 				rsm = NULL;
12182 				SOCKBUF_LOCK(sb);
12183 				goto send;
12184 			}
12185 		} else
12186 			len = rsm->r_end - rsm->r_start;
12187 		if ((bbr->rc_resends_use_tso == 0) &&
12188 		    (len > maxseg)) {
12189 			len = maxseg;
12190 			more_to_rxt = 1;
12191 		}
12192 		sb_offset = rsm->r_start - tp->snd_una;
12193 		if (len > 0) {
12194 			sack_rxmit = 1;
12195 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
12196 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12197 			    min(len, maxseg));
12198 		} else {
12199 			/* I dont think this can happen */
12200 			rsm = NULL;
12201 			goto recheck_resend;
12202 		}
12203 		BBR_STAT_INC(bbr_resends_set);
12204 	} else if (bbr->r_ctl.rc_tlp_send) {
12205 		/*
12206 		 * Tail loss probe
12207 		 */
12208 		doing_tlp = 1;
12209 		rsm = bbr->r_ctl.rc_tlp_send;
12210 		bbr->r_ctl.rc_tlp_send = NULL;
12211 		sack_rxmit = 1;
12212 		len = rsm->r_end - rsm->r_start;
12213 		if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12214 			len = maxseg;
12215 
12216 		if (SEQ_GT(tp->snd_una, rsm->r_start)) {
12217 #ifdef BBR_INVARIANTS
12218 			panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u",
12219 			    tp, bbr, tp->snd_una, rsm, rsm->r_start);
12220 #else
12221 			/* TSNH */
12222 			rsm = NULL;
12223 			goto recheck_resend;
12224 #endif
12225 		}
12226 		sb_offset = rsm->r_start - tp->snd_una;
12227 		BBR_STAT_INC(bbr_tlp_set);
12228 	}
12229 	/*
12230 	 * Enforce a connection sendmap count limit if set
12231 	 * as long as we are not retransmiting.
12232 	 */
12233 	if ((rsm == NULL) &&
12234 	    (V_tcp_map_entries_limit > 0) &&
12235 	    (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
12236 		BBR_STAT_INC(bbr_alloc_limited);
12237 		if (!bbr->alloc_limit_reported) {
12238 			bbr->alloc_limit_reported = 1;
12239 			BBR_STAT_INC(bbr_alloc_limited_conns);
12240 		}
12241 		goto just_return_nolock;
12242 	}
12243 #ifdef BBR_INVARIANTS
12244 	if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) {
12245 		panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u",
12246 		    tp, bbr, rsm, sb_offset, len);
12247 	}
12248 #endif
12249 	/*
12250 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
12251 	 * state flags.
12252 	 */
12253 	if (tp->t_flags & TF_NEEDFIN && (rsm == NULL))
12254 		flags |= TH_FIN;
12255 	if (tp->t_flags & TF_NEEDSYN)
12256 		flags |= TH_SYN;
12257 
12258 	if (rsm && (rsm->r_flags & BBR_HAS_FIN)) {
12259 		/* we are retransmitting the fin */
12260 		len--;
12261 		if (len) {
12262 			/*
12263 			 * When retransmitting data do *not* include the
12264 			 * FIN. This could happen from a TLP probe if we
12265 			 * allowed data with a FIN.
12266 			 */
12267 			flags &= ~TH_FIN;
12268 		}
12269 	} else if (rsm) {
12270 		if (flags & TH_FIN)
12271 			flags &= ~TH_FIN;
12272 	}
12273 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12274 		void *end_rsm;
12275 
12276 		end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
12277 		if (end_rsm)
12278 			kern_prefetch(end_rsm, &prefetch_rsm);
12279 		prefetch_rsm = 1;
12280 	}
12281 	SOCKBUF_LOCK(sb);
12282 	/*
12283 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
12284 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12285 	 * negative length.  This can also occur when TCP opens up its
12286 	 * congestion window while receiving additional duplicate acks after
12287 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
12288 	 * the fast-retransmit.
12289 	 *
12290 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
12291 	 * set to snd_una, the sb_offset will be 0, and the length may wind
12292 	 * up 0.
12293 	 *
12294 	 * If sack_rxmit is true we are retransmitting from the scoreboard
12295 	 * in which case len is already set.
12296 	 */
12297 	if (sack_rxmit == 0) {
12298 		uint32_t avail;
12299 
12300 		avail = sbavail(sb);
12301 		if (SEQ_GT(tp->snd_max, tp->snd_una))
12302 			sb_offset = tp->snd_max - tp->snd_una;
12303 		else
12304 			sb_offset = 0;
12305 		if (bbr->rc_tlp_new_data) {
12306 			/* TLP is forcing out new data */
12307 			uint32_t tlplen;
12308 
12309 			doing_tlp = 1;
12310 			tlplen = maxseg;
12311 
12312 			if (tlplen > (uint32_t)(avail - sb_offset)) {
12313 				tlplen = (uint32_t)(avail - sb_offset);
12314 			}
12315 			if (tlplen > tp->snd_wnd) {
12316 				len = tp->snd_wnd;
12317 			} else {
12318 				len = tlplen;
12319 			}
12320 			bbr->rc_tlp_new_data = 0;
12321 		} else {
12322 			len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts);
12323 			if ((len < p_maxseg) &&
12324 			    (bbr->rc_in_persist == 0) &&
12325 			    (ctf_outstanding(tp) >= (2 * p_maxseg)) &&
12326 			    ((avail - sb_offset) >= p_maxseg)) {
12327 				/*
12328 				 * We are not completing whats in the socket
12329 				 * buffer (i.e. there is at least a segment
12330 				 * waiting to send) and we have 2 or more
12331 				 * segments outstanding. There is no sense
12332 				 * of sending a little piece. Lets defer and
12333 				 * and wait until we can send a whole
12334 				 * segment.
12335 				 */
12336 				len = 0;
12337 			}
12338 			if (bbr->rc_in_persist) {
12339 				/*
12340 				 * We are in persists, figure out if
12341 				 * a retransmit is available (maybe the previous
12342 				 * persists we sent) or if we have to send new
12343 				 * data.
12344 				 */
12345 				rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
12346 				if (rsm) {
12347 					len = rsm->r_end - rsm->r_start;
12348 					if (rsm->r_flags & BBR_HAS_FIN)
12349 						len--;
12350 					if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12351 						len = maxseg;
12352 					if (len > 1)
12353 						BBR_STAT_INC(bbr_persist_reneg);
12354 					/*
12355 					 * XXXrrs we could force the len to
12356 					 * 1 byte here to cause the chunk to
12357 					 * split apart.. but that would then
12358 					 * mean we always retransmit it as
12359 					 * one byte even after the window
12360 					 * opens.
12361 					 */
12362 					sack_rxmit = 1;
12363 					sb_offset = rsm->r_start - tp->snd_una;
12364 				} else {
12365 					/*
12366 					 * First time through in persists or peer
12367 					 * acked our one byte. Though we do have
12368 					 * to have something in the sb.
12369 					 */
12370 					len = 1;
12371 					sb_offset = 0;
12372 					if (avail == 0)
12373 					    len = 0;
12374 				}
12375 			}
12376 		}
12377 	}
12378 	if (prefetch_so_done == 0) {
12379 		kern_prefetch(so, &prefetch_so_done);
12380 		prefetch_so_done = 1;
12381 	}
12382 	/*
12383 	 * Lop off SYN bit if it has already been sent.  However, if this is
12384 	 * SYN-SENT state and if segment contains data and if we don't know
12385 	 * that foreign host supports TAO, suppress sending segment.
12386 	 */
12387 	if ((flags & TH_SYN) && (rsm == NULL) &&
12388 	    SEQ_GT(tp->snd_max, tp->snd_una)) {
12389 		if (tp->t_state != TCPS_SYN_RECEIVED)
12390 			flags &= ~TH_SYN;
12391 		/*
12392 		 * When sending additional segments following a TFO SYN|ACK,
12393 		 * do not include the SYN bit.
12394 		 */
12395 		if (IS_FASTOPEN(tp->t_flags) &&
12396 		    (tp->t_state == TCPS_SYN_RECEIVED))
12397 			flags &= ~TH_SYN;
12398 		sb_offset--, len++;
12399 		if (sbavail(sb) == 0)
12400 			len = 0;
12401 	} else if ((flags & TH_SYN) && rsm) {
12402 		/*
12403 		 * Subtract one from the len for the SYN being
12404 		 * retransmitted.
12405 		 */
12406 		len--;
12407 	}
12408 	/*
12409 	 * Be careful not to send data and/or FIN on SYN segments. This
12410 	 * measure is needed to prevent interoperability problems with not
12411 	 * fully conformant TCP implementations.
12412 	 */
12413 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12414 		len = 0;
12415 		flags &= ~TH_FIN;
12416 	}
12417 	/*
12418 	 * On TFO sockets, ensure no data is sent in the following cases:
12419 	 *
12420 	 *  - When retransmitting SYN|ACK on a passively-created socket
12421 	 *  - When retransmitting SYN on an actively created socket
12422 	 *  - When sending a zero-length cookie (cookie request) on an
12423 	 *    actively created socket
12424 	 *  - When the socket is in the CLOSED state (RST is being sent)
12425 	 */
12426 	if (IS_FASTOPEN(tp->t_flags) &&
12427 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12428 	     ((tp->t_state == TCPS_SYN_SENT) &&
12429 	      (tp->t_tfo_client_cookie_len == 0)) ||
12430 	     (flags & TH_RST))) {
12431 		len = 0;
12432 		sack_rxmit = 0;
12433 		rsm = NULL;
12434 	}
12435 	/* Without fast-open there should never be data sent on a SYN */
12436 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags)))
12437 		len = 0;
12438 	if (len <= 0) {
12439 		/*
12440 		 * If FIN has been sent but not acked, but we haven't been
12441 		 * called to retransmit, len will be < 0.  Otherwise, window
12442 		 * shrank after we sent into it.  If window shrank to 0,
12443 		 * cancel pending retransmit, pull snd_nxt back to (closed)
12444 		 * window, and set the persist timer if it isn't already
12445 		 * going.  If the window didn't close completely, just wait
12446 		 * for an ACK.
12447 		 *
12448 		 * We also do a general check here to ensure that we will
12449 		 * set the persist timer when we have data to send, but a
12450 		 * 0-byte window. This makes sure the persist timer is set
12451 		 * even if the packet hits one of the "goto send" lines
12452 		 * below.
12453 		 */
12454 		len = 0;
12455 		if ((tp->snd_wnd == 0) &&
12456 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12457 		    (tp->snd_una == tp->snd_max) &&
12458 		    (sb_offset < (int)sbavail(sb))) {
12459 			/*
12460 			 * Not enough room in the rwnd to send
12461 			 * a paced segment out.
12462 			 */
12463 			bbr_enter_persist(tp, bbr, cts, __LINE__);
12464 		}
12465 	} else if ((rsm == NULL) &&
12466 		   (doing_tlp == 0) &&
12467 		   (len < bbr->r_ctl.rc_pace_max_segs)) {
12468 		/*
12469 		 * We are not sending a full segment for
12470 		 * some reason. Should we not send anything (think
12471 		 * sws or persists)?
12472 		 */
12473 		if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12474 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12475 		    (len < (int)(sbavail(sb) - sb_offset))) {
12476 			/*
12477 			 * Here the rwnd is less than
12478 			 * the pacing size, this is not a retransmit,
12479 			 * we are established and
12480 			 * the send is not the last in the socket buffer
12481 			 * lets not send, and possibly enter persists.
12482 			 */
12483 			len = 0;
12484 			if (tp->snd_max == tp->snd_una)
12485 				bbr_enter_persist(tp, bbr, cts, __LINE__);
12486 		} else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) &&
12487 			   (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12488 						 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12489 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12490 			   (len < bbr_minseg(bbr))) {
12491 			/*
12492 			 * Here we are not retransmitting, and
12493 			 * the cwnd is not so small that we could
12494 			 * not send at least a min size (rxt timer
12495 			 * not having gone off), We have 2 segments or
12496 			 * more already in flight, its not the tail end
12497 			 * of the socket buffer  and the cwnd is blocking
12498 			 * us from sending out minimum pacing segment size.
12499 			 * Lets not send anything.
12500 			 */
12501 			bbr->rc_cwnd_limited = 1;
12502 			len = 0;
12503 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12504 			    min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12505 			   (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12506 						 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12507 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12508 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
12509 			/*
12510 			 * Here we have a send window but we have
12511 			 * filled it up and we can't send another pacing segment.
12512 			 * We also have in flight more than 2 segments
12513 			 * and we are not completing the sb i.e. we allow
12514 			 * the last bytes of the sb to go out even if
12515 			 * its not a full pacing segment.
12516 			 */
12517 			len = 0;
12518 		}
12519 	}
12520 	/* len will be >= 0 after this point. */
12521 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12522 	tcp_sndbuf_autoscale(tp, so, sendwin);
12523 	/*
12524 	 *
12525 	 */
12526 	if (bbr->rc_in_persist &&
12527 	    len &&
12528 	    (rsm == NULL) &&
12529 	    (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) {
12530 		/*
12531 		 * We are in persist, not doing a retransmit and don't have enough space
12532 		 * yet to send a full TSO. So is it at the end of the sb
12533 		 * if so we need to send else nuke to 0 and don't send.
12534 		 */
12535 		int sbleft;
12536 		if (sbavail(sb) > sb_offset)
12537 			sbleft = sbavail(sb) - sb_offset;
12538 		else
12539 			sbleft = 0;
12540 		if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) {
12541 			/* not at end of sb lets not send */
12542 			len = 0;
12543 		}
12544 	}
12545 	/*
12546 	 * Decide if we can use TCP Segmentation Offloading (if supported by
12547 	 * hardware).
12548 	 *
12549 	 * TSO may only be used if we are in a pure bulk sending state.  The
12550 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12551 	 * options prevent using TSO.  With TSO the TCP header is the same
12552 	 * (except for the sequence number) for all generated packets.  This
12553 	 * makes it impossible to transmit any options which vary per
12554 	 * generated segment or packet.
12555 	 *
12556 	 * IPv4 handling has a clear separation of ip options and ip header
12557 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen()
12558 	 * does the right thing below to provide length of just ip options
12559 	 * and thus checking for ipoptlen is enough to decide if ip options
12560 	 * are present.
12561 	 */
12562 #ifdef INET6
12563 	if (isipv6)
12564 		ipoptlen = ip6_optlen(inp);
12565 	else
12566 #endif
12567 	if (inp->inp_options)
12568 		ipoptlen = inp->inp_options->m_len -
12569 		    offsetof(struct ipoption, ipopt_list);
12570 	else
12571 		ipoptlen = 0;
12572 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12573 	/*
12574 	 * Pre-calculate here as we save another lookup into the darknesses
12575 	 * of IPsec that way and can actually decide if TSO is ok.
12576 	 */
12577 #ifdef INET6
12578 	if (isipv6 && IPSEC_ENABLED(ipv6))
12579 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
12580 #ifdef INET
12581 	else
12582 #endif
12583 #endif				/* INET6 */
12584 #ifdef INET
12585 	if (IPSEC_ENABLED(ipv4))
12586 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
12587 #endif				/* INET */
12588 #endif				/* IPSEC */
12589 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12590 	ipoptlen += ipsec_optlen;
12591 #endif
12592 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
12593 	    (len > maxseg) &&
12594 	    (tp->t_port == 0) &&
12595 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
12596 	    tp->rcv_numsacks == 0 &&
12597 	    ipoptlen == 0)
12598 		tso = 1;
12599 
12600 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12601 	    (long)TCP_MAXWIN << tp->rcv_scale);
12602 	/*
12603 	 * Sender silly window avoidance.   We transmit under the following
12604 	 * conditions when len is non-zero:
12605 	 *
12606 	 * - We have a full segment (or more with TSO) - This is the last
12607 	 * buffer in a write()/send() and we are either idle or running
12608 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
12609 	 * then 1/2 the maximum send window's worth of data (receiver may be
12610 	 * limited the window size) - we need to retransmit
12611 	 */
12612 	if (rsm)
12613 		goto send;
12614 	if (len) {
12615 		if (sack_rxmit)
12616 			goto send;
12617 		if (len >= p_maxseg)
12618 			goto send;
12619 		/*
12620 		 * NOTE! on localhost connections an 'ack' from the remote
12621 		 * end may occur synchronously with the output and cause us
12622 		 * to flush a buffer queued with moretocome.  XXX
12623 		 *
12624 		 */
12625 		if (((tp->t_flags & TF_MORETOCOME) == 0) &&	/* normal case */
12626 		    ((tp->t_flags & TF_NODELAY) ||
12627 		    ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) &&
12628 		    (tp->t_flags & TF_NOPUSH) == 0) {
12629 			goto send;
12630 		}
12631 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
12632 			goto send;
12633 		}
12634 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
12635 			goto send;
12636 		}
12637 	}
12638 	/*
12639 	 * Sending of standalone window updates.
12640 	 *
12641 	 * Window updates are important when we close our window due to a
12642 	 * full socket buffer and are opening it again after the application
12643 	 * reads data from it.  Once the window has opened again and the
12644 	 * remote end starts to send again the ACK clock takes over and
12645 	 * provides the most current window information.
12646 	 *
12647 	 * We must avoid the silly window syndrome whereas every read from
12648 	 * the receive buffer, no matter how small, causes a window update
12649 	 * to be sent.  We also should avoid sending a flurry of window
12650 	 * updates when the socket buffer had queued a lot of data and the
12651 	 * application is doing small reads.
12652 	 *
12653 	 * Prevent a flurry of pointless window updates by only sending an
12654 	 * update when we can increase the advertized window by more than
12655 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
12656 	 * full or is very small be more aggressive and send an update
12657 	 * whenever we can increase by two mss sized segments. In all other
12658 	 * situations the ACK's to new incoming data will carry further
12659 	 * window increases.
12660 	 *
12661 	 * Don't send an independent window update if a delayed ACK is
12662 	 * pending (it will get piggy-backed on it) or the remote side
12663 	 * already has done a half-close and won't send more data.  Skip
12664 	 * this if the connection is in T/TCP half-open state.
12665 	 */
12666 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
12667 	    !(tp->t_flags & TF_DELACK) &&
12668 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
12669 		/* Check to see if we should do a window update */
12670 		if (bbr_window_update_needed(tp, so, recwin, maxseg))
12671 			goto send;
12672 	}
12673 	/*
12674 	 * Send if we owe the peer an ACK, RST, SYN.  ACKNOW
12675 	 * is also a catch-all for the retransmit timer timeout case.
12676 	 */
12677 	if (tp->t_flags & TF_ACKNOW) {
12678 		goto send;
12679 	}
12680 	if (flags & TH_RST) {
12681 		/* Always send a RST if one is due */
12682 		goto send;
12683 	}
12684 	if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) {
12685 		goto send;
12686 	}
12687 	/*
12688 	 * If our state indicates that FIN should be sent and we have not
12689 	 * yet done so, then we need to send.
12690 	 */
12691 	if (flags & TH_FIN &&
12692 	    ((tp->t_flags & TF_SENTFIN) == 0)) {
12693 		goto send;
12694 	}
12695 	/*
12696 	 * No reason to send a segment, just return.
12697 	 */
12698 just_return:
12699 	SOCKBUF_UNLOCK(sb);
12700 just_return_nolock:
12701 	if (tot_len)
12702 		slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
12703 	if (bbr->rc_no_pacing)
12704 		slot = 0;
12705 	if (tot_len == 0) {
12706 		if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >=
12707 		    tp->snd_wnd) {
12708 			BBR_STAT_INC(bbr_rwnd_limited);
12709 			app_limited = BBR_JR_RWND_LIMITED;
12710 			bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12711 			if ((bbr->rc_in_persist == 0) &&
12712 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
12713 			    (tp->snd_max == tp->snd_una) &&
12714 			    sbavail(&so->so_snd)) {
12715 				/* No send window.. we must enter persist */
12716 				bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
12717 			}
12718 		} else if (ctf_outstanding(tp) >= sbavail(sb)) {
12719 			BBR_STAT_INC(bbr_app_limited);
12720 			app_limited = BBR_JR_APP_LIMITED;
12721 			bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12722 		} else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12723 						 bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) {
12724 			BBR_STAT_INC(bbr_cwnd_limited);
12725  			app_limited = BBR_JR_CWND_LIMITED;
12726 			bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12727 									bbr->r_ctl.rc_lost_bytes)));
12728 			bbr->rc_cwnd_limited = 1;
12729 		} else {
12730 			BBR_STAT_INC(bbr_app_limited);
12731 			app_limited = BBR_JR_APP_LIMITED;
12732 			bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12733 		}
12734 		bbr->r_ctl.rc_hptsi_agg_delay = 0;
12735 		bbr->r_agg_early_set = 0;
12736 		bbr->r_ctl.rc_agg_early = 0;
12737 		bbr->r_ctl.rc_last_delay_val = 0;
12738 	} else if (bbr->rc_use_google == 0)
12739 		bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12740 	/* Are we app limited? */
12741 	if ((app_limited == BBR_JR_APP_LIMITED) ||
12742 	    (app_limited == BBR_JR_RWND_LIMITED)) {
12743 		/**
12744 		 * We are application limited.
12745 		 */
12746 		bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12747 								       bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered);
12748 	}
12749 	if (tot_len == 0)
12750 		counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1);
12751 	/* Dont update the time if we did not send */
12752 	bbr->r_ctl.rc_last_delay_val = 0;
12753 	bbr->rc_output_starts_timer = 1;
12754 	bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len);
12755 	bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len);
12756 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12757 		/* Make sure snd_nxt is drug up */
12758 		tp->snd_nxt = tp->snd_max;
12759 	}
12760 	return (error);
12761 
12762 send:
12763 	if (doing_tlp == 0) {
12764 		/*
12765 		 * Data not a TLP, and its not the rxt firing. If it is the
12766 		 * rxt firing, we want to leave the tlp_in_progress flag on
12767 		 * so we don't send another TLP. It has to be a rack timer
12768 		 * or normal send (response to acked data) to clear the tlp
12769 		 * in progress flag.
12770 		 */
12771 		bbr->rc_tlp_in_progress = 0;
12772 		bbr->rc_tlp_rtx_out = 0;
12773 	} else {
12774 		/*
12775 		 * Its a TLP.
12776 		 */
12777 		bbr->rc_tlp_in_progress = 1;
12778 	}
12779 	bbr_timer_cancel(bbr, __LINE__, cts);
12780 	if (rsm == NULL) {
12781 		if (sbused(sb) > 0) {
12782 			/*
12783 			 * This is sub-optimal. We only send a stand alone
12784 			 * FIN on its own segment.
12785 			 */
12786 			if (flags & TH_FIN) {
12787 				flags &= ~TH_FIN;
12788 				if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) {
12789 					/* Lets not send this */
12790 					slot = 0;
12791 					goto just_return;
12792 				}
12793 			}
12794 		}
12795 	} else {
12796 		/*
12797 		 * We do *not* send a FIN on a retransmit if it has data.
12798 		 * The if clause here where len > 1 should never come true.
12799 		 */
12800 		if ((len > 0) &&
12801 		    (((rsm->r_flags & BBR_HAS_FIN) == 0) &&
12802 		    (flags & TH_FIN))) {
12803 			flags &= ~TH_FIN;
12804 			len--;
12805 		}
12806 	}
12807 	SOCKBUF_LOCK_ASSERT(sb);
12808 	if (len > 0) {
12809 		if ((tp->snd_una == tp->snd_max) &&
12810 		    (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
12811 			/*
12812 			 * This qualifies as a RTT_PROBE session since we
12813 			 * drop the data outstanding to nothing and waited
12814 			 * more than bbr_rtt_probe_time.
12815 			 */
12816 			bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
12817 			bbr_set_reduced_rtt(bbr, cts, __LINE__);
12818 		}
12819 		if (len >= maxseg)
12820 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
12821 		else
12822 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
12823 	}
12824 	/*
12825 	 * Before ESTABLISHED, force sending of initial options unless TCP
12826 	 * set not to do any options. NOTE: we assume that the IP/TCP header
12827 	 * plus TCP options always fit in a single mbuf, leaving room for a
12828 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
12829 	 * + optlen <= MCLBYTES
12830 	 */
12831 	optlen = 0;
12832 #ifdef INET6
12833 	if (isipv6)
12834 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12835 	else
12836 #endif
12837 		hdrlen = sizeof(struct tcpiphdr);
12838 
12839 	/*
12840 	 * Compute options for segment. We only have to care about SYN and
12841 	 * established connection segments.  Options for SYN-ACK segments
12842 	 * are handled in TCP syncache.
12843 	 */
12844 	to.to_flags = 0;
12845 	local_options = 0;
12846 	if ((tp->t_flags & TF_NOOPT) == 0) {
12847 		/* Maximum segment size. */
12848 		if (flags & TH_SYN) {
12849 			to.to_mss = tcp_mssopt(&inp->inp_inc);
12850 			if (tp->t_port)
12851 				to.to_mss -= V_tcp_udp_tunneling_overhead;
12852 			to.to_flags |= TOF_MSS;
12853 			/*
12854 			 * On SYN or SYN|ACK transmits on TFO connections,
12855 			 * only include the TFO option if it is not a
12856 			 * retransmit, as the presence of the TFO option may
12857 			 * have caused the original SYN or SYN|ACK to have
12858 			 * been dropped by a middlebox.
12859 			 */
12860 			if (IS_FASTOPEN(tp->t_flags) &&
12861 			    (tp->t_rxtshift == 0)) {
12862 				if (tp->t_state == TCPS_SYN_RECEIVED) {
12863 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
12864 					to.to_tfo_cookie =
12865 					    (u_int8_t *)&tp->t_tfo_cookie.server;
12866 					to.to_flags |= TOF_FASTOPEN;
12867 					wanted_cookie = 1;
12868 				} else if (tp->t_state == TCPS_SYN_SENT) {
12869 					to.to_tfo_len =
12870 					    tp->t_tfo_client_cookie_len;
12871 					to.to_tfo_cookie =
12872 					    tp->t_tfo_cookie.client;
12873 					to.to_flags |= TOF_FASTOPEN;
12874 					wanted_cookie = 1;
12875 				}
12876 			}
12877 		}
12878 		/* Window scaling. */
12879 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
12880 			to.to_wscale = tp->request_r_scale;
12881 			to.to_flags |= TOF_SCALE;
12882 		}
12883 		/* Timestamps. */
12884 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
12885 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
12886 			to.to_tsval = 	tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset;
12887 			to.to_tsecr = tp->ts_recent;
12888 			to.to_flags |= TOF_TS;
12889 			local_options += TCPOLEN_TIMESTAMP + 2;
12890 		}
12891 		/* Set receive buffer autosizing timestamp. */
12892 		if (tp->rfbuf_ts == 0 &&
12893 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
12894 			tp->rfbuf_ts = 	tcp_tv_to_mssectick(&bbr->rc_tv);
12895 		/* Selective ACK's. */
12896 		if (flags & TH_SYN)
12897 			to.to_flags |= TOF_SACKPERM;
12898 		else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
12899 		    tp->rcv_numsacks > 0) {
12900 			to.to_flags |= TOF_SACK;
12901 			to.to_nsacks = tp->rcv_numsacks;
12902 			to.to_sacks = (u_char *)tp->sackblks;
12903 		}
12904 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
12905 		/* TCP-MD5 (RFC2385). */
12906 		if (tp->t_flags & TF_SIGNATURE)
12907 			to.to_flags |= TOF_SIGNATURE;
12908 #endif				/* TCP_SIGNATURE */
12909 
12910 		/* Processing the options. */
12911 		hdrlen += (optlen = tcp_addoptions(&to, opt));
12912 		/*
12913 		 * If we wanted a TFO option to be added, but it was unable
12914 		 * to fit, ensure no data is sent.
12915 		 */
12916 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
12917 		    !(to.to_flags & TOF_FASTOPEN))
12918 			len = 0;
12919 	}
12920 	if (tp->t_port) {
12921 		if (V_tcp_udp_tunneling_port == 0) {
12922 			/* The port was removed?? */
12923 			SOCKBUF_UNLOCK(&so->so_snd);
12924 			return (EHOSTUNREACH);
12925 		}
12926 		hdrlen += sizeof(struct udphdr);
12927 	}
12928 #ifdef INET6
12929 	if (isipv6)
12930 		ipoptlen = ip6_optlen(inp);
12931 	else
12932 #endif
12933 	if (inp->inp_options)
12934 		ipoptlen = inp->inp_options->m_len -
12935 		    offsetof(struct ipoption, ipopt_list);
12936 	else
12937 		ipoptlen = 0;
12938 	ipoptlen = 0;
12939 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12940 	ipoptlen += ipsec_optlen;
12941 #endif
12942 	if (bbr->rc_last_options != local_options) {
12943 		/*
12944 		 * Cache the options length this generally does not change
12945 		 * on a connection. We use this to calculate TSO.
12946 		 */
12947 		bbr->rc_last_options = local_options;
12948 	}
12949 	maxseg = tp->t_maxseg - (ipoptlen + optlen);
12950 	p_maxseg = min(maxseg, pace_max_segs);
12951 	/*
12952 	 * Adjust data length if insertion of options will bump the packet
12953 	 * length beyond the t_maxseg length. Clear the FIN bit because we
12954 	 * cut off the tail of the segment.
12955 	 */
12956 	if (len > maxseg) {
12957 		if (len != 0 && (flags & TH_FIN)) {
12958 			flags &= ~TH_FIN;
12959 		}
12960 		if (tso) {
12961 			uint32_t moff;
12962 			int32_t max_len;
12963 
12964 			/* extract TSO information */
12965 			if_hw_tsomax = tp->t_tsomax;
12966 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
12967 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
12968 			KASSERT(ipoptlen == 0,
12969 			    ("%s: TSO can't do IP options", __func__));
12970 
12971 			/*
12972 			 * Check if we should limit by maximum payload
12973 			 * length:
12974 			 */
12975 			if (if_hw_tsomax != 0) {
12976 				/* compute maximum TSO length */
12977 				max_len = (if_hw_tsomax - hdrlen -
12978 				    max_linkhdr);
12979 				if (max_len <= 0) {
12980 					len = 0;
12981 				} else if (len > max_len) {
12982 					len = max_len;
12983 				}
12984 			}
12985 			/*
12986 			 * Prevent the last segment from being fractional
12987 			 * unless the send sockbuf can be emptied:
12988 			 */
12989 			if ((sb_offset + len) < sbavail(sb)) {
12990 				moff = len % (uint32_t)maxseg;
12991 				if (moff != 0) {
12992 					len -= moff;
12993 				}
12994 			}
12995 			/*
12996 			 * In case there are too many small fragments don't
12997 			 * use TSO:
12998 			 */
12999 			if (len <= maxseg) {
13000 				len = maxseg;
13001 				tso = 0;
13002 			}
13003 		} else {
13004 			/* Not doing TSO */
13005 			if (optlen + ipoptlen >= tp->t_maxseg) {
13006 				/*
13007 				 * Since we don't have enough space to put
13008 				 * the IP header chain and the TCP header in
13009 				 * one packet as required by RFC 7112, don't
13010 				 * send it. Also ensure that at least one
13011 				 * byte of the payload can be put into the
13012 				 * TCP segment.
13013 				 */
13014 				SOCKBUF_UNLOCK(&so->so_snd);
13015 				error = EMSGSIZE;
13016 				sack_rxmit = 0;
13017 				goto out;
13018 			}
13019 			len = maxseg;
13020 		}
13021 	} else {
13022 		/* Not doing TSO */
13023 		if_hw_tsomaxsegcount = 0;
13024 		tso = 0;
13025 	}
13026 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
13027 	    ("%s: len > IP_MAXPACKET", __func__));
13028 #ifdef DIAGNOSTIC
13029 #ifdef INET6
13030 	if (max_linkhdr + hdrlen > MCLBYTES)
13031 #else
13032 	if (max_linkhdr + hdrlen > MHLEN)
13033 #endif
13034 		panic("tcphdr too big");
13035 #endif
13036 	/*
13037 	 * This KASSERT is here to catch edge cases at a well defined place.
13038 	 * Before, those had triggered (random) panic conditions further
13039 	 * down.
13040 	 */
13041 #ifdef BBR_INVARIANTS
13042 	if (sack_rxmit) {
13043 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
13044 			panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u",
13045 			    rsm, tp, bbr, rsm->r_start, tp->snd_una);
13046 		}
13047 	}
13048 #endif
13049 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
13050 	if ((len == 0) &&
13051 	    (flags & TH_FIN) &&
13052 	    (sbused(sb))) {
13053 		/*
13054 		 * We have outstanding data, don't send a fin by itself!.
13055 		 */
13056 		slot = 0;
13057 		goto just_return;
13058 	}
13059 	/*
13060 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
13061 	 * and initialize the header from the template for sends on this
13062 	 * connection.
13063 	 */
13064 	if (len) {
13065 		uint32_t moff;
13066 
13067 		/*
13068 		 * We place a limit on sending with hptsi.
13069 		 */
13070 		if ((rsm == NULL) && len > pace_max_segs)
13071 			len = pace_max_segs;
13072 		if (len <= maxseg)
13073 			tso = 0;
13074 #ifdef INET6
13075 		if (MHLEN < hdrlen + max_linkhdr)
13076 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
13077 		else
13078 #endif
13079 			m = m_gethdr(M_NOWAIT, MT_DATA);
13080 
13081 		if (m == NULL) {
13082 			BBR_STAT_INC(bbr_failed_mbuf_aloc);
13083 			bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13084 			SOCKBUF_UNLOCK(sb);
13085 			error = ENOBUFS;
13086 			sack_rxmit = 0;
13087 			goto out;
13088 		}
13089 		m->m_data += max_linkhdr;
13090 		m->m_len = hdrlen;
13091 		/*
13092 		 * Start the m_copy functions from the closest mbuf to the
13093 		 * sb_offset in the socket buffer chain.
13094 		 */
13095 		if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) {
13096 #ifdef BBR_INVARIANTS
13097 			if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0)))
13098 				panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u",
13099 				    tp, bbr, len, sb_offset, sbavail(sb), rsm,
13100 				    doing_retran_from,
13101 				    picked_up_retran,
13102 				    doing_tlp);
13103 
13104 #endif
13105 			/*
13106 			 * In this messed up situation we have two choices,
13107 			 * a) pretend the send worked, and just start timers
13108 			 * and what not (not good since that may lead us
13109 			 * back here a lot). <or> b) Send the lowest segment
13110 			 * in the map. <or> c) Drop the connection. Lets do
13111 			 * <b> which if it continues to happen will lead to
13112 			 * <c> via timeouts.
13113 			 */
13114 			BBR_STAT_INC(bbr_offset_recovery);
13115 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
13116 			sb_offset = 0;
13117 			if (rsm == NULL) {
13118 				sack_rxmit = 0;
13119 				len = sbavail(sb);
13120 			} else {
13121 				sack_rxmit = 1;
13122 				if (rsm->r_start != tp->snd_una) {
13123 					/*
13124 					 * Things are really messed up, <c>
13125 					 * is the only thing to do.
13126 					 */
13127 					BBR_STAT_INC(bbr_offset_drop);
13128 					SOCKBUF_UNLOCK(sb);
13129 					(void)m_free(m);
13130 					return (-EFAULT); /* tcp_drop() */
13131 				}
13132 				len = rsm->r_end - rsm->r_start;
13133 			}
13134 			if (len > sbavail(sb))
13135 				len = sbavail(sb);
13136 			if (len > maxseg)
13137 				len = maxseg;
13138 		}
13139 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
13140 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
13141 			m_copydata(mb, moff, (int)len,
13142 			    mtod(m, caddr_t)+hdrlen);
13143 			if (rsm == NULL)
13144 				sbsndptr_adv(sb, mb, len);
13145 			m->m_len += len;
13146 		} else {
13147 			struct sockbuf *msb;
13148 
13149 			if (rsm)
13150 				msb = NULL;
13151 			else
13152 				msb = sb;
13153 #ifdef BBR_INVARIANTS
13154 			if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) {
13155 				if (rsm) {
13156 					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 ",
13157 					    tp, bbr, len, moff,
13158 					    sbavail(sb), rsm,
13159 					    tp->snd_una, rsm->r_flags, rsm->r_start,
13160 					    doing_retran_from,
13161 					    picked_up_retran,
13162 					    doing_tlp, sack_rxmit);
13163 				} else {
13164 					panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u",
13165 					    tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una);
13166 				}
13167 			}
13168 #endif
13169 			m->m_next = tcp_m_copym(
13170 				mb, moff, &len,
13171 				if_hw_tsomaxsegcount,
13172 				if_hw_tsomaxsegsize, msb,
13173 				((rsm == NULL) ? hw_tls : 0)
13174 #ifdef NETFLIX_COPY_ARGS
13175 				, NULL, NULL
13176 #endif
13177 				);
13178 			if (len <= maxseg) {
13179 				/*
13180 				 * Must have ran out of mbufs for the copy
13181 				 * shorten it to no longer need tso. Lets
13182 				 * not put on sendalot since we are low on
13183 				 * mbufs.
13184 				 */
13185 				tso = 0;
13186 			}
13187 			if (m->m_next == NULL) {
13188 				SOCKBUF_UNLOCK(sb);
13189 				(void)m_free(m);
13190 				error = ENOBUFS;
13191 				sack_rxmit = 0;
13192 				goto out;
13193 			}
13194 		}
13195 #ifdef BBR_INVARIANTS
13196 		if (tso && len < maxseg) {
13197 			panic("tp:%p tso on, but len:%d < maxseg:%d",
13198 			    tp, len, maxseg);
13199 		}
13200 		if (tso && if_hw_tsomaxsegcount) {
13201 			int32_t seg_cnt = 0;
13202 			struct mbuf *foo;
13203 
13204 			foo = m;
13205 			while (foo) {
13206 				seg_cnt++;
13207 				foo = foo->m_next;
13208 			}
13209 			if (seg_cnt > if_hw_tsomaxsegcount) {
13210 				panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount);
13211 			}
13212 		}
13213 #endif
13214 		/*
13215 		 * If we're sending everything we've got, set PUSH. (This
13216 		 * will keep happy those implementations which only give
13217 		 * data to the user when a buffer fills or a PUSH comes in.)
13218 		 */
13219 		if (sb_offset + len == sbused(sb) &&
13220 		    sbused(sb) &&
13221 		    !(flags & TH_SYN)) {
13222 			flags |= TH_PUSH;
13223 		}
13224 		SOCKBUF_UNLOCK(sb);
13225 	} else {
13226 		SOCKBUF_UNLOCK(sb);
13227 		if (tp->t_flags & TF_ACKNOW)
13228 			KMOD_TCPSTAT_INC(tcps_sndacks);
13229 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
13230 			KMOD_TCPSTAT_INC(tcps_sndctrl);
13231 		else
13232 			KMOD_TCPSTAT_INC(tcps_sndwinup);
13233 
13234 		m = m_gethdr(M_NOWAIT, MT_DATA);
13235 		if (m == NULL) {
13236 			BBR_STAT_INC(bbr_failed_mbuf_aloc);
13237 			bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13238 			error = ENOBUFS;
13239 			/* Fudge the send time since we could not send */
13240 			sack_rxmit = 0;
13241 			goto out;
13242 		}
13243 #ifdef INET6
13244 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13245 		    MHLEN >= hdrlen) {
13246 			M_ALIGN(m, hdrlen);
13247 		} else
13248 #endif
13249 			m->m_data += max_linkhdr;
13250 		m->m_len = hdrlen;
13251 	}
13252 	SOCKBUF_UNLOCK_ASSERT(sb);
13253 	m->m_pkthdr.rcvif = (struct ifnet *)0;
13254 #ifdef MAC
13255 	mac_inpcb_create_mbuf(inp, m);
13256 #endif
13257 #ifdef INET6
13258 	if (isipv6) {
13259 		ip6 = mtod(m, struct ip6_hdr *);
13260 		if (tp->t_port) {
13261 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
13262 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13263 			udp->uh_dport = tp->t_port;
13264 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
13265 			udp->uh_ulen = htons(ulen);
13266 			th = (struct tcphdr *)(udp + 1);
13267 		} else {
13268 			th = (struct tcphdr *)(ip6 + 1);
13269 		}
13270 		tcpip_fillheaders(inp, tp->t_port, ip6, th);
13271 	} else
13272 #endif				/* INET6 */
13273 	{
13274 		ip = mtod(m, struct ip *);
13275 		if (tp->t_port) {
13276 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
13277 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13278 			udp->uh_dport = tp->t_port;
13279 			ulen = hdrlen + len - sizeof(struct ip);
13280 			udp->uh_ulen = htons(ulen);
13281 			th = (struct tcphdr *)(udp + 1);
13282 		} else {
13283 			th = (struct tcphdr *)(ip + 1);
13284 		}
13285 		tcpip_fillheaders(inp, tp->t_port, ip, th);
13286 	}
13287 	/*
13288 	 * If we are doing retransmissions, then snd_nxt will not reflect
13289 	 * the first unsent octet.  For ACK only packets, we do not want the
13290 	 * sequence number of the retransmitted packet, we want the sequence
13291 	 * number of the next unsent octet.  So, if there is no data (and no
13292 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
13293 	 * ti_seq.  But if we are in persist state, snd_max might reflect
13294 	 * one byte beyond the right edge of the window, so use snd_nxt in
13295 	 * that case, since we know we aren't doing a retransmission.
13296 	 * (retransmit and persist are mutually exclusive...)
13297 	 */
13298 	if (sack_rxmit == 0) {
13299 		if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) {
13300 			/* New data (including new persists) */
13301 			th->th_seq = htonl(tp->snd_max);
13302 			bbr_seq = tp->snd_max;
13303 		} else if (flags & TH_SYN) {
13304 			/* Syn's always send from iss */
13305 			th->th_seq = htonl(tp->iss);
13306 			bbr_seq = tp->iss;
13307 		} else if (flags & TH_FIN) {
13308 			if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) {
13309 				/*
13310 				 * If we sent the fin already its 1 minus
13311 				 * snd_max
13312 				 */
13313 				th->th_seq = (htonl(tp->snd_max - 1));
13314 				bbr_seq = (tp->snd_max - 1);
13315 			} else {
13316 				/* First time FIN use snd_max */
13317 				th->th_seq = htonl(tp->snd_max);
13318 				bbr_seq = tp->snd_max;
13319 			}
13320 		} else {
13321 			/*
13322 			 * len == 0 and not persist we use snd_max, sending
13323 			 * an ack unless we have sent the fin then its 1
13324 			 * minus.
13325 			 */
13326 			/*
13327 			 * XXXRRS Question if we are in persists and we have
13328 			 * nothing outstanding to send and we have not sent
13329 			 * a FIN, we will send an ACK. In such a case it
13330 			 * might be better to send (tp->snd_una - 1) which
13331 			 * would force the peer to ack.
13332 			 */
13333 			if (tp->t_flags & TF_SENTFIN) {
13334 				th->th_seq = htonl(tp->snd_max - 1);
13335 				bbr_seq = (tp->snd_max - 1);
13336 			} else {
13337 				th->th_seq = htonl(tp->snd_max);
13338 				bbr_seq = tp->snd_max;
13339 			}
13340 		}
13341 	} else {
13342 		/* All retransmits use the rsm to guide the send */
13343 		th->th_seq = htonl(rsm->r_start);
13344 		bbr_seq = rsm->r_start;
13345 	}
13346 	th->th_ack = htonl(tp->rcv_nxt);
13347 	if (optlen) {
13348 		bcopy(opt, th + 1, optlen);
13349 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13350 	}
13351 	tcp_set_flags(th, flags);
13352 	/*
13353 	 * Calculate receive window.  Don't shrink window, but avoid silly
13354 	 * window syndrome.
13355 	 */
13356 	if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) &&
13357 				  recwin < maxseg)))
13358 		recwin = 0;
13359 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13360 	    recwin < (tp->rcv_adv - tp->rcv_nxt))
13361 		recwin = (tp->rcv_adv - tp->rcv_nxt);
13362 	if (recwin > TCP_MAXWIN << tp->rcv_scale)
13363 		recwin = TCP_MAXWIN << tp->rcv_scale;
13364 
13365 	/*
13366 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13367 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
13368 	 * handled in syncache.
13369 	 */
13370 	if (flags & TH_SYN)
13371 		th->th_win = htons((u_short)
13372 		    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13373 	else {
13374 		/* Avoid shrinking window with window scaling. */
13375 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
13376 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13377 	}
13378 	/*
13379 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13380 	 * window.  This may cause the remote transmitter to stall.  This
13381 	 * flag tells soreceive() to disable delayed acknowledgements when
13382 	 * draining the buffer.  This can occur if the receiver is
13383 	 * attempting to read more data than can be buffered prior to
13384 	 * transmitting on the connection.
13385 	 */
13386 	if (th->th_win == 0) {
13387 		tp->t_sndzerowin++;
13388 		tp->t_flags |= TF_RXWIN0SENT;
13389 	} else
13390 		tp->t_flags &= ~TF_RXWIN0SENT;
13391 	/*
13392 	 * We don't support urgent data, but drag along
13393 	 * the pointer in case of a stack switch.
13394 	 */
13395 	tp->snd_up = tp->snd_una;
13396 
13397 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13398 	if (to.to_flags & TOF_SIGNATURE) {
13399 		/*
13400 		 * Calculate MD5 signature and put it into the place
13401 		 * determined before. NOTE: since TCP options buffer doesn't
13402 		 * point into mbuf's data, calculate offset and use it.
13403 		 */
13404 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13405 		    (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13406 			/*
13407 			 * Do not send segment if the calculation of MD5
13408 			 * digest has failed.
13409 			 */
13410 			goto out;
13411 		}
13412 	}
13413 #endif
13414 
13415 	/*
13416 	 * Put TCP length in extended header, and then checksum extended
13417 	 * header and data.
13418 	 */
13419 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
13420 #ifdef INET6
13421 	if (isipv6) {
13422 		/*
13423 		 * ip6_plen is not need to be filled now, and will be filled
13424 		 * in ip6_output.
13425 		 */
13426 		if (tp->t_port) {
13427 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13428 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13429 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13430 			th->th_sum = htons(0);
13431 			UDPSTAT_INC(udps_opackets);
13432 		} else {
13433 			csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13434 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13435 			th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
13436 			    optlen + len, IPPROTO_TCP, 0);
13437 		}
13438 	}
13439 #endif
13440 #if defined(INET6) && defined(INET)
13441 	else
13442 #endif
13443 #ifdef INET
13444 	{
13445 		if (tp->t_port) {
13446 			m->m_pkthdr.csum_flags = CSUM_UDP;
13447 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13448 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13449 			    ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13450 			th->th_sum = htons(0);
13451 			UDPSTAT_INC(udps_opackets);
13452 		} else {
13453 			csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP;
13454 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13455 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
13456 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13457 			    IPPROTO_TCP + len + optlen));
13458 		}
13459 		/* IP version must be set here for ipv4/ipv6 checking later */
13460 		KASSERT(ip->ip_v == IPVERSION,
13461 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
13462 	}
13463 #endif
13464 
13465 	/*
13466 	 * Enable TSO and specify the size of the segments. The TCP pseudo
13467 	 * header checksum is always provided. XXX: Fixme: This is currently
13468 	 * not the case for IPv6.
13469 	 */
13470 	if (tso) {
13471 		KASSERT(len > maxseg,
13472 		    ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg));
13473 		m->m_pkthdr.csum_flags |= CSUM_TSO;
13474 		csum_flags |= CSUM_TSO;
13475 		m->m_pkthdr.tso_segsz = maxseg;
13476 	}
13477 	KASSERT(len + hdrlen == m_length(m, NULL),
13478 	    ("%s: mbuf chain different than expected: %d + %u != %u",
13479 	    __func__, len, hdrlen, m_length(m, NULL)));
13480 
13481 #ifdef TCP_HHOOK
13482 	/* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */
13483 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
13484 #endif
13485 
13486 	/* Log to the black box */
13487 	if (tcp_bblogging_on(tp)) {
13488 		union tcp_log_stackspecific log;
13489 
13490 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
13491 		/* Record info on type of transmission */
13492 		log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay;
13493 		log.u_bbr.flex2 = (bbr->r_recovery_bw << 3);
13494 		log.u_bbr.flex3 = maxseg;
13495 		log.u_bbr.flex4 = delay_calc;
13496 		log.u_bbr.flex5 = bbr->rc_past_init_win;
13497 		log.u_bbr.flex5 <<= 1;
13498 		log.u_bbr.flex5 |= bbr->rc_no_pacing;
13499 		log.u_bbr.flex5 <<= 29;
13500 		log.u_bbr.flex5 |= tp->t_maxseg;
13501 		log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs;
13502 		log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr);
13503 		/* lets poke in the low and the high here for debugging */
13504 		log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
13505 		if (rsm || sack_rxmit) {
13506 			if (doing_tlp)
13507 				log.u_bbr.flex8 = 2;
13508 			else
13509 				log.u_bbr.flex8 = 1;
13510 		} else {
13511 			log.u_bbr.flex8 = 0;
13512 		}
13513 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13514 		    len, &log, false, NULL, NULL, 0, tv);
13515 	} else {
13516 		lgb = NULL;
13517 	}
13518 	/*
13519 	 * Fill in IP length and desired time to live and send to IP level.
13520 	 * There should be a better way to handle ttl and tos; we could keep
13521 	 * them in the template, but need a way to checksum without them.
13522 	 */
13523 	/*
13524 	 * m->m_pkthdr.len should have been set before cksum calcuration,
13525 	 * because in6_cksum() need it.
13526 	 */
13527 #ifdef INET6
13528 	if (isipv6) {
13529 		/*
13530 		 * we separately set hoplimit for every segment, since the
13531 		 * user might want to change the value via setsockopt. Also,
13532 		 * desired default hop limit might be changed via Neighbor
13533 		 * Discovery.
13534 		 */
13535 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13536 
13537 		/*
13538 		 * Set the packet size here for the benefit of DTrace
13539 		 * probes. ip6_output() will set it properly; it's supposed
13540 		 * to include the option header lengths as well.
13541 		 */
13542 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13543 
13544 		if (V_path_mtu_discovery && maxseg > V_tcp_minmss)
13545 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13546 		else
13547 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13548 
13549 		if (tp->t_state == TCPS_SYN_SENT)
13550 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13551 
13552 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13553 		/* TODO: IPv6 IP6TOS_ECT bit on */
13554 		error = ip6_output(m, inp->in6p_outputopts,
13555 		    &inp->inp_route6,
13556 		    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13557 		    NULL, NULL, inp);
13558 
13559 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
13560 			mtu = inp->inp_route6.ro_nh->nh_mtu;
13561 	}
13562 #endif				/* INET6 */
13563 #if defined(INET) && defined(INET6)
13564 	else
13565 #endif
13566 #ifdef INET
13567 	{
13568 		ip->ip_len = htons(m->m_pkthdr.len);
13569 #ifdef INET6
13570 		if (isipv6)
13571 			ip->ip_ttl = in6_selecthlim(inp, NULL);
13572 #endif				/* INET6 */
13573 		/*
13574 		 * If we do path MTU discovery, then we set DF on every
13575 		 * packet. This might not be the best thing to do according
13576 		 * to RFC3390 Section 2. However the tcp hostcache migitates
13577 		 * the problem so it affects only the first tcp connection
13578 		 * with a host.
13579 		 *
13580 		 * NB: Don't set DF on small MTU/MSS to have a safe
13581 		 * fallback.
13582 		 */
13583 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13584 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13585 			if (tp->t_port == 0 || len < V_tcp_minmss) {
13586 				ip->ip_off |= htons(IP_DF);
13587 			}
13588 		} else {
13589 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13590 		}
13591 
13592 		if (tp->t_state == TCPS_SYN_SENT)
13593 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
13594 
13595 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
13596 
13597 		error = ip_output(m, inp->inp_options, &inp->inp_route,
13598 		    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
13599 		    inp);
13600 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
13601 			mtu = inp->inp_route.ro_nh->nh_mtu;
13602 	}
13603 #endif				/* INET */
13604 out:
13605 
13606 	if (lgb) {
13607 		lgb->tlb_errno = error;
13608 		lgb = NULL;
13609 	}
13610 	/*
13611 	 * In transmit state, time the transmission and arrange for the
13612 	 * retransmit.  In persist state, just set snd_max.
13613 	 */
13614 	if (error == 0) {
13615 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
13616 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13617 		    (tp->t_flags & TF_SACK_PERMIT) &&
13618 		    tp->rcv_numsacks > 0)
13619 			tcp_clean_dsack_blocks(tp);
13620 		/* We sent an ack clear the bbr_segs_rcvd count */
13621 		bbr->output_error_seen = 0;
13622 		bbr->oerror_cnt = 0;
13623 		bbr->bbr_segs_rcvd = 0;
13624 		if (len == 0)
13625 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1);
13626 		/* Do accounting for new sends */
13627 		if ((len > 0) && (rsm == NULL)) {
13628 			int idx;
13629 			if (tp->snd_una == tp->snd_max) {
13630 				/*
13631 				 * Special case to match google, when
13632 				 * nothing is in flight the delivered
13633 				 * time does get updated to the current
13634 				 * time (see tcp_rate_bsd.c).
13635 				 */
13636 				bbr->r_ctl.rc_del_time = cts;
13637 			}
13638 			if (len >= maxseg) {
13639 				idx = (len / maxseg) + 3;
13640 				if (idx >= TCP_MSS_ACCT_ATIMER)
13641 					counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1);
13642 				else
13643 					counter_u64_add(bbr_out_size[idx], 1);
13644 			} else {
13645 				/* smaller than a MSS */
13646 				idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options);
13647 				if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV)
13648 					idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1);
13649 				counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1);
13650 			}
13651 		}
13652 	}
13653 	abandon = 0;
13654 	/*
13655 	 * We must do the send accounting before we log the output,
13656 	 * otherwise the state of the rsm could change and we account to the
13657 	 * wrong bucket.
13658 	 */
13659 	if (len > 0) {
13660 		bbr_do_send_accounting(tp, bbr, rsm, len, error);
13661 		if (error == 0) {
13662 			if (tp->snd_una == tp->snd_max)
13663 				bbr->r_ctl.rc_tlp_rxt_last_time = cts;
13664 		}
13665 	}
13666 	bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error,
13667 	    cts, mb, &abandon, rsm, 0, sb);
13668 	if (abandon) {
13669 		/*
13670 		 * If bbr_log_output destroys the TCB or sees a TH_RST being
13671 		 * sent we should hit this condition.
13672 		 */
13673 		return (0);
13674 	}
13675 	if (bbr->rc_in_persist == 0) {
13676 		/*
13677 		 * Advance snd_nxt over sequence space of this segment.
13678 		 */
13679 		if (error)
13680 			/* We don't log or do anything with errors */
13681 			goto skip_upd;
13682 
13683 		if (tp->snd_una == tp->snd_max &&
13684 		    (len || (flags & (TH_SYN | TH_FIN)))) {
13685 			/*
13686 			 * Update the time we just added data since none was
13687 			 * outstanding.
13688 			 */
13689 			bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13690 			bbr->rc_tp->t_acktime  = ticks;
13691 		}
13692 		if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) {
13693 			if (flags & TH_SYN) {
13694 				/*
13695 				 * Smack the snd_max to iss + 1
13696 				 * if its a FO we will add len below.
13697 				 */
13698 				tp->snd_max = tp->iss + 1;
13699 			}
13700 			if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13701 				tp->snd_max++;
13702 				tp->t_flags |= TF_SENTFIN;
13703 			}
13704 		}
13705 		if (sack_rxmit == 0)
13706 			tp->snd_max += len;
13707 skip_upd:
13708 		if ((error == 0) && len)
13709 			tot_len += len;
13710 	} else {
13711 		/* Persists case */
13712 		int32_t xlen = len;
13713 
13714 		if (error)
13715 			goto nomore;
13716 
13717 		if (flags & TH_SYN)
13718 			++xlen;
13719 		if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13720 			++xlen;
13721 			tp->t_flags |= TF_SENTFIN;
13722 		}
13723 		if (xlen && (tp->snd_una == tp->snd_max)) {
13724 			/*
13725 			 * Update the time we just added data since none was
13726 			 * outstanding.
13727 			 */
13728 			bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13729 			bbr->rc_tp->t_acktime = ticks;
13730 		}
13731 		if (sack_rxmit == 0)
13732 			tp->snd_max += xlen;
13733 		tot_len += (len + optlen + ipoptlen);
13734 	}
13735 nomore:
13736 	if (error) {
13737 		/*
13738 		 * Failures do not advance the seq counter above. For the
13739 		 * case of ENOBUFS we will fall out and become ack-clocked.
13740 		 * capping the cwnd at the current flight.
13741 		 * Everything else will just have to retransmit with the timer
13742 		 * (no pacer).
13743 		 */
13744 		SOCKBUF_UNLOCK_ASSERT(sb);
13745 		BBR_STAT_INC(bbr_saw_oerr);
13746 		/* Clear all delay/early tracks */
13747 		bbr->r_ctl.rc_hptsi_agg_delay = 0;
13748 		bbr->r_ctl.rc_agg_early = 0;
13749 		bbr->r_agg_early_set = 0;
13750 		bbr->output_error_seen = 1;
13751 		if (bbr->oerror_cnt < 0xf)
13752 			bbr->oerror_cnt++;
13753 		if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) {
13754 			/* drop the session */
13755 			return (-ENETDOWN);
13756 		}
13757 		switch (error) {
13758 		case ENOBUFS:
13759 			/*
13760 			 * Make this guy have to get ack's to send
13761 			 * more but lets make sure we don't
13762 			 * slam him below a T-O (1MSS).
13763 			 */
13764 			if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
13765 				tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13766 								    bbr->r_ctl.rc_lost_bytes)) - maxseg;
13767 				if (tp->snd_cwnd < maxseg)
13768 					tp->snd_cwnd = maxseg;
13769 			}
13770 			slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt;
13771 			BBR_STAT_INC(bbr_saw_enobuf);
13772 			if (bbr->bbr_hdrw_pacing)
13773 				counter_u64_add(bbr_hdwr_pacing_enobuf, 1);
13774 			else
13775 				counter_u64_add(bbr_nohdwr_pacing_enobuf, 1);
13776 			/*
13777 			 * Here even in the enobuf's case we want to do our
13778 			 * state update. The reason being we may have been
13779 			 * called by the input function. If so we have had
13780 			 * things change.
13781 			 */
13782 			error = 0;
13783 			goto enobufs;
13784 		case EMSGSIZE:
13785 			/*
13786 			 * For some reason the interface we used initially
13787 			 * to send segments changed to another or lowered
13788 			 * its MTU. If TSO was active we either got an
13789 			 * interface without TSO capabilits or TSO was
13790 			 * turned off. If we obtained mtu from ip_output()
13791 			 * then update it and try again.
13792 			 */
13793 			/* Turn on tracing (or try to) */
13794 			{
13795 				int old_maxseg;
13796 
13797 				old_maxseg = tp->t_maxseg;
13798 				BBR_STAT_INC(bbr_saw_emsgsiz);
13799 				bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts);
13800 				if (mtu != 0)
13801 					tcp_mss_update(tp, -1, mtu, NULL, NULL);
13802 				if (old_maxseg <= tp->t_maxseg) {
13803 					/* Huh it did not shrink? */
13804 					tp->t_maxseg = old_maxseg - 40;
13805 					bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts);
13806 				}
13807 				/*
13808 				 * Nuke all other things that can interfere
13809 				 * with slot
13810 				 */
13811 				if ((tot_len + len) && (len >= tp->t_maxseg)) {
13812 					slot = bbr_get_pacing_delay(bbr,
13813 					    bbr->r_ctl.rc_bbr_hptsi_gain,
13814 					    (tot_len + len), cts, 0);
13815 					if (slot < bbr_error_base_paceout)
13816 						slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
13817 				} else
13818 					slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
13819 				bbr->rc_output_starts_timer = 1;
13820 				bbr_start_hpts_timer(bbr, tp, cts, 10, slot,
13821 				    tot_len);
13822 				return (error);
13823 			}
13824 		case EPERM:
13825 			tp->t_softerror = error;
13826 			/* Fall through */
13827 		case EHOSTDOWN:
13828 		case EHOSTUNREACH:
13829 		case ENETDOWN:
13830 		case ENETUNREACH:
13831 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
13832 				tp->t_softerror = error;
13833 			}
13834 			/* FALLTHROUGH */
13835 		default:
13836 			slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt;
13837 			bbr->rc_output_starts_timer = 1;
13838 			bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0);
13839 			return (error);
13840 		}
13841 #ifdef STATS
13842 	} else if (((tp->t_flags & TF_GPUTINPROG) == 0) &&
13843 		    len &&
13844 		    (rsm == NULL) &&
13845 	    (bbr->rc_in_persist == 0)) {
13846 		tp->gput_seq = bbr_seq;
13847 		tp->gput_ack = bbr_seq +
13848 		    min(sbavail(&so->so_snd) - sb_offset, sendwin);
13849 		tp->gput_ts = cts;
13850 		tp->t_flags |= TF_GPUTINPROG;
13851 #endif
13852 	}
13853 	KMOD_TCPSTAT_INC(tcps_sndtotal);
13854 	if ((bbr->bbr_hdw_pace_ena) &&
13855 	    (bbr->bbr_attempt_hdwr_pace == 0) &&
13856 	    (bbr->rc_past_init_win) &&
13857 	    (bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
13858 	    (get_filter_value(&bbr->r_ctl.rc_delrate)) &&
13859 	    (inp->inp_route.ro_nh &&
13860 	     inp->inp_route.ro_nh->nh_ifp)) {
13861 		/*
13862 		 * We are past the initial window and
13863 		 * have at least one measurement so we
13864 		 * could use hardware pacing if its available.
13865 		 * We have an interface and we have not attempted
13866 		 * to setup hardware pacing, lets try to now.
13867 		 */
13868 		uint64_t rate_wanted;
13869 		int err = 0;
13870 
13871 		rate_wanted = bbr_get_hardware_rate(bbr);
13872 		bbr->bbr_attempt_hdwr_pace = 1;
13873 		bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp,
13874 						      inp->inp_route.ro_nh->nh_ifp,
13875 						      rate_wanted,
13876 						      (RS_PACING_GEQ|RS_PACING_SUB_OK),
13877 						      &err, NULL);
13878 		if (bbr->r_ctl.crte) {
13879 			bbr_type_log_hdwr_pacing(bbr,
13880 						 bbr->r_ctl.crte->ptbl->rs_ifp,
13881 						 rate_wanted,
13882 						 bbr->r_ctl.crte->rate,
13883 						 __LINE__, cts, err);
13884 			BBR_STAT_INC(bbr_hdwr_rl_add_ok);
13885 			counter_u64_add(bbr_flows_nohdwr_pacing, -1);
13886 			counter_u64_add(bbr_flows_whdwr_pacing, 1);
13887 			bbr->bbr_hdrw_pacing = 1;
13888 			/* Now what is our gain status? */
13889 			if (bbr->r_ctl.crte->rate < rate_wanted) {
13890 				/* We have a problem */
13891 				bbr_setup_less_of_rate(bbr, cts,
13892 						       bbr->r_ctl.crte->rate, rate_wanted);
13893 			} else {
13894 				/* We are good */
13895 				bbr->gain_is_limited = 0;
13896 				bbr->skip_gain = 0;
13897 			}
13898 			tcp_bbr_tso_size_check(bbr, cts);
13899 		} else {
13900 			bbr_type_log_hdwr_pacing(bbr,
13901 						 inp->inp_route.ro_nh->nh_ifp,
13902 						 rate_wanted,
13903 						 0,
13904 						 __LINE__, cts, err);
13905 			BBR_STAT_INC(bbr_hdwr_rl_add_fail);
13906 		}
13907 	}
13908 	if (bbr->bbr_hdrw_pacing) {
13909 		/*
13910 		 * Worry about cases where the route
13911 		 * changes or something happened that we
13912 		 * lost our hardware pacing possibly during
13913 		 * the last ip_output call.
13914 		 */
13915 		if (inp->inp_snd_tag == NULL) {
13916 			/* A change during ip output disabled hw pacing? */
13917 			bbr->bbr_hdrw_pacing = 0;
13918 		} else if ((inp->inp_route.ro_nh == NULL) ||
13919 		    (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) {
13920 			/*
13921 			 * We had an interface or route change,
13922 			 * detach from the current hdwr pacing
13923 			 * and setup to re-attempt next go
13924 			 * round.
13925 			 */
13926 			bbr->bbr_hdrw_pacing = 0;
13927 			bbr->bbr_attempt_hdwr_pace = 0;
13928 			tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
13929 			tcp_bbr_tso_size_check(bbr, cts);
13930 		}
13931 	}
13932 	/*
13933 	 * Data sent (as far as we can tell). If this advertises a larger
13934 	 * window than any other segment, then remember the size of the
13935 	 * advertised window. Any pending ACK has now been sent.
13936 	 */
13937 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
13938 		tp->rcv_adv = tp->rcv_nxt + recwin;
13939 
13940 	tp->last_ack_sent = tp->rcv_nxt;
13941 	if ((error == 0) &&
13942 	    (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) &&
13943 	    (doing_tlp == 0) &&
13944 	    (tso == 0) &&
13945 	    (len > 0) &&
13946 	    ((flags & TH_RST) == 0) &&
13947 	    ((flags & TH_SYN) == 0) &&
13948 	    (IN_RECOVERY(tp->t_flags) == 0) &&
13949 	    (bbr->rc_in_persist == 0) &&
13950 	    (tot_len < bbr->r_ctl.rc_pace_max_segs)) {
13951 		/*
13952 		 * For non-tso we need to goto again until we have sent out
13953 		 * enough data to match what we are hptsi out every hptsi
13954 		 * interval.
13955 		 */
13956 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
13957 			/* Make sure snd_nxt is drug up */
13958 			tp->snd_nxt = tp->snd_max;
13959 		}
13960 		if (rsm != NULL) {
13961 			rsm = NULL;
13962 			goto skip_again;
13963 		}
13964 		rsm = NULL;
13965 		sack_rxmit = 0;
13966 		tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
13967 		goto again;
13968 	}
13969 skip_again:
13970 	if ((error == 0) && (flags & TH_FIN))
13971 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
13972 	if ((error == 0) && (flags & TH_RST))
13973 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13974 	if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) {
13975 		/*
13976 		 * Calculate/Re-Calculate the hptsi slot in usecs based on
13977 		 * what we have sent so far
13978 		 */
13979 		slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
13980 		if (bbr->rc_no_pacing)
13981 			slot = 0;
13982 	}
13983 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
13984 enobufs:
13985 	if (bbr->rc_use_google == 0)
13986 		bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
13987 	bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13988 							bbr->r_ctl.rc_lost_bytes)));
13989 	bbr->rc_output_starts_timer = 1;
13990 	if (bbr->bbr_use_rack_cheat &&
13991 	    (more_to_rxt ||
13992 	     ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) {
13993 		/* Rack cheats and shotguns out all rxt's 1ms apart */
13994 		if (slot > 1000)
13995 			slot = 1000;
13996 	}
13997 	if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) {
13998 		/*
13999 		 * We don't change the tso size until some number of sends
14000 		 * to give the hardware commands time to get down
14001 		 * to the interface.
14002 		 */
14003 		bbr->r_ctl.bbr_hdwr_cnt_noset_snt++;
14004 		if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) {
14005 			bbr->hw_pacing_set = 1;
14006 			tcp_bbr_tso_size_check(bbr, cts);
14007 		}
14008 	}
14009 	bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len);
14010 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14011 		/* Make sure snd_nxt is drug up */
14012 		tp->snd_nxt = tp->snd_max;
14013 	}
14014 	return (error);
14015 
14016 }
14017 
14018 /*
14019  * See bbr_output_wtime() for return values.
14020  */
14021 static int
14022 bbr_output(struct tcpcb *tp)
14023 {
14024 	int32_t ret;
14025 	struct timeval tv;
14026 
14027 	NET_EPOCH_ASSERT();
14028 
14029 	INP_WLOCK_ASSERT(tptoinpcb(tp));
14030 	(void)tcp_get_usecs(&tv);
14031 	ret = bbr_output_wtime(tp, &tv);
14032 	return (ret);
14033 }
14034 
14035 static void
14036 bbr_mtu_chg(struct tcpcb *tp)
14037 {
14038 	struct tcp_bbr *bbr;
14039 	struct bbr_sendmap *rsm, *frsm = NULL;
14040 	uint32_t maxseg;
14041 
14042 	/*
14043 	 * The MTU has changed. a) Clear the sack filter. b) Mark everything
14044 	 * over the current size as SACK_PASS so a retransmit will occur.
14045 	 */
14046 
14047 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14048 	maxseg = tp->t_maxseg - bbr->rc_last_options;
14049 	sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
14050 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
14051 		/* Don't mess with ones acked (by sack?) */
14052 		if (rsm->r_flags & BBR_ACKED)
14053 			continue;
14054 		if ((rsm->r_end - rsm->r_start) > maxseg) {
14055 			/*
14056 			 * We mark sack-passed on all the previous large
14057 			 * sends we did. This will force them to retransmit.
14058 			 */
14059 			rsm->r_flags |= BBR_SACK_PASSED;
14060 			if (((rsm->r_flags & BBR_MARKED_LOST) == 0) &&
14061 			    bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) {
14062 				bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
14063 				bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
14064 				rsm->r_flags |= BBR_MARKED_LOST;
14065 			}
14066 			if (frsm == NULL)
14067 				frsm = rsm;
14068 		}
14069 	}
14070 	if (frsm) {
14071 		bbr->r_ctl.rc_resend = frsm;
14072 	}
14073 }
14074 
14075 static int
14076 bbr_pru_options(struct tcpcb *tp, int flags)
14077 {
14078 	if (flags & PRUS_OOB)
14079 		return (EOPNOTSUPP);
14080 	return (0);
14081 }
14082 
14083 static void
14084 bbr_switch_failed(struct tcpcb *tp)
14085 {
14086 	/*
14087 	 * If a switch fails we only need to
14088 	 * make sure mbuf_queuing is still in place.
14089 	 * We also need to make sure we are still in
14090 	 * ticks granularity (though we should probably
14091 	 * change bbr to go to USECs).
14092 	 *
14093 	 * For timers we need to see if we are still in the
14094 	 * pacer (if our flags are up) if so we are good, if
14095 	 * not we need to get back into the pacer.
14096 	 */
14097 	struct inpcb *inp = tptoinpcb(tp);
14098 	struct timeval tv;
14099 	uint32_t cts;
14100 	uint32_t toval;
14101 	struct tcp_bbr *bbr;
14102 	struct hpts_diag diag;
14103 
14104 	inp->inp_flags2 |= INP_CANNOT_DO_ECN;
14105 	inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14106 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_TICKS);
14107 	if (inp->inp_in_hpts) {
14108 		return;
14109 	}
14110 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14111 	cts = tcp_get_usecs(&tv);
14112 	if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14113 		if (TSTMP_GT(bbr->rc_pacer_started, cts)) {
14114 			toval = bbr->rc_pacer_started - cts;
14115 		} else {
14116 			/* one slot please */
14117 			toval = HPTS_TICKS_PER_SLOT;
14118 		}
14119 	} else if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14120 		if (TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
14121 			toval = bbr->r_ctl.rc_timer_exp - cts;
14122 		} else {
14123 			/* one slot please */
14124 			toval = HPTS_TICKS_PER_SLOT;
14125 		}
14126 	} else
14127 		toval = HPTS_TICKS_PER_SLOT;
14128 	(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(toval),
14129 				   __LINE__, &diag);
14130 	bbr_log_hpts_diag(bbr, cts, &diag);
14131 }
14132 
14133 struct tcp_function_block __tcp_bbr = {
14134 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
14135 	.tfb_tcp_output = bbr_output,
14136 	.tfb_do_queued_segments = ctf_do_queued_segments,
14137 	.tfb_do_segment_nounlock = bbr_do_segment_nounlock,
14138 	.tfb_tcp_do_segment = bbr_do_segment,
14139 	.tfb_tcp_ctloutput = bbr_ctloutput,
14140 	.tfb_tcp_fb_init = bbr_init,
14141 	.tfb_tcp_fb_fini = bbr_fini,
14142 	.tfb_tcp_timer_stop_all = bbr_stopall,
14143 	.tfb_tcp_rexmit_tmr = bbr_remxt_tmr,
14144 	.tfb_tcp_handoff_ok = bbr_handoff_ok,
14145 	.tfb_tcp_mtu_chg = bbr_mtu_chg,
14146 	.tfb_pru_options = bbr_pru_options,
14147 	.tfb_switch_failed = bbr_switch_failed,
14148 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
14149 };
14150 
14151 /*
14152  * bbr_ctloutput() must drop the inpcb lock before performing copyin on
14153  * socket option arguments.  When it re-acquires the lock after the copy, it
14154  * has to revalidate that the connection is still valid for the socket
14155  * option.
14156  */
14157 static int
14158 bbr_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
14159 {
14160 	struct epoch_tracker et;
14161 	struct tcpcb *tp;
14162 	struct tcp_bbr *bbr;
14163 	int32_t error = 0, optval;
14164 
14165 	switch (sopt->sopt_level) {
14166 	case IPPROTO_IPV6:
14167 	case IPPROTO_IP:
14168 		return (tcp_default_ctloutput(inp, sopt));
14169 	}
14170 
14171 	switch (sopt->sopt_name) {
14172 	case TCP_RACK_PACE_MAX_SEG:
14173 	case TCP_RACK_MIN_TO:
14174 	case TCP_RACK_REORD_THRESH:
14175 	case TCP_RACK_REORD_FADE:
14176 	case TCP_RACK_TLP_THRESH:
14177 	case TCP_RACK_PKT_DELAY:
14178 	case TCP_BBR_ALGORITHM:
14179 	case TCP_BBR_TSLIMITS:
14180 	case TCP_BBR_IWINTSO:
14181 	case TCP_BBR_RECFORCE:
14182 	case TCP_BBR_STARTUP_PG:
14183 	case TCP_BBR_DRAIN_PG:
14184 	case TCP_BBR_RWND_IS_APP:
14185 	case TCP_BBR_PROBE_RTT_INT:
14186 	case TCP_BBR_PROBE_RTT_GAIN:
14187 	case TCP_BBR_PROBE_RTT_LEN:
14188 	case TCP_BBR_STARTUP_LOSS_EXIT:
14189 	case TCP_BBR_USEDEL_RATE:
14190 	case TCP_BBR_MIN_RTO:
14191 	case TCP_BBR_MAX_RTO:
14192 	case TCP_BBR_PACE_PER_SEC:
14193 	case TCP_DELACK:
14194 	case TCP_BBR_PACE_DEL_TAR:
14195 	case TCP_BBR_SEND_IWND_IN_TSO:
14196 	case TCP_BBR_EXTRA_STATE:
14197 	case TCP_BBR_UTTER_MAX_TSO:
14198 	case TCP_BBR_MIN_TOPACEOUT:
14199 	case TCP_BBR_FLOOR_MIN_TSO:
14200 	case TCP_BBR_TSTMP_RAISES:
14201 	case TCP_BBR_POLICER_DETECT:
14202 	case TCP_BBR_USE_RACK_CHEAT:
14203 	case TCP_DATA_AFTER_CLOSE:
14204 	case TCP_BBR_HDWR_PACE:
14205 	case TCP_BBR_PACE_SEG_MAX:
14206 	case TCP_BBR_PACE_SEG_MIN:
14207 	case TCP_BBR_PACE_CROSS:
14208 	case TCP_BBR_PACE_OH:
14209 #ifdef NETFLIX_PEAKRATE
14210 	case TCP_MAXPEAKRATE:
14211 #endif
14212 	case TCP_BBR_TMR_PACE_OH:
14213 	case TCP_BBR_RACK_RTT_USE:
14214 	case TCP_BBR_RETRAN_WTSO:
14215 		break;
14216 	default:
14217 		return (tcp_default_ctloutput(inp, sopt));
14218 		break;
14219 	}
14220 	INP_WUNLOCK(inp);
14221 	error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14222 	if (error)
14223 		return (error);
14224 	INP_WLOCK(inp);
14225 	if (inp->inp_flags & INP_DROPPED) {
14226 		INP_WUNLOCK(inp);
14227 		return (ECONNRESET);
14228 	}
14229 	tp = intotcpcb(inp);
14230 	if (tp->t_fb != &__tcp_bbr) {
14231 		INP_WUNLOCK(inp);
14232 		return (ENOPROTOOPT);
14233 	}
14234 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14235 	switch (sopt->sopt_name) {
14236 	case TCP_BBR_PACE_PER_SEC:
14237 		BBR_OPTS_INC(tcp_bbr_pace_per_sec);
14238 		bbr->r_ctl.bbr_hptsi_per_second = optval;
14239 		break;
14240 	case TCP_BBR_PACE_DEL_TAR:
14241 		BBR_OPTS_INC(tcp_bbr_pace_del_tar);
14242 		bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval;
14243 		break;
14244 	case TCP_BBR_PACE_SEG_MAX:
14245 		BBR_OPTS_INC(tcp_bbr_pace_seg_max);
14246 		bbr->r_ctl.bbr_hptsi_segments_max = optval;
14247 		break;
14248 	case TCP_BBR_PACE_SEG_MIN:
14249 		BBR_OPTS_INC(tcp_bbr_pace_seg_min);
14250 		bbr->r_ctl.bbr_hptsi_bytes_min = optval;
14251 		break;
14252 	case TCP_BBR_PACE_CROSS:
14253 		BBR_OPTS_INC(tcp_bbr_pace_cross);
14254 		bbr->r_ctl.bbr_cross_over = optval;
14255 		break;
14256 	case TCP_BBR_ALGORITHM:
14257 		BBR_OPTS_INC(tcp_bbr_algorithm);
14258 		if (optval && (bbr->rc_use_google == 0)) {
14259 			/* Turn on the google mode */
14260 			bbr_google_mode_on(bbr);
14261 			if ((optval > 3) && (optval < 500)) {
14262 				/*
14263 				 * Must be at least greater than .3%
14264 				 * and must be less than 50.0%.
14265 				 */
14266 				bbr->r_ctl.bbr_google_discount = optval;
14267 			}
14268 		} else if ((optval == 0) && (bbr->rc_use_google == 1)) {
14269 			/* Turn off the google mode */
14270 			bbr_google_mode_off(bbr);
14271 		}
14272 		break;
14273 	case TCP_BBR_TSLIMITS:
14274 		BBR_OPTS_INC(tcp_bbr_tslimits);
14275 		if (optval == 1)
14276 			bbr->rc_use_ts_limit = 1;
14277 		else if (optval == 0)
14278 			bbr->rc_use_ts_limit = 0;
14279 		else
14280 			error = EINVAL;
14281 		break;
14282 
14283 	case TCP_BBR_IWINTSO:
14284 		BBR_OPTS_INC(tcp_bbr_iwintso);
14285 		if ((optval >= 0) && (optval < 128)) {
14286 			uint32_t twin;
14287 
14288 			bbr->rc_init_win = optval;
14289 			twin = bbr_initial_cwnd(bbr, tp);
14290 			if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd))
14291 				tp->snd_cwnd = twin;
14292 			else
14293 				error = EBUSY;
14294 		} else
14295 			error = EINVAL;
14296 		break;
14297 	case TCP_BBR_STARTUP_PG:
14298 		BBR_OPTS_INC(tcp_bbr_startup_pg);
14299 		if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) {
14300 			bbr->r_ctl.rc_startup_pg = optval;
14301 			if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
14302 				bbr->r_ctl.rc_bbr_hptsi_gain = optval;
14303 			}
14304 		} else
14305 			error = EINVAL;
14306 		break;
14307 	case TCP_BBR_DRAIN_PG:
14308 		BBR_OPTS_INC(tcp_bbr_drain_pg);
14309 		if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE))
14310 			bbr->r_ctl.rc_drain_pg = optval;
14311 		else
14312 			error = EINVAL;
14313 		break;
14314 	case TCP_BBR_PROBE_RTT_LEN:
14315 		BBR_OPTS_INC(tcp_bbr_probertt_len);
14316 		if (optval <= 1)
14317 			reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND));
14318 		else
14319 			error = EINVAL;
14320 		break;
14321 	case TCP_BBR_PROBE_RTT_GAIN:
14322 		BBR_OPTS_INC(tcp_bbr_probertt_gain);
14323 		if (optval <= BBR_UNIT)
14324 			bbr->r_ctl.bbr_rttprobe_gain_val = optval;
14325 		else
14326 			error = EINVAL;
14327 		break;
14328 	case TCP_BBR_PROBE_RTT_INT:
14329 		BBR_OPTS_INC(tcp_bbr_probe_rtt_int);
14330 		if (optval > 1000)
14331 			bbr->r_ctl.rc_probertt_int = optval;
14332 		else
14333 			error = EINVAL;
14334 		break;
14335 	case TCP_BBR_MIN_TOPACEOUT:
14336 		BBR_OPTS_INC(tcp_bbr_topaceout);
14337 		if (optval == 0) {
14338 			bbr->no_pacing_until = 0;
14339 			bbr->rc_no_pacing = 0;
14340 		} else if (optval <= 0x00ff) {
14341 			bbr->no_pacing_until = optval;
14342 			if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) &&
14343 			    (bbr->rc_bbr_state == BBR_STATE_STARTUP)){
14344 				/* Turn on no pacing */
14345 				bbr->rc_no_pacing = 1;
14346 			}
14347 		} else
14348 			error = EINVAL;
14349 		break;
14350 	case TCP_BBR_STARTUP_LOSS_EXIT:
14351 		BBR_OPTS_INC(tcp_bbr_startup_loss_exit);
14352 		bbr->rc_loss_exit = optval;
14353 		break;
14354 	case TCP_BBR_USEDEL_RATE:
14355 		error = EINVAL;
14356 		break;
14357 	case TCP_BBR_MIN_RTO:
14358 		BBR_OPTS_INC(tcp_bbr_min_rto);
14359 		bbr->r_ctl.rc_min_rto_ms = optval;
14360 		break;
14361 	case TCP_BBR_MAX_RTO:
14362 		BBR_OPTS_INC(tcp_bbr_max_rto);
14363 		bbr->rc_max_rto_sec = optval;
14364 		break;
14365 	case TCP_RACK_MIN_TO:
14366 		/* Minimum time between rack t-o's in ms */
14367 		BBR_OPTS_INC(tcp_rack_min_to);
14368 		bbr->r_ctl.rc_min_to = optval;
14369 		break;
14370 	case TCP_RACK_REORD_THRESH:
14371 		/* RACK reorder threshold (shift amount) */
14372 		BBR_OPTS_INC(tcp_rack_reord_thresh);
14373 		if ((optval > 0) && (optval < 31))
14374 			bbr->r_ctl.rc_reorder_shift = optval;
14375 		else
14376 			error = EINVAL;
14377 		break;
14378 	case TCP_RACK_REORD_FADE:
14379 		/* Does reordering fade after ms time */
14380 		BBR_OPTS_INC(tcp_rack_reord_fade);
14381 		bbr->r_ctl.rc_reorder_fade = optval;
14382 		break;
14383 	case TCP_RACK_TLP_THRESH:
14384 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14385 		BBR_OPTS_INC(tcp_rack_tlp_thresh);
14386 		if (optval)
14387 			bbr->rc_tlp_threshold = optval;
14388 		else
14389 			error = EINVAL;
14390 		break;
14391 	case TCP_BBR_USE_RACK_CHEAT:
14392 		BBR_OPTS_INC(tcp_use_rackcheat);
14393 		if (bbr->rc_use_google) {
14394 			error = EINVAL;
14395 			break;
14396 		}
14397 		BBR_OPTS_INC(tcp_rack_cheat);
14398 		if (optval)
14399 			bbr->bbr_use_rack_cheat = 1;
14400 		else
14401 			bbr->bbr_use_rack_cheat = 0;
14402 		break;
14403 	case TCP_BBR_FLOOR_MIN_TSO:
14404 		BBR_OPTS_INC(tcp_utter_max_tso);
14405 		if ((optval >= 0) && (optval < 40))
14406 			bbr->r_ctl.bbr_hptsi_segments_floor = optval;
14407 		else
14408 			error = EINVAL;
14409 		break;
14410 	case TCP_BBR_UTTER_MAX_TSO:
14411 		BBR_OPTS_INC(tcp_utter_max_tso);
14412 		if ((optval >= 0) && (optval < 0xffff))
14413 			bbr->r_ctl.bbr_utter_max = optval;
14414 		else
14415 			error = EINVAL;
14416 		break;
14417 
14418 	case TCP_BBR_EXTRA_STATE:
14419 		BBR_OPTS_INC(tcp_extra_state);
14420 		if (optval)
14421 			bbr->rc_use_idle_restart = 1;
14422 		else
14423 			bbr->rc_use_idle_restart = 0;
14424 		break;
14425 	case TCP_BBR_SEND_IWND_IN_TSO:
14426 		BBR_OPTS_INC(tcp_iwnd_tso);
14427 		if (optval) {
14428 			bbr->bbr_init_win_cheat = 1;
14429 			if (bbr->rc_past_init_win == 0) {
14430 				uint32_t cts;
14431 				cts = tcp_get_usecs(&bbr->rc_tv);
14432 				tcp_bbr_tso_size_check(bbr, cts);
14433 			}
14434 		} else
14435 			bbr->bbr_init_win_cheat = 0;
14436 		break;
14437 	case TCP_BBR_HDWR_PACE:
14438 		BBR_OPTS_INC(tcp_hdwr_pacing);
14439 		if (optval){
14440 			bbr->bbr_hdw_pace_ena = 1;
14441 			bbr->bbr_attempt_hdwr_pace = 0;
14442 		} else {
14443 			bbr->bbr_hdw_pace_ena = 0;
14444 #ifdef RATELIMIT
14445 			if (bbr->r_ctl.crte != NULL) {
14446 				tcp_rel_pacing_rate(bbr->r_ctl.crte, tp);
14447 				bbr->r_ctl.crte = NULL;
14448 			}
14449 #endif
14450 		}
14451 		break;
14452 
14453 	case TCP_DELACK:
14454 		BBR_OPTS_INC(tcp_delack);
14455 		if (optval < 100) {
14456 			if (optval == 0) /* off */
14457 				tp->t_delayed_ack = 0;
14458 			else if (optval == 1) /* on which is 2 */
14459 				tp->t_delayed_ack = 2;
14460 			else /* higher than 2 and less than 100 */
14461 				tp->t_delayed_ack = optval;
14462 			if (tp->t_flags & TF_DELACK) {
14463 				tp->t_flags &= ~TF_DELACK;
14464 				tp->t_flags |= TF_ACKNOW;
14465 				NET_EPOCH_ENTER(et);
14466 				bbr_output(tp);
14467 				NET_EPOCH_EXIT(et);
14468 			}
14469 		} else
14470 			error = EINVAL;
14471 		break;
14472 	case TCP_RACK_PKT_DELAY:
14473 		/* RACK added ms i.e. rack-rtt + reord + N */
14474 		BBR_OPTS_INC(tcp_rack_pkt_delay);
14475 		bbr->r_ctl.rc_pkt_delay = optval;
14476 		break;
14477 #ifdef NETFLIX_PEAKRATE
14478 	case TCP_MAXPEAKRATE:
14479 		BBR_OPTS_INC(tcp_maxpeak);
14480 		error = tcp_set_maxpeakrate(tp, optval);
14481 		if (!error)
14482 			tp->t_peakrate_thr = tp->t_maxpeakrate;
14483 		break;
14484 #endif
14485 	case TCP_BBR_RETRAN_WTSO:
14486 		BBR_OPTS_INC(tcp_retran_wtso);
14487 		if (optval)
14488 			bbr->rc_resends_use_tso = 1;
14489 		else
14490 			bbr->rc_resends_use_tso = 0;
14491 		break;
14492 	case TCP_DATA_AFTER_CLOSE:
14493 		BBR_OPTS_INC(tcp_data_ac);
14494 		if (optval)
14495 			bbr->rc_allow_data_af_clo = 1;
14496 		else
14497 			bbr->rc_allow_data_af_clo = 0;
14498 		break;
14499 	case TCP_BBR_POLICER_DETECT:
14500 		BBR_OPTS_INC(tcp_policer_det);
14501 		if (bbr->rc_use_google == 0)
14502 			error = EINVAL;
14503 		else if (optval)
14504 			bbr->r_use_policer = 1;
14505 		else
14506 			bbr->r_use_policer = 0;
14507 		break;
14508 
14509 	case TCP_BBR_TSTMP_RAISES:
14510 		BBR_OPTS_INC(tcp_ts_raises);
14511 		if (optval)
14512 			bbr->ts_can_raise = 1;
14513 		else
14514 			bbr->ts_can_raise = 0;
14515 		break;
14516 	case TCP_BBR_TMR_PACE_OH:
14517 		BBR_OPTS_INC(tcp_pacing_oh_tmr);
14518 		if (bbr->rc_use_google) {
14519 			error = EINVAL;
14520 		} else {
14521 			if (optval)
14522 				bbr->r_ctl.rc_incr_tmrs = 1;
14523 			else
14524 				bbr->r_ctl.rc_incr_tmrs = 0;
14525 		}
14526 		break;
14527 	case TCP_BBR_PACE_OH:
14528 		BBR_OPTS_INC(tcp_pacing_oh);
14529 		if (bbr->rc_use_google) {
14530 			error = EINVAL;
14531 		} else {
14532 			if (optval > (BBR_INCL_TCP_OH|
14533 				      BBR_INCL_IP_OH|
14534 				      BBR_INCL_ENET_OH)) {
14535 				error = EINVAL;
14536 				break;
14537 			}
14538 			if (optval & BBR_INCL_TCP_OH)
14539 				bbr->r_ctl.rc_inc_tcp_oh = 1;
14540 			else
14541 				bbr->r_ctl.rc_inc_tcp_oh = 0;
14542 			if (optval & BBR_INCL_IP_OH)
14543 				bbr->r_ctl.rc_inc_ip_oh = 1;
14544 			else
14545 				bbr->r_ctl.rc_inc_ip_oh = 0;
14546 			if (optval & BBR_INCL_ENET_OH)
14547 				bbr->r_ctl.rc_inc_enet_oh = 1;
14548 			else
14549 				bbr->r_ctl.rc_inc_enet_oh = 0;
14550 		}
14551 		break;
14552 	default:
14553 		return (tcp_default_ctloutput(inp, sopt));
14554 		break;
14555 	}
14556 #ifdef NETFLIX_STATS
14557 	tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14558 #endif
14559 	INP_WUNLOCK(inp);
14560 	return (error);
14561 }
14562 
14563 /*
14564  * return 0 on success, error-num on failure
14565  */
14566 static int
14567 bbr_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
14568 {
14569 	struct tcpcb *tp;
14570 	struct tcp_bbr *bbr;
14571 	int32_t error, optval;
14572 
14573 	tp = intotcpcb(inp);
14574 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14575 	if (bbr == NULL) {
14576 		INP_WUNLOCK(inp);
14577 		return (EINVAL);
14578 	}
14579 	/*
14580 	 * Because all our options are either boolean or an int, we can just
14581 	 * pull everything into optval and then unlock and copy. If we ever
14582 	 * add a option that is not a int, then this will have quite an
14583 	 * impact to this routine.
14584 	 */
14585 	switch (sopt->sopt_name) {
14586 	case TCP_BBR_PACE_PER_SEC:
14587 		optval = bbr->r_ctl.bbr_hptsi_per_second;
14588 		break;
14589 	case TCP_BBR_PACE_DEL_TAR:
14590 		optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar;
14591 		break;
14592 	case TCP_BBR_PACE_SEG_MAX:
14593 		optval = bbr->r_ctl.bbr_hptsi_segments_max;
14594 		break;
14595 	case TCP_BBR_MIN_TOPACEOUT:
14596 		optval = bbr->no_pacing_until;
14597 		break;
14598 	case TCP_BBR_PACE_SEG_MIN:
14599 		optval = bbr->r_ctl.bbr_hptsi_bytes_min;
14600 		break;
14601 	case TCP_BBR_PACE_CROSS:
14602 		optval = bbr->r_ctl.bbr_cross_over;
14603 		break;
14604 	case TCP_BBR_ALGORITHM:
14605 		optval = bbr->rc_use_google;
14606 		break;
14607 	case TCP_BBR_TSLIMITS:
14608 		optval = bbr->rc_use_ts_limit;
14609 		break;
14610 	case TCP_BBR_IWINTSO:
14611 		optval = bbr->rc_init_win;
14612 		break;
14613 	case TCP_BBR_STARTUP_PG:
14614 		optval = bbr->r_ctl.rc_startup_pg;
14615 		break;
14616 	case TCP_BBR_DRAIN_PG:
14617 		optval = bbr->r_ctl.rc_drain_pg;
14618 		break;
14619 	case TCP_BBR_PROBE_RTT_INT:
14620 		optval = bbr->r_ctl.rc_probertt_int;
14621 		break;
14622 	case TCP_BBR_PROBE_RTT_LEN:
14623 		optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND);
14624 		break;
14625 	case TCP_BBR_PROBE_RTT_GAIN:
14626 		optval = bbr->r_ctl.bbr_rttprobe_gain_val;
14627 		break;
14628 	case TCP_BBR_STARTUP_LOSS_EXIT:
14629 		optval = bbr->rc_loss_exit;
14630 		break;
14631 	case TCP_BBR_USEDEL_RATE:
14632 		error = EINVAL;
14633 		break;
14634 	case TCP_BBR_MIN_RTO:
14635 		optval = bbr->r_ctl.rc_min_rto_ms;
14636 		break;
14637 	case TCP_BBR_MAX_RTO:
14638 		optval = bbr->rc_max_rto_sec;
14639 		break;
14640 	case TCP_RACK_PACE_MAX_SEG:
14641 		/* Max segments in a pace */
14642 		optval = bbr->r_ctl.rc_pace_max_segs;
14643 		break;
14644 	case TCP_RACK_MIN_TO:
14645 		/* Minimum time between rack t-o's in ms */
14646 		optval = bbr->r_ctl.rc_min_to;
14647 		break;
14648 	case TCP_RACK_REORD_THRESH:
14649 		/* RACK reorder threshold (shift amount) */
14650 		optval = bbr->r_ctl.rc_reorder_shift;
14651 		break;
14652 	case TCP_RACK_REORD_FADE:
14653 		/* Does reordering fade after ms time */
14654 		optval = bbr->r_ctl.rc_reorder_fade;
14655 		break;
14656 	case TCP_BBR_USE_RACK_CHEAT:
14657 		/* Do we use the rack cheat for rxt */
14658 		optval = bbr->bbr_use_rack_cheat;
14659 		break;
14660 	case TCP_BBR_FLOOR_MIN_TSO:
14661 		optval = bbr->r_ctl.bbr_hptsi_segments_floor;
14662 		break;
14663 	case TCP_BBR_UTTER_MAX_TSO:
14664 		optval = bbr->r_ctl.bbr_utter_max;
14665 		break;
14666 	case TCP_BBR_SEND_IWND_IN_TSO:
14667 		/* Do we send TSO size segments initially */
14668 		optval = bbr->bbr_init_win_cheat;
14669 		break;
14670 	case TCP_BBR_EXTRA_STATE:
14671 		optval = bbr->rc_use_idle_restart;
14672 		break;
14673 	case TCP_RACK_TLP_THRESH:
14674 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14675 		optval = bbr->rc_tlp_threshold;
14676 		break;
14677 	case TCP_RACK_PKT_DELAY:
14678 		/* RACK added ms i.e. rack-rtt + reord + N */
14679 		optval = bbr->r_ctl.rc_pkt_delay;
14680 		break;
14681 	case TCP_BBR_RETRAN_WTSO:
14682 		optval = bbr->rc_resends_use_tso;
14683 		break;
14684 	case TCP_DATA_AFTER_CLOSE:
14685 		optval = bbr->rc_allow_data_af_clo;
14686 		break;
14687 	case TCP_DELACK:
14688 		optval = tp->t_delayed_ack;
14689 		break;
14690 	case TCP_BBR_HDWR_PACE:
14691 		optval = bbr->bbr_hdw_pace_ena;
14692 		break;
14693 	case TCP_BBR_POLICER_DETECT:
14694 		optval = bbr->r_use_policer;
14695 		break;
14696 	case TCP_BBR_TSTMP_RAISES:
14697 		optval = bbr->ts_can_raise;
14698 		break;
14699 	case TCP_BBR_TMR_PACE_OH:
14700 		optval = bbr->r_ctl.rc_incr_tmrs;
14701 		break;
14702 	case TCP_BBR_PACE_OH:
14703 		optval = 0;
14704 		if (bbr->r_ctl.rc_inc_tcp_oh)
14705 			optval |= BBR_INCL_TCP_OH;
14706 		if (bbr->r_ctl.rc_inc_ip_oh)
14707 			optval |= BBR_INCL_IP_OH;
14708 		if (bbr->r_ctl.rc_inc_enet_oh)
14709 			optval |= BBR_INCL_ENET_OH;
14710 		break;
14711 	default:
14712 		return (tcp_default_ctloutput(inp, sopt));
14713 		break;
14714 	}
14715 	INP_WUNLOCK(inp);
14716 	error = sooptcopyout(sopt, &optval, sizeof optval);
14717 	return (error);
14718 }
14719 
14720 /*
14721  * return 0 on success, error-num on failure
14722  */
14723 static int
14724 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt)
14725 {
14726 	if (sopt->sopt_dir == SOPT_SET) {
14727 		return (bbr_set_sockopt(inp, sopt));
14728 	} else if (sopt->sopt_dir == SOPT_GET) {
14729 		return (bbr_get_sockopt(inp, sopt));
14730 	} else {
14731 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
14732 	}
14733 }
14734 
14735 static const char *bbr_stack_names[] = {
14736 	__XSTRING(STACKNAME),
14737 #ifdef STACKALIAS
14738 	__XSTRING(STACKALIAS),
14739 #endif
14740 };
14741 
14742 static bool bbr_mod_inited = false;
14743 
14744 static int
14745 tcp_addbbr(module_t mod, int32_t type, void *data)
14746 {
14747 	int32_t err = 0;
14748 	int num_stacks;
14749 
14750 	switch (type) {
14751 	case MOD_LOAD:
14752 		printf("Attempting to load " __XSTRING(MODNAME) "\n");
14753 		bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
14754 		    sizeof(struct bbr_sendmap),
14755 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
14756 		bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
14757 		    sizeof(struct tcp_bbr),
14758 		    NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
14759 		sysctl_ctx_init(&bbr_sysctl_ctx);
14760 		bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
14761 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
14762 		    OID_AUTO,
14763 #ifdef STACKALIAS
14764 		    __XSTRING(STACKALIAS),
14765 #else
14766 		    __XSTRING(STACKNAME),
14767 #endif
14768 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
14769 		    "");
14770 		if (bbr_sysctl_root == NULL) {
14771 			printf("Failed to add sysctl node\n");
14772 			err = EFAULT;
14773 			goto free_uma;
14774 		}
14775 		bbr_init_sysctls();
14776 		num_stacks = nitems(bbr_stack_names);
14777 		err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK,
14778 		    bbr_stack_names, &num_stacks);
14779 		if (err) {
14780 			printf("Failed to register %s stack name for "
14781 			    "%s module\n", bbr_stack_names[num_stacks],
14782 			    __XSTRING(MODNAME));
14783 			sysctl_ctx_free(&bbr_sysctl_ctx);
14784 	free_uma:
14785 			uma_zdestroy(bbr_zone);
14786 			uma_zdestroy(bbr_pcb_zone);
14787 			bbr_counter_destroy();
14788 			printf("Failed to register " __XSTRING(MODNAME)
14789 			    " module err:%d\n", err);
14790 			return (err);
14791 		}
14792 		tcp_lro_reg_mbufq();
14793 		bbr_mod_inited = true;
14794 		printf(__XSTRING(MODNAME) " is now available\n");
14795 		break;
14796 	case MOD_QUIESCE:
14797 		err = deregister_tcp_functions(&__tcp_bbr, true, false);
14798 		break;
14799 	case MOD_UNLOAD:
14800 		err = deregister_tcp_functions(&__tcp_bbr, false, true);
14801 		if (err == EBUSY)
14802 			break;
14803 		if (bbr_mod_inited) {
14804 			uma_zdestroy(bbr_zone);
14805 			uma_zdestroy(bbr_pcb_zone);
14806 			sysctl_ctx_free(&bbr_sysctl_ctx);
14807 			bbr_counter_destroy();
14808 			printf(__XSTRING(MODNAME)
14809 			    " is now no longer available\n");
14810 			bbr_mod_inited = false;
14811 		}
14812 		tcp_lro_dereg_mbufq();
14813 		err = 0;
14814 		break;
14815 	default:
14816 		return (EOPNOTSUPP);
14817 	}
14818 	return (err);
14819 }
14820 
14821 static moduledata_t tcp_bbr = {
14822 	.name = __XSTRING(MODNAME),
14823 	    .evhand = tcp_addbbr,
14824 	    .priv = 0
14825 };
14826 
14827 MODULE_VERSION(MODNAME, 1);
14828 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
14829 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
14830