xref: /freebsd/sys/netinet/tcp_stacks/bbr.c (revision 120e232f1ae386f0ce413b27d60b3d52c568c58e)
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 #include "opt_inet.h"
35 #include "opt_inet6.h"
36 #include "opt_ipsec.h"
37 #include "opt_ratelimit.h"
38 #include <sys/param.h>
39 #include <sys/arb.h>
40 #include <sys/module.h>
41 #include <sys/kernel.h>
42 #include <sys/libkern.h>
43 #ifdef TCP_HHOOK
44 #include <sys/hhook.h>
45 #endif
46 #include <sys/malloc.h>
47 #include <sys/mbuf.h>
48 #include <sys/proc.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #ifdef STATS
54 #include <sys/qmath.h>
55 #include <sys/tree.h>
56 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
57 #endif
58 #include <sys/refcount.h>
59 #include <sys/queue.h>
60 #include <sys/eventhandler.h>
61 #include <sys/smp.h>
62 #include <sys/kthread.h>
63 #include <sys/lock.h>
64 #include <sys/mutex.h>
65 #include <sys/tim_filter.h>
66 #include <sys/time.h>
67 #include <sys/protosw.h>
68 #include <vm/uma.h>
69 #include <sys/kern_prefetch.h>
70 
71 #include <net/route.h>
72 #include <net/route/nhop.h>
73 #include <net/vnet.h>
74 
75 #define TCPSTATES		/* for logging */
76 
77 #include <netinet/in.h>
78 #include <netinet/in_kdtrace.h>
79 #include <netinet/in_pcb.h>
80 #include <netinet/ip.h>
81 #include <netinet/ip_var.h>
82 #include <netinet/ip6.h>
83 #include <netinet6/in6_pcb.h>
84 #include <netinet6/ip6_var.h>
85 #define	TCPOUTFLAGS
86 #include <netinet/tcp.h>
87 #include <netinet/tcp_fsm.h>
88 #include <netinet/tcp_seq.h>
89 #include <netinet/tcp_timer.h>
90 #include <netinet/tcp_var.h>
91 #include <netinet/tcpip.h>
92 #include <netinet/tcp_hpts.h>
93 #include <netinet/cc/cc.h>
94 #include <netinet/tcp_log_buf.h>
95 #include <netinet/tcp_ratelimit.h>
96 #include <netinet/tcp_lro.h>
97 #ifdef TCP_OFFLOAD
98 #include <netinet/tcp_offload.h>
99 #endif
100 #ifdef INET6
101 #include <netinet6/tcp6_var.h>
102 #endif
103 #include <netinet/tcp_fastopen.h>
104 
105 #include <netipsec/ipsec_support.h>
106 #include <net/if.h>
107 #include <net/if_var.h>
108 #include <net/ethernet.h>
109 
110 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
111 #include <netipsec/ipsec.h>
112 #include <netipsec/ipsec6.h>
113 #endif				/* IPSEC */
114 
115 #include <netinet/udp.h>
116 #include <netinet/udp_var.h>
117 #include <machine/in_cksum.h>
118 
119 #ifdef MAC
120 #include <security/mac/mac_framework.h>
121 #endif
122 
123 #include "sack_filter.h"
124 #include "tcp_bbr.h"
125 #include "rack_bbr_common.h"
126 uma_zone_t bbr_zone;
127 uma_zone_t bbr_pcb_zone;
128 
129 struct sysctl_ctx_list bbr_sysctl_ctx;
130 struct sysctl_oid *bbr_sysctl_root;
131 
132 #define	TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \
133 	(tv) = (value); \
134 	if ((u_long)(tv) < (u_long)(tvmin)) \
135 		(tv) = (tvmin); \
136 	if ((u_long)(tv) > (u_long)(tvmax)) \
137 		(tv) = (tvmax); \
138 } while(0)
139 
140 /*#define BBR_INVARIANT 1*/
141 
142 /*
143  * initial window
144  */
145 static uint32_t bbr_def_init_win = 10;
146 static int32_t bbr_persist_min = 250000;	/* 250ms */
147 static int32_t bbr_persist_max = 1000000;	/* 1 Second */
148 static int32_t bbr_cwnd_may_shrink = 0;
149 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP;
150 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT;
151 static int32_t bbr_hardware_pacing_limit = 8000;
152 static int32_t bbr_quanta = 3;	/* How much extra quanta do we get? */
153 static int32_t bbr_no_retran = 0;
154 
155 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */
156 static int32_t bbr_max_net_error_cnt = 10;
157 /* Should the following be dynamic too -- loss wise */
158 static int32_t bbr_rtt_gain_thresh = 0;
159 /* Measurement controls */
160 static int32_t bbr_use_google_algo = 1;
161 static int32_t bbr_ts_limiting = 1;
162 static int32_t bbr_ts_can_raise = 0;
163 static int32_t bbr_do_red = 600;
164 static int32_t bbr_red_scale = 20000;
165 static int32_t bbr_red_mul = 1;
166 static int32_t bbr_red_div = 2;
167 static int32_t bbr_red_growth_restrict = 1;
168 static int32_t  bbr_target_is_bbunit = 0;
169 static int32_t bbr_drop_limit = 0;
170 /*
171  * How much gain do we need to see to
172  * stay in startup?
173  */
174 static int32_t bbr_marks_rxt_sack_passed = 0;
175 static int32_t bbr_start_exit = 25;
176 static int32_t bbr_low_start_exit = 25;	/* When we are in reduced gain */
177 static int32_t bbr_startup_loss_thresh = 2000;	/* 20.00% loss */
178 static int32_t bbr_hptsi_max_mul = 1;	/* These two mul/div assure a min pacing */
179 static int32_t bbr_hptsi_max_div = 2;	/* time, 0 means turned off. We need this
180 					 * if we go back ever to where the pacer
181 					 * has priority over timers.
182 					 */
183 static int32_t bbr_policer_call_from_rack_to = 0;
184 static int32_t bbr_policer_detection_enabled = 1;
185 static int32_t bbr_min_measurements_req = 1;	/* We need at least 2
186 						 * measurements before we are
187 						 * "good" note that 2 == 1.
188 						 * This is because we use a >
189 						 * comparison. This means if
190 						 * min_measure was 0, it takes
191 						 * num-measures > min(0) and
192 						 * you get 1 measurement and
193 						 * you are good. Set to 1, you
194 						 * have to have two
195 						 * measurements (this is done
196 						 * to prevent it from being ok
197 						 * to have no measurements). */
198 static int32_t bbr_no_pacing_until = 4;
199 
200 static int32_t bbr_min_usec_delta = 20000;	/* 20,000 usecs */
201 static int32_t bbr_min_peer_delta = 20;		/* 20 units */
202 static int32_t bbr_delta_percent = 150;		/* 15.0 % */
203 
204 static int32_t bbr_target_cwnd_mult_limit = 8;
205 /*
206  * bbr_cwnd_min_val is the number of
207  * segments we hold to in the RTT probe
208  * state typically 4.
209  */
210 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS;
211 
212 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS;
213 
214 static int32_t bbr_gain_to_target = 1;
215 static int32_t bbr_gain_gets_extra_too = 1;
216 /*
217  * bbr_high_gain is the 2/ln(2) value we need
218  * to double the sending rate in startup. This
219  * is used for both cwnd and hptsi gain's.
220  */
221 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1;
222 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1;
223 static int32_t bbr_use_lower_gain_in_startup = 1;
224 
225 /* thresholds for reduction on drain in sub-states/drain */
226 static int32_t bbr_drain_rtt = BBR_SRTT;
227 static int32_t bbr_drain_floor = 88;
228 static int32_t google_allow_early_out = 1;
229 static int32_t google_consider_lost = 1;
230 static int32_t bbr_drain_drop_mul = 4;
231 static int32_t bbr_drain_drop_div = 5;
232 static int32_t bbr_rand_ot = 50;
233 static int32_t bbr_can_force_probertt = 0;
234 static int32_t bbr_can_adjust_probertt = 1;
235 static int32_t bbr_probertt_sets_rtt = 0;
236 static int32_t bbr_can_use_ts_for_rtt = 1;
237 static int32_t bbr_is_ratio = 0;
238 static int32_t bbr_sub_drain_app_limit = 1;
239 static int32_t bbr_prtt_slam_cwnd = 1;
240 static int32_t bbr_sub_drain_slam_cwnd = 1;
241 static int32_t bbr_slam_cwnd_in_main_drain = 1;
242 static int32_t bbr_filter_len_sec = 6;	/* How long does the rttProp filter
243 					 * hold */
244 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4);
245 /*
246  * bbr_drain_gain is the reverse of the high_gain
247  * designed to drain back out the standing queue
248  * that is formed in startup by causing a larger
249  * hptsi gain and thus drainging the packets
250  * in flight.
251  */
252 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885;
253 static int32_t bbr_rttprobe_gain = 192;
254 
255 /*
256  * The cwnd_gain is the default cwnd gain applied when
257  * calculating a target cwnd. Note that the cwnd is
258  * a secondary factor in the way BBR works (see the
259  * paper and think about it, it will take some time).
260  * Basically the hptsi_gain spreads the packets out
261  * so you never get more than BDP to the peer even
262  * if the cwnd is high. In our implemenation that
263  * means in non-recovery/retransmission scenarios
264  * cwnd will never be reached by the flight-size.
265  */
266 static int32_t bbr_cwnd_gain = BBR_UNIT * 2;
267 static int32_t bbr_tlp_type_to_use = BBR_SRTT;
268 static int32_t bbr_delack_time = 100000;	/* 100ms in useconds */
269 static int32_t bbr_sack_not_required = 0;	/* set to one to allow non-sack to use bbr */
270 static int32_t bbr_initial_bw_bps = 62500;	/* 500kbps in bytes ps */
271 static int32_t bbr_ignore_data_after_close = 1;
272 static int16_t bbr_hptsi_gain[] = {
273 	(BBR_UNIT *5 / 4),
274 	(BBR_UNIT * 3 / 4),
275 	BBR_UNIT,
276 	BBR_UNIT,
277 	BBR_UNIT,
278 	BBR_UNIT,
279 	BBR_UNIT,
280 	BBR_UNIT
281 };
282 int32_t bbr_use_rack_resend_cheat = 1;
283 int32_t bbr_sends_full_iwnd = 1;
284 
285 #define BBR_HPTSI_GAIN_MAX 8
286 /*
287  * The BBR module incorporates a number of
288  * TCP ideas that have been put out into the IETF
289  * over the last few years:
290  * - Yuchung Cheng's RACK TCP (for which its named) that
291  *    will stop us using the number of dup acks and instead
292  *    use time as the gage of when we retransmit.
293  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
294  *    of Dukkipati et.al.
295  * - Van Jacobson's et.al BBR.
296  *
297  * RACK depends on SACK, so if an endpoint arrives that
298  * cannot do SACK the state machine below will shuttle the
299  * connection back to using the "default" TCP stack that is
300  * in FreeBSD.
301  *
302  * To implement BBR and RACK the original TCP stack was first decomposed
303  * into a functional state machine with individual states
304  * for each of the possible TCP connection states. The do_segment
305  * functions role in life is to mandate the connection supports SACK
306  * initially and then assure that the RACK state matches the conenction
307  * state before calling the states do_segment function. Data processing
308  * of inbound segments also now happens in the hpts_do_segment in general
309  * with only one exception. This is so we can keep the connection on
310  * a single CPU.
311  *
312  * Each state is simplified due to the fact that the original do_segment
313  * has been decomposed and we *know* what state we are in (no
314  * switches on the state) and all tests for SACK are gone. This
315  * greatly simplifies what each state does.
316  *
317  * TCP output is also over-written with a new version since it
318  * must maintain the new rack scoreboard and has had hptsi
319  * integrated as a requirment. Still todo is to eliminate the
320  * use of the callout_() system and use the hpts for all
321  * timers as well.
322  */
323 static uint32_t bbr_rtt_probe_time = 200000;	/* 200ms in micro seconds */
324 static uint32_t bbr_rtt_probe_cwndtarg = 4;	/* How many mss's outstanding */
325 static const int32_t bbr_min_req_free = 2;	/* The min we must have on the
326 						 * free list */
327 static int32_t bbr_tlp_thresh = 1;
328 static int32_t bbr_reorder_thresh = 2;
329 static int32_t bbr_reorder_fade = 60000000;	/* 0 - never fade, def
330 						 * 60,000,000 - 60 seconds */
331 static int32_t bbr_pkt_delay = 1000;
332 static int32_t bbr_min_to = 1000;	/* Number of usec's minimum timeout */
333 static int32_t bbr_incr_timers = 1;
334 
335 static int32_t bbr_tlp_min = 10000;	/* 10ms in usecs */
336 static int32_t bbr_delayed_ack_time = 200000;	/* 200ms in usecs */
337 static int32_t bbr_exit_startup_at_loss = 1;
338 
339 /*
340  * bbr_lt_bw_ratio is 1/8th
341  * bbr_lt_bw_diff is  < 4 Kbit/sec
342  */
343 static uint64_t bbr_lt_bw_diff = 4000 / 8;	/* In bytes per second */
344 static uint64_t bbr_lt_bw_ratio = 8;	/* For 1/8th */
345 static uint32_t bbr_lt_bw_max_rtts = 48;	/* How many rtt's do we use
346 						 * the lt_bw for */
347 static uint32_t bbr_lt_intvl_min_rtts = 4;	/* Min num of RTT's to measure
348 						 * lt_bw */
349 static int32_t bbr_lt_intvl_fp = 0;		/* False positive epoch diff */
350 static int32_t bbr_lt_loss_thresh = 196;	/* Lost vs delivered % */
351 static int32_t bbr_lt_fd_thresh = 100;		/* false detection % */
352 
353 static int32_t bbr_verbose_logging = 0;
354 /*
355  * Currently regular tcp has a rto_min of 30ms
356  * the backoff goes 12 times so that ends up
357  * being a total of 122.850 seconds before a
358  * connection is killed.
359  */
360 static int32_t bbr_rto_min_ms = 30;	/* 30ms same as main freebsd */
361 static int32_t bbr_rto_max_sec = 4;	/* 4 seconds */
362 
363 /****************************************************/
364 /* DEFAULT TSO SIZING  (cpu performance impacting)  */
365 /****************************************************/
366 /* What amount is our formula using to get TSO size */
367 static int32_t bbr_hptsi_per_second = 1000;
368 
369 /*
370  * For hptsi under bbr_cross_over connections what is delay
371  * target 7ms (in usec) combined with a seg_max of 2
372  * gets us close to identical google behavior in
373  * TSO size selection (possibly more 1MSS sends).
374  */
375 static int32_t bbr_hptsi_segments_delay_tar = 7000;
376 
377 /* Does pacing delay include overhead's in its time calculations? */
378 static int32_t bbr_include_enet_oh = 0;
379 static int32_t bbr_include_ip_oh = 1;
380 static int32_t bbr_include_tcp_oh = 1;
381 static int32_t bbr_google_discount = 10;
382 
383 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */
384 static int32_t bbr_state_is_pkt_epoch = 0;
385 static int32_t bbr_state_drain_2_tar = 1;
386 /* What is the max the 0 - bbr_cross_over MBPS TSO target
387  * can reach using our delay target. Note that this
388  * value becomes the floor for the cross over
389  * algorithm.
390  */
391 static int32_t bbr_hptsi_segments_max = 2;
392 static int32_t bbr_hptsi_segments_floor = 1;
393 static int32_t bbr_hptsi_utter_max = 0;
394 
395 /* What is the min the 0 - bbr_cross-over MBPS  TSO target can be */
396 static int32_t bbr_hptsi_bytes_min = 1460;
397 static int32_t bbr_all_get_min = 0;
398 
399 /* Cross over point from algo-a to algo-b */
400 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS;
401 
402 /* Do we deal with our restart state? */
403 static int32_t bbr_uses_idle_restart = 0;
404 static int32_t bbr_idle_restart_threshold = 100000;	/* 100ms in useconds */
405 
406 /* Do we allow hardware pacing? */
407 static int32_t bbr_allow_hdwr_pacing = 0;
408 static int32_t bbr_hdwr_pace_adjust = 2;	/* multipler when we calc the tso size */
409 static int32_t bbr_hdwr_pace_floor = 1;
410 static int32_t bbr_hdwr_pacing_delay_cnt = 10;
411 
412 /****************************************************/
413 static int32_t bbr_resends_use_tso = 0;
414 static int32_t bbr_tlp_max_resend = 2;
415 static int32_t bbr_sack_block_limit = 128;
416 
417 #define  BBR_MAX_STAT 19
418 counter_u64_t bbr_state_time[BBR_MAX_STAT];
419 counter_u64_t bbr_state_lost[BBR_MAX_STAT];
420 counter_u64_t bbr_state_resend[BBR_MAX_STAT];
421 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE];
422 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE];
423 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE];
424 counter_u64_t bbr_flows_whdwr_pacing;
425 counter_u64_t bbr_flows_nohdwr_pacing;
426 
427 counter_u64_t bbr_nohdwr_pacing_enobuf;
428 counter_u64_t bbr_hdwr_pacing_enobuf;
429 
430 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr);
431 
432 /*
433  * Static defintions we need for forward declarations.
434  */
435 static uint32_t
436 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain,
437 		      uint32_t useconds_time, uint64_t bw);
438 static uint32_t
439 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain);
440 static void
441 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win);
442 static void
443 bbr_set_probebw_gains(struct tcp_bbr *bbr,  uint32_t cts, uint32_t losses);
444 static void
445 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line,
446 		    int dolog);
447 static uint32_t
448 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain);
449 static void
450 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch,
451 		 int32_t pkt_epoch, uint32_t losses);
452 static uint32_t
453 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts,
454 		     struct bbr_sendmap *rsm);
455 static uint32_t
456 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp);
457 static uint32_t
458 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
459 		    struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts);
460 static void
461 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts,
462 		 int32_t line);
463 static void
464 bbr_set_state_target(struct tcp_bbr *bbr, int line);
465 static void
466 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line);
467 static void
468 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick,
469 		       int event, int line);
470 static void
471 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts);
472 static void
473 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts);
474 static void
475 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
476 		    uint32_t rtt, uint32_t line, uint8_t is_start,
477 		    uint16_t set);
478 static struct bbr_sendmap *
479 bbr_find_lowest_rsm(struct tcp_bbr *bbr);
480 static __inline uint32_t
481 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type);
482 static void
483 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot,
484 		 uint8_t which);
485 static void
486 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts,
487 		  uint32_t time_since_sent, uint32_t srtt,
488 		  uint32_t thresh, uint32_t to);
489 static void
490 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag);
491 static void
492 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot,
493 		    uint32_t del_by, uint32_t cts, uint32_t sloton,
494 		    uint32_t prev_delay);
495 static void
496 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts,
497 		  int32_t line);
498 static void
499 bbr_stop_all_timers(struct tcpcb *tp, struct tcp_bbr *bbr);
500 static void
501 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts);
502 static void
503 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts);
504 static void
505 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts);
506 static void
507 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
508 			  uint32_t cts, uint32_t usecs, uint64_t bw,
509 			  uint32_t override, int mod);
510 static int bbr_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
511 
512 static inline uint8_t
bbr_state_val(struct tcp_bbr * bbr)513 bbr_state_val(struct tcp_bbr *bbr)
514 {
515 	return(bbr->rc_bbr_substate);
516 }
517 
518 static inline uint32_t
get_min_cwnd(struct tcp_bbr * bbr)519 get_min_cwnd(struct tcp_bbr *bbr)
520 {
521 	int mss;
522 
523 	mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options),
524 		  bbr->r_ctl.rc_pace_max_segs);
525 	if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED)
526 		return (bbr_cwnd_min_val_hs * mss);
527 	else
528 		return (bbr_cwnd_min_val * mss);
529 }
530 
531 static uint32_t
bbr_get_persists_timer_val(struct tcpcb * tp,struct tcp_bbr * bbr)532 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr)
533 {
534 	uint64_t srtt, var;
535 	uint64_t ret_val;
536 
537 	bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
538 	if (tp->t_srtt == 0) {
539 		srtt = (uint64_t)BBR_INITIAL_RTO;
540 		var = 0;
541 	} else {
542 		srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
543 		var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT);
544 	}
545 	TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]),
546 	    bbr_persist_min, bbr_persist_max);
547 	return ((uint32_t)ret_val);
548 }
549 
550 static uint32_t
bbr_timer_start(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts)551 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
552 {
553 	/*
554 	 * Start the FR timer, we do this based on getting the first one in
555 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
556 	 * events we need to stop the running timer (if its running) before
557 	 * starting the new one.
558 	 */
559 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
560 	int32_t idx;
561 	int32_t is_tlp_timer = 0;
562 	struct bbr_sendmap *rsm;
563 
564 	if (bbr->rc_all_timers_stopped) {
565 		/* All timers have been stopped none are to run */
566 		return (0);
567 	}
568 	if (bbr->rc_in_persist) {
569 		/* We can't start any timer in persists */
570 		return (bbr_get_persists_timer_val(tp, bbr));
571 	}
572 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
573 	if ((rsm == NULL) ||
574 	    ((tp->t_flags & TF_SACK_PERMIT) == 0) ||
575 	    (tp->t_state < TCPS_ESTABLISHED)) {
576 		/* Nothing on the send map */
577 activate_rxt:
578 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
579 		    sbavail(&tptosocket(tp)->so_snd)) {
580 			uint64_t tov;
581 
582 			time_since_sent = 0;
583 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
584 			if (rsm) {
585 				idx = rsm->r_rtr_cnt - 1;
586 				if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
587 					tstmp_touse = rsm->r_tim_lastsent[idx];
588 				else
589 					tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
590 				if (TSTMP_GT(tstmp_touse, cts))
591 				    time_since_sent = cts - tstmp_touse;
592 			}
593 			bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
594 			if (tp->t_srtt == 0)
595 				tov = BBR_INITIAL_RTO;
596 			else
597 				tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) +
598 				    ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT);
599 			if (tp->t_rxtshift)
600 				tov *= tcp_backoff[tp->t_rxtshift];
601 			if (tov > time_since_sent)
602 				tov -= time_since_sent;
603 			else
604 				tov = bbr->r_ctl.rc_min_to;
605 			TCPT_RANGESET_NOSLOP(to, tov,
606 			    (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC),
607 			    (bbr->rc_max_rto_sec * USECS_IN_SECOND));
608 			bbr_log_timer_var(bbr, 2, cts, 0, bbr_get_rtt(bbr, BBR_SRTT), 0, to);
609 			return (to);
610 		}
611 		return (0);
612 	}
613 	if (rsm->r_flags & BBR_ACKED) {
614 		rsm = bbr_find_lowest_rsm(bbr);
615 		if (rsm == NULL) {
616 			/* No lowest? */
617 			goto activate_rxt;
618 		}
619 	}
620 	/* Convert from ms to usecs */
621 	if (rsm->r_flags & BBR_SACK_PASSED) {
622 		if ((tp->t_flags & TF_SENTFIN) &&
623 		    ((tp->snd_max - tp->snd_una) == 1) &&
624 		    (rsm->r_flags & BBR_HAS_FIN)) {
625 			/*
626 			 * We don't start a bbr rack timer if all we have is
627 			 * a FIN outstanding.
628 			 */
629 			goto activate_rxt;
630 		}
631 		srtt = bbr_get_rtt(bbr, BBR_RTT_RACK);
632 		thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm);
633 		idx = rsm->r_rtr_cnt - 1;
634 		exp = rsm->r_tim_lastsent[idx] + thresh;
635 		if (SEQ_GEQ(exp, cts)) {
636 			to = exp - cts;
637 			if (to < bbr->r_ctl.rc_min_to) {
638 				to = bbr->r_ctl.rc_min_to;
639 			}
640 		} else {
641 			to = bbr->r_ctl.rc_min_to;
642 		}
643 	} else {
644 		/* Ok we need to do a TLP not RACK */
645 		if (bbr->rc_tlp_in_progress != 0) {
646 			/*
647 			 * The previous send was a TLP.
648 			 */
649 			goto activate_rxt;
650 		}
651 		rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
652 		if (rsm == NULL) {
653 			/* We found no rsm to TLP with. */
654 			goto activate_rxt;
655 		}
656 		if (rsm->r_flags & BBR_HAS_FIN) {
657 			/* If its a FIN we don't do TLP */
658 			rsm = NULL;
659 			goto activate_rxt;
660 		}
661 		time_since_sent = 0;
662 		idx = rsm->r_rtr_cnt - 1;
663 		if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
664 			tstmp_touse = rsm->r_tim_lastsent[idx];
665 		else
666 			tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
667 		if (TSTMP_GT(tstmp_touse, cts))
668 		    time_since_sent = cts - tstmp_touse;
669 		is_tlp_timer = 1;
670 		srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use);
671 		thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts);
672 		if (thresh > time_since_sent)
673 			to = thresh - time_since_sent;
674 		else
675 			to = bbr->r_ctl.rc_min_to;
676 		if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
677 			/*
678 			 * If the TLP time works out to larger than the max
679 			 * RTO lets not do TLP.. just RTO.
680 			 */
681 			goto activate_rxt;
682 		}
683 		if ((bbr->rc_tlp_rtx_out == 1) &&
684 		    (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) {
685 			/*
686 			 * Second retransmit of the same TLP
687 			 * lets not.
688 			 */
689 			bbr->rc_tlp_rtx_out = 0;
690 			goto activate_rxt;
691 		}
692 		if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) {
693 			/*
694 			 * The tail is no longer the last one I did a probe
695 			 * on
696 			 */
697 			bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
698 			bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
699 		}
700 	}
701 	if (is_tlp_timer == 0) {
702 		BBR_STAT_INC(bbr_to_arm_rack);
703 		bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
704 	} else {
705 		bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to);
706 		if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
707 			/*
708 			 * We have exceeded how many times we can retran the
709 			 * current TLP timer, switch to the RTO timer.
710 			 */
711 			goto activate_rxt;
712 		} else {
713 			BBR_STAT_INC(bbr_to_arm_tlp);
714 			bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
715 		}
716 	}
717 	return (to);
718 }
719 
720 static inline int32_t
bbr_minseg(struct tcp_bbr * bbr)721 bbr_minseg(struct tcp_bbr *bbr)
722 {
723 	return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options);
724 }
725 
726 static void
bbr_start_hpts_timer(struct tcp_bbr * bbr,struct tcpcb * tp,uint32_t cts,int32_t frm,int32_t slot,uint32_t tot_len)727 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len)
728 {
729 	struct inpcb *inp = tptoinpcb(tp);
730 	struct hpts_diag diag;
731 	uint32_t delayed_ack = 0;
732 	uint32_t left = 0;
733 	uint32_t hpts_timeout;
734 	uint8_t stopped;
735 	int32_t delay_calc = 0;
736 	uint32_t prev_delay = 0;
737 
738 	if (tcp_in_hpts(tp)) {
739 		/* A previous call is already set up */
740 		return;
741 	}
742 	if ((tp->t_state == TCPS_CLOSED) ||
743 	    (tp->t_state == TCPS_LISTEN)) {
744 		return;
745 	}
746 	stopped = bbr->rc_tmr_stopped;
747 	if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
748 		left = bbr->r_ctl.rc_timer_exp - cts;
749 	}
750 	bbr->r_ctl.rc_hpts_flags = 0;
751 	bbr->r_ctl.rc_timer_exp = 0;
752 	prev_delay = bbr->r_ctl.rc_last_delay_val;
753 	if (bbr->r_ctl.rc_last_delay_val &&
754 	    (slot == 0)) {
755 		/*
756 		 * If a previous pacer delay was in place we
757 		 * are not coming from the output side (where
758 		 * we calculate a delay, more likely a timer).
759 		 */
760 		slot = bbr->r_ctl.rc_last_delay_val;
761 		if (TSTMP_GT(cts, bbr->rc_pacer_started)) {
762 			/* Compensate for time passed  */
763 			delay_calc = cts - bbr->rc_pacer_started;
764 			if (delay_calc <= slot)
765 				slot -= delay_calc;
766 		}
767 	}
768 	/* Do we have early to make up for by pushing out the pacing time? */
769 	if (bbr->r_agg_early_set) {
770 		bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2);
771 		slot += bbr->r_ctl.rc_agg_early;
772 		bbr->r_ctl.rc_agg_early = 0;
773 		bbr->r_agg_early_set = 0;
774 	}
775 	/* Are we running a total debt that needs to be compensated for? */
776 	if (bbr->r_ctl.rc_hptsi_agg_delay) {
777 		if (slot > bbr->r_ctl.rc_hptsi_agg_delay) {
778 			/* We nuke the delay */
779 			slot -= bbr->r_ctl.rc_hptsi_agg_delay;
780 			bbr->r_ctl.rc_hptsi_agg_delay = 0;
781 		} else {
782 			/* We nuke some of the delay, put in a minimal 100usecs  */
783 			bbr->r_ctl.rc_hptsi_agg_delay -= slot;
784 			bbr->r_ctl.rc_last_delay_val = slot = 100;
785 		}
786 	}
787 	bbr->r_ctl.rc_last_delay_val = slot;
788 	hpts_timeout = bbr_timer_start(tp, bbr, cts);
789 	if (tp->t_flags & TF_DELACK) {
790 		if (bbr->rc_in_persist == 0) {
791 			delayed_ack = bbr_delack_time;
792 		} else {
793 			/*
794 			 * We are in persists and have
795 			 * gotten a new data element.
796 			 */
797 			if (hpts_timeout > bbr_delack_time) {
798 				/*
799 				 * Lets make the persists timer (which acks)
800 				 * be the smaller of hpts_timeout and bbr_delack_time.
801 				 */
802 				hpts_timeout = bbr_delack_time;
803 			}
804 		}
805 	}
806 	if (delayed_ack &&
807 	    ((hpts_timeout == 0) ||
808 	     (delayed_ack < hpts_timeout))) {
809 		/* We need a Delayed ack timer */
810 		bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
811 		hpts_timeout = delayed_ack;
812 	}
813 	if (slot) {
814 		/* Mark that we have a pacing timer up */
815 		BBR_STAT_INC(bbr_paced_segments);
816 		bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
817 	}
818 	/*
819 	 * If no timers are going to run and we will fall off thfe hptsi
820 	 * wheel, we resort to a keep-alive timer if its configured.
821 	 */
822 	if ((hpts_timeout == 0) &&
823 	    (slot == 0)) {
824 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
825 		    (tp->t_state <= TCPS_CLOSING)) {
826 			/*
827 			 * Ok we have no timer (persists, rack, tlp, rxt  or
828 			 * del-ack), we don't have segments being paced. So
829 			 * all that is left is the keepalive timer.
830 			 */
831 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
832 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
833 			} else {
834 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
835 			}
836 			bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
837 		}
838 	}
839 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
840 	    (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
841 		/*
842 		 * RACK, TLP, persists and RXT timers all are restartable
843 		 * based on actions input .. i.e we received a packet (ack
844 		 * or sack) and that changes things (rw, or snd_una etc).
845 		 * Thus we can restart them with a new value. For
846 		 * keep-alive, delayed_ack we keep track of what was left
847 		 * and restart the timer with a smaller value.
848 		 */
849 		if (left < hpts_timeout)
850 			hpts_timeout = left;
851 	}
852 	if (bbr->r_ctl.rc_incr_tmrs && slot &&
853 	    (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
854 		/*
855 		 * If configured to do so, and the timer is either
856 		 * the TLP or RXT timer, we need to increase the timeout
857 		 * by the pacing time. Consider the bottleneck at my
858 		 * machine as an example, we are sending something
859 		 * to start a TLP on. The last packet won't be emitted
860 		 * fully until the pacing time (the bottleneck will hold
861 		 * the data in place). Once the packet is emitted that
862 		 * is when we want to start waiting for the TLP. This
863 		 * is most evident with hardware pacing (where the nic
864 		 * is holding the packet(s) before emitting). But it
865 		 * can also show up in the network so we do it for all
866 		 * cases. Technically we would take off one packet from
867 		 * this extra delay but this is easier and being more
868 		 * conservative is probably better.
869 		 */
870 		hpts_timeout += slot;
871 	}
872 	if (hpts_timeout) {
873 		/*
874 		 * Hack alert for now we can't time-out over 2147 seconds (a
875 		 * bit more than 35min)
876 		 */
877 		if (hpts_timeout > 0x7ffffffe)
878 			hpts_timeout = 0x7ffffffe;
879 		bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
880 	} else
881 		bbr->r_ctl.rc_timer_exp = 0;
882 	if ((slot) &&
883 	    (bbr->rc_use_google ||
884 	     bbr->output_error_seen ||
885 	     (slot <= hpts_timeout))  ) {
886 		/*
887 		 * Tell LRO that it can queue packets while
888 		 * we pace.
889 		 */
890 		bbr->rc_tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
891 		if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
892 		    (bbr->rc_cwnd_limited == 0)) {
893 			/*
894 			 * If we are not cwnd limited and we
895 			 * are running a rack timer we put on
896 			 * the do not disturbe even for sack.
897 			 */
898 			tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
899 		} else
900 			tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
901 		bbr->rc_pacer_started = cts;
902 
903 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
904 					   __LINE__, &diag);
905 		bbr->rc_timer_first = 0;
906 		bbr->bbr_timer_src = frm;
907 		bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1);
908 		bbr_log_hpts_diag(bbr, cts, &diag);
909 	} else if (hpts_timeout) {
910 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
911 					   __LINE__, &diag);
912 		/*
913 		 * We add the flag here as well if the slot is set,
914 		 * since hpts will call in to clear the queue first before
915 		 * calling the output routine (which does our timers).
916 		 * We don't want to set the flag if its just a timer
917 		 * else the arrival of data might (that causes us
918 		 * to send more) might get delayed. Imagine being
919 		 * on a keep-alive timer and a request comes in for
920 		 * more data.
921 		 */
922 		if (slot)
923 			bbr->rc_pacer_started = cts;
924 		if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
925 		    (bbr->rc_cwnd_limited == 0)) {
926 			/*
927 			 * For a rack timer, don't wake us even
928 			 * if a sack arrives as long as we are
929 			 * not cwnd limited.
930 			 */
931 			tp->t_flags2 |= (TF2_MBUF_QUEUE_READY |
932 			    TF2_DONT_SACK_QUEUE);
933 		} else {
934 			/* All other timers wake us up */
935 			tp->t_flags2 &= ~(TF2_MBUF_QUEUE_READY |
936 			    TF2_DONT_SACK_QUEUE);
937 		}
938 		bbr->bbr_timer_src = frm;
939 		bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0);
940 		bbr_log_hpts_diag(bbr, cts, &diag);
941 		bbr->rc_timer_first = 1;
942 	}
943 	bbr->rc_tmr_stopped = 0;
944 	bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay);
945 }
946 
947 static void
bbr_timer_audit(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts,struct sockbuf * sb)948 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb)
949 {
950 	/*
951 	 * We received an ack, and then did not call send or were bounced
952 	 * out due to the hpts was running. Now a timer is up as well, is it
953 	 * the right timer?
954 	 */
955 	struct inpcb *inp;
956 	struct bbr_sendmap *rsm;
957 	uint32_t hpts_timeout;
958 	int tmr_up;
959 
960 	tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
961 	if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
962 		return;
963 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
964 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
965 	    (tmr_up == PACE_TMR_RXT)) {
966 		/* Should be an RXT */
967 		return;
968 	}
969 	inp = bbr->rc_inp;
970 	if (rsm == NULL) {
971 		/* Nothing outstanding? */
972 		if (tp->t_flags & TF_DELACK) {
973 			if (tmr_up == PACE_TMR_DELACK)
974 				/*
975 				 * We are supposed to have delayed ack up
976 				 * and we do
977 				 */
978 				return;
979 		} else if (((V_tcp_always_keepalive ||
980 			    inp->inp_socket->so_options & SO_KEEPALIVE) &&
981 			    (tp->t_state <= TCPS_CLOSING)) &&
982 			    (tmr_up == PACE_TMR_KEEP) &&
983 		    (tp->snd_max == tp->snd_una)) {
984 			/* We should have keep alive up and we do */
985 			return;
986 		}
987 	}
988 	if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) {
989 		if ((tp->t_flags & TF_SENTFIN) &&
990 		    ((tp->snd_max - tp->snd_una) == 1) &&
991 		    (rsm->r_flags & BBR_HAS_FIN)) {
992 			/* needs to be a RXT */
993 			if (tmr_up == PACE_TMR_RXT)
994 				return;
995 			else
996 				goto wrong_timer;
997 		} else if (tmr_up == PACE_TMR_RACK)
998 			return;
999 		else
1000 			goto wrong_timer;
1001 	} else if (rsm && (tmr_up == PACE_TMR_RACK)) {
1002 		/* Rack timer has priority if we have data out */
1003 		return;
1004 	} else if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1005 		    ((tmr_up == PACE_TMR_TLP) ||
1006 	    (tmr_up == PACE_TMR_RXT))) {
1007 		/*
1008 		 * Either a TLP or RXT is fine if no sack-passed is in place
1009 		 * and data is outstanding.
1010 		 */
1011 		return;
1012 	} else if (tmr_up == PACE_TMR_DELACK) {
1013 		/*
1014 		 * If the delayed ack was going to go off before the
1015 		 * rtx/tlp/rack timer were going to expire, then that would
1016 		 * be the timer in control. Note we don't check the time
1017 		 * here trusting the code is correct.
1018 		 */
1019 		return;
1020 	}
1021 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1022 	    ((tmr_up == PACE_TMR_RXT) ||
1023 	     (tmr_up == PACE_TMR_TLP) ||
1024 	     (tmr_up == PACE_TMR_RACK))) {
1025 		/*
1026 		 * We have outstanding data and
1027 		 * we *do* have a RACK, TLP or RXT
1028 		 * timer running. We won't restart
1029 		 * anything here since thats probably ok we
1030 		 * will get called with some timer here shortly.
1031 		 */
1032 		return;
1033 	}
1034 	/*
1035 	 * Ok the timer originally started is not what we want now. We will
1036 	 * force the hpts to be stopped if any, and restart with the slot
1037 	 * set to what was in the saved slot.
1038 	 */
1039 wrong_timer:
1040 	if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) {
1041 		if (tcp_in_hpts(tp))
1042 			tcp_hpts_remove(tp);
1043 		bbr_timer_cancel(bbr, __LINE__, cts);
1044 		bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val,
1045 		    0);
1046 	} else {
1047 		/*
1048 		 * Output is hptsi so we just need to switch the type of
1049 		 * timer. We don't bother with keep-alive, since when we
1050 		 * jump through the output, it will start the keep-alive if
1051 		 * nothing is sent.
1052 		 *
1053 		 * We only need a delayed-ack added and or the hpts_timeout.
1054 		 */
1055 		hpts_timeout = bbr_timer_start(tp, bbr, cts);
1056 		if (tp->t_flags & TF_DELACK) {
1057 			if (hpts_timeout == 0) {
1058 				hpts_timeout = bbr_delack_time;
1059 				bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1060 			}
1061 			else if (hpts_timeout > bbr_delack_time) {
1062 				hpts_timeout = bbr_delack_time;
1063 				bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1064 			}
1065 		}
1066 		if (hpts_timeout) {
1067 			if (hpts_timeout > 0x7ffffffe)
1068 				hpts_timeout = 0x7ffffffe;
1069 			bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
1070 		}
1071 	}
1072 }
1073 
1074 int32_t bbr_clear_lost = 0;
1075 
1076 /*
1077  * Considers the two time values now (cts) and earlier.
1078  * If cts is smaller than earlier, we could have
1079  * had a sequence wrap (our counter wraps every
1080  * 70 min or so) or it could be just clock skew
1081  * getting us two different time values. Clock skew
1082  * will show up within 10ms or so. So in such
1083  * a case (where cts is behind earlier time by
1084  * less than 10ms) we return 0. Otherwise we
1085  * return the true difference between them.
1086  */
1087 static inline uint32_t
bbr_calc_time(uint32_t cts,uint32_t earlier_time)1088 bbr_calc_time(uint32_t cts, uint32_t earlier_time) {
1089 	/*
1090 	 * Given two timestamps, the current time stamp cts, and some other
1091 	 * time-stamp taken in theory earlier return the difference. The
1092 	 * trick is here sometimes locking will get the other timestamp
1093 	 * after the cts. If this occurs we need to return 0.
1094 	 */
1095 	if (TSTMP_GEQ(cts, earlier_time))
1096 		return (cts - earlier_time);
1097 	/*
1098 	 * cts is behind earlier_time if its less than 10ms consider it 0.
1099 	 * If its more than 10ms difference then we had a time wrap. Else
1100 	 * its just the normal locking foo. I wonder if we should not go to
1101 	 * 64bit TS and get rid of this issue.
1102 	 */
1103 	if (TSTMP_GEQ((cts + 10000), earlier_time))
1104 		return (0);
1105 	/*
1106 	 * Ok the time must have wrapped. So we need to answer a large
1107 	 * amount of time, which the normal subtraction should do.
1108 	 */
1109 	return (cts - earlier_time);
1110 }
1111 
1112 static int
sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS)1113 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS)
1114 {
1115 	uint32_t stat;
1116 	int32_t error;
1117 
1118 	error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t));
1119 	if (error || req->newptr == NULL)
1120 		return error;
1121 
1122 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
1123 	if (error)
1124 		return (error);
1125 	if (stat == 1) {
1126 #ifdef BBR_INVARIANTS
1127 		printf("Clearing BBR lost counters\n");
1128 #endif
1129 		COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT);
1130 		COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT);
1131 		COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT);
1132 	} else if (stat == 2) {
1133 #ifdef BBR_INVARIANTS
1134 		printf("Clearing BBR option counters\n");
1135 #endif
1136 		COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE);
1137 	} else if (stat == 3) {
1138 #ifdef BBR_INVARIANTS
1139 		printf("Clearing BBR stats counters\n");
1140 #endif
1141 		COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE);
1142 	} else if (stat == 4) {
1143 #ifdef BBR_INVARIANTS
1144 		printf("Clearing BBR out-size counters\n");
1145 #endif
1146 		COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE);
1147 	}
1148 	bbr_clear_lost = 0;
1149 	return (0);
1150 }
1151 
1152 static void
bbr_init_sysctls(void)1153 bbr_init_sysctls(void)
1154 {
1155 	struct sysctl_oid *bbr_probertt;
1156 	struct sysctl_oid *bbr_hptsi;
1157 	struct sysctl_oid *bbr_measure;
1158 	struct sysctl_oid *bbr_cwnd;
1159 	struct sysctl_oid *bbr_timeout;
1160 	struct sysctl_oid *bbr_states;
1161 	struct sysctl_oid *bbr_startup;
1162 	struct sysctl_oid *bbr_policer;
1163 
1164 	/* Probe rtt controls */
1165 	bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1166 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1167 	    OID_AUTO,
1168 	    "probertt",
1169 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1170 	    "");
1171 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1172 	    SYSCTL_CHILDREN(bbr_probertt),
1173 	    OID_AUTO, "gain", CTLFLAG_RW,
1174 	    &bbr_rttprobe_gain, 192,
1175 	    "What is the filter gain drop in probe_rtt (0=disable)?");
1176 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1177 	    SYSCTL_CHILDREN(bbr_probertt),
1178 	    OID_AUTO, "cwnd", CTLFLAG_RW,
1179 	    &bbr_rtt_probe_cwndtarg, 4,
1180 	    "How many mss's are outstanding during probe-rtt");
1181 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1182 	    SYSCTL_CHILDREN(bbr_probertt),
1183 	    OID_AUTO, "int", CTLFLAG_RW,
1184 	    &bbr_rtt_probe_limit, 4000000,
1185 	    "If RTT has not shrank in this many micro-seconds enter probe-rtt");
1186 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1187 	    SYSCTL_CHILDREN(bbr_probertt),
1188 	    OID_AUTO, "mintime", CTLFLAG_RW,
1189 	    &bbr_rtt_probe_time, 200000,
1190 	    "How many microseconds in probe-rtt");
1191 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1192 	    SYSCTL_CHILDREN(bbr_probertt),
1193 	    OID_AUTO, "filter_len_sec", CTLFLAG_RW,
1194 	    &bbr_filter_len_sec, 6,
1195 	    "How long in seconds does the rttProp filter run?");
1196 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1197 	    SYSCTL_CHILDREN(bbr_probertt),
1198 	    OID_AUTO, "drain_rtt", CTLFLAG_RW,
1199 	    &bbr_drain_rtt, BBR_SRTT,
1200 	    "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?");
1201 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1202 	    SYSCTL_CHILDREN(bbr_probertt),
1203 	    OID_AUTO, "can_force", CTLFLAG_RW,
1204 	    &bbr_can_force_probertt, 0,
1205 	    "If we keep setting new low rtt's but delay going in probe-rtt can we force in??");
1206 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1207 	    SYSCTL_CHILDREN(bbr_probertt),
1208 	    OID_AUTO, "enter_sets_force", CTLFLAG_RW,
1209 	    &bbr_probertt_sets_rtt, 0,
1210 	    "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?");
1211 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1212 	    SYSCTL_CHILDREN(bbr_probertt),
1213 	    OID_AUTO, "can_adjust", CTLFLAG_RW,
1214 	    &bbr_can_adjust_probertt, 1,
1215 	    "Can we dynamically adjust the probe-rtt limits and times?");
1216 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1217 	    SYSCTL_CHILDREN(bbr_probertt),
1218 	    OID_AUTO, "is_ratio", CTLFLAG_RW,
1219 	    &bbr_is_ratio, 0,
1220 	    "is the limit to filter a ratio?");
1221 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1222 	    SYSCTL_CHILDREN(bbr_probertt),
1223 	    OID_AUTO, "use_cwnd", CTLFLAG_RW,
1224 	    &bbr_prtt_slam_cwnd, 0,
1225 	    "Should we set/recover cwnd?");
1226 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1227 	    SYSCTL_CHILDREN(bbr_probertt),
1228 	    OID_AUTO, "can_use_ts", CTLFLAG_RW,
1229 	    &bbr_can_use_ts_for_rtt, 1,
1230 	    "Can we use the ms timestamp if available for retransmistted rtt calculations?");
1231 
1232 	/* Pacing controls */
1233 	bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1234 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1235 	    OID_AUTO,
1236 	    "pacing",
1237 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1238 	    "");
1239 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1240 	    SYSCTL_CHILDREN(bbr_hptsi),
1241 	    OID_AUTO, "hw_pacing", CTLFLAG_RW,
1242 	    &bbr_allow_hdwr_pacing, 1,
1243 	    "Do we allow hardware pacing?");
1244 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1245 	    SYSCTL_CHILDREN(bbr_hptsi),
1246 	    OID_AUTO, "hw_pacing_limit", CTLFLAG_RW,
1247 	    &bbr_hardware_pacing_limit, 4000,
1248 	    "Do we have a limited number of connections for pacing chelsio (0=no limit)?");
1249 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1250 	    SYSCTL_CHILDREN(bbr_hptsi),
1251 	    OID_AUTO, "hw_pacing_adj", CTLFLAG_RW,
1252 	    &bbr_hdwr_pace_adjust, 2,
1253 	    "Multiplier to calculated tso size?");
1254 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1255 	    SYSCTL_CHILDREN(bbr_hptsi),
1256 	    OID_AUTO, "hw_pacing_floor", CTLFLAG_RW,
1257 	    &bbr_hdwr_pace_floor, 1,
1258 	    "Do we invoke the hardware pacing floor?");
1259 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1260 	    SYSCTL_CHILDREN(bbr_hptsi),
1261 	    OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW,
1262 	    &bbr_hdwr_pacing_delay_cnt, 10,
1263 	    "How many packets must be sent after hdwr pacing is enabled");
1264 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1265 	    SYSCTL_CHILDREN(bbr_hptsi),
1266 	    OID_AUTO, "bw_cross", CTLFLAG_RW,
1267 	    &bbr_cross_over, 3000000,
1268 	    "What is the point where we cross over to linux like TSO size set");
1269 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1270 	    SYSCTL_CHILDREN(bbr_hptsi),
1271 	    OID_AUTO, "seg_deltarg", CTLFLAG_RW,
1272 	    &bbr_hptsi_segments_delay_tar, 7000,
1273 	    "What is the worse case delay target for hptsi < 48Mbp connections");
1274 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1275 	    SYSCTL_CHILDREN(bbr_hptsi),
1276 	    OID_AUTO, "enet_oh", CTLFLAG_RW,
1277 	    &bbr_include_enet_oh, 0,
1278 	    "Do we include the ethernet overhead in calculating pacing delay?");
1279 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1280 	    SYSCTL_CHILDREN(bbr_hptsi),
1281 	    OID_AUTO, "ip_oh", CTLFLAG_RW,
1282 	    &bbr_include_ip_oh, 1,
1283 	    "Do we include the IP overhead in calculating pacing delay?");
1284 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1285 	    SYSCTL_CHILDREN(bbr_hptsi),
1286 	    OID_AUTO, "tcp_oh", CTLFLAG_RW,
1287 	    &bbr_include_tcp_oh, 0,
1288 	    "Do we include the TCP overhead in calculating pacing delay?");
1289 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1290 	    SYSCTL_CHILDREN(bbr_hptsi),
1291 	    OID_AUTO, "google_discount", CTLFLAG_RW,
1292 	    &bbr_google_discount, 10,
1293 	    "What is the default google discount percentage wise for pacing (11 = 1.1%%)?");
1294 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1295 	    SYSCTL_CHILDREN(bbr_hptsi),
1296 	    OID_AUTO, "all_get_min", CTLFLAG_RW,
1297 	    &bbr_all_get_min, 0,
1298 	    "If you are less than a MSS do you just get the min?");
1299 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1300 	    SYSCTL_CHILDREN(bbr_hptsi),
1301 	    OID_AUTO, "tso_min", CTLFLAG_RW,
1302 	    &bbr_hptsi_bytes_min, 1460,
1303 	    "For 0 -> 24Mbps what is floor number of segments for TSO");
1304 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1305 	    SYSCTL_CHILDREN(bbr_hptsi),
1306 	    OID_AUTO, "seg_tso_max", CTLFLAG_RW,
1307 	    &bbr_hptsi_segments_max, 6,
1308 	    "For 0 -> 24Mbps what is top number of segments for TSO");
1309 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1310 	    SYSCTL_CHILDREN(bbr_hptsi),
1311 	    OID_AUTO, "seg_floor", CTLFLAG_RW,
1312 	    &bbr_hptsi_segments_floor, 1,
1313 	    "Minimum TSO size we will fall too in segments");
1314 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1315 	    SYSCTL_CHILDREN(bbr_hptsi),
1316 	    OID_AUTO, "utter_max", CTLFLAG_RW,
1317 	    &bbr_hptsi_utter_max, 0,
1318 	    "The absolute maximum that any pacing (outside of hardware) can be");
1319 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1320 	    SYSCTL_CHILDREN(bbr_hptsi),
1321 	    OID_AUTO, "seg_divisor", CTLFLAG_RW,
1322 	    &bbr_hptsi_per_second, 100,
1323 	    "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps ");
1324 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1325 	    SYSCTL_CHILDREN(bbr_hptsi),
1326 	    OID_AUTO, "srtt_mul", CTLFLAG_RW,
1327 	    &bbr_hptsi_max_mul, 1,
1328 	    "The multiplier for pace len max");
1329 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1330 	    SYSCTL_CHILDREN(bbr_hptsi),
1331 	    OID_AUTO, "srtt_div", CTLFLAG_RW,
1332 	    &bbr_hptsi_max_div, 2,
1333 	    "The divisor for pace len max");
1334 	/* Measurement controls */
1335 	bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1336 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1337 	    OID_AUTO,
1338 	    "measure",
1339 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1340 	    "Measurement controls");
1341 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1342 	    SYSCTL_CHILDREN(bbr_measure),
1343 	    OID_AUTO, "min_i_bw", CTLFLAG_RW,
1344 	    &bbr_initial_bw_bps, 62500,
1345 	    "Minimum initial b/w in bytes per second");
1346 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1347 	    SYSCTL_CHILDREN(bbr_measure),
1348 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1349 	    &bbr_sack_not_required, 0,
1350 	    "Do we allow bbr to run on connections not supporting SACK?");
1351 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1352 	    SYSCTL_CHILDREN(bbr_measure),
1353 	    OID_AUTO, "use_google", CTLFLAG_RW,
1354 	    &bbr_use_google_algo, 0,
1355 	    "Use has close to google V1.0 has possible?");
1356 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1357 	    SYSCTL_CHILDREN(bbr_measure),
1358 	    OID_AUTO, "ts_limiting", CTLFLAG_RW,
1359 	    &bbr_ts_limiting, 1,
1360 	    "Do we attempt to use the peers timestamp to limit b/w caculations?");
1361 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1362 	    SYSCTL_CHILDREN(bbr_measure),
1363 	    OID_AUTO, "ts_can_raise", CTLFLAG_RW,
1364 	    &bbr_ts_can_raise, 0,
1365 	    "Can we raise the b/w via timestamp b/w calculation?");
1366 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1367 	    SYSCTL_CHILDREN(bbr_measure),
1368 	    OID_AUTO, "ts_delta", CTLFLAG_RW,
1369 	    &bbr_min_usec_delta, 20000,
1370 	    "How long in usec between ts of our sends in ts validation code?");
1371 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1372 	    SYSCTL_CHILDREN(bbr_measure),
1373 	    OID_AUTO, "ts_peer_delta", CTLFLAG_RW,
1374 	    &bbr_min_peer_delta, 20,
1375 	    "What min numerical value should be between the peer deltas?");
1376 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1377 	    SYSCTL_CHILDREN(bbr_measure),
1378 	    OID_AUTO, "ts_delta_percent", CTLFLAG_RW,
1379 	    &bbr_delta_percent, 150,
1380 	    "What percentage (150 = 15.0) do we allow variance for?");
1381 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1382 	    SYSCTL_CHILDREN(bbr_measure),
1383 	    OID_AUTO, "min_measure_good_bw", CTLFLAG_RW,
1384 	    &bbr_min_measurements_req, 1,
1385 	    "What is the minimum measurement count we need before we switch to our b/w estimate");
1386 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1387 	    SYSCTL_CHILDREN(bbr_measure),
1388 	    OID_AUTO, "min_measure_before_pace", CTLFLAG_RW,
1389 	    &bbr_no_pacing_until, 4,
1390 	    "How many pkt-epoch's (0 is off) do we need before pacing is on?");
1391 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1392 	    SYSCTL_CHILDREN(bbr_measure),
1393 	    OID_AUTO, "quanta", CTLFLAG_RW,
1394 	    &bbr_quanta, 2,
1395 	    "Extra quanta to add when calculating the target (ID section 4.2.3.2).");
1396 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1397 	    SYSCTL_CHILDREN(bbr_measure),
1398 	    OID_AUTO, "noretran", CTLFLAG_RW,
1399 	    &bbr_no_retran, 0,
1400 	    "Should google mode not use retransmission measurements for the b/w estimation?");
1401 	/* State controls */
1402 	bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1403 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1404 	    OID_AUTO,
1405 	    "states",
1406 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1407 	    "State controls");
1408 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1409 	    SYSCTL_CHILDREN(bbr_states),
1410 	    OID_AUTO, "idle_restart", CTLFLAG_RW,
1411 	    &bbr_uses_idle_restart, 0,
1412 	    "Do we use a new special idle_restart state to ramp back up quickly?");
1413 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1414 	    SYSCTL_CHILDREN(bbr_states),
1415 	    OID_AUTO, "idle_restart_threshold", CTLFLAG_RW,
1416 	    &bbr_idle_restart_threshold, 100000,
1417 	    "How long must we be idle before we restart??");
1418 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1419 	    SYSCTL_CHILDREN(bbr_states),
1420 	    OID_AUTO, "use_pkt_epoch", CTLFLAG_RW,
1421 	    &bbr_state_is_pkt_epoch, 0,
1422 	    "Do we use a pkt-epoch for substate if 0 rttProp?");
1423 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1424 	    SYSCTL_CHILDREN(bbr_states),
1425 	    OID_AUTO, "startup_rtt_gain", CTLFLAG_RW,
1426 	    &bbr_rtt_gain_thresh, 0,
1427 	    "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?");
1428 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1429 	    SYSCTL_CHILDREN(bbr_states),
1430 	    OID_AUTO, "drain_floor", CTLFLAG_RW,
1431 	    &bbr_drain_floor, 88,
1432 	    "What is the lowest we can drain (pg) too?");
1433 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1434 	    SYSCTL_CHILDREN(bbr_states),
1435 	    OID_AUTO, "drain_2_target", CTLFLAG_RW,
1436 	    &bbr_state_drain_2_tar, 1,
1437 	    "Do we drain to target in drain substate?");
1438 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1439 	    SYSCTL_CHILDREN(bbr_states),
1440 	    OID_AUTO, "gain_2_target", CTLFLAG_RW,
1441 	    &bbr_gain_to_target, 1,
1442 	    "Does probe bw gain to target??");
1443 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1444 	    SYSCTL_CHILDREN(bbr_states),
1445 	    OID_AUTO, "gain_extra_time", CTLFLAG_RW,
1446 	    &bbr_gain_gets_extra_too, 1,
1447 	    "Does probe bw gain get the extra time too?");
1448 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1449 	    SYSCTL_CHILDREN(bbr_states),
1450 	    OID_AUTO, "ld_div", CTLFLAG_RW,
1451 	    &bbr_drain_drop_div, 5,
1452 	    "Long drain drop divider?");
1453 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1454 	    SYSCTL_CHILDREN(bbr_states),
1455 	    OID_AUTO, "ld_mul", CTLFLAG_RW,
1456 	    &bbr_drain_drop_mul, 4,
1457 	    "Long drain drop multiplier?");
1458 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1459 	    SYSCTL_CHILDREN(bbr_states),
1460 	    OID_AUTO, "rand_ot_disc", CTLFLAG_RW,
1461 	    &bbr_rand_ot, 50,
1462 	    "Random discount of the ot?");
1463 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1464 	    SYSCTL_CHILDREN(bbr_states),
1465 	    OID_AUTO, "dr_filter_life", CTLFLAG_RW,
1466 	    &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT,
1467 	    "How many packet-epochs does the b/w delivery rate last?");
1468 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1469 	    SYSCTL_CHILDREN(bbr_states),
1470 	    OID_AUTO, "subdrain_applimited", CTLFLAG_RW,
1471 	    &bbr_sub_drain_app_limit, 0,
1472 	    "Does our sub-state drain invoke app limited if its long?");
1473 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1474 	    SYSCTL_CHILDREN(bbr_states),
1475 	    OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW,
1476 	    &bbr_sub_drain_slam_cwnd, 0,
1477 	    "Should we set/recover cwnd for sub-state drain?");
1478 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1479 	    SYSCTL_CHILDREN(bbr_states),
1480 	    OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW,
1481 	    &bbr_slam_cwnd_in_main_drain, 0,
1482 	    "Should we set/recover cwnd for main-state drain?");
1483 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1484 	    SYSCTL_CHILDREN(bbr_states),
1485 	    OID_AUTO, "google_gets_earlyout", CTLFLAG_RW,
1486 	    &google_allow_early_out, 1,
1487 	    "Should we allow google probe-bw/drain to exit early at flight target?");
1488 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1489 	    SYSCTL_CHILDREN(bbr_states),
1490 	    OID_AUTO, "google_exit_loss", CTLFLAG_RW,
1491 	    &google_consider_lost, 1,
1492 	    "Should we have losses exit gain of probebw in google mode??");
1493 	/* Startup controls */
1494 	bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1495 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1496 	    OID_AUTO,
1497 	    "startup",
1498 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1499 	    "Startup controls");
1500 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1501 	    SYSCTL_CHILDREN(bbr_startup),
1502 	    OID_AUTO, "cheat_iwnd", CTLFLAG_RW,
1503 	    &bbr_sends_full_iwnd, 1,
1504 	    "Do we not pace but burst out initial windows has our TSO size?");
1505 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1506 	    SYSCTL_CHILDREN(bbr_startup),
1507 	    OID_AUTO, "loss_threshold", CTLFLAG_RW,
1508 	    &bbr_startup_loss_thresh, 2000,
1509 	    "In startup what is the loss threshold in a pe that will exit us from startup?");
1510 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1511 	    SYSCTL_CHILDREN(bbr_startup),
1512 	    OID_AUTO, "use_lowerpg", CTLFLAG_RW,
1513 	    &bbr_use_lower_gain_in_startup, 1,
1514 	    "Should we use a lower hptsi gain if we see loss in startup?");
1515 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1516 	    SYSCTL_CHILDREN(bbr_startup),
1517 	    OID_AUTO, "gain", CTLFLAG_RW,
1518 	    &bbr_start_exit, 25,
1519 	    "What gain percent do we need to see to stay in startup??");
1520 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1521 	    SYSCTL_CHILDREN(bbr_startup),
1522 	    OID_AUTO, "low_gain", CTLFLAG_RW,
1523 	    &bbr_low_start_exit, 15,
1524 	    "What gain percent do we need to see to stay in the lower gain startup??");
1525 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1526 	    SYSCTL_CHILDREN(bbr_startup),
1527 	    OID_AUTO, "loss_exit", CTLFLAG_RW,
1528 	    &bbr_exit_startup_at_loss, 1,
1529 	    "Should we exit startup at loss in an epoch if we are not gaining?");
1530 	/* CWND controls */
1531 	bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1532 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1533 	    OID_AUTO,
1534 	    "cwnd",
1535 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1536 	    "Cwnd controls");
1537 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1538 	    SYSCTL_CHILDREN(bbr_cwnd),
1539 	    OID_AUTO, "tar_rtt", CTLFLAG_RW,
1540 	    &bbr_cwndtarget_rtt_touse, 0,
1541 	    "Target cwnd rtt measurement to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?");
1542 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1543 	    SYSCTL_CHILDREN(bbr_cwnd),
1544 	    OID_AUTO, "may_shrink", CTLFLAG_RW,
1545 	    &bbr_cwnd_may_shrink, 0,
1546 	    "Can the cwnd shrink if it would grow to more than the target?");
1547 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1548 	    SYSCTL_CHILDREN(bbr_cwnd),
1549 	    OID_AUTO, "max_target_limit", CTLFLAG_RW,
1550 	    &bbr_target_cwnd_mult_limit, 8,
1551 	    "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?");
1552 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1553 	    SYSCTL_CHILDREN(bbr_cwnd),
1554 	    OID_AUTO, "highspeed_min", CTLFLAG_RW,
1555 	    &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS,
1556 	    "What is the high-speed min cwnd (rttProp under 1ms)");
1557 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1558 	    SYSCTL_CHILDREN(bbr_cwnd),
1559 	    OID_AUTO, "lowspeed_min", CTLFLAG_RW,
1560 	    &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS,
1561 	    "What is the min cwnd (rttProp > 1ms)");
1562 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1563 	    SYSCTL_CHILDREN(bbr_cwnd),
1564 	    OID_AUTO, "initwin", CTLFLAG_RW,
1565 	    &bbr_def_init_win, 10,
1566 	    "What is the BBR initial window, if 0 use tcp version");
1567 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1568 	    SYSCTL_CHILDREN(bbr_cwnd),
1569 	    OID_AUTO, "do_loss_red", CTLFLAG_RW,
1570 	    &bbr_do_red, 600,
1571 	    "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?");
1572 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1573 	    SYSCTL_CHILDREN(bbr_cwnd),
1574 	    OID_AUTO, "red_scale", CTLFLAG_RW,
1575 	    &bbr_red_scale, 20000,
1576 	    "What RTT do we scale with?");
1577 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1578 	    SYSCTL_CHILDREN(bbr_cwnd),
1579 	    OID_AUTO, "red_growslow", CTLFLAG_RW,
1580 	    &bbr_red_growth_restrict, 1,
1581 	    "Do we restrict cwnd growth for whats in flight?");
1582 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1583 	    SYSCTL_CHILDREN(bbr_cwnd),
1584 	    OID_AUTO, "red_div", CTLFLAG_RW,
1585 	    &bbr_red_div, 2,
1586 	    "If we reduce whats the divisor?");
1587 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1588 	    SYSCTL_CHILDREN(bbr_cwnd),
1589 	    OID_AUTO, "red_mul", CTLFLAG_RW,
1590 	    &bbr_red_mul, 1,
1591 	    "If we reduce whats the mulitiplier?");
1592 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1593 	    SYSCTL_CHILDREN(bbr_cwnd),
1594 	    OID_AUTO, "target_is_unit", CTLFLAG_RW,
1595 	    &bbr_target_is_bbunit, 0,
1596 	    "Is the state target the pacing_gain or BBR_UNIT?");
1597 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1598 	    SYSCTL_CHILDREN(bbr_cwnd),
1599 	    OID_AUTO, "drop_limit", CTLFLAG_RW,
1600 	    &bbr_drop_limit, 0,
1601 	    "Number of segments limit for drop (0=use min_cwnd w/flight)?");
1602 
1603 	/* Timeout controls */
1604 	bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1605 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1606 	    OID_AUTO,
1607 	    "timeout",
1608 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1609 	    "Time out controls");
1610 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1611 	    SYSCTL_CHILDREN(bbr_timeout),
1612 	    OID_AUTO, "delack", CTLFLAG_RW,
1613 	    &bbr_delack_time, 100000,
1614 	    "BBR's delayed ack time");
1615 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1616 	    SYSCTL_CHILDREN(bbr_timeout),
1617 	    OID_AUTO, "tlp_uses", CTLFLAG_RW,
1618 	    &bbr_tlp_type_to_use, 3,
1619 	    "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt");
1620 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1621 	    SYSCTL_CHILDREN(bbr_timeout),
1622 	    OID_AUTO, "persmin", CTLFLAG_RW,
1623 	    &bbr_persist_min, 250000,
1624 	    "What is the minimum time in microseconds between persists");
1625 	SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1626 	    SYSCTL_CHILDREN(bbr_timeout),
1627 	    OID_AUTO, "persmax", CTLFLAG_RW,
1628 	    &bbr_persist_max, 1000000,
1629 	    "What is the largest delay in microseconds between persists");
1630 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1631 	    SYSCTL_CHILDREN(bbr_timeout),
1632 	    OID_AUTO, "tlp_minto", CTLFLAG_RW,
1633 	    &bbr_tlp_min, 10000,
1634 	    "TLP Min timeout in usecs");
1635 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1636 	    SYSCTL_CHILDREN(bbr_timeout),
1637 	    OID_AUTO, "tlp_dack_time", CTLFLAG_RW,
1638 	    &bbr_delayed_ack_time, 200000,
1639 	    "TLP delayed ack compensation value");
1640 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1641 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1642 	    OID_AUTO, "minrto", CTLFLAG_RW,
1643 	    &bbr_rto_min_ms, 30,
1644 	    "Minimum RTO in ms");
1645 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1646 	    SYSCTL_CHILDREN(bbr_timeout),
1647 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1648 	    &bbr_rto_max_sec, 4,
1649 	    "Maximum RTO in seconds -- should be at least as large as min_rto");
1650 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1651 	    SYSCTL_CHILDREN(bbr_timeout),
1652 	    OID_AUTO, "tlp_retry", CTLFLAG_RW,
1653 	    &bbr_tlp_max_resend, 2,
1654 	    "How many times does TLP retry a single segment or multiple with no ACK");
1655 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1656 	    SYSCTL_CHILDREN(bbr_timeout),
1657 	    OID_AUTO, "minto", CTLFLAG_RW,
1658 	    &bbr_min_to, 1000,
1659 	    "Minimum rack timeout in useconds");
1660 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1661 	    SYSCTL_CHILDREN(bbr_timeout),
1662 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1663 	    &bbr_pkt_delay, 1000,
1664 	    "Extra RACK time (in useconds) besides reordering thresh");
1665 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1666 	    SYSCTL_CHILDREN(bbr_timeout),
1667 	    OID_AUTO, "incr_tmrs", CTLFLAG_RW,
1668 	    &bbr_incr_timers, 1,
1669 	    "Increase the RXT/TLP timer by the pacing time used?");
1670 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1671 	    SYSCTL_CHILDREN(bbr_timeout),
1672 	    OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW,
1673 	    &bbr_marks_rxt_sack_passed, 0,
1674 	    "Mark sack passed on all those not ack'd when a RXT hits?");
1675 	/* Policer controls */
1676 	bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1677 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1678 	    OID_AUTO,
1679 	    "policer",
1680 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1681 	    "Policer controls");
1682 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1683 	    SYSCTL_CHILDREN(bbr_policer),
1684 	    OID_AUTO, "detect_enable", CTLFLAG_RW,
1685 	    &bbr_policer_detection_enabled, 1,
1686 	    "Is policer detection enabled??");
1687 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1688 	    SYSCTL_CHILDREN(bbr_policer),
1689 	    OID_AUTO, "min_pes", CTLFLAG_RW,
1690 	    &bbr_lt_intvl_min_rtts, 4,
1691 	    "Minimum number of PE's?");
1692 	SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1693 	    SYSCTL_CHILDREN(bbr_policer),
1694 	    OID_AUTO, "bwdiff", CTLFLAG_RW,
1695 	    &bbr_lt_bw_diff, (4000/8),
1696 	    "Minimal bw diff?");
1697 	SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1698 	    SYSCTL_CHILDREN(bbr_policer),
1699 	    OID_AUTO, "bwratio", CTLFLAG_RW,
1700 	    &bbr_lt_bw_ratio, 8,
1701 	    "Minimal bw diff?");
1702 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1703 	    SYSCTL_CHILDREN(bbr_policer),
1704 	    OID_AUTO, "from_rack_rxt", CTLFLAG_RW,
1705 	    &bbr_policer_call_from_rack_to, 0,
1706 	    "Do we call the policer detection code from a rack-timeout?");
1707 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1708 	    SYSCTL_CHILDREN(bbr_policer),
1709 	    OID_AUTO, "false_postive", CTLFLAG_RW,
1710 	    &bbr_lt_intvl_fp, 0,
1711 	    "What packet epoch do we do false-positive detection at (0=no)?");
1712 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1713 	    SYSCTL_CHILDREN(bbr_policer),
1714 	    OID_AUTO, "loss_thresh", CTLFLAG_RW,
1715 	    &bbr_lt_loss_thresh, 196,
1716 	    "Loss threshold 196 = 19.6%?");
1717 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1718 	    SYSCTL_CHILDREN(bbr_policer),
1719 	    OID_AUTO, "false_postive_thresh", CTLFLAG_RW,
1720 	    &bbr_lt_fd_thresh, 100,
1721 	    "What percentage is the false detection threshold (150=15.0)?");
1722 	/* All the rest */
1723 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1724 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1725 	    OID_AUTO, "cheat_rxt", CTLFLAG_RW,
1726 	    &bbr_use_rack_resend_cheat, 0,
1727 	    "Do we burst 1ms between sends on retransmissions (like rack)?");
1728 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1729 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1730 	    OID_AUTO, "error_paceout", CTLFLAG_RW,
1731 	    &bbr_error_base_paceout, 10000,
1732 	    "When we hit an error what is the min to pace out in usec's?");
1733 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1734 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1735 	    OID_AUTO, "kill_paceout", CTLFLAG_RW,
1736 	    &bbr_max_net_error_cnt, 10,
1737 	    "When we hit this many errors in a row, kill the session?");
1738 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1739 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1740 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1741 	    &bbr_ignore_data_after_close, 1,
1742 	    "Do we hold off sending a RST until all pending data is ack'd");
1743 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1744 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1745 	    OID_AUTO, "resend_use_tso", CTLFLAG_RW,
1746 	    &bbr_resends_use_tso, 0,
1747 	    "Can resends use TSO?");
1748 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1749 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1750 	    OID_AUTO, "sblklimit", CTLFLAG_RW,
1751 	    &bbr_sack_block_limit, 128,
1752 	    "When do we start ignoring small sack blocks");
1753 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1754 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1755 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1756 	    &bbr_verbose_logging, 0,
1757 	    "Should BBR black box logging be verbose");
1758 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1759 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1760 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1761 	    &bbr_reorder_thresh, 2,
1762 	    "What factor for rack will be added when seeing reordering (shift right)");
1763 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1764 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1765 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1766 	    &bbr_reorder_fade, 0,
1767 	    "Does reorder detection fade, if so how many ms (0 means never)");
1768 	SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1769 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1770 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1771 	    &bbr_tlp_thresh, 1,
1772 	    "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1773 	/* Stats and counters */
1774 	/* The pacing counters for hdwr/software can't be in the array */
1775 	bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1776 	bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1777 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1778 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1779 	    OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD,
1780 	    &bbr_hdwr_pacing_enobuf,
1781 	    "Total number of enobufs for hardware paced flows");
1782 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1783 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1784 	    OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD,
1785 	    &bbr_nohdwr_pacing_enobuf,
1786 	    "Total number of enobufs for non-hardware paced flows");
1787 
1788 	bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK);
1789 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1790 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1791 	    OID_AUTO, "hdwr_pacing", CTLFLAG_RD,
1792 	    &bbr_flows_whdwr_pacing,
1793 	    "Total number of hardware paced flows");
1794 	bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK);
1795 	SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1796 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1797 	    OID_AUTO, "software_pacing", CTLFLAG_RD,
1798 	    &bbr_flows_nohdwr_pacing,
1799 	    "Total number of software paced flows");
1800 	COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK);
1801 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1802 	    OID_AUTO, "stats", CTLFLAG_RD,
1803 	    bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats");
1804 	COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK);
1805 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1806 	    OID_AUTO, "opts", CTLFLAG_RD,
1807 	    bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats");
1808 	COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK);
1809 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1810 	    OID_AUTO, "lost", CTLFLAG_RD,
1811 	    bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur");
1812 	COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK);
1813 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1814 	    OID_AUTO, "stateresend", CTLFLAG_RD,
1815 	    bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend");
1816 	COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK);
1817 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1818 	    OID_AUTO, "statetime", CTLFLAG_RD,
1819 	    bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states");
1820 	COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1821 	SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1822 	    OID_AUTO, "outsize", CTLFLAG_RD,
1823 	    bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls");
1824 	SYSCTL_ADD_PROC(&bbr_sysctl_ctx,
1825 	    SYSCTL_CHILDREN(bbr_sysctl_root),
1826 	    OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1827 	    &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters");
1828 }
1829 
1830 static void
bbr_counter_destroy(void)1831 bbr_counter_destroy(void)
1832 {
1833 	COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE);
1834 	COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE);
1835 	COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE);
1836 	COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT);
1837 	COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT);
1838 	COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT);
1839 	counter_u64_free(bbr_nohdwr_pacing_enobuf);
1840 	counter_u64_free(bbr_hdwr_pacing_enobuf);
1841 	counter_u64_free(bbr_flows_whdwr_pacing);
1842 	counter_u64_free(bbr_flows_nohdwr_pacing);
1843 
1844 }
1845 
1846 static __inline void
bbr_fill_in_logging_data(struct tcp_bbr * bbr,struct tcp_log_bbr * l,uint32_t cts)1847 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts)
1848 {
1849 	memset(l, 0, sizeof(union tcp_log_stackspecific));
1850 	l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate;
1851 	l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate);
1852 	l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
1853 	l->bw_inuse = bbr_get_bw(bbr);
1854 	l->inflight = ctf_flight_size(bbr->rc_tp,
1855 			  (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
1856 	l->applimited = bbr->r_ctl.r_app_limited_until;
1857 	l->delivered = bbr->r_ctl.rc_delivered;
1858 	l->timeStamp = cts;
1859 	l->lost = bbr->r_ctl.rc_lost;
1860 	l->bbr_state = bbr->rc_bbr_state;
1861 	l->bbr_substate = bbr_state_val(bbr);
1862 	l->epoch = bbr->r_ctl.rc_rtt_epoch;
1863 	l->lt_epoch = bbr->r_ctl.rc_lt_epoch;
1864 	l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
1865 	l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain;
1866 	l->inhpts = tcp_in_hpts(bbr->rc_tp);
1867 	l->use_lt_bw = bbr->rc_lt_use_bw;
1868 	l->pkts_out = bbr->r_ctl.rc_flight_at_input;
1869 	l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch;
1870 }
1871 
1872 static void
bbr_log_type_bw_reduce(struct tcp_bbr * bbr,int reason)1873 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason)
1874 {
1875 	if (tcp_bblogging_on(bbr->rc_tp)) {
1876 		union tcp_log_stackspecific log;
1877 
1878 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1879 		log.u_bbr.flex1 = 0;
1880 		log.u_bbr.flex2 = 0;
1881 		log.u_bbr.flex5 = 0;
1882 		log.u_bbr.flex3 = 0;
1883 		log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate;
1884 		log.u_bbr.flex7 = reason;
1885 		log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt;
1886 		log.u_bbr.flex8 = 0;
1887 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1888 		    &bbr->rc_inp->inp_socket->so_rcv,
1889 		    &bbr->rc_inp->inp_socket->so_snd,
1890 		    BBR_LOG_BW_RED_EV, 0,
1891 		    0, &log, false, &bbr->rc_tv);
1892 	}
1893 }
1894 
1895 static void
bbr_log_type_rwnd_collapse(struct tcp_bbr * bbr,int seq,int mode,uint32_t count)1896 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count)
1897 {
1898 	if (tcp_bblogging_on(bbr->rc_tp)) {
1899 		union tcp_log_stackspecific log;
1900 
1901 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1902 		log.u_bbr.flex1 = seq;
1903 		log.u_bbr.flex2 = count;
1904 		log.u_bbr.flex8 = mode;
1905 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1906 		    &bbr->rc_inp->inp_socket->so_rcv,
1907 		    &bbr->rc_inp->inp_socket->so_snd,
1908 		    BBR_LOG_LOWGAIN, 0,
1909 		    0, &log, false, &bbr->rc_tv);
1910 	}
1911 }
1912 
1913 static void
bbr_log_type_just_return(struct tcp_bbr * bbr,uint32_t cts,uint32_t tlen,uint8_t hpts_calling,uint8_t reason,uint32_t p_maxseg,int len)1914 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling,
1915     uint8_t reason, uint32_t p_maxseg, int len)
1916 {
1917 	if (tcp_bblogging_on(bbr->rc_tp)) {
1918 		union tcp_log_stackspecific log;
1919 
1920 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1921 		log.u_bbr.flex1 = p_maxseg;
1922 		log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags;
1923 		log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
1924 		log.u_bbr.flex4 = reason;
1925 		log.u_bbr.flex5 = bbr->rc_in_persist;
1926 		log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val;
1927 		log.u_bbr.flex7 = p_maxseg;
1928 		log.u_bbr.flex8 = bbr->rc_in_persist;
1929 		log.u_bbr.pkts_out = 0;
1930 		log.u_bbr.applimited = len;
1931 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1932 		    &bbr->rc_inp->inp_socket->so_rcv,
1933 		    &bbr->rc_inp->inp_socket->so_snd,
1934 		    BBR_LOG_JUSTRET, 0,
1935 		    tlen, &log, false, &bbr->rc_tv);
1936 	}
1937 }
1938 
1939 static void
bbr_log_type_enter_rec(struct tcp_bbr * bbr,uint32_t seq)1940 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq)
1941 {
1942 	if (tcp_bblogging_on(bbr->rc_tp)) {
1943 		union tcp_log_stackspecific log;
1944 
1945 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1946 		log.u_bbr.flex1 = seq;
1947 		log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
1948 		log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start;
1949 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1950 		    &bbr->rc_inp->inp_socket->so_rcv,
1951 		    &bbr->rc_inp->inp_socket->so_snd,
1952 		    BBR_LOG_ENTREC, 0,
1953 		    0, &log, false, &bbr->rc_tv);
1954 	}
1955 }
1956 
1957 static void
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)1958 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)
1959 {
1960 	if (tcp_bblogging_on(tp)) {
1961 		union tcp_log_stackspecific log;
1962 
1963 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1964 		log.u_bbr.flex1 = tso;
1965 		log.u_bbr.flex2 = maxseg;
1966 		log.u_bbr.flex3 = mtu;
1967 		log.u_bbr.flex4 = csum_flags;
1968 		TCP_LOG_EVENTP(tp, NULL,
1969 		    &bbr->rc_inp->inp_socket->so_rcv,
1970 		    &bbr->rc_inp->inp_socket->so_snd,
1971 		    BBR_LOG_MSGSIZE, 0,
1972 		    0, &log, false, &bbr->rc_tv);
1973 	}
1974 }
1975 
1976 static void
bbr_log_flowend(struct tcp_bbr * bbr)1977 bbr_log_flowend(struct tcp_bbr *bbr)
1978 {
1979 	if (tcp_bblogging_on(bbr->rc_tp)) {
1980 		union tcp_log_stackspecific log;
1981 		struct sockbuf *r, *s;
1982 		struct timeval tv;
1983 
1984 		if (bbr->rc_inp->inp_socket) {
1985 			r = &bbr->rc_inp->inp_socket->so_rcv;
1986 			s = &bbr->rc_inp->inp_socket->so_snd;
1987 		} else {
1988 			r = s = NULL;
1989 		}
1990 		bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv));
1991 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1992 		    r, s,
1993 		    TCP_LOG_FLOWEND, 0,
1994 		    0, &log, false, &tv);
1995 	}
1996 }
1997 
1998 static void
bbr_log_pkt_epoch(struct tcp_bbr * bbr,uint32_t cts,uint32_t line,uint32_t lost,uint32_t del)1999 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line,
2000     uint32_t lost, uint32_t del)
2001 {
2002 	if (tcp_bblogging_on(bbr->rc_tp)) {
2003 		union tcp_log_stackspecific log;
2004 
2005 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2006 		log.u_bbr.flex1 = lost;
2007 		log.u_bbr.flex2 = del;
2008 		log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw;
2009 		log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt;
2010 		log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2011 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2012 		log.u_bbr.flex7 = line;
2013 		log.u_bbr.flex8 = 0;
2014 		log.u_bbr.inflight = bbr->r_ctl.r_measurement_count;
2015 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2016 		    &bbr->rc_inp->inp_socket->so_rcv,
2017 		    &bbr->rc_inp->inp_socket->so_snd,
2018 		    BBR_LOG_PKT_EPOCH, 0,
2019 		    0, &log, false, &bbr->rc_tv);
2020 	}
2021 }
2022 
2023 static void
bbr_log_time_epoch(struct tcp_bbr * bbr,uint32_t cts,uint32_t line,uint32_t epoch_time)2024 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time)
2025 {
2026 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2027 		union tcp_log_stackspecific log;
2028 
2029 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2030 		log.u_bbr.flex1 = bbr->r_ctl.rc_lost;
2031 		log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat;
2032 		log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat;
2033 		log.u_bbr.flex7 = line;
2034 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2035 		    &bbr->rc_inp->inp_socket->so_rcv,
2036 		    &bbr->rc_inp->inp_socket->so_snd,
2037 		    BBR_LOG_TIME_EPOCH, 0,
2038 		    0, &log, false, &bbr->rc_tv);
2039 	}
2040 }
2041 
2042 static void
bbr_log_set_of_state_target(struct tcp_bbr * bbr,uint32_t new_tar,int line,int meth)2043 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth)
2044 {
2045 	if (tcp_bblogging_on(bbr->rc_tp)) {
2046 		union tcp_log_stackspecific log;
2047 
2048 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2049 		log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2050 		log.u_bbr.flex2 = new_tar;
2051 		log.u_bbr.flex3 = line;
2052 		log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2053 		log.u_bbr.flex5 = bbr_quanta;
2054 		log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs;
2055 		log.u_bbr.flex7 = bbr->rc_last_options;
2056 		log.u_bbr.flex8 = meth;
2057 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2058 		    &bbr->rc_inp->inp_socket->so_rcv,
2059 		    &bbr->rc_inp->inp_socket->so_snd,
2060 		    BBR_LOG_STATE_TARGET, 0,
2061 		    0, &log, false, &bbr->rc_tv);
2062 	}
2063 
2064 }
2065 
2066 static void
bbr_log_type_statechange(struct tcp_bbr * bbr,uint32_t cts,int32_t line)2067 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2068 {
2069 	if (tcp_bblogging_on(bbr->rc_tp)) {
2070 		union tcp_log_stackspecific log;
2071 
2072 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2073 		log.u_bbr.flex1 = line;
2074 		log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2075 		log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2076 		if (bbr_state_is_pkt_epoch)
2077 			log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
2078 		else
2079 			log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP);
2080 		log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2081 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2082 		log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000);
2083 		log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra;
2084 		log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state;
2085 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2086 		    &bbr->rc_inp->inp_socket->so_rcv,
2087 		    &bbr->rc_inp->inp_socket->so_snd,
2088 		    BBR_LOG_STATE, 0,
2089 		    0, &log, false, &bbr->rc_tv);
2090 	}
2091 }
2092 
2093 static void
bbr_log_rtt_shrinks(struct tcp_bbr * bbr,uint32_t cts,uint32_t applied,uint32_t rtt,uint32_t line,uint8_t reas,uint16_t cond)2094 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
2095 		    uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond)
2096 {
2097 	if (tcp_bblogging_on(bbr->rc_tp)) {
2098 		union tcp_log_stackspecific log;
2099 
2100 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2101 		log.u_bbr.flex1 = line;
2102 		log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2103 		log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt;
2104 		log.u_bbr.flex4 = applied;
2105 		log.u_bbr.flex5 = rtt;
2106 		log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2107 		log.u_bbr.flex7 = cond;
2108 		log.u_bbr.flex8 = reas;
2109 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2110 		    &bbr->rc_inp->inp_socket->so_rcv,
2111 		    &bbr->rc_inp->inp_socket->so_snd,
2112 		    BBR_LOG_RTT_SHRINKS, 0,
2113 		    0, &log, false, &bbr->rc_tv);
2114 	}
2115 }
2116 
2117 static void
bbr_log_type_exit_rec(struct tcp_bbr * bbr)2118 bbr_log_type_exit_rec(struct tcp_bbr *bbr)
2119 {
2120 	if (tcp_bblogging_on(bbr->rc_tp)) {
2121 		union tcp_log_stackspecific log;
2122 
2123 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2124 		log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start;
2125 		log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
2126 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2127 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2128 		    &bbr->rc_inp->inp_socket->so_rcv,
2129 		    &bbr->rc_inp->inp_socket->so_snd,
2130 		    BBR_LOG_EXITREC, 0,
2131 		    0, &log, false, &bbr->rc_tv);
2132 	}
2133 }
2134 
2135 static void
bbr_log_type_cwndupd(struct tcp_bbr * bbr,uint32_t bytes_this_ack,uint32_t chg,uint32_t prev_acked,int32_t meth,uint32_t target,uint32_t th_ack,int32_t line)2136 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg,
2137     uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line)
2138 {
2139 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2140 		union tcp_log_stackspecific log;
2141 
2142 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2143 		log.u_bbr.flex1 = line;
2144 		log.u_bbr.flex2 = prev_acked;
2145 		log.u_bbr.flex3 = bytes_this_ack;
2146 		log.u_bbr.flex4 = chg;
2147 		log.u_bbr.flex5 = th_ack;
2148 		log.u_bbr.flex6 = target;
2149 		log.u_bbr.flex8 = meth;
2150 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2151 		    &bbr->rc_inp->inp_socket->so_rcv,
2152 		    &bbr->rc_inp->inp_socket->so_snd,
2153 		    BBR_LOG_CWND, 0,
2154 		    0, &log, false, &bbr->rc_tv);
2155 	}
2156 }
2157 
2158 static void
bbr_log_rtt_sample(struct tcp_bbr * bbr,uint32_t rtt,uint32_t tsin)2159 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin)
2160 {
2161 	/*
2162 	 * Log the rtt sample we are applying to the srtt algorithm in
2163 	 * useconds.
2164 	 */
2165 	if (tcp_bblogging_on(bbr->rc_tp)) {
2166 		union tcp_log_stackspecific log;
2167 
2168 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2169 		log.u_bbr.flex1 = rtt;
2170 		log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time;
2171 		log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay;
2172 		log.u_bbr.flex4 = bbr->rc_tp->ts_offset;
2173 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2174 		log.u_bbr.pkts_out = tcp_tv_to_msec(&bbr->rc_tv);
2175 		log.u_bbr.flex6 = tsin;
2176 		log.u_bbr.flex7 = 0;
2177 		log.u_bbr.flex8 = bbr->rc_ack_was_delayed;
2178 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2179 		    &bbr->rc_inp->inp_socket->so_rcv,
2180 		    &bbr->rc_inp->inp_socket->so_snd,
2181 		    TCP_LOG_RTT, 0,
2182 		    0, &log, false, &bbr->rc_tv);
2183 	}
2184 }
2185 
2186 static void
bbr_log_type_pesist(struct tcp_bbr * bbr,uint32_t cts,uint32_t time_in,int32_t line,uint8_t enter_exit)2187 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit)
2188 {
2189 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2190 		union tcp_log_stackspecific log;
2191 
2192 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2193 		log.u_bbr.flex1 = time_in;
2194 		log.u_bbr.flex2 = line;
2195 		log.u_bbr.flex8 = enter_exit;
2196 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2197 		    &bbr->rc_inp->inp_socket->so_rcv,
2198 		    &bbr->rc_inp->inp_socket->so_snd,
2199 		    BBR_LOG_PERSIST, 0,
2200 		    0, &log, false, &bbr->rc_tv);
2201 	}
2202 }
2203 static void
bbr_log_ack_clear(struct tcp_bbr * bbr,uint32_t cts)2204 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts)
2205 {
2206 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2207 		union tcp_log_stackspecific log;
2208 
2209 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2210 		log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age;
2211 		log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2212 		log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2213 		log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time;
2214 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2215 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2216 		    &bbr->rc_inp->inp_socket->so_rcv,
2217 		    &bbr->rc_inp->inp_socket->so_snd,
2218 		    BBR_LOG_ACKCLEAR, 0,
2219 		    0, &log, false, &bbr->rc_tv);
2220 	}
2221 }
2222 
2223 static void
bbr_log_ack_event(struct tcp_bbr * bbr,struct tcphdr * th,struct tcpopt * to,uint32_t tlen,uint16_t nsegs,uint32_t cts,int32_t nxt_pkt,struct mbuf * m)2224 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen,
2225 		  uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m)
2226 {
2227 	if (tcp_bblogging_on(bbr->rc_tp)) {
2228 		union tcp_log_stackspecific log;
2229 		struct timeval tv;
2230 
2231 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2232 		log.u_bbr.flex1 = nsegs;
2233 		log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes;
2234 		if (m) {
2235 			struct timespec ts;
2236 
2237 			log.u_bbr.flex3 = m->m_flags;
2238 			if (m->m_flags & M_TSTMP) {
2239 				mbuf_tstmp2timespec(m, &ts);
2240 				tv.tv_sec = ts.tv_sec;
2241 				tv.tv_usec = ts.tv_nsec / 1000;
2242 				log.u_bbr.lt_epoch = tcp_tv_to_usec(&tv);
2243 			} else {
2244 				log.u_bbr.lt_epoch = 0;
2245 			}
2246 			if (m->m_flags & M_TSTMP_LRO) {
2247 				mbuf_tstmp2timeval(m, &tv);
2248 				log.u_bbr.flex5 = tcp_tv_to_usec(&tv);
2249 			} else {
2250 				/* No arrival timestamp */
2251 				log.u_bbr.flex5 = 0;
2252 			}
2253 
2254 			log.u_bbr.pkts_out = tcp_get_usecs(&tv);
2255 		} else {
2256 			log.u_bbr.flex3 = 0;
2257 			log.u_bbr.flex5 = 0;
2258 			log.u_bbr.flex6 = 0;
2259 			log.u_bbr.pkts_out = 0;
2260 		}
2261 		log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state;
2262 		log.u_bbr.flex7 = bbr->r_wanted_output;
2263 		log.u_bbr.flex8 = bbr->rc_in_persist;
2264 		TCP_LOG_EVENTP(bbr->rc_tp, th,
2265 		    &bbr->rc_inp->inp_socket->so_rcv,
2266 		    &bbr->rc_inp->inp_socket->so_snd,
2267 		    TCP_LOG_IN, 0,
2268 		    tlen, &log, true, &bbr->rc_tv);
2269 	}
2270 }
2271 
2272 static void
bbr_log_doseg_done(struct tcp_bbr * bbr,uint32_t cts,int32_t nxt_pkt,int32_t did_out)2273 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out)
2274 {
2275 	if (tcp_bblogging_on(bbr->rc_tp)) {
2276 		union tcp_log_stackspecific log;
2277 
2278 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2279 		log.u_bbr.flex1 = did_out;
2280 		log.u_bbr.flex2 = nxt_pkt;
2281 		log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val;
2282 		log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2283 		log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp;
2284 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes;
2285 		log.u_bbr.flex7 = bbr->r_wanted_output;
2286 		log.u_bbr.flex8 = bbr->rc_in_persist;
2287 		log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay;
2288 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2289 		    &bbr->rc_inp->inp_socket->so_rcv,
2290 		    &bbr->rc_inp->inp_socket->so_snd,
2291 		    BBR_LOG_DOSEG_DONE, 0,
2292 		    0, &log, true, &bbr->rc_tv);
2293 	}
2294 }
2295 
2296 static void
bbr_log_enobuf_jmp(struct tcp_bbr * bbr,uint32_t len,uint32_t cts,int32_t line,uint32_t o_len,uint32_t segcnt,uint32_t segsiz)2297 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts,
2298     int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz)
2299 {
2300 	if (tcp_bblogging_on(bbr->rc_tp)) {
2301 		union tcp_log_stackspecific log;
2302 
2303 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2304 		log.u_bbr.flex1 = line;
2305 		log.u_bbr.flex2 = o_len;
2306 		log.u_bbr.flex3 = segcnt;
2307 		log.u_bbr.flex4 = segsiz;
2308 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2309 		    &bbr->rc_inp->inp_socket->so_rcv,
2310 		    &bbr->rc_inp->inp_socket->so_snd,
2311 		    BBR_LOG_ENOBUF_JMP, ENOBUFS,
2312 		    len, &log, true, &bbr->rc_tv);
2313 	}
2314 }
2315 
2316 static void
bbr_log_to_processing(struct tcp_bbr * bbr,uint32_t cts,int32_t ret,int32_t timers,uint8_t hpts_calling)2317 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling)
2318 {
2319 	if (tcp_bblogging_on(bbr->rc_tp)) {
2320 		union tcp_log_stackspecific log;
2321 
2322 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2323 		log.u_bbr.flex1 = timers;
2324 		log.u_bbr.flex2 = ret;
2325 		log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
2326 		log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2327 		log.u_bbr.flex5 = cts;
2328 		log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2329 		log.u_bbr.flex8 = hpts_calling;
2330 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2331 		    &bbr->rc_inp->inp_socket->so_rcv,
2332 		    &bbr->rc_inp->inp_socket->so_snd,
2333 		    BBR_LOG_TO_PROCESS, 0,
2334 		    0, &log, false, &bbr->rc_tv);
2335 	}
2336 }
2337 
2338 static void
bbr_log_to_event(struct tcp_bbr * bbr,uint32_t cts,int32_t to_num)2339 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num)
2340 {
2341 	if (tcp_bblogging_on(bbr->rc_tp)) {
2342 		union tcp_log_stackspecific log;
2343 		uint64_t ar;
2344 
2345 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2346 		log.u_bbr.flex1 = bbr->bbr_timer_src;
2347 		log.u_bbr.flex2 = 0;
2348 		log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2349 		ar = (uintptr_t)(bbr->r_ctl.rc_resend);
2350 		ar >>= 32;
2351 		ar &= 0x00000000ffffffff;
2352 		log.u_bbr.flex4 = (uint32_t)ar;
2353 		ar = (uintptr_t)bbr->r_ctl.rc_resend;
2354 		ar &= 0x00000000ffffffff;
2355 		log.u_bbr.flex5 = (uint32_t)ar;
2356 		log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2357 		log.u_bbr.flex8 = to_num;
2358 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2359 		    &bbr->rc_inp->inp_socket->so_rcv,
2360 		    &bbr->rc_inp->inp_socket->so_snd,
2361 		    BBR_LOG_RTO, 0,
2362 		    0, &log, false, &bbr->rc_tv);
2363 	}
2364 }
2365 
2366 static void
bbr_log_startup_event(struct tcp_bbr * bbr,uint32_t cts,uint32_t flex1,uint32_t flex2,uint32_t flex3,uint8_t reason)2367 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason)
2368 {
2369 	if (tcp_bblogging_on(bbr->rc_tp)) {
2370 		union tcp_log_stackspecific log;
2371 
2372 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2373 		log.u_bbr.flex1 = flex1;
2374 		log.u_bbr.flex2 = flex2;
2375 		log.u_bbr.flex3 = flex3;
2376 		log.u_bbr.flex4 = 0;
2377 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2378 		log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2379 		log.u_bbr.flex8 = reason;
2380 		log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw;
2381 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2382 		    &bbr->rc_inp->inp_socket->so_rcv,
2383 		    &bbr->rc_inp->inp_socket->so_snd,
2384 		    BBR_LOG_REDUCE, 0,
2385 		    0, &log, false, &bbr->rc_tv);
2386 	}
2387 }
2388 
2389 static void
bbr_log_hpts_diag(struct tcp_bbr * bbr,uint32_t cts,struct hpts_diag * diag)2390 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag)
2391 {
2392 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2393 		union tcp_log_stackspecific log;
2394 
2395 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2396 		log.u_bbr.flex1 = diag->p_nxt_slot;
2397 		log.u_bbr.flex2 = diag->p_cur_slot;
2398 		log.u_bbr.flex3 = diag->slot_req;
2399 		log.u_bbr.flex4 = diag->inp_hptsslot;
2400 		log.u_bbr.flex5 = diag->slot_remaining;
2401 		log.u_bbr.flex6 = diag->need_new_to;
2402 		log.u_bbr.flex7 = diag->p_hpts_active;
2403 		log.u_bbr.flex8 = diag->p_on_min_sleep;
2404 		/* Hijack other fields as needed  */
2405 		log.u_bbr.epoch = diag->have_slept;
2406 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
2407 		log.u_bbr.pkts_out = diag->co_ret;
2408 		log.u_bbr.applimited = diag->hpts_sleep_time;
2409 		log.u_bbr.delivered = diag->p_prev_slot;
2410 		log.u_bbr.inflight = diag->p_runningslot;
2411 		log.u_bbr.bw_inuse = diag->wheel_slot;
2412 		log.u_bbr.rttProp = diag->wheel_cts;
2413 		log.u_bbr.delRate = diag->maxslots;
2414 		log.u_bbr.cur_del_rate = diag->p_curtick;
2415 		log.u_bbr.cur_del_rate <<= 32;
2416 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
2417 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2418 		    &bbr->rc_inp->inp_socket->so_rcv,
2419 		    &bbr->rc_inp->inp_socket->so_snd,
2420 		    BBR_LOG_HPTSDIAG, 0,
2421 		    0, &log, false, &bbr->rc_tv);
2422 	}
2423 }
2424 
2425 static void
bbr_log_timer_var(struct tcp_bbr * bbr,int mode,uint32_t cts,uint32_t time_since_sent,uint32_t srtt,uint32_t thresh,uint32_t to)2426 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
2427     uint32_t thresh, uint32_t to)
2428 {
2429 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2430 		union tcp_log_stackspecific log;
2431 
2432 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2433 		log.u_bbr.flex1 = bbr->rc_tp->t_rttvar;
2434 		log.u_bbr.flex2 = time_since_sent;
2435 		log.u_bbr.flex3 = srtt;
2436 		log.u_bbr.flex4 = thresh;
2437 		log.u_bbr.flex5 = to;
2438 		log.u_bbr.flex6 = bbr->rc_tp->t_srtt;
2439 		log.u_bbr.flex8 = mode;
2440 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2441 		    &bbr->rc_inp->inp_socket->so_rcv,
2442 		    &bbr->rc_inp->inp_socket->so_snd,
2443 		    BBR_LOG_TIMERPREP, 0,
2444 		    0, &log, false, &bbr->rc_tv);
2445 	}
2446 }
2447 
2448 static void
bbr_log_pacing_delay_calc(struct tcp_bbr * bbr,uint16_t gain,uint32_t len,uint32_t cts,uint32_t usecs,uint64_t bw,uint32_t override,int mod)2449 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
2450     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod)
2451 {
2452 	if (tcp_bblogging_on(bbr->rc_tp)) {
2453 		union tcp_log_stackspecific log;
2454 
2455 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2456 		log.u_bbr.flex1 = usecs;
2457 		log.u_bbr.flex2 = len;
2458 		log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff);
2459 		log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff);
2460 		if (override)
2461 			log.u_bbr.flex5 = (1 << 2);
2462 		else
2463 			log.u_bbr.flex5 = 0;
2464 		log.u_bbr.flex6 = override;
2465 		log.u_bbr.flex7 = gain;
2466 		log.u_bbr.flex8 = mod;
2467 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2468 		    &bbr->rc_inp->inp_socket->so_rcv,
2469 		    &bbr->rc_inp->inp_socket->so_snd,
2470 		    BBR_LOG_HPTSI_CALC, 0,
2471 		    len, &log, false, &bbr->rc_tv);
2472 	}
2473 }
2474 
2475 static void
bbr_log_to_start(struct tcp_bbr * bbr,uint32_t cts,uint32_t to,int32_t slot,uint8_t which)2476 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2477 {
2478 	if (tcp_bblogging_on(bbr->rc_tp)) {
2479 		union tcp_log_stackspecific log;
2480 
2481 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2482 
2483 		log.u_bbr.flex1 = bbr->bbr_timer_src;
2484 		log.u_bbr.flex2 = to;
2485 		log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2486 		log.u_bbr.flex4 = slot;
2487 		log.u_bbr.flex5 = bbr->rc_tp->t_hpts_slot;
2488 		log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2489 		log.u_bbr.pkts_out = bbr->rc_tp->t_flags2;
2490 		log.u_bbr.flex8 = which;
2491 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2492 		    &bbr->rc_inp->inp_socket->so_rcv,
2493 		    &bbr->rc_inp->inp_socket->so_snd,
2494 		    BBR_LOG_TIMERSTAR, 0,
2495 		    0, &log, false, &bbr->rc_tv);
2496 	}
2497 }
2498 
2499 static void
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)2500 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)
2501 {
2502 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2503 		union tcp_log_stackspecific log;
2504 
2505 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2506 		log.u_bbr.flex1 = thresh;
2507 		log.u_bbr.flex2 = lro;
2508 		log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts;
2509 		log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
2510 		log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2511 		log.u_bbr.flex6 = srtt;
2512 		log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift;
2513 		log.u_bbr.flex8 = frm;
2514 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2515 		    &bbr->rc_inp->inp_socket->so_rcv,
2516 		    &bbr->rc_inp->inp_socket->so_snd,
2517 		    BBR_LOG_THRESH_CALC, 0,
2518 		    0, &log, false, &bbr->rc_tv);
2519 	}
2520 }
2521 
2522 static void
bbr_log_to_cancel(struct tcp_bbr * bbr,int32_t line,uint32_t cts,uint8_t hpts_removed)2523 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed)
2524 {
2525 	if (tcp_bblogging_on(bbr->rc_tp)) {
2526 		union tcp_log_stackspecific log;
2527 
2528 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2529 		log.u_bbr.flex1 = line;
2530 		log.u_bbr.flex2 = bbr->bbr_timer_src;
2531 		log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2532 		log.u_bbr.flex4 = bbr->rc_in_persist;
2533 		log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2534 		log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2535 		log.u_bbr.flex8 = hpts_removed;
2536 		log.u_bbr.pkts_out = bbr->rc_pacer_started;
2537 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2538 		    &bbr->rc_inp->inp_socket->so_rcv,
2539 		    &bbr->rc_inp->inp_socket->so_snd,
2540 		    BBR_LOG_TIMERCANC, 0,
2541 		    0, &log, false, &bbr->rc_tv);
2542 	}
2543 }
2544 
2545 static void
bbr_log_tstmp_validation(struct tcp_bbr * bbr,uint64_t peer_delta,uint64_t delta)2546 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta)
2547 {
2548 	if (tcp_bblogging_on(bbr->rc_tp)) {
2549 		union tcp_log_stackspecific log;
2550 
2551 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2552 		log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio;
2553 		log.u_bbr.flex2 = (peer_delta >> 32);
2554 		log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff);
2555 		log.u_bbr.flex4 = (delta >> 32);
2556 		log.u_bbr.flex5 = (delta & 0x00000000ffffffff);
2557 		log.u_bbr.flex7 = bbr->rc_ts_clock_set;
2558 		log.u_bbr.flex8 = bbr->rc_ts_cant_be_used;
2559 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2560 		    &bbr->rc_inp->inp_socket->so_rcv,
2561 		    &bbr->rc_inp->inp_socket->so_snd,
2562 		    BBR_LOG_TSTMP_VAL, 0,
2563 		    0, &log, false, &bbr->rc_tv);
2564 	}
2565 }
2566 
2567 static void
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)2568 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)
2569 {
2570 	if (tcp_bblogging_on(bbr->rc_tp)) {
2571 		union tcp_log_stackspecific log;
2572 
2573 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2574 		log.u_bbr.flex1 = tsosz;
2575 		log.u_bbr.flex2 = tls;
2576 		log.u_bbr.flex3 = tcp_min_hptsi_time;
2577 		log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min;
2578 		log.u_bbr.flex5 = old_val;
2579 		log.u_bbr.flex6 = maxseg;
2580 		log.u_bbr.flex7 = bbr->rc_no_pacing;
2581 		log.u_bbr.flex7 <<= 1;
2582 		log.u_bbr.flex7 |= bbr->rc_past_init_win;
2583 		if (hdwr)
2584 			log.u_bbr.flex8 = 0x80 | bbr->rc_use_google;
2585 		else
2586 			log.u_bbr.flex8 = bbr->rc_use_google;
2587 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2588 		    &bbr->rc_inp->inp_socket->so_rcv,
2589 		    &bbr->rc_inp->inp_socket->so_snd,
2590 		    BBR_LOG_BBRTSO, 0,
2591 		    0, &log, false, &bbr->rc_tv);
2592 	}
2593 }
2594 
2595 static void
bbr_log_type_rsmclear(struct tcp_bbr * bbr,uint32_t cts,struct bbr_sendmap * rsm,uint32_t flags,uint32_t line)2596 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm,
2597 		      uint32_t flags, uint32_t line)
2598 {
2599 	if (tcp_bblogging_on(bbr->rc_tp)) {
2600 		union tcp_log_stackspecific log;
2601 
2602 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2603 		log.u_bbr.flex1 = line;
2604 		log.u_bbr.flex2 = rsm->r_start;
2605 		log.u_bbr.flex3 = rsm->r_end;
2606 		log.u_bbr.flex4 = rsm->r_delivered;
2607 		log.u_bbr.flex5 = rsm->r_rtr_cnt;
2608 		log.u_bbr.flex6 = rsm->r_dupack;
2609 		log.u_bbr.flex7 = rsm->r_tim_lastsent[0];
2610 		log.u_bbr.flex8 = rsm->r_flags;
2611 		/* Hijack the pkts_out fids */
2612 		log.u_bbr.applimited = flags;
2613 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2614 		    &bbr->rc_inp->inp_socket->so_rcv,
2615 		    &bbr->rc_inp->inp_socket->so_snd,
2616 		    BBR_RSM_CLEARED, 0,
2617 		    0, &log, false, &bbr->rc_tv);
2618 	}
2619 }
2620 
2621 static void
bbr_log_type_bbrupd(struct tcp_bbr * bbr,uint8_t flex8,uint32_t cts,uint32_t flex3,uint32_t flex2,uint32_t flex5,uint32_t flex6,uint32_t pkts_out,int flex7,uint32_t flex4,uint32_t flex1)2622 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts,
2623     uint32_t flex3, uint32_t flex2, uint32_t flex5,
2624     uint32_t flex6, uint32_t pkts_out, int flex7,
2625     uint32_t flex4, uint32_t flex1)
2626 {
2627 
2628 	if (tcp_bblogging_on(bbr->rc_tp)) {
2629 		union tcp_log_stackspecific log;
2630 
2631 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2632 		log.u_bbr.flex1 = flex1;
2633 		log.u_bbr.flex2 = flex2;
2634 		log.u_bbr.flex3 = flex3;
2635 		log.u_bbr.flex4 = flex4;
2636 		log.u_bbr.flex5 = flex5;
2637 		log.u_bbr.flex6 = flex6;
2638 		log.u_bbr.flex7 = flex7;
2639 		/* Hijack the pkts_out fids */
2640 		log.u_bbr.pkts_out = pkts_out;
2641 		log.u_bbr.flex8 = flex8;
2642 		if (bbr->rc_ack_was_delayed)
2643 			log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay;
2644 		else
2645 			log.u_bbr.epoch = 0;
2646 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2647 		    &bbr->rc_inp->inp_socket->so_rcv,
2648 		    &bbr->rc_inp->inp_socket->so_snd,
2649 		    BBR_LOG_BBRUPD, 0,
2650 		    flex2, &log, false, &bbr->rc_tv);
2651 	}
2652 }
2653 
2654 static void
bbr_log_type_ltbw(struct tcp_bbr * bbr,uint32_t cts,int32_t reason,uint32_t newbw,uint32_t obw,uint32_t diff,uint32_t tim)2655 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason,
2656 	uint32_t newbw, uint32_t obw, uint32_t diff,
2657 	uint32_t tim)
2658 {
2659 	if (/*bbr_verbose_logging && */tcp_bblogging_on(bbr->rc_tp)) {
2660 		union tcp_log_stackspecific log;
2661 
2662 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2663 		log.u_bbr.flex1 = reason;
2664 		log.u_bbr.flex2 = newbw;
2665 		log.u_bbr.flex3 = obw;
2666 		log.u_bbr.flex4 = diff;
2667 		log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost;
2668 		log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del;
2669 		log.u_bbr.flex7 = bbr->rc_lt_is_sampling;
2670 		log.u_bbr.pkts_out = tim;
2671 		log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw;
2672 		if (bbr->rc_lt_use_bw == 0)
2673 			log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
2674 		else
2675 			log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
2676 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2677 		    &bbr->rc_inp->inp_socket->so_rcv,
2678 		    &bbr->rc_inp->inp_socket->so_snd,
2679 		    BBR_LOG_BWSAMP, 0,
2680 		    0, &log, false, &bbr->rc_tv);
2681 	}
2682 }
2683 
2684 static inline void
bbr_log_progress_event(struct tcp_bbr * bbr,struct tcpcb * tp,uint32_t tick,int event,int line)2685 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line)
2686 {
2687 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2688 		union tcp_log_stackspecific log;
2689 
2690 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2691 		log.u_bbr.flex1 = line;
2692 		log.u_bbr.flex2 = tick;
2693 		log.u_bbr.flex3 = tp->t_maxunacktime;
2694 		log.u_bbr.flex4 = tp->t_acktime;
2695 		log.u_bbr.flex8 = event;
2696 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2697 		    &bbr->rc_inp->inp_socket->so_rcv,
2698 		    &bbr->rc_inp->inp_socket->so_snd,
2699 		    BBR_LOG_PROGRESS, 0,
2700 		    0, &log, false, &bbr->rc_tv);
2701 	}
2702 }
2703 
2704 static void
bbr_type_log_hdwr_pacing(struct tcp_bbr * bbr,const struct ifnet * ifp,uint64_t rate,uint64_t hw_rate,int line,uint32_t cts,int error)2705 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp,
2706 			 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts,
2707 			 int error)
2708 {
2709 	if (tcp_bblogging_on(bbr->rc_tp)) {
2710 		union tcp_log_stackspecific log;
2711 		uint64_t ifp64 = (uintptr_t)ifp;
2712 
2713 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2714 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2715 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2716 		log.u_bbr.flex3 = ((ifp64  >> 32) & 0x00000000ffffffff);
2717 		log.u_bbr.flex4 = (ifp64 & 0x00000000ffffffff);
2718 		log.u_bbr.bw_inuse = rate;
2719 		log.u_bbr.flex5 = line;
2720 		log.u_bbr.flex6 = error;
2721 		log.u_bbr.flex8 = bbr->skip_gain;
2722 		log.u_bbr.flex8 <<= 1;
2723 		log.u_bbr.flex8 |= bbr->gain_is_limited;
2724 		log.u_bbr.flex8 <<= 1;
2725 		log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing;
2726 		log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
2727 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2728 		    &bbr->rc_inp->inp_socket->so_rcv,
2729 		    &bbr->rc_inp->inp_socket->so_snd,
2730 		    BBR_LOG_HDWR_PACE, 0,
2731 		    0, &log, false, &bbr->rc_tv);
2732 	}
2733 }
2734 
2735 static void
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)2736 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)
2737 {
2738 	if (tcp_bblogging_on(bbr->rc_tp)) {
2739 		union tcp_log_stackspecific log;
2740 
2741 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2742 		log.u_bbr.flex1 = slot;
2743 		log.u_bbr.flex2 = del_by;
2744 		log.u_bbr.flex3 = prev_delay;
2745 		log.u_bbr.flex4 = line;
2746 		log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val;
2747 		log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay;
2748 		log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags);
2749 		log.u_bbr.flex8 = bbr->rc_in_persist;
2750 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2751 		    &bbr->rc_inp->inp_socket->so_rcv,
2752 		    &bbr->rc_inp->inp_socket->so_snd,
2753 		    BBR_LOG_BBRSND, 0,
2754 		    len, &log, false, &bbr->rc_tv);
2755 	}
2756 }
2757 
2758 static void
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)2759 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)
2760 {
2761 	if (tcp_bblogging_on(bbr->rc_tp)) {
2762 		union tcp_log_stackspecific log;
2763 
2764 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2765 		log.u_bbr.flex1 = bbr->r_ctl.rc_delivered;
2766 		log.u_bbr.flex2 = 0;
2767 		log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt;
2768 		log.u_bbr.flex4 = end;
2769 		log.u_bbr.flex5 = seq;
2770 		log.u_bbr.flex6 = t;
2771 		log.u_bbr.flex7 = match;
2772 		log.u_bbr.flex8 = flags;
2773 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2774 		    &bbr->rc_inp->inp_socket->so_rcv,
2775 		    &bbr->rc_inp->inp_socket->so_snd,
2776 		    BBR_LOG_BBRRTT, 0,
2777 		    0, &log, false, &bbr->rc_tv);
2778 	}
2779 }
2780 
2781 static void
bbr_log_exit_gain(struct tcp_bbr * bbr,uint32_t cts,int32_t entry_method)2782 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method)
2783 {
2784 	if (tcp_bblogging_on(bbr->rc_tp)) {
2785 		union tcp_log_stackspecific log;
2786 
2787 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2788 		log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2789 		log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
2790 		log.u_bbr.flex3 = bbr->r_ctl.gain_epoch;
2791 		log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2792 		log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs;
2793 		log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight;
2794 		log.u_bbr.flex7 = 0;
2795 		log.u_bbr.flex8 = entry_method;
2796 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2797 		    &bbr->rc_inp->inp_socket->so_rcv,
2798 		    &bbr->rc_inp->inp_socket->so_snd,
2799 		    BBR_LOG_EXIT_GAIN, 0,
2800 		    0, &log, false, &bbr->rc_tv);
2801 	}
2802 }
2803 
2804 static void
bbr_log_settings_change(struct tcp_bbr * bbr,int settings_desired)2805 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired)
2806 {
2807 	if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) {
2808 		union tcp_log_stackspecific log;
2809 
2810 		bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2811 		/* R-HU */
2812 		log.u_bbr.flex1 = 0;
2813 		log.u_bbr.flex2 = 0;
2814 		log.u_bbr.flex3 = 0;
2815 		log.u_bbr.flex4 = 0;
2816 		log.u_bbr.flex7 = 0;
2817 		log.u_bbr.flex8 = settings_desired;
2818 
2819 		TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2820 		    &bbr->rc_inp->inp_socket->so_rcv,
2821 		    &bbr->rc_inp->inp_socket->so_snd,
2822 		    BBR_LOG_SETTINGS_CHG, 0,
2823 		    0, &log, false, &bbr->rc_tv);
2824 	}
2825 }
2826 
2827 /*
2828  * Returns the bw from the our filter.
2829  */
2830 static inline uint64_t
bbr_get_full_bw(struct tcp_bbr * bbr)2831 bbr_get_full_bw(struct tcp_bbr *bbr)
2832 {
2833 	uint64_t bw;
2834 
2835 	bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2836 
2837 	return (bw);
2838 }
2839 
2840 static inline void
bbr_set_pktepoch(struct tcp_bbr * bbr,uint32_t cts,int32_t line)2841 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2842 {
2843 	uint64_t calclr;
2844 	uint32_t lost, del;
2845 
2846 	if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch)
2847 		lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch;
2848 	else
2849 		lost = 0;
2850 	del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del;
2851 	if (lost == 0)  {
2852 		calclr = 0;
2853 	} else if (del) {
2854 		calclr = lost;
2855 		calclr *= (uint64_t)1000;
2856 		calclr /= (uint64_t)del;
2857 	} else {
2858 		/* Nothing delivered? 100.0% loss */
2859 		calclr = 1000;
2860 	}
2861 	bbr->r_ctl.rc_pkt_epoch_loss_rate =  (uint32_t)calclr;
2862 	if (IN_RECOVERY(bbr->rc_tp->t_flags))
2863 		bbr->r_ctl.recovery_lr += (uint32_t)calclr;
2864 	bbr->r_ctl.rc_pkt_epoch++;
2865 	if (bbr->rc_no_pacing &&
2866 	    (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) {
2867 		bbr->rc_no_pacing = 0;
2868 		tcp_bbr_tso_size_check(bbr, cts);
2869 	}
2870 	bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time);
2871 	bbr->r_ctl.rc_pkt_epoch_time = cts;
2872 	/* What was our loss rate */
2873 	bbr_log_pkt_epoch(bbr, cts, line, lost, del);
2874 	bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered;
2875 	bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost;
2876 }
2877 
2878 static inline void
bbr_set_epoch(struct tcp_bbr * bbr,uint32_t cts,int32_t line)2879 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2880 {
2881 	uint32_t epoch_time;
2882 
2883 	/* Tick the RTT clock */
2884 	bbr->r_ctl.rc_rtt_epoch++;
2885 	epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start;
2886 	bbr_log_time_epoch(bbr, cts, line, epoch_time);
2887 	bbr->r_ctl.rc_rcv_epoch_start = cts;
2888 }
2889 
2890 static inline void
bbr_isit_a_pkt_epoch(struct tcp_bbr * bbr,uint32_t cts,struct bbr_sendmap * rsm,int32_t line,int32_t cum_acked)2891 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked)
2892 {
2893 	if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) {
2894 		bbr->rc_is_pkt_epoch_now = 1;
2895 	}
2896 }
2897 
2898 /*
2899  * Returns the bw from either the b/w filter
2900  * or from the lt_bw (if the connection is being
2901  * policed).
2902  */
2903 static inline uint64_t
__bbr_get_bw(struct tcp_bbr * bbr)2904 __bbr_get_bw(struct tcp_bbr *bbr)
2905 {
2906 	uint64_t bw, min_bw;
2907 	uint64_t rtt;
2908 	int gm_measure_cnt = 1;
2909 
2910 	/*
2911 	 * For startup we make, like google, a
2912 	 * minimum b/w. This is generated from the
2913 	 * IW and the rttProp. We do fall back to srtt
2914 	 * if for some reason (initial handshake) we don't
2915 	 * have a rttProp. We, in the worst case, fall back
2916 	 * to the configured min_bw (rc_initial_hptsi_bw).
2917 	 */
2918 	if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
2919 		/* Attempt first to use rttProp */
2920 		rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2921 		if (rtt && (rtt < 0xffffffff)) {
2922 measure:
2923 			min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2924 				((uint64_t)1000000);
2925 			min_bw /= rtt;
2926 			if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2927 				min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2928 			}
2929 
2930 		} else if (bbr->rc_tp->t_srtt != 0) {
2931 			/* No rttProp, use srtt? */
2932 			rtt = bbr_get_rtt(bbr, BBR_SRTT);
2933 			goto measure;
2934 		} else {
2935 			min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2936 		}
2937 	} else
2938 		min_bw = 0;
2939 
2940 	if ((bbr->rc_past_init_win == 0) &&
2941 	    (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp)))
2942 		bbr->rc_past_init_win = 1;
2943 	if ((bbr->rc_use_google)  && (bbr->r_ctl.r_measurement_count >= 1))
2944 		gm_measure_cnt = 0;
2945 	if (gm_measure_cnt &&
2946 	    ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) ||
2947 	     (bbr->rc_past_init_win == 0))) {
2948 		/* For google we use our guess rate until we get 1 measurement */
2949 
2950 use_initial_window:
2951 		rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2952 		if (rtt && (rtt < 0xffffffff)) {
2953 			/*
2954 			 * We have an RTT measurement. Use that in
2955 			 * combination with our initial window to calculate
2956 			 * a b/w.
2957 			 */
2958 			bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2959 				((uint64_t)1000000);
2960 			bw /= rtt;
2961 			if (bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2962 				bw = bbr->r_ctl.rc_initial_hptsi_bw;
2963 			}
2964 		} else {
2965 			/* Drop back to the 40 and punt to a default */
2966 			bw = bbr->r_ctl.rc_initial_hptsi_bw;
2967 		}
2968 		if (bw < 1)
2969 			/* Probably should panic */
2970 			bw = 1;
2971 		if (bw > min_bw)
2972 			return (bw);
2973 		else
2974 			return (min_bw);
2975 	}
2976 	if (bbr->rc_lt_use_bw)
2977 		bw = bbr->r_ctl.rc_lt_bw;
2978 	else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0))
2979 		bw = bbr->r_ctl.red_bw;
2980 	else
2981 		bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2982 	if (bw == 0) {
2983 		/* We should not be at 0, go to the initial window then  */
2984 		goto use_initial_window;
2985 	}
2986 	if (bw < min_bw)
2987 		bw = min_bw;
2988 	return (bw);
2989 }
2990 
2991 static inline uint64_t
bbr_get_bw(struct tcp_bbr * bbr)2992 bbr_get_bw(struct tcp_bbr *bbr)
2993 {
2994 	uint64_t bw;
2995 
2996 	bw = __bbr_get_bw(bbr);
2997 	return (bw);
2998 }
2999 
3000 static inline void
bbr_reset_lt_bw_interval(struct tcp_bbr * bbr,uint32_t cts)3001 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts)
3002 {
3003 	bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch;
3004 	bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time;
3005 	bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3006 	bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3007 }
3008 
3009 static inline void
bbr_reset_lt_bw_sampling(struct tcp_bbr * bbr,uint32_t cts)3010 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts)
3011 {
3012 	bbr->rc_lt_is_sampling = 0;
3013 	bbr->rc_lt_use_bw = 0;
3014 	bbr->r_ctl.rc_lt_bw = 0;
3015 	bbr_reset_lt_bw_interval(bbr, cts);
3016 }
3017 
3018 static inline void
bbr_lt_bw_samp_done(struct tcp_bbr * bbr,uint64_t bw,uint32_t cts,uint32_t timin)3019 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin)
3020 {
3021 	uint64_t diff;
3022 
3023 	/* Do we have a previous sample? */
3024 	if (bbr->r_ctl.rc_lt_bw) {
3025 		/* Get the diff in bytes per second */
3026 		if (bbr->r_ctl.rc_lt_bw > bw)
3027 			diff = bbr->r_ctl.rc_lt_bw - bw;
3028 		else
3029 			diff = bw - bbr->r_ctl.rc_lt_bw;
3030 		if ((diff <= bbr_lt_bw_diff) ||
3031 		    (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) {
3032 			/* Consider us policed */
3033 			uint32_t saved_bw;
3034 
3035 			saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw;
3036 			bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2;	/* average of two */
3037 			bbr->rc_lt_use_bw = 1;
3038 			bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
3039 			/*
3040 			 * Use pkt based epoch for measuring length of
3041 			 * policer up
3042 			 */
3043 			bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch;
3044 			/*
3045 			 * reason 4 is we need to start consider being
3046 			 * policed
3047 			 */
3048 			bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin);
3049 			return;
3050 		}
3051 	}
3052 	bbr->r_ctl.rc_lt_bw = bw;
3053 	bbr_reset_lt_bw_interval(bbr, cts);
3054 	bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin);
3055 }
3056 
3057 static void
bbr_randomize_extra_state_time(struct tcp_bbr * bbr)3058 bbr_randomize_extra_state_time(struct tcp_bbr *bbr)
3059 {
3060 	uint32_t ran, deduct;
3061 
3062 	ran = arc4random_uniform(bbr_rand_ot);
3063 	if (ran) {
3064 		deduct = bbr->r_ctl.rc_level_state_extra / ran;
3065 		bbr->r_ctl.rc_level_state_extra -= deduct;
3066 	}
3067 }
3068 /*
3069  * Return randomly the starting state
3070  * to use in probebw.
3071  */
3072 static uint8_t
bbr_pick_probebw_substate(struct tcp_bbr * bbr,uint32_t cts)3073 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts)
3074 {
3075 	uint32_t ran;
3076 	uint8_t ret_val;
3077 
3078 	/* Initialize the offset to 0 */
3079 	bbr->r_ctl.rc_exta_time_gd = 0;
3080 	bbr->rc_hit_state_1 = 0;
3081 	bbr->r_ctl.rc_level_state_extra = 0;
3082 	ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1));
3083 	/*
3084 	 * The math works funny here :) the return value is used to set the
3085 	 * substate and then the state change is called which increments by
3086 	 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when
3087 	 * we fully enter the state. Note that the (8 - 1 - ran) assures that
3088 	 * we return 1 - 7, so we dont return 0 and end up starting in
3089 	 * state 1 (DRAIN).
3090 	 */
3091 	ret_val = BBR_SUBSTATE_COUNT - 1 - ran;
3092 	/* Set an epoch */
3093 	if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP))
3094 		bbr_set_epoch(bbr, cts, __LINE__);
3095 
3096 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
3097 	return (ret_val);
3098 }
3099 
3100 static void
bbr_lt_bw_sampling(struct tcp_bbr * bbr,uint32_t cts,int32_t loss_detected)3101 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected)
3102 {
3103 	uint32_t diff, d_time;
3104 	uint64_t del_time, bw, lost, delivered;
3105 
3106 	if (bbr->r_use_policer == 0)
3107 		return;
3108 	if (bbr->rc_lt_use_bw) {
3109 		/* We are using lt bw do we stop yet? */
3110 		diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
3111 		if (diff > bbr_lt_bw_max_rtts) {
3112 			/* Reset it all */
3113 reset_all:
3114 			bbr_reset_lt_bw_sampling(bbr, cts);
3115 			if (bbr->rc_filled_pipe) {
3116 				bbr_set_epoch(bbr, cts, __LINE__);
3117 				bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
3118 				bbr_substate_change(bbr, cts, __LINE__, 0);
3119 				bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
3120 				bbr_log_type_statechange(bbr, cts, __LINE__);
3121 			} else {
3122 				/*
3123 				 * This should not happen really
3124 				 * unless we remove the startup/drain
3125 				 * restrictions above.
3126 				 */
3127 				bbr->rc_bbr_state = BBR_STATE_STARTUP;
3128 				bbr_set_epoch(bbr, cts, __LINE__);
3129 				bbr->r_ctl.rc_bbr_state_time = cts;
3130 				bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
3131 				bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
3132 				bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
3133 				bbr_set_state_target(bbr, __LINE__);
3134 				bbr_log_type_statechange(bbr, cts, __LINE__);
3135 			}
3136 			/* reason 0 is to stop using lt-bw */
3137 			bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0);
3138 			return;
3139 		}
3140 		if (bbr_lt_intvl_fp == 0) {
3141 			/* Not doing false-positive detection */
3142 			return;
3143 		}
3144 		/* False positive detection */
3145 		if (diff == bbr_lt_intvl_fp) {
3146 			/* At bbr_lt_intvl_fp we record the lost */
3147 			bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3148 			bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3149 		} else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) {
3150 			/* Now is our loss rate still high? */
3151 			lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3152 			delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3153 			if ((delivered == 0) ||
3154 			    (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) {
3155 				/* No still below our threshold */
3156 				bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0);
3157 			} else {
3158 				/* Yikes its still high, it must be a false positive */
3159 				bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0);
3160 				goto reset_all;
3161 			}
3162 		}
3163 		return;
3164 	}
3165 	/*
3166 	 * Wait for the first loss before sampling, to let the policer
3167 	 * exhaust its tokens and estimate the steady-state rate allowed by
3168 	 * the policer. Starting samples earlier includes bursts that
3169 	 * over-estimate the bw.
3170 	 */
3171 	if (bbr->rc_lt_is_sampling == 0) {
3172 		/* reason 1 is to begin doing the sampling  */
3173 		if (loss_detected == 0)
3174 			return;
3175 		bbr_reset_lt_bw_interval(bbr, cts);
3176 		bbr->rc_lt_is_sampling = 1;
3177 		bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0);
3178 		return;
3179 	}
3180 	/* Now how long were we delivering long term last> */
3181 	if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time))
3182 		d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time;
3183 	else
3184 		d_time = 0;
3185 
3186 	/* To avoid underestimates, reset sampling if we run out of data. */
3187 	if (bbr->r_ctl.r_app_limited_until) {
3188 		/* Can not measure in app-limited state */
3189 		bbr_reset_lt_bw_sampling(bbr, cts);
3190 		/* reason 2 is to reset sampling due to app limits  */
3191 		bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time);
3192 		return;
3193 	}
3194 	diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
3195 	if (diff < bbr_lt_intvl_min_rtts) {
3196 		/*
3197 		 * need more samples (we don't
3198 		 * start on a round like linux so
3199 		 * we need 1 more).
3200 		 */
3201 		/* 6 is not_enough time or no-loss */
3202 		bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3203 		return;
3204 	}
3205 	if (diff > (4 * bbr_lt_intvl_min_rtts)) {
3206 		/*
3207 		 * For now if we wait too long, reset all sampling. We need
3208 		 * to do some research here, its possible that we should
3209 		 * base this on how much loss as occurred.. something like
3210 		 * if its under 10% (or some thresh) reset all otherwise
3211 		 * don't.  Thats for phase II I guess.
3212 		 */
3213 		bbr_reset_lt_bw_sampling(bbr, cts);
3214  		/* reason 3 is to reset sampling due too long of sampling */
3215 		bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3216 		return;
3217 	}
3218 	/*
3219 	 * End sampling interval when a packet is lost, so we estimate the
3220 	 * policer tokens were exhausted. Stopping the sampling before the
3221 	 * tokens are exhausted under-estimates the policed rate.
3222 	 */
3223 	if (loss_detected == 0) {
3224 		/* 6 is not_enough time or no-loss */
3225 		bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3226 		return;
3227 	}
3228 	/* Calculate packets lost and delivered in sampling interval. */
3229 	lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3230 	delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3231 	if ((delivered == 0) ||
3232 	    (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) {
3233 		bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time);
3234 		return;
3235 	}
3236 	if (d_time < 1000) {
3237 		/* Not enough time. wait */
3238 		/* 6 is not_enough time or no-loss */
3239 		bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3240 		return;
3241 	}
3242 	if (d_time >= (0xffffffff / USECS_IN_MSEC)) {
3243 		/* Too long */
3244 		bbr_reset_lt_bw_sampling(bbr, cts);
3245  		/* reason 3 is to reset sampling due too long of sampling */
3246 		bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3247 		return;
3248 	}
3249 	del_time = d_time;
3250 	bw = delivered;
3251 	bw *= (uint64_t)USECS_IN_SECOND;
3252 	bw /= del_time;
3253 	bbr_lt_bw_samp_done(bbr, bw, cts, d_time);
3254 }
3255 
3256 /*
3257  * Allocate a sendmap from our zone.
3258  */
3259 static struct bbr_sendmap *
bbr_alloc(struct tcp_bbr * bbr)3260 bbr_alloc(struct tcp_bbr *bbr)
3261 {
3262 	struct bbr_sendmap *rsm;
3263 
3264 	BBR_STAT_INC(bbr_to_alloc);
3265 	rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO));
3266 	if (rsm) {
3267 		bbr->r_ctl.rc_num_maps_alloced++;
3268 		return (rsm);
3269 	}
3270 	if (bbr->r_ctl.rc_free_cnt) {
3271 		BBR_STAT_INC(bbr_to_alloc_emerg);
3272 		rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
3273 		TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
3274 		bbr->r_ctl.rc_free_cnt--;
3275 		return (rsm);
3276 	}
3277 	BBR_STAT_INC(bbr_to_alloc_failed);
3278 	return (NULL);
3279 }
3280 
3281 static struct bbr_sendmap *
bbr_alloc_full_limit(struct tcp_bbr * bbr)3282 bbr_alloc_full_limit(struct tcp_bbr *bbr)
3283 {
3284 	if ((V_tcp_map_entries_limit > 0) &&
3285 	    (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3286 		BBR_STAT_INC(bbr_alloc_limited);
3287 		if (!bbr->alloc_limit_reported) {
3288 			bbr->alloc_limit_reported = 1;
3289 			BBR_STAT_INC(bbr_alloc_limited_conns);
3290 		}
3291 		return (NULL);
3292 	}
3293 	return (bbr_alloc(bbr));
3294 }
3295 
3296 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3297 static struct bbr_sendmap *
bbr_alloc_limit(struct tcp_bbr * bbr,uint8_t limit_type)3298 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type)
3299 {
3300 	struct bbr_sendmap *rsm;
3301 
3302 	if (limit_type) {
3303 		/* currently there is only one limit type */
3304 		if (V_tcp_map_split_limit > 0 &&
3305 		    bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
3306 			BBR_STAT_INC(bbr_split_limited);
3307 			if (!bbr->alloc_limit_reported) {
3308 				bbr->alloc_limit_reported = 1;
3309 				BBR_STAT_INC(bbr_alloc_limited_conns);
3310 			}
3311 			return (NULL);
3312 		}
3313 	}
3314 
3315 	/* allocate and mark in the limit type, if set */
3316 	rsm = bbr_alloc(bbr);
3317 	if (rsm != NULL && limit_type) {
3318 		rsm->r_limit_type = limit_type;
3319 		bbr->r_ctl.rc_num_split_allocs++;
3320 	}
3321 	return (rsm);
3322 }
3323 
3324 static void
bbr_free(struct tcp_bbr * bbr,struct bbr_sendmap * rsm)3325 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
3326 {
3327 	if (rsm->r_limit_type) {
3328 		/* currently there is only one limit type */
3329 		bbr->r_ctl.rc_num_split_allocs--;
3330 	}
3331 	if (rsm->r_is_smallmap)
3332 		bbr->r_ctl.rc_num_small_maps_alloced--;
3333 	if (bbr->r_ctl.rc_tlp_send == rsm)
3334 		bbr->r_ctl.rc_tlp_send = NULL;
3335 	if (bbr->r_ctl.rc_resend == rsm) {
3336 		bbr->r_ctl.rc_resend = NULL;
3337 	}
3338 	if (bbr->r_ctl.rc_next == rsm)
3339 		bbr->r_ctl.rc_next = NULL;
3340 	if (bbr->r_ctl.rc_sacklast == rsm)
3341 		bbr->r_ctl.rc_sacklast = NULL;
3342 	if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
3343 		memset(rsm, 0, sizeof(struct bbr_sendmap));
3344 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
3345 		rsm->r_limit_type = 0;
3346 		bbr->r_ctl.rc_free_cnt++;
3347 		return;
3348 	}
3349 	bbr->r_ctl.rc_num_maps_alloced--;
3350 	uma_zfree(bbr_zone, rsm);
3351 }
3352 
3353 /*
3354  * Returns the BDP.
3355  */
3356 static uint64_t
bbr_get_bw_delay_prod(uint64_t rtt,uint64_t bw)3357 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) {
3358 	/*
3359 	 * Calculate the bytes in flight needed given the bw (in bytes per
3360 	 * second) and the specifyed rtt in useconds. We need to put out the
3361 	 * returned value per RTT to match that rate. Gain will normally
3362 	 * raise it up from there.
3363 	 *
3364 	 * This should not overflow as long as the bandwidth is below 1
3365 	 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller
3366 	 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30).
3367 	 */
3368 	uint64_t usec_per_sec;
3369 
3370 	usec_per_sec = USECS_IN_SECOND;
3371 	return ((rtt * bw) / usec_per_sec);
3372 }
3373 
3374 /*
3375  * Return the initial cwnd.
3376  */
3377 static uint32_t
bbr_initial_cwnd(struct tcp_bbr * bbr,struct tcpcb * tp)3378 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp)
3379 {
3380 	uint32_t i_cwnd;
3381 
3382 	if (bbr->rc_init_win) {
3383 		i_cwnd = bbr->rc_init_win * tp->t_maxseg;
3384 	} else if (V_tcp_initcwnd_segments)
3385 		i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg),
3386 		    max(2 * tp->t_maxseg, 14600));
3387 	else if (V_tcp_do_rfc3390)
3388 		i_cwnd = min(4 * tp->t_maxseg,
3389 		    max(2 * tp->t_maxseg, 4380));
3390 	else {
3391 		/* Per RFC5681 Section 3.1 */
3392 		if (tp->t_maxseg > 2190)
3393 			i_cwnd = 2 * tp->t_maxseg;
3394 		else if (tp->t_maxseg > 1095)
3395 			i_cwnd = 3 * tp->t_maxseg;
3396 		else
3397 			i_cwnd = 4 * tp->t_maxseg;
3398 	}
3399 	return (i_cwnd);
3400 }
3401 
3402 /*
3403  * Given a specified gain, return the target
3404  * cwnd based on that gain.
3405  */
3406 static uint32_t
bbr_get_raw_target_cwnd(struct tcp_bbr * bbr,uint32_t gain,uint64_t bw)3407 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw)
3408 {
3409 	uint64_t bdp, rtt;
3410 	uint32_t cwnd;
3411 
3412 	if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) ||
3413 	    (bbr_get_full_bw(bbr) == 0)) {
3414 		/* No measurements yet */
3415 		return (bbr_initial_cwnd(bbr, bbr->rc_tp));
3416 	}
3417 	/*
3418 	 * Get bytes per RTT needed (rttProp is normally in
3419 	 * bbr_cwndtarget_rtt_touse)
3420 	 */
3421 	rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse);
3422 	/* Get the bdp from the two values */
3423 	bdp = bbr_get_bw_delay_prod(rtt, bw);
3424 	/* Now apply the gain */
3425 	cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT));
3426 
3427 	return (cwnd);
3428 }
3429 
3430 static uint32_t
bbr_get_target_cwnd(struct tcp_bbr * bbr,uint64_t bw,uint32_t gain)3431 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain)
3432 {
3433 	uint32_t cwnd, mss;
3434 
3435 	mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
3436 	/* Get the base cwnd with gain rounded to a mss */
3437 	cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss);
3438 	/*
3439 	 * Add in N (2 default since we do not have a
3440 	 * fq layer to trap packets in) quanta's per the I-D
3441 	 * section 4.2.3.2 quanta adjust.
3442 	 */
3443 	cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs);
3444 	if (bbr->rc_use_google) {
3445 		if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3446 		   (bbr_state_val(bbr) == BBR_SUB_GAIN)) {
3447 			/*
3448 			 * The linux implementation adds
3449 			 * an extra 2 x mss in gain cycle which
3450 			 * is documented no-where except in the code.
3451 			 * so we add more for Neal undocumented feature
3452 			 */
3453 			cwnd += 2 * mss;
3454 		}
3455  		if ((cwnd / mss) & 0x1) {
3456 			/* Round up for odd num mss */
3457 			cwnd += mss;
3458 		}
3459 	}
3460 	/* Are we below the min cwnd? */
3461 	if (cwnd < get_min_cwnd(bbr))
3462 		return (get_min_cwnd(bbr));
3463 	return (cwnd);
3464 }
3465 
3466 static uint16_t
bbr_gain_adjust(struct tcp_bbr * bbr,uint16_t gain)3467 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain)
3468 {
3469 	if (gain < 1)
3470 		gain = 1;
3471 	return (gain);
3472 }
3473 
3474 static uint32_t
bbr_get_header_oh(struct tcp_bbr * bbr)3475 bbr_get_header_oh(struct tcp_bbr *bbr)
3476 {
3477 	int seg_oh;
3478 
3479 	seg_oh = 0;
3480 	if (bbr->r_ctl.rc_inc_tcp_oh) {
3481 		/* Do we include TCP overhead? */
3482 		seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr));
3483 	}
3484 	if (bbr->r_ctl.rc_inc_ip_oh) {
3485 		/* Do we include IP overhead? */
3486 #ifdef INET6
3487 		if (bbr->r_is_v6) {
3488 			seg_oh += sizeof(struct ip6_hdr);
3489 		} else
3490 #endif
3491 		{
3492 
3493 #ifdef INET
3494 			seg_oh += sizeof(struct ip);
3495 #endif
3496 		}
3497 	}
3498 	if (bbr->r_ctl.rc_inc_enet_oh) {
3499 		/* Do we include the ethernet overhead?  */
3500 		seg_oh += sizeof(struct ether_header);
3501 	}
3502 	return(seg_oh);
3503 }
3504 
3505 static uint32_t
bbr_get_pacing_length(struct tcp_bbr * bbr,uint16_t gain,uint32_t useconds_time,uint64_t bw)3506 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw)
3507 {
3508 	uint64_t divor, res, tim;
3509 
3510 	if (useconds_time == 0)
3511 		return (0);
3512 	gain = bbr_gain_adjust(bbr, gain);
3513 	divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT;
3514 	tim = useconds_time;
3515 	res = (tim * bw * gain) / divor;
3516 	if (res == 0)
3517 		res = 1;
3518 	return ((uint32_t)res);
3519 }
3520 
3521 /*
3522  * Given a gain and a length return the delay in useconds that
3523  * should be used to evenly space out packets
3524  * on the connection (based on the gain factor).
3525  */
3526 static uint32_t
bbr_get_pacing_delay(struct tcp_bbr * bbr,uint16_t gain,int32_t len,uint32_t cts,int nolog)3527 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog)
3528 {
3529 	uint64_t bw, lentim, res;
3530 	uint32_t usecs, srtt, over = 0;
3531 	uint32_t seg_oh, num_segs, maxseg;
3532 
3533 	if (len == 0)
3534 		return (0);
3535 
3536 	maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
3537 	num_segs = (len + maxseg - 1) / maxseg;
3538 	if (bbr->rc_use_google == 0) {
3539 		seg_oh = bbr_get_header_oh(bbr);
3540 		len += (num_segs * seg_oh);
3541 	}
3542 	gain = bbr_gain_adjust(bbr, gain);
3543 	bw = bbr_get_bw(bbr);
3544 	if (bbr->rc_use_google) {
3545 		uint64_t cbw;
3546 
3547 		/*
3548 		 * Reduce the b/w by the google discount
3549 		 * factor 10 = 1%.
3550 		 */
3551 		cbw = bw *  (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount);
3552 		cbw /= (uint64_t)1000;
3553 		/* We don't apply a discount if it results in 0 */
3554 		if (cbw > 0)
3555 			bw = cbw;
3556 	}
3557 	lentim = ((uint64_t)len *
3558 		  (uint64_t)USECS_IN_SECOND *
3559 		  (uint64_t)BBR_UNIT);
3560 	res = lentim / ((uint64_t)gain * bw);
3561 	if (res == 0)
3562 		res = 1;
3563 	usecs = (uint32_t)res;
3564 	srtt = bbr_get_rtt(bbr, BBR_SRTT);
3565 	if (bbr_hptsi_max_mul && bbr_hptsi_max_div &&
3566 	    (bbr->rc_use_google == 0) &&
3567 	    (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) {
3568 		/*
3569 		 * We cannot let the delay be more than 1/2 the srtt time.
3570 		 * Otherwise we cannot pace out or send properly.
3571 		 */
3572 		over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div;
3573 		BBR_STAT_INC(bbr_hpts_min_time);
3574 	}
3575 	if (!nolog)
3576 		bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1);
3577 	return (usecs);
3578 }
3579 
3580 static void
bbr_ack_received(struct tcpcb * tp,struct tcp_bbr * bbr,struct tcphdr * th,uint32_t bytes_this_ack,uint32_t sack_changed,uint32_t prev_acked,int32_t line,uint32_t losses)3581 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack,
3582 		 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses)
3583 {
3584 	uint64_t bw;
3585 	uint32_t cwnd, target_cwnd, saved_bytes, maxseg;
3586 	int32_t meth;
3587 
3588 	INP_WLOCK_ASSERT(tptoinpcb(tp));
3589 
3590 #ifdef STATS
3591 	if ((tp->t_flags & TF_GPUTINPROG) &&
3592 	    SEQ_GEQ(th->th_ack, tp->gput_ack)) {
3593 		/*
3594 		 * Strech acks and compressed acks will cause this to
3595 		 * oscillate but we are doing it the same way as the main
3596 		 * stack so it will be compariable (though possibly not
3597 		 * ideal).
3598 		 */
3599 		int32_t cgput;
3600 		int64_t gput, time_stamp;
3601 
3602 		gput = (int64_t) (th->th_ack - tp->gput_seq) * 8;
3603 		time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000));
3604 		cgput = gput / time_stamp;
3605 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3606 					 cgput);
3607 		if (tp->t_stats_gput_prev > 0)
3608 			stats_voi_update_abs_s32(tp->t_stats,
3609 						 VOI_TCP_GPUT_ND,
3610 						 ((gput - tp->t_stats_gput_prev) * 100) /
3611 						 tp->t_stats_gput_prev);
3612 		tp->t_flags &= ~TF_GPUTINPROG;
3613 		tp->t_stats_gput_prev = cgput;
3614 	}
3615 #endif
3616 	if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3617 	    ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
3618 		/* We don't change anything in probe-rtt */
3619 		return;
3620 	}
3621 	maxseg = tp->t_maxseg - bbr->rc_last_options;
3622 	saved_bytes = bytes_this_ack;
3623 	bytes_this_ack += sack_changed;
3624 	if (bytes_this_ack > prev_acked) {
3625 		bytes_this_ack -= prev_acked;
3626 		/*
3627 		 * A byte ack'd gives us a full mss
3628 		 * to be like linux i.e. they count packets.
3629 		 */
3630 		if ((bytes_this_ack < maxseg) && bbr->rc_use_google)
3631 			bytes_this_ack = maxseg;
3632 	} else {
3633 		/* Unlikely */
3634 		bytes_this_ack = 0;
3635 	}
3636 	cwnd = tp->snd_cwnd;
3637 	bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3638 	if (bw)
3639 		target_cwnd = bbr_get_target_cwnd(bbr,
3640 						  bw,
3641 						  (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain);
3642 	else
3643 		target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp);
3644 	if (IN_RECOVERY(tp->t_flags) &&
3645 	    (bbr->bbr_prev_in_rec == 0)) {
3646 		/*
3647 		 * We are entering recovery and
3648 		 * thus packet conservation.
3649 		 */
3650 		bbr->pkt_conservation = 1;
3651 		bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime;
3652 		cwnd = ctf_flight_size(tp,
3653 				       (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3654 			bytes_this_ack;
3655 	}
3656 	if (IN_RECOVERY(tp->t_flags)) {
3657 		uint32_t flight;
3658 
3659 		bbr->bbr_prev_in_rec = 1;
3660 		if (cwnd > losses) {
3661 			cwnd -= losses;
3662 			if (cwnd < maxseg)
3663 				cwnd = maxseg;
3664 		} else
3665 			cwnd = maxseg;
3666 		flight = ctf_flight_size(tp,
3667 					 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3668 		bbr_log_type_cwndupd(bbr, flight, 0,
3669 				     losses, 10, 0, 0, line);
3670 		if (bbr->pkt_conservation) {
3671 			uint32_t time_in;
3672 
3673 			if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start))
3674 				time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start;
3675 			else
3676 				time_in = 0;
3677 
3678 			if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
3679 				/* Clear packet conservation after an rttProp */
3680 				bbr->pkt_conservation = 0;
3681 			} else {
3682 				if ((flight + bytes_this_ack) > cwnd)
3683 					cwnd = flight + bytes_this_ack;
3684 				if (cwnd < get_min_cwnd(bbr))
3685 					cwnd = get_min_cwnd(bbr);
3686 				tp->snd_cwnd = cwnd;
3687 				bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed,
3688 						     prev_acked, 1, target_cwnd, th->th_ack, line);
3689 				return;
3690 			}
3691 		}
3692 	} else
3693 		bbr->bbr_prev_in_rec = 0;
3694 	if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) {
3695 		bbr->r_ctl.restrict_growth--;
3696 		if (bytes_this_ack > maxseg)
3697 			bytes_this_ack = maxseg;
3698 	}
3699 	if (bbr->rc_filled_pipe) {
3700 		/*
3701 		 * Here we have exited startup and filled the pipe. We will
3702 		 * thus allow the cwnd to shrink to the target. We hit here
3703 		 * mostly.
3704 		 */
3705 		uint32_t s_cwnd;
3706 
3707 		meth = 2;
3708 		s_cwnd = min((cwnd + bytes_this_ack), target_cwnd);
3709 		if (s_cwnd > cwnd)
3710 			cwnd = s_cwnd;
3711 		else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing)
3712 			cwnd = s_cwnd;
3713 	} else {
3714 		/*
3715 		 * Here we are still in startup, we increase cwnd by what
3716 		 * has been acked.
3717 		 */
3718 		if ((cwnd < target_cwnd) ||
3719 		    (bbr->rc_past_init_win == 0)) {
3720 			meth = 3;
3721 			cwnd += bytes_this_ack;
3722 		} else {
3723 			/*
3724 			 * Method 4 means we are at target so no gain in
3725 			 * startup and past the initial window.
3726 			 */
3727 			meth = 4;
3728 		}
3729 	}
3730 	tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr));
3731 	bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line);
3732 }
3733 
3734 static void
tcp_bbr_partialack(struct tcpcb * tp)3735 tcp_bbr_partialack(struct tcpcb *tp)
3736 {
3737 	struct tcp_bbr *bbr;
3738 
3739 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3740 	INP_WLOCK_ASSERT(tptoinpcb(tp));
3741 	if (ctf_flight_size(tp,
3742 		(bbr->r_ctl.rc_sacked  + bbr->r_ctl.rc_lost_bytes)) <=
3743 	    tp->snd_cwnd) {
3744 		bbr->r_wanted_output = 1;
3745 	}
3746 }
3747 
3748 static void
bbr_post_recovery(struct tcpcb * tp)3749 bbr_post_recovery(struct tcpcb *tp)
3750 {
3751 	struct tcp_bbr *bbr;
3752 	uint32_t  flight;
3753 
3754 	INP_WLOCK_ASSERT(tptoinpcb(tp));
3755 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3756 	/*
3757 	 * Here we just exit recovery.
3758 	 */
3759 	EXIT_RECOVERY(tp->t_flags);
3760 	/* Lock in our b/w reduction for the specified number of pkt-epochs */
3761 	bbr->r_recovery_bw = 0;
3762 	tp->snd_recover = tp->snd_una;
3763 	tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3764 	bbr->pkt_conservation = 0;
3765 	if (bbr->rc_use_google == 0) {
3766 		/*
3767 		 * For non-google mode lets
3768 		 * go ahead and make sure we clear
3769 		 * the recovery state so if we
3770 		 * bounce back in to recovery we
3771 		 * will do PC.
3772 		 */
3773 		bbr->bbr_prev_in_rec = 0;
3774 	}
3775 	bbr_log_type_exit_rec(bbr);
3776 	if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3777 		tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
3778 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__);
3779 	} else {
3780 		/* For probe-rtt case lets fix up its saved_cwnd */
3781 		if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) {
3782 			bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent;
3783 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__);
3784 		}
3785 	}
3786 	flight = ctf_flight_size(tp,
3787 		     (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3788 	if ((bbr->rc_use_google == 0) &&
3789 	    bbr_do_red) {
3790 		uint64_t val, lr2use;
3791 		uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd;
3792 		uint32_t *cwnd_p;
3793 
3794 		if (bbr_get_rtt(bbr, BBR_SRTT)) {
3795 			val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000);
3796 			val /= bbr_get_rtt(bbr, BBR_SRTT);
3797 			ratio = (uint32_t)val;
3798 		} else
3799 			ratio = 1000;
3800 
3801 		bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div,
3802 				     bbr->r_ctl.recovery_lr, 21,
3803 				     ratio,
3804 				     bbr->r_ctl.rc_red_cwnd_pe,
3805 				     __LINE__);
3806 		if ((ratio < bbr_do_red) || (bbr_do_red == 0))
3807 			goto done;
3808 		if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3809 		     bbr_prtt_slam_cwnd) ||
3810 		    (bbr_sub_drain_slam_cwnd &&
3811 		     (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3812 		     bbr->rc_hit_state_1 &&
3813 		     (bbr_state_val(bbr) == BBR_SUB_DRAIN)) ||
3814 		    ((bbr->rc_bbr_state == BBR_STATE_DRAIN) &&
3815 		     bbr_slam_cwnd_in_main_drain)) {
3816 			/*
3817 			 * Here we must poke at the saved cwnd
3818 			 * as well as the cwnd.
3819 			 */
3820 			cwnd = bbr->r_ctl.rc_saved_cwnd;
3821 			cwnd_p = &bbr->r_ctl.rc_saved_cwnd;
3822 		} else {
3823  			cwnd = tp->snd_cwnd;
3824 			cwnd_p = &tp->snd_cwnd;
3825 		}
3826 		maxseg = tp->t_maxseg - bbr->rc_last_options;
3827 		/* Add the overall lr with the recovery lr */
3828 		if (bbr->r_ctl.rc_lost == 0)
3829 			lr2use = 0;
3830 		else if (bbr->r_ctl.rc_delivered == 0)
3831 			lr2use = 1000;
3832 		else {
3833 			lr2use = (uint64_t)bbr->r_ctl.rc_lost * (uint64_t)1000;
3834 			lr2use /= bbr->r_ctl.rc_delivered;
3835 		}
3836 		lr2use += bbr->r_ctl.recovery_lr;
3837 		acks_inflight = (flight / (maxseg * 2));
3838 		if (bbr_red_scale) {
3839 			lr2use *= bbr_get_rtt(bbr, BBR_SRTT);
3840 			lr2use /= bbr_red_scale;
3841 			if ((bbr_red_growth_restrict) &&
3842 			    ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1))
3843 			    bbr->r_ctl.restrict_growth += acks_inflight;
3844 		}
3845 		if (lr2use) {
3846 			val = (uint64_t)cwnd * lr2use;
3847 			val /= 1000;
3848 			if (cwnd > val)
3849 				newcwnd = roundup((cwnd - val), maxseg);
3850 			else
3851 				newcwnd = maxseg;
3852 		} else {
3853 			val = (uint64_t)cwnd * (uint64_t)bbr_red_mul;
3854 			val /= (uint64_t)bbr_red_div;
3855 			newcwnd = roundup((uint32_t)val, maxseg);
3856 		}
3857 		/* with standard delayed acks how many acks can I expect? */
3858 		if (bbr_drop_limit == 0) {
3859 			/*
3860 			 * Anticpate how much we will
3861 			 * raise the cwnd based on the acks.
3862 			 */
3863 			if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) {
3864 				/* We do enforce the min (with the acks) */
3865 				newcwnd = (get_min_cwnd(bbr) - acks_inflight);
3866 			}
3867 		} else {
3868 			/*
3869 			 * A strict drop limit of N is inplace
3870 			 */
3871 			if (newcwnd < (bbr_drop_limit * maxseg)) {
3872 				newcwnd = bbr_drop_limit * maxseg;
3873 			}
3874 		}
3875 		/* For the next N acks do we restrict the growth */
3876 		*cwnd_p = newcwnd;
3877 		if (tp->snd_cwnd > newcwnd)
3878 			tp->snd_cwnd = newcwnd;
3879 		bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22,
3880 				     (uint32_t)lr2use,
3881 				     bbr_get_rtt(bbr, BBR_SRTT), __LINE__);
3882 		bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch;
3883 	}
3884 done:
3885 	bbr->r_ctl.recovery_lr = 0;
3886 	if (flight <= tp->snd_cwnd) {
3887 		bbr->r_wanted_output = 1;
3888 	}
3889 	tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3890 }
3891 
3892 static void
bbr_setup_red_bw(struct tcp_bbr * bbr,uint32_t cts)3893 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts)
3894 {
3895 	bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3896 	/* Limit the drop in b/w to 1/2 our current filter. */
3897 	if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate)
3898 		bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate;
3899 	if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2))
3900 		bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2;
3901 	tcp_bbr_tso_size_check(bbr, cts);
3902 }
3903 
3904 static void
bbr_cong_signal(struct tcpcb * tp,struct tcphdr * th,uint32_t type,struct bbr_sendmap * rsm)3905 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm)
3906 {
3907 	struct tcp_bbr *bbr;
3908 
3909 	INP_WLOCK_ASSERT(tptoinpcb(tp));
3910 #ifdef STATS
3911 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
3912 #endif
3913 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3914 	switch (type) {
3915 	case CC_NDUPACK:
3916 		if (!IN_RECOVERY(tp->t_flags)) {
3917 			tp->snd_recover = tp->snd_max;
3918 			/* Start a new epoch */
3919 			bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
3920 			if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) {
3921 				/*
3922 				 * Move forward the lt epoch
3923 				 * so it won't count the truncated
3924 				 * epoch.
3925 				 */
3926 				bbr->r_ctl.rc_lt_epoch++;
3927 			}
3928 			if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
3929 				/*
3930 				 * Just like the policer detection code
3931 				 * if we are in startup we must push
3932 				 * forward the last startup epoch
3933 				 * to hide the truncated PE.
3934 				 */
3935 				bbr->r_ctl.rc_bbr_last_startup_epoch++;
3936 			}
3937 			bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd;
3938 			ENTER_RECOVERY(tp->t_flags);
3939 			bbr->rc_tlp_rtx_out = 0;
3940 			bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate;
3941 			tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3942 			if (tcp_in_hpts(bbr->rc_tp) &&
3943 			    ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) {
3944 				/*
3945 				 * When we enter recovery, we need to restart
3946 				 * any timers. This may mean we gain an agg
3947 				 * early, which will be made up for at the last
3948 				 * rxt out.
3949 				 */
3950 				bbr->rc_timer_first = 1;
3951 				bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
3952 			}
3953 			/*
3954 			 * Calculate a new cwnd based on to the current
3955 			 * delivery rate with no gain. We get the bdp
3956 			 * without gaining it up like we normally would and
3957 			 * we use the last cur_del_rate.
3958 			 */
3959 			if ((bbr->rc_use_google == 0) &&
3960 			    (bbr->r_ctl.bbr_rttprobe_gain_val ||
3961 			     (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) {
3962 				tp->snd_cwnd = ctf_flight_size(tp,
3963 					           (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3964 					(tp->t_maxseg - bbr->rc_last_options);
3965 				if (tp->snd_cwnd < get_min_cwnd(bbr)) {
3966 					/* We always gate to min cwnd */
3967 					tp->snd_cwnd = get_min_cwnd(bbr);
3968 				}
3969 				bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__);
3970 			}
3971 			bbr_log_type_enter_rec(bbr, rsm->r_start);
3972 		}
3973 		break;
3974 	case CC_RTO_ERR:
3975 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
3976 		/* RTO was unnecessary, so reset everything. */
3977 		bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime);
3978 		if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3979 			tp->snd_cwnd = tp->snd_cwnd_prev;
3980 			tp->snd_ssthresh = tp->snd_ssthresh_prev;
3981 			tp->snd_recover = tp->snd_recover_prev;
3982 			tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
3983 			bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__);
3984 		}
3985 		tp->t_badrxtwin = 0;
3986 		break;
3987 	}
3988 }
3989 
3990 /*
3991  * Indicate whether this ack should be delayed.  We can delay the ack if
3992  * following conditions are met:
3993  *	- There is no delayed ack timer in progress.
3994  *	- Our last ack wasn't a 0-sized window. We never want to delay
3995  *	  the ack that opens up a 0-sized window.
3996  *	- LRO wasn't used for this segment. We make sure by checking that the
3997  *	  segment size is not larger than the MSS.
3998  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
3999  *	  connection.
4000  *	- The data being acked is less than a full segment (a stretch ack
4001  *        of more than a segment we should ack.
4002  *      - nsegs is 1 (if its more than that we received more than 1 ack).
4003  */
4004 #define DELAY_ACK(tp, bbr, nsegs)				\
4005 	(((tp->t_flags & TF_RXWIN0SENT) == 0) &&		\
4006 	 ((tp->t_flags & TF_DELACK) == 0) && 		 	\
4007 	 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) &&	\
4008 	 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4009 
4010 /*
4011  * Return the lowest RSM in the map of
4012  * packets still in flight that is not acked.
4013  * This should normally find on the first one
4014  * since we remove packets from the send
4015  * map after they are marked ACKED.
4016  */
4017 static struct bbr_sendmap *
bbr_find_lowest_rsm(struct tcp_bbr * bbr)4018 bbr_find_lowest_rsm(struct tcp_bbr *bbr)
4019 {
4020 	struct bbr_sendmap *rsm;
4021 
4022 	/*
4023 	 * Walk the time-order transmitted list looking for an rsm that is
4024 	 * not acked. This will be the one that was sent the longest time
4025 	 * ago that is still outstanding.
4026 	 */
4027 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) {
4028 		if (rsm->r_flags & BBR_ACKED) {
4029 			continue;
4030 		}
4031 		goto finish;
4032 	}
4033 finish:
4034 	return (rsm);
4035 }
4036 
4037 static struct bbr_sendmap *
bbr_find_high_nonack(struct tcp_bbr * bbr,struct bbr_sendmap * rsm)4038 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
4039 {
4040 	struct bbr_sendmap *prsm;
4041 
4042 	/*
4043 	 * Walk the sequence order list backward until we hit and arrive at
4044 	 * the highest seq not acked. In theory when this is called it
4045 	 * should be the last segment (which it was not).
4046 	 */
4047 	prsm = rsm;
4048 	TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4049 		if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) {
4050 			continue;
4051 		}
4052 		return (prsm);
4053 	}
4054 	return (NULL);
4055 }
4056 
4057 /*
4058  * Returns to the caller the number of microseconds that
4059  * the packet can be outstanding before we think we
4060  * should have had an ack returned.
4061  */
4062 static uint32_t
bbr_calc_thresh_rack(struct tcp_bbr * bbr,uint32_t srtt,uint32_t cts,struct bbr_sendmap * rsm)4063 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm)
4064 {
4065 	/*
4066 	 * lro is the flag we use to determine if we have seen reordering.
4067 	 * If it gets set we have seen reordering. The reorder logic either
4068 	 * works in one of two ways:
4069 	 *
4070 	 * If reorder-fade is configured, then we track the last time we saw
4071 	 * re-ordering occur. If we reach the point where enough time as
4072 	 * passed we no longer consider reordering has occuring.
4073 	 *
4074 	 * Or if reorder-face is 0, then once we see reordering we consider
4075 	 * the connection to alway be subject to reordering and just set lro
4076 	 * to 1.
4077 	 *
4078 	 * In the end if lro is non-zero we add the extra time for
4079 	 * reordering in.
4080 	 */
4081 	int32_t lro;
4082 	uint32_t thresh, t_rxtcur;
4083 
4084 	if (srtt == 0)
4085 		srtt = 1;
4086 	if (bbr->r_ctl.rc_reorder_ts) {
4087 		if (bbr->r_ctl.rc_reorder_fade) {
4088 			if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) {
4089 				lro = cts - bbr->r_ctl.rc_reorder_ts;
4090 				if (lro == 0) {
4091 					/*
4092 					 * No time as passed since the last
4093 					 * reorder, mark it as reordering.
4094 					 */
4095 					lro = 1;
4096 				}
4097 			} else {
4098 				/* Negative time? */
4099 				lro = 0;
4100 			}
4101 			if (lro > bbr->r_ctl.rc_reorder_fade) {
4102 				/* Turn off reordering seen too */
4103 				bbr->r_ctl.rc_reorder_ts = 0;
4104 				lro = 0;
4105 			}
4106 		} else {
4107 			/* Reodering does not fade */
4108 			lro = 1;
4109 		}
4110 	} else {
4111 		lro = 0;
4112 	}
4113 	thresh = srtt + bbr->r_ctl.rc_pkt_delay;
4114 	if (lro) {
4115 		/* It must be set, if not you get 1/4 rtt */
4116 		if (bbr->r_ctl.rc_reorder_shift)
4117 			thresh += (srtt >> bbr->r_ctl.rc_reorder_shift);
4118 		else
4119 			thresh += (srtt >> 2);
4120 	} else {
4121 		thresh += 1000;
4122 	}
4123 	/* We don't let the rack timeout be above a RTO */
4124 	if ((bbr->rc_tp)->t_srtt == 0)
4125 		t_rxtcur = BBR_INITIAL_RTO;
4126 	else
4127 		t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
4128 	if (thresh > t_rxtcur) {
4129 		thresh = t_rxtcur;
4130 	}
4131 	/* And we don't want it above the RTO max either */
4132 	if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4133 		thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4134 	}
4135 	bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK);
4136 	return (thresh);
4137 }
4138 
4139 /*
4140  * Return to the caller the amount of time in mico-seconds
4141  * that should be used for the TLP timer from the last
4142  * send time of this packet.
4143  */
4144 static uint32_t
bbr_calc_thresh_tlp(struct tcpcb * tp,struct tcp_bbr * bbr,struct bbr_sendmap * rsm,uint32_t srtt,uint32_t cts)4145 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
4146     struct bbr_sendmap *rsm, uint32_t srtt,
4147     uint32_t cts)
4148 {
4149 	uint32_t thresh, len, maxseg, t_rxtcur;
4150 	struct bbr_sendmap *prsm;
4151 
4152 	if (srtt == 0)
4153 		srtt = 1;
4154 	if (bbr->rc_tlp_threshold)
4155 		thresh = srtt + (srtt / bbr->rc_tlp_threshold);
4156 	else
4157 		thresh = (srtt * 2);
4158 	maxseg = tp->t_maxseg - bbr->rc_last_options;
4159 	/* Get the previous sent packet, if any  */
4160 	len = rsm->r_end - rsm->r_start;
4161 
4162 	/* 2.1 behavior */
4163 	prsm = TAILQ_PREV(rsm, bbr_head, r_tnext);
4164 	if (prsm && (len <= maxseg)) {
4165 		/*
4166 		 * Two packets outstanding, thresh should be (2*srtt) +
4167 		 * possible inter-packet delay (if any).
4168 		 */
4169 		uint32_t inter_gap = 0;
4170 		int idx, nidx;
4171 
4172 		idx = rsm->r_rtr_cnt - 1;
4173 		nidx = prsm->r_rtr_cnt - 1;
4174 		if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4175 			/* Yes it was sent later (or at the same time) */
4176 			inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4177 		}
4178 		thresh += inter_gap;
4179 	} else if (len <= maxseg) {
4180 		/*
4181 		 * Possibly compensate for delayed-ack.
4182 		 */
4183 		uint32_t alt_thresh;
4184 
4185 		alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time;
4186 		if (alt_thresh > thresh)
4187 			thresh = alt_thresh;
4188 	}
4189 	/* Not above the current  RTO */
4190 	if (tp->t_srtt == 0)
4191 		t_rxtcur = BBR_INITIAL_RTO;
4192 	else
4193 		t_rxtcur = TICKS_2_USEC(tp->t_rxtcur);
4194 
4195 	bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP);
4196 	/* Not above an RTO */
4197 	if (thresh > t_rxtcur) {
4198 		thresh = t_rxtcur;
4199 	}
4200 	/* Not above a RTO max */
4201 	if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4202 		thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4203 	}
4204 	/* And now apply the user TLP min */
4205 	if (thresh < bbr_tlp_min) {
4206 		thresh = bbr_tlp_min;
4207 	}
4208 	return (thresh);
4209 }
4210 
4211 /*
4212  * Return one of three RTTs to use (in microseconds).
4213  */
4214 static __inline uint32_t
bbr_get_rtt(struct tcp_bbr * bbr,int32_t rtt_type)4215 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type)
4216 {
4217 	uint32_t f_rtt;
4218 	uint32_t srtt;
4219 
4220 	f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
4221 	if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) {
4222 		/* We have no rtt at all */
4223 		if (bbr->rc_tp->t_srtt == 0)
4224 			f_rtt = BBR_INITIAL_RTO;
4225 		else
4226 			f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4227 		/*
4228 		 * Since we don't know how good the rtt is apply a
4229 		 * delayed-ack min
4230 		 */
4231 		if (f_rtt < bbr_delayed_ack_time) {
4232 			f_rtt = bbr_delayed_ack_time;
4233 		}
4234 	}
4235 	/* Take the filter version or last measured pkt-rtt */
4236 	if (rtt_type == BBR_RTT_PROP) {
4237 		srtt = f_rtt;
4238 	} else if (rtt_type == BBR_RTT_PKTRTT) {
4239 		if (bbr->r_ctl.rc_pkt_epoch_rtt) {
4240 			srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
4241 		} else {
4242 			/* No pkt rtt yet */
4243 			srtt = f_rtt;
4244 		}
4245 	} else if (rtt_type == BBR_RTT_RACK) {
4246 		srtt = bbr->r_ctl.rc_last_rtt;
4247 		/* We need to add in any internal delay for our timer */
4248 		if (bbr->rc_ack_was_delayed)
4249 			srtt += bbr->r_ctl.rc_ack_hdwr_delay;
4250 	} else if (rtt_type == BBR_SRTT) {
4251 		srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4252 	} else {
4253 		/* TSNH */
4254 		srtt = f_rtt;
4255 #ifdef BBR_INVARIANTS
4256 		panic("Unknown rtt request type %d", rtt_type);
4257 #endif
4258 	}
4259 	return (srtt);
4260 }
4261 
4262 static int
bbr_is_lost(struct tcp_bbr * bbr,struct bbr_sendmap * rsm,uint32_t cts)4263 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts)
4264 {
4265 	uint32_t thresh;
4266 
4267 	thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK),
4268 				      cts, rsm);
4269 	if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) {
4270 		/* It is lost (past time) */
4271 		return (1);
4272 	}
4273 	return (0);
4274 }
4275 
4276 /*
4277  * Return a sendmap if we need to retransmit something.
4278  */
4279 static struct bbr_sendmap *
bbr_check_recovery_mode(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts)4280 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4281 {
4282 	/*
4283 	 * Check to see that we don't need to fall into recovery. We will
4284 	 * need to do so if our oldest transmit is past the time we should
4285 	 * have had an ack.
4286 	 */
4287 
4288 	struct bbr_sendmap *rsm;
4289 	int32_t idx;
4290 
4291 	if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) {
4292 		/* Nothing outstanding that we know of */
4293 		return (NULL);
4294 	}
4295 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
4296 	if (rsm == NULL) {
4297 		/* Nothing in the transmit map */
4298 		return (NULL);
4299 	}
4300 	if (tp->t_flags & TF_SENTFIN) {
4301 		/* Fin restricted, don't find anything once a fin is sent */
4302 		return (NULL);
4303 	}
4304 	if (rsm->r_flags & BBR_ACKED) {
4305 		/*
4306 		 * Ok the first one is acked (this really should not happen
4307 		 * since we remove the from the tmap once they are acked)
4308 		 */
4309 		rsm = bbr_find_lowest_rsm(bbr);
4310 		if (rsm == NULL)
4311 			return (NULL);
4312 	}
4313 	idx = rsm->r_rtr_cnt - 1;
4314 	if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) {
4315 		/* Send timestamp is the same or less? can't be ready */
4316 		return (NULL);
4317 	}
4318 	/* Get our RTT time */
4319 	if (bbr_is_lost(bbr, rsm, cts) &&
4320 	    ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
4321 	     (rsm->r_flags & BBR_SACK_PASSED))) {
4322 		if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4323 			rsm->r_flags |= BBR_MARKED_LOST;
4324 			bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4325 			bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4326 		}
4327 		bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm);
4328 #ifdef BBR_INVARIANTS
4329 		if ((rsm->r_end - rsm->r_start) == 0)
4330 			panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm);
4331 #endif
4332 		return (rsm);
4333 	}
4334 	return (NULL);
4335 }
4336 
4337 /*
4338  * RACK Timer, here we simply do logging and house keeping.
4339  * the normal bbr_output_wtime() function will call the
4340  * appropriate thing to check if we need to do a RACK retransmit.
4341  * We return 1, saying don't proceed with bbr_output_wtime only
4342  * when all timers have been stopped (destroyed PCB?).
4343  */
4344 static int
bbr_timeout_rack(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts)4345 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4346 {
4347 	/*
4348 	 * This timer simply provides an internal trigger to send out data.
4349 	 * The check_recovery_mode call will see if there are needed
4350 	 * retransmissions, if so we will enter fast-recovery. The output
4351 	 * call may or may not do the same thing depending on sysctl
4352 	 * settings.
4353 	 */
4354 	uint32_t lost;
4355 
4356 	if (bbr->rc_all_timers_stopped) {
4357 		return (1);
4358 	}
4359 	if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4360 		/* Its not time yet */
4361 		return (0);
4362 	}
4363 	BBR_STAT_INC(bbr_to_tot);
4364 	lost = bbr->r_ctl.rc_lost;
4365 	if (bbr->r_state && (bbr->r_state != tp->t_state))
4366 		bbr_set_state(tp, bbr, 0);
4367 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK);
4368 	if (bbr->r_ctl.rc_resend == NULL) {
4369 		/* Lets do the check here */
4370 		bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
4371 	}
4372 	if (bbr_policer_call_from_rack_to)
4373 		bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4374 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
4375 	return (0);
4376 }
4377 
4378 static __inline void
bbr_clone_rsm(struct tcp_bbr * bbr,struct bbr_sendmap * nrsm,struct bbr_sendmap * rsm,uint32_t start)4379 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start)
4380 {
4381 	int idx;
4382 
4383 	nrsm->r_start = start;
4384 	nrsm->r_end = rsm->r_end;
4385 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
4386 	nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed;
4387 	nrsm->r_flags = rsm->r_flags;
4388 	/* We don't transfer forward the SYN flag */
4389 	nrsm->r_flags &= ~BBR_HAS_SYN;
4390 	/* We move forward the FIN flag, not that this should happen */
4391 	rsm->r_flags &= ~BBR_HAS_FIN;
4392 	nrsm->r_dupack = rsm->r_dupack;
4393 	nrsm->r_rtr_bytes = 0;
4394 	nrsm->r_is_gain = rsm->r_is_gain;
4395 	nrsm->r_is_drain = rsm->r_is_drain;
4396 	nrsm->r_delivered = rsm->r_delivered;
4397 	nrsm->r_ts_valid = rsm->r_ts_valid;
4398 	nrsm->r_del_ack_ts = rsm->r_del_ack_ts;
4399 	nrsm->r_del_time = rsm->r_del_time;
4400 	nrsm->r_app_limited = rsm->r_app_limited;
4401 	nrsm->r_first_sent_time = rsm->r_first_sent_time;
4402 	nrsm->r_flight_at_send = rsm->r_flight_at_send;
4403 	/* We split a piece the lower section looses any just_ret flag. */
4404 	nrsm->r_bbr_state = rsm->r_bbr_state;
4405 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
4406 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
4407 	}
4408 	rsm->r_end = nrsm->r_start;
4409 	idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
4410 	idx /= 8;
4411 	/* Check if we got too small */
4412 	if ((rsm->r_is_smallmap == 0) &&
4413 	    ((rsm->r_end - rsm->r_start) <= idx)) {
4414 		bbr->r_ctl.rc_num_small_maps_alloced++;
4415 		rsm->r_is_smallmap = 1;
4416 	}
4417 	/* Check the new one as well */
4418 	if ((nrsm->r_end - nrsm->r_start) <= idx) {
4419 		bbr->r_ctl.rc_num_small_maps_alloced++;
4420 		nrsm->r_is_smallmap = 1;
4421 	}
4422 }
4423 
4424 static int
bbr_sack_mergable(struct bbr_sendmap * at,uint32_t start,uint32_t end)4425 bbr_sack_mergable(struct bbr_sendmap *at,
4426 		  uint32_t start, uint32_t end)
4427 {
4428 	/*
4429 	 * Given a sack block defined by
4430 	 * start and end, and a current position
4431 	 * at. Return 1 if either side of at
4432 	 * would show that the block is mergable
4433 	 * to that side. A block to be mergable
4434 	 * must have overlap with the start/end
4435 	 * and be in the SACK'd state.
4436 	 */
4437 	struct bbr_sendmap *l_rsm;
4438 	struct bbr_sendmap *r_rsm;
4439 
4440 	/* first get the either side blocks */
4441 	l_rsm = TAILQ_PREV(at, bbr_head, r_next);
4442 	r_rsm = TAILQ_NEXT(at, r_next);
4443 	if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) {
4444 		/* Potentially mergeable */
4445 		if ((l_rsm->r_end == start) ||
4446 		    (SEQ_LT(start, l_rsm->r_end) &&
4447 		     SEQ_GT(end, l_rsm->r_end))) {
4448 			    /*
4449 			     * map blk   |------|
4450 			     * sack blk         |------|
4451 			     * <or>
4452 			     * map blk   |------|
4453 			     * sack blk      |------|
4454 			     */
4455 			    return (1);
4456 		    }
4457 	}
4458 	if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) {
4459 		/* Potentially mergeable */
4460 		if ((r_rsm->r_start == end) ||
4461 		    (SEQ_LT(start, r_rsm->r_start) &&
4462 		     SEQ_GT(end, r_rsm->r_start))) {
4463 			/*
4464 			 * map blk          |---------|
4465 			 * sack blk    |----|
4466 			 * <or>
4467 			 * map blk          |---------|
4468 			 * sack blk    |-------|
4469 			 */
4470 			return (1);
4471 		}
4472 	}
4473 	return (0);
4474 }
4475 
4476 static struct bbr_sendmap *
bbr_merge_rsm(struct tcp_bbr * bbr,struct bbr_sendmap * l_rsm,struct bbr_sendmap * r_rsm)4477 bbr_merge_rsm(struct tcp_bbr *bbr,
4478 	      struct bbr_sendmap *l_rsm,
4479 	      struct bbr_sendmap *r_rsm)
4480 {
4481 	/*
4482 	 * We are merging two ack'd RSM's,
4483 	 * the l_rsm is on the left (lower seq
4484 	 * values) and the r_rsm is on the right
4485 	 * (higher seq value). The simplest way
4486 	 * to merge these is to move the right
4487 	 * one into the left. I don't think there
4488 	 * is any reason we need to try to find
4489 	 * the oldest (or last oldest retransmitted).
4490 	 */
4491 	l_rsm->r_end = r_rsm->r_end;
4492 	if (l_rsm->r_dupack < r_rsm->r_dupack)
4493 		l_rsm->r_dupack = r_rsm->r_dupack;
4494 	if (r_rsm->r_rtr_bytes)
4495 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
4496 	if (r_rsm->r_in_tmap) {
4497 		/* This really should not happen */
4498 		TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext);
4499 	}
4500 	if (r_rsm->r_app_limited)
4501 		l_rsm->r_app_limited = r_rsm->r_app_limited;
4502 	/* Now the flags */
4503 	if (r_rsm->r_flags & BBR_HAS_FIN)
4504 		l_rsm->r_flags |= BBR_HAS_FIN;
4505 	if (r_rsm->r_flags & BBR_TLP)
4506 		l_rsm->r_flags |= BBR_TLP;
4507 	if (r_rsm->r_flags & BBR_RWND_COLLAPSED)
4508 		l_rsm->r_flags |= BBR_RWND_COLLAPSED;
4509 	if (r_rsm->r_flags & BBR_MARKED_LOST) {
4510 		/* This really should not happen */
4511 		bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start;
4512 	}
4513 	TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next);
4514 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
4515 		/* Transfer the split limit to the map we free */
4516 		r_rsm->r_limit_type = l_rsm->r_limit_type;
4517 		l_rsm->r_limit_type = 0;
4518 	}
4519 	bbr_free(bbr, r_rsm);
4520 	return(l_rsm);
4521 }
4522 
4523 /*
4524  * TLP Timer, here we simply setup what segment we want to
4525  * have the TLP expire on, the normal bbr_output_wtime() will then
4526  * send it out.
4527  *
4528  * We return 1, saying don't proceed with bbr_output_wtime only
4529  * when all timers have been stopped (destroyed PCB?).
4530  */
4531 static int
bbr_timeout_tlp(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts)4532 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4533 {
4534 	/*
4535 	 * Tail Loss Probe.
4536 	 */
4537 	struct bbr_sendmap *rsm = NULL;
4538 	struct socket *so;
4539 	uint32_t amm;
4540 	uint32_t out, avail;
4541 	uint32_t maxseg;
4542 	int collapsed_win = 0;
4543 
4544 	if (bbr->rc_all_timers_stopped) {
4545 		return (1);
4546 	}
4547 	if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4548 		/* Its not time yet */
4549 		return (0);
4550 	}
4551 	if (ctf_progress_timeout_check(tp, true)) {
4552 		bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4553 		return (-ETIMEDOUT);	/* tcp_drop() */
4554 	}
4555 	/* Did we somehow get into persists? */
4556 	if (bbr->rc_in_persist) {
4557 		return (0);
4558 	}
4559 	if (bbr->r_state && (bbr->r_state != tp->t_state))
4560 		bbr_set_state(tp, bbr, 0);
4561 	BBR_STAT_INC(bbr_tlp_tot);
4562 	maxseg = tp->t_maxseg - bbr->rc_last_options;
4563 	/*
4564 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
4565 	 * need to figure out how to force a full MSS segment out.
4566 	 */
4567 	so = tptosocket(tp);
4568 	avail = sbavail(&so->so_snd);
4569 	out = ctf_outstanding(tp);
4570 	if (out > tp->snd_wnd) {
4571 		/* special case, we need a retransmission */
4572 		collapsed_win = 1;
4573 		goto need_retran;
4574 	}
4575 	if (avail > out) {
4576 		/* New data is available */
4577 		amm = avail - out;
4578 		if (amm > maxseg) {
4579 			amm = maxseg;
4580 		} else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) {
4581 			/* not enough to fill a MTU and no-delay is off */
4582 			goto need_retran;
4583 		}
4584 		/* Set the send-new override */
4585 		if ((out + amm) <= tp->snd_wnd) {
4586 			bbr->rc_tlp_new_data = 1;
4587 		} else {
4588 			goto need_retran;
4589 		}
4590 		bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4591 		bbr->r_ctl.rc_last_tlp_seq = tp->snd_max;
4592 		bbr->r_ctl.rc_tlp_send = NULL;
4593 		/* cap any slots */
4594 		BBR_STAT_INC(bbr_tlp_newdata);
4595 		goto send;
4596 	}
4597 need_retran:
4598 	/*
4599 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
4600 	 * optionally the first un-acked segment.
4601 	 */
4602 	if (collapsed_win == 0) {
4603 		rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
4604 		if (rsm && (rsm->r_flags & (BBR_ACKED | BBR_HAS_FIN))) {
4605 			rsm = bbr_find_high_nonack(bbr, rsm);
4606 		}
4607 		if (rsm == NULL) {
4608 			goto restore;
4609 		}
4610 	} else {
4611 		/*
4612 		 * We must find the last segment
4613 		 * that was acceptable by the client.
4614 		 */
4615 		TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4616 			if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) {
4617 				/* Found one */
4618 				break;
4619 			}
4620 		}
4621 		if (rsm == NULL) {
4622 			/* None? if so send the first */
4623 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4624 			if (rsm == NULL)
4625 				goto restore;
4626 		}
4627 	}
4628 	if ((rsm->r_end - rsm->r_start) > maxseg) {
4629 		/*
4630 		 * We need to split this the last segment in two.
4631 		 */
4632 		struct bbr_sendmap *nrsm;
4633 
4634 		nrsm = bbr_alloc_full_limit(bbr);
4635 		if (nrsm == NULL) {
4636 			/*
4637 			 * We can't get memory to split, we can either just
4638 			 * not split it. Or retransmit the whole piece, lets
4639 			 * do the large send (BTLP :-) ).
4640 			 */
4641 			goto go_for_it;
4642 		}
4643 		bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg));
4644 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
4645 		if (rsm->r_in_tmap) {
4646 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
4647 			nrsm->r_in_tmap = 1;
4648 		}
4649 		rsm->r_flags &= (~BBR_HAS_FIN);
4650 		rsm = nrsm;
4651 	}
4652 go_for_it:
4653 	bbr->r_ctl.rc_tlp_send = rsm;
4654 	bbr->rc_tlp_rtx_out = 1;
4655 	if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) {
4656 		bbr->r_ctl.rc_tlp_seg_send_cnt++;
4657 		tp->t_rxtshift++;
4658 	} else {
4659 		bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
4660 		bbr->r_ctl.rc_tlp_seg_send_cnt = 1;
4661 	}
4662 send:
4663 	if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
4664 		/*
4665 		 * Can't [re]/transmit a segment we have retransmitted the
4666 		 * max times. We need the retransmit timer to take over.
4667 		 */
4668 restore:
4669 		bbr->rc_tlp_new_data = 0;
4670 		bbr->r_ctl.rc_tlp_send = NULL;
4671 		if (rsm)
4672 			rsm->r_flags &= ~BBR_TLP;
4673 		BBR_STAT_INC(bbr_tlp_retran_fail);
4674 		return (0);
4675 	} else if (rsm) {
4676 		rsm->r_flags |= BBR_TLP;
4677 	}
4678 	if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) &&
4679 	    (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) {
4680 		/*
4681 		 * We have retransmitted to many times for TLP. Switch to
4682 		 * the regular RTO timer
4683 		 */
4684 		goto restore;
4685 	}
4686 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP);
4687 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
4688 	return (0);
4689 }
4690 
4691 /*
4692  * Delayed ack Timer, here we simply need to setup the
4693  * ACK_NOW flag and remove the DELACK flag. From there
4694  * the output routine will send the ack out.
4695  *
4696  * We only return 1, saying don't proceed, if all timers
4697  * are stopped (destroyed PCB?).
4698  */
4699 static int
bbr_timeout_delack(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts)4700 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4701 {
4702 	if (bbr->rc_all_timers_stopped) {
4703 		return (1);
4704 	}
4705 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK);
4706 	tp->t_flags &= ~TF_DELACK;
4707 	tp->t_flags |= TF_ACKNOW;
4708 	KMOD_TCPSTAT_INC(tcps_delack);
4709 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4710 	return (0);
4711 }
4712 
4713 /*
4714  * Here we send a KEEP-ALIVE like probe to the
4715  * peer, we do not send data.
4716  *
4717  * We only return 1, saying don't proceed, if all timers
4718  * are stopped (destroyed PCB?).
4719  */
4720 static int
bbr_timeout_persist(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts)4721 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4722 {
4723 	struct tcptemp *t_template;
4724 	int32_t retval = 1;
4725 
4726 	if (bbr->rc_all_timers_stopped) {
4727 		return (1);
4728 	}
4729 	if (bbr->rc_in_persist == 0)
4730 		return (0);
4731 
4732 	/*
4733 	 * Persistence timer into zero window. Force a byte to be output, if
4734 	 * possible.
4735 	 */
4736 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST);
4737 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
4738 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
4739 	/*
4740 	 * Have we exceeded the user specified progress time?
4741 	 */
4742 	if (ctf_progress_timeout_check(tp, true)) {
4743 		bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4744 		return (-ETIMEDOUT);	/* tcp_drop() */
4745 	}
4746 	/*
4747 	 * Hack: if the peer is dead/unreachable, we do not time out if the
4748 	 * window is closed.  After a full backoff, drop the connection if
4749 	 * the idle time (no responses to probes) reaches the maximum
4750 	 * backoff that we would use if retransmitting.
4751 	 */
4752 	if (tp->t_rxtshift >= V_tcp_retries &&
4753 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
4754 	    ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
4755 		KMOD_TCPSTAT_INC(tcps_persistdrop);
4756 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4757 		return (-ETIMEDOUT);	/* tcp_drop() */
4758 	}
4759 	if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) &&
4760 	    tp->snd_una == tp->snd_max) {
4761 		bbr_exit_persist(tp, bbr, cts, __LINE__);
4762 		retval = 0;
4763 		goto out;
4764 	}
4765 	/*
4766 	 * If the user has closed the socket then drop a persisting
4767 	 * connection after a much reduced timeout.
4768 	 */
4769 	if (tp->t_state > TCPS_CLOSE_WAIT &&
4770 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
4771 		KMOD_TCPSTAT_INC(tcps_persistdrop);
4772 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4773 		return (-ETIMEDOUT);	/* tcp_drop() */
4774 	}
4775 	t_template = tcpip_maketemplate(bbr->rc_inp);
4776 	if (t_template) {
4777 		tcp_respond(tp, t_template->tt_ipgen,
4778 			    &t_template->tt_t, (struct mbuf *)NULL,
4779 			    tp->rcv_nxt, tp->snd_una - 1, 0);
4780 		/* This sends an ack */
4781 		if (tp->t_flags & TF_DELACK)
4782 			tp->t_flags &= ~TF_DELACK;
4783 		free(t_template, M_TEMP);
4784 	}
4785 	if (tp->t_rxtshift < V_tcp_retries)
4786 		tp->t_rxtshift++;
4787 	bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0);
4788 out:
4789 	return (retval);
4790 }
4791 
4792 /*
4793  * If a keepalive goes off, we had no other timers
4794  * happening. We always return 1 here since this
4795  * routine either drops the connection or sends
4796  * out a segment with respond.
4797  */
4798 static int
bbr_timeout_keepalive(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts)4799 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4800 {
4801 	struct tcptemp *t_template;
4802 	struct inpcb *inp = tptoinpcb(tp);
4803 
4804 	if (bbr->rc_all_timers_stopped) {
4805 		return (1);
4806 	}
4807 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
4808 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP);
4809 	/*
4810 	 * Keep-alive timer went off; send something or drop connection if
4811 	 * idle for too long.
4812 	 */
4813 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
4814 	if (tp->t_state < TCPS_ESTABLISHED)
4815 		goto dropit;
4816 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4817 	    tp->t_state <= TCPS_CLOSING) {
4818 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
4819 			goto dropit;
4820 		/*
4821 		 * Send a packet designed to force a response if the peer is
4822 		 * up and reachable: either an ACK if the connection is
4823 		 * still alive, or an RST if the peer has closed the
4824 		 * connection due to timeout or reboot. Using sequence
4825 		 * number tp->snd_una-1 causes the transmitted zero-length
4826 		 * segment to lie outside the receive window; by the
4827 		 * protocol spec, this requires the correspondent TCP to
4828 		 * respond.
4829 		 */
4830 		KMOD_TCPSTAT_INC(tcps_keepprobe);
4831 		t_template = tcpip_maketemplate(inp);
4832 		if (t_template) {
4833 			tcp_respond(tp, t_template->tt_ipgen,
4834 			    &t_template->tt_t, (struct mbuf *)NULL,
4835 			    tp->rcv_nxt, tp->snd_una - 1, 0);
4836 			free(t_template, M_TEMP);
4837 		}
4838 	}
4839 	bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0);
4840 	return (1);
4841 dropit:
4842 	KMOD_TCPSTAT_INC(tcps_keepdrops);
4843 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
4844 	return (-ETIMEDOUT);	/* tcp_drop() */
4845 }
4846 
4847 /*
4848  * Retransmit helper function, clear up all the ack
4849  * flags and take care of important book keeping.
4850  */
4851 static void
bbr_remxt_tmr(struct tcpcb * tp)4852 bbr_remxt_tmr(struct tcpcb *tp)
4853 {
4854 	/*
4855 	 * The retransmit timer went off, all sack'd blocks must be
4856 	 * un-acked.
4857 	 */
4858 	struct bbr_sendmap *rsm, *trsm = NULL;
4859 	struct tcp_bbr *bbr;
4860 	uint32_t cts, lost;
4861 
4862 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
4863 	cts = tcp_get_usecs(&bbr->rc_tv);
4864 	lost = bbr->r_ctl.rc_lost;
4865 	if (bbr->r_state && (bbr->r_state != tp->t_state))
4866 		bbr_set_state(tp, bbr, 0);
4867 
4868 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
4869 		if (rsm->r_flags & BBR_ACKED) {
4870 			uint32_t old_flags;
4871 
4872 			rsm->r_dupack = 0;
4873 			if (rsm->r_in_tmap == 0) {
4874 				/* We must re-add it back to the tlist */
4875 				if (trsm == NULL) {
4876 					TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
4877 				} else {
4878 					TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext);
4879 				}
4880 				rsm->r_in_tmap = 1;
4881 			}
4882 			old_flags = rsm->r_flags;
4883 			rsm->r_flags |= BBR_RXT_CLEARED;
4884 			rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS);
4885 			bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
4886 		} else {
4887 			if ((tp->t_state < TCPS_ESTABLISHED) &&
4888 			    (rsm->r_start == tp->snd_una)) {
4889 				/*
4890 				 * Special case for TCP FO. Where
4891 				 * we sent more data beyond the snd_max.
4892 				 * We don't mark that as lost and stop here.
4893 				 */
4894 				break;
4895 			}
4896 			if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4897 				bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4898 				bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4899 			}
4900 			if (bbr_marks_rxt_sack_passed) {
4901 				/*
4902 				 * With this option, we will rack out
4903 				 * in 1ms increments the rest of the packets.
4904 				 */
4905 				rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST;
4906 				rsm->r_flags &= ~BBR_WAS_SACKPASS;
4907 			} else {
4908 				/*
4909 				 * With this option we only mark them lost
4910 				 * and remove all sack'd markings. We will run
4911 				 * another RXT or a TLP. This will cause
4912 				 * us to eventually send more based on what
4913 				 * ack's come in.
4914 				 */
4915 				rsm->r_flags |= BBR_MARKED_LOST;
4916 				rsm->r_flags &= ~BBR_WAS_SACKPASS;
4917 				rsm->r_flags &= ~BBR_SACK_PASSED;
4918 			}
4919 		}
4920 		trsm = rsm;
4921 	}
4922 	bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4923 	/* Clear the count (we just un-acked them) */
4924 	bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR);
4925 	bbr->rc_tlp_new_data = 0;
4926 	bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4927 	/* zap the behindness on a rxt */
4928 	bbr->r_ctl.rc_hptsi_agg_delay = 0;
4929 	bbr->r_agg_early_set = 0;
4930 	bbr->r_ctl.rc_agg_early = 0;
4931 	bbr->rc_tlp_rtx_out = 0;
4932 	bbr->r_ctl.rc_sacked = 0;
4933 	bbr->r_ctl.rc_sacklast = NULL;
4934 	bbr->r_timer_override = 1;
4935 	bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4936 }
4937 
4938 /*
4939  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
4940  * we will setup to retransmit the lowest seq number outstanding.
4941  */
4942 static int
bbr_timeout_rxt(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts)4943 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4944 {
4945 	struct inpcb *inp = tptoinpcb(tp);
4946 	int32_t rexmt;
4947 	int32_t retval = 0;
4948 	bool isipv6;
4949 
4950 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
4951 	if (bbr->rc_all_timers_stopped) {
4952 		return (1);
4953 	}
4954 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
4955 	    (tp->snd_una == tp->snd_max)) {
4956 		/* Nothing outstanding .. nothing to do */
4957 		return (0);
4958 	}
4959 	/*
4960 	 * Retransmission timer went off.  Message has not been acked within
4961 	 * retransmit interval.  Back off to a longer retransmit interval
4962 	 * and retransmit one segment.
4963 	 */
4964 	if (ctf_progress_timeout_check(tp, true)) {
4965 		bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4966 		return (-ETIMEDOUT);	/* tcp_drop() */
4967 	}
4968 	bbr_remxt_tmr(tp);
4969 	if ((bbr->r_ctl.rc_resend == NULL) ||
4970 	    ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) {
4971 		/*
4972 		 * If the rwnd collapsed on
4973 		 * the one we are retransmitting
4974 		 * it does not count against the
4975 		 * rxt count.
4976 		 */
4977 		tp->t_rxtshift++;
4978 	}
4979 	if (tp->t_rxtshift > V_tcp_retries) {
4980 		tp->t_rxtshift = V_tcp_retries;
4981 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
4982 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
4983 		/* XXXGL: previously t_softerror was casted to uint16_t */
4984 		MPASS(tp->t_softerror >= 0);
4985 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
4986 		return (retval);	/* tcp_drop() */
4987 	}
4988 	if (tp->t_state == TCPS_SYN_SENT) {
4989 		/*
4990 		 * If the SYN was retransmitted, indicate CWND to be limited
4991 		 * to 1 segment in cc_conn_init().
4992 		 */
4993 		tp->snd_cwnd = 1;
4994 	} else if (tp->t_rxtshift == 1) {
4995 		/*
4996 		 * first retransmit; record ssthresh and cwnd so they can be
4997 		 * recovered if this turns out to be a "bad" retransmit. A
4998 		 * retransmit is considered "bad" if an ACK for this segment
4999 		 * is received within RTT/2 interval; the assumption here is
5000 		 * that the ACK was already in flight.  See "On Estimating
5001 		 * End-to-End Network Path Properties" by Allman and Paxson
5002 		 * for more details.
5003 		 */
5004 		tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5005 		if (!IN_RECOVERY(tp->t_flags)) {
5006 			tp->snd_cwnd_prev = tp->snd_cwnd;
5007 			tp->snd_ssthresh_prev = tp->snd_ssthresh;
5008 			tp->snd_recover_prev = tp->snd_recover;
5009 			tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5010 			tp->t_flags |= TF_PREVVALID;
5011 		} else {
5012 			tp->t_flags &= ~TF_PREVVALID;
5013 		}
5014 		tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5015 	} else {
5016 		tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5017 		tp->t_flags &= ~TF_PREVVALID;
5018 	}
5019 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
5020 	if ((tp->t_state == TCPS_SYN_SENT) ||
5021 	    (tp->t_state == TCPS_SYN_RECEIVED))
5022 		rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift];
5023 	else
5024 		rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5025 	TCPT_RANGESET(tp->t_rxtcur, rexmt,
5026 	    MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms),
5027 	    MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
5028 	/*
5029 	 * We enter the path for PLMTUD if connection is established or, if
5030 	 * connection is FIN_WAIT_1 status, reason for the last is that if
5031 	 * amount of data we send is very small, we could send it in couple
5032 	 * of packets and process straight to FIN. In that case we won't
5033 	 * catch ESTABLISHED state.
5034 	 */
5035 #ifdef INET6
5036 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
5037 #else
5038 	isipv6 = false;
5039 #endif
5040 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
5041 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
5042 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
5043 	    ((tp->t_state == TCPS_ESTABLISHED) ||
5044 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
5045 		/*
5046 		 * Idea here is that at each stage of mtu probe (usually,
5047 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
5048 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5049 		 * should take care of that.
5050 		 */
5051 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5052 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5053 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5054 		    tp->t_rxtshift % 2 == 0)) {
5055 			/*
5056 			 * Enter Path MTU Black-hole Detection mechanism: -
5057 			 * Disable Path MTU Discovery (IP "DF" bit). -
5058 			 * Reduce MTU to lower value than what we negotiated
5059 			 * with peer.
5060 			 */
5061 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5062 				/*
5063 				 * Record that we may have found a black
5064 				 * hole.
5065 				 */
5066 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5067 				/* Keep track of previous MSS. */
5068 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5069 			}
5070 			/*
5071 			 * Reduce the MSS to blackhole value or to the
5072 			 * default in an attempt to retransmit.
5073 			 */
5074 #ifdef INET6
5075 			isipv6 = bbr->r_is_v6;
5076 			if (isipv6 &&
5077 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5078 				/* Use the sysctl tuneable blackhole MSS. */
5079 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5080 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5081 			} else if (isipv6) {
5082 				/* Use the default MSS. */
5083 				tp->t_maxseg = V_tcp_v6mssdflt;
5084 				/*
5085 				 * Disable Path MTU Discovery when we switch
5086 				 * to minmss.
5087 				 */
5088 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5089 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5090 			}
5091 #endif
5092 #if defined(INET6) && defined(INET)
5093 			else
5094 #endif
5095 #ifdef INET
5096 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5097 				/* Use the sysctl tuneable blackhole MSS. */
5098 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5099 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5100 			} else {
5101 				/* Use the default MSS. */
5102 				tp->t_maxseg = V_tcp_mssdflt;
5103 				/*
5104 				 * Disable Path MTU Discovery when we switch
5105 				 * to minmss.
5106 				 */
5107 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5108 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5109 			}
5110 #endif
5111 		} else {
5112 			/*
5113 			 * If further retransmissions are still unsuccessful
5114 			 * with a lowered MTU, maybe this isn't a blackhole
5115 			 * and we restore the previous MSS and blackhole
5116 			 * detection flags. The limit '6' is determined by
5117 			 * giving each probe stage (1448, 1188, 524) 2
5118 			 * chances to recover.
5119 			 */
5120 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5121 			    (tp->t_rxtshift >= 6)) {
5122 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5123 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5124 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5125 				if (tp->t_maxseg < V_tcp_mssdflt) {
5126 					/*
5127 					 * The MSS is so small we should not
5128 					 * process incoming SACK's since we are
5129 					 * subject to attack in such a case.
5130 					 */
5131 					tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
5132 				} else {
5133 					tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
5134 				}
5135 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5136 			}
5137 		}
5138 	}
5139 	/*
5140 	 * Disable RFC1323 and SACK if we haven't got any response to our
5141 	 * third SYN to work-around some broken terminal servers (most of
5142 	 * which have hopefully been retired) that have bad VJ header
5143 	 * compression code which trashes TCP segments containing
5144 	 * unknown-to-them TCP options.
5145 	 */
5146 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
5147 	    (tp->t_rxtshift == 3))
5148 		tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT);
5149 	/*
5150 	 * If we backed off this far, our srtt estimate is probably bogus.
5151 	 * Clobber it so we'll take the next rtt measurement as our srtt;
5152 	 * move the current srtt into rttvar to keep the current retransmit
5153 	 * times until then.
5154 	 */
5155 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5156 #ifdef INET6
5157 		if (bbr->r_is_v6)
5158 			in6_losing(inp);
5159 		else
5160 #endif
5161 			in_losing(inp);
5162 		tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5163 		tp->t_srtt = 0;
5164 	}
5165 	sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
5166 	tp->snd_recover = tp->snd_max;
5167 	tp->t_flags |= TF_ACKNOW;
5168 	tp->t_rtttime = 0;
5169 
5170 	return (retval);
5171 }
5172 
5173 static int
bbr_process_timers(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts,uint8_t hpts_calling)5174 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling)
5175 {
5176 	int32_t ret = 0;
5177 	int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5178 
5179 	if (timers == 0) {
5180 		return (0);
5181 	}
5182 	if (tp->t_state == TCPS_LISTEN) {
5183 		/* no timers on listen sockets */
5184 		if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5185 			return (0);
5186 		return (1);
5187 	}
5188 	if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
5189 		uint32_t left;
5190 
5191 		if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5192 			ret = -1;
5193 			bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5194 			return (0);
5195 		}
5196 		if (hpts_calling == 0) {
5197 			ret = -2;
5198 			bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5199 			return (0);
5200 		}
5201 		/*
5202 		 * Ok our timer went off early and we are not paced false
5203 		 * alarm, go back to sleep.
5204 		 */
5205 		left = bbr->r_ctl.rc_timer_exp - cts;
5206 		ret = -3;
5207 		bbr_log_to_processing(bbr, cts, ret, left, hpts_calling);
5208 		tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(left));
5209 		return (1);
5210 	}
5211 	bbr->rc_tmr_stopped = 0;
5212 	bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5213 	if (timers & PACE_TMR_DELACK) {
5214 		ret = bbr_timeout_delack(tp, bbr, cts);
5215 	} else if (timers & PACE_TMR_PERSIT) {
5216 		ret = bbr_timeout_persist(tp, bbr, cts);
5217 	} else if (timers & PACE_TMR_RACK) {
5218 		bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5219 		ret = bbr_timeout_rack(tp, bbr, cts);
5220 	} else if (timers & PACE_TMR_TLP) {
5221 		bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5222 		ret = bbr_timeout_tlp(tp, bbr, cts);
5223 	} else if (timers & PACE_TMR_RXT) {
5224 		bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5225 		ret = bbr_timeout_rxt(tp, bbr, cts);
5226 	} else if (timers & PACE_TMR_KEEP) {
5227 		ret = bbr_timeout_keepalive(tp, bbr, cts);
5228 	}
5229 	bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling);
5230 	return (ret);
5231 }
5232 
5233 static void
bbr_timer_cancel(struct tcp_bbr * bbr,int32_t line,uint32_t cts)5234 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts)
5235 {
5236 	if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5237 		uint8_t hpts_removed = 0;
5238 
5239 		if (tcp_in_hpts(bbr->rc_tp) &&
5240 		    (bbr->rc_timer_first == 1)) {
5241 			/*
5242 			 * If we are canceling timer's when we have the
5243 			 * timer ahead of the output being paced. We also
5244 			 * must remove ourselves from the hpts.
5245 			 */
5246 			hpts_removed = 1;
5247 			tcp_hpts_remove(bbr->rc_tp);
5248 			if (bbr->r_ctl.rc_last_delay_val) {
5249 				/* Update the last hptsi delay too */
5250 				uint32_t time_since_send;
5251 
5252 				if (TSTMP_GT(cts, bbr->rc_pacer_started))
5253 					time_since_send = cts - bbr->rc_pacer_started;
5254 				else
5255 					time_since_send = 0;
5256 				if (bbr->r_ctl.rc_last_delay_val > time_since_send) {
5257 					/* Cut down our slot time */
5258 					bbr->r_ctl.rc_last_delay_val -= time_since_send;
5259 				} else {
5260 					bbr->r_ctl.rc_last_delay_val = 0;
5261 				}
5262 				bbr->rc_pacer_started = cts;
5263 			}
5264 		}
5265 		bbr->rc_timer_first = 0;
5266 		bbr_log_to_cancel(bbr, line, cts, hpts_removed);
5267 		bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5268 		bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5269 	}
5270 }
5271 
5272 static int
bbr_stopall(struct tcpcb * tp)5273 bbr_stopall(struct tcpcb *tp)
5274 {
5275 	struct tcp_bbr *bbr;
5276 
5277 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
5278 	bbr->rc_all_timers_stopped = 1;
5279 
5280 	tcp_hpts_remove(tp);
5281 
5282 	return (0);
5283 }
5284 
5285 static uint32_t
bbr_get_earliest_send_outstanding(struct tcp_bbr * bbr,struct bbr_sendmap * u_rsm,uint32_t cts)5286 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts)
5287 {
5288 	struct bbr_sendmap *rsm;
5289 
5290 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
5291 	if ((rsm == NULL) || (u_rsm == rsm))
5292 		return (cts);
5293 	return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
5294 }
5295 
5296 static void
bbr_update_rsm(struct tcpcb * tp,struct tcp_bbr * bbr,struct bbr_sendmap * rsm,uint32_t cts,uint32_t pacing_time)5297 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr,
5298      struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time)
5299 {
5300 	int32_t idx;
5301 
5302 	rsm->r_rtr_cnt++;
5303 	rsm->r_dupack = 0;
5304 	if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) {
5305 		rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS;
5306 		rsm->r_flags |= BBR_OVERMAX;
5307 	}
5308 	if (rsm->r_flags & BBR_RWND_COLLAPSED) {
5309 		/* Take off the collapsed flag at rxt */
5310 		rsm->r_flags &= ~BBR_RWND_COLLAPSED;
5311 	}
5312 	if (rsm->r_flags & BBR_MARKED_LOST) {
5313 		/* We have retransmitted, its no longer lost */
5314 		rsm->r_flags &= ~BBR_MARKED_LOST;
5315 		bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
5316 	}
5317 	if (rsm->r_flags & BBR_RXT_CLEARED) {
5318 		/*
5319 		 * We hit a RXT timer on it and
5320 		 * we cleared the "acked" flag.
5321 		 * We now have it going back into
5322 		 * flight, we can remove the cleared
5323 		 * flag and possibly do accounting on
5324 		 * this piece.
5325 		 */
5326 		rsm->r_flags &= ~BBR_RXT_CLEARED;
5327 	}
5328 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) {
5329 		bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5330 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5331 	}
5332 	idx = rsm->r_rtr_cnt - 1;
5333 	rsm->r_tim_lastsent[idx] = cts;
5334 	rsm->r_pacing_delay = pacing_time;
5335 	rsm->r_delivered = bbr->r_ctl.rc_delivered;
5336 	rsm->r_ts_valid = bbr->rc_ts_valid;
5337 	if (bbr->rc_ts_valid)
5338 		rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5339 	if (bbr->r_ctl.r_app_limited_until)
5340 		rsm->r_app_limited = 1;
5341 	else
5342 		rsm->r_app_limited = 0;
5343 	if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5344 		rsm->r_bbr_state = bbr_state_val(bbr);
5345 	else
5346 		rsm->r_bbr_state = 8;
5347 	if (rsm->r_flags & BBR_ACKED) {
5348 		/* Problably MTU discovery messing with us */
5349 		uint32_t old_flags;
5350 
5351 		old_flags = rsm->r_flags;
5352 		rsm->r_flags &= ~BBR_ACKED;
5353 		bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
5354 		bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5355 		if (bbr->r_ctl.rc_sacked == 0)
5356 			bbr->r_ctl.rc_sacklast = NULL;
5357 	}
5358 	if (rsm->r_in_tmap) {
5359 		TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5360 	}
5361 	TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5362 	rsm->r_in_tmap = 1;
5363 	if (rsm->r_flags & BBR_SACK_PASSED) {
5364 		/* We have retransmitted due to the SACK pass */
5365 		rsm->r_flags &= ~BBR_SACK_PASSED;
5366 		rsm->r_flags |= BBR_WAS_SACKPASS;
5367 	}
5368 	rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5369 	rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5370 						(bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5371 	bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
5372 	if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5373 		rsm->r_is_gain = 1;
5374 		rsm->r_is_drain = 0;
5375 	} else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5376 		rsm->r_is_drain = 1;
5377 		rsm->r_is_gain = 0;
5378 	} else {
5379 		rsm->r_is_drain = 0;
5380 		rsm->r_is_gain = 0;
5381 	}
5382 	rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */
5383 }
5384 
5385 /*
5386  * Returns 0, or the sequence where we stopped
5387  * updating. We also update the lenp to be the amount
5388  * of data left.
5389  */
5390 
5391 static uint32_t
bbr_update_entry(struct tcpcb * tp,struct tcp_bbr * bbr,struct bbr_sendmap * rsm,uint32_t cts,int32_t * lenp,uint32_t pacing_time)5392 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr,
5393     struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time)
5394 {
5395 	/*
5396 	 * We (re-)transmitted starting at rsm->r_start for some length
5397 	 * (possibly less than r_end.
5398 	 */
5399 	struct bbr_sendmap *nrsm;
5400 	uint32_t c_end;
5401 	int32_t len;
5402 
5403 	len = *lenp;
5404 	c_end = rsm->r_start + len;
5405 	if (SEQ_GEQ(c_end, rsm->r_end)) {
5406 		/*
5407 		 * We retransmitted the whole piece or more than the whole
5408 		 * slopping into the next rsm.
5409 		 */
5410 		bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5411 		if (c_end == rsm->r_end) {
5412 			*lenp = 0;
5413 			return (0);
5414 		} else {
5415 			int32_t act_len;
5416 
5417 			/* Hangs over the end return whats left */
5418 			act_len = rsm->r_end - rsm->r_start;
5419 			*lenp = (len - act_len);
5420 			return (rsm->r_end);
5421 		}
5422 		/* We don't get out of this block. */
5423 	}
5424 	/*
5425 	 * Here we retransmitted less than the whole thing which means we
5426 	 * have to split this into what was transmitted and what was not.
5427 	 */
5428 	nrsm = bbr_alloc_full_limit(bbr);
5429 	if (nrsm == NULL) {
5430 		*lenp = 0;
5431 		return (0);
5432 	}
5433 	/*
5434 	 * So here we are going to take the original rsm and make it what we
5435 	 * retransmitted. nrsm will be the tail portion we did not
5436 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
5437 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
5438 	 * 1, 6 and the new piece will be 6, 11.
5439 	 */
5440 	bbr_clone_rsm(bbr, nrsm, rsm, c_end);
5441 	TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
5442 	nrsm->r_dupack = 0;
5443 	if (rsm->r_in_tmap) {
5444 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5445 		nrsm->r_in_tmap = 1;
5446 	}
5447 	rsm->r_flags &= (~BBR_HAS_FIN);
5448 	bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5449 	*lenp = 0;
5450 	return (0);
5451 }
5452 
5453 static uint64_t
bbr_get_hardware_rate(struct tcp_bbr * bbr)5454 bbr_get_hardware_rate(struct tcp_bbr *bbr)
5455 {
5456 	uint64_t bw;
5457 
5458 	bw = bbr_get_bw(bbr);
5459 	bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN];
5460 	bw /= (uint64_t)BBR_UNIT;
5461 	return(bw);
5462 }
5463 
5464 static void
bbr_setup_less_of_rate(struct tcp_bbr * bbr,uint32_t cts,uint64_t act_rate,uint64_t rate_wanted)5465 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts,
5466 		       uint64_t act_rate, uint64_t rate_wanted)
5467 {
5468 	/*
5469 	 * We could not get a full gains worth
5470 	 * of rate.
5471 	 */
5472 	if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) {
5473 		/* we can't even get the real rate */
5474 		uint64_t red;
5475 
5476 		bbr->skip_gain = 1;
5477 		bbr->gain_is_limited = 0;
5478 		red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate;
5479 		if (red)
5480 			filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts);
5481 	} else {
5482 		/* We can use a lower gain */
5483 		bbr->skip_gain = 0;
5484 		bbr->gain_is_limited = 1;
5485 	}
5486 }
5487 
5488 static void
bbr_update_hardware_pacing_rate(struct tcp_bbr * bbr,uint32_t cts)5489 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts)
5490 {
5491 	const struct tcp_hwrate_limit_table *nrte;
5492 	int error, rate = -1;
5493 
5494 	if (bbr->r_ctl.crte == NULL)
5495 		return;
5496 	if ((bbr->rc_inp->inp_route.ro_nh == NULL) ||
5497 	    (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) {
5498 		/* Lost our routes? */
5499 		/* Clear the way for a re-attempt */
5500 		bbr->bbr_attempt_hdwr_pace = 0;
5501 lost_rate:
5502 		bbr->gain_is_limited = 0;
5503 		bbr->skip_gain = 0;
5504 		bbr->bbr_hdrw_pacing = 0;
5505 		counter_u64_add(bbr_flows_whdwr_pacing, -1);
5506 		counter_u64_add(bbr_flows_nohdwr_pacing, 1);
5507 		tcp_bbr_tso_size_check(bbr, cts);
5508 		return;
5509 	}
5510 	rate = bbr_get_hardware_rate(bbr);
5511 	nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte,
5512 				   bbr->rc_tp,
5513 				   bbr->rc_inp->inp_route.ro_nh->nh_ifp,
5514 				   rate,
5515 				   (RS_PACING_GEQ|RS_PACING_SUB_OK),
5516 				   &error, NULL);
5517 	if (nrte == NULL) {
5518 		goto lost_rate;
5519 	}
5520 	if (nrte != bbr->r_ctl.crte) {
5521 		bbr->r_ctl.crte = nrte;
5522 		if (error == 0)  {
5523 			BBR_STAT_INC(bbr_hdwr_rl_mod_ok);
5524 			if (bbr->r_ctl.crte->rate < rate) {
5525 				/* We have a problem */
5526 				bbr_setup_less_of_rate(bbr, cts,
5527 						       bbr->r_ctl.crte->rate, rate);
5528 			} else {
5529 				/* We are good */
5530 				bbr->gain_is_limited = 0;
5531 				bbr->skip_gain = 0;
5532 			}
5533 		} else {
5534 			/* A failure should release the tag */
5535 			BBR_STAT_INC(bbr_hdwr_rl_mod_fail);
5536 			bbr->gain_is_limited = 0;
5537 			bbr->skip_gain = 0;
5538 			bbr->bbr_hdrw_pacing = 0;
5539 		}
5540 		bbr_type_log_hdwr_pacing(bbr,
5541 					 bbr->r_ctl.crte->ptbl->rs_ifp,
5542 					 rate,
5543 					 bbr->r_ctl.crte->rate,
5544 					 __LINE__,
5545 					 cts,
5546 					 error);
5547 	}
5548 }
5549 
5550 static void
bbr_adjust_for_hw_pacing(struct tcp_bbr * bbr,uint32_t cts)5551 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts)
5552 {
5553 	/*
5554 	 * If we have hardware pacing support
5555 	 * we need to factor that in for our
5556 	 * TSO size.
5557 	 */
5558 	const struct tcp_hwrate_limit_table *rlp;
5559 	uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay;
5560 
5561 	if ((bbr->bbr_hdrw_pacing == 0) ||
5562 	    (IN_RECOVERY(bbr->rc_tp->t_flags)) ||
5563 	    (bbr->r_ctl.crte == NULL))
5564 		return;
5565 	if (bbr->hw_pacing_set == 0) {
5566 		/* Not yet by the hdwr pacing count delay */
5567 		return;
5568 	}
5569 	if (bbr_hdwr_pace_adjust == 0) {
5570 		/* No adjustment */
5571 		return;
5572 	}
5573 	rlp = bbr->r_ctl.crte;
5574 	if (bbr->rc_tp->t_maxseg > bbr->rc_last_options)
5575 		maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5576 	else
5577 		maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5578 	/*
5579 	 * So lets first get the
5580 	 * time we will take between
5581 	 * TSO sized sends currently without
5582 	 * hardware help.
5583 	 */
5584 	cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT,
5585 		        bbr->r_ctl.rc_pace_max_segs, cts, 1);
5586 	hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg;
5587 	hdwr_delay *= rlp->time_between;
5588 	if (cur_delay > hdwr_delay)
5589 		delta = cur_delay - hdwr_delay;
5590 	else
5591 		delta = 0;
5592 	bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay,
5593 			     (bbr->r_ctl.rc_pace_max_segs / maxseg),
5594 			     1);
5595 	if (delta &&
5596 	    (delta < (max(rlp->time_between,
5597 			  bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) {
5598 		/*
5599 		 * Now lets divide by the pacing
5600 		 * time between each segment the
5601 		 * hardware sends rounding up and
5602 		 * derive a bytes from that. We multiply
5603 		 * that by bbr_hdwr_pace_adjust to get
5604 		 * more bang for our buck.
5605 		 *
5606 		 * The goal is to have the software pacer
5607 		 * waiting no more than an additional
5608 		 * pacing delay if we can (without the
5609 		 * compensation i.e. x bbr_hdwr_pace_adjust).
5610 		 */
5611 		seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between),
5612 			     (bbr->r_ctl.rc_pace_max_segs/maxseg));
5613 		seg_sz *= bbr_hdwr_pace_adjust;
5614 		if (bbr_hdwr_pace_floor &&
5615 		    (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5616 			/* Currently hardware paces
5617 			 * out rs_min_seg segments at a time.
5618 			 * We need to make sure we always send at least
5619 			 * a full burst of bbr_hdwr_pace_floor down.
5620 			 */
5621 			seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5622 		}
5623 		seg_sz *= maxseg;
5624 	} else if (delta == 0) {
5625 		/*
5626 		 * The highest pacing rate is
5627 		 * above our b/w gained. This means
5628 		 * we probably are going quite fast at
5629 		 * the hardware highest rate. Lets just multiply
5630 		 * the calculated TSO size by the
5631 		 * multiplier factor (its probably
5632 		 * 4 segments in the default config for
5633 		 * mlx).
5634 		 */
5635 		seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust;
5636 		if (bbr_hdwr_pace_floor &&
5637 		    (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5638 			/* Currently hardware paces
5639 			 * out rs_min_seg segments at a time.
5640 			 * We need to make sure we always send at least
5641 			 * a full burst of bbr_hdwr_pace_floor down.
5642 			 */
5643 			seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5644 		}
5645 	} else {
5646 		/*
5647 		 * The pacing time difference is so
5648 		 * big that the hardware will
5649 		 * pace out more rapidly then we
5650 		 * really want and then we
5651 		 * will have a long delay. Lets just keep
5652 		 * the same TSO size so its as if
5653 		 * we were not using hdwr pacing (we
5654 		 * just gain a bit of spacing from the
5655 		 * hardware if seg_sz > 1).
5656 		 */
5657 		seg_sz = bbr->r_ctl.rc_pace_max_segs;
5658 	}
5659 	if (seg_sz > bbr->r_ctl.rc_pace_max_segs)
5660 		new_tso = seg_sz;
5661 	else
5662 		new_tso = bbr->r_ctl.rc_pace_max_segs;
5663 	if (new_tso >= (PACE_MAX_IP_BYTES-maxseg))
5664 		new_tso = PACE_MAX_IP_BYTES - maxseg;
5665 
5666 	if (new_tso != bbr->r_ctl.rc_pace_max_segs) {
5667 		bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0);
5668 		bbr->r_ctl.rc_pace_max_segs = new_tso;
5669 	}
5670 }
5671 
5672 static void
tcp_bbr_tso_size_check(struct tcp_bbr * bbr,uint32_t cts)5673 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts)
5674 {
5675 	uint64_t bw;
5676 	uint32_t old_tso = 0, new_tso;
5677 	uint32_t maxseg, bytes;
5678 	uint32_t tls_seg=0;
5679 	/*
5680 	 * Google/linux uses the following algorithm to determine
5681 	 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18):
5682 	 *
5683 	 *  bytes = bw_in_bytes_per_second / 1000
5684 	 *  bytes = min(bytes, 64k)
5685 	 *  tso_segs = bytes / MSS
5686 	 *  if (bw < 1.2Mbs)
5687 	 *      min_tso_segs = 1
5688 	 *  else
5689 	 *	min_tso_segs = 2
5690 	 * tso_segs = max(tso_segs, min_tso_segs)
5691 	 *
5692 	 * * Note apply a device specific limit (we apply this in the
5693 	 *   tcp_m_copym).
5694 	 * Note that before the initial measurement is made google bursts out
5695 	 * a full iwnd just like new-reno/cubic.
5696 	 *
5697 	 * We do not use this algorithm. Instead we
5698 	 * use a two phased approach:
5699 	 *
5700 	 *  if ( bw <= per-tcb-cross-over)
5701 	 *     goal_tso =  calculate how much with this bw we
5702 	 *                 can send in goal-time seconds.
5703 	 *     if (goal_tso > mss)
5704 	 *         seg = goal_tso / mss
5705 	 *         tso = seg * mss
5706 	 *     else
5707 	 *         tso = mss
5708 	 *     if (tso > per-tcb-max)
5709 	 *         tso = per-tcb-max
5710 	 *  else if ( bw > 512Mbps)
5711 	 *     tso = max-tso (64k/mss)
5712 	 *  else
5713 	 *     goal_tso = bw / per-tcb-divsor
5714 	 *     seg = (goal_tso + mss-1)/mss
5715 	 *     tso = seg * mss
5716 	 *
5717 	 * if (tso < per-tcb-floor)
5718 	 *    tso = per-tcb-floor
5719 	 * if (tso > per-tcb-utter_max)
5720 	 *    tso = per-tcb-utter_max
5721 	 *
5722 	 * Note the default per-tcb-divisor is 1000 (same as google).
5723 	 * the goal cross over is 30Mbps however. To recreate googles
5724 	 * algorithm you need to set:
5725 	 *
5726 	 * cross-over = 23,168,000 bps
5727 	 * goal-time = 18000
5728 	 * per-tcb-max = 2
5729 	 * per-tcb-divisor = 1000
5730 	 * per-tcb-floor = 1
5731 	 *
5732 	 * This will get you "google bbr" behavior with respect to tso size.
5733 	 *
5734 	 * Note we do set anything TSO size until we are past the initial
5735 	 * window. Before that we gnerally use either a single MSS
5736 	 * or we use the full IW size (so we burst a IW at a time)
5737 	 */
5738 
5739 	if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) {
5740 		maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5741 	} else {
5742 		maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5743 	}
5744 	old_tso = bbr->r_ctl.rc_pace_max_segs;
5745 	if (bbr->rc_past_init_win == 0) {
5746 		/*
5747 		 * Not enough data has been acknowledged to make a
5748 		 * judgement. Set up the initial TSO based on if we
5749 		 * are sending a full IW at once or not.
5750 		 */
5751 		if (bbr->rc_use_google)
5752 			bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2);
5753 		else if (bbr->bbr_init_win_cheat)
5754 			bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp);
5755 		else
5756 			bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5757 		if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg)
5758 			bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg;
5759 		if (bbr->r_ctl.rc_pace_max_segs == 0) {
5760 			bbr->r_ctl.rc_pace_max_segs = maxseg;
5761 		}
5762 		bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0);
5763 			bbr_adjust_for_hw_pacing(bbr, cts);
5764 		return;
5765 	}
5766 	/**
5767 	 * Now lets set the TSO goal based on our delivery rate in
5768 	 * bytes per second. Note we only do this if
5769 	 * we have acked at least the initial cwnd worth of data.
5770 	 */
5771 	bw = bbr_get_bw(bbr);
5772 	if (IN_RECOVERY(bbr->rc_tp->t_flags) &&
5773 	     (bbr->rc_use_google == 0)) {
5774 		/* We clamp to one MSS in recovery */
5775 		new_tso = maxseg;
5776 	} else if (bbr->rc_use_google) {
5777 		int min_tso_segs;
5778 
5779 		/* Google considers the gain too */
5780 		if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) {
5781 			bw *= bbr->r_ctl.rc_bbr_hptsi_gain;
5782 			bw /= BBR_UNIT;
5783 		}
5784 		bytes = bw / 1024;
5785 		if (bytes > (64 * 1024))
5786 			bytes = 64 * 1024;
5787 		new_tso = bytes / maxseg;
5788 		if (bw < ONE_POINT_TWO_MEG)
5789 			min_tso_segs = 1;
5790 		else
5791 			min_tso_segs = 2;
5792 		if (new_tso < min_tso_segs)
5793 			new_tso = min_tso_segs;
5794 		new_tso *= maxseg;
5795 	} else if (bbr->rc_no_pacing) {
5796 		new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg;
5797 	} else if (bw <= bbr->r_ctl.bbr_cross_over) {
5798 		/*
5799 		 * Calculate the worse case b/w TSO if we are inserting no
5800 		 * more than a delay_target number of TSO's.
5801 		 */
5802 		uint32_t tso_len, min_tso;
5803 
5804 		tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw);
5805 		if (tso_len > maxseg) {
5806 			new_tso = tso_len / maxseg;
5807 			if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max)
5808 				new_tso = bbr->r_ctl.bbr_hptsi_segments_max;
5809 			new_tso *= maxseg;
5810 		} else {
5811 			/*
5812 			 * less than a full sized frame yikes.. long rtt or
5813 			 * low bw?
5814 			 */
5815 			min_tso = bbr_minseg(bbr);
5816 			if ((tso_len > min_tso) && (bbr_all_get_min == 0))
5817 				new_tso = rounddown(tso_len, min_tso);
5818 			else
5819 				new_tso = min_tso;
5820 		}
5821 	} else if (bw > FIVETWELVE_MBPS) {
5822 		/*
5823 		 * This guy is so fast b/w wise that we can TSO as large as
5824 		 * possible of segments that the NIC will allow.
5825 		 */
5826 		new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5827 	} else {
5828 		/*
5829 		 * This formula is based on attempting to send a segment or
5830 		 * more every bbr_hptsi_per_second. The default is 1000
5831 		 * which means you are targeting what you can send every 1ms
5832 		 * based on the peers bw.
5833 		 *
5834 		 * If the number drops to say 500, then you are looking more
5835 		 * at 2ms and you will raise how much we send in a single
5836 		 * TSO thus saving CPU (less bbr_output_wtime() calls). The
5837 		 * trade off of course is you will send more at once and
5838 		 * thus tend to clump up the sends into larger "bursts"
5839 		 * building a queue.
5840 		 */
5841 		bw /= bbr->r_ctl.bbr_hptsi_per_second;
5842 		new_tso = roundup(bw, (uint64_t)maxseg);
5843 		/*
5844 		 * Gate the floor to match what our lower than 48Mbps
5845 		 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus
5846 		 * becomes the floor for this calculation.
5847 		 */
5848 		if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg))
5849 			new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg);
5850 	}
5851 	if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor)))
5852 		new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor;
5853 	if (new_tso > PACE_MAX_IP_BYTES)
5854 		new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5855 	/* Enforce an utter maximum. */
5856 	if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) {
5857 		new_tso = bbr->r_ctl.bbr_utter_max * maxseg;
5858 	}
5859 	if (old_tso != new_tso) {
5860 		/* Only log changes */
5861 		bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0);
5862 		bbr->r_ctl.rc_pace_max_segs = new_tso;
5863 	}
5864 	/* We have hardware pacing! */
5865 	bbr_adjust_for_hw_pacing(bbr, cts);
5866 }
5867 
5868 static void
bbr_log_output(struct tcp_bbr * bbr,struct tcpcb * tp,struct tcpopt * to,int32_t len,uint32_t seq_out,uint16_t th_flags,int32_t err,uint32_t cts,struct mbuf * mb,int32_t * abandon,struct bbr_sendmap * hintrsm,uint32_t delay_calc,struct sockbuf * sb)5869 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len,
5870     uint32_t seq_out, uint16_t th_flags, int32_t err, uint32_t cts,
5871     struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc,
5872     struct sockbuf *sb)
5873 {
5874 
5875 	struct bbr_sendmap *rsm, *nrsm;
5876 	register uint32_t snd_max, snd_una;
5877 	uint32_t pacing_time;
5878 	/*
5879 	 * Add to the RACK log of packets in flight or retransmitted. If
5880 	 * there is a TS option we will use the TS echoed, if not we will
5881 	 * grab a TS.
5882 	 *
5883 	 * Retransmissions will increment the count and move the ts to its
5884 	 * proper place. Note that if options do not include TS's then we
5885 	 * won't be able to effectively use the ACK for an RTT on a retran.
5886 	 *
5887 	 * Notes about r_start and r_end. Lets consider a send starting at
5888 	 * sequence 1 for 10 bytes. In such an example the r_start would be
5889 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
5890 	 * This means that r_end is actually the first sequence for the next
5891 	 * slot (11).
5892 	 *
5893 	 */
5894 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5895 	if (err) {
5896 		/*
5897 		 * We don't log errors -- we could but snd_max does not
5898 		 * advance in this case either.
5899 		 */
5900 		return;
5901 	}
5902 	if (th_flags & TH_RST) {
5903 		/*
5904 		 * We don't log resets and we return immediately from
5905 		 * sending
5906 		 */
5907 		*abandon = 1;
5908 		return;
5909 	}
5910 	snd_una = tp->snd_una;
5911 	if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) {
5912 		/*
5913 		 * The call to bbr_log_output is made before bumping
5914 		 * snd_max. This means we can record one extra byte on a SYN
5915 		 * or FIN if seq_out is adding more on and a FIN is present
5916 		 * (and we are not resending).
5917 		 */
5918 		if ((th_flags & TH_SYN) && (tp->iss == seq_out))
5919 			len++;
5920 		if (th_flags & TH_FIN)
5921 			len++;
5922 	}
5923 	if (SEQ_LEQ((seq_out + len), snd_una)) {
5924 		/* Are sending an old segment to induce an ack (keep-alive)? */
5925 		return;
5926 	}
5927 	if (SEQ_LT(seq_out, snd_una)) {
5928 		/* huh? should we panic? */
5929 		uint32_t end;
5930 
5931 		end = seq_out + len;
5932 		seq_out = snd_una;
5933 		len = end - seq_out;
5934 	}
5935 	snd_max = tp->snd_max;
5936 	if (len == 0) {
5937 		/* We don't log zero window probes */
5938 		return;
5939 	}
5940 	pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1);
5941 	/* First question is it a retransmission? */
5942 	if (seq_out == snd_max) {
5943 again:
5944 		rsm = bbr_alloc(bbr);
5945 		if (rsm == NULL) {
5946 			return;
5947 		}
5948 		rsm->r_flags = 0;
5949 		if (th_flags & TH_SYN)
5950 			rsm->r_flags |= BBR_HAS_SYN;
5951 		if (th_flags & TH_FIN)
5952 			rsm->r_flags |= BBR_HAS_FIN;
5953 		rsm->r_tim_lastsent[0] = cts;
5954 		rsm->r_rtr_cnt = 1;
5955 		rsm->r_rtr_bytes = 0;
5956 		rsm->r_start = seq_out;
5957 		rsm->r_end = rsm->r_start + len;
5958 		rsm->r_dupack = 0;
5959 		rsm->r_delivered = bbr->r_ctl.rc_delivered;
5960 		rsm->r_pacing_delay = pacing_time;
5961 		rsm->r_ts_valid = bbr->rc_ts_valid;
5962 		if (bbr->rc_ts_valid)
5963 			rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5964 		rsm->r_del_time = bbr->r_ctl.rc_del_time;
5965 		if (bbr->r_ctl.r_app_limited_until)
5966 			rsm->r_app_limited = 1;
5967 		else
5968 			rsm->r_app_limited = 0;
5969 		rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5970 		rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5971 						(bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5972 		/*
5973 		 * Here we must also add in this rsm since snd_max
5974 		 * is updated after we return from a new send.
5975 		 */
5976 		rsm->r_flight_at_send += len;
5977 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
5978 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5979 		rsm->r_in_tmap = 1;
5980 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5981 			rsm->r_bbr_state = bbr_state_val(bbr);
5982 		else
5983 			rsm->r_bbr_state = 8;
5984 		if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5985 			rsm->r_is_gain = 1;
5986 			rsm->r_is_drain = 0;
5987 		} else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5988 			rsm->r_is_drain = 1;
5989 			rsm->r_is_gain = 0;
5990 		} else {
5991 			rsm->r_is_drain = 0;
5992 			rsm->r_is_gain = 0;
5993 		}
5994 		return;
5995 	}
5996 	/*
5997 	 * If we reach here its a retransmission and we need to find it.
5998 	 */
5999 more:
6000 	if (hintrsm && (hintrsm->r_start == seq_out)) {
6001 		rsm = hintrsm;
6002 		hintrsm = NULL;
6003 	} else if (bbr->r_ctl.rc_next) {
6004 		/* We have a hint from a previous run */
6005 		rsm = bbr->r_ctl.rc_next;
6006 	} else {
6007 		/* No hints sorry */
6008 		rsm = NULL;
6009 	}
6010 	if ((rsm) && (rsm->r_start == seq_out)) {
6011 		/*
6012 		 * We used rc_next or hintrsm  to retransmit, hopefully the
6013 		 * likely case.
6014 		 */
6015 		seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6016 		if (len == 0) {
6017 			return;
6018 		} else {
6019 			goto more;
6020 		}
6021 	}
6022 	/* Ok it was not the last pointer go through it the hard way. */
6023 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6024 		if (rsm->r_start == seq_out) {
6025 			seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6026 			bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
6027 			if (len == 0) {
6028 				return;
6029 			} else {
6030 				continue;
6031 			}
6032 		}
6033 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6034 			/* Transmitted within this piece */
6035 			/*
6036 			 * Ok we must split off the front and then let the
6037 			 * update do the rest
6038 			 */
6039 			nrsm = bbr_alloc_full_limit(bbr);
6040 			if (nrsm == NULL) {
6041 				bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
6042 				return;
6043 			}
6044 			/*
6045 			 * copy rsm to nrsm and then trim the front of rsm
6046 			 * to not include this part.
6047 			 */
6048 			bbr_clone_rsm(bbr, nrsm, rsm, seq_out);
6049 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
6050 			if (rsm->r_in_tmap) {
6051 				TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6052 				nrsm->r_in_tmap = 1;
6053 			}
6054 			rsm->r_flags &= (~BBR_HAS_FIN);
6055 			seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time);
6056 			if (len == 0) {
6057 				return;
6058 			}
6059 		}
6060 	}
6061 	/*
6062 	 * Hmm not found in map did they retransmit both old and on into the
6063 	 * new?
6064 	 */
6065 	if (seq_out == tp->snd_max) {
6066 		goto again;
6067 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
6068 #ifdef BBR_INVARIANTS
6069 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6070 		    seq_out, len, tp->snd_una, tp->snd_max);
6071 		printf("Starting Dump of all rack entries\n");
6072 		TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6073 			printf("rsm:%p start:%u end:%u\n",
6074 			    rsm, rsm->r_start, rsm->r_end);
6075 		}
6076 		printf("Dump complete\n");
6077 		panic("seq_out not found rack:%p tp:%p",
6078 		    bbr, tp);
6079 #endif
6080 	} else {
6081 #ifdef BBR_INVARIANTS
6082 		/*
6083 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
6084 		 * flag)
6085 		 */
6086 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6087 		    seq_out, len, tp->snd_max, tp);
6088 #endif
6089 	}
6090 }
6091 
6092 static void
bbr_collapse_rtt(struct tcpcb * tp,struct tcp_bbr * bbr,int32_t rtt)6093 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt)
6094 {
6095 	/*
6096 	 * Collapse timeout back the cum-ack moved.
6097 	 */
6098 	tp->t_rxtshift = 0;
6099 	tp->t_softerror = 0;
6100 }
6101 
6102 static void
tcp_bbr_xmit_timer(struct tcp_bbr * bbr,uint32_t rtt_usecs,uint32_t rsm_send_time,uint32_t r_start,uint32_t tsin)6103 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin)
6104 {
6105 	bbr->rtt_valid = 1;
6106 	bbr->r_ctl.cur_rtt = rtt_usecs;
6107 	bbr->r_ctl.ts_in = tsin;
6108 	if (rsm_send_time)
6109 		bbr->r_ctl.cur_rtt_send_time = rsm_send_time;
6110 }
6111 
6112 static void
bbr_make_timestamp_determination(struct tcp_bbr * bbr)6113 bbr_make_timestamp_determination(struct tcp_bbr *bbr)
6114 {
6115 	/**
6116 	 * We have in our bbr control:
6117 	 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp).
6118 	 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts).
6119 	 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts)
6120 	 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time)
6121 	 *
6122 	 * Now we can calculate the time between the sends by doing:
6123 	 *
6124 	 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts
6125 	 *
6126 	 * And the peer's time between receiving them by doing:
6127 	 *
6128 	 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp
6129 	 *
6130 	 * We want to figure out if the timestamp values are in msec, 10msec or usec.
6131 	 * We also may find that we can't use the timestamps if say we see
6132 	 * that the peer_delta indicates that though we may have taken 10ms to
6133 	 * pace out the data, it only saw 1ms between the two packets. This would
6134 	 * indicate that somewhere on the path is a batching entity that is giving
6135 	 * out time-slices of the actual b/w. This would mean we could not use
6136 	 * reliably the peers timestamps.
6137 	 *
6138 	 * We expect delta > peer_delta initially. Until we figure out the
6139 	 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio.
6140 	 * If we place 1000 there then its a ms vs our usec. If we place 10000 there
6141 	 * then its 10ms vs our usec. If the peer is running a usec clock we would
6142 	 * put a 1 there. If the value is faster then ours, we will disable the
6143 	 * use of timestamps (though we could revist this later if we find it to be not
6144 	 * just an isolated one or two flows)).
6145 	 *
6146 	 * To detect the batching middle boxes we will come up with our compensation and
6147 	 * if with it in place, we find the peer is drastically off (by some margin) in
6148 	 * the smaller direction, then we will assume the worst case and disable use of timestamps.
6149 	 *
6150 	 */
6151 	uint64_t delta, peer_delta, delta_up;
6152 
6153 	delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts;
6154 	if (delta < bbr_min_usec_delta) {
6155 		/*
6156 		 * Have not seen a min amount of time
6157 		 * between our send times so we can
6158 		 * make a determination of the timestamp
6159 		 * yet.
6160 		 */
6161 		return;
6162 	}
6163 	peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp;
6164 	if (peer_delta < bbr_min_peer_delta) {
6165 		/*
6166 		 * We may have enough in the form of
6167 		 * our delta but the peers number
6168 		 * has not changed that much. It could
6169 		 * be its clock ratio is such that
6170 		 * we need more data (10ms tick) or
6171 		 * there may be other compression scenarios
6172 		 * going on. In any event we need the
6173 		 * spread to be larger.
6174 		 */
6175 		return;
6176 	}
6177 	/* Ok lets first see which way our delta is going */
6178 	if (peer_delta > delta) {
6179 		/* Very unlikely, the peer without
6180 		 * compensation shows that it saw
6181 		 * the two sends arrive further apart
6182 		 * then we saw then in micro-seconds.
6183 		 */
6184 		if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) {
6185 			/* well it looks like the peer is a micro-second clock. */
6186 			bbr->rc_ts_clock_set = 1;
6187 			bbr->r_ctl.bbr_peer_tsratio = 1;
6188 		} else {
6189 			bbr->rc_ts_cant_be_used = 1;
6190 			bbr->rc_ts_clock_set = 1;
6191 		}
6192 		return;
6193 	}
6194 	/* Ok we know that the peer_delta is smaller than our send distance */
6195 	bbr->rc_ts_clock_set = 1;
6196 	/* First question is it within the percentage that they are using usec time? */
6197 	delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent;
6198 	if ((peer_delta + delta_up) >= delta) {
6199 		/* Its a usec clock */
6200 		bbr->r_ctl.bbr_peer_tsratio = 1;
6201 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6202 		return;
6203 	}
6204 	/* Ok if not usec, what about 10usec (though unlikely)? */
6205 	delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent;
6206 	if (((peer_delta * 10) + delta_up) >= delta) {
6207 		bbr->r_ctl.bbr_peer_tsratio = 10;
6208 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6209 		return;
6210 	}
6211 	/* And what about 100usec (though again unlikely)? */
6212 	delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent;
6213 	if (((peer_delta * 100) + delta_up) >= delta) {
6214 		bbr->r_ctl.bbr_peer_tsratio = 100;
6215 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6216 		return;
6217 	}
6218 	/* And how about 1 msec (the most likely one)? */
6219 	delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent;
6220 	if (((peer_delta * 1000) + delta_up) >= delta) {
6221 		bbr->r_ctl.bbr_peer_tsratio = 1000;
6222 		bbr_log_tstmp_validation(bbr, peer_delta, delta);
6223 		return;
6224 	}
6225 	/* Ok if not msec could it be 10 msec? */
6226 	delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent;
6227 	if (((peer_delta * 10000) + delta_up) >= delta) {
6228 		bbr->r_ctl.bbr_peer_tsratio = 10000;
6229 		return;
6230 	}
6231 	/* If we fall down here the clock tick so slowly we can't use it */
6232 	bbr->rc_ts_cant_be_used = 1;
6233 	bbr->r_ctl.bbr_peer_tsratio = 0;
6234 	bbr_log_tstmp_validation(bbr, peer_delta, delta);
6235 }
6236 
6237 /*
6238  * Collect new round-trip time estimate
6239  * and update averages and current timeout.
6240  */
6241 static void
tcp_bbr_xmit_timer_commit(struct tcp_bbr * bbr,struct tcpcb * tp,uint32_t cts)6242 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts)
6243 {
6244 	int32_t delta;
6245 	uint32_t rtt, tsin;
6246 	int32_t rtt_ticks;
6247 
6248 	if (bbr->rtt_valid == 0)
6249 		/* No valid sample */
6250 		return;
6251 
6252 	rtt = bbr->r_ctl.cur_rtt;
6253 	tsin = bbr->r_ctl.ts_in;
6254 	if (bbr->rc_prtt_set_ts) {
6255 		/*
6256 		 * We are to force feed the rttProp filter due
6257 		 * to an entry into PROBE_RTT. This assures
6258 		 * that the times are sync'd between when we
6259 		 * go into PROBE_RTT and the filter expiration.
6260 		 *
6261 		 * Google does not use a true filter, so they do
6262 		 * this implicitly since they only keep one value
6263 		 * and when they enter probe-rtt they update the
6264 		 * value to the newest rtt.
6265 		 */
6266 		uint32_t rtt_prop;
6267 
6268 		bbr->rc_prtt_set_ts = 0;
6269 		rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
6270 		if (rtt > rtt_prop)
6271 			filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts);
6272 		else
6273 			apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6274 	}
6275 #ifdef STATS
6276 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rtt));
6277 #endif
6278 	if (bbr->rc_ack_was_delayed)
6279 		rtt += bbr->r_ctl.rc_ack_hdwr_delay;
6280 
6281 	if (rtt < bbr->r_ctl.rc_lowest_rtt)
6282 		bbr->r_ctl.rc_lowest_rtt = rtt;
6283 	bbr_log_rtt_sample(bbr, rtt, tsin);
6284 	if (bbr->r_init_rtt) {
6285 		/*
6286 		 * The initial rtt is not-trusted, nuke it and lets get
6287 		 * our first valid measurement in.
6288 		 */
6289 		bbr->r_init_rtt = 0;
6290 		tp->t_srtt = 0;
6291 	}
6292 	if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) {
6293 		/*
6294 		 * So we have not yet figured out
6295 		 * what the peers TSTMP value is
6296 		 * in (most likely ms). We need a
6297 		 * series of cum-ack's to determine
6298 		 * this reliably.
6299 		 */
6300 		if (bbr->rc_ack_is_cumack) {
6301 			if (bbr->rc_ts_data_set) {
6302 				/* Lets attempt to determine the timestamp granularity. */
6303 				bbr_make_timestamp_determination(bbr);
6304 			} else {
6305 				bbr->rc_ts_data_set = 1;
6306 				bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts;
6307 				bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time;
6308 			}
6309 		} else {
6310 			/*
6311 			 * We have to have consecutive acks
6312 			 * reset any "filled" state to none.
6313 			 */
6314 			bbr->rc_ts_data_set = 0;
6315 		}
6316 	}
6317 	/* Round it up */
6318 	rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1)));
6319 	if (tp->t_srtt != 0) {
6320 		/*
6321 		 * srtt is stored as fixed point with 5 bits after the
6322 		 * binary point (i.e., scaled by 8).  The following magic is
6323 		 * equivalent to the smoothing algorithm in rfc793 with an
6324 		 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6325 		 * Adjust rtt to origin 0.
6326 		 */
6327 
6328 		delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT)
6329 		    - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6330 
6331 		tp->t_srtt += delta;
6332 		if (tp->t_srtt <= 0)
6333 			tp->t_srtt = 1;
6334 
6335 		/*
6336 		 * We accumulate a smoothed rtt variance (actually, a
6337 		 * smoothed mean difference), then set the retransmit timer
6338 		 * to smoothed rtt + 4 times the smoothed variance. rttvar
6339 		 * is stored as fixed point with 4 bits after the binary
6340 		 * point (scaled by 16).  The following is equivalent to
6341 		 * rfc793 smoothing with an alpha of .75 (rttvar =
6342 		 * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
6343 		 * wired-in beta.
6344 		 */
6345 		if (delta < 0)
6346 			delta = -delta;
6347 		delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6348 		tp->t_rttvar += delta;
6349 		if (tp->t_rttvar <= 0)
6350 			tp->t_rttvar = 1;
6351 	} else {
6352 		/*
6353 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
6354 		 * variance to half the rtt (so our first retransmit happens
6355 		 * at 3*rtt).
6356 		 */
6357 		tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT;
6358 		tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1);
6359 	}
6360 	KMOD_TCPSTAT_INC(tcps_rttupdated);
6361 	if (tp->t_rttupdated < UCHAR_MAX)
6362 		tp->t_rttupdated++;
6363 #ifdef STATS
6364 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks));
6365 #endif
6366 	/*
6367 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
6368 	 * way we do the smoothing, srtt and rttvar will each average +1/2
6369 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
6370 	 * tick of rounding and 1 extra tick because of +-1/2 tick
6371 	 * uncertainty in the firing of the timer.  The bias will give us
6372 	 * exactly the 1.5 tick we need.  But, because the bias is
6373 	 * statistical, we have to test that we don't drop below the minimum
6374 	 * feasible timer (which is 2 ticks).
6375 	 */
6376 	TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6377 	    max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2),
6378 	    MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
6379 
6380 	/*
6381 	 * We received an ack for a packet that wasn't retransmitted; it is
6382 	 * probably safe to discard any error indications we've received
6383 	 * recently.  This isn't quite right, but close enough for now (a
6384 	 * route might have failed after we sent a segment, and the return
6385 	 * path might not be symmetrical).
6386 	 */
6387 	tp->t_softerror = 0;
6388 	rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
6389 	if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt)
6390 		bbr->r_ctl.bbr_smallest_srtt_this_state = rtt;
6391 }
6392 
6393 static void
bbr_set_reduced_rtt(struct tcp_bbr * bbr,uint32_t cts,uint32_t line)6394 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line)
6395 {
6396 	bbr->r_ctl.rc_rtt_shrinks = cts;
6397 	if (bbr_can_force_probertt &&
6398 	    (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
6399 	    ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
6400 		/*
6401 		 * We should enter probe-rtt its been too long
6402 		 * since we have been there.
6403 		 */
6404 		bbr_enter_probe_rtt(bbr, cts, __LINE__);
6405 	} else
6406 		bbr_check_probe_rtt_limits(bbr, cts);
6407 }
6408 
6409 static void
tcp_bbr_commit_bw(struct tcp_bbr * bbr,uint32_t cts)6410 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts)
6411 {
6412 	uint64_t orig_bw;
6413 
6414 	if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) {
6415 		/* We never apply a zero measurement */
6416 		bbr_log_type_bbrupd(bbr, 20, cts, 0, 0,
6417 				    0, 0, 0, 0, 0, 0);
6418 		return;
6419 	}
6420 	if (bbr->r_ctl.r_measurement_count < 0xffffffff)
6421 		bbr->r_ctl.r_measurement_count++;
6422 	orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
6423 	apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch);
6424 	bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw,
6425 			    (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate),
6426 			    0, 0, 0, 0, 0, 0);
6427 	if (orig_bw &&
6428 	    (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) {
6429 		if (bbr->bbr_hdrw_pacing) {
6430 			/*
6431 			 * Apply a new rate to the hardware
6432 			 * possibly.
6433 			 */
6434 			bbr_update_hardware_pacing_rate(bbr, cts);
6435 		}
6436 		bbr_set_state_target(bbr, __LINE__);
6437 		tcp_bbr_tso_size_check(bbr, cts);
6438 		if (bbr->r_recovery_bw)  {
6439 			bbr_setup_red_bw(bbr, cts);
6440 			bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW);
6441 		}
6442 	} else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate))
6443 		tcp_bbr_tso_size_check(bbr, cts);
6444 }
6445 
6446 static void
bbr_nf_measurement(struct tcp_bbr * bbr,struct bbr_sendmap * rsm,uint32_t rtt,uint32_t cts)6447 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6448 {
6449 	if (bbr->rc_in_persist == 0) {
6450 		/* We log only when not in persist */
6451 		/* Translate to a Bytes Per Second */
6452 		uint64_t tim, bw, ts_diff, ts_bw;
6453 		uint32_t delivered;
6454 
6455 		if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6456 			tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6457 		else
6458 			tim = 1;
6459 		/*
6460 		 * Now that we have processed the tim (skipping the sample
6461 		 * or possibly updating the time, go ahead and
6462 		 * calculate the cdr.
6463 		 */
6464 		delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6465 		bw = (uint64_t)delivered;
6466 		bw *= (uint64_t)USECS_IN_SECOND;
6467 		bw /= tim;
6468 		if (bw == 0) {
6469 			/* We must have a calculatable amount */
6470 			return;
6471 		}
6472 		/*
6473 		 * If we are using this b/w shove it in now so we
6474 		 * can see in the trace viewer if it gets over-ridden.
6475 		 */
6476 		if (rsm->r_ts_valid &&
6477 		    bbr->rc_ts_valid &&
6478 		    bbr->rc_ts_clock_set &&
6479 		    (bbr->rc_ts_cant_be_used == 0) &&
6480 		    bbr->rc_use_ts_limit) {
6481 			ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1);
6482 			ts_diff *= bbr->r_ctl.bbr_peer_tsratio;
6483 			if ((delivered == 0) ||
6484 			    (rtt < 1000)) {
6485 				/* Can't use the ts */
6486 				bbr_log_type_bbrupd(bbr, 61, cts,
6487 						    ts_diff,
6488 						    bbr->r_ctl.last_inbound_ts,
6489 						    rsm->r_del_ack_ts, 0,
6490 						    0, 0, 0, delivered);
6491 			} else {
6492 				ts_bw = (uint64_t)delivered;
6493 				ts_bw *= (uint64_t)USECS_IN_SECOND;
6494 				ts_bw /= ts_diff;
6495 				bbr_log_type_bbrupd(bbr, 62, cts,
6496 						    (ts_bw >> 32),
6497 						    (ts_bw & 0xffffffff), 0, 0,
6498 						    0, 0, ts_diff, delivered);
6499 				if ((bbr->ts_can_raise) &&
6500 				    (ts_bw > bw)) {
6501 					bbr_log_type_bbrupd(bbr, 8, cts,
6502 							    delivered,
6503 							    ts_diff,
6504 							    (bw >> 32),
6505 							    (bw & 0x00000000ffffffff),
6506 							    0, 0, 0, 0);
6507 					bw = ts_bw;
6508 				} else if (ts_bw && (ts_bw < bw)) {
6509 					bbr_log_type_bbrupd(bbr, 7, cts,
6510 							    delivered,
6511 							    ts_diff,
6512 							    (bw >> 32),
6513 							    (bw & 0x00000000ffffffff),
6514 							    0, 0, 0, 0);
6515 					bw = ts_bw;
6516 				}
6517 			}
6518 		}
6519 		if (rsm->r_first_sent_time &&
6520 		    TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6521 			uint64_t sbw, sti;
6522 			/*
6523 			 * We use what was in flight at the time of our
6524 			 * send  and the size of this send to figure
6525 			 * out what we have been sending at (amount).
6526 			 * For the time we take from the time of
6527 			 * the send of the first send outstanding
6528 			 * until this send plus this sends pacing
6529 			 * time. This gives us a good calculation
6530 			 * as to the rate we have been sending at.
6531 			 */
6532 
6533 			sbw = (uint64_t)(rsm->r_flight_at_send);
6534 			sbw *= (uint64_t)USECS_IN_SECOND;
6535 			sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6536 			sti += rsm->r_pacing_delay;
6537 			sbw /= sti;
6538 			if (sbw < bw) {
6539 				bbr_log_type_bbrupd(bbr, 6, cts,
6540 						    delivered,
6541 						    (uint32_t)sti,
6542 						    (bw >> 32),
6543 						    (uint32_t)bw,
6544 						    rsm->r_first_sent_time, 0, (sbw >> 32),
6545 						    (uint32_t)sbw);
6546 				bw = sbw;
6547 			}
6548 		}
6549 		/* Use the google algorithm for b/w measurements */
6550 		bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6551 		if ((rsm->r_app_limited == 0) ||
6552 		    (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) {
6553 			tcp_bbr_commit_bw(bbr, cts);
6554 			bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6555 					    0, 0, 0, 0,  bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6556 		}
6557 	}
6558 }
6559 
6560 static void
bbr_google_measurement(struct tcp_bbr * bbr,struct bbr_sendmap * rsm,uint32_t rtt,uint32_t cts)6561 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6562 {
6563 	if (bbr->rc_in_persist == 0) {
6564 		/* We log only when not in persist */
6565 		/* Translate to a Bytes Per Second */
6566 		uint64_t tim, bw;
6567 		uint32_t delivered;
6568 		int no_apply = 0;
6569 
6570 		if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6571 			tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6572 		else
6573 			tim = 1;
6574 		/*
6575 		 * Now that we have processed the tim (skipping the sample
6576 		 * or possibly updating the time, go ahead and
6577 		 * calculate the cdr.
6578 		 */
6579 		delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6580 		bw = (uint64_t)delivered;
6581 		bw *= (uint64_t)USECS_IN_SECOND;
6582 		bw /= tim;
6583 		if (tim < bbr->r_ctl.rc_lowest_rtt) {
6584 			bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6585 					    tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6586 
6587 			no_apply = 1;
6588 		}
6589 		/*
6590 		 * If we are using this b/w shove it in now so we
6591 		 * can see in the trace viewer if it gets over-ridden.
6592 		 */
6593 		bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6594 		/* Gate by the sending rate */
6595 		if (rsm->r_first_sent_time &&
6596 		    TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6597 			uint64_t sbw, sti;
6598 			/*
6599 			 * We use what was in flight at the time of our
6600 			 * send  and the size of this send to figure
6601 			 * out what we have been sending at (amount).
6602 			 * For the time we take from the time of
6603 			 * the send of the first send outstanding
6604 			 * until this send plus this sends pacing
6605 			 * time. This gives us a good calculation
6606 			 * as to the rate we have been sending at.
6607 			 */
6608 
6609 			sbw = (uint64_t)(rsm->r_flight_at_send);
6610 			sbw *= (uint64_t)USECS_IN_SECOND;
6611 			sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6612 			sti += rsm->r_pacing_delay;
6613 			sbw /= sti;
6614 			if (sbw < bw) {
6615 				bbr_log_type_bbrupd(bbr, 6, cts,
6616 						    delivered,
6617 						    (uint32_t)sti,
6618 						    (bw >> 32),
6619 						    (uint32_t)bw,
6620 						    rsm->r_first_sent_time, 0, (sbw >> 32),
6621 						    (uint32_t)sbw);
6622 				bw = sbw;
6623 			}
6624 			if ((sti > tim) &&
6625 			    (sti < bbr->r_ctl.rc_lowest_rtt)) {
6626 				bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6627 						    (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6628 				no_apply = 1;
6629 			} else
6630 				no_apply = 0;
6631 		}
6632 		bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6633 		if ((no_apply == 0) &&
6634 		    ((rsm->r_app_limited == 0) ||
6635 		     (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) {
6636 			tcp_bbr_commit_bw(bbr, cts);
6637 			bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6638 					    0, 0, 0, 0, bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6639 		}
6640 	}
6641 }
6642 
6643 static void
bbr_update_bbr_info(struct tcp_bbr * bbr,struct bbr_sendmap * rsm,uint32_t rtt,uint32_t cts,uint32_t tsin,uint32_t uts,int32_t match,uint32_t rsm_send_time,int32_t ack_type,struct tcpopt * to)6644 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin,
6645     uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to)
6646 {
6647 	uint64_t old_rttprop;
6648 
6649 	/* Update our delivery time and amount */
6650 	bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start);
6651 	bbr->r_ctl.rc_del_time = cts;
6652 	if (rtt == 0) {
6653 		/*
6654 		 * 0 means its a retransmit, for now we don't use these for
6655 		 * the rest of BBR.
6656 		 */
6657 		return;
6658 	}
6659 	if ((bbr->rc_use_google == 0) &&
6660 	    (match != BBR_RTT_BY_EXACTMATCH) &&
6661 	    (match != BBR_RTT_BY_TIMESTAMP)){
6662 		/*
6663 		 * We get a lot of rtt updates, lets not pay attention to
6664 		 * any that are not an exact match. That way we don't have
6665 		 * to worry about timestamps and the whole nonsense of
6666 		 * unsure if its a retransmission etc (if we ever had the
6667 		 * timestamp fixed to always have the last thing sent this
6668 		 * would not be a issue).
6669 		 */
6670 		return;
6671 	}
6672 	if ((bbr_no_retran && bbr->rc_use_google) &&
6673 	    (match != BBR_RTT_BY_EXACTMATCH) &&
6674 	    (match != BBR_RTT_BY_TIMESTAMP)){
6675 		/*
6676 		 * We only do measurements in google mode
6677 		 * with bbr_no_retran on for sure things.
6678 		 */
6679 		return;
6680 	}
6681 	/* Only update srtt if we know by exact match */
6682 	tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin);
6683 	if (ack_type == BBR_CUM_ACKED)
6684 		bbr->rc_ack_is_cumack = 1;
6685 	else
6686 		bbr->rc_ack_is_cumack = 0;
6687 	old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP);
6688 	/*
6689 	 * Note the following code differs to the original
6690 	 * BBR spec. It calls for <= not <. However after a
6691 	 * long discussion in email with Neal, he acknowledged
6692 	 * that it should be < than so that we will have flows
6693 	 * going into probe-rtt (we were seeing cases where that
6694 	 * did not happen and caused ugly things to occur). We
6695 	 * have added this agreed upon fix to our code base.
6696 	 */
6697 	if (rtt < old_rttprop) {
6698 		/* Update when we last saw a rtt drop */
6699 		bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0);
6700 		bbr_set_reduced_rtt(bbr, cts, __LINE__);
6701 	}
6702 	bbr_log_type_bbrrttprop(bbr, rtt, rsm->r_end, uts, cts,
6703 	    match, rsm->r_start, rsm->r_flags);
6704 	apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6705 	if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) {
6706 		/*
6707 		 * The RTT-prop moved, reset the target (may be a
6708 		 * nop for some states).
6709 		 */
6710 		bbr_set_state_target(bbr, __LINE__);
6711 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)
6712 			bbr_log_rtt_shrinks(bbr, cts, 0, 0,
6713 					    __LINE__, BBR_RTTS_NEW_TARGET, 0);
6714 		else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP))
6715 			/* It went up */
6716 			bbr_check_probe_rtt_limits(bbr, cts);
6717 	}
6718 	if ((bbr->rc_use_google == 0) &&
6719 	    (match == BBR_RTT_BY_TIMESTAMP)) {
6720 		/*
6721 		 * We don't do b/w update with
6722 		 * these since they are not really
6723 		 * reliable.
6724 		 */
6725 		return;
6726 	}
6727 	if (bbr->r_ctl.r_app_limited_until &&
6728 	    (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) {
6729 		/* We are no longer app-limited */
6730 		bbr->r_ctl.r_app_limited_until = 0;
6731 	}
6732 	if (bbr->rc_use_google) {
6733 		bbr_google_measurement(bbr, rsm, rtt, cts);
6734 	} else {
6735 		bbr_nf_measurement(bbr, rsm, rtt, cts);
6736 	}
6737 }
6738 
6739 /*
6740  * Convert a timestamp that the main stack
6741  * uses (milliseconds) into one that bbr uses
6742  * (microseconds). Return that converted timestamp.
6743  */
6744 static uint32_t
bbr_ts_convert(uint32_t cts)6745 bbr_ts_convert(uint32_t cts) {
6746 	uint32_t sec, msec;
6747 
6748 	sec = cts / MS_IN_USEC;
6749 	msec = cts - (MS_IN_USEC * sec);
6750 	return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC));
6751 }
6752 
6753 /*
6754  * Return 0 if we did not update the RTT time, return
6755  * 1 if we did.
6756  */
6757 static int
bbr_update_rtt(struct tcpcb * tp,struct tcp_bbr * bbr,struct bbr_sendmap * rsm,struct tcpopt * to,uint32_t cts,int32_t ack_type,uint32_t th_ack)6758 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr,
6759     struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack)
6760 {
6761 	int32_t i;
6762 	uint32_t t, uts = 0;
6763 
6764 	if ((rsm->r_flags & BBR_ACKED) ||
6765 	    (rsm->r_flags & BBR_WAS_RENEGED) ||
6766 	    (rsm->r_flags & BBR_RXT_CLEARED)) {
6767 		/* Already done */
6768 		return (0);
6769 	}
6770 	if (rsm->r_rtt_not_allowed) {
6771 		/* Not allowed */
6772 		return (0);
6773 	}
6774 	if (rsm->r_rtr_cnt == 1) {
6775 		/*
6776 		 * Only one transmit. Hopefully the normal case.
6777 		 */
6778 		if (TSTMP_GT(cts, rsm->r_tim_lastsent[0]))
6779 			t = cts - rsm->r_tim_lastsent[0];
6780 		else
6781 			t = 1;
6782 		bbr->r_ctl.rc_last_rtt = t;
6783 		bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6784 				    BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to);
6785 		return (1);
6786 	}
6787 	/* Convert to usecs */
6788 	if ((bbr_can_use_ts_for_rtt == 1) &&
6789 	    (bbr->rc_use_google == 1) &&
6790 	    (ack_type == BBR_CUM_ACKED) &&
6791 	    (to->to_flags & TOF_TS) &&
6792 	    (to->to_tsecr != 0)) {
6793 		t = tcp_tv_to_msec(&bbr->rc_tv) - to->to_tsecr;
6794 		if (t < 1)
6795 			t = 1;
6796 		t *= MS_IN_USEC;
6797 		bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6798 				    BBR_RTT_BY_TIMESTAMP,
6799 				    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)],
6800 				    ack_type, to);
6801 		return (1);
6802 	}
6803 	uts = bbr_ts_convert(to->to_tsecr);
6804 	if ((to->to_flags & TOF_TS) &&
6805 	    (to->to_tsecr != 0) &&
6806 	    (ack_type == BBR_CUM_ACKED) &&
6807 	    ((rsm->r_flags & BBR_OVERMAX) == 0)) {
6808 		/*
6809 		 * Now which timestamp does it match? In this block the ACK
6810 		 * may be coming from a previous transmission.
6811 		 */
6812 		uint32_t fudge;
6813 
6814 		fudge = BBR_TIMER_FUDGE;
6815 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
6816 			if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) &&
6817 			    (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) {
6818 				if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6819 					t = cts - rsm->r_tim_lastsent[i];
6820 				else
6821 					t = 1;
6822 				bbr->r_ctl.rc_last_rtt = t;
6823 				bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING,
6824 						    rsm->r_tim_lastsent[i], ack_type, to);
6825 				if ((i + 1) < rsm->r_rtr_cnt) {
6826 					/* Likely */
6827 					return (0);
6828 				} else if (rsm->r_flags & BBR_TLP) {
6829 					bbr->rc_tlp_rtx_out = 0;
6830 				}
6831 				return (1);
6832 			}
6833 		}
6834 		/* Fall through if we can't find a matching timestamp */
6835 	}
6836 	/*
6837 	 * Ok its a SACK block that we retransmitted. or a windows
6838 	 * machine without timestamps. We can tell nothing from the
6839 	 * time-stamp since its not there or the time the peer last
6840 	 * received a segment that moved forward its cum-ack point.
6841 	 *
6842 	 * Lets look at the last retransmit and see what we can tell
6843 	 * (with BBR for space we only keep 2 note we have to keep
6844 	 * at least 2 so the map can not be condensed more).
6845 	 */
6846 	i = rsm->r_rtr_cnt - 1;
6847 	if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6848 		t = cts - rsm->r_tim_lastsent[i];
6849 	else
6850 		goto not_sure;
6851 	if (t < bbr->r_ctl.rc_lowest_rtt) {
6852 		/*
6853 		 * We retransmitted and the ack came back in less
6854 		 * than the smallest rtt we have observed in the
6855 		 * windowed rtt. We most likey did an improper
6856 		 * retransmit as outlined in 4.2 Step 3 point 2 in
6857 		 * the rack-draft.
6858 		 *
6859 		 * Use the prior transmission to update all the
6860 		 * information as long as there is only one prior
6861 		 * transmission.
6862 		 */
6863 		if ((rsm->r_flags & BBR_OVERMAX) == 0) {
6864 #ifdef BBR_INVARIANTS
6865 			if (rsm->r_rtr_cnt == 1)
6866 				panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags);
6867 #endif
6868 			i = rsm->r_rtr_cnt - 2;
6869 			if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6870 				t = cts - rsm->r_tim_lastsent[i];
6871 			else
6872 				t = 1;
6873 			bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET,
6874 					    rsm->r_tim_lastsent[i], ack_type, to);
6875 			return (0);
6876 		} else {
6877 			/*
6878 			 * Too many prior transmissions, just
6879 			 * updated BBR delivered
6880 			 */
6881 not_sure:
6882 			bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6883 					    BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
6884 		}
6885 	} else {
6886 		/*
6887 		 * We retransmitted it and the retransmit did the
6888 		 * job.
6889 		 */
6890 		if (rsm->r_flags & BBR_TLP)
6891 			bbr->rc_tlp_rtx_out = 0;
6892 		if ((rsm->r_flags & BBR_OVERMAX) == 0)
6893 			bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts,
6894 					    BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to);
6895 		else
6896 			bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6897 					    BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
6898 		return (1);
6899 	}
6900 	return (0);
6901 }
6902 
6903 /*
6904  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
6905  */
6906 static void
bbr_log_sack_passed(struct tcpcb * tp,struct tcp_bbr * bbr,struct bbr_sendmap * rsm)6907 bbr_log_sack_passed(struct tcpcb *tp,
6908     struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
6909 {
6910 	struct bbr_sendmap *nrsm;
6911 
6912 	nrsm = rsm;
6913 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap,
6914 	    bbr_head, r_tnext) {
6915 		if (nrsm == rsm) {
6916 			/* Skip original segment he is acked */
6917 			continue;
6918 		}
6919 		if (nrsm->r_flags & BBR_ACKED) {
6920 			/* Skip ack'd segments */
6921 			continue;
6922 		}
6923 		if (nrsm->r_flags & BBR_SACK_PASSED) {
6924 			/*
6925 			 * We found one that is already marked
6926 			 * passed, we have been here before and
6927 			 * so all others below this are marked.
6928 			 */
6929 			break;
6930 		}
6931 		BBR_STAT_INC(bbr_sack_passed);
6932 		nrsm->r_flags |= BBR_SACK_PASSED;
6933 		if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) &&
6934 		    bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) {
6935 			bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start;
6936 			bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start;
6937 			nrsm->r_flags |= BBR_MARKED_LOST;
6938 		}
6939 		nrsm->r_flags &= ~BBR_WAS_SACKPASS;
6940 	}
6941 }
6942 
6943 /*
6944  * Returns the number of bytes that were
6945  * newly ack'd by sack blocks.
6946  */
6947 static uint32_t
bbr_proc_sack_blk(struct tcpcb * tp,struct tcp_bbr * bbr,struct sackblk * sack,struct tcpopt * to,struct bbr_sendmap ** prsm,uint32_t cts)6948 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack,
6949     struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts)
6950 {
6951 	int32_t times = 0;
6952 	uint32_t start, end, changed = 0;
6953 	struct bbr_sendmap *rsm, *nrsm;
6954 	int32_t used_ref = 1;
6955 	uint8_t went_back = 0, went_fwd = 0;
6956 
6957 	start = sack->start;
6958 	end = sack->end;
6959 	rsm = *prsm;
6960 	if (rsm == NULL)
6961 		used_ref = 0;
6962 
6963 	/* Do we locate the block behind where we last were? */
6964 	if (rsm && SEQ_LT(start, rsm->r_start)) {
6965 		went_back = 1;
6966 		TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
6967 			if (SEQ_GEQ(start, rsm->r_start) &&
6968 			    SEQ_LT(start, rsm->r_end)) {
6969 				goto do_rest_ofb;
6970 			}
6971 		}
6972 	}
6973 start_at_beginning:
6974 	went_fwd = 1;
6975 	/*
6976 	 * Ok lets locate the block where this guy is fwd from rsm (if its
6977 	 * set)
6978 	 */
6979 	TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) {
6980 		if (SEQ_GEQ(start, rsm->r_start) &&
6981 		    SEQ_LT(start, rsm->r_end)) {
6982 			break;
6983 		}
6984 	}
6985 do_rest_ofb:
6986 	if (rsm == NULL) {
6987 		/*
6988 		 * This happens when we get duplicate sack blocks with the
6989 		 * same end. For example SACK 4: 100 SACK 3: 100 The sort
6990 		 * will not change there location so we would just start at
6991 		 * the end of the first one and get lost.
6992 		 */
6993 		if (tp->t_flags & TF_SENTFIN) {
6994 			/*
6995 			 * Check to see if we have not logged the FIN that
6996 			 * went out.
6997 			 */
6998 			nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
6999 			if (nrsm && (nrsm->r_end + 1) == tp->snd_max) {
7000 				/*
7001 				 * Ok we did not get the FIN logged.
7002 				 */
7003 				nrsm->r_end++;
7004 				rsm = nrsm;
7005 				goto do_rest_ofb;
7006 			}
7007 		}
7008 		if (times == 1) {
7009 #ifdef BBR_INVARIANTS
7010 			panic("tp:%p bbr:%p sack:%p to:%p prsm:%p",
7011 			    tp, bbr, sack, to, prsm);
7012 #else
7013 			goto out;
7014 #endif
7015 		}
7016 		times++;
7017 		BBR_STAT_INC(bbr_sack_proc_restart);
7018 		rsm = NULL;
7019 		goto start_at_beginning;
7020 	}
7021 	/* Ok we have an ACK for some piece of rsm */
7022 	if (rsm->r_start != start) {
7023 		/*
7024 		 * Need to split this in two pieces the before and after.
7025 		 */
7026 		if (bbr_sack_mergable(rsm, start, end))
7027 			nrsm = bbr_alloc_full_limit(bbr);
7028 		else
7029 			nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7030 		if (nrsm == NULL) {
7031 			/* We could not allocate ignore the sack */
7032 			struct sackblk blk;
7033 
7034 			blk.start = start;
7035 			blk.end = end;
7036 			sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7037 			goto out;
7038 		}
7039 		bbr_clone_rsm(bbr, nrsm, rsm, start);
7040 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7041 		if (rsm->r_in_tmap) {
7042 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7043 			nrsm->r_in_tmap = 1;
7044 		}
7045 		rsm->r_flags &= (~BBR_HAS_FIN);
7046 		rsm = nrsm;
7047 	}
7048 	if (SEQ_GEQ(end, rsm->r_end)) {
7049 		/*
7050 		 * The end of this block is either beyond this guy or right
7051 		 * at this guy.
7052 		 */
7053 		if ((rsm->r_flags & BBR_ACKED) == 0) {
7054 			bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7055 			changed += (rsm->r_end - rsm->r_start);
7056 			bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7057 			bbr_log_sack_passed(tp, bbr, rsm);
7058 			if (rsm->r_flags & BBR_MARKED_LOST) {
7059 				bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7060 			}
7061 			/* Is Reordering occuring? */
7062 			if (rsm->r_flags & BBR_SACK_PASSED) {
7063 				BBR_STAT_INC(bbr_reorder_seen);
7064 				bbr->r_ctl.rc_reorder_ts = cts;
7065 				if (rsm->r_flags & BBR_MARKED_LOST) {
7066 					bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7067 					if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7068 						/* LT sampling also needs adjustment */
7069 						bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7070 				}
7071 			}
7072 			rsm->r_flags |= BBR_ACKED;
7073 			rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7074 			if (rsm->r_in_tmap) {
7075 				TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7076 				rsm->r_in_tmap = 0;
7077 			}
7078 		}
7079 		bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7080 		if (end == rsm->r_end) {
7081 			/* This block only - done */
7082 			goto out;
7083 		}
7084 		/* There is more not coverend by this rsm move on */
7085 		start = rsm->r_end;
7086 		nrsm = TAILQ_NEXT(rsm, r_next);
7087 		rsm = nrsm;
7088 		times = 0;
7089 		goto do_rest_ofb;
7090 	}
7091 	if (rsm->r_flags & BBR_ACKED) {
7092 		/* Been here done that */
7093 		goto out;
7094 	}
7095 	/* Ok we need to split off this one at the tail */
7096 	if (bbr_sack_mergable(rsm, start, end))
7097 		nrsm = bbr_alloc_full_limit(bbr);
7098 	else
7099 		nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7100 	if (nrsm == NULL) {
7101 		/* failed XXXrrs what can we do but loose the sack info? */
7102 		struct sackblk blk;
7103 
7104 		blk.start = start;
7105 		blk.end = end;
7106 		sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7107 		goto out;
7108 	}
7109 	/* Clone it */
7110 	bbr_clone_rsm(bbr, nrsm, rsm, end);
7111 	/* The sack block does not cover this guy fully */
7112 	rsm->r_flags &= (~BBR_HAS_FIN);
7113 	TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7114 	if (rsm->r_in_tmap) {
7115 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7116 		nrsm->r_in_tmap = 1;
7117 	}
7118 	nrsm->r_dupack = 0;
7119 	bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7120 	bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7121 	changed += (rsm->r_end - rsm->r_start);
7122 	bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7123 	bbr_log_sack_passed(tp, bbr, rsm);
7124 	/* Is Reordering occuring? */
7125 	if (rsm->r_flags & BBR_MARKED_LOST) {
7126 		bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7127 	}
7128 	if (rsm->r_flags & BBR_SACK_PASSED) {
7129 		BBR_STAT_INC(bbr_reorder_seen);
7130 		bbr->r_ctl.rc_reorder_ts = cts;
7131 		if (rsm->r_flags & BBR_MARKED_LOST) {
7132 			bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7133 			if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7134 				/* LT sampling also needs adjustment */
7135 				bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7136 		}
7137 	}
7138 	rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7139 	rsm->r_flags |= BBR_ACKED;
7140 	if (rsm->r_in_tmap) {
7141 		TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7142 		rsm->r_in_tmap = 0;
7143 	}
7144 out:
7145 	if (rsm && (rsm->r_flags & BBR_ACKED)) {
7146 		/*
7147 		 * Now can we merge this newly acked
7148 		 * block with either the previous or
7149 		 * next block?
7150 		 */
7151 		nrsm = TAILQ_NEXT(rsm, r_next);
7152 		if (nrsm &&
7153 		    (nrsm->r_flags & BBR_ACKED)) {
7154 			/* yep this and next can be merged */
7155 			rsm = bbr_merge_rsm(bbr, rsm, nrsm);
7156 		}
7157 		/* Now what about the previous? */
7158 		nrsm = TAILQ_PREV(rsm, bbr_head, r_next);
7159 		if (nrsm &&
7160 		    (nrsm->r_flags & BBR_ACKED)) {
7161 			/* yep the previous and this can be merged */
7162 			rsm = bbr_merge_rsm(bbr, nrsm, rsm);
7163 		}
7164 	}
7165 	if (used_ref == 0) {
7166 		BBR_STAT_INC(bbr_sack_proc_all);
7167 	} else {
7168 		BBR_STAT_INC(bbr_sack_proc_short);
7169 	}
7170 	if (went_fwd && went_back) {
7171 		BBR_STAT_INC(bbr_sack_search_both);
7172 	} else if (went_fwd) {
7173 		BBR_STAT_INC(bbr_sack_search_fwd);
7174 	} else if (went_back) {
7175 		BBR_STAT_INC(bbr_sack_search_back);
7176 	}
7177 	/* Save off where the next seq is */
7178 	if (rsm)
7179 		bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next);
7180 	else
7181 		bbr->r_ctl.rc_sacklast = NULL;
7182 	*prsm = rsm;
7183 	return (changed);
7184 }
7185 
7186 static void inline
bbr_peer_reneges(struct tcp_bbr * bbr,struct bbr_sendmap * rsm,tcp_seq th_ack)7187 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack)
7188 {
7189 	struct bbr_sendmap *tmap;
7190 
7191 	BBR_STAT_INC(bbr_reneges_seen);
7192 	tmap = NULL;
7193 	while (rsm && (rsm->r_flags & BBR_ACKED)) {
7194 		/* Its no longer sacked, mark it so */
7195 		uint32_t oflags;
7196 		bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7197 #ifdef BBR_INVARIANTS
7198 		if (rsm->r_in_tmap) {
7199 			panic("bbr:%p rsm:%p flags:0x%x in tmap?",
7200 			    bbr, rsm, rsm->r_flags);
7201 		}
7202 #endif
7203 		oflags = rsm->r_flags;
7204 		if (rsm->r_flags & BBR_MARKED_LOST) {
7205 			bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7206 			bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7207 			if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7208 				/* LT sampling also needs adjustment */
7209 				bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7210 		}
7211 		rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST);
7212 		rsm->r_flags |= BBR_WAS_RENEGED;
7213 		rsm->r_flags |= BBR_RXT_CLEARED;
7214 		bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__);
7215 		/* Rebuild it into our tmap */
7216 		if (tmap == NULL) {
7217 			TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7218 			tmap = rsm;
7219 		} else {
7220 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7221 			tmap = rsm;
7222 		}
7223 		tmap->r_in_tmap = 1;
7224 		/*
7225 		 * XXXrrs Delivered? Should we do anything here?
7226 		 *
7227 		 * Of course we don't on a rxt timeout so maybe its ok that
7228 		 * we don't?
7229 		 *
7230 		 * For now lets not.
7231 		 */
7232 		rsm = TAILQ_NEXT(rsm, r_next);
7233 	}
7234 	/*
7235 	 * Now lets possibly clear the sack filter so we start recognizing
7236 	 * sacks that cover this area.
7237 	 */
7238 	sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack);
7239 }
7240 
7241 static void
bbr_log_syn(struct tcpcb * tp,struct tcpopt * to)7242 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to)
7243 {
7244 	struct tcp_bbr *bbr;
7245 	struct bbr_sendmap *rsm;
7246 	uint32_t cts;
7247 
7248 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7249 	cts = bbr->r_ctl.rc_rcvtime;
7250 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7251 	if (rsm && (rsm->r_flags & BBR_HAS_SYN)) {
7252 		if ((rsm->r_end - rsm->r_start) <= 1) {
7253 			/* Log out the SYN completely */
7254 			bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7255 			rsm->r_rtr_bytes = 0;
7256 			TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7257 			if (rsm->r_in_tmap) {
7258 				TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7259 				rsm->r_in_tmap = 0;
7260 			}
7261 			if (bbr->r_ctl.rc_next == rsm) {
7262 				/* scoot along the marker */
7263 				bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7264 			}
7265 			if (to != NULL)
7266 				bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0);
7267 			bbr_free(bbr, rsm);
7268 		} else {
7269 			/* There is more (Fast open)? strip out SYN. */
7270 			rsm->r_flags &= ~BBR_HAS_SYN;
7271 			rsm->r_start++;
7272 		}
7273 	}
7274 }
7275 
7276 /*
7277  * Returns the number of bytes that were
7278  * acknowledged by SACK blocks.
7279  */
7280 
7281 static uint32_t
bbr_log_ack(struct tcpcb * tp,struct tcpopt * to,struct tcphdr * th,uint32_t * prev_acked)7282 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th,
7283     uint32_t *prev_acked)
7284 {
7285 	uint32_t changed, last_seq, entered_recovery = 0;
7286 	struct tcp_bbr *bbr;
7287 	struct bbr_sendmap *rsm;
7288 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7289 	register uint32_t th_ack;
7290 	int32_t i, j, k, new_sb, num_sack_blks = 0;
7291 	uint32_t cts, acked, ack_point, sack_changed = 0;
7292 	uint32_t p_maxseg, maxseg, p_acked = 0;
7293 
7294 	INP_WLOCK_ASSERT(tptoinpcb(tp));
7295 	if (tcp_get_flags(th) & TH_RST) {
7296 		/* We don't log resets */
7297 		return (0);
7298 	}
7299 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7300 	cts = bbr->r_ctl.rc_rcvtime;
7301 
7302 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7303 	changed = 0;
7304 	maxseg = tp->t_maxseg - bbr->rc_last_options;
7305 	p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg);
7306 	th_ack = th->th_ack;
7307 	if (SEQ_GT(th_ack, tp->snd_una)) {
7308 		bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__);
7309 		bbr->rc_tp->t_acktime = ticks;
7310 	}
7311 	if (SEQ_LEQ(th_ack, tp->snd_una)) {
7312 		/* Only sent here for sack processing */
7313 		goto proc_sack;
7314 	}
7315 	if (rsm && SEQ_GT(th_ack, rsm->r_start)) {
7316 		changed = th_ack - rsm->r_start;
7317 	} else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) {
7318 		/*
7319 		 * For the SYN incoming case we will not have called
7320 		 * tcp_output for the sending of the SYN, so there will be
7321 		 * no map. All other cases should probably be a panic.
7322 		 */
7323 		if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) {
7324 			/*
7325 			 * We have a timestamp that can be used to generate
7326 			 * an initial RTT.
7327 			 */
7328 			uint32_t ts, now, rtt;
7329 
7330 			ts = bbr_ts_convert(to->to_tsecr);
7331 			now = bbr_ts_convert(tcp_tv_to_msec(&bbr->rc_tv));
7332 			rtt = now - ts;
7333 			if (rtt < 1)
7334 				rtt = 1;
7335 			bbr_log_type_bbrrttprop(bbr, rtt,
7336 						tp->iss, 0, cts,
7337 						BBR_RTT_BY_TIMESTAMP, tp->iss, 0);
7338 			apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
7339 			changed = 1;
7340 			bbr->r_wanted_output = 1;
7341 			goto out;
7342 		}
7343 		goto proc_sack;
7344 	} else if (rsm == NULL) {
7345 		goto out;
7346 	}
7347 	if (changed) {
7348 		/*
7349 		 * The ACK point is advancing to th_ack, we must drop off
7350 		 * the packets in the rack log and calculate any eligble
7351 		 * RTT's.
7352 		 */
7353 		bbr->r_wanted_output = 1;
7354 more:
7355 		if (rsm == NULL) {
7356 			if (tp->t_flags & TF_SENTFIN) {
7357 				/* if we send a FIN we will not hav a map */
7358 				goto proc_sack;
7359 			}
7360 #ifdef BBR_INVARIANTS
7361 			panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n",
7362 			    tp,
7363 			    th, tp->t_state, bbr,
7364 			    tp->snd_una, tp->snd_max, changed);
7365 #endif
7366 			goto proc_sack;
7367 		}
7368 	}
7369 	if (SEQ_LT(th_ack, rsm->r_start)) {
7370 		/* Huh map is missing this */
7371 #ifdef BBR_INVARIANTS
7372 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n",
7373 		    rsm->r_start,
7374 		    th_ack, tp->t_state,
7375 		    bbr->r_state, bbr);
7376 		panic("th-ack is bad bbr:%p tp:%p", bbr, tp);
7377 #endif
7378 		goto proc_sack;
7379 	} else if (th_ack == rsm->r_start) {
7380 		/* None here to ack */
7381 		goto proc_sack;
7382 	}
7383 	/*
7384 	 * Clear the dup ack counter, it will
7385 	 * either be freed or if there is some
7386 	 * remaining we need to start it at zero.
7387 	 */
7388 	rsm->r_dupack = 0;
7389 	/* Now do we consume the whole thing? */
7390 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
7391 		/* Its all consumed. */
7392 		uint32_t left;
7393 
7394 		if (rsm->r_flags & BBR_ACKED) {
7395 			/*
7396 			 * It was acked on the scoreboard -- remove it from
7397 			 * total
7398 			 */
7399 			p_acked += (rsm->r_end - rsm->r_start);
7400 			bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7401 			if (bbr->r_ctl.rc_sacked == 0)
7402 				bbr->r_ctl.rc_sacklast = NULL;
7403 		} else {
7404 			bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack);
7405 			if (rsm->r_flags & BBR_MARKED_LOST) {
7406 				bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7407 			}
7408 			if (rsm->r_flags & BBR_SACK_PASSED) {
7409 				/*
7410 				 * There are acked segments ACKED on the
7411 				 * scoreboard further up. We are seeing
7412 				 * reordering.
7413 				 */
7414 				BBR_STAT_INC(bbr_reorder_seen);
7415 				bbr->r_ctl.rc_reorder_ts = cts;
7416 				if (rsm->r_flags & BBR_MARKED_LOST) {
7417 					bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7418 					if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7419 						/* LT sampling also needs adjustment */
7420 						bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7421 				}
7422 			}
7423 			rsm->r_flags &= ~BBR_MARKED_LOST;
7424 		}
7425 		bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7426 		rsm->r_rtr_bytes = 0;
7427 		TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7428 		if (rsm->r_in_tmap) {
7429 			TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7430 			rsm->r_in_tmap = 0;
7431 		}
7432 		if (bbr->r_ctl.rc_next == rsm) {
7433 			/* scoot along the marker */
7434 			bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7435 		}
7436 		bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7437 		/* Adjust the packet counts */
7438 		left = th_ack - rsm->r_end;
7439 		/* Free back to zone */
7440 		bbr_free(bbr, rsm);
7441 		if (left) {
7442 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7443 			goto more;
7444 		}
7445 		goto proc_sack;
7446 	}
7447 	if (rsm->r_flags & BBR_ACKED) {
7448 		/*
7449 		 * It was acked on the scoreboard -- remove it from total
7450 		 * for the part being cum-acked.
7451 		 */
7452 		p_acked += (rsm->r_end - rsm->r_start);
7453 		bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7454 		if (bbr->r_ctl.rc_sacked == 0)
7455 			bbr->r_ctl.rc_sacklast = NULL;
7456 	} else {
7457 		/*
7458 		 * It was acked up to th_ack point for the first time
7459 		 */
7460 		struct bbr_sendmap lrsm;
7461 
7462 		memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap));
7463 		lrsm.r_end = th_ack;
7464 		bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack);
7465 	}
7466 	if ((rsm->r_flags & BBR_MARKED_LOST) &&
7467 	    ((rsm->r_flags & BBR_ACKED) == 0)) {
7468 		/*
7469 		 * It was marked lost and partly ack'd now
7470 		 * for the first time. We lower the rc_lost_bytes
7471 		 * and still leave it MARKED.
7472 		 */
7473 		bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start;
7474 	}
7475 	bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7476 	bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7477 	rsm->r_rtr_bytes = 0;
7478 	/* adjust packet count */
7479 	rsm->r_start = th_ack;
7480 proc_sack:
7481 	/* Check for reneging */
7482 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7483 	if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) {
7484 		/*
7485 		 * The peer has moved snd_una up to the edge of this send,
7486 		 * i.e. one that it had previously acked. The only way that
7487 		 * can be true if the peer threw away data (space issues)
7488 		 * that it had previously sacked (else it would have given
7489 		 * us snd_una up to (rsm->r_end). We need to undo the acked
7490 		 * markings here.
7491 		 *
7492 		 * Note we have to look to make sure th_ack is our
7493 		 * rsm->r_start in case we get an old ack where th_ack is
7494 		 * behind snd_una.
7495 		 */
7496 		bbr_peer_reneges(bbr, rsm, th->th_ack);
7497 	}
7498 	if ((to->to_flags & TOF_SACK) == 0) {
7499 		/* We are done nothing left to log */
7500 		goto out;
7501 	}
7502 	rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7503 	if (rsm) {
7504 		last_seq = rsm->r_end;
7505 	} else {
7506 		last_seq = tp->snd_max;
7507 	}
7508 	/* Sack block processing */
7509 	if (SEQ_GT(th_ack, tp->snd_una))
7510 		ack_point = th_ack;
7511 	else
7512 		ack_point = tp->snd_una;
7513 	for (i = 0; i < to->to_nsacks; i++) {
7514 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
7515 		    &sack, sizeof(sack));
7516 		sack.start = ntohl(sack.start);
7517 		sack.end = ntohl(sack.end);
7518 		if (SEQ_GT(sack.end, sack.start) &&
7519 		    SEQ_GT(sack.start, ack_point) &&
7520 		    SEQ_LT(sack.start, tp->snd_max) &&
7521 		    SEQ_GT(sack.end, ack_point) &&
7522 		    SEQ_LEQ(sack.end, tp->snd_max)) {
7523 			if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) &&
7524 			    (SEQ_LT(sack.end, last_seq)) &&
7525 			    ((sack.end - sack.start) < (p_maxseg / 8))) {
7526 				/*
7527 				 * Not the last piece and its smaller than
7528 				 * 1/8th of a p_maxseg. We ignore this.
7529 				 */
7530 				BBR_STAT_INC(bbr_runt_sacks);
7531 				continue;
7532 			}
7533 			sack_blocks[num_sack_blks] = sack;
7534 			num_sack_blks++;
7535 		} else if (SEQ_LEQ(sack.start, th_ack) &&
7536 		    SEQ_LEQ(sack.end, th_ack)) {
7537 			/*
7538 			 * Its a D-SACK block.
7539 			 */
7540 			tcp_record_dsack(tp, sack.start, sack.end, 0);
7541 		}
7542 	}
7543 	if (num_sack_blks == 0)
7544 		goto out;
7545 	/*
7546 	 * Sort the SACK blocks so we can update the rack scoreboard with
7547 	 * just one pass.
7548 	 */
7549 	new_sb = sack_filter_blks(tp, &bbr->r_ctl.bbr_sf, sack_blocks,
7550 				  num_sack_blks, th->th_ack);
7551 	ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks);
7552 	BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks);
7553 	BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb));
7554 	num_sack_blks = new_sb;
7555 	if (num_sack_blks < 2) {
7556 		goto do_sack_work;
7557 	}
7558 	/* Sort the sacks */
7559 	for (i = 0; i < num_sack_blks; i++) {
7560 		for (j = i + 1; j < num_sack_blks; j++) {
7561 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7562 				sack = sack_blocks[i];
7563 				sack_blocks[i] = sack_blocks[j];
7564 				sack_blocks[j] = sack;
7565 			}
7566 		}
7567 	}
7568 	/*
7569 	 * Now are any of the sack block ends the same (yes some
7570 	 * implememtations send these)?
7571 	 */
7572 again:
7573 	if (num_sack_blks > 1) {
7574 		for (i = 0; i < num_sack_blks; i++) {
7575 			for (j = i + 1; j < num_sack_blks; j++) {
7576 				if (sack_blocks[i].end == sack_blocks[j].end) {
7577 					/*
7578 					 * Ok these two have the same end we
7579 					 * want the smallest end and then
7580 					 * throw away the larger and start
7581 					 * again.
7582 					 */
7583 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7584 						/*
7585 						 * The second block covers
7586 						 * more area use that
7587 						 */
7588 						sack_blocks[i].start = sack_blocks[j].start;
7589 					}
7590 					/*
7591 					 * Now collapse out the dup-sack and
7592 					 * lower the count
7593 					 */
7594 					for (k = (j + 1); k < num_sack_blks; k++) {
7595 						sack_blocks[j].start = sack_blocks[k].start;
7596 						sack_blocks[j].end = sack_blocks[k].end;
7597 						j++;
7598 					}
7599 					num_sack_blks--;
7600 					goto again;
7601 				}
7602 			}
7603 		}
7604 	}
7605 do_sack_work:
7606 	rsm = bbr->r_ctl.rc_sacklast;
7607 	for (i = 0; i < num_sack_blks; i++) {
7608 		acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts);
7609 		if (acked) {
7610 			bbr->r_wanted_output = 1;
7611 			changed += acked;
7612 			sack_changed += acked;
7613 		}
7614 	}
7615 out:
7616 	*prev_acked = p_acked;
7617 	if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) {
7618 		/*
7619 		 * Ok we have a high probability that we need to go in to
7620 		 * recovery since we have data sack'd
7621 		 */
7622 		struct bbr_sendmap *rsm;
7623 
7624 		rsm = bbr_check_recovery_mode(tp, bbr, cts);
7625 		if (rsm) {
7626 			/* Enter recovery */
7627 			entered_recovery = 1;
7628 			bbr->r_wanted_output = 1;
7629 			/*
7630 			 * When we enter recovery we need to assure we send
7631 			 * one packet.
7632 			 */
7633 			if (bbr->r_ctl.rc_resend == NULL) {
7634 				bbr->r_ctl.rc_resend = rsm;
7635 			}
7636 		}
7637 	}
7638 	if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) {
7639 		/*
7640 		 * See if we need to rack-retransmit anything if so set it
7641 		 * up as the thing to resend assuming something else is not
7642 		 * already in that position.
7643 		 */
7644 		if (bbr->r_ctl.rc_resend == NULL) {
7645 			bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
7646 		}
7647 	}
7648 	/*
7649 	 * We return the amount that changed via sack, this is used by the
7650 	 * ack-received code to augment what was changed between th_ack <->
7651 	 * snd_una.
7652 	 */
7653 	return (sack_changed);
7654 }
7655 
7656 static void
bbr_strike_dupack(struct tcp_bbr * bbr)7657 bbr_strike_dupack(struct tcp_bbr *bbr)
7658 {
7659 	struct bbr_sendmap *rsm;
7660 
7661 	rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
7662 	if (rsm && (rsm->r_dupack < 0xff)) {
7663 		rsm->r_dupack++;
7664 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD)
7665 			bbr->r_wanted_output = 1;
7666 	}
7667 }
7668 
7669 /*
7670  * Return value of 1, we do not need to call bbr_process_data().
7671  * return value of 0, bbr_process_data can be called.
7672  * For ret_val if its 0 the TCB is locked and valid, if its non-zero
7673  * its unlocked and probably unsafe to touch the TCB.
7674  */
7675 static int
bbr_process_ack(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,uint32_t tiwin,int32_t tlen,int32_t * ofia,int32_t thflags,int32_t * ret_val)7676 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
7677     struct tcpcb *tp, struct tcpopt *to,
7678     uint32_t tiwin, int32_t tlen,
7679     int32_t * ofia, int32_t thflags, int32_t * ret_val)
7680 {
7681 	int32_t ourfinisacked = 0;
7682 	int32_t acked_amount;
7683 	uint16_t nsegs;
7684 	int32_t acked;
7685 	uint32_t lost, sack_changed = 0;
7686 	struct mbuf *mfree;
7687 	struct tcp_bbr *bbr;
7688 	uint32_t prev_acked = 0;
7689 
7690 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7691 	lost = bbr->r_ctl.rc_lost;
7692 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
7693 	if (SEQ_GEQ(tp->snd_una, tp->iss + (65535 << tp->snd_scale))) {
7694 		/* Checking SEG.ACK against ISS is definitely redundant. */
7695 		tp->t_flags2 |= TF2_NO_ISS_CHECK;
7696 	}
7697 	if (!V_tcp_insecure_ack) {
7698 		tcp_seq seq_min;
7699 		bool ghost_ack_check;
7700 
7701 		if (tp->t_flags2 & TF2_NO_ISS_CHECK) {
7702 			/* Check for too old ACKs (RFC 5961, Section 5.2). */
7703 			seq_min = tp->snd_una - tp->max_sndwnd;
7704 			ghost_ack_check = false;
7705 		} else {
7706 			if (SEQ_GT(tp->iss + 1, tp->snd_una - tp->max_sndwnd)) {
7707 				/* Checking for ghost ACKs is stricter. */
7708 				seq_min = tp->iss + 1;
7709 				ghost_ack_check = true;
7710 			} else {
7711 				/*
7712 				 * Checking for too old ACKs (RFC 5961,
7713 				 * Section 5.2) is stricter.
7714 				 */
7715 				seq_min = tp->snd_una - tp->max_sndwnd;
7716 				ghost_ack_check = false;
7717 			}
7718 		}
7719 		if (SEQ_LT(th->th_ack, seq_min)) {
7720 			if (ghost_ack_check)
7721 				TCPSTAT_INC(tcps_rcvghostack);
7722 			else
7723 				TCPSTAT_INC(tcps_rcvacktooold);
7724 			/* Send challenge ACK. */
7725 			ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
7726 			bbr->r_wanted_output = 1;
7727 			return (1);
7728 		}
7729 	}
7730 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
7731 		ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
7732 		bbr->r_wanted_output = 1;
7733 		return (1);
7734 	}
7735 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
7736 		/* Process the ack */
7737 		if (bbr->rc_in_persist)
7738 			tp->t_rxtshift = 0;
7739 		if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
7740 			bbr_strike_dupack(bbr);
7741 		sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
7742 	}
7743 	bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost));
7744 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
7745 		/*
7746 		 * Old ack, behind the last one rcv'd or a duplicate ack
7747 		 * with SACK info.
7748 		 */
7749 		if (th->th_ack == tp->snd_una) {
7750 			bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0);
7751 			if (bbr->r_state == TCPS_SYN_SENT) {
7752 				/*
7753 				 * Special case on where we sent SYN. When
7754 				 * the SYN-ACK is processed in syn_sent
7755 				 * state it bumps the snd_una. This causes
7756 				 * us to hit here even though we did ack 1
7757 				 * byte.
7758 				 *
7759 				 * Go through the nothing left case so we
7760 				 * send data.
7761 				 */
7762 				goto nothing_left;
7763 			}
7764 		}
7765 		return (0);
7766 	}
7767 	/*
7768 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
7769 	 * something we sent.
7770 	 */
7771 	if (tp->t_flags & TF_NEEDSYN) {
7772 		/*
7773 		 * T/TCP: Connection was half-synchronized, and our SYN has
7774 		 * been ACK'd (so connection is now fully synchronized).  Go
7775 		 * to non-starred state, increment snd_una for ACK of SYN,
7776 		 * and check if we can do window scaling.
7777 		 */
7778 		tp->t_flags &= ~TF_NEEDSYN;
7779 		tp->snd_una++;
7780 		/* Do window scaling? */
7781 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
7782 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
7783 			tp->rcv_scale = tp->request_r_scale;
7784 			/* Send window already scaled. */
7785 		}
7786 	}
7787 	INP_WLOCK_ASSERT(tptoinpcb(tp));
7788 
7789 	acked = BYTES_THIS_ACK(tp, th);
7790 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
7791 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
7792 
7793 	/*
7794 	 * If we just performed our first retransmit, and the ACK arrives
7795 	 * within our recovery window, then it was a mistake to do the
7796 	 * retransmit in the first place.  Recover our original cwnd and
7797 	 * ssthresh, and proceed to transmit where we left off.
7798 	 */
7799 	if (tp->t_flags & TF_PREVVALID) {
7800 		tp->t_flags &= ~TF_PREVVALID;
7801 		if (tp->t_rxtshift == 1 &&
7802 		    (int)(ticks - tp->t_badrxtwin) < 0)
7803 			bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
7804 	}
7805 	SOCK_SENDBUF_LOCK(so);
7806 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
7807 	tp->snd_wnd -= acked_amount;
7808 	mfree = sbcut_locked(&so->so_snd, acked_amount);
7809 	/* NB: sowwakeup_locked() does an implicit unlock. */
7810 	sowwakeup_locked(so);
7811 	m_freem(mfree);
7812 	if (SEQ_GT(th->th_ack, tp->snd_una)) {
7813 		bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
7814 	}
7815 	tp->snd_una = th->th_ack;
7816 	bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost));
7817 	if (IN_RECOVERY(tp->t_flags)) {
7818 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
7819 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
7820 			tcp_bbr_partialack(tp);
7821 		} else {
7822 			bbr_post_recovery(tp);
7823 		}
7824 	}
7825 	if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
7826 		tp->snd_recover = tp->snd_una;
7827 	}
7828 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
7829 		tp->snd_nxt = tp->snd_max;
7830 	}
7831 	if (tp->snd_una == tp->snd_max) {
7832 		/* Nothing left outstanding */
7833 nothing_left:
7834 		bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
7835 		if (sbavail(&so->so_snd) == 0)
7836 			bbr->rc_tp->t_acktime = 0;
7837 		if ((sbused(&so->so_snd) == 0) &&
7838 		    (tp->t_flags & TF_SENTFIN)) {
7839 			ourfinisacked = 1;
7840 		}
7841 		bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
7842 		if (bbr->rc_in_persist == 0) {
7843 			bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
7844 		}
7845 		sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
7846 		bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
7847 		/*
7848 		 * We invalidate the last ack here since we
7849 		 * don't want to transfer forward the time
7850 		 * for our sum's calculations.
7851 		 */
7852 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7853 		    (sbavail(&so->so_snd) == 0) &&
7854 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
7855 			/*
7856 			 * The socket was gone and the peer sent data, time
7857 			 * to reset him.
7858 			 */
7859 			*ret_val = 1;
7860 			tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
7861 			/* tcp_close will kill the inp pre-log the Reset */
7862 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
7863 			tp = tcp_close(tp);
7864 			ctf_do_dropwithreset(m, tp, th, tlen);
7865 			BBR_STAT_INC(bbr_dropped_af_data);
7866 			return (1);
7867 		}
7868 		/* Set need output so persist might get set */
7869 		bbr->r_wanted_output = 1;
7870 	}
7871 	if (ofia)
7872 		*ofia = ourfinisacked;
7873 	return (0);
7874 }
7875 
7876 static void
bbr_enter_persist(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts,int32_t line)7877 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
7878 {
7879 	if (bbr->rc_in_persist == 0) {
7880 		bbr_timer_cancel(bbr, __LINE__, cts);
7881 		bbr->r_ctl.rc_last_delay_val = 0;
7882 		tp->t_rxtshift = 0;
7883 		bbr->rc_in_persist = 1;
7884 		bbr->r_ctl.rc_went_idle_time = cts;
7885 		/* We should be capped when rw went to 0 but just in case */
7886 		bbr_log_type_pesist(bbr, cts, 0, line, 1);
7887 		/* Time freezes for the state, so do the accounting now */
7888 		if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
7889 			uint32_t time_in;
7890 
7891 			time_in = cts - bbr->r_ctl.rc_bbr_state_time;
7892 			if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
7893 				int32_t idx;
7894 
7895 				idx = bbr_state_val(bbr);
7896 				counter_u64_add(bbr_state_time[(idx + 5)], time_in);
7897 			} else {
7898 				counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
7899 			}
7900 		}
7901 		bbr->r_ctl.rc_bbr_state_time = cts;
7902 	}
7903 }
7904 
7905 static void
bbr_restart_after_idle(struct tcp_bbr * bbr,uint32_t cts,uint32_t idle_time)7906 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time)
7907 {
7908 	/*
7909 	 * Note that if idle time does not exceed our
7910 	 * threshold, we do nothing continuing the state
7911 	 * transitions we were last walking through.
7912 	 */
7913 	if (idle_time >= bbr_idle_restart_threshold) {
7914 		if (bbr->rc_use_idle_restart) {
7915 			bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT;
7916 			/*
7917 			 * Set our target using BBR_UNIT, so
7918 			 * we increase at a dramatic rate but
7919 			 * we stop when we get the pipe
7920 			 * full again for our current b/w estimate.
7921 			 */
7922 			bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
7923 			bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
7924 			bbr_set_state_target(bbr, __LINE__);
7925 			/* Now setup our gains to ramp up */
7926 			bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
7927 			bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
7928 			bbr_log_type_statechange(bbr, cts, __LINE__);
7929 		} else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
7930 			bbr_substate_change(bbr, cts, __LINE__, 1);
7931 		}
7932 	}
7933 }
7934 
7935 static void
bbr_exit_persist(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts,int32_t line)7936 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
7937 {
7938 	uint32_t idle_time;
7939 
7940 	if (bbr->rc_in_persist == 0)
7941 		return;
7942 	idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time);
7943 	bbr->rc_in_persist = 0;
7944 	bbr->rc_hit_state_1 = 0;
7945 	bbr->r_ctl.rc_del_time = cts;
7946 	/*
7947 	 * We invalidate the last ack here since we
7948 	 * don't want to transfer forward the time
7949 	 * for our sum's calculations.
7950 	 */
7951 	if (tcp_in_hpts(bbr->rc_tp)) {
7952 		tcp_hpts_remove(bbr->rc_tp);
7953 		bbr->rc_timer_first = 0;
7954 		bbr->r_ctl.rc_hpts_flags = 0;
7955 		bbr->r_ctl.rc_last_delay_val = 0;
7956 		bbr->r_ctl.rc_hptsi_agg_delay = 0;
7957 		bbr->r_agg_early_set = 0;
7958 		bbr->r_ctl.rc_agg_early = 0;
7959 	}
7960 	bbr_log_type_pesist(bbr, cts, idle_time, line, 0);
7961 	if (idle_time >= bbr_rtt_probe_time) {
7962 		/*
7963 		 * This qualifies as a RTT_PROBE session since we drop the
7964 		 * data outstanding to nothing and waited more than
7965 		 * bbr_rtt_probe_time.
7966 		 */
7967 		bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0);
7968 		bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts;
7969 	}
7970 	tp->t_rxtshift = 0;
7971 	/*
7972 	 * If in probeBW and we have persisted more than an RTT lets do
7973 	 * special handling.
7974 	 */
7975 	/* Force a time based epoch */
7976 	bbr_set_epoch(bbr, cts, __LINE__);
7977 	/*
7978 	 * Setup the lost so we don't count anything against the guy
7979 	 * we have been stuck with during persists.
7980 	 */
7981 	bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
7982 	/* Time un-freezes for the state */
7983 	bbr->r_ctl.rc_bbr_state_time = cts;
7984 	if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) ||
7985 	    (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) {
7986 		/*
7987 		 * If we are going back to probe-bw
7988 		 * or probe_rtt, we may need to possibly
7989 		 * do a fast restart.
7990 		 */
7991 		bbr_restart_after_idle(bbr, cts, idle_time);
7992 	}
7993 }
7994 
7995 static void
bbr_collapsed_window(struct tcp_bbr * bbr)7996 bbr_collapsed_window(struct tcp_bbr *bbr)
7997 {
7998 	/*
7999 	 * Now we must walk the
8000 	 * send map and divide the
8001 	 * ones left stranded. These
8002 	 * guys can't cause us to abort
8003 	 * the connection and are really
8004 	 * "unsent". However if a buggy
8005 	 * client actually did keep some
8006 	 * of the data i.e. collapsed the win
8007 	 * and refused to ack and then opened
8008 	 * the win and acked that data. We would
8009 	 * get into an ack war, the simplier
8010 	 * method then of just pretending we
8011 	 * did not send those segments something
8012 	 * won't work.
8013 	 */
8014 	struct bbr_sendmap *rsm, *nrsm;
8015 	tcp_seq max_seq;
8016 	uint32_t maxseg;
8017 	int can_split = 0;
8018 	int fnd = 0;
8019 
8020 	maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
8021 	max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd;
8022 	bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0);
8023 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
8024 		/* Find the first seq past or at maxseq */
8025 		if (rsm->r_flags & BBR_RWND_COLLAPSED)
8026 			rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8027 		if (SEQ_GEQ(max_seq, rsm->r_start) &&
8028 		    SEQ_GEQ(rsm->r_end, max_seq)) {
8029 			fnd = 1;
8030 			break;
8031 		}
8032 	}
8033 	bbr->rc_has_collapsed = 0;
8034 	if (!fnd) {
8035 		/* Nothing to do strange */
8036 		return;
8037 	}
8038 	/*
8039 	 * Now can we split?
8040 	 *
8041 	 * We don't want to split if splitting
8042 	 * would generate too many small segments
8043 	 * less we let an attacker fragment our
8044 	 * send_map and leave us out of memory.
8045 	 */
8046 	if ((max_seq != rsm->r_start) &&
8047 	    (max_seq != rsm->r_end)){
8048 		/* can we split? */
8049 		int res1, res2;
8050 
8051 		res1 = max_seq - rsm->r_start;
8052 		res2 = rsm->r_end - max_seq;
8053 		if ((res1 >= (maxseg/8)) &&
8054 		    (res2 >= (maxseg/8))) {
8055 			/* No small pieces here */
8056 			can_split = 1;
8057 		} else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) {
8058 			/* We are under the limit */
8059 			can_split = 1;
8060 		}
8061 	}
8062 	/* Ok do we need to split this rsm? */
8063 	if (max_seq == rsm->r_start) {
8064 		/* It's this guy no split required */
8065 		nrsm = rsm;
8066 	} else if (max_seq == rsm->r_end) {
8067 		/* It's the next one no split required. */
8068 		nrsm = TAILQ_NEXT(rsm, r_next);
8069 		if (nrsm == NULL) {
8070 			/* Huh? */
8071 			return;
8072 		}
8073 	} else if (can_split && SEQ_LT(max_seq, rsm->r_end)) {
8074 		/* yep we need to split it */
8075 		nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
8076 		if (nrsm == NULL) {
8077 			/* failed XXXrrs what can we do mark the whole? */
8078 			nrsm = rsm;
8079 			goto no_split;
8080 		}
8081 		/* Clone it */
8082 		bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0);
8083 		bbr_clone_rsm(bbr, nrsm, rsm, max_seq);
8084 		TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
8085 		if (rsm->r_in_tmap) {
8086 			TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8087 			nrsm->r_in_tmap = 1;
8088 		}
8089 	} else {
8090 		/*
8091 		 * Split not allowed just start here just
8092 		 * use this guy.
8093 		 */
8094 		nrsm = rsm;
8095 	}
8096 no_split:
8097 	BBR_STAT_INC(bbr_collapsed_win);
8098 	/* reuse fnd as a count */
8099 	fnd = 0;
8100 	TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) {
8101 		nrsm->r_flags |= BBR_RWND_COLLAPSED;
8102 		fnd++;
8103 		bbr->rc_has_collapsed = 1;
8104 	}
8105 	bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd);
8106 }
8107 
8108 static void
bbr_un_collapse_window(struct tcp_bbr * bbr)8109 bbr_un_collapse_window(struct tcp_bbr *bbr)
8110 {
8111 	struct bbr_sendmap *rsm;
8112 	int cleared = 0;
8113 
8114 	TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
8115 		if (rsm->r_flags & BBR_RWND_COLLAPSED) {
8116 			/* Clear the flag */
8117 			rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8118 			cleared++;
8119 		} else
8120 			break;
8121 	}
8122 	bbr_log_type_rwnd_collapse(bbr,
8123 				   (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared);
8124 	bbr->rc_has_collapsed = 0;
8125 }
8126 
8127 /*
8128  * Return value of 1, the TCB is unlocked and most
8129  * likely gone, return value of 0, the TCB is still
8130  * locked.
8131  */
8132 static int
bbr_process_data(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt)8133 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8134     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8135     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8136 {
8137 	/*
8138 	 * Update window information. Don't look at window if no ACK: TAC's
8139 	 * send garbage on first SYN.
8140 	 */
8141 	uint16_t nsegs;
8142 	int32_t tfo_syn;
8143 	struct tcp_bbr *bbr;
8144 
8145 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8146 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8147 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8148 	if ((thflags & TH_ACK) &&
8149 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8150 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8151 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8152 		/* keep track of pure window updates */
8153 		if (tlen == 0 &&
8154 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8155 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
8156 		tp->snd_wnd = tiwin;
8157 		tp->snd_wl1 = th->th_seq;
8158 		tp->snd_wl2 = th->th_ack;
8159 		if (tp->snd_wnd > tp->max_sndwnd)
8160 			tp->max_sndwnd = tp->snd_wnd;
8161 		bbr->r_wanted_output = 1;
8162 	} else if (thflags & TH_ACK) {
8163 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8164 			tp->snd_wnd = tiwin;
8165 			tp->snd_wl1 = th->th_seq;
8166 			tp->snd_wl2 = th->th_ack;
8167 		}
8168 	}
8169 	if (tp->snd_wnd < ctf_outstanding(tp))
8170 		/* The peer collapsed its window on us */
8171 		bbr_collapsed_window(bbr);
8172  	else if (bbr->rc_has_collapsed)
8173 		bbr_un_collapse_window(bbr);
8174 	/* Was persist timer active and now we have window space? */
8175 	if ((bbr->rc_in_persist != 0) &&
8176 	    (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8177 				bbr_minseg(bbr)))) {
8178 		/*
8179 		 * Make the rate persist at end of persist mode if idle long
8180 		 * enough
8181 		 */
8182 		bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8183 
8184 		/* Make sure we output to start the timer */
8185 		bbr->r_wanted_output = 1;
8186 	}
8187 	/* Do we need to enter persist? */
8188 	if ((bbr->rc_in_persist == 0) &&
8189 	    (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8190 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
8191 	    (tp->snd_max == tp->snd_una) &&
8192 	    sbavail(&so->so_snd) &&
8193 	    (sbavail(&so->so_snd) > tp->snd_wnd)) {
8194 		/* No send window.. we must enter persist */
8195 		bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8196 	}
8197 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8198 		m_freem(m);
8199 		return (0);
8200 	}
8201 	/*
8202 	 * We don't support urgent data but
8203 	 * drag along the up just to make sure
8204 	 * if there is a stack switch no one
8205 	 * is surprised.
8206 	 */
8207 	tp->rcv_up = tp->rcv_nxt;
8208 
8209 	/*
8210 	 * Process the segment text, merging it into the TCP sequencing
8211 	 * queue, and arranging for acknowledgment of receipt if necessary.
8212 	 * This process logically involves adjusting tp->rcv_wnd as data is
8213 	 * presented to the user (this happens in tcp_usrreq.c, case
8214 	 * PRU_RCVD).  If a FIN has already been received on this connection
8215 	 * then we just ignore the text.
8216 	 */
8217 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8218 	    (tp->t_flags & TF_FASTOPEN));
8219 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
8220 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8221 		tcp_seq save_start = th->th_seq;
8222 		tcp_seq save_rnxt  = tp->rcv_nxt;
8223 		int     save_tlen  = tlen;
8224 
8225 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8226 		/*
8227 		 * Insert segment which includes th into TCP reassembly
8228 		 * queue with control block tp.  Set thflags to whether
8229 		 * reassembly now includes a segment with FIN.  This handles
8230 		 * the common case inline (segment is the next to be
8231 		 * received on an established connection, and the queue is
8232 		 * empty), avoiding linkage into and removal from the queue
8233 		 * and repetition of various conversions. Set DELACK for
8234 		 * segments received in order, but ack immediately when
8235 		 * segments are out of order (so fast retransmit can work).
8236 		 */
8237 		if (th->th_seq == tp->rcv_nxt &&
8238 		    SEGQ_EMPTY(tp) &&
8239 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
8240 		    tfo_syn)) {
8241 #ifdef NETFLIX_SB_LIMITS
8242 			u_int mcnt, appended;
8243 
8244 			if (so->so_rcv.sb_shlim) {
8245 				mcnt = m_memcnt(m);
8246 				appended = 0;
8247 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8248 				    CFO_NOSLEEP, NULL) == false) {
8249 					counter_u64_add(tcp_sb_shlim_fails, 1);
8250 					m_freem(m);
8251 					return (0);
8252 				}
8253 			}
8254 
8255 #endif
8256 			if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) {
8257 				bbr->bbr_segs_rcvd += max(1, nsegs);
8258 				tp->t_flags |= TF_DELACK;
8259 				bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8260 			} else {
8261 				bbr->r_wanted_output = 1;
8262 				tp->t_flags |= TF_ACKNOW;
8263 			}
8264 			tp->rcv_nxt += tlen;
8265 			if (tlen &&
8266 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8267 			    (tp->t_fbyte_in == 0)) {
8268 				tp->t_fbyte_in = ticks;
8269 				if (tp->t_fbyte_in == 0)
8270 					tp->t_fbyte_in = 1;
8271 				if (tp->t_fbyte_out && tp->t_fbyte_in)
8272 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8273 			}
8274 			thflags = tcp_get_flags(th) & TH_FIN;
8275 			KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8276 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8277 			SOCK_RECVBUF_LOCK(so);
8278 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
8279 				m_freem(m);
8280 			else
8281 #ifdef NETFLIX_SB_LIMITS
8282 				appended =
8283 #endif
8284 					sbappendstream_locked(&so->so_rcv, m, 0);
8285 			/* NB: sorwakeup_locked() does an implicit unlock. */
8286 			sorwakeup_locked(so);
8287 #ifdef NETFLIX_SB_LIMITS
8288 			if (so->so_rcv.sb_shlim && appended != mcnt)
8289 				counter_fo_release(so->so_rcv.sb_shlim,
8290 				    mcnt - appended);
8291 #endif
8292 
8293 		} else {
8294 			/*
8295 			 * XXX: Due to the header drop above "th" is
8296 			 * theoretically invalid by now.  Fortunately
8297 			 * m_adj() doesn't actually frees any mbufs when
8298 			 * trimming from the head.
8299 			 */
8300 			tcp_seq temp = save_start;
8301 
8302 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
8303 			tp->t_flags |= TF_ACKNOW;
8304 			if (tp->t_flags & TF_WAKESOR) {
8305 				tp->t_flags &= ~TF_WAKESOR;
8306 				/* NB: sorwakeup_locked() does an implicit unlock. */
8307 				sorwakeup_locked(so);
8308 			}
8309 		}
8310 		if ((tp->t_flags & TF_SACK_PERMIT) &&
8311 		    (save_tlen > 0) &&
8312 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
8313 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8314 				/*
8315 				 * DSACK actually handled in the fastpath
8316 				 * above.
8317 				 */
8318 				tcp_update_sack_list(tp, save_start,
8319 				    save_start + save_tlen);
8320 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8321 				if ((tp->rcv_numsacks >= 1) &&
8322 				    (tp->sackblks[0].end == save_start)) {
8323 					/*
8324 					 * Partial overlap, recorded at todrop
8325 					 * above.
8326 					 */
8327 					tcp_update_sack_list(tp,
8328 					    tp->sackblks[0].start,
8329 					    tp->sackblks[0].end);
8330 				} else {
8331 					tcp_update_dsack_list(tp, save_start,
8332 					    save_start + save_tlen);
8333 				}
8334 			} else if (tlen >= save_tlen) {
8335 				/* Update of sackblks. */
8336 				tcp_update_dsack_list(tp, save_start,
8337 				    save_start + save_tlen);
8338 			} else if (tlen > 0) {
8339 				tcp_update_dsack_list(tp, save_start,
8340 				    save_start + tlen);
8341 			}
8342 		}
8343 	} else {
8344 		m_freem(m);
8345 		thflags &= ~TH_FIN;
8346 	}
8347 
8348 	/*
8349 	 * If FIN is received ACK the FIN and let the user know that the
8350 	 * connection is closing.
8351 	 */
8352 	if (thflags & TH_FIN) {
8353 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8354 			/* The socket upcall is handled by socantrcvmore. */
8355 			socantrcvmore(so);
8356 			/*
8357 			 * If connection is half-synchronized (ie NEEDSYN
8358 			 * flag on) then delay ACK, so it may be piggybacked
8359 			 * when SYN is sent. Otherwise, since we received a
8360 			 * FIN then no more input can be expected, send ACK
8361 			 * now.
8362 			 */
8363 			if (tp->t_flags & TF_NEEDSYN) {
8364 				tp->t_flags |= TF_DELACK;
8365 				bbr_timer_cancel(bbr,
8366 				    __LINE__, bbr->r_ctl.rc_rcvtime);
8367 			} else {
8368 				tp->t_flags |= TF_ACKNOW;
8369 			}
8370 			tp->rcv_nxt++;
8371 		}
8372 		switch (tp->t_state) {
8373 			/*
8374 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
8375 			 * CLOSE_WAIT state.
8376 			 */
8377 		case TCPS_SYN_RECEIVED:
8378 			tp->t_starttime = ticks;
8379 			/* FALLTHROUGH */
8380 		case TCPS_ESTABLISHED:
8381 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
8382 			break;
8383 
8384 			/*
8385 			 * If still in FIN_WAIT_1 STATE FIN has not been
8386 			 * acked so enter the CLOSING state.
8387 			 */
8388 		case TCPS_FIN_WAIT_1:
8389 			tcp_state_change(tp, TCPS_CLOSING);
8390 			break;
8391 
8392 			/*
8393 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
8394 			 * starting the time-wait timer, turning off the
8395 			 * other standard timers.
8396 			 */
8397 		case TCPS_FIN_WAIT_2:
8398 			bbr->rc_timer_first = 1;
8399 			bbr_timer_cancel(bbr,
8400 			    __LINE__, bbr->r_ctl.rc_rcvtime);
8401 			tcp_twstart(tp);
8402 			return (1);
8403 		}
8404 	}
8405 	/*
8406 	 * Return any desired output.
8407 	 */
8408 	if ((tp->t_flags & TF_ACKNOW) ||
8409 	    (sbavail(&so->so_snd) > ctf_outstanding(tp))) {
8410 		bbr->r_wanted_output = 1;
8411 	}
8412 	return (0);
8413 }
8414 
8415 /*
8416  * Here nothing is really faster, its just that we
8417  * have broken out the fast-data path also just like
8418  * the fast-ack. Return 1 if we processed the packet
8419  * return 0 if you need to take the "slow-path".
8420  */
8421 static int
bbr_do_fastnewdata(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t nxt_pkt)8422 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8423     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8424     uint32_t tiwin, int32_t nxt_pkt)
8425 {
8426 	uint16_t nsegs;
8427 	int32_t newsize = 0;	/* automatic sockbuf scaling */
8428 	struct tcp_bbr *bbr;
8429 #ifdef NETFLIX_SB_LIMITS
8430 	u_int mcnt, appended;
8431 #endif
8432 
8433 	/* On the hpts and we would have called output */
8434 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8435 
8436 	/*
8437 	 * If last ACK falls within this segment's sequence numbers, record
8438 	 * the timestamp. NOTE that the test is modified according to the
8439 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8440 	 */
8441 	if (bbr->r_ctl.rc_resend != NULL) {
8442 		return (0);
8443 	}
8444 	if (tiwin && tiwin != tp->snd_wnd) {
8445 		return (0);
8446 	}
8447 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8448 		return (0);
8449 	}
8450 	if (__predict_false((to->to_flags & TOF_TS) &&
8451 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8452 		return (0);
8453 	}
8454 	if (__predict_false((th->th_ack != tp->snd_una))) {
8455 		return (0);
8456 	}
8457 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8458 		return (0);
8459 	}
8460 	if ((to->to_flags & TOF_TS) != 0 &&
8461 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8462 		tp->ts_recent_age = tcp_tv_to_msec(&bbr->rc_tv);
8463 		tp->ts_recent = to->to_tsval;
8464 	}
8465 	/*
8466 	 * This is a pure, in-sequence data packet with nothing on the
8467 	 * reassembly queue and we have enough buffer space to take it.
8468 	 */
8469 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8470 
8471 #ifdef NETFLIX_SB_LIMITS
8472 	if (so->so_rcv.sb_shlim) {
8473 		mcnt = m_memcnt(m);
8474 		appended = 0;
8475 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8476 		    CFO_NOSLEEP, NULL) == false) {
8477 			counter_u64_add(tcp_sb_shlim_fails, 1);
8478 			m_freem(m);
8479 			return (1);
8480 		}
8481 	}
8482 #endif
8483 	/* Clean receiver SACK report if present */
8484 	if (tp->rcv_numsacks)
8485 		tcp_clean_sackreport(tp);
8486 	KMOD_TCPSTAT_INC(tcps_preddat);
8487 	tp->rcv_nxt += tlen;
8488 	if (tlen &&
8489 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8490 	    (tp->t_fbyte_in == 0)) {
8491 		tp->t_fbyte_in = ticks;
8492 		if (tp->t_fbyte_in == 0)
8493 			tp->t_fbyte_in = 1;
8494 		if (tp->t_fbyte_out && tp->t_fbyte_in)
8495 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8496 	}
8497 	/*
8498 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8499 	 */
8500 	tp->snd_wl1 = th->th_seq;
8501 	/*
8502 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8503 	 */
8504 	tp->rcv_up = tp->rcv_nxt;
8505 	KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8506 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8507 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8508 
8509 	/* Add data to socket buffer. */
8510 	SOCK_RECVBUF_LOCK(so);
8511 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8512 		m_freem(m);
8513 	} else {
8514 		/*
8515 		 * Set new socket buffer size. Give up when limit is
8516 		 * reached.
8517 		 */
8518 		if (newsize)
8519 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
8520 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8521 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8522 
8523 #ifdef NETFLIX_SB_LIMITS
8524 		appended =
8525 #endif
8526 			sbappendstream_locked(&so->so_rcv, m, 0);
8527 		ctf_calc_rwin(so, tp);
8528 	}
8529 	/* NB: sorwakeup_locked() does an implicit unlock. */
8530 	sorwakeup_locked(so);
8531 #ifdef NETFLIX_SB_LIMITS
8532 	if (so->so_rcv.sb_shlim && mcnt != appended)
8533 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8534 #endif
8535 	if (DELAY_ACK(tp, bbr, nsegs)) {
8536 		bbr->bbr_segs_rcvd += max(1, nsegs);
8537 		tp->t_flags |= TF_DELACK;
8538 		bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8539 	} else {
8540 		bbr->r_wanted_output = 1;
8541 		tp->t_flags |= TF_ACKNOW;
8542 	}
8543 	return (1);
8544 }
8545 
8546 /*
8547  * This subfunction is used to try to highly optimize the
8548  * fast path. We again allow window updates that are
8549  * in sequence to remain in the fast-path. We also add
8550  * in the __predict's to attempt to help the compiler.
8551  * Note that if we return a 0, then we can *not* process
8552  * it and the caller should push the packet into the
8553  * slow-path. If we return 1, then all is well and
8554  * the packet is fully processed.
8555  */
8556 static int
bbr_fastack(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t nxt_pkt,uint8_t iptos)8557 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8558     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8559     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
8560 {
8561 	int32_t acked;
8562 	uint16_t nsegs;
8563 	uint32_t sack_changed;
8564 	uint32_t prev_acked = 0;
8565 	struct tcp_bbr *bbr;
8566 
8567 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8568 		/* Old ack, behind (or duplicate to) the last one rcv'd */
8569 		return (0);
8570 	}
8571 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
8572 		/* Above what we have sent? */
8573 		return (0);
8574 	}
8575 	if (__predict_false(tiwin == 0)) {
8576 		/* zero window */
8577 		return (0);
8578 	}
8579 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
8580 		/* We need a SYN or a FIN, unlikely.. */
8581 		return (0);
8582 	}
8583 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
8584 		/* Timestamp is behind .. old ack with seq wrap? */
8585 		return (0);
8586 	}
8587 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
8588 		/* Still recovering */
8589 		return (0);
8590 	}
8591 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8592 	if (__predict_false(bbr->r_ctl.rc_resend != NULL)) {
8593 		/* We are retransmitting */
8594 		return (0);
8595 	}
8596 	if (__predict_false(bbr->rc_in_persist != 0)) {
8597 		/* In persist mode */
8598 		return (0);
8599 	}
8600 	if (bbr->r_ctl.rc_sacked) {
8601 		/* We have sack holes on our scoreboard */
8602 		return (0);
8603 	}
8604 	/* Ok if we reach here, we can process a fast-ack */
8605 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8606 	sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
8607 	/*
8608 	 * We never detect loss in fast ack [we can't
8609 	 * have a sack and can't be in recovery so
8610 	 * we always pass 0 (nothing detected)].
8611 	 */
8612 	bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0);
8613 	/* Did the window get updated? */
8614 	if (tiwin != tp->snd_wnd) {
8615 		tp->snd_wnd = tiwin;
8616 		tp->snd_wl1 = th->th_seq;
8617 		if (tp->snd_wnd > tp->max_sndwnd)
8618 			tp->max_sndwnd = tp->snd_wnd;
8619 	}
8620 	/* Do we need to exit persists? */
8621 	if ((bbr->rc_in_persist != 0) &&
8622 	    (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8623 			       bbr_minseg(bbr)))) {
8624 		bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8625 		bbr->r_wanted_output = 1;
8626 	}
8627 	/* Do we need to enter persists? */
8628 	if ((bbr->rc_in_persist == 0) &&
8629 	    (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8630 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
8631 	    (tp->snd_max == tp->snd_una) &&
8632 	    sbavail(&so->so_snd) &&
8633 	    (sbavail(&so->so_snd) > tp->snd_wnd)) {
8634 		/* No send window.. we must enter persist */
8635 		bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8636 	}
8637 	/*
8638 	 * If last ACK falls within this segment's sequence numbers, record
8639 	 * the timestamp. NOTE that the test is modified according to the
8640 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8641 	 */
8642 	if ((to->to_flags & TOF_TS) != 0 &&
8643 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8644 		tp->ts_recent_age = bbr->r_ctl.rc_rcvtime;
8645 		tp->ts_recent = to->to_tsval;
8646 	}
8647 	/*
8648 	 * This is a pure ack for outstanding data.
8649 	 */
8650 	KMOD_TCPSTAT_INC(tcps_predack);
8651 
8652 	/*
8653 	 * "bad retransmit" recovery.
8654 	 */
8655 	if (tp->t_flags & TF_PREVVALID) {
8656 		tp->t_flags &= ~TF_PREVVALID;
8657 		if (tp->t_rxtshift == 1 &&
8658 		    (int)(ticks - tp->t_badrxtwin) < 0)
8659 			bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
8660 	}
8661 	/*
8662 	 * Recalculate the transmit timer / rtt.
8663 	 *
8664 	 * Some boxes send broken timestamp replies during the SYN+ACK
8665 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
8666 	 * and blow up the retransmit timer.
8667 	 */
8668 	acked = BYTES_THIS_ACK(tp, th);
8669 
8670 #ifdef TCP_HHOOK
8671 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
8672 	hhook_run_tcp_est_in(tp, th, to);
8673 #endif
8674 
8675 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
8676 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
8677 	sbdrop(&so->so_snd, acked);
8678 
8679 	if (SEQ_GT(th->th_ack, tp->snd_una))
8680 		bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
8681 	tp->snd_una = th->th_ack;
8682 	if (tp->snd_wnd < ctf_outstanding(tp))
8683 		/* The peer collapsed its window on us */
8684 		bbr_collapsed_window(bbr);
8685 	else if (bbr->rc_has_collapsed)
8686 		bbr_un_collapse_window(bbr);
8687 
8688 	if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8689 		tp->snd_recover = tp->snd_una;
8690 	}
8691 	bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0);
8692 	/*
8693 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
8694 	 */
8695 	tp->snd_wl2 = th->th_ack;
8696 	m_freem(m);
8697 	/*
8698 	 * If all outstanding data are acked, stop retransmit timer,
8699 	 * otherwise restart timer using current (possibly backed-off)
8700 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
8701 	 * If data are ready to send, let tcp_output decide between more
8702 	 * output or persist.
8703 	 * Wake up the socket if we have room to write more.
8704 	 */
8705 	sowwakeup(so);
8706 	if (tp->snd_una == tp->snd_max) {
8707 		/* Nothing left outstanding */
8708 		bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8709 		if (sbavail(&so->so_snd) == 0)
8710 			bbr->rc_tp->t_acktime = 0;
8711 		bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8712 		if (bbr->rc_in_persist == 0) {
8713 			bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8714 		}
8715 		sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8716 		bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8717 		/*
8718 		 * We invalidate the last ack here since we
8719 		 * don't want to transfer forward the time
8720 		 * for our sum's calculations.
8721 		 */
8722 		bbr->r_wanted_output = 1;
8723 	}
8724 	if (sbavail(&so->so_snd)) {
8725 		bbr->r_wanted_output = 1;
8726 	}
8727 	return (1);
8728 }
8729 
8730 /*
8731  * Return value of 1, the TCB is unlocked and most
8732  * likely gone, return value of 0, the TCB is still
8733  * locked.
8734  */
8735 static int
bbr_do_syn_sent(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)8736 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
8737     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8738     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
8739 {
8740 	int32_t todrop;
8741 	int32_t ourfinisacked = 0;
8742 	struct tcp_bbr *bbr;
8743 	int32_t ret_val = 0;
8744 
8745 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8746 
8747 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8748 	ctf_calc_rwin(so, tp);
8749 	/*
8750 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
8751 	 * SYN, drop the input. if seg contains a RST, then drop the
8752 	 * connection. if seg does not contain SYN, then drop it. Otherwise
8753 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
8754 	 * tp->irs if seg contains ack then advance tp->snd_una. BRR does
8755 	 * not support ECN so we will not say we are capable. if SYN has
8756 	 * been acked change to ESTABLISHED else SYN_RCVD state arrange for
8757 	 * segment to be acked (eventually) continue processing rest of
8758 	 * data/controls, beginning with URG
8759 	 */
8760 	if ((thflags & TH_ACK) &&
8761 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
8762 	    SEQ_GT(th->th_ack, tp->snd_max))) {
8763 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8764 		ctf_do_dropwithreset(m, tp, th, tlen);
8765 		return (1);
8766 	}
8767 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
8768 		TCP_PROBE5(connect__refused, NULL, tp,
8769 		    mtod(m, const char *), tp, th);
8770 		tp = tcp_drop(tp, ECONNREFUSED);
8771 		ctf_do_drop(m, tp);
8772 		return (1);
8773 	}
8774 	if (thflags & TH_RST) {
8775 		ctf_do_drop(m, tp);
8776 		return (1);
8777 	}
8778 	if (!(thflags & TH_SYN)) {
8779 		ctf_do_drop(m, tp);
8780 		return (1);
8781 	}
8782 	tp->irs = th->th_seq;
8783 	tcp_rcvseqinit(tp);
8784 	if (thflags & TH_ACK) {
8785 		int tfo_partial = 0;
8786 
8787 		KMOD_TCPSTAT_INC(tcps_connects);
8788 		soisconnected(so);
8789 #ifdef MAC
8790 		mac_socketpeer_set_from_mbuf(m, so);
8791 #endif
8792 		/* Do window scaling on this connection? */
8793 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
8794 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
8795 			tp->rcv_scale = tp->request_r_scale;
8796 		}
8797 		tp->rcv_adv += min(tp->rcv_wnd,
8798 		    TCP_MAXWIN << tp->rcv_scale);
8799 		/*
8800 		 * If not all the data that was sent in the TFO SYN
8801 		 * has been acked, resend the remainder right away.
8802 		 */
8803 		if ((tp->t_flags & TF_FASTOPEN) &&
8804 		    (tp->snd_una != tp->snd_max)) {
8805 			tp->snd_nxt = th->th_ack;
8806 			tfo_partial = 1;
8807 		}
8808 		/*
8809 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
8810 		 * will be turned on later.
8811 		 */
8812 		if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) {
8813 			bbr->bbr_segs_rcvd += 1;
8814 			tp->t_flags |= TF_DELACK;
8815 			bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8816 		} else {
8817 			bbr->r_wanted_output = 1;
8818 			tp->t_flags |= TF_ACKNOW;
8819 		}
8820 		if (SEQ_GT(th->th_ack, tp->iss)) {
8821 			/*
8822 			 * The SYN is acked
8823 			 * handle it specially.
8824 			 */
8825 			bbr_log_syn(tp, to);
8826 		}
8827 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
8828 			/*
8829 			 * We advance snd_una for the
8830 			 * fast open case. If th_ack is
8831 			 * acknowledging data beyond
8832 			 * snd_una we can't just call
8833 			 * ack-processing since the
8834 			 * data stream in our send-map
8835 			 * will start at snd_una + 1 (one
8836 			 * beyond the SYN). If its just
8837 			 * equal we don't need to do that
8838 			 * and there is no send_map.
8839 			 */
8840 			tp->snd_una++;
8841 		}
8842 		/*
8843 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
8844 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
8845 		 */
8846 		tp->t_starttime = ticks;
8847 		if (tp->t_flags & TF_NEEDFIN) {
8848 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
8849 			tp->t_flags &= ~TF_NEEDFIN;
8850 			thflags &= ~TH_SYN;
8851 		} else {
8852 			tcp_state_change(tp, TCPS_ESTABLISHED);
8853 			TCP_PROBE5(connect__established, NULL, tp,
8854 			    mtod(m, const char *), tp, th);
8855 			cc_conn_init(tp);
8856 		}
8857 	} else {
8858 		/*
8859 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
8860 		 * open.  If segment contains CC option and there is a
8861 		 * cached CC, apply TAO test. If it succeeds, connection is *
8862 		 * half-synchronized. Otherwise, do 3-way handshake:
8863 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
8864 		 * there was no CC option, clear cached CC value.
8865 		 */
8866 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
8867 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
8868 	}
8869 	/*
8870 	 * Advance th->th_seq to correspond to first data byte. If data,
8871 	 * trim to stay within window, dropping FIN if necessary.
8872 	 */
8873 	th->th_seq++;
8874 	if (tlen > tp->rcv_wnd) {
8875 		todrop = tlen - tp->rcv_wnd;
8876 		m_adj(m, -todrop);
8877 		tlen = tp->rcv_wnd;
8878 		thflags &= ~TH_FIN;
8879 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
8880 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
8881 	}
8882 	tp->snd_wl1 = th->th_seq - 1;
8883 	tp->rcv_up = th->th_seq;
8884 	/*
8885 	 * Client side of transaction: already sent SYN and data. If the
8886 	 * remote host used T/TCP to validate the SYN, our data will be
8887 	 * ACK'd; if so, enter normal data segment processing in the middle
8888 	 * of step 5, ack processing. Otherwise, goto step 6.
8889 	 */
8890 	if (thflags & TH_ACK) {
8891 		if ((to->to_flags & TOF_TS) != 0) {
8892 			uint32_t t, rtt;
8893 
8894 			t = tcp_tv_to_msec(&bbr->rc_tv);
8895 			if (TSTMP_GEQ(t, to->to_tsecr)) {
8896 				rtt = t - to->to_tsecr;
8897 				if (rtt == 0) {
8898 					rtt = 1;
8899 				}
8900 				rtt *= MS_IN_USEC;
8901 				tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
8902 				apply_filter_min_small(&bbr->r_ctl.rc_rttprop,
8903 						       rtt, bbr->r_ctl.rc_rcvtime);
8904 			}
8905 		}
8906 		if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
8907 			return (ret_val);
8908 		/* We may have changed to FIN_WAIT_1 above */
8909 		if (tp->t_state == TCPS_FIN_WAIT_1) {
8910 			/*
8911 			 * In FIN_WAIT_1 STATE in addition to the processing
8912 			 * for the ESTABLISHED state if our FIN is now
8913 			 * acknowledged then enter FIN_WAIT_2.
8914 			 */
8915 			if (ourfinisacked) {
8916 				/*
8917 				 * If we can't receive any more data, then
8918 				 * closing user can proceed. Starting the
8919 				 * timer is contrary to the specification,
8920 				 * but if we don't get a FIN we'll hang
8921 				 * forever.
8922 				 *
8923 				 * XXXjl: we should release the tp also, and
8924 				 * use a compressed state.
8925 				 */
8926 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8927 					soisdisconnected(so);
8928 					tcp_timer_activate(tp, TT_2MSL,
8929 					    (tcp_fast_finwait2_recycle ?
8930 					    tcp_finwait2_timeout :
8931 					    TP_MAXIDLE(tp)));
8932 				}
8933 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
8934 			}
8935 		}
8936 	}
8937 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
8938 	    tiwin, thflags, nxt_pkt));
8939 }
8940 
8941 /*
8942  * Return value of 1, the TCB is unlocked and most
8943  * likely gone, return value of 0, the TCB is still
8944  * locked.
8945  */
8946 static int
bbr_do_syn_recv(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)8947 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
8948 		struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8949 		uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
8950 {
8951 	int32_t ourfinisacked = 0;
8952 	int32_t ret_val;
8953 	struct tcp_bbr *bbr;
8954 
8955 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8956 
8957 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8958 	ctf_calc_rwin(so, tp);
8959 	if ((thflags & TH_RST) ||
8960 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
8961 		return (ctf_process_rst(m, th, so, tp));
8962 	if ((thflags & TH_ACK) &&
8963 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
8964 	     SEQ_GT(th->th_ack, tp->snd_max))) {
8965 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8966 		ctf_do_dropwithreset(m, tp, th, tlen);
8967 		return (1);
8968 	}
8969 	if (tp->t_flags & TF_FASTOPEN) {
8970 		/*
8971 		 * When a TFO connection is in SYN_RECEIVED, the only valid
8972 		 * packets are the initial SYN, a retransmit/copy of the
8973 		 * initial SYN (possibly with a subset of the original
8974 		 * data), a valid ACK, a FIN, or a RST.
8975 		 */
8976 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
8977 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8978 			ctf_do_dropwithreset(m, tp, th, tlen);
8979 			return (1);
8980 		} else if (thflags & TH_SYN) {
8981 			/* non-initial SYN is ignored */
8982 			if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
8983 			    (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
8984 			    (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
8985 				ctf_do_drop(m, NULL);
8986 				return (0);
8987 			}
8988 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
8989 			ctf_do_drop(m, NULL);
8990 			return (0);
8991 		}
8992 	}
8993 	/*
8994 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
8995 	 * it's less than ts_recent, drop it.
8996 	 */
8997 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
8998 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
8999 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9000 			return (ret_val);
9001 	}
9002 	/*
9003 	 * In the SYN-RECEIVED state, validate that the packet belongs to
9004 	 * this connection before trimming the data to fit the receive
9005 	 * window.  Check the sequence number versus IRS since we know the
9006 	 * sequence numbers haven't wrapped.  This is a partial fix for the
9007 	 * "LAND" DoS attack.
9008 	 */
9009 	if (SEQ_LT(th->th_seq, tp->irs)) {
9010 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9011 		ctf_do_dropwithreset(m, tp, th, tlen);
9012 		return (1);
9013 	}
9014 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9015 		return (ret_val);
9016 	}
9017 	/*
9018 	 * If last ACK falls within this segment's sequence numbers, record
9019 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9020 	 * from the latest proposal of the tcplw@cray.com list (Braden
9021 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9022 	 * with our earlier PAWS tests, so this check should be solely
9023 	 * predicated on the sequence space of this segment. 3) That we
9024 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9025 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9026 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9027 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9028 	 * p.869. In such cases, we can still calculate the RTT correctly
9029 	 * when RCV.NXT == Last.ACK.Sent.
9030 	 */
9031 	if ((to->to_flags & TOF_TS) != 0 &&
9032 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9033 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9034 		    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9035 		tp->ts_recent_age = tcp_tv_to_msec(&bbr->rc_tv);
9036 		tp->ts_recent = to->to_tsval;
9037 	}
9038 	tp->snd_wnd = tiwin;
9039 	/*
9040 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9041 	 * is on (half-synchronized state), then queue data for later
9042 	 * processing; else drop segment and return.
9043 	 */
9044 	if ((thflags & TH_ACK) == 0) {
9045 		if (tp->t_flags & TF_FASTOPEN) {
9046 			cc_conn_init(tp);
9047 		}
9048 		return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9049 					 tiwin, thflags, nxt_pkt));
9050 	}
9051 	KMOD_TCPSTAT_INC(tcps_connects);
9052 	if (tp->t_flags & TF_SONOTCONN) {
9053 		tp->t_flags &= ~TF_SONOTCONN;
9054 		soisconnected(so);
9055 	}
9056 	/* Do window scaling? */
9057 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9058 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9059 		tp->rcv_scale = tp->request_r_scale;
9060 	}
9061 	/*
9062 	 * ok for the first time in lets see if we can use the ts to figure
9063 	 * out what the initial RTT was.
9064 	 */
9065 	if ((to->to_flags & TOF_TS) != 0) {
9066 		uint32_t t, rtt;
9067 
9068 		t = tcp_tv_to_msec(&bbr->rc_tv);
9069 		if (TSTMP_GEQ(t, to->to_tsecr)) {
9070 			rtt = t - to->to_tsecr;
9071 			if (rtt == 0) {
9072 				rtt = 1;
9073 			}
9074 			rtt *= MS_IN_USEC;
9075 			tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9076 			apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime);
9077 		}
9078 	}
9079 	/* Drop off any SYN in the send map (probably not there)  */
9080 	if (thflags & TH_ACK)
9081 		bbr_log_syn(tp, to);
9082 	if ((tp->t_flags & TF_FASTOPEN) && tp->t_tfo_pending) {
9083 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9084 		tp->t_tfo_pending = NULL;
9085 	}
9086 	/*
9087 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
9088 	 * FIN-WAIT-1
9089 	 */
9090 	tp->t_starttime = ticks;
9091 	if (tp->t_flags & TF_NEEDFIN) {
9092 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
9093 		tp->t_flags &= ~TF_NEEDFIN;
9094 	} else {
9095 		tcp_state_change(tp, TCPS_ESTABLISHED);
9096 		TCP_PROBE5(accept__established, NULL, tp,
9097 			   mtod(m, const char *), tp, th);
9098 		/*
9099 		 * TFO connections call cc_conn_init() during SYN
9100 		 * processing.  Calling it again here for such connections
9101 		 * is not harmless as it would undo the snd_cwnd reduction
9102 		 * that occurs when a TFO SYN|ACK is retransmitted.
9103 		 */
9104 		if (!(tp->t_flags & TF_FASTOPEN))
9105 			cc_conn_init(tp);
9106 	}
9107 	/*
9108 	 * Account for the ACK of our SYN prior to
9109 	 * regular ACK processing below, except for
9110 	 * simultaneous SYN, which is handled later.
9111 	 */
9112 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
9113 		tp->snd_una++;
9114 	/*
9115 	 * If segment contains data or ACK, will call tcp_reass() later; if
9116 	 * not, do so now to pass queued data to user.
9117 	 */
9118 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
9119 		(void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9120 			(struct mbuf *)0);
9121 		if (tp->t_flags & TF_WAKESOR) {
9122 			tp->t_flags &= ~TF_WAKESOR;
9123 			/* NB: sorwakeup_locked() does an implicit unlock. */
9124 			sorwakeup_locked(so);
9125 		}
9126 	}
9127 	tp->snd_wl1 = th->th_seq - 1;
9128 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9129 		return (ret_val);
9130 	}
9131 	if (tp->t_state == TCPS_FIN_WAIT_1) {
9132 		/* We could have went to FIN_WAIT_1 (or EST) above */
9133 		/*
9134 		 * In FIN_WAIT_1 STATE in addition to the processing for the
9135 		 * ESTABLISHED state if our FIN is now acknowledged then
9136 		 * enter FIN_WAIT_2.
9137 		 */
9138 		if (ourfinisacked) {
9139 			/*
9140 			 * If we can't receive any more data, then closing
9141 			 * user can proceed. Starting the timer is contrary
9142 			 * to the specification, but if we don't get a FIN
9143 			 * we'll hang forever.
9144 			 *
9145 			 * XXXjl: we should release the tp also, and use a
9146 			 * compressed state.
9147 			 */
9148 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9149 				soisdisconnected(so);
9150 				tcp_timer_activate(tp, TT_2MSL,
9151 						   (tcp_fast_finwait2_recycle ?
9152 						    tcp_finwait2_timeout :
9153 						    TP_MAXIDLE(tp)));
9154 			}
9155 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
9156 		}
9157 	}
9158 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9159 				 tiwin, thflags, nxt_pkt));
9160 }
9161 
9162 /*
9163  * Return value of 1, the TCB is unlocked and most
9164  * likely gone, return value of 0, the TCB is still
9165  * locked.
9166  */
9167 static int
bbr_do_established(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9168 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9169     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9170     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9171 {
9172 	struct tcp_bbr *bbr;
9173 	int32_t ret_val;
9174 
9175 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9176 
9177 	/*
9178 	 * Header prediction: check for the two common cases of a
9179 	 * uni-directional data xfer.  If the packet has no control flags,
9180 	 * is in-sequence, the window didn't change and we're not
9181 	 * retransmitting, it's a candidate.  If the length is zero and the
9182 	 * ack moved forward, we're the sender side of the xfer.  Just free
9183 	 * the data acked & wake any higher level process that was blocked
9184 	 * waiting for space.  If the length is non-zero and the ack didn't
9185 	 * move, we're the receiver side.  If we're getting packets in-order
9186 	 * (the reassembly queue is empty), add the data toc The socket
9187 	 * buffer and note that we need a delayed ack. Make sure that the
9188 	 * hidden state-flags are also off. Since we check for
9189 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9190 	 */
9191 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9192 	if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) {
9193 		/*
9194 		 * If we have delived under 4 segments increase the initial
9195 		 * window if raised by the peer. We use this to determine
9196 		 * dynamic and static rwnd's at the end of a connection.
9197 		 */
9198 		bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd);
9199 	}
9200 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9201 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) &&
9202 	    __predict_true(SEGQ_EMPTY(tp)) &&
9203 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
9204 		if (tlen == 0) {
9205 			if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9206 			    tiwin, nxt_pkt, iptos)) {
9207 				return (0);
9208 			}
9209 		} else {
9210 			if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9211 			    tiwin, nxt_pkt)) {
9212 				return (0);
9213 			}
9214 		}
9215 	}
9216 	ctf_calc_rwin(so, tp);
9217 
9218 	if ((thflags & TH_RST) ||
9219 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9220 		return (ctf_process_rst(m, th, so, tp));
9221 	/*
9222 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9223 	 * synchronized state.
9224 	 */
9225 	if (thflags & TH_SYN) {
9226 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9227 		return (ret_val);
9228 	}
9229 	/*
9230 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9231 	 * it's less than ts_recent, drop it.
9232 	 */
9233 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9234 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9235 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9236 			return (ret_val);
9237 	}
9238 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9239 		return (ret_val);
9240 	}
9241 	/*
9242 	 * If last ACK falls within this segment's sequence numbers, record
9243 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9244 	 * from the latest proposal of the tcplw@cray.com list (Braden
9245 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9246 	 * with our earlier PAWS tests, so this check should be solely
9247 	 * predicated on the sequence space of this segment. 3) That we
9248 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9249 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9250 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9251 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9252 	 * p.869. In such cases, we can still calculate the RTT correctly
9253 	 * when RCV.NXT == Last.ACK.Sent.
9254 	 */
9255 	if ((to->to_flags & TOF_TS) != 0 &&
9256 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9257 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9258 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9259 		tp->ts_recent_age = tcp_tv_to_msec(&bbr->rc_tv);
9260 		tp->ts_recent = to->to_tsval;
9261 	}
9262 	/*
9263 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9264 	 * is on (half-synchronized state), then queue data for later
9265 	 * processing; else drop segment and return.
9266 	 */
9267 	if ((thflags & TH_ACK) == 0) {
9268 		if (tp->t_flags & TF_NEEDSYN) {
9269 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9270 			    tiwin, thflags, nxt_pkt));
9271 		} else if (tp->t_flags & TF_ACKNOW) {
9272 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9273 			bbr->r_wanted_output = 1;
9274 			return (ret_val);
9275 		} else {
9276 			ctf_do_drop(m, NULL);
9277 			return (0);
9278 		}
9279 	}
9280 	/*
9281 	 * Ack processing.
9282 	 */
9283 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9284 		return (ret_val);
9285 	}
9286 	if (sbavail(&so->so_snd)) {
9287 		if (ctf_progress_timeout_check(tp, true)) {
9288 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9289 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
9290 			return (1);
9291 		}
9292 	}
9293 	/* State changes only happen in bbr_process_data() */
9294 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9295 	    tiwin, thflags, nxt_pkt));
9296 }
9297 
9298 /*
9299  * Return value of 1, the TCB is unlocked and most
9300  * likely gone, return value of 0, the TCB is still
9301  * locked.
9302  */
9303 static int
bbr_do_close_wait(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9304 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9305     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9306     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9307 {
9308 	struct tcp_bbr *bbr;
9309 	int32_t ret_val;
9310 
9311 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9312 
9313 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9314 	ctf_calc_rwin(so, tp);
9315 	if ((thflags & TH_RST) ||
9316 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9317 		return (ctf_process_rst(m, th, so, tp));
9318 	/*
9319 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9320 	 * synchronized state.
9321 	 */
9322 	if (thflags & TH_SYN) {
9323 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9324 		return (ret_val);
9325 	}
9326 	/*
9327 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9328 	 * it's less than ts_recent, drop it.
9329 	 */
9330 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9331 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9332 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9333 			return (ret_val);
9334 	}
9335 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9336 		return (ret_val);
9337 	}
9338 	/*
9339 	 * If last ACK falls within this segment's sequence numbers, record
9340 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9341 	 * from the latest proposal of the tcplw@cray.com list (Braden
9342 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9343 	 * with our earlier PAWS tests, so this check should be solely
9344 	 * predicated on the sequence space of this segment. 3) That we
9345 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9346 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9347 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9348 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9349 	 * p.869. In such cases, we can still calculate the RTT correctly
9350 	 * when RCV.NXT == Last.ACK.Sent.
9351 	 */
9352 	if ((to->to_flags & TOF_TS) != 0 &&
9353 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9354 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9355 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9356 		tp->ts_recent_age = tcp_tv_to_msec(&bbr->rc_tv);
9357 		tp->ts_recent = to->to_tsval;
9358 	}
9359 	/*
9360 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9361 	 * is on (half-synchronized state), then queue data for later
9362 	 * processing; else drop segment and return.
9363 	 */
9364 	if ((thflags & TH_ACK) == 0) {
9365 		if (tp->t_flags & TF_NEEDSYN) {
9366 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9367 			    tiwin, thflags, nxt_pkt));
9368 		} else if (tp->t_flags & TF_ACKNOW) {
9369 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9370 			bbr->r_wanted_output = 1;
9371 			return (ret_val);
9372 		} else {
9373 			ctf_do_drop(m, NULL);
9374 			return (0);
9375 		}
9376 	}
9377 	/*
9378 	 * Ack processing.
9379 	 */
9380 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9381 		return (ret_val);
9382 	}
9383 	if (sbavail(&so->so_snd)) {
9384 		if (ctf_progress_timeout_check(tp, true)) {
9385 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9386 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
9387 			return (1);
9388 		}
9389 	}
9390 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9391 	    tiwin, thflags, nxt_pkt));
9392 }
9393 
9394 static int
bbr_check_data_after_close(struct mbuf * m,struct tcp_bbr * bbr,struct tcpcb * tp,int32_t * tlen,struct tcphdr * th,struct socket * so)9395 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr,
9396     struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so)
9397 {
9398 
9399 	if (bbr->rc_allow_data_af_clo == 0) {
9400 close_now:
9401 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9402 		/* tcp_close will kill the inp pre-log the Reset */
9403 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9404 		tp = tcp_close(tp);
9405 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
9406 		ctf_do_dropwithreset(m, tp, th, *tlen);
9407 		return (1);
9408 	}
9409 	if (sbavail(&so->so_snd) == 0)
9410 		goto close_now;
9411 	/* Ok we allow data that is ignored and a followup reset */
9412 	tp->rcv_nxt = th->th_seq + *tlen;
9413 	tp->t_flags2 |= TF2_DROP_AF_DATA;
9414 	bbr->r_wanted_output = 1;
9415 	*tlen = 0;
9416 	return (0);
9417 }
9418 
9419 /*
9420  * Return value of 1, the TCB is unlocked and most
9421  * likely gone, return value of 0, the TCB is still
9422  * locked.
9423  */
9424 static int
bbr_do_fin_wait_1(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9425 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9426     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9427     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9428 {
9429 	int32_t ourfinisacked = 0;
9430 	int32_t ret_val;
9431 	struct tcp_bbr *bbr;
9432 
9433 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9434 
9435 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9436 	ctf_calc_rwin(so, tp);
9437 	if ((thflags & TH_RST) ||
9438 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9439 		return (ctf_process_rst(m, th, so, tp));
9440 	/*
9441 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9442 	 * synchronized state.
9443 	 */
9444 	if (thflags & TH_SYN) {
9445 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9446 		return (ret_val);
9447 	}
9448 	/*
9449 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9450 	 * it's less than ts_recent, drop it.
9451 	 */
9452 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9453 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9454 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9455 			return (ret_val);
9456 	}
9457 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9458 		return (ret_val);
9459 	}
9460 	/*
9461 	 * If new data are received on a connection after the user processes
9462 	 * are gone, then RST the other end.
9463 	 * We call a new function now so we might continue and setup
9464 	 * to reset at all data being ack'd.
9465 	 */
9466 	if ((tp->t_flags & TF_CLOSED) && tlen &&
9467 	    bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9468 		return (1);
9469 	/*
9470 	 * If last ACK falls within this segment's sequence numbers, record
9471 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9472 	 * from the latest proposal of the tcplw@cray.com list (Braden
9473 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9474 	 * with our earlier PAWS tests, so this check should be solely
9475 	 * predicated on the sequence space of this segment. 3) That we
9476 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9477 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9478 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9479 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9480 	 * p.869. In such cases, we can still calculate the RTT correctly
9481 	 * when RCV.NXT == Last.ACK.Sent.
9482 	 */
9483 	if ((to->to_flags & TOF_TS) != 0 &&
9484 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9485 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9486 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9487 		tp->ts_recent_age = tcp_tv_to_msec(&bbr->rc_tv);
9488 		tp->ts_recent = to->to_tsval;
9489 	}
9490 	/*
9491 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9492 	 * is on (half-synchronized state), then queue data for later
9493 	 * processing; else drop segment and return.
9494 	 */
9495 	if ((thflags & TH_ACK) == 0) {
9496 		if (tp->t_flags & TF_NEEDSYN) {
9497 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9498 			    tiwin, thflags, nxt_pkt));
9499 		} else if (tp->t_flags & TF_ACKNOW) {
9500 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9501 			bbr->r_wanted_output = 1;
9502 			return (ret_val);
9503 		} else {
9504 			ctf_do_drop(m, NULL);
9505 			return (0);
9506 		}
9507 	}
9508 	/*
9509 	 * Ack processing.
9510 	 */
9511 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9512 		return (ret_val);
9513 	}
9514 	if (ourfinisacked) {
9515 		/*
9516 		 * If we can't receive any more data, then closing user can
9517 		 * proceed. Starting the timer is contrary to the
9518 		 * specification, but if we don't get a FIN we'll hang
9519 		 * forever.
9520 		 *
9521 		 * XXXjl: we should release the tp also, and use a
9522 		 * compressed state.
9523 		 */
9524 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9525 			soisdisconnected(so);
9526 			tcp_timer_activate(tp, TT_2MSL,
9527 			    (tcp_fast_finwait2_recycle ?
9528 			    tcp_finwait2_timeout :
9529 			    TP_MAXIDLE(tp)));
9530 		}
9531 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
9532 	}
9533 	if (sbavail(&so->so_snd)) {
9534 		if (ctf_progress_timeout_check(tp, true)) {
9535 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9536 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
9537 			return (1);
9538 		}
9539 	}
9540 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9541 	    tiwin, thflags, nxt_pkt));
9542 }
9543 
9544 /*
9545  * Return value of 1, the TCB is unlocked and most
9546  * likely gone, return value of 0, the TCB is still
9547  * locked.
9548  */
9549 static int
bbr_do_closing(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9550 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9551     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9552     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9553 {
9554 	int32_t ourfinisacked = 0;
9555 	int32_t ret_val;
9556 	struct tcp_bbr *bbr;
9557 
9558 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9559 
9560 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9561 	ctf_calc_rwin(so, tp);
9562 	if ((thflags & TH_RST) ||
9563 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9564 		return (ctf_process_rst(m, th, so, tp));
9565 	/*
9566 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9567 	 * synchronized state.
9568 	 */
9569 	if (thflags & TH_SYN) {
9570 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9571 		return (ret_val);
9572 	}
9573 	/*
9574 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9575 	 * it's less than ts_recent, drop it.
9576 	 */
9577 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9578 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9579 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9580 			return (ret_val);
9581 	}
9582 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9583 		return (ret_val);
9584 	}
9585 	/*
9586 	 * If last ACK falls within this segment's sequence numbers, record
9587 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9588 	 * from the latest proposal of the tcplw@cray.com list (Braden
9589 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9590 	 * with our earlier PAWS tests, so this check should be solely
9591 	 * predicated on the sequence space of this segment. 3) That we
9592 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9593 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9594 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9595 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9596 	 * p.869. In such cases, we can still calculate the RTT correctly
9597 	 * when RCV.NXT == Last.ACK.Sent.
9598 	 */
9599 	if ((to->to_flags & TOF_TS) != 0 &&
9600 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9601 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9602 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9603 		tp->ts_recent_age = tcp_tv_to_msec(&bbr->rc_tv);
9604 		tp->ts_recent = to->to_tsval;
9605 	}
9606 	/*
9607 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9608 	 * is on (half-synchronized state), then queue data for later
9609 	 * processing; else drop segment and return.
9610 	 */
9611 	if ((thflags & TH_ACK) == 0) {
9612 		if (tp->t_flags & TF_NEEDSYN) {
9613 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9614 			    tiwin, thflags, nxt_pkt));
9615 		} else if (tp->t_flags & TF_ACKNOW) {
9616 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9617 			bbr->r_wanted_output = 1;
9618 			return (ret_val);
9619 		} else {
9620 			ctf_do_drop(m, NULL);
9621 			return (0);
9622 		}
9623 	}
9624 	/*
9625 	 * Ack processing.
9626 	 */
9627 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9628 		return (ret_val);
9629 	}
9630 	if (ourfinisacked) {
9631 		tcp_twstart(tp);
9632 		m_freem(m);
9633 		return (1);
9634 	}
9635 	if (sbavail(&so->so_snd)) {
9636 		if (ctf_progress_timeout_check(tp, true)) {
9637 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9638 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
9639 			return (1);
9640 		}
9641 	}
9642 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9643 	    tiwin, thflags, nxt_pkt));
9644 }
9645 
9646 /*
9647  * Return value of 1, the TCB is unlocked and most
9648  * likely gone, return value of 0, the TCB is still
9649  * locked.
9650  */
9651 static int
bbr_do_lastack(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9652 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9653     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9654     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9655 {
9656 	int32_t ourfinisacked = 0;
9657 	int32_t ret_val;
9658 	struct tcp_bbr *bbr;
9659 
9660 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9661 
9662 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9663 	ctf_calc_rwin(so, tp);
9664 	if ((thflags & TH_RST) ||
9665 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9666 		return (ctf_process_rst(m, th, so, tp));
9667 	/*
9668 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9669 	 * synchronized state.
9670 	 */
9671 	if (thflags & TH_SYN) {
9672 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9673 		return (ret_val);
9674 	}
9675 	/*
9676 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9677 	 * it's less than ts_recent, drop it.
9678 	 */
9679 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9680 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9681 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9682 			return (ret_val);
9683 	}
9684 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9685 		return (ret_val);
9686 	}
9687 	/*
9688 	 * If last ACK falls within this segment's sequence numbers, record
9689 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9690 	 * from the latest proposal of the tcplw@cray.com list (Braden
9691 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9692 	 * with our earlier PAWS tests, so this check should be solely
9693 	 * predicated on the sequence space of this segment. 3) That we
9694 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9695 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9696 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9697 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9698 	 * p.869. In such cases, we can still calculate the RTT correctly
9699 	 * when RCV.NXT == Last.ACK.Sent.
9700 	 */
9701 	if ((to->to_flags & TOF_TS) != 0 &&
9702 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9703 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9704 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9705 		tp->ts_recent_age = tcp_tv_to_msec(&bbr->rc_tv);
9706 		tp->ts_recent = to->to_tsval;
9707 	}
9708 	/*
9709 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9710 	 * is on (half-synchronized state), then queue data for later
9711 	 * processing; else drop segment and return.
9712 	 */
9713 	if ((thflags & TH_ACK) == 0) {
9714 		if (tp->t_flags & TF_NEEDSYN) {
9715 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9716 			    tiwin, thflags, nxt_pkt));
9717 		} else if (tp->t_flags & TF_ACKNOW) {
9718 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9719 			bbr->r_wanted_output = 1;
9720 			return (ret_val);
9721 		} else {
9722 			ctf_do_drop(m, NULL);
9723 			return (0);
9724 		}
9725 	}
9726 	/*
9727 	 * case TCPS_LAST_ACK: Ack processing.
9728 	 */
9729 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9730 		return (ret_val);
9731 	}
9732 	if (ourfinisacked) {
9733 		tp = tcp_close(tp);
9734 		ctf_do_drop(m, tp);
9735 		return (1);
9736 	}
9737 	if (sbavail(&so->so_snd)) {
9738 		if (ctf_progress_timeout_check(tp, true)) {
9739 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9740 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
9741 			return (1);
9742 		}
9743 	}
9744 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9745 	    tiwin, thflags, nxt_pkt));
9746 }
9747 
9748 /*
9749  * Return value of 1, the TCB is unlocked and most
9750  * likely gone, return value of 0, the TCB is still
9751  * locked.
9752  */
9753 static int
bbr_do_fin_wait_2(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)9754 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
9755     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9756     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9757 {
9758 	int32_t ourfinisacked = 0;
9759 	int32_t ret_val;
9760 	struct tcp_bbr *bbr;
9761 
9762 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9763 
9764 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9765 	ctf_calc_rwin(so, tp);
9766 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
9767 	if ((thflags & TH_RST) ||
9768 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9769 		return (ctf_process_rst(m, th, so, tp));
9770 
9771 	/*
9772 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9773 	 * synchronized state.
9774 	 */
9775 	if (thflags & TH_SYN) {
9776 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
9777 		return (ret_val);
9778 	}
9779 	/*
9780 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9781 	 * it's less than ts_recent, drop it.
9782 	 */
9783 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9784 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9785 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9786 			return (ret_val);
9787 	}
9788 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9789 		return (ret_val);
9790 	}
9791 	/*
9792 	 * If new data are received on a connection after the user processes
9793 	 * are gone, then we may RST the other end depending on the outcome
9794 	 * of bbr_check_data_after_close.
9795 	 * We call a new function now so we might continue and setup
9796 	 * to reset at all data being ack'd.
9797 	 */
9798 	if ((tp->t_flags & TF_CLOSED) && tlen &&
9799 	    bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9800 		return (1);
9801 	/*
9802 	 * If last ACK falls within this segment's sequence numbers, record
9803 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9804 	 * from the latest proposal of the tcplw@cray.com list (Braden
9805 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9806 	 * with our earlier PAWS tests, so this check should be solely
9807 	 * predicated on the sequence space of this segment. 3) That we
9808 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9809 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9810 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9811 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9812 	 * p.869. In such cases, we can still calculate the RTT correctly
9813 	 * when RCV.NXT == Last.ACK.Sent.
9814 	 */
9815 	if ((to->to_flags & TOF_TS) != 0 &&
9816 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9817 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9818 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9819 		tp->ts_recent_age = tcp_tv_to_msec(&bbr->rc_tv);
9820 		tp->ts_recent = to->to_tsval;
9821 	}
9822 	/*
9823 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9824 	 * is on (half-synchronized state), then queue data for later
9825 	 * processing; else drop segment and return.
9826 	 */
9827 	if ((thflags & TH_ACK) == 0) {
9828 		if (tp->t_flags & TF_NEEDSYN) {
9829 			return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9830 			    tiwin, thflags, nxt_pkt));
9831 		} else if (tp->t_flags & TF_ACKNOW) {
9832 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9833 			bbr->r_wanted_output = 1;
9834 			return (ret_val);
9835 		} else {
9836 			ctf_do_drop(m, NULL);
9837 			return (0);
9838 		}
9839 	}
9840 	/*
9841 	 * Ack processing.
9842 	 */
9843 	if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9844 		return (ret_val);
9845 	}
9846 	if (sbavail(&so->so_snd)) {
9847 		if (ctf_progress_timeout_check(tp, true)) {
9848 			bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9849 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
9850 			return (1);
9851 		}
9852 	}
9853 	return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9854 	    tiwin, thflags, nxt_pkt));
9855 }
9856 
9857 static void
bbr_stop_all_timers(struct tcpcb * tp,struct tcp_bbr * bbr)9858 bbr_stop_all_timers(struct tcpcb *tp, struct tcp_bbr *bbr)
9859 {
9860 	/*
9861 	 * Assure no timers are running.
9862 	 */
9863 	if (tcp_timer_active(tp, TT_PERSIST)) {
9864 		/* We enter in persists, set the flag appropriately */
9865 		bbr->rc_in_persist = 1;
9866 	}
9867 	if (tcp_in_hpts(bbr->rc_tp)) {
9868 		tcp_hpts_remove(bbr->rc_tp);
9869 	}
9870 }
9871 
9872 static void
bbr_google_mode_on(struct tcp_bbr * bbr)9873 bbr_google_mode_on(struct tcp_bbr *bbr)
9874 {
9875 	bbr->rc_use_google = 1;
9876 	bbr->rc_no_pacing = 0;
9877 	bbr->r_ctl.bbr_google_discount = bbr_google_discount;
9878 	bbr->r_use_policer = bbr_policer_detection_enabled;
9879 	bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
9880 	bbr->bbr_use_rack_cheat = 0;
9881 	bbr->r_ctl.rc_incr_tmrs = 0;
9882 	bbr->r_ctl.rc_inc_tcp_oh = 0;
9883 	bbr->r_ctl.rc_inc_ip_oh = 0;
9884 	bbr->r_ctl.rc_inc_enet_oh = 0;
9885 	reset_time(&bbr->r_ctl.rc_delrate,
9886 		   BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
9887 	reset_time_small(&bbr->r_ctl.rc_rttprop,
9888 			 (11 * USECS_IN_SECOND));
9889 	tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
9890 }
9891 
9892 static void
bbr_google_mode_off(struct tcp_bbr * bbr)9893 bbr_google_mode_off(struct tcp_bbr *bbr)
9894 {
9895 	bbr->rc_use_google = 0;
9896 	bbr->r_ctl.bbr_google_discount = 0;
9897 	bbr->no_pacing_until = bbr_no_pacing_until;
9898 	bbr->r_use_policer = 0;
9899 	if (bbr->no_pacing_until)
9900 		bbr->rc_no_pacing = 1;
9901 	else
9902 		bbr->rc_no_pacing = 0;
9903 	if (bbr_use_rack_resend_cheat)
9904 		bbr->bbr_use_rack_cheat = 1;
9905 	else
9906 		bbr->bbr_use_rack_cheat = 0;
9907 	if (bbr_incr_timers)
9908 		bbr->r_ctl.rc_incr_tmrs = 1;
9909 	else
9910 		bbr->r_ctl.rc_incr_tmrs = 0;
9911 	if (bbr_include_tcp_oh)
9912 		bbr->r_ctl.rc_inc_tcp_oh = 1;
9913 	else
9914 		bbr->r_ctl.rc_inc_tcp_oh = 0;
9915 	if (bbr_include_ip_oh)
9916 		bbr->r_ctl.rc_inc_ip_oh = 1;
9917 	else
9918 		bbr->r_ctl.rc_inc_ip_oh = 0;
9919 	if (bbr_include_enet_oh)
9920 		bbr->r_ctl.rc_inc_enet_oh = 1;
9921 	else
9922 		bbr->r_ctl.rc_inc_enet_oh = 0;
9923 	bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
9924 	reset_time(&bbr->r_ctl.rc_delrate,
9925 		   bbr_num_pktepo_for_del_limit);
9926 	reset_time_small(&bbr->r_ctl.rc_rttprop,
9927 			 (bbr_filter_len_sec * USECS_IN_SECOND));
9928 	tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
9929 }
9930 /*
9931  * Return 0 on success, non-zero on failure
9932  * which indicates the error (usually no memory).
9933  */
9934 static int
bbr_init(struct tcpcb * tp,void ** ptr)9935 bbr_init(struct tcpcb *tp, void **ptr)
9936 {
9937 	struct inpcb *inp = tptoinpcb(tp);
9938 	struct tcp_bbr *bbr = NULL;
9939 	uint32_t cts;
9940 
9941 	tcp_hpts_init(tp);
9942 
9943 	*ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO));
9944 	if (*ptr == NULL) {
9945 		/*
9946 		 * We need to allocate memory but cant. The INP and INP_INFO
9947 		 * locks and they are recursive (happens during setup. So a
9948 		 * scheme to drop the locks fails :(
9949 		 *
9950 		 */
9951 		return (ENOMEM);
9952 	}
9953 	bbr = (struct tcp_bbr *)*ptr;
9954 	bbr->rtt_valid = 0;
9955 	tp->t_flags2 |= TF2_CANNOT_DO_ECN;
9956 	tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
9957 	/* Take off any undesired flags */
9958 	tp->t_flags2 &= ~TF2_MBUF_QUEUE_READY;
9959 	tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
9960 	tp->t_flags2 &= ~TF2_MBUF_ACKCMP;
9961 	tp->t_flags2 &= ~TF2_MBUF_L_ACKS;
9962 
9963 	TAILQ_INIT(&bbr->r_ctl.rc_map);
9964 	TAILQ_INIT(&bbr->r_ctl.rc_free);
9965 	TAILQ_INIT(&bbr->r_ctl.rc_tmap);
9966 	bbr->rc_tp = tp;
9967 	bbr->rc_inp = inp;
9968 	cts = tcp_get_usecs(&bbr->rc_tv);
9969 	tp->t_acktime = 0;
9970 	bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close;
9971 	bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade;
9972 	bbr->rc_tlp_threshold = bbr_tlp_thresh;
9973 	bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh;
9974 	bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay;
9975 	bbr->r_ctl.rc_min_to = bbr_min_to;
9976 	bbr->rc_bbr_state = BBR_STATE_STARTUP;
9977 	bbr->r_ctl.bbr_lost_at_state = 0;
9978 	bbr->r_ctl.rc_lost_at_startup = 0;
9979 	bbr->rc_all_timers_stopped = 0;
9980 	bbr->r_ctl.rc_bbr_lastbtlbw = 0;
9981 	bbr->r_ctl.rc_pkt_epoch_del = 0;
9982 	bbr->r_ctl.rc_pkt_epoch = 0;
9983 	bbr->r_ctl.rc_lowest_rtt = 0xffffffff;
9984 	bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain;
9985 	bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
9986 	bbr->r_ctl.rc_went_idle_time = cts;
9987 	bbr->rc_pacer_started = cts;
9988 	bbr->r_ctl.rc_pkt_epoch_time = cts;
9989 	bbr->r_ctl.rc_rcvtime = cts;
9990 	bbr->r_ctl.rc_bbr_state_time = cts;
9991 	bbr->r_ctl.rc_del_time = cts;
9992 	bbr->r_ctl.rc_tlp_rxt_last_time = cts;
9993 	bbr->r_ctl.last_in_probertt = cts;
9994 	bbr->skip_gain = 0;
9995 	bbr->gain_is_limited = 0;
9996 	bbr->no_pacing_until = bbr_no_pacing_until;
9997 	if (bbr->no_pacing_until)
9998 		bbr->rc_no_pacing = 1;
9999 	if (bbr_use_google_algo) {
10000 		bbr->rc_no_pacing = 0;
10001 		bbr->rc_use_google = 1;
10002 		bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10003 		bbr->r_use_policer = bbr_policer_detection_enabled;
10004 	} else {
10005 		bbr->rc_use_google = 0;
10006 		bbr->r_ctl.bbr_google_discount = 0;
10007 		bbr->r_use_policer = 0;
10008 	}
10009 	if (bbr_ts_limiting)
10010 		bbr->rc_use_ts_limit = 1;
10011 	else
10012 		bbr->rc_use_ts_limit = 0;
10013 	if (bbr_ts_can_raise)
10014 		bbr->ts_can_raise = 1;
10015 	else
10016 		bbr->ts_can_raise = 0;
10017 	if (V_tcp_delack_enabled == 1)
10018 		tp->t_delayed_ack = 2;
10019 	else if (V_tcp_delack_enabled == 0)
10020 		tp->t_delayed_ack = 0;
10021 	else if (V_tcp_delack_enabled < 100)
10022 		tp->t_delayed_ack = V_tcp_delack_enabled;
10023 	else
10024 		tp->t_delayed_ack = 2;
10025 	if (bbr->rc_use_google == 0)
10026 		bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10027 	else
10028 		bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10029 	bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms;
10030 	bbr->rc_max_rto_sec = bbr_rto_max_sec;
10031 	bbr->rc_init_win = bbr_def_init_win;
10032 	if (tp->t_flags & TF_REQ_TSTMP)
10033 		bbr->rc_last_options = TCP_TS_OVERHEAD;
10034 	bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options;
10035 	bbr->r_ctl.rc_high_rwnd = tp->snd_wnd;
10036 	bbr->r_init_rtt = 1;
10037 
10038 	counter_u64_add(bbr_flows_nohdwr_pacing, 1);
10039 	if (bbr_allow_hdwr_pacing)
10040 		bbr->bbr_hdw_pace_ena = 1;
10041 	else
10042 		bbr->bbr_hdw_pace_ena = 0;
10043 	if (bbr_sends_full_iwnd)
10044 		bbr->bbr_init_win_cheat = 1;
10045 	else
10046 		bbr->bbr_init_win_cheat = 0;
10047 	bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max;
10048 	bbr->r_ctl.rc_drain_pg = bbr_drain_gain;
10049 	bbr->r_ctl.rc_startup_pg = bbr_high_gain;
10050 	bbr->rc_loss_exit = bbr_exit_startup_at_loss;
10051 	bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain;
10052 	bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second;
10053 	bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar;
10054 	bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max;
10055 	bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor;
10056 	bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min;
10057 	bbr->r_ctl.bbr_cross_over = bbr_cross_over;
10058 	bbr->r_ctl.rc_rtt_shrinks = cts;
10059 	if (bbr->rc_use_google) {
10060 		setup_time_filter(&bbr->r_ctl.rc_delrate,
10061 				  FILTER_TYPE_MAX,
10062 				  BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10063 		setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10064 					FILTER_TYPE_MIN, (11 * USECS_IN_SECOND));
10065 	} else {
10066 		setup_time_filter(&bbr->r_ctl.rc_delrate,
10067 				  FILTER_TYPE_MAX,
10068 				  bbr_num_pktepo_for_del_limit);
10069 		setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10070 					FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND));
10071 	}
10072 	bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0);
10073 	if (bbr_uses_idle_restart)
10074 		bbr->rc_use_idle_restart = 1;
10075 	else
10076 		bbr->rc_use_idle_restart = 0;
10077 	bbr->r_ctl.rc_bbr_cur_del_rate = 0;
10078 	bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps;
10079 	if (bbr_resends_use_tso)
10080 		bbr->rc_resends_use_tso = 1;
10081 	if (tp->snd_una != tp->snd_max) {
10082 		/* Create a send map for the current outstanding data */
10083 		struct bbr_sendmap *rsm;
10084 
10085 		rsm = bbr_alloc(bbr);
10086 		if (rsm == NULL) {
10087 			uma_zfree(bbr_pcb_zone, *ptr);
10088 			*ptr = NULL;
10089 			return (ENOMEM);
10090 		}
10091 		rsm->r_rtt_not_allowed = 1;
10092 		rsm->r_tim_lastsent[0] = cts;
10093 		rsm->r_rtr_cnt = 1;
10094 		rsm->r_rtr_bytes = 0;
10095 		rsm->r_start = tp->snd_una;
10096 		rsm->r_end = tp->snd_max;
10097 		rsm->r_dupack = 0;
10098 		rsm->r_delivered = bbr->r_ctl.rc_delivered;
10099 		rsm->r_ts_valid = 0;
10100 		rsm->r_del_ack_ts = tp->ts_recent;
10101 		rsm->r_del_time = cts;
10102 		if (bbr->r_ctl.r_app_limited_until)
10103 			rsm->r_app_limited = 1;
10104 		else
10105 			rsm->r_app_limited = 0;
10106 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
10107 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
10108 		rsm->r_in_tmap = 1;
10109 		if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
10110 			rsm->r_bbr_state = bbr_state_val(bbr);
10111 		else
10112 			rsm->r_bbr_state = 8;
10113 	}
10114 	if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0))
10115 		bbr->bbr_use_rack_cheat = 1;
10116 	if (bbr_incr_timers && (bbr->rc_use_google == 0))
10117 		bbr->r_ctl.rc_incr_tmrs = 1;
10118 	if (bbr_include_tcp_oh && (bbr->rc_use_google == 0))
10119 		bbr->r_ctl.rc_inc_tcp_oh = 1;
10120 	if (bbr_include_ip_oh && (bbr->rc_use_google == 0))
10121 		bbr->r_ctl.rc_inc_ip_oh = 1;
10122 	if (bbr_include_enet_oh && (bbr->rc_use_google == 0))
10123 		bbr->r_ctl.rc_inc_enet_oh = 1;
10124 
10125 	bbr_log_type_statechange(bbr, cts, __LINE__);
10126 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
10127 	    (tp->t_srtt)) {
10128 		uint32_t rtt;
10129 
10130 		rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
10131 		apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
10132 	}
10133 	/* announce the settings and state */
10134 	bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT);
10135 	tcp_bbr_tso_size_check(bbr, cts);
10136 	/*
10137 	 * Now call the generic function to start a timer. This will place
10138 	 * the TCB on the hptsi wheel if a timer is needed with appropriate
10139 	 * flags.
10140 	 */
10141 	bbr_stop_all_timers(tp, bbr);
10142 	/*
10143 	 * Validate the timers are not in usec, if they are convert.
10144 	 * BBR should in theory move to USEC and get rid of a
10145 	 * lot of the TICKS_2 calls.. but for now we stay
10146 	 * with tick timers.
10147 	 */
10148 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_TICKS);
10149 	TCPT_RANGESET(tp->t_rxtcur,
10150 	    ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1,
10151 	    tp->t_rttmin, tcp_rexmit_max);
10152 	bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0);
10153 	return (0);
10154 }
10155 
10156 /*
10157  * Return 0 if we can accept the connection. Return
10158  * non-zero if we can't handle the connection. A EAGAIN
10159  * means you need to wait until the connection is up.
10160  * a EADDRNOTAVAIL means we can never handle the connection
10161  * (no SACK).
10162  */
10163 static int
bbr_handoff_ok(struct tcpcb * tp)10164 bbr_handoff_ok(struct tcpcb *tp)
10165 {
10166 	if ((tp->t_state == TCPS_CLOSED) ||
10167 	    (tp->t_state == TCPS_LISTEN)) {
10168 		/* Sure no problem though it may not stick */
10169 		return (0);
10170 	}
10171 	if ((tp->t_state == TCPS_SYN_SENT) ||
10172 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
10173 		/*
10174 		 * We really don't know you have to get to ESTAB or beyond
10175 		 * to tell.
10176 		 */
10177 		return (EAGAIN);
10178 	}
10179 	if (tp->t_flags & TF_SENTFIN)
10180 		return (EINVAL);
10181 	if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) {
10182 		return (0);
10183 	}
10184 	/*
10185 	 * If we reach here we don't do SACK on this connection so we can
10186 	 * never do rack.
10187 	 */
10188 	return (EINVAL);
10189 }
10190 
10191 static void
bbr_fini(struct tcpcb * tp,int32_t tcb_is_purged)10192 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10193 {
10194 	if (tp->t_fb_ptr) {
10195 		uint32_t calc;
10196 		struct tcp_bbr *bbr;
10197 		struct bbr_sendmap *rsm;
10198 
10199 		bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10200 		if (bbr->r_ctl.crte)
10201 			tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
10202 		bbr_log_flowend(bbr);
10203 		bbr->rc_tp = NULL;
10204 		if (bbr->bbr_hdrw_pacing)
10205 			counter_u64_add(bbr_flows_whdwr_pacing, -1);
10206 		else
10207 			counter_u64_add(bbr_flows_nohdwr_pacing, -1);
10208 		if (bbr->r_ctl.crte != NULL) {
10209 			tcp_rel_pacing_rate(bbr->r_ctl.crte, tp);
10210 			bbr->r_ctl.crte = NULL;
10211 		}
10212 		rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10213 		while (rsm) {
10214 			TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
10215 			uma_zfree(bbr_zone, rsm);
10216 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10217 		}
10218 		rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10219 		while (rsm) {
10220 			TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
10221 			uma_zfree(bbr_zone, rsm);
10222 			rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10223 		}
10224 		calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd;
10225 		if (calc > (bbr->r_ctl.rc_init_rwnd / 10))
10226 			BBR_STAT_INC(bbr_dynamic_rwnd);
10227 		else
10228 			BBR_STAT_INC(bbr_static_rwnd);
10229 		bbr->r_ctl.rc_free_cnt = 0;
10230 		uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10231 		tp->t_fb_ptr = NULL;
10232 	}
10233 	/* Make sure snd_nxt is correctly set */
10234 	tp->snd_nxt = tp->snd_max;
10235 }
10236 
10237 static void
bbr_set_state(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t win)10238 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win)
10239 {
10240 	switch (tp->t_state) {
10241 	case TCPS_SYN_SENT:
10242 		bbr->r_state = TCPS_SYN_SENT;
10243 		bbr->r_substate = bbr_do_syn_sent;
10244 		break;
10245 	case TCPS_SYN_RECEIVED:
10246 		bbr->r_state = TCPS_SYN_RECEIVED;
10247 		bbr->r_substate = bbr_do_syn_recv;
10248 		break;
10249 	case TCPS_ESTABLISHED:
10250 		bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd);
10251 		bbr->r_state = TCPS_ESTABLISHED;
10252 		bbr->r_substate = bbr_do_established;
10253 		break;
10254 	case TCPS_CLOSE_WAIT:
10255 		bbr->r_state = TCPS_CLOSE_WAIT;
10256 		bbr->r_substate = bbr_do_close_wait;
10257 		break;
10258 	case TCPS_FIN_WAIT_1:
10259 		bbr->r_state = TCPS_FIN_WAIT_1;
10260 		bbr->r_substate = bbr_do_fin_wait_1;
10261 		break;
10262 	case TCPS_CLOSING:
10263 		bbr->r_state = TCPS_CLOSING;
10264 		bbr->r_substate = bbr_do_closing;
10265 		break;
10266 	case TCPS_LAST_ACK:
10267 		bbr->r_state = TCPS_LAST_ACK;
10268 		bbr->r_substate = bbr_do_lastack;
10269 		break;
10270 	case TCPS_FIN_WAIT_2:
10271 		bbr->r_state = TCPS_FIN_WAIT_2;
10272 		bbr->r_substate = bbr_do_fin_wait_2;
10273 		break;
10274 	case TCPS_LISTEN:
10275 	case TCPS_CLOSED:
10276 	case TCPS_TIME_WAIT:
10277 	default:
10278 		break;
10279 	};
10280 }
10281 
10282 static void
bbr_substate_change(struct tcp_bbr * bbr,uint32_t cts,int32_t line,int dolog)10283 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog)
10284 {
10285 	/*
10286 	 * Now what state are we going into now? Is there adjustments
10287 	 * needed?
10288 	 */
10289 	int32_t old_state;
10290 
10291 	old_state = bbr_state_val(bbr);
10292 	if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) {
10293 		/* Save the lowest srtt we saw in our end of the sub-state */
10294 		bbr->rc_hit_state_1 = 0;
10295 		if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff)
10296 			bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state;
10297 	}
10298 	bbr->rc_bbr_substate++;
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
bbr_set_probebw_google_gains(struct tcp_bbr * bbr,uint32_t cts,uint32_t losses)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
bbr_set_probebw_gains(struct tcp_bbr * bbr,uint32_t cts,uint32_t losses)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
bbr_get_a_state_target(struct tcp_bbr * bbr,uint32_t gain)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
bbr_set_state_target(struct tcp_bbr * bbr,int line)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
bbr_enter_probe_rtt(struct tcp_bbr * bbr,uint32_t cts,int32_t line)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
bbr_check_probe_rtt_limits(struct tcp_bbr * bbr,uint32_t cts)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
bbr_exit_probe_rtt(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t cts)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
bbr_should_enter_probe_rtt(struct tcp_bbr * bbr,uint32_t cts)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
bbr_google_startup(struct tcp_bbr * bbr,uint32_t cts,int32_t pkt_epoch)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
bbr_state_startup(struct tcp_bbr * bbr,uint32_t cts,int32_t epoch,int32_t pkt_epoch)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
bbr_state_change(struct tcp_bbr * bbr,uint32_t cts,int32_t epoch,int32_t pkt_epoch,uint32_t losses)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
bbr_check_bbr_for_state(struct tcp_bbr * bbr,uint32_t cts,int32_t line,uint32_t losses)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
bbr_do_segment_nounlock(struct tcpcb * tp,struct mbuf * m,struct tcphdr * th,int32_t drop_hdrlen,int32_t tlen,uint8_t iptos,int32_t nxt_pkt,struct timeval * tv)11274 bbr_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
11275     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
11276     struct timeval *tv)
11277 {
11278 	struct inpcb *inp = tptoinpcb(tp);
11279 	struct socket *so = tptosocket(tp);
11280 	int32_t thflags, retval;
11281 	uint32_t cts, lcts;
11282 	uint32_t tiwin;
11283 	struct tcpopt to;
11284 	struct tcp_bbr *bbr;
11285 	struct bbr_sendmap *rsm;
11286 	struct timeval ltv;
11287 	int32_t did_out = 0;
11288 	uint16_t nsegs;
11289 	int32_t prev_state;
11290 	uint32_t lost;
11291 
11292 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
11293 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11294 	/* add in our stats */
11295 	kern_prefetch(bbr, &prev_state);
11296 	prev_state = 0;
11297 	thflags = tcp_get_flags(th);
11298 	/*
11299 	 * If this is either a state-changing packet or current state isn't
11300 	 * established, we require a write lock on tcbinfo.  Otherwise, we
11301 	 * allow the tcbinfo to be in either alocked or unlocked, as the
11302 	 * caller may have unnecessarily acquired a write lock due to a
11303 	 * race.
11304 	 */
11305 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11306 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
11307 	    __func__));
11308 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
11309 	    __func__));
11310 
11311 	tp->t_rcvtime = ticks;
11312 	/*
11313 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
11314 	 * the scale is zero.
11315 	 */
11316 	tiwin = th->th_win << tp->snd_scale;
11317 #ifdef STATS
11318 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
11319 #endif
11320 
11321 	if (m->m_flags & M_TSTMP) {
11322 		/* Prefer the hardware timestamp if present */
11323 		struct timespec ts;
11324 
11325 		mbuf_tstmp2timespec(m, &ts);
11326 		bbr->rc_tv.tv_sec = ts.tv_sec;
11327 		bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11328 		bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usec(&bbr->rc_tv);
11329 	} else if (m->m_flags & M_TSTMP_LRO) {
11330 		/* Next the arrival timestamp */
11331 		struct timespec ts;
11332 
11333 		mbuf_tstmp2timespec(m, &ts);
11334 		bbr->rc_tv.tv_sec = ts.tv_sec;
11335 		bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11336 		bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usec(&bbr->rc_tv);
11337 	} else {
11338 		/*
11339 		 * Ok just get the current time.
11340 		 */
11341 		bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv);
11342 	}
11343 	/*
11344 	 * Parse options on any incoming segment.
11345 	 */
11346 	tcp_dooptions(&to, (u_char *)(th + 1),
11347 	    (th->th_off << 2) - sizeof(struct tcphdr),
11348 	    (thflags & TH_SYN) ? TO_SYN : 0);
11349 	if (tp->t_flags2 & TF2_PROC_SACK_PROHIBIT) {
11350 		/*
11351 		 * We don't look at sack's from the
11352 		 * peer because the MSS is too small which
11353 		 * can subject us to an attack.
11354 		 */
11355 		to.to_flags &= ~TOF_SACK;
11356 	}
11357 	/*
11358 	 * If timestamps were negotiated during SYN/ACK and a
11359 	 * segment without a timestamp is received, silently drop
11360 	 * the segment, unless it is a RST segment or missing timestamps are
11361 	 * tolerated.
11362 	 * See section 3.2 of RFC 7323.
11363 	 */
11364 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
11365 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
11366 		retval = 0;
11367 		m_freem(m);
11368 		goto done_with_input;
11369 	}
11370 	/*
11371 	 * If echoed timestamp is later than the current time, fall back to
11372 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
11373 	 * were used when this connection was established.
11374 	 */
11375 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
11376 		to.to_tsecr -= tp->ts_offset;
11377 		if (TSTMP_GT(to.to_tsecr, tcp_tv_to_msec(&bbr->rc_tv)))
11378 			to.to_tsecr = 0;
11379 	}
11380 	/*
11381 	 * If its the first time in we need to take care of options and
11382 	 * verify we can do SACK for rack!
11383 	 */
11384 	if (bbr->r_state == 0) {
11385 		/*
11386 		 * Process options only when we get SYN/ACK back. The SYN
11387 		 * case for incoming connections is handled in tcp_syncache.
11388 		 * According to RFC1323 the window field in a SYN (i.e., a
11389 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
11390 		 * this is traditional behavior, may need to be cleaned up.
11391 		 */
11392 		if (bbr->rc_inp == NULL) {
11393 			bbr->rc_inp = inp;
11394 		}
11395 		/*
11396 		 * We need to init rc_inp here since its not init'd when
11397 		 * bbr_init is called
11398 		 */
11399 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
11400 			if ((to.to_flags & TOF_SCALE) &&
11401 			    (tp->t_flags & TF_REQ_SCALE)) {
11402 				tp->t_flags |= TF_RCVD_SCALE;
11403 				tp->snd_scale = to.to_wscale;
11404 			} else
11405 				tp->t_flags &= ~TF_REQ_SCALE;
11406 			/*
11407 			 * Initial send window.  It will be updated with the
11408 			 * next incoming segment to the scaled value.
11409 			 */
11410 			tp->snd_wnd = th->th_win;
11411 			if ((to.to_flags & TOF_TS) &&
11412 			    (tp->t_flags & TF_REQ_TSTMP)) {
11413 				tp->t_flags |= TF_RCVD_TSTMP;
11414 				tp->ts_recent = to.to_tsval;
11415 				tp->ts_recent_age = tcp_tv_to_msec(&bbr->rc_tv);
11416 			} else
11417 			    tp->t_flags &= ~TF_REQ_TSTMP;
11418 			if (to.to_flags & TOF_MSS)
11419 				tcp_mss(tp, to.to_mss);
11420 			if ((tp->t_flags & TF_SACK_PERMIT) &&
11421 			    (to.to_flags & TOF_SACKPERM) == 0)
11422 				tp->t_flags &= ~TF_SACK_PERMIT;
11423 			if (tp->t_flags & TF_FASTOPEN) {
11424 				if (to.to_flags & TOF_FASTOPEN) {
11425 					uint16_t mss;
11426 
11427 					if (to.to_flags & TOF_MSS)
11428 						mss = to.to_mss;
11429 					else
11430 						if ((inp->inp_vflag & INP_IPV6) != 0)
11431 							mss = TCP6_MSS;
11432 						else
11433 							mss = TCP_MSS;
11434 					tcp_fastopen_update_cache(tp, mss,
11435 					    to.to_tfo_len, to.to_tfo_cookie);
11436 				} else
11437 					tcp_fastopen_disable_path(tp);
11438 			}
11439 		}
11440 		/*
11441 		 * At this point we are at the initial call. Here we decide
11442 		 * if we are doing RACK or not. We do this by seeing if
11443 		 * TF_SACK_PERMIT is set, if not rack is *not* possible and
11444 		 * we switch to the default code.
11445 		 */
11446 		if ((tp->t_flags & TF_SACK_PERMIT) == 0) {
11447 			/* Bail */
11448 			tcp_switch_back_to_default(tp);
11449 			(*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
11450 			    tlen, iptos);
11451 			return (1);
11452 		}
11453 		/* Set the flag */
11454 		bbr->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
11455 		tcp_set_hpts(tp);
11456 		sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack);
11457 	}
11458 	if (thflags & TH_ACK) {
11459 		/* Track ack types */
11460 		if (to.to_flags & TOF_SACK)
11461 			BBR_STAT_INC(bbr_acks_with_sacks);
11462 		else
11463 			BBR_STAT_INC(bbr_plain_acks);
11464 	}
11465 	/*
11466 	 * This is the one exception case where we set the rack state
11467 	 * always. All other times (timers etc) we must have a rack-state
11468 	 * set (so we assure we have done the checks above for SACK).
11469 	 */
11470 	if (thflags & TH_FIN)
11471 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
11472 	if (bbr->r_state != tp->t_state)
11473 		bbr_set_state(tp, bbr, tiwin);
11474 
11475 	if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL)
11476 		kern_prefetch(rsm, &prev_state);
11477 	prev_state = bbr->r_state;
11478 	bbr->rc_ack_was_delayed = 0;
11479 	lost = bbr->r_ctl.rc_lost;
11480 	bbr->rc_is_pkt_epoch_now = 0;
11481 	if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) {
11482 		/* Get the real time into lcts and figure the real delay */
11483 		lcts = tcp_get_usecs(&ltv);
11484 		if (TSTMP_GT(lcts, cts)) {
11485 			bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts;
11486 			bbr->rc_ack_was_delayed = 1;
11487 			if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay,
11488 				     bbr->r_ctl.highest_hdwr_delay))
11489 				bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay;
11490 		} else {
11491 			bbr->r_ctl.rc_ack_hdwr_delay = 0;
11492 			bbr->rc_ack_was_delayed = 0;
11493 		}
11494 	} else {
11495 		bbr->r_ctl.rc_ack_hdwr_delay = 0;
11496 		bbr->rc_ack_was_delayed = 0;
11497 	}
11498 	bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m);
11499 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
11500 		retval = 0;
11501 		m_freem(m);
11502 		goto done_with_input;
11503 	}
11504 	/*
11505 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
11506 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
11507 	 */
11508 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
11509 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
11510 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11511 		ctf_do_dropwithreset_conn(m, tp, th, tlen);
11512 		return (1);
11513 	}
11514 	if (tiwin > bbr->r_ctl.rc_high_rwnd)
11515 		bbr->r_ctl.rc_high_rwnd = tiwin;
11516 	bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp,
11517 					    (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11518 	bbr->rtt_valid = 0;
11519 	if (to.to_flags & TOF_TS) {
11520 		bbr->rc_ts_valid = 1;
11521 		bbr->r_ctl.last_inbound_ts = to.to_tsval;
11522 	} else {
11523 		bbr->rc_ts_valid = 0;
11524 		bbr->r_ctl.last_inbound_ts = 0;
11525 	}
11526 	retval = (*bbr->r_substate) (m, th, so,
11527 	    tp, &to, drop_hdrlen,
11528 	    tlen, tiwin, thflags, nxt_pkt, iptos);
11529 	if (nxt_pkt == 0)
11530 		BBR_STAT_INC(bbr_rlock_left_ret0);
11531 	else
11532 		BBR_STAT_INC(bbr_rlock_left_ret1);
11533 	if (retval == 0) {
11534 		/*
11535 		 * If retval is 1 the tcb is unlocked and most likely the tp
11536 		 * is gone.
11537 		 */
11538 		INP_WLOCK_ASSERT(inp);
11539 		tcp_bbr_xmit_timer_commit(bbr, tp, cts);
11540 		if (bbr->rc_is_pkt_epoch_now)
11541 			bbr_set_pktepoch(bbr, cts, __LINE__);
11542 		bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost));
11543 		if (nxt_pkt == 0) {
11544 			if ((bbr->r_wanted_output != 0) ||
11545 			    (tp->t_flags & TF_ACKNOW)) {
11546 
11547 				bbr->rc_output_starts_timer = 0;
11548 				did_out = 1;
11549 				if (tcp_output(tp) < 0)
11550 					return (1);
11551 			} else
11552 				bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0);
11553 		}
11554 		if ((nxt_pkt == 0) &&
11555 		    ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11556 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
11557 		     (tp->t_flags & TF_DELACK) ||
11558 		     ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11559 		      (tp->t_state <= TCPS_CLOSING)))) {
11560 			/*
11561 			 * We could not send (probably in the hpts but
11562 			 * stopped the timer)?
11563 			 */
11564 			if ((tp->snd_max == tp->snd_una) &&
11565 			    ((tp->t_flags & TF_DELACK) == 0) &&
11566 			    (tcp_in_hpts(tp)) &&
11567 			    (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11568 				/*
11569 				 * keep alive not needed if we are hptsi
11570 				 * output yet
11571 				 */
11572 				;
11573 			} else {
11574 				if (tcp_in_hpts(tp)) {
11575 					tcp_hpts_remove(tp);
11576 					if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11577 					    (TSTMP_GT(lcts, bbr->rc_pacer_started))) {
11578 						uint32_t del;
11579 
11580 						del = lcts - bbr->rc_pacer_started;
11581 						if (bbr->r_ctl.rc_last_delay_val > del) {
11582 							BBR_STAT_INC(bbr_force_timer_start);
11583 							bbr->r_ctl.rc_last_delay_val -= del;
11584 							bbr->rc_pacer_started = lcts;
11585 						} else {
11586 							/* We are late */
11587 							bbr->r_ctl.rc_last_delay_val = 0;
11588 							BBR_STAT_INC(bbr_force_output);
11589 							if (tcp_output(tp) < 0)
11590 								return (1);
11591 						}
11592 					}
11593 				}
11594 				bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val,
11595 				    0);
11596 			}
11597 		} else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) {
11598 			/* Do we have the correct timer running? */
11599 			bbr_timer_audit(tp, bbr, lcts, &so->so_snd);
11600 		}
11601 		/* Clear the flag, it may have been cleared by output but we may not have  */
11602 		if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
11603 			tp->t_flags2 &= ~TF2_HPTS_CALLS;
11604 		/* Do we have a new state */
11605 		if (bbr->r_state != tp->t_state)
11606 			bbr_set_state(tp, bbr, tiwin);
11607 done_with_input:
11608 		bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out);
11609 		if (did_out)
11610 			bbr->r_wanted_output = 0;
11611 	}
11612 	return (retval);
11613 }
11614 
11615 static void
bbr_do_segment(struct tcpcb * tp,struct mbuf * m,struct tcphdr * th,int32_t drop_hdrlen,int32_t tlen,uint8_t iptos)11616 bbr_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
11617     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11618 {
11619 	struct timeval tv;
11620 	int retval;
11621 
11622 	/* First lets see if we have old packets */
11623 	if (!STAILQ_EMPTY(&tp->t_inqueue)) {
11624 		if (ctf_do_queued_segments(tp, 1)) {
11625 			m_freem(m);
11626 			return;
11627 		}
11628 	}
11629 	if (m->m_flags & M_TSTMP_LRO) {
11630 		mbuf_tstmp2timeval(m, &tv);
11631 	} else {
11632 		/* Should not be should we kassert instead? */
11633 		tcp_get_usecs(&tv);
11634 	}
11635 	retval = bbr_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos,
11636 	    0, &tv);
11637 	if (retval == 0) {
11638 		INP_WUNLOCK(tptoinpcb(tp));
11639 	}
11640 }
11641 
11642 /*
11643  * Return how much data can be sent without violating the
11644  * cwnd or rwnd.
11645  */
11646 
11647 static inline uint32_t
bbr_what_can_we_send(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t sendwin,uint32_t avail,int32_t sb_offset,uint32_t cts)11648 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin,
11649     uint32_t avail, int32_t sb_offset, uint32_t cts)
11650 {
11651 	uint32_t len;
11652 
11653 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
11654 		/* We never want to go over our peers rcv-window */
11655 		len = 0;
11656 	} else {
11657 		uint32_t flight;
11658 
11659 		flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11660 		if (flight >= sendwin) {
11661 			/*
11662 			 * We have in flight what we are allowed by cwnd (if
11663 			 * it was rwnd blocking it would have hit above out
11664 			 * >= tp->snd_wnd).
11665 			 */
11666 			return (0);
11667 		}
11668 		len = sendwin - flight;
11669 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11670 			/* We would send too much (beyond the rwnd) */
11671 			len = tp->snd_wnd - ctf_outstanding(tp);
11672 		}
11673 		if ((len + sb_offset) > avail) {
11674 			/*
11675 			 * We don't have that much in the SB, how much is
11676 			 * there?
11677 			 */
11678 			len = avail - sb_offset;
11679 		}
11680 	}
11681 	return (len);
11682 }
11683 
11684 static inline void
bbr_do_send_accounting(struct tcpcb * tp,struct tcp_bbr * bbr,struct bbr_sendmap * rsm,int32_t len,int32_t error)11685 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11686 {
11687 	if (error) {
11688 		return;
11689 	}
11690 	if (rsm) {
11691 		if (rsm->r_flags & BBR_TLP) {
11692 			/*
11693 			 * TLP should not count in retran count, but in its
11694 			 * own bin
11695 			 */
11696 			KMOD_TCPSTAT_INC(tcps_tlpresends);
11697 			KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len);
11698 		} else {
11699 			/* Retransmit */
11700 			tp->t_sndrexmitpack++;
11701 			KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
11702 			KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
11703 #ifdef STATS
11704 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
11705 			    len);
11706 #endif
11707 		}
11708 		/*
11709 		 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is
11710 		 * sub-state
11711 		 */
11712 		counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len);
11713 		if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) {
11714 			/* Non probe_bw log in 1, 2, or 4. */
11715 			counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len);
11716 		} else {
11717 			/*
11718 			 * Log our probe state 3, and log also 5-13 to show
11719 			 * us the recovery sub-state for the send. This
11720 			 * means that 3 == (5+6+7+8+9+10+11+12+13)
11721 			 */
11722 			counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len);
11723 			counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len);
11724 		}
11725 		/* Place in both 16's the totals of retransmitted */
11726 		counter_u64_add(bbr_state_lost[16], len);
11727 		counter_u64_add(bbr_state_resend[16], len);
11728 		/* Place in 17's the total sent */
11729 		counter_u64_add(bbr_state_resend[17], len);
11730 		counter_u64_add(bbr_state_lost[17], len);
11731 
11732 	} else {
11733 		/* New sends */
11734 		KMOD_TCPSTAT_INC(tcps_sndpack);
11735 		KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
11736 		/* Place in 17's the total sent */
11737 		counter_u64_add(bbr_state_resend[17], len);
11738 		counter_u64_add(bbr_state_lost[17], len);
11739 #ifdef STATS
11740 		stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
11741 		    len);
11742 #endif
11743 	}
11744 }
11745 
11746 static void
bbr_cwnd_limiting(struct tcpcb * tp,struct tcp_bbr * bbr,uint32_t in_level)11747 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level)
11748 {
11749 	if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) {
11750 		/*
11751 		 * Limit the cwnd to not be above N x the target plus whats
11752 		 * is outstanding. The target is based on the current b/w
11753 		 * estimate.
11754 		 */
11755 		uint32_t target;
11756 
11757 		target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT);
11758 		target += ctf_outstanding(tp);
11759 		target *= bbr_target_cwnd_mult_limit;
11760 		if (tp->snd_cwnd > target)
11761 			tp->snd_cwnd = target;
11762 		bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__);
11763 	}
11764 }
11765 
11766 static int
bbr_window_update_needed(struct tcpcb * tp,struct socket * so,uint32_t recwin,int32_t maxseg)11767 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg)
11768 {
11769 	/*
11770 	 * "adv" is the amount we could increase the window, taking into
11771 	 * account that we are limited by TCP_MAXWIN << tp->rcv_scale.
11772 	 */
11773 	int32_t adv;
11774 	int32_t oldwin;
11775 
11776 	adv = recwin;
11777 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
11778 		oldwin = (tp->rcv_adv - tp->rcv_nxt);
11779 		if (adv > oldwin)
11780 			adv -= oldwin;
11781 		else {
11782 			/* We can't increase the window */
11783 			adv = 0;
11784 		}
11785 	} else
11786 		oldwin = 0;
11787 
11788 	/*
11789 	 * If the new window size ends up being the same as or less
11790 	 * than the old size when it is scaled, then don't force
11791 	 * a window update.
11792 	 */
11793 	if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
11794 		return (0);
11795 
11796 	if (adv >= (2 * maxseg) &&
11797 	    (adv >= (so->so_rcv.sb_hiwat / 4) ||
11798 	    recwin <= (so->so_rcv.sb_hiwat / 8) ||
11799 	    so->so_rcv.sb_hiwat <= 8 * maxseg)) {
11800 		return (1);
11801 	}
11802 	if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat)
11803 		return (1);
11804 	return (0);
11805 }
11806 
11807 /*
11808  * Return 0 on success and a errno on failure to send.
11809  * Note that a 0 return may not mean we sent anything
11810  * if the TCB was on the hpts. A non-zero return
11811  * does indicate the error we got from ip[6]_output.
11812  */
11813 static int
bbr_output_wtime(struct tcpcb * tp,const struct timeval * tv)11814 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv)
11815 {
11816 	struct socket *so;
11817 	int32_t len;
11818 	uint32_t cts;
11819 	uint32_t recwin, sendwin;
11820 	int32_t sb_offset;
11821 	int32_t flags, abandon, error = 0;
11822 	struct tcp_log_buffer *lgb;
11823 	struct mbuf *m;
11824 	struct mbuf *mb;
11825 	uint32_t if_hw_tsomaxsegcount = 0;
11826 	uint32_t if_hw_tsomaxsegsize = 0;
11827 	uint32_t if_hw_tsomax = 0;
11828 	struct ip *ip = NULL;
11829 	struct tcp_bbr *bbr;
11830 	struct tcphdr *th;
11831 	struct udphdr *udp = NULL;
11832 	u_char opt[TCP_MAXOLEN];
11833 	unsigned ipoptlen, optlen, hdrlen;
11834 	unsigned ulen;
11835 	uint32_t bbr_seq;
11836 	uint32_t delay_calc=0;
11837 	uint8_t doing_tlp = 0;
11838 	uint8_t local_options;
11839 #ifdef BBR_INVARIANTS
11840 	uint8_t doing_retran_from = 0;
11841 	uint8_t picked_up_retran = 0;
11842 #endif
11843 	uint8_t wanted_cookie = 0;
11844 	uint8_t more_to_rxt=0;
11845 	int32_t prefetch_so_done = 0;
11846 	int32_t prefetch_rsm = 0;
11847 	uint32_t tot_len = 0;
11848 	uint32_t maxseg, pace_max_segs, p_maxseg;
11849 	int32_t csum_flags = 0;
11850  	int32_t hw_tls;
11851 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
11852 	unsigned ipsec_optlen = 0;
11853 
11854 #endif
11855 	volatile int32_t sack_rxmit;
11856 	struct bbr_sendmap *rsm = NULL;
11857 	int32_t tso, mtu;
11858 	struct tcpopt to;
11859 	int32_t slot = 0;
11860 	struct inpcb *inp;
11861 	struct sockbuf *sb;
11862 	bool hpts_calling;
11863 #ifdef INET6
11864 	struct ip6_hdr *ip6 = NULL;
11865 	int32_t isipv6;
11866 #endif
11867 	uint8_t app_limited = BBR_JR_SENT_DATA;
11868 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11869 	/* We take a cache hit here */
11870 	memcpy(&bbr->rc_tv, tv, sizeof(struct timeval));
11871 	cts = tcp_tv_to_usec(&bbr->rc_tv);
11872 	inp = bbr->rc_inp;
11873 	hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
11874 	tp->t_flags2 &= ~TF2_HPTS_CALLS;
11875 	so = inp->inp_socket;
11876 	sb = &so->so_snd;
11877 	if (tp->t_nic_ktls_xmit)
11878  		hw_tls = 1;
11879  	else
11880  		hw_tls = 0;
11881 	kern_prefetch(sb, &maxseg);
11882 	maxseg = tp->t_maxseg - bbr->rc_last_options;
11883 	if (bbr_minseg(bbr) < maxseg) {
11884 		tcp_bbr_tso_size_check(bbr, cts);
11885 	}
11886 	/* Remove any flags that indicate we are pacing on the inp  */
11887 	pace_max_segs = bbr->r_ctl.rc_pace_max_segs;
11888 	p_maxseg = min(maxseg, pace_max_segs);
11889 	INP_WLOCK_ASSERT(inp);
11890 #ifdef TCP_OFFLOAD
11891 	if (tp->t_flags & TF_TOE)
11892 		return (tcp_offload_output(tp));
11893 #endif
11894 
11895 #ifdef INET6
11896 	if (bbr->r_state) {
11897 		/* Use the cache line loaded if possible */
11898 		isipv6 = bbr->r_is_v6;
11899 	} else {
11900 		isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
11901 	}
11902 #endif
11903 	if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
11904 	    tcp_in_hpts(tp)) {
11905 		/*
11906 		 * We are on the hpts for some timer but not hptsi output.
11907 		 * Possibly remove from the hpts so we can send/recv etc.
11908 		 */
11909 		if ((tp->t_flags & TF_ACKNOW) == 0) {
11910 			/*
11911 			 * No immediate demand right now to send an ack, but
11912 			 * the user may have read, making room for new data
11913 			 * (a window update). If so we may want to cancel
11914 			 * whatever timer is running (KEEP/DEL-ACK?) and
11915 			 * continue to send out a window update. Or we may
11916 			 * have gotten more data into the socket buffer to
11917 			 * send.
11918 			 */
11919 			recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
11920 				      (long)TCP_MAXWIN << tp->rcv_scale);
11921 			if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) &&
11922 			    ((tcp_outflags[tp->t_state] & TH_RST) == 0) &&
11923 			    ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <=
11924 			    (tp->snd_max - tp->snd_una))) {
11925 				/*
11926 				 * Nothing new to send and no window update
11927 				 * is needed to send. Lets just return and
11928 				 * let the timer-run off.
11929 				 */
11930 				return (0);
11931 			}
11932 		}
11933 		tcp_hpts_remove(tp);
11934 		bbr_timer_cancel(bbr, __LINE__, cts);
11935 	}
11936 	if (bbr->r_ctl.rc_last_delay_val) {
11937 		/* Calculate a rough delay for early escape to sending  */
11938 		if (SEQ_GT(cts, bbr->rc_pacer_started))
11939 			delay_calc = cts - bbr->rc_pacer_started;
11940 		if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
11941 			delay_calc -= bbr->r_ctl.rc_last_delay_val;
11942 		else
11943 			delay_calc = 0;
11944 	}
11945 	/* Mark that we have called bbr_output(). */
11946 	if ((bbr->r_timer_override) ||
11947 	    (tp->t_state < TCPS_ESTABLISHED)) {
11948 		/* Timeouts or early states are exempt */
11949 		if (tcp_in_hpts(tp))
11950 			tcp_hpts_remove(tp);
11951 	} else if (tcp_in_hpts(tp)) {
11952 		if ((bbr->r_ctl.rc_last_delay_val) &&
11953 		    (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11954 		    delay_calc) {
11955 			/*
11956 			 * We were being paced for output and the delay has
11957 			 * already exceeded when we were supposed to be
11958 			 * called, lets go ahead and pull out of the hpts
11959 			 * and call output.
11960 			 */
11961 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1);
11962 			bbr->r_ctl.rc_last_delay_val = 0;
11963 			tcp_hpts_remove(tp);
11964 		} else if (tp->t_state == TCPS_CLOSED) {
11965 			bbr->r_ctl.rc_last_delay_val = 0;
11966 			tcp_hpts_remove(tp);
11967 		} else {
11968 			/*
11969 			 * On the hpts, you shall not pass! even if ACKNOW
11970 			 * is on, we will when the hpts fires, unless of
11971 			 * course we are overdue.
11972 			 */
11973 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1);
11974 			return (0);
11975 		}
11976 	}
11977 	bbr->rc_cwnd_limited = 0;
11978 	if (bbr->r_ctl.rc_last_delay_val) {
11979 		/* recalculate the real delay and deal with over/under  */
11980 		if (SEQ_GT(cts, bbr->rc_pacer_started))
11981 			delay_calc = cts - bbr->rc_pacer_started;
11982 		else
11983 			delay_calc = 0;
11984 		if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
11985 			/* Setup the delay which will be added in */
11986 			delay_calc -= bbr->r_ctl.rc_last_delay_val;
11987 		else {
11988 			/*
11989 			 * We are early setup to adjust
11990 			 * our slot time.
11991 			 */
11992 			uint64_t merged_val;
11993 
11994 			bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc);
11995 			bbr->r_agg_early_set = 1;
11996 			if (bbr->r_ctl.rc_hptsi_agg_delay) {
11997 				if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) {
11998 					/* Nope our previous late cancels out the early */
11999 					bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early;
12000 					bbr->r_agg_early_set = 0;
12001 					bbr->r_ctl.rc_agg_early = 0;
12002 				} else {
12003 					bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay;
12004 					bbr->r_ctl.rc_hptsi_agg_delay = 0;
12005 				}
12006 			}
12007 			merged_val = bbr->rc_pacer_started;
12008 			merged_val <<= 32;
12009 			merged_val |= bbr->r_ctl.rc_last_delay_val;
12010 			bbr_log_pacing_delay_calc(bbr, hpts_calling,
12011 						 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val,
12012 						 bbr->r_agg_early_set, 3);
12013 			bbr->r_ctl.rc_last_delay_val = 0;
12014 			BBR_STAT_INC(bbr_early);
12015 			delay_calc = 0;
12016 		}
12017 	} else {
12018 		/* We were not delayed due to hptsi */
12019 		if (bbr->r_agg_early_set)
12020 			bbr->r_ctl.rc_agg_early = 0;
12021 		bbr->r_agg_early_set = 0;
12022 		delay_calc = 0;
12023 	}
12024 	if (delay_calc) {
12025 		/*
12026 		 * We had a hptsi delay which means we are falling behind on
12027 		 * sending at the expected rate. Calculate an extra amount
12028 		 * of data we can send, if any, to put us back on track.
12029 		 */
12030 		if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay)
12031 			bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff;
12032 		else
12033 			bbr->r_ctl.rc_hptsi_agg_delay += delay_calc;
12034 	}
12035 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12036 	if ((tp->snd_una == tp->snd_max) &&
12037 	    (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
12038 	    (sbavail(sb))) {
12039 		/*
12040 		 * Ok we have been idle with nothing outstanding
12041 		 * we possibly need to start fresh with either a new
12042 		 * suite of states or a fast-ramp up.
12043 		 */
12044 		bbr_restart_after_idle(bbr,
12045 				       cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time));
12046 	}
12047 	/*
12048 	 * Now was there a hptsi delay where we are behind? We only count
12049 	 * being behind if: a) We are not in recovery. b) There was a delay.
12050 	 * <and> c) We had room to send something.
12051 	 *
12052 	 */
12053 	if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
12054 		int retval;
12055 
12056 		retval = bbr_process_timers(tp, bbr, cts, hpts_calling);
12057 		if (retval != 0) {
12058 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1);
12059 			/*
12060 			 * If timers want tcp_drop(), then pass error out,
12061 			 * otherwise suppress it.
12062 			 */
12063 			return (retval < 0 ? retval : 0);
12064 		}
12065 	}
12066 	bbr->rc_tp->t_flags2 &= ~TF2_MBUF_QUEUE_READY;
12067 	if (hpts_calling &&
12068 	    (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12069 		bbr->r_ctl.rc_last_delay_val = 0;
12070 	}
12071 	bbr->r_timer_override = 0;
12072 	bbr->r_wanted_output = 0;
12073 	/*
12074 	 * For TFO connections in SYN_RECEIVED, only allow the initial
12075 	 * SYN|ACK and those sent by the retransmit timer.
12076 	 */
12077 	if ((tp->t_flags & TF_FASTOPEN) &&
12078 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
12079 	     (tp->t_state == TCPS_SYN_SENT)) &&
12080 	    SEQ_GT(tp->snd_max, tp->snd_una) &&	/* initial SYN or SYN|ACK sent */
12081 	    (tp->t_rxtshift == 0)) {	/* not a retransmit */
12082 		len = 0;
12083 		goto just_return_nolock;
12084 	}
12085 	/*
12086 	 * Before sending anything check for a state update. For hpts
12087 	 * calling without input this is important. If its input calling
12088 	 * then this was already done.
12089 	 */
12090 	if (bbr->rc_use_google == 0)
12091 		bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12092 again:
12093 	/*
12094 	 * If we've recently taken a timeout, snd_max will be greater than
12095 	 * snd_max. BBR in general does not pay much attention to snd_nxt
12096 	 * for historic reasons the persist timer still uses it. This means
12097 	 * we have to look at it. All retransmissions that are not persits
12098 	 * use the rsm that needs to be sent so snd_nxt is ignored. At the
12099 	 * end of this routine we pull snd_nxt always up to snd_max.
12100 	 */
12101 	doing_tlp = 0;
12102 #ifdef BBR_INVARIANTS
12103 	doing_retran_from = picked_up_retran = 0;
12104 #endif
12105 	error = 0;
12106 	tso = 0;
12107 	slot = 0;
12108 	mtu = 0;
12109 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12110 	sb_offset = tp->snd_max - tp->snd_una;
12111 	flags = tcp_outflags[tp->t_state];
12112 	sack_rxmit = 0;
12113 	len = 0;
12114 	rsm = NULL;
12115 	if (flags & TH_RST) {
12116 		SOCK_SENDBUF_LOCK(so);
12117 		goto send;
12118 	}
12119 recheck_resend:
12120 	while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
12121 		/* We need to always have one in reserve */
12122 		rsm = bbr_alloc(bbr);
12123 		if (rsm == NULL) {
12124 			error = ENOMEM;
12125 			/* Lie to get on the hpts */
12126 			tot_len = tp->t_maxseg;
12127 			if (hpts_calling)
12128 				/* Retry in a ms */
12129 				slot = 1001;
12130 			goto just_return_nolock;
12131 		}
12132 		TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
12133 		bbr->r_ctl.rc_free_cnt++;
12134 		rsm = NULL;
12135 	}
12136 	/* What do we send, a resend? */
12137 	if (bbr->r_ctl.rc_resend == NULL) {
12138 		/* Check for rack timeout */
12139 		bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
12140 		if (bbr->r_ctl.rc_resend) {
12141 #ifdef BBR_INVARIANTS
12142 			picked_up_retran = 1;
12143 #endif
12144 			bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend);
12145 		}
12146 	}
12147 	if (bbr->r_ctl.rc_resend) {
12148 		rsm = bbr->r_ctl.rc_resend;
12149 #ifdef BBR_INVARIANTS
12150 		doing_retran_from = 1;
12151 #endif
12152 		/* Remove any TLP flags its a RACK or T-O */
12153 		rsm->r_flags &= ~BBR_TLP;
12154 		bbr->r_ctl.rc_resend = NULL;
12155 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12156 #ifdef BBR_INVARIANTS
12157 			panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n",
12158 			    tp, bbr, rsm, rsm->r_start, tp->snd_una);
12159 			goto recheck_resend;
12160 #else
12161 			/* TSNH */
12162 			rsm = NULL;
12163 			goto recheck_resend;
12164 #endif
12165 		}
12166 		if (rsm->r_flags & BBR_HAS_SYN) {
12167 			/* Only retransmit a SYN by itself */
12168 			len = 0;
12169 			if ((flags & TH_SYN) == 0) {
12170 				/* Huh something is wrong */
12171 				rsm->r_start++;
12172 				if (rsm->r_start == rsm->r_end) {
12173 					/* Clean it up, somehow we missed the ack? */
12174 					bbr_log_syn(tp, NULL);
12175 				} else {
12176 					/* TFO with data? */
12177 					rsm->r_flags &= ~BBR_HAS_SYN;
12178 					len = rsm->r_end - rsm->r_start;
12179 				}
12180 			} else {
12181 				/* Retransmitting SYN */
12182 				rsm = NULL;
12183 				SOCK_SENDBUF_LOCK(so);
12184 				goto send;
12185 			}
12186 		} else
12187 			len = rsm->r_end - rsm->r_start;
12188 		if ((bbr->rc_resends_use_tso == 0) &&
12189 		    (len > maxseg)) {
12190 			len = maxseg;
12191 			more_to_rxt = 1;
12192 		}
12193 		sb_offset = rsm->r_start - tp->snd_una;
12194 		if (len > 0) {
12195 			sack_rxmit = 1;
12196 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
12197 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12198 			    min(len, maxseg));
12199 		} else {
12200 			/* I dont think this can happen */
12201 			rsm = NULL;
12202 			goto recheck_resend;
12203 		}
12204 		BBR_STAT_INC(bbr_resends_set);
12205 	} else if (bbr->r_ctl.rc_tlp_send) {
12206 		/*
12207 		 * Tail loss probe
12208 		 */
12209 		doing_tlp = 1;
12210 		rsm = bbr->r_ctl.rc_tlp_send;
12211 		bbr->r_ctl.rc_tlp_send = NULL;
12212 		sack_rxmit = 1;
12213 		len = rsm->r_end - rsm->r_start;
12214 		if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12215 			len = maxseg;
12216 
12217 		if (SEQ_GT(tp->snd_una, rsm->r_start)) {
12218 #ifdef BBR_INVARIANTS
12219 			panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u",
12220 			    tp, bbr, tp->snd_una, rsm, rsm->r_start);
12221 #else
12222 			/* TSNH */
12223 			rsm = NULL;
12224 			goto recheck_resend;
12225 #endif
12226 		}
12227 		sb_offset = rsm->r_start - tp->snd_una;
12228 		BBR_STAT_INC(bbr_tlp_set);
12229 	}
12230 	/*
12231 	 * Enforce a connection sendmap count limit if set
12232 	 * as long as we are not retransmiting.
12233 	 */
12234 	if ((rsm == NULL) &&
12235 	    (V_tcp_map_entries_limit > 0) &&
12236 	    (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
12237 		BBR_STAT_INC(bbr_alloc_limited);
12238 		if (!bbr->alloc_limit_reported) {
12239 			bbr->alloc_limit_reported = 1;
12240 			BBR_STAT_INC(bbr_alloc_limited_conns);
12241 		}
12242 		goto just_return_nolock;
12243 	}
12244 #ifdef BBR_INVARIANTS
12245 	if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) {
12246 		panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u",
12247 		    tp, bbr, rsm, sb_offset, len);
12248 	}
12249 #endif
12250 	/*
12251 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
12252 	 * state flags.
12253 	 */
12254 	if (tp->t_flags & TF_NEEDFIN && (rsm == NULL))
12255 		flags |= TH_FIN;
12256 	if (tp->t_flags & TF_NEEDSYN)
12257 		flags |= TH_SYN;
12258 
12259 	if (rsm && (rsm->r_flags & BBR_HAS_FIN)) {
12260 		/* we are retransmitting the fin */
12261 		len--;
12262 		if (len) {
12263 			/*
12264 			 * When retransmitting data do *not* include the
12265 			 * FIN. This could happen from a TLP probe if we
12266 			 * allowed data with a FIN.
12267 			 */
12268 			flags &= ~TH_FIN;
12269 		}
12270 	} else if (rsm) {
12271 		if (flags & TH_FIN)
12272 			flags &= ~TH_FIN;
12273 	}
12274 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12275 		void *end_rsm;
12276 
12277 		end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
12278 		if (end_rsm)
12279 			kern_prefetch(end_rsm, &prefetch_rsm);
12280 		prefetch_rsm = 1;
12281 	}
12282 	SOCK_SENDBUF_LOCK(so);
12283 	/*
12284 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
12285 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12286 	 * negative length.  This can also occur when TCP opens up its
12287 	 * congestion window while receiving additional duplicate acks after
12288 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
12289 	 * the fast-retransmit.
12290 	 *
12291 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
12292 	 * set to snd_una, the sb_offset will be 0, and the length may wind
12293 	 * up 0.
12294 	 *
12295 	 * If sack_rxmit is true we are retransmitting from the scoreboard
12296 	 * in which case len is already set.
12297 	 */
12298 	if (sack_rxmit == 0) {
12299 		uint32_t avail;
12300 
12301 		avail = sbavail(sb);
12302 		if (SEQ_GT(tp->snd_max, tp->snd_una))
12303 			sb_offset = tp->snd_max - tp->snd_una;
12304 		else
12305 			sb_offset = 0;
12306 		if (bbr->rc_tlp_new_data) {
12307 			/* TLP is forcing out new data */
12308 			uint32_t tlplen;
12309 
12310 			doing_tlp = 1;
12311 			tlplen = maxseg;
12312 
12313 			if (tlplen > (uint32_t)(avail - sb_offset)) {
12314 				tlplen = (uint32_t)(avail - sb_offset);
12315 			}
12316 			if (tlplen > tp->snd_wnd) {
12317 				len = tp->snd_wnd;
12318 			} else {
12319 				len = tlplen;
12320 			}
12321 			bbr->rc_tlp_new_data = 0;
12322 		} else {
12323 			len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts);
12324 			if ((len < p_maxseg) &&
12325 			    (bbr->rc_in_persist == 0) &&
12326 			    (ctf_outstanding(tp) >= (2 * p_maxseg)) &&
12327 			    ((avail - sb_offset) >= p_maxseg)) {
12328 				/*
12329 				 * We are not completing whats in the socket
12330 				 * buffer (i.e. there is at least a segment
12331 				 * waiting to send) and we have 2 or more
12332 				 * segments outstanding. There is no sense
12333 				 * of sending a little piece. Lets defer and
12334 				 * and wait until we can send a whole
12335 				 * segment.
12336 				 */
12337 				len = 0;
12338 			}
12339 			if (bbr->rc_in_persist) {
12340 				/*
12341 				 * We are in persists, figure out if
12342 				 * a retransmit is available (maybe the previous
12343 				 * persists we sent) or if we have to send new
12344 				 * data.
12345 				 */
12346 				rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
12347 				if (rsm) {
12348 					len = rsm->r_end - rsm->r_start;
12349 					if (rsm->r_flags & BBR_HAS_FIN)
12350 						len--;
12351 					if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12352 						len = maxseg;
12353 					if (len > 1)
12354 						BBR_STAT_INC(bbr_persist_reneg);
12355 					/*
12356 					 * XXXrrs we could force the len to
12357 					 * 1 byte here to cause the chunk to
12358 					 * split apart.. but that would then
12359 					 * mean we always retransmit it as
12360 					 * one byte even after the window
12361 					 * opens.
12362 					 */
12363 					sack_rxmit = 1;
12364 					sb_offset = rsm->r_start - tp->snd_una;
12365 				} else {
12366 					/*
12367 					 * First time through in persists or peer
12368 					 * acked our one byte. Though we do have
12369 					 * to have something in the sb.
12370 					 */
12371 					len = 1;
12372 					sb_offset = 0;
12373 					if (avail == 0)
12374 					    len = 0;
12375 				}
12376 			}
12377 		}
12378 	}
12379 	if (prefetch_so_done == 0) {
12380 		kern_prefetch(so, &prefetch_so_done);
12381 		prefetch_so_done = 1;
12382 	}
12383 	/*
12384 	 * Lop off SYN bit if it has already been sent.  However, if this is
12385 	 * SYN-SENT state and if segment contains data and if we don't know
12386 	 * that foreign host supports TAO, suppress sending segment.
12387 	 */
12388 	if ((flags & TH_SYN) && (rsm == NULL) &&
12389 	    SEQ_GT(tp->snd_max, tp->snd_una)) {
12390 		if (tp->t_state != TCPS_SYN_RECEIVED)
12391 			flags &= ~TH_SYN;
12392 		/*
12393 		 * When sending additional segments following a TFO SYN|ACK,
12394 		 * do not include the SYN bit.
12395 		 */
12396 		if ((tp->t_flags & TF_FASTOPEN) &&
12397 		    (tp->t_state == TCPS_SYN_RECEIVED))
12398 			flags &= ~TH_SYN;
12399 		sb_offset--, len++;
12400 		if (sbavail(sb) == 0)
12401 			len = 0;
12402 	} else if ((flags & TH_SYN) && rsm) {
12403 		/*
12404 		 * Subtract one from the len for the SYN being
12405 		 * retransmitted.
12406 		 */
12407 		len--;
12408 	}
12409 	/*
12410 	 * Be careful not to send data and/or FIN on SYN segments. This
12411 	 * measure is needed to prevent interoperability problems with not
12412 	 * fully conformant TCP implementations.
12413 	 */
12414 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12415 		len = 0;
12416 		flags &= ~TH_FIN;
12417 	}
12418 	/*
12419 	 * On TFO sockets, ensure no data is sent in the following cases:
12420 	 *
12421 	 *  - When retransmitting SYN|ACK on a passively-created socket
12422 	 *  - When retransmitting SYN on an actively created socket
12423 	 *  - When sending a zero-length cookie (cookie request) on an
12424 	 *    actively created socket
12425 	 *  - When the socket is in the CLOSED state (RST is being sent)
12426 	 */
12427 	if ((tp->t_flags & TF_FASTOPEN) &&
12428 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12429 	     ((tp->t_state == TCPS_SYN_SENT) &&
12430 	      (tp->t_tfo_client_cookie_len == 0)) ||
12431 	     (flags & TH_RST))) {
12432 		len = 0;
12433 		sack_rxmit = 0;
12434 		rsm = NULL;
12435 	}
12436 	/* Without fast-open there should never be data sent on a SYN */
12437 	if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN))
12438 		len = 0;
12439 	if (len <= 0) {
12440 		/*
12441 		 * If FIN has been sent but not acked, but we haven't been
12442 		 * called to retransmit, len will be < 0.  Otherwise, window
12443 		 * shrank after we sent into it.  If window shrank to 0,
12444 		 * cancel pending retransmit, pull snd_nxt back to (closed)
12445 		 * window, and set the persist timer if it isn't already
12446 		 * going.  If the window didn't close completely, just wait
12447 		 * for an ACK.
12448 		 *
12449 		 * We also do a general check here to ensure that we will
12450 		 * set the persist timer when we have data to send, but a
12451 		 * 0-byte window. This makes sure the persist timer is set
12452 		 * even if the packet hits one of the "goto send" lines
12453 		 * below.
12454 		 */
12455 		len = 0;
12456 		if ((tp->snd_wnd == 0) &&
12457 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12458 		    (tp->snd_una == tp->snd_max) &&
12459 		    (sb_offset < (int)sbavail(sb))) {
12460 			/*
12461 			 * Not enough room in the rwnd to send
12462 			 * a paced segment out.
12463 			 */
12464 			bbr_enter_persist(tp, bbr, cts, __LINE__);
12465 		}
12466 	} else if ((rsm == NULL) &&
12467 		   (doing_tlp == 0) &&
12468 		   (len < bbr->r_ctl.rc_pace_max_segs)) {
12469 		/*
12470 		 * We are not sending a full segment for
12471 		 * some reason. Should we not send anything (think
12472 		 * sws or persists)?
12473 		 */
12474 		if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12475 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12476 		    (len < (int)(sbavail(sb) - sb_offset))) {
12477 			/*
12478 			 * Here the rwnd is less than
12479 			 * the pacing size, this is not a retransmit,
12480 			 * we are established and
12481 			 * the send is not the last in the socket buffer
12482 			 * lets not send, and possibly enter persists.
12483 			 */
12484 			len = 0;
12485 			if (tp->snd_max == tp->snd_una)
12486 				bbr_enter_persist(tp, bbr, cts, __LINE__);
12487 		} else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) &&
12488 			   (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12489 						 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12490 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12491 			   (len < bbr_minseg(bbr))) {
12492 			/*
12493 			 * Here we are not retransmitting, and
12494 			 * the cwnd is not so small that we could
12495 			 * not send at least a min size (rxt timer
12496 			 * not having gone off), We have 2 segments or
12497 			 * more already in flight, its not the tail end
12498 			 * of the socket buffer  and the cwnd is blocking
12499 			 * us from sending out minimum pacing segment size.
12500 			 * Lets not send anything.
12501 			 */
12502 			bbr->rc_cwnd_limited = 1;
12503 			len = 0;
12504 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12505 			    min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12506 			   (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12507 						 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12508 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12509 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
12510 			/*
12511 			 * Here we have a send window but we have
12512 			 * filled it up and we can't send another pacing segment.
12513 			 * We also have in flight more than 2 segments
12514 			 * and we are not completing the sb i.e. we allow
12515 			 * the last bytes of the sb to go out even if
12516 			 * its not a full pacing segment.
12517 			 */
12518 			len = 0;
12519 		}
12520 	}
12521 	/* len will be >= 0 after this point. */
12522 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12523 	tcp_sndbuf_autoscale(tp, so, sendwin);
12524 	/*
12525 	 *
12526 	 */
12527 	if (bbr->rc_in_persist &&
12528 	    len &&
12529 	    (rsm == NULL) &&
12530 	    (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) {
12531 		/*
12532 		 * We are in persist, not doing a retransmit and don't have enough space
12533 		 * yet to send a full TSO. So is it at the end of the sb
12534 		 * if so we need to send else nuke to 0 and don't send.
12535 		 */
12536 		int sbleft;
12537 		if (sbavail(sb) > sb_offset)
12538 			sbleft = sbavail(sb) - sb_offset;
12539 		else
12540 			sbleft = 0;
12541 		if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) {
12542 			/* not at end of sb lets not send */
12543 			len = 0;
12544 		}
12545 	}
12546 	/*
12547 	 * Decide if we can use TCP Segmentation Offloading (if supported by
12548 	 * hardware).
12549 	 *
12550 	 * TSO may only be used if we are in a pure bulk sending state.  The
12551 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12552 	 * options prevent using TSO.  With TSO the TCP header is the same
12553 	 * (except for the sequence number) for all generated packets.  This
12554 	 * makes it impossible to transmit any options which vary per
12555 	 * generated segment or packet.
12556 	 *
12557 	 * IPv4 handling has a clear separation of ip options and ip header
12558 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen()
12559 	 * does the right thing below to provide length of just ip options
12560 	 * and thus checking for ipoptlen is enough to decide if ip options
12561 	 * are present.
12562 	 */
12563 #ifdef INET6
12564 	if (isipv6)
12565 		ipoptlen = ip6_optlen(inp);
12566 	else
12567 #endif
12568 	if (inp->inp_options)
12569 		ipoptlen = inp->inp_options->m_len -
12570 		    offsetof(struct ipoption, ipopt_list);
12571 	else
12572 		ipoptlen = 0;
12573 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12574 	/*
12575 	 * Pre-calculate here as we save another lookup into the darknesses
12576 	 * of IPsec that way and can actually decide if TSO is ok.
12577 	 */
12578 #ifdef INET6
12579 	if (isipv6 && IPSEC_ENABLED(ipv6))
12580 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
12581 #ifdef INET
12582 	else
12583 #endif
12584 #endif				/* INET6 */
12585 #ifdef INET
12586 	if (IPSEC_ENABLED(ipv4))
12587 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
12588 #endif				/* INET */
12589 #endif				/* IPSEC */
12590 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12591 	ipoptlen += ipsec_optlen;
12592 #endif
12593 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
12594 	    (len > maxseg) &&
12595 	    (tp->t_port == 0) &&
12596 	    ((tp->t_flags & TF_SIGNATURE) == 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 	SOCK_SENDBUF_UNLOCK(so);
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 	SOCK_SENDBUF_LOCK_ASSERT(so);
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 ((tp->t_flags & TF_FASTOPEN) &&
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_msec(&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_msec(&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 ((tp->t_flags & TF_FASTOPEN) && 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 			SOCK_SENDBUF_UNLOCK(so);
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 				SOCK_SENDBUF_UNLOCK(so);
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 			SOCK_SENDBUF_UNLOCK(so);
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 					SOCK_SENDBUF_UNLOCK(so);
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 			if (len <= maxseg) {
13175 				/*
13176 				 * Must have ran out of mbufs for the copy
13177 				 * shorten it to no longer need tso. Lets
13178 				 * not put on sendalot since we are low on
13179 				 * mbufs.
13180 				 */
13181 				tso = 0;
13182 			}
13183 			if (m->m_next == NULL) {
13184 				SOCK_SENDBUF_UNLOCK(so);
13185 				(void)m_free(m);
13186 				error = ENOBUFS;
13187 				sack_rxmit = 0;
13188 				goto out;
13189 			}
13190 		}
13191 #ifdef BBR_INVARIANTS
13192 		if (tso && len < maxseg) {
13193 			panic("tp:%p tso on, but len:%d < maxseg:%d",
13194 			    tp, len, maxseg);
13195 		}
13196 		if (tso && if_hw_tsomaxsegcount) {
13197 			int32_t seg_cnt = 0;
13198 			struct mbuf *foo;
13199 
13200 			foo = m;
13201 			while (foo) {
13202 				seg_cnt++;
13203 				foo = foo->m_next;
13204 			}
13205 			if (seg_cnt > if_hw_tsomaxsegcount) {
13206 				panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount);
13207 			}
13208 		}
13209 #endif
13210 		/*
13211 		 * If we're sending everything we've got, set PUSH. (This
13212 		 * will keep happy those implementations which only give
13213 		 * data to the user when a buffer fills or a PUSH comes in.)
13214 		 */
13215 		if (sb_offset + len == sbused(sb) &&
13216 		    sbused(sb) &&
13217 		    !(flags & TH_SYN)) {
13218 			flags |= TH_PUSH;
13219 		}
13220 		SOCK_SENDBUF_UNLOCK(so);
13221 	} else {
13222 		SOCK_SENDBUF_UNLOCK(so);
13223 		if (tp->t_flags & TF_ACKNOW)
13224 			KMOD_TCPSTAT_INC(tcps_sndacks);
13225 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
13226 			KMOD_TCPSTAT_INC(tcps_sndctrl);
13227 		else
13228 			KMOD_TCPSTAT_INC(tcps_sndwinup);
13229 
13230 		m = m_gethdr(M_NOWAIT, MT_DATA);
13231 		if (m == NULL) {
13232 			BBR_STAT_INC(bbr_failed_mbuf_aloc);
13233 			bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13234 			error = ENOBUFS;
13235 			/* Fudge the send time since we could not send */
13236 			sack_rxmit = 0;
13237 			goto out;
13238 		}
13239 #ifdef INET6
13240 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13241 		    MHLEN >= hdrlen) {
13242 			M_ALIGN(m, hdrlen);
13243 		} else
13244 #endif
13245 			m->m_data += max_linkhdr;
13246 		m->m_len = hdrlen;
13247 	}
13248 	SOCK_SENDBUF_UNLOCK_ASSERT(so);
13249 	m->m_pkthdr.rcvif = (struct ifnet *)0;
13250 #ifdef MAC
13251 	mac_inpcb_create_mbuf(inp, m);
13252 #endif
13253 #ifdef INET6
13254 	if (isipv6) {
13255 		ip6 = mtod(m, struct ip6_hdr *);
13256 		if (tp->t_port) {
13257 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
13258 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13259 			udp->uh_dport = tp->t_port;
13260 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
13261 			udp->uh_ulen = htons(ulen);
13262 			th = (struct tcphdr *)(udp + 1);
13263 		} else {
13264 			th = (struct tcphdr *)(ip6 + 1);
13265 		}
13266 		tcpip_fillheaders(inp, tp->t_port, ip6, th);
13267 	} else
13268 #endif				/* INET6 */
13269 	{
13270 		ip = mtod(m, struct ip *);
13271 		if (tp->t_port) {
13272 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
13273 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13274 			udp->uh_dport = tp->t_port;
13275 			ulen = hdrlen + len - sizeof(struct ip);
13276 			udp->uh_ulen = htons(ulen);
13277 			th = (struct tcphdr *)(udp + 1);
13278 		} else {
13279 			th = (struct tcphdr *)(ip + 1);
13280 		}
13281 		tcpip_fillheaders(inp, tp->t_port, ip, th);
13282 	}
13283 	/*
13284 	 * If we are doing retransmissions, then snd_nxt will not reflect
13285 	 * the first unsent octet.  For ACK only packets, we do not want the
13286 	 * sequence number of the retransmitted packet, we want the sequence
13287 	 * number of the next unsent octet.  So, if there is no data (and no
13288 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
13289 	 * ti_seq.  But if we are in persist state, snd_max might reflect
13290 	 * one byte beyond the right edge of the window, so use snd_nxt in
13291 	 * that case, since we know we aren't doing a retransmission.
13292 	 * (retransmit and persist are mutually exclusive...)
13293 	 */
13294 	if (sack_rxmit == 0) {
13295 		if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) {
13296 			/* New data (including new persists) */
13297 			th->th_seq = htonl(tp->snd_max);
13298 			bbr_seq = tp->snd_max;
13299 		} else if (flags & TH_SYN) {
13300 			/* Syn's always send from iss */
13301 			th->th_seq = htonl(tp->iss);
13302 			bbr_seq = tp->iss;
13303 		} else if (flags & TH_FIN) {
13304 			if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) {
13305 				/*
13306 				 * If we sent the fin already its 1 minus
13307 				 * snd_max
13308 				 */
13309 				th->th_seq = (htonl(tp->snd_max - 1));
13310 				bbr_seq = (tp->snd_max - 1);
13311 			} else {
13312 				/* First time FIN use snd_max */
13313 				th->th_seq = htonl(tp->snd_max);
13314 				bbr_seq = tp->snd_max;
13315 			}
13316 		} else {
13317 			/*
13318 			 * len == 0 and not persist we use snd_max, sending
13319 			 * an ack unless we have sent the fin then its 1
13320 			 * minus.
13321 			 */
13322 			/*
13323 			 * XXXRRS Question if we are in persists and we have
13324 			 * nothing outstanding to send and we have not sent
13325 			 * a FIN, we will send an ACK. In such a case it
13326 			 * might be better to send (tp->snd_una - 1) which
13327 			 * would force the peer to ack.
13328 			 */
13329 			if (tp->t_flags & TF_SENTFIN) {
13330 				th->th_seq = htonl(tp->snd_max - 1);
13331 				bbr_seq = (tp->snd_max - 1);
13332 			} else {
13333 				th->th_seq = htonl(tp->snd_max);
13334 				bbr_seq = tp->snd_max;
13335 			}
13336 		}
13337 	} else {
13338 		/* All retransmits use the rsm to guide the send */
13339 		th->th_seq = htonl(rsm->r_start);
13340 		bbr_seq = rsm->r_start;
13341 	}
13342 	th->th_ack = htonl(tp->rcv_nxt);
13343 	if (optlen) {
13344 		bcopy(opt, th + 1, optlen);
13345 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13346 	}
13347 	tcp_set_flags(th, flags);
13348 	/*
13349 	 * Calculate receive window.  Don't shrink window, but avoid silly
13350 	 * window syndrome.
13351 	 */
13352 	if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) &&
13353 				  recwin < maxseg)))
13354 		recwin = 0;
13355 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13356 	    recwin < (tp->rcv_adv - tp->rcv_nxt))
13357 		recwin = (tp->rcv_adv - tp->rcv_nxt);
13358 	if (recwin > TCP_MAXWIN << tp->rcv_scale)
13359 		recwin = TCP_MAXWIN << tp->rcv_scale;
13360 
13361 	/*
13362 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13363 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
13364 	 * handled in syncache.
13365 	 */
13366 	if (flags & TH_SYN)
13367 		th->th_win = htons((u_short)
13368 		    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13369 	else {
13370 		/* Avoid shrinking window with window scaling. */
13371 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
13372 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13373 	}
13374 	/*
13375 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13376 	 * window.  This may cause the remote transmitter to stall.  This
13377 	 * flag tells soreceive() to disable delayed acknowledgements when
13378 	 * draining the buffer.  This can occur if the receiver is
13379 	 * attempting to read more data than can be buffered prior to
13380 	 * transmitting on the connection.
13381 	 */
13382 	if (th->th_win == 0) {
13383 		tp->t_sndzerowin++;
13384 		tp->t_flags |= TF_RXWIN0SENT;
13385 	} else
13386 		tp->t_flags &= ~TF_RXWIN0SENT;
13387 	/*
13388 	 * We don't support urgent data, but drag along
13389 	 * the pointer in case of a stack switch.
13390 	 */
13391 	tp->snd_up = tp->snd_una;
13392 	/*
13393 	 * Put TCP length in extended header, and then checksum extended
13394 	 * header and data.
13395 	 */
13396 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
13397 
13398 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13399 	if (to.to_flags & TOF_SIGNATURE) {
13400 		/*
13401 		 * Calculate MD5 signature and put it into the place
13402 		 * determined before. NOTE: since TCP options buffer doesn't
13403 		 * point into mbuf's data, calculate offset and use it.
13404 		 */
13405 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13406 		    (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13407 			/*
13408 			 * Do not send segment if the calculation of MD5
13409 			 * digest has failed.
13410 			 */
13411 			goto out;
13412 		}
13413 	}
13414 #endif
13415 
13416 #ifdef INET6
13417 	if (isipv6) {
13418 		/*
13419 		 * ip6_plen is not need to be filled now, and will be filled
13420 		 * in ip6_output.
13421 		 */
13422 		if (tp->t_port) {
13423 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13424 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13425 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13426 			th->th_sum = htons(0);
13427 			UDPSTAT_INC(udps_opackets);
13428 		} else {
13429 			csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13430 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13431 			th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
13432 			    optlen + len, IPPROTO_TCP, 0);
13433 		}
13434 	}
13435 #endif
13436 #if defined(INET6) && defined(INET)
13437 	else
13438 #endif
13439 #ifdef INET
13440 	{
13441 		if (tp->t_port) {
13442 			m->m_pkthdr.csum_flags = CSUM_UDP;
13443 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13444 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13445 			    ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13446 			th->th_sum = htons(0);
13447 			UDPSTAT_INC(udps_opackets);
13448 		} else {
13449 			csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP;
13450 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13451 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
13452 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13453 			    IPPROTO_TCP + len + optlen));
13454 		}
13455 		/* IP version must be set here for ipv4/ipv6 checking later */
13456 		KASSERT(ip->ip_v == IPVERSION,
13457 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
13458 	}
13459 #endif
13460 
13461 	/*
13462 	 * Enable TSO and specify the size of the segments. The TCP pseudo
13463 	 * header checksum is always provided. XXX: Fixme: This is currently
13464 	 * not the case for IPv6.
13465 	 */
13466 	if (tso) {
13467 		KASSERT(len > maxseg,
13468 		    ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg));
13469 		m->m_pkthdr.csum_flags |= CSUM_TSO;
13470 		csum_flags |= CSUM_TSO;
13471 		m->m_pkthdr.tso_segsz = maxseg;
13472 	}
13473 	KASSERT(len + hdrlen == m_length(m, NULL),
13474 	    ("%s: mbuf chain different than expected: %d + %u != %u",
13475 	    __func__, len, hdrlen, m_length(m, NULL)));
13476 
13477 #ifdef TCP_HHOOK
13478 	/* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */
13479 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
13480 #endif
13481 
13482 	/* Log to the black box */
13483 	if (tcp_bblogging_on(tp)) {
13484 		union tcp_log_stackspecific log;
13485 
13486 		bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
13487 		/* Record info on type of transmission */
13488 		log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay;
13489 		log.u_bbr.flex2 = (bbr->r_recovery_bw << 3);
13490 		log.u_bbr.flex3 = maxseg;
13491 		log.u_bbr.flex4 = delay_calc;
13492 		log.u_bbr.flex5 = bbr->rc_past_init_win;
13493 		log.u_bbr.flex5 <<= 1;
13494 		log.u_bbr.flex5 |= bbr->rc_no_pacing;
13495 		log.u_bbr.flex5 <<= 29;
13496 		log.u_bbr.flex5 |= tp->t_maxseg;
13497 		log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs;
13498 		log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr);
13499 		/* lets poke in the low and the high here for debugging */
13500 		log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
13501 		if (rsm || sack_rxmit) {
13502 			if (doing_tlp)
13503 				log.u_bbr.flex8 = 2;
13504 			else
13505 				log.u_bbr.flex8 = 1;
13506 		} else {
13507 			log.u_bbr.flex8 = 0;
13508 		}
13509 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13510 		    len, &log, false, NULL, NULL, 0, tv);
13511 	} else {
13512 		lgb = NULL;
13513 	}
13514 	/*
13515 	 * Fill in IP length and desired time to live and send to IP level.
13516 	 * There should be a better way to handle ttl and tos; we could keep
13517 	 * them in the template, but need a way to checksum without them.
13518 	 */
13519 	/*
13520 	 * m->m_pkthdr.len should have been set before cksum calcuration,
13521 	 * because in6_cksum() need it.
13522 	 */
13523 #ifdef INET6
13524 	if (isipv6) {
13525 		/*
13526 		 * we separately set hoplimit for every segment, since the
13527 		 * user might want to change the value via setsockopt. Also,
13528 		 * desired default hop limit might be changed via Neighbor
13529 		 * Discovery.
13530 		 */
13531 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13532 
13533 		/*
13534 		 * Set the packet size here for the benefit of DTrace
13535 		 * probes. ip6_output() will set it properly; it's supposed
13536 		 * to include the option header lengths as well.
13537 		 */
13538 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13539 
13540 		if (V_path_mtu_discovery && maxseg > V_tcp_minmss)
13541 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13542 		else
13543 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13544 
13545 		if (tp->t_state == TCPS_SYN_SENT)
13546 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13547 
13548 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13549 		/* TODO: IPv6 IP6TOS_ECT bit on */
13550 		error = ip6_output(m, inp->in6p_outputopts,
13551 		    &inp->inp_route6,
13552 		    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13553 		    NULL, NULL, inp);
13554 
13555 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
13556 			mtu = inp->inp_route6.ro_nh->nh_mtu;
13557 	}
13558 #endif				/* INET6 */
13559 #if defined(INET) && defined(INET6)
13560 	else
13561 #endif
13562 #ifdef INET
13563 	{
13564 		ip->ip_len = htons(m->m_pkthdr.len);
13565 #ifdef INET6
13566 		if (isipv6)
13567 			ip->ip_ttl = in6_selecthlim(inp, NULL);
13568 #endif				/* INET6 */
13569 		/*
13570 		 * If we do path MTU discovery, then we set DF on every
13571 		 * packet. This might not be the best thing to do according
13572 		 * to RFC3390 Section 2. However the tcp hostcache migitates
13573 		 * the problem so it affects only the first tcp connection
13574 		 * with a host.
13575 		 *
13576 		 * NB: Don't set DF on small MTU/MSS to have a safe
13577 		 * fallback.
13578 		 */
13579 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13580 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13581 			if (tp->t_port == 0 || len < V_tcp_minmss) {
13582 				ip->ip_off |= htons(IP_DF);
13583 			}
13584 		} else {
13585 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13586 		}
13587 
13588 		if (tp->t_state == TCPS_SYN_SENT)
13589 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
13590 
13591 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
13592 
13593 		error = ip_output(m, inp->inp_options, &inp->inp_route,
13594 		    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
13595 		    inp);
13596 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
13597 			mtu = inp->inp_route.ro_nh->nh_mtu;
13598 	}
13599 #endif				/* INET */
13600 	if (lgb) {
13601 		lgb->tlb_errno = error;
13602 		lgb = NULL;
13603 	}
13604 
13605 out:
13606 	/*
13607 	 * In transmit state, time the transmission and arrange for the
13608 	 * retransmit.  In persist state, just set snd_max.
13609 	 */
13610 	if (error == 0) {
13611 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
13612 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13613 		    (tp->t_flags & TF_SACK_PERMIT) &&
13614 		    tp->rcv_numsacks > 0)
13615 			tcp_clean_dsack_blocks(tp);
13616 		/* We sent an ack clear the bbr_segs_rcvd count */
13617 		bbr->output_error_seen = 0;
13618 		bbr->oerror_cnt = 0;
13619 		bbr->bbr_segs_rcvd = 0;
13620 		if (len == 0)
13621 			counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1);
13622 		/* Do accounting for new sends */
13623 		if ((len > 0) && (rsm == NULL)) {
13624 			int idx;
13625 			if (tp->snd_una == tp->snd_max) {
13626 				/*
13627 				 * Special case to match google, when
13628 				 * nothing is in flight the delivered
13629 				 * time does get updated to the current
13630 				 * time (see tcp_rate_bsd.c).
13631 				 */
13632 				bbr->r_ctl.rc_del_time = cts;
13633 			}
13634 			if (len >= maxseg) {
13635 				idx = (len / maxseg) + 3;
13636 				if (idx >= TCP_MSS_ACCT_ATIMER)
13637 					counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1);
13638 				else
13639 					counter_u64_add(bbr_out_size[idx], 1);
13640 			} else {
13641 				/* smaller than a MSS */
13642 				idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options);
13643 				if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV)
13644 					idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1);
13645 				counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1);
13646 			}
13647 		}
13648 	}
13649 	abandon = 0;
13650 	/*
13651 	 * We must do the send accounting before we log the output,
13652 	 * otherwise the state of the rsm could change and we account to the
13653 	 * wrong bucket.
13654 	 */
13655 	if (len > 0) {
13656 		bbr_do_send_accounting(tp, bbr, rsm, len, error);
13657 		if (error == 0) {
13658 			if (tp->snd_una == tp->snd_max)
13659 				bbr->r_ctl.rc_tlp_rxt_last_time = cts;
13660 		}
13661 	}
13662 	bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error,
13663 	    cts, mb, &abandon, rsm, 0, sb);
13664 	if (abandon) {
13665 		/*
13666 		 * If bbr_log_output destroys the TCB or sees a TH_RST being
13667 		 * sent we should hit this condition.
13668 		 */
13669 		return (0);
13670 	}
13671 	if (bbr->rc_in_persist == 0) {
13672 		/*
13673 		 * Advance snd_nxt over sequence space of this segment.
13674 		 */
13675 		if (error)
13676 			/* We don't log or do anything with errors */
13677 			goto skip_upd;
13678 
13679 		if (tp->snd_una == tp->snd_max &&
13680 		    (len || (flags & (TH_SYN | TH_FIN)))) {
13681 			/*
13682 			 * Update the time we just added data since none was
13683 			 * outstanding.
13684 			 */
13685 			bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13686 			bbr->rc_tp->t_acktime  = ticks;
13687 		}
13688 		if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) {
13689 			if (flags & TH_SYN) {
13690 				/*
13691 				 * Smack the snd_max to iss + 1
13692 				 * if its a FO we will add len below.
13693 				 */
13694 				tp->snd_max = tp->iss + 1;
13695 			}
13696 			if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13697 				tp->snd_max++;
13698 				tp->t_flags |= TF_SENTFIN;
13699 			}
13700 		}
13701 		if (sack_rxmit == 0)
13702 			tp->snd_max += len;
13703 skip_upd:
13704 		if ((error == 0) && len)
13705 			tot_len += len;
13706 	} else {
13707 		/* Persists case */
13708 		int32_t xlen = len;
13709 
13710 		if (error)
13711 			goto nomore;
13712 
13713 		if (flags & TH_SYN)
13714 			++xlen;
13715 		if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13716 			++xlen;
13717 			tp->t_flags |= TF_SENTFIN;
13718 		}
13719 		if (xlen && (tp->snd_una == tp->snd_max)) {
13720 			/*
13721 			 * Update the time we just added data since none was
13722 			 * outstanding.
13723 			 */
13724 			bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13725 			bbr->rc_tp->t_acktime = ticks;
13726 		}
13727 		if (sack_rxmit == 0)
13728 			tp->snd_max += xlen;
13729 		tot_len += (len + optlen + ipoptlen);
13730 	}
13731 nomore:
13732 	if (error) {
13733 		/*
13734 		 * Failures do not advance the seq counter above. For the
13735 		 * case of ENOBUFS we will fall out and become ack-clocked.
13736 		 * capping the cwnd at the current flight.
13737 		 * Everything else will just have to retransmit with the timer
13738 		 * (no pacer).
13739 		 */
13740 		SOCK_SENDBUF_UNLOCK_ASSERT(so);
13741 		BBR_STAT_INC(bbr_saw_oerr);
13742 		/* Clear all delay/early tracks */
13743 		bbr->r_ctl.rc_hptsi_agg_delay = 0;
13744 		bbr->r_ctl.rc_agg_early = 0;
13745 		bbr->r_agg_early_set = 0;
13746 		bbr->output_error_seen = 1;
13747 		if (bbr->oerror_cnt < 0xf)
13748 			bbr->oerror_cnt++;
13749 		if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) {
13750 			/* drop the session */
13751 			return (-ENETDOWN);
13752 		}
13753 		switch (error) {
13754 		case ENOBUFS:
13755 			/*
13756 			 * Make this guy have to get ack's to send
13757 			 * more but lets make sure we don't
13758 			 * slam him below a T-O (1MSS).
13759 			 */
13760 			if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
13761 				tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13762 								    bbr->r_ctl.rc_lost_bytes)) - maxseg;
13763 				if (tp->snd_cwnd < maxseg)
13764 					tp->snd_cwnd = maxseg;
13765 			}
13766 			slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt;
13767 			BBR_STAT_INC(bbr_saw_enobuf);
13768 			if (bbr->bbr_hdrw_pacing)
13769 				counter_u64_add(bbr_hdwr_pacing_enobuf, 1);
13770 			else
13771 				counter_u64_add(bbr_nohdwr_pacing_enobuf, 1);
13772 			/*
13773 			 * Here even in the enobuf's case we want to do our
13774 			 * state update. The reason being we may have been
13775 			 * called by the input function. If so we have had
13776 			 * things change.
13777 			 */
13778 			error = 0;
13779 			goto enobufs;
13780 		case EMSGSIZE:
13781 			/*
13782 			 * For some reason the interface we used initially
13783 			 * to send segments changed to another or lowered
13784 			 * its MTU. If TSO was active we either got an
13785 			 * interface without TSO capabilits or TSO was
13786 			 * turned off. If we obtained mtu from ip_output()
13787 			 * then update it and try again.
13788 			 */
13789 			/* Turn on tracing (or try to) */
13790 			{
13791 				int old_maxseg;
13792 
13793 				old_maxseg = tp->t_maxseg;
13794 				BBR_STAT_INC(bbr_saw_emsgsiz);
13795 				bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts);
13796 				if (mtu != 0)
13797 					tcp_mss_update(tp, -1, mtu, NULL, NULL);
13798 				if (old_maxseg <= tp->t_maxseg) {
13799 					/* Huh it did not shrink? */
13800 					tp->t_maxseg = old_maxseg - 40;
13801 					if (tp->t_maxseg < V_tcp_mssdflt) {
13802 						/*
13803 						 * The MSS is so small we should not
13804 						 * process incoming SACK's since we are
13805 						 * subject to attack in such a case.
13806 						 */
13807 						tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
13808 					} else {
13809 						tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
13810 					}
13811 					bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts);
13812 				}
13813 				/*
13814 				 * Nuke all other things that can interfere
13815 				 * with slot
13816 				 */
13817 				if ((tot_len + len) && (len >= tp->t_maxseg)) {
13818 					slot = bbr_get_pacing_delay(bbr,
13819 					    bbr->r_ctl.rc_bbr_hptsi_gain,
13820 					    (tot_len + len), cts, 0);
13821 					if (slot < bbr_error_base_paceout)
13822 						slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
13823 				} else
13824 					slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
13825 				bbr->rc_output_starts_timer = 1;
13826 				bbr_start_hpts_timer(bbr, tp, cts, 10, slot,
13827 				    tot_len);
13828 				return (error);
13829 			}
13830 		case EPERM:
13831 		case EACCES:
13832 			tp->t_softerror = error;
13833 			/* FALLTHROUGH */
13834 		case EHOSTDOWN:
13835 		case EHOSTUNREACH:
13836 		case ENETDOWN:
13837 		case ENETUNREACH:
13838 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
13839 				tp->t_softerror = error;
13840 				error = 0;
13841 			}
13842 			/* FALLTHROUGH */
13843 		default:
13844 			slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt;
13845 			bbr->rc_output_starts_timer = 1;
13846 			bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0);
13847 			return (error);
13848 		}
13849 #ifdef STATS
13850 	} else if (((tp->t_flags & TF_GPUTINPROG) == 0) &&
13851 		    len &&
13852 		    (rsm == NULL) &&
13853 	    (bbr->rc_in_persist == 0)) {
13854 		tp->gput_seq = bbr_seq;
13855 		tp->gput_ack = bbr_seq +
13856 		    min(sbavail(&so->so_snd) - sb_offset, sendwin);
13857 		tp->gput_ts = cts;
13858 		tp->t_flags |= TF_GPUTINPROG;
13859 #endif
13860 	}
13861 	KMOD_TCPSTAT_INC(tcps_sndtotal);
13862 	if ((bbr->bbr_hdw_pace_ena) &&
13863 	    (bbr->bbr_attempt_hdwr_pace == 0) &&
13864 	    (bbr->rc_past_init_win) &&
13865 	    (bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
13866 	    (get_filter_value(&bbr->r_ctl.rc_delrate)) &&
13867 	    (inp->inp_route.ro_nh &&
13868 	     inp->inp_route.ro_nh->nh_ifp)) {
13869 		/*
13870 		 * We are past the initial window and
13871 		 * have at least one measurement so we
13872 		 * could use hardware pacing if its available.
13873 		 * We have an interface and we have not attempted
13874 		 * to setup hardware pacing, lets try to now.
13875 		 */
13876 		uint64_t rate_wanted;
13877 		int err = 0;
13878 
13879 		rate_wanted = bbr_get_hardware_rate(bbr);
13880 		bbr->bbr_attempt_hdwr_pace = 1;
13881 		bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp,
13882 						      inp->inp_route.ro_nh->nh_ifp,
13883 						      rate_wanted,
13884 						      (RS_PACING_GEQ|RS_PACING_SUB_OK),
13885 						      &err, NULL);
13886 		if (bbr->r_ctl.crte) {
13887 			bbr_type_log_hdwr_pacing(bbr,
13888 						 bbr->r_ctl.crte->ptbl->rs_ifp,
13889 						 rate_wanted,
13890 						 bbr->r_ctl.crte->rate,
13891 						 __LINE__, cts, err);
13892 			BBR_STAT_INC(bbr_hdwr_rl_add_ok);
13893 			counter_u64_add(bbr_flows_nohdwr_pacing, -1);
13894 			counter_u64_add(bbr_flows_whdwr_pacing, 1);
13895 			bbr->bbr_hdrw_pacing = 1;
13896 			/* Now what is our gain status? */
13897 			if (bbr->r_ctl.crte->rate < rate_wanted) {
13898 				/* We have a problem */
13899 				bbr_setup_less_of_rate(bbr, cts,
13900 						       bbr->r_ctl.crte->rate, rate_wanted);
13901 			} else {
13902 				/* We are good */
13903 				bbr->gain_is_limited = 0;
13904 				bbr->skip_gain = 0;
13905 			}
13906 			tcp_bbr_tso_size_check(bbr, cts);
13907 		} else {
13908 			bbr_type_log_hdwr_pacing(bbr,
13909 						 inp->inp_route.ro_nh->nh_ifp,
13910 						 rate_wanted,
13911 						 0,
13912 						 __LINE__, cts, err);
13913 			BBR_STAT_INC(bbr_hdwr_rl_add_fail);
13914 		}
13915 	}
13916 	if (bbr->bbr_hdrw_pacing) {
13917 		/*
13918 		 * Worry about cases where the route
13919 		 * changes or something happened that we
13920 		 * lost our hardware pacing possibly during
13921 		 * the last ip_output call.
13922 		 */
13923 		if (inp->inp_snd_tag == NULL) {
13924 			/* A change during ip output disabled hw pacing? */
13925 			bbr->bbr_hdrw_pacing = 0;
13926 		} else if ((inp->inp_route.ro_nh == NULL) ||
13927 		    (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) {
13928 			/*
13929 			 * We had an interface or route change,
13930 			 * detach from the current hdwr pacing
13931 			 * and setup to re-attempt next go
13932 			 * round.
13933 			 */
13934 			bbr->bbr_hdrw_pacing = 0;
13935 			bbr->bbr_attempt_hdwr_pace = 0;
13936 			tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
13937 			tcp_bbr_tso_size_check(bbr, cts);
13938 		}
13939 	}
13940 	/*
13941 	 * Data sent (as far as we can tell). If this advertises a larger
13942 	 * window than any other segment, then remember the size of the
13943 	 * advertised window. Any pending ACK has now been sent.
13944 	 */
13945 	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
13946 		tp->rcv_adv = tp->rcv_nxt + recwin;
13947 
13948 	tp->last_ack_sent = tp->rcv_nxt;
13949 	if ((error == 0) &&
13950 	    (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) &&
13951 	    (doing_tlp == 0) &&
13952 	    (tso == 0) &&
13953 	    (len > 0) &&
13954 	    ((flags & TH_RST) == 0) &&
13955 	    ((flags & TH_SYN) == 0) &&
13956 	    (IN_RECOVERY(tp->t_flags) == 0) &&
13957 	    (bbr->rc_in_persist == 0) &&
13958 	    (tot_len < bbr->r_ctl.rc_pace_max_segs)) {
13959 		/*
13960 		 * For non-tso we need to goto again until we have sent out
13961 		 * enough data to match what we are hptsi out every hptsi
13962 		 * interval.
13963 		 */
13964 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
13965 			/* Make sure snd_nxt is drug up */
13966 			tp->snd_nxt = tp->snd_max;
13967 		}
13968 		if (rsm != NULL) {
13969 			rsm = NULL;
13970 			goto skip_again;
13971 		}
13972 		rsm = NULL;
13973 		sack_rxmit = 0;
13974 		tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
13975 		goto again;
13976 	}
13977 skip_again:
13978 	if ((error == 0) && (flags & TH_FIN))
13979 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
13980 	if ((error == 0) && (flags & TH_RST))
13981 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13982 	if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) {
13983 		/*
13984 		 * Calculate/Re-Calculate the hptsi slot in usecs based on
13985 		 * what we have sent so far
13986 		 */
13987 		slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
13988 		if (bbr->rc_no_pacing)
13989 			slot = 0;
13990 	}
13991 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
13992 enobufs:
13993 	if (bbr->rc_use_google == 0)
13994 		bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
13995 	bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13996 							bbr->r_ctl.rc_lost_bytes)));
13997 	bbr->rc_output_starts_timer = 1;
13998 	if (bbr->bbr_use_rack_cheat &&
13999 	    (more_to_rxt ||
14000 	     ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) {
14001 		/* Rack cheats and shotguns out all rxt's 1ms apart */
14002 		if (slot > 1000)
14003 			slot = 1000;
14004 	}
14005 	if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) {
14006 		/*
14007 		 * We don't change the tso size until some number of sends
14008 		 * to give the hardware commands time to get down
14009 		 * to the interface.
14010 		 */
14011 		bbr->r_ctl.bbr_hdwr_cnt_noset_snt++;
14012 		if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) {
14013 			bbr->hw_pacing_set = 1;
14014 			tcp_bbr_tso_size_check(bbr, cts);
14015 		}
14016 	}
14017 	bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len);
14018 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14019 		/* Make sure snd_nxt is drug up */
14020 		tp->snd_nxt = tp->snd_max;
14021 	}
14022 	return (error);
14023 
14024 }
14025 
14026 /*
14027  * See bbr_output_wtime() for return values.
14028  */
14029 static int
bbr_output(struct tcpcb * tp)14030 bbr_output(struct tcpcb *tp)
14031 {
14032 	int32_t ret;
14033 	struct timeval tv;
14034 
14035 	NET_EPOCH_ASSERT();
14036 
14037 	INP_WLOCK_ASSERT(tptoinpcb(tp));
14038 	(void)tcp_get_usecs(&tv);
14039 	ret = bbr_output_wtime(tp, &tv);
14040 	return (ret);
14041 }
14042 
14043 static void
bbr_mtu_chg(struct tcpcb * tp)14044 bbr_mtu_chg(struct tcpcb *tp)
14045 {
14046 	struct tcp_bbr *bbr;
14047 	struct bbr_sendmap *rsm, *frsm = NULL;
14048 	uint32_t maxseg;
14049 
14050 	/*
14051 	 * The MTU has changed. a) Clear the sack filter. b) Mark everything
14052 	 * over the current size as SACK_PASS so a retransmit will occur.
14053 	 */
14054 
14055 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14056 	maxseg = tp->t_maxseg - bbr->rc_last_options;
14057 	sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
14058 	TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
14059 		/* Don't mess with ones acked (by sack?) */
14060 		if (rsm->r_flags & BBR_ACKED)
14061 			continue;
14062 		if ((rsm->r_end - rsm->r_start) > maxseg) {
14063 			/*
14064 			 * We mark sack-passed on all the previous large
14065 			 * sends we did. This will force them to retransmit.
14066 			 */
14067 			rsm->r_flags |= BBR_SACK_PASSED;
14068 			if (((rsm->r_flags & BBR_MARKED_LOST) == 0) &&
14069 			    bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) {
14070 				bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
14071 				bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
14072 				rsm->r_flags |= BBR_MARKED_LOST;
14073 			}
14074 			if (frsm == NULL)
14075 				frsm = rsm;
14076 		}
14077 	}
14078 	if (frsm) {
14079 		bbr->r_ctl.rc_resend = frsm;
14080 	}
14081 }
14082 
14083 static int
bbr_pru_options(struct tcpcb * tp,int flags)14084 bbr_pru_options(struct tcpcb *tp, int flags)
14085 {
14086 	if (flags & PRUS_OOB)
14087 		return (EOPNOTSUPP);
14088 	return (0);
14089 }
14090 
14091 static void
bbr_switch_failed(struct tcpcb * tp)14092 bbr_switch_failed(struct tcpcb *tp)
14093 {
14094 	/*
14095 	 * If a switch fails we only need to
14096 	 * make sure mbuf_queuing is still in place.
14097 	 * We also need to make sure we are still in
14098 	 * ticks granularity (though we should probably
14099 	 * change bbr to go to USECs).
14100 	 *
14101 	 * For timers we need to see if we are still in the
14102 	 * pacer (if our flags are up) if so we are good, if
14103 	 * not we need to get back into the pacer.
14104 	 */
14105 	struct timeval tv;
14106 	uint32_t cts;
14107 	uint32_t toval;
14108 	struct tcp_bbr *bbr;
14109 	struct hpts_diag diag;
14110 
14111 	tp->t_flags2 |= TF2_CANNOT_DO_ECN;
14112 	tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14113 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_TICKS);
14114 	if (tp->t_in_hpts > IHPTS_NONE) {
14115 		return;
14116 	}
14117 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14118 	cts = tcp_get_usecs(&tv);
14119 	if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14120 		if (TSTMP_GT(bbr->rc_pacer_started, cts)) {
14121 			toval = bbr->rc_pacer_started - cts;
14122 		} else {
14123 			/* one slot please */
14124 			toval = HPTS_USECS_PER_SLOT;
14125 		}
14126 	} else if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14127 		if (TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
14128 			toval = bbr->r_ctl.rc_timer_exp - cts;
14129 		} else {
14130 			/* one slot please */
14131 			toval = HPTS_USECS_PER_SLOT;
14132 		}
14133 	} else
14134 		toval = HPTS_USECS_PER_SLOT;
14135 	(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14136 				   __LINE__, &diag);
14137 	bbr_log_hpts_diag(bbr, cts, &diag);
14138 }
14139 
14140 struct tcp_function_block __tcp_bbr = {
14141 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
14142 	.tfb_tcp_output = bbr_output,
14143 	.tfb_do_queued_segments = ctf_do_queued_segments,
14144 	.tfb_do_segment_nounlock = bbr_do_segment_nounlock,
14145 	.tfb_tcp_do_segment = bbr_do_segment,
14146 	.tfb_tcp_ctloutput = bbr_ctloutput,
14147 	.tfb_tcp_fb_init = bbr_init,
14148 	.tfb_tcp_fb_fini = bbr_fini,
14149 	.tfb_tcp_timer_stop_all = bbr_stopall,
14150 	.tfb_tcp_rexmit_tmr = bbr_remxt_tmr,
14151 	.tfb_tcp_handoff_ok = bbr_handoff_ok,
14152 	.tfb_tcp_mtu_chg = bbr_mtu_chg,
14153 	.tfb_pru_options = bbr_pru_options,
14154 	.tfb_switch_failed = bbr_switch_failed,
14155 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP | TCP_FUNC_DEFAULT_OK,
14156 };
14157 
14158 /*
14159  * bbr_ctloutput() must drop the inpcb lock before performing copyin on
14160  * socket option arguments.  When it re-acquires the lock after the copy, it
14161  * has to revalidate that the connection is still valid for the socket
14162  * option.
14163  */
14164 static int
bbr_set_sockopt(struct tcpcb * tp,struct sockopt * sopt)14165 bbr_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
14166 {
14167 	struct epoch_tracker et;
14168 	struct inpcb *inp = tptoinpcb(tp);
14169 	struct tcp_bbr *bbr;
14170 	int32_t error = 0, optval;
14171 
14172 	switch (sopt->sopt_level) {
14173 	case IPPROTO_IPV6:
14174 	case IPPROTO_IP:
14175 		return (tcp_default_ctloutput(tp, sopt));
14176 	}
14177 
14178 	switch (sopt->sopt_name) {
14179 	case TCP_RACK_PACE_MAX_SEG:
14180 	case TCP_RACK_MIN_TO:
14181 	case TCP_RACK_REORD_THRESH:
14182 	case TCP_RACK_REORD_FADE:
14183 	case TCP_RACK_TLP_THRESH:
14184 	case TCP_RACK_PKT_DELAY:
14185 	case TCP_BBR_ALGORITHM:
14186 	case TCP_BBR_TSLIMITS:
14187 	case TCP_BBR_IWINTSO:
14188 	case TCP_BBR_STARTUP_PG:
14189 	case TCP_BBR_DRAIN_PG:
14190 	case TCP_BBR_PROBE_RTT_INT:
14191 	case TCP_BBR_PROBE_RTT_GAIN:
14192 	case TCP_BBR_PROBE_RTT_LEN:
14193 	case TCP_BBR_STARTUP_LOSS_EXIT:
14194 	case TCP_BBR_USEDEL_RATE:
14195 	case TCP_BBR_MIN_RTO:
14196 	case TCP_BBR_MAX_RTO:
14197 	case TCP_BBR_PACE_PER_SEC:
14198 	case TCP_DELACK:
14199 	case TCP_BBR_PACE_DEL_TAR:
14200 	case TCP_BBR_SEND_IWND_IN_TSO:
14201 	case TCP_BBR_EXTRA_STATE:
14202 	case TCP_BBR_UTTER_MAX_TSO:
14203 	case TCP_BBR_MIN_TOPACEOUT:
14204 	case TCP_BBR_FLOOR_MIN_TSO:
14205 	case TCP_BBR_TSTMP_RAISES:
14206 	case TCP_BBR_POLICER_DETECT:
14207 	case TCP_BBR_USE_RACK_CHEAT:
14208 	case TCP_DATA_AFTER_CLOSE:
14209 	case TCP_BBR_HDWR_PACE:
14210 	case TCP_BBR_PACE_SEG_MAX:
14211 	case TCP_BBR_PACE_SEG_MIN:
14212 	case TCP_BBR_PACE_CROSS:
14213 	case TCP_BBR_PACE_OH:
14214 	case TCP_BBR_TMR_PACE_OH:
14215 	case TCP_BBR_RACK_RTT_USE:
14216 	case TCP_BBR_RETRAN_WTSO:
14217 		break;
14218 	default:
14219 		return (tcp_default_ctloutput(tp, sopt));
14220 		break;
14221 	}
14222 	INP_WUNLOCK(inp);
14223 	error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14224 	if (error)
14225 		return (error);
14226 	INP_WLOCK(inp);
14227 	if (inp->inp_flags & INP_DROPPED) {
14228 		INP_WUNLOCK(inp);
14229 		return (ECONNRESET);
14230 	}
14231 	if (tp->t_fb != &__tcp_bbr) {
14232 		INP_WUNLOCK(inp);
14233 		return (ENOPROTOOPT);
14234 	}
14235 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14236 	switch (sopt->sopt_name) {
14237 	case TCP_BBR_PACE_PER_SEC:
14238 		BBR_OPTS_INC(tcp_bbr_pace_per_sec);
14239 		bbr->r_ctl.bbr_hptsi_per_second = optval;
14240 		break;
14241 	case TCP_BBR_PACE_DEL_TAR:
14242 		BBR_OPTS_INC(tcp_bbr_pace_del_tar);
14243 		bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval;
14244 		break;
14245 	case TCP_BBR_PACE_SEG_MAX:
14246 		BBR_OPTS_INC(tcp_bbr_pace_seg_max);
14247 		bbr->r_ctl.bbr_hptsi_segments_max = optval;
14248 		break;
14249 	case TCP_BBR_PACE_SEG_MIN:
14250 		BBR_OPTS_INC(tcp_bbr_pace_seg_min);
14251 		bbr->r_ctl.bbr_hptsi_bytes_min = optval;
14252 		break;
14253 	case TCP_BBR_PACE_CROSS:
14254 		BBR_OPTS_INC(tcp_bbr_pace_cross);
14255 		bbr->r_ctl.bbr_cross_over = optval;
14256 		break;
14257 	case TCP_BBR_ALGORITHM:
14258 		BBR_OPTS_INC(tcp_bbr_algorithm);
14259 		if (optval && (bbr->rc_use_google == 0)) {
14260 			/* Turn on the google mode */
14261 			bbr_google_mode_on(bbr);
14262 			if ((optval > 3) && (optval < 500)) {
14263 				/*
14264 				 * Must be at least greater than .3%
14265 				 * and must be less than 50.0%.
14266 				 */
14267 				bbr->r_ctl.bbr_google_discount = optval;
14268 			}
14269 		} else if ((optval == 0) && (bbr->rc_use_google == 1)) {
14270 			/* Turn off the google mode */
14271 			bbr_google_mode_off(bbr);
14272 		}
14273 		break;
14274 	case TCP_BBR_TSLIMITS:
14275 		BBR_OPTS_INC(tcp_bbr_tslimits);
14276 		if (optval == 1)
14277 			bbr->rc_use_ts_limit = 1;
14278 		else if (optval == 0)
14279 			bbr->rc_use_ts_limit = 0;
14280 		else
14281 			error = EINVAL;
14282 		break;
14283 
14284 	case TCP_BBR_IWINTSO:
14285 		BBR_OPTS_INC(tcp_bbr_iwintso);
14286 		if ((optval >= 0) && (optval < 128)) {
14287 			uint32_t twin;
14288 
14289 			bbr->rc_init_win = optval;
14290 			twin = bbr_initial_cwnd(bbr, tp);
14291 			if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd))
14292 				tp->snd_cwnd = twin;
14293 			else
14294 				error = EBUSY;
14295 		} else
14296 			error = EINVAL;
14297 		break;
14298 	case TCP_BBR_STARTUP_PG:
14299 		BBR_OPTS_INC(tcp_bbr_startup_pg);
14300 		if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) {
14301 			bbr->r_ctl.rc_startup_pg = optval;
14302 			if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
14303 				bbr->r_ctl.rc_bbr_hptsi_gain = optval;
14304 			}
14305 		} else
14306 			error = EINVAL;
14307 		break;
14308 	case TCP_BBR_DRAIN_PG:
14309 		BBR_OPTS_INC(tcp_bbr_drain_pg);
14310 		if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE))
14311 			bbr->r_ctl.rc_drain_pg = optval;
14312 		else
14313 			error = EINVAL;
14314 		break;
14315 	case TCP_BBR_PROBE_RTT_LEN:
14316 		BBR_OPTS_INC(tcp_bbr_probertt_len);
14317 		if (optval <= 1)
14318 			reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND));
14319 		else
14320 			error = EINVAL;
14321 		break;
14322 	case TCP_BBR_PROBE_RTT_GAIN:
14323 		BBR_OPTS_INC(tcp_bbr_probertt_gain);
14324 		if (optval <= BBR_UNIT)
14325 			bbr->r_ctl.bbr_rttprobe_gain_val = optval;
14326 		else
14327 			error = EINVAL;
14328 		break;
14329 	case TCP_BBR_PROBE_RTT_INT:
14330 		BBR_OPTS_INC(tcp_bbr_probe_rtt_int);
14331 		if (optval > 1000)
14332 			bbr->r_ctl.rc_probertt_int = optval;
14333 		else
14334 			error = EINVAL;
14335 		break;
14336 	case TCP_BBR_MIN_TOPACEOUT:
14337 		BBR_OPTS_INC(tcp_bbr_topaceout);
14338 		if (optval == 0) {
14339 			bbr->no_pacing_until = 0;
14340 			bbr->rc_no_pacing = 0;
14341 		} else if (optval <= 0x00ff) {
14342 			bbr->no_pacing_until = optval;
14343 			if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) &&
14344 			    (bbr->rc_bbr_state == BBR_STATE_STARTUP)){
14345 				/* Turn on no pacing */
14346 				bbr->rc_no_pacing = 1;
14347 			}
14348 		} else
14349 			error = EINVAL;
14350 		break;
14351 	case TCP_BBR_STARTUP_LOSS_EXIT:
14352 		BBR_OPTS_INC(tcp_bbr_startup_loss_exit);
14353 		bbr->rc_loss_exit = optval;
14354 		break;
14355 	case TCP_BBR_USEDEL_RATE:
14356 		error = EINVAL;
14357 		break;
14358 	case TCP_BBR_MIN_RTO:
14359 		BBR_OPTS_INC(tcp_bbr_min_rto);
14360 		bbr->r_ctl.rc_min_rto_ms = optval;
14361 		break;
14362 	case TCP_BBR_MAX_RTO:
14363 		BBR_OPTS_INC(tcp_bbr_max_rto);
14364 		bbr->rc_max_rto_sec = optval;
14365 		break;
14366 	case TCP_RACK_MIN_TO:
14367 		/* Minimum time between rack t-o's in ms */
14368 		BBR_OPTS_INC(tcp_rack_min_to);
14369 		bbr->r_ctl.rc_min_to = optval;
14370 		break;
14371 	case TCP_RACK_REORD_THRESH:
14372 		/* RACK reorder threshold (shift amount) */
14373 		BBR_OPTS_INC(tcp_rack_reord_thresh);
14374 		if ((optval > 0) && (optval < 31))
14375 			bbr->r_ctl.rc_reorder_shift = optval;
14376 		else
14377 			error = EINVAL;
14378 		break;
14379 	case TCP_RACK_REORD_FADE:
14380 		/* Does reordering fade after ms time */
14381 		BBR_OPTS_INC(tcp_rack_reord_fade);
14382 		bbr->r_ctl.rc_reorder_fade = optval;
14383 		break;
14384 	case TCP_RACK_TLP_THRESH:
14385 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14386 		BBR_OPTS_INC(tcp_rack_tlp_thresh);
14387 		if (optval)
14388 			bbr->rc_tlp_threshold = optval;
14389 		else
14390 			error = EINVAL;
14391 		break;
14392 	case TCP_BBR_USE_RACK_CHEAT:
14393 		BBR_OPTS_INC(tcp_use_rackcheat);
14394 		if (bbr->rc_use_google) {
14395 			error = EINVAL;
14396 			break;
14397 		}
14398 		BBR_OPTS_INC(tcp_rack_cheat);
14399 		if (optval)
14400 			bbr->bbr_use_rack_cheat = 1;
14401 		else
14402 			bbr->bbr_use_rack_cheat = 0;
14403 		break;
14404 	case TCP_BBR_FLOOR_MIN_TSO:
14405 		BBR_OPTS_INC(tcp_utter_max_tso);
14406 		if ((optval >= 0) && (optval < 40))
14407 			bbr->r_ctl.bbr_hptsi_segments_floor = optval;
14408 		else
14409 			error = EINVAL;
14410 		break;
14411 	case TCP_BBR_UTTER_MAX_TSO:
14412 		BBR_OPTS_INC(tcp_utter_max_tso);
14413 		if ((optval >= 0) && (optval < 0xffff))
14414 			bbr->r_ctl.bbr_utter_max = optval;
14415 		else
14416 			error = EINVAL;
14417 		break;
14418 
14419 	case TCP_BBR_EXTRA_STATE:
14420 		BBR_OPTS_INC(tcp_extra_state);
14421 		if (optval)
14422 			bbr->rc_use_idle_restart = 1;
14423 		else
14424 			bbr->rc_use_idle_restart = 0;
14425 		break;
14426 	case TCP_BBR_SEND_IWND_IN_TSO:
14427 		BBR_OPTS_INC(tcp_iwnd_tso);
14428 		if (optval) {
14429 			bbr->bbr_init_win_cheat = 1;
14430 			if (bbr->rc_past_init_win == 0) {
14431 				uint32_t cts;
14432 				cts = tcp_get_usecs(&bbr->rc_tv);
14433 				tcp_bbr_tso_size_check(bbr, cts);
14434 			}
14435 		} else
14436 			bbr->bbr_init_win_cheat = 0;
14437 		break;
14438 	case TCP_BBR_HDWR_PACE:
14439 		BBR_OPTS_INC(tcp_hdwr_pacing);
14440 		if (optval){
14441 			bbr->bbr_hdw_pace_ena = 1;
14442 			bbr->bbr_attempt_hdwr_pace = 0;
14443 		} else {
14444 			bbr->bbr_hdw_pace_ena = 0;
14445 #ifdef RATELIMIT
14446 			if (bbr->r_ctl.crte != NULL) {
14447 				tcp_rel_pacing_rate(bbr->r_ctl.crte, tp);
14448 				bbr->r_ctl.crte = NULL;
14449 			}
14450 #endif
14451 		}
14452 		break;
14453 
14454 	case TCP_DELACK:
14455 		BBR_OPTS_INC(tcp_delack);
14456 		if (optval < 100) {
14457 			if (optval == 0) /* off */
14458 				tp->t_delayed_ack = 0;
14459 			else if (optval == 1) /* on which is 2 */
14460 				tp->t_delayed_ack = 2;
14461 			else /* higher than 2 and less than 100 */
14462 				tp->t_delayed_ack = optval;
14463 			if (tp->t_flags & TF_DELACK) {
14464 				tp->t_flags &= ~TF_DELACK;
14465 				tp->t_flags |= TF_ACKNOW;
14466 				NET_EPOCH_ENTER(et);
14467 				bbr_output(tp);
14468 				NET_EPOCH_EXIT(et);
14469 			}
14470 		} else
14471 			error = EINVAL;
14472 		break;
14473 	case TCP_RACK_PKT_DELAY:
14474 		/* RACK added ms i.e. rack-rtt + reord + N */
14475 		BBR_OPTS_INC(tcp_rack_pkt_delay);
14476 		bbr->r_ctl.rc_pkt_delay = optval;
14477 		break;
14478 
14479 	case TCP_BBR_RETRAN_WTSO:
14480 		BBR_OPTS_INC(tcp_retran_wtso);
14481 		if (optval)
14482 			bbr->rc_resends_use_tso = 1;
14483 		else
14484 			bbr->rc_resends_use_tso = 0;
14485 		break;
14486 	case TCP_DATA_AFTER_CLOSE:
14487 		BBR_OPTS_INC(tcp_data_ac);
14488 		if (optval)
14489 			bbr->rc_allow_data_af_clo = 1;
14490 		else
14491 			bbr->rc_allow_data_af_clo = 0;
14492 		break;
14493 	case TCP_BBR_POLICER_DETECT:
14494 		BBR_OPTS_INC(tcp_policer_det);
14495 		if (bbr->rc_use_google == 0)
14496 			error = EINVAL;
14497 		else if (optval)
14498 			bbr->r_use_policer = 1;
14499 		else
14500 			bbr->r_use_policer = 0;
14501 		break;
14502 
14503 	case TCP_BBR_TSTMP_RAISES:
14504 		BBR_OPTS_INC(tcp_ts_raises);
14505 		if (optval)
14506 			bbr->ts_can_raise = 1;
14507 		else
14508 			bbr->ts_can_raise = 0;
14509 		break;
14510 	case TCP_BBR_TMR_PACE_OH:
14511 		BBR_OPTS_INC(tcp_pacing_oh_tmr);
14512 		if (bbr->rc_use_google) {
14513 			error = EINVAL;
14514 		} else {
14515 			if (optval)
14516 				bbr->r_ctl.rc_incr_tmrs = 1;
14517 			else
14518 				bbr->r_ctl.rc_incr_tmrs = 0;
14519 		}
14520 		break;
14521 	case TCP_BBR_PACE_OH:
14522 		BBR_OPTS_INC(tcp_pacing_oh);
14523 		if (bbr->rc_use_google) {
14524 			error = EINVAL;
14525 		} else {
14526 			if (optval > (BBR_INCL_TCP_OH|
14527 				      BBR_INCL_IP_OH|
14528 				      BBR_INCL_ENET_OH)) {
14529 				error = EINVAL;
14530 				break;
14531 			}
14532 			if (optval & BBR_INCL_TCP_OH)
14533 				bbr->r_ctl.rc_inc_tcp_oh = 1;
14534 			else
14535 				bbr->r_ctl.rc_inc_tcp_oh = 0;
14536 			if (optval & BBR_INCL_IP_OH)
14537 				bbr->r_ctl.rc_inc_ip_oh = 1;
14538 			else
14539 				bbr->r_ctl.rc_inc_ip_oh = 0;
14540 			if (optval & BBR_INCL_ENET_OH)
14541 				bbr->r_ctl.rc_inc_enet_oh = 1;
14542 			else
14543 				bbr->r_ctl.rc_inc_enet_oh = 0;
14544 		}
14545 		break;
14546 	default:
14547 		return (tcp_default_ctloutput(tp, sopt));
14548 		break;
14549 	}
14550 	tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14551 	INP_WUNLOCK(inp);
14552 	return (error);
14553 }
14554 
14555 /*
14556  * return 0 on success, error-num on failure
14557  */
14558 static int
bbr_get_sockopt(struct tcpcb * tp,struct sockopt * sopt)14559 bbr_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
14560 {
14561 	struct inpcb *inp = tptoinpcb(tp);
14562 	struct tcp_bbr *bbr;
14563 	uint64_t loptval;
14564 	int32_t error, optval;
14565 
14566 	bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14567 	if (bbr == NULL) {
14568 		INP_WUNLOCK(inp);
14569 		return (EINVAL);
14570 	}
14571 	/*
14572 	 * Because all our options are either boolean or an int, we can just
14573 	 * pull everything into optval and then unlock and copy. If we ever
14574 	 * add a option that is not a int, then this will have quite an
14575 	 * impact to this routine.
14576 	 */
14577 	switch (sopt->sopt_name) {
14578 	case TCP_BBR_PACE_PER_SEC:
14579 		optval = bbr->r_ctl.bbr_hptsi_per_second;
14580 		break;
14581 	case TCP_BBR_PACE_DEL_TAR:
14582 		optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar;
14583 		break;
14584 	case TCP_BBR_PACE_SEG_MAX:
14585 		optval = bbr->r_ctl.bbr_hptsi_segments_max;
14586 		break;
14587 	case TCP_BBR_MIN_TOPACEOUT:
14588 		optval = bbr->no_pacing_until;
14589 		break;
14590 	case TCP_BBR_PACE_SEG_MIN:
14591 		optval = bbr->r_ctl.bbr_hptsi_bytes_min;
14592 		break;
14593 	case TCP_BBR_PACE_CROSS:
14594 		optval = bbr->r_ctl.bbr_cross_over;
14595 		break;
14596 	case TCP_BBR_ALGORITHM:
14597 		optval = bbr->rc_use_google;
14598 		break;
14599 	case TCP_BBR_TSLIMITS:
14600 		optval = bbr->rc_use_ts_limit;
14601 		break;
14602 	case TCP_BBR_IWINTSO:
14603 		optval = bbr->rc_init_win;
14604 		break;
14605 	case TCP_BBR_STARTUP_PG:
14606 		optval = bbr->r_ctl.rc_startup_pg;
14607 		break;
14608 	case TCP_BBR_DRAIN_PG:
14609 		optval = bbr->r_ctl.rc_drain_pg;
14610 		break;
14611 	case TCP_BBR_PROBE_RTT_INT:
14612 		optval = bbr->r_ctl.rc_probertt_int;
14613 		break;
14614 	case TCP_BBR_PROBE_RTT_LEN:
14615 		optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND);
14616 		break;
14617 	case TCP_BBR_PROBE_RTT_GAIN:
14618 		optval = bbr->r_ctl.bbr_rttprobe_gain_val;
14619 		break;
14620 	case TCP_BBR_STARTUP_LOSS_EXIT:
14621 		optval = bbr->rc_loss_exit;
14622 		break;
14623 	case TCP_BBR_USEDEL_RATE:
14624 		loptval = get_filter_value(&bbr->r_ctl.rc_delrate);
14625 		break;
14626 	case TCP_BBR_MIN_RTO:
14627 		optval = bbr->r_ctl.rc_min_rto_ms;
14628 		break;
14629 	case TCP_BBR_MAX_RTO:
14630 		optval = bbr->rc_max_rto_sec;
14631 		break;
14632 	case TCP_RACK_PACE_MAX_SEG:
14633 		/* Max segments in a pace */
14634 		optval = bbr->r_ctl.rc_pace_max_segs;
14635 		break;
14636 	case TCP_RACK_MIN_TO:
14637 		/* Minimum time between rack t-o's in ms */
14638 		optval = bbr->r_ctl.rc_min_to;
14639 		break;
14640 	case TCP_RACK_REORD_THRESH:
14641 		/* RACK reorder threshold (shift amount) */
14642 		optval = bbr->r_ctl.rc_reorder_shift;
14643 		break;
14644 	case TCP_RACK_REORD_FADE:
14645 		/* Does reordering fade after ms time */
14646 		optval = bbr->r_ctl.rc_reorder_fade;
14647 		break;
14648 	case TCP_BBR_USE_RACK_CHEAT:
14649 		/* Do we use the rack cheat for rxt */
14650 		optval = bbr->bbr_use_rack_cheat;
14651 		break;
14652 	case TCP_BBR_FLOOR_MIN_TSO:
14653 		optval = bbr->r_ctl.bbr_hptsi_segments_floor;
14654 		break;
14655 	case TCP_BBR_UTTER_MAX_TSO:
14656 		optval = bbr->r_ctl.bbr_utter_max;
14657 		break;
14658 	case TCP_BBR_SEND_IWND_IN_TSO:
14659 		/* Do we send TSO size segments initially */
14660 		optval = bbr->bbr_init_win_cheat;
14661 		break;
14662 	case TCP_BBR_EXTRA_STATE:
14663 		optval = bbr->rc_use_idle_restart;
14664 		break;
14665 	case TCP_RACK_TLP_THRESH:
14666 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14667 		optval = bbr->rc_tlp_threshold;
14668 		break;
14669 	case TCP_RACK_PKT_DELAY:
14670 		/* RACK added ms i.e. rack-rtt + reord + N */
14671 		optval = bbr->r_ctl.rc_pkt_delay;
14672 		break;
14673 	case TCP_BBR_RETRAN_WTSO:
14674 		optval = bbr->rc_resends_use_tso;
14675 		break;
14676 	case TCP_DATA_AFTER_CLOSE:
14677 		optval = bbr->rc_allow_data_af_clo;
14678 		break;
14679 	case TCP_DELACK:
14680 		optval = tp->t_delayed_ack;
14681 		break;
14682 	case TCP_BBR_HDWR_PACE:
14683 		optval = bbr->bbr_hdw_pace_ena;
14684 		break;
14685 	case TCP_BBR_POLICER_DETECT:
14686 		optval = bbr->r_use_policer;
14687 		break;
14688 	case TCP_BBR_TSTMP_RAISES:
14689 		optval = bbr->ts_can_raise;
14690 		break;
14691 	case TCP_BBR_TMR_PACE_OH:
14692 		optval = bbr->r_ctl.rc_incr_tmrs;
14693 		break;
14694 	case TCP_BBR_PACE_OH:
14695 		optval = 0;
14696 		if (bbr->r_ctl.rc_inc_tcp_oh)
14697 			optval |= BBR_INCL_TCP_OH;
14698 		if (bbr->r_ctl.rc_inc_ip_oh)
14699 			optval |= BBR_INCL_IP_OH;
14700 		if (bbr->r_ctl.rc_inc_enet_oh)
14701 			optval |= BBR_INCL_ENET_OH;
14702 		break;
14703 	default:
14704 		return (tcp_default_ctloutput(tp, sopt));
14705 		break;
14706 	}
14707 	INP_WUNLOCK(inp);
14708 	if (sopt->sopt_name == TCP_BBR_USEDEL_RATE)
14709 		error = sooptcopyout(sopt, &loptval, sizeof loptval);
14710 	else
14711 		error = sooptcopyout(sopt, &optval, sizeof optval);
14712 	return (error);
14713 }
14714 
14715 /*
14716  * return 0 on success, error-num on failure
14717  */
14718 static int
bbr_ctloutput(struct tcpcb * tp,struct sockopt * sopt)14719 bbr_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
14720 {
14721 	if (sopt->sopt_dir == SOPT_SET) {
14722 		return (bbr_set_sockopt(tp, sopt));
14723 	} else if (sopt->sopt_dir == SOPT_GET) {
14724 		return (bbr_get_sockopt(tp, sopt));
14725 	} else {
14726 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
14727 	}
14728 }
14729 
14730 static const char *bbr_stack_names[] = {
14731 	__XSTRING(STACKNAME),
14732 #ifdef STACKALIAS
14733 	__XSTRING(STACKALIAS),
14734 #endif
14735 };
14736 
14737 static bool bbr_mod_inited = false;
14738 
14739 static int
tcp_addbbr(module_t mod,int32_t type,void * data)14740 tcp_addbbr(module_t mod, int32_t type, void *data)
14741 {
14742 	int32_t err = 0;
14743 	int num_stacks;
14744 
14745 	switch (type) {
14746 	case MOD_LOAD:
14747 		printf("Attempting to load " __XSTRING(MODNAME) "\n");
14748 		bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
14749 		    sizeof(struct bbr_sendmap),
14750 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
14751 		bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
14752 		    sizeof(struct tcp_bbr),
14753 		    NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
14754 		sysctl_ctx_init(&bbr_sysctl_ctx);
14755 		bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
14756 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
14757 		    OID_AUTO,
14758 #ifdef STACKALIAS
14759 		    __XSTRING(STACKALIAS),
14760 #else
14761 		    __XSTRING(STACKNAME),
14762 #endif
14763 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
14764 		    "");
14765 		if (bbr_sysctl_root == NULL) {
14766 			printf("Failed to add sysctl node\n");
14767 			err = EFAULT;
14768 			goto free_uma;
14769 		}
14770 		bbr_init_sysctls();
14771 		num_stacks = nitems(bbr_stack_names);
14772 		err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK,
14773 		    bbr_stack_names, &num_stacks);
14774 		if (err) {
14775 			printf("Failed to register %s stack name for "
14776 			    "%s module\n", bbr_stack_names[num_stacks],
14777 			    __XSTRING(MODNAME));
14778 			sysctl_ctx_free(&bbr_sysctl_ctx);
14779 	free_uma:
14780 			uma_zdestroy(bbr_zone);
14781 			uma_zdestroy(bbr_pcb_zone);
14782 			bbr_counter_destroy();
14783 			printf("Failed to register " __XSTRING(MODNAME)
14784 			    " module err:%d\n", err);
14785 			return (err);
14786 		}
14787 		tcp_lro_reg_mbufq();
14788 		bbr_mod_inited = true;
14789 		printf(__XSTRING(MODNAME) " is now available\n");
14790 		break;
14791 	case MOD_QUIESCE:
14792 		err = deregister_tcp_functions(&__tcp_bbr, true, false);
14793 		break;
14794 	case MOD_UNLOAD:
14795 		err = deregister_tcp_functions(&__tcp_bbr, false, true);
14796 		if (err == EBUSY)
14797 			break;
14798 		if (bbr_mod_inited) {
14799 			uma_zdestroy(bbr_zone);
14800 			uma_zdestroy(bbr_pcb_zone);
14801 			sysctl_ctx_free(&bbr_sysctl_ctx);
14802 			bbr_counter_destroy();
14803 			printf(__XSTRING(MODNAME)
14804 			    " is now no longer available\n");
14805 			bbr_mod_inited = false;
14806 		}
14807 		tcp_lro_dereg_mbufq();
14808 		err = 0;
14809 		break;
14810 	default:
14811 		return (EOPNOTSUPP);
14812 	}
14813 	return (err);
14814 }
14815 
14816 static moduledata_t tcp_bbr = {
14817 	.name = __XSTRING(MODNAME),
14818 	    .evhand = tcp_addbbr,
14819 	    .priv = 0
14820 };
14821 
14822 MODULE_VERSION(MODNAME, 1);
14823 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
14824 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
14825