xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision da5137ab) 
1 /*-
2  * Copyright (c) 2016-2020 Netflix, Inc.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include "opt_inet.h"
31 #include "opt_inet6.h"
32 #include "opt_ipsec.h"
33 #include "opt_tcpdebug.h"
34 #include "opt_ratelimit.h"
35 #include "opt_kern_tls.h"
36 #include <sys/param.h>
37 #include <sys/arb.h>
38 #include <sys/module.h>
39 #include <sys/kernel.h>
40 #ifdef TCP_HHOOK
41 #include <sys/hhook.h>
42 #endif
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/mbuf.h>
48 #include <sys/proc.h>		/* for proc0 declaration */
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 #else
58 #include <sys/tree.h>
59 #endif
60 #include <sys/refcount.h>
61 #include <sys/queue.h>
62 #include <sys/tim_filter.h>
63 #include <sys/smp.h>
64 #include <sys/kthread.h>
65 #include <sys/kern_prefetch.h>
66 #include <sys/protosw.h>
67 #ifdef TCP_ACCOUNTING
68 #include <sys/sched.h>
69 #include <machine/cpu.h>
70 #endif
71 #include <vm/uma.h>
72 
73 #include <net/route.h>
74 #include <net/route/nhop.h>
75 #include <net/vnet.h>
76 
77 #define TCPSTATES		/* for logging */
78 
79 #include <netinet/in.h>
80 #include <netinet/in_kdtrace.h>
81 #include <netinet/in_pcb.h>
82 #include <netinet/ip.h>
83 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
84 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
85 #include <netinet/ip_var.h>
86 #include <netinet/ip6.h>
87 #include <netinet6/in6_pcb.h>
88 #include <netinet6/ip6_var.h>
89 #include <netinet/tcp.h>
90 #define	TCPOUTFLAGS
91 #include <netinet/tcp_fsm.h>
92 #include <netinet/tcp_log_buf.h>
93 #include <netinet/tcp_seq.h>
94 #include <netinet/tcp_timer.h>
95 #include <netinet/tcp_var.h>
96 #include <netinet/tcp_syncache.h>
97 #include <netinet/tcp_hpts.h>
98 #include <netinet/tcp_ratelimit.h>
99 #include <netinet/tcp_accounting.h>
100 #include <netinet/tcpip.h>
101 #include <netinet/cc/cc.h>
102 #include <netinet/cc/cc_newreno.h>
103 #include <netinet/tcp_fastopen.h>
104 #include <netinet/tcp_lro.h>
105 #ifdef NETFLIX_SHARED_CWND
106 #include <netinet/tcp_shared_cwnd.h>
107 #endif
108 #ifdef TCPDEBUG
109 #include <netinet/tcp_debug.h>
110 #endif				/* TCPDEBUG */
111 #ifdef TCP_OFFLOAD
112 #include <netinet/tcp_offload.h>
113 #endif
114 #ifdef INET6
115 #include <netinet6/tcp6_var.h>
116 #endif
117 #include <netinet/tcp_ecn.h>
118 
119 #include <netipsec/ipsec_support.h>
120 
121 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
122 #include <netipsec/ipsec.h>
123 #include <netipsec/ipsec6.h>
124 #endif				/* IPSEC */
125 
126 #include <netinet/udp.h>
127 #include <netinet/udp_var.h>
128 #include <machine/in_cksum.h>
129 
130 #ifdef MAC
131 #include <security/mac/mac_framework.h>
132 #endif
133 #include "sack_filter.h"
134 #include "tcp_rack.h"
135 #include "rack_bbr_common.h"
136 
137 uma_zone_t rack_zone;
138 uma_zone_t rack_pcb_zone;
139 
140 #ifndef TICKS2SBT
141 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
142 #endif
143 
144 VNET_DECLARE(uint32_t, newreno_beta);
145 VNET_DECLARE(uint32_t, newreno_beta_ecn);
146 #define V_newreno_beta VNET(newreno_beta)
147 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
148 
149 
150 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
151 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
152 
153 struct sysctl_ctx_list rack_sysctl_ctx;
154 struct sysctl_oid *rack_sysctl_root;
155 
156 #define CUM_ACKED 1
157 #define SACKED 2
158 
159 /*
160  * The RACK module incorporates a number of
161  * TCP ideas that have been put out into the IETF
162  * over the last few years:
163  * - Matt Mathis's Rate Halving which slowly drops
164  *    the congestion window so that the ack clock can
165  *    be maintained during a recovery.
166  * - Yuchung Cheng's RACK TCP (for which its named) that
167  *    will stop us using the number of dup acks and instead
168  *    use time as the gage of when we retransmit.
169  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
170  *    of Dukkipati et.al.
171  * RACK depends on SACK, so if an endpoint arrives that
172  * cannot do SACK the state machine below will shuttle the
173  * connection back to using the "default" TCP stack that is
174  * in FreeBSD.
175  *
176  * To implement RACK the original TCP stack was first decomposed
177  * into a functional state machine with individual states
178  * for each of the possible TCP connection states. The do_segment
179  * functions role in life is to mandate the connection supports SACK
180  * initially and then assure that the RACK state matches the conenction
181  * state before calling the states do_segment function. Each
182  * state is simplified due to the fact that the original do_segment
183  * has been decomposed and we *know* what state we are in (no
184  * switches on the state) and all tests for SACK are gone. This
185  * greatly simplifies what each state does.
186  *
187  * TCP output is also over-written with a new version since it
188  * must maintain the new rack scoreboard.
189  *
190  */
191 static int32_t rack_tlp_thresh = 1;
192 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
193 static int32_t rack_tlp_use_greater = 1;
194 static int32_t rack_reorder_thresh = 2;
195 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
196 						 * - 60 seconds */
197 static uint8_t rack_req_measurements = 1;
198 /* Attack threshold detections */
199 static uint32_t rack_highest_sack_thresh_seen = 0;
200 static uint32_t rack_highest_move_thresh_seen = 0;
201 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
202 static int32_t rack_hw_pace_extra_slots = 2;	/* 2 extra MSS time betweens */
203 static int32_t rack_hw_rate_caps = 1; /* 1; */
204 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
205 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
206 static int32_t rack_hw_up_only = 1;
207 static int32_t rack_stats_gets_ms_rtt = 1;
208 static int32_t rack_prr_addbackmax = 2;
209 static int32_t rack_do_hystart = 0;
210 static int32_t rack_apply_rtt_with_reduced_conf = 0;
211 
212 static int32_t rack_pkt_delay = 1000;
213 static int32_t rack_send_a_lot_in_prr = 1;
214 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
215 static int32_t rack_verbose_logging = 0;
216 static int32_t rack_ignore_data_after_close = 1;
217 static int32_t rack_enable_shared_cwnd = 1;
218 static int32_t rack_use_cmp_acks = 1;
219 static int32_t rack_use_fsb = 1;
220 static int32_t rack_use_rfo = 1;
221 static int32_t rack_use_rsm_rfo = 1;
222 static int32_t rack_max_abc_post_recovery = 2;
223 static int32_t rack_client_low_buf = 0;
224 static int32_t rack_dsack_std_based = 0x3;	/* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */
225 #ifdef TCP_ACCOUNTING
226 static int32_t rack_tcp_accounting = 0;
227 #endif
228 static int32_t rack_limits_scwnd = 1;
229 static int32_t rack_enable_mqueue_for_nonpaced = 0;
230 static int32_t rack_disable_prr = 0;
231 static int32_t use_rack_rr = 1;
232 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
233 static int32_t rack_persist_min = 250000;	/* 250usec */
234 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
235 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
236 static int32_t rack_default_init_window = 0;	/* Use system default */
237 static int32_t rack_limit_time_with_srtt = 0;
238 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
239 static int32_t rack_enobuf_hw_boost_mult = 2;	/* How many times the hw rate we boost slot using time_between */
240 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
241 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
242 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
243 /*
244  * Currently regular tcp has a rto_min of 30ms
245  * the backoff goes 12 times so that ends up
246  * being a total of 122.850 seconds before a
247  * connection is killed.
248  */
249 static uint32_t rack_def_data_window = 20;
250 static uint32_t rack_goal_bdp = 2;
251 static uint32_t rack_min_srtts = 1;
252 static uint32_t rack_min_measure_usec = 0;
253 static int32_t rack_tlp_min = 10000;	/* 10ms */
254 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
255 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
256 static const int32_t rack_free_cache = 2;
257 static int32_t rack_hptsi_segments = 40;
258 static int32_t rack_rate_sample_method = USE_RTT_LOW;
259 static int32_t rack_pace_every_seg = 0;
260 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
261 static int32_t rack_slot_reduction = 4;
262 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
263 static int32_t rack_cwnd_block_ends_measure = 0;
264 static int32_t rack_rwnd_block_ends_measure = 0;
265 static int32_t rack_def_profile = 0;
266 
267 static int32_t rack_lower_cwnd_at_tlp = 0;
268 static int32_t rack_limited_retran = 0;
269 static int32_t rack_always_send_oldest = 0;
270 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
271 
272 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
273 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
274 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
275 
276 /* Probertt */
277 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
278 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
279 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
280 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
281 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
282 
283 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
284 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
285 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
286 static uint32_t rack_probertt_use_min_rtt_exit = 0;
287 static uint32_t rack_probe_rtt_sets_cwnd = 0;
288 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
289 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
290 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
291 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
292 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
293 static uint32_t rack_probertt_filter_life = 10000000;
294 static uint32_t rack_probertt_lower_within = 10;
295 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
296 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
297 static int32_t rack_probertt_clear_is = 1;
298 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
299 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
300 
301 /* Part of pacing */
302 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
303 
304 /* Timely information */
305 /* Combine these two gives the range of 'no change' to bw */
306 /* ie the up/down provide the upper and lower bound */
307 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
308 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
309 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
310 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
311 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
312 static int32_t rack_gp_decrease_per = 20;	/* 20% decrease in multipler */
313 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multipler */
314 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
315 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
316 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
317 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
318 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
319 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
320 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
321 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
322 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
323 static int32_t rack_use_max_for_nobackoff = 0;
324 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
325 static int32_t rack_timely_no_stopping = 0;
326 static int32_t rack_down_raise_thresh = 100;
327 static int32_t rack_req_segs = 1;
328 static uint64_t rack_bw_rate_cap = 0;
329 
330 /* Weird delayed ack mode */
331 static int32_t rack_use_imac_dack = 0;
332 /* Rack specific counters */
333 counter_u64_t rack_saw_enobuf;
334 counter_u64_t rack_saw_enobuf_hw;
335 counter_u64_t rack_saw_enetunreach;
336 counter_u64_t rack_persists_sends;
337 counter_u64_t rack_persists_acks;
338 counter_u64_t rack_persists_loss;
339 counter_u64_t rack_persists_lost_ends;
340 #ifdef INVARIANTS
341 counter_u64_t rack_adjust_map_bw;
342 #endif
343 /* Tail loss probe counters */
344 counter_u64_t rack_tlp_tot;
345 counter_u64_t rack_tlp_newdata;
346 counter_u64_t rack_tlp_retran;
347 counter_u64_t rack_tlp_retran_bytes;
348 counter_u64_t rack_to_tot;
349 counter_u64_t rack_hot_alloc;
350 counter_u64_t rack_to_alloc;
351 counter_u64_t rack_to_alloc_hard;
352 counter_u64_t rack_to_alloc_emerg;
353 counter_u64_t rack_to_alloc_limited;
354 counter_u64_t rack_alloc_limited_conns;
355 counter_u64_t rack_split_limited;
356 
357 counter_u64_t rack_multi_single_eq;
358 counter_u64_t rack_proc_non_comp_ack;
359 
360 counter_u64_t rack_fto_send;
361 counter_u64_t rack_fto_rsm_send;
362 counter_u64_t rack_nfto_resend;
363 counter_u64_t rack_non_fto_send;
364 counter_u64_t rack_extended_rfo;
365 
366 counter_u64_t rack_sack_proc_all;
367 counter_u64_t rack_sack_proc_short;
368 counter_u64_t rack_sack_proc_restart;
369 counter_u64_t rack_sack_attacks_detected;
370 counter_u64_t rack_sack_attacks_reversed;
371 counter_u64_t rack_sack_used_next_merge;
372 counter_u64_t rack_sack_splits;
373 counter_u64_t rack_sack_used_prev_merge;
374 counter_u64_t rack_sack_skipped_acked;
375 counter_u64_t rack_ack_total;
376 counter_u64_t rack_express_sack;
377 counter_u64_t rack_sack_total;
378 counter_u64_t rack_move_none;
379 counter_u64_t rack_move_some;
380 
381 counter_u64_t rack_input_idle_reduces;
382 counter_u64_t rack_collapsed_win;
383 counter_u64_t rack_try_scwnd;
384 counter_u64_t rack_hw_pace_init_fail;
385 counter_u64_t rack_hw_pace_lost;
386 
387 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
388 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
389 
390 
391 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
392 
393 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
394 	(tv) = (value) + slop;	 \
395 	if ((u_long)(tv) < (u_long)(tvmin)) \
396 		(tv) = (tvmin); \
397 	if ((u_long)(tv) > (u_long)(tvmax)) \
398 		(tv) = (tvmax); \
399 } while (0)
400 
401 static void
402 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
403 
404 static int
405 rack_process_ack(struct mbuf *m, struct tcphdr *th,
406     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
407     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
408 static int
409 rack_process_data(struct mbuf *m, struct tcphdr *th,
410     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
411     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
412 static void
413 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
414    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
415 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
416 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
417     uint8_t limit_type);
418 static struct rack_sendmap *
419 rack_check_recovery_mode(struct tcpcb *tp,
420     uint32_t tsused);
421 static void
422 rack_cong_signal(struct tcpcb *tp,
423 		 uint32_t type, uint32_t ack);
424 static void rack_counter_destroy(void);
425 static int
426 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt);
427 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
428 static void
429 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
430 static void
431 rack_do_segment(struct mbuf *m, struct tcphdr *th,
432     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
433     uint8_t iptos);
434 static void rack_dtor(void *mem, int32_t size, void *arg);
435 static void
436 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
437     uint32_t flex1, uint32_t flex2,
438     uint32_t flex3, uint32_t flex4,
439     uint32_t flex5, uint32_t flex6,
440     uint16_t flex7, uint8_t mod);
441 
442 static void
443 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
444    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
445    struct rack_sendmap *rsm, uint8_t quality);
446 static struct rack_sendmap *
447 rack_find_high_nonack(struct tcp_rack *rack,
448     struct rack_sendmap *rsm);
449 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
450 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
451 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
452 static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt);
453 static void
454 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
455 			    tcp_seq th_ack, int line, uint8_t quality);
456 static uint32_t
457 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
458 static int32_t rack_handoff_ok(struct tcpcb *tp);
459 static int32_t rack_init(struct tcpcb *tp);
460 static void rack_init_sysctls(void);
461 static void
462 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
463     struct tcphdr *th, int entered_rec, int dup_ack_struck);
464 static void
465 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
466     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
467     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
468 
469 static void
470 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
471     struct rack_sendmap *rsm);
472 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
473 static int32_t rack_output(struct tcpcb *tp);
474 
475 static uint32_t
476 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
477     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
478     uint32_t cts, int *moved_two);
479 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
480 static void rack_remxt_tmr(struct tcpcb *tp);
481 static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt);
482 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
483 static int32_t rack_stopall(struct tcpcb *tp);
484 static void
485 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
486     uint32_t delta);
487 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
488 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
489 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
490 static uint32_t
491 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
492     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
493 static void
494 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
495     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
496 static int
497 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
498     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
499 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
500 static int
501 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
502     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
503     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
504 static int
505 rack_do_closing(struct mbuf *m, struct tcphdr *th,
506     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
507     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
508 static int
509 rack_do_established(struct mbuf *m, struct tcphdr *th,
510     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
511     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
512 static int
513 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
514     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
515     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
516 static int
517 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
518     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
519     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
520 static int
521 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
522     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
523     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
524 static int
525 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
526     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
527     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
528 static int
529 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
530     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
531     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
532 static int
533 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
534     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
535     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
536 struct rack_sendmap *
537 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
538     uint32_t tsused);
539 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
540     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
541 static void
542      tcp_rack_partialack(struct tcpcb *tp);
543 static int
544 rack_set_profile(struct tcp_rack *rack, int prof);
545 static void
546 rack_apply_deferred_options(struct tcp_rack *rack);
547 
548 int32_t rack_clear_counter=0;
549 
550 static void
551 rack_set_cc_pacing(struct tcp_rack *rack)
552 {
553 	struct sockopt sopt;
554 	struct cc_newreno_opts opt;
555 	struct newreno old, *ptr;
556 	struct tcpcb *tp;
557 	int error;
558 
559 	if (rack->rc_pacing_cc_set)
560 		return;
561 
562 	tp = rack->rc_tp;
563 	if (tp->cc_algo == NULL) {
564 		/* Tcb is leaving */
565 		printf("No cc algorithm?\n");
566 		return;
567 	}
568 	rack->rc_pacing_cc_set = 1;
569 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
570 		/* Not new-reno we can't play games with beta! */
571 		goto out;
572 	}
573 	ptr = ((struct newreno *)tp->ccv->cc_data);
574 	if (CC_ALGO(tp)->ctl_output == NULL)  {
575 		/* Huh, why does new_reno no longer have a set function? */
576 		goto out;
577 	}
578 	if (ptr == NULL) {
579 		/* Just the default values */
580 		old.beta = V_newreno_beta_ecn;
581 		old.beta_ecn = V_newreno_beta_ecn;
582 		old.newreno_flags = 0;
583 	} else {
584 		old.beta = ptr->beta;
585 		old.beta_ecn = ptr->beta_ecn;
586 		old.newreno_flags = ptr->newreno_flags;
587 	}
588 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
589 	sopt.sopt_dir = SOPT_SET;
590 	opt.name = CC_NEWRENO_BETA;
591 	opt.val = rack->r_ctl.rc_saved_beta.beta;
592 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
593 	if (error)  {
594 		goto out;
595 	}
596 	/*
597 	 * Hack alert we need to set in our newreno_flags
598 	 * so that Abe behavior is also applied.
599 	 */
600 	((struct newreno *)tp->ccv->cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
601 	opt.name = CC_NEWRENO_BETA_ECN;
602 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
603 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
604 	if (error) {
605 		goto out;
606 	}
607 	/* Save off the original values for restoral */
608 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
609 out:
610 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
611 		union tcp_log_stackspecific log;
612 		struct timeval tv;
613 
614 		ptr = ((struct newreno *)tp->ccv->cc_data);
615 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
616 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
617 		if (ptr) {
618 			log.u_bbr.flex1 = ptr->beta;
619 			log.u_bbr.flex2 = ptr->beta_ecn;
620 			log.u_bbr.flex3 = ptr->newreno_flags;
621 		}
622 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
623 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
624 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
625 		log.u_bbr.flex7 = rack->gp_ready;
626 		log.u_bbr.flex7 <<= 1;
627 		log.u_bbr.flex7 |= rack->use_fixed_rate;
628 		log.u_bbr.flex7 <<= 1;
629 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
630 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
631 		log.u_bbr.flex8 = 3;
632 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
633 			       0, &log, false, NULL, NULL, 0, &tv);
634 	}
635 }
636 
637 static void
638 rack_undo_cc_pacing(struct tcp_rack *rack)
639 {
640 	struct newreno old, *ptr;
641 	struct tcpcb *tp;
642 
643 	if (rack->rc_pacing_cc_set == 0)
644 		return;
645 	tp = rack->rc_tp;
646 	rack->rc_pacing_cc_set = 0;
647 	if (tp->cc_algo == NULL)
648 		/* Tcb is leaving */
649 		return;
650 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
651 		/* Not new-reno nothing to do! */
652 		return;
653 	}
654 	ptr = ((struct newreno *)tp->ccv->cc_data);
655 	if (ptr == NULL) {
656 		/*
657 		 * This happens at rack_fini() if the
658 		 * cc module gets freed on us. In that
659 		 * case we loose our "new" settings but
660 		 * thats ok, since the tcb is going away anyway.
661 		 */
662 		return;
663 	}
664 	/* Grab out our set values */
665 	memcpy(&old, ptr, sizeof(struct newreno));
666 	/* Copy back in the original values */
667 	memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
668 	/* Now save back the values we had set in (for when pacing is restored) */
669 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
670 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
671 		union tcp_log_stackspecific log;
672 		struct timeval tv;
673 
674 		ptr = ((struct newreno *)tp->ccv->cc_data);
675 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
676 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
677 		log.u_bbr.flex1 = ptr->beta;
678 		log.u_bbr.flex2 = ptr->beta_ecn;
679 		log.u_bbr.flex3 = ptr->newreno_flags;
680 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
681 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
682 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
683 		log.u_bbr.flex7 = rack->gp_ready;
684 		log.u_bbr.flex7 <<= 1;
685 		log.u_bbr.flex7 |= rack->use_fixed_rate;
686 		log.u_bbr.flex7 <<= 1;
687 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
688 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
689 		log.u_bbr.flex8 = 4;
690 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
691 			       0, &log, false, NULL, NULL, 0, &tv);
692 	}
693 }
694 
695 #ifdef NETFLIX_PEAKRATE
696 static inline void
697 rack_update_peakrate_thr(struct tcpcb *tp)
698 {
699 	/* Keep in mind that t_maxpeakrate is in B/s. */
700 	uint64_t peak;
701 	peak = uqmax((tp->t_maxseg * 2),
702 		     (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
703 	tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
704 }
705 #endif
706 
707 static int
708 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
709 {
710 	uint32_t stat;
711 	int32_t error;
712 
713 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
714 	if (error || req->newptr == NULL)
715 		return error;
716 
717 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
718 	if (error)
719 		return (error);
720 	if (stat == 1) {
721 #ifdef INVARIANTS
722 		printf("Clearing RACK counters\n");
723 #endif
724 		counter_u64_zero(rack_tlp_tot);
725 		counter_u64_zero(rack_tlp_newdata);
726 		counter_u64_zero(rack_tlp_retran);
727 		counter_u64_zero(rack_tlp_retran_bytes);
728 		counter_u64_zero(rack_to_tot);
729 		counter_u64_zero(rack_saw_enobuf);
730 		counter_u64_zero(rack_saw_enobuf_hw);
731 		counter_u64_zero(rack_saw_enetunreach);
732 		counter_u64_zero(rack_persists_sends);
733 		counter_u64_zero(rack_persists_acks);
734 		counter_u64_zero(rack_persists_loss);
735 		counter_u64_zero(rack_persists_lost_ends);
736 #ifdef INVARIANTS
737 		counter_u64_zero(rack_adjust_map_bw);
738 #endif
739 		counter_u64_zero(rack_to_alloc_hard);
740 		counter_u64_zero(rack_to_alloc_emerg);
741 		counter_u64_zero(rack_sack_proc_all);
742 		counter_u64_zero(rack_fto_send);
743 		counter_u64_zero(rack_fto_rsm_send);
744 		counter_u64_zero(rack_extended_rfo);
745 		counter_u64_zero(rack_hw_pace_init_fail);
746 		counter_u64_zero(rack_hw_pace_lost);
747 		counter_u64_zero(rack_non_fto_send);
748 		counter_u64_zero(rack_nfto_resend);
749 		counter_u64_zero(rack_sack_proc_short);
750 		counter_u64_zero(rack_sack_proc_restart);
751 		counter_u64_zero(rack_to_alloc);
752 		counter_u64_zero(rack_to_alloc_limited);
753 		counter_u64_zero(rack_alloc_limited_conns);
754 		counter_u64_zero(rack_split_limited);
755 		counter_u64_zero(rack_multi_single_eq);
756 		counter_u64_zero(rack_proc_non_comp_ack);
757 		counter_u64_zero(rack_sack_attacks_detected);
758 		counter_u64_zero(rack_sack_attacks_reversed);
759 		counter_u64_zero(rack_sack_used_next_merge);
760 		counter_u64_zero(rack_sack_used_prev_merge);
761 		counter_u64_zero(rack_sack_splits);
762 		counter_u64_zero(rack_sack_skipped_acked);
763 		counter_u64_zero(rack_ack_total);
764 		counter_u64_zero(rack_express_sack);
765 		counter_u64_zero(rack_sack_total);
766 		counter_u64_zero(rack_move_none);
767 		counter_u64_zero(rack_move_some);
768 		counter_u64_zero(rack_try_scwnd);
769 		counter_u64_zero(rack_collapsed_win);
770 	}
771 	rack_clear_counter = 0;
772 	return (0);
773 }
774 
775 static void
776 rack_init_sysctls(void)
777 {
778 	struct sysctl_oid *rack_counters;
779 	struct sysctl_oid *rack_attack;
780 	struct sysctl_oid *rack_pacing;
781 	struct sysctl_oid *rack_timely;
782 	struct sysctl_oid *rack_timers;
783 	struct sysctl_oid *rack_tlp;
784 	struct sysctl_oid *rack_misc;
785 	struct sysctl_oid *rack_features;
786 	struct sysctl_oid *rack_measure;
787 	struct sysctl_oid *rack_probertt;
788 	struct sysctl_oid *rack_hw_pacing;
789 
790 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
791 	    SYSCTL_CHILDREN(rack_sysctl_root),
792 	    OID_AUTO,
793 	    "sack_attack",
794 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
795 	    "Rack Sack Attack Counters and Controls");
796 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
797 	    SYSCTL_CHILDREN(rack_sysctl_root),
798 	    OID_AUTO,
799 	    "stats",
800 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
801 	    "Rack Counters");
802 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
803 	    SYSCTL_CHILDREN(rack_sysctl_root),
804 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
805 	    &rack_rate_sample_method , USE_RTT_LOW,
806 	    "What method should we use for rate sampling 0=high, 1=low ");
807 	/* Probe rtt related controls */
808 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
809 	    SYSCTL_CHILDREN(rack_sysctl_root),
810 	    OID_AUTO,
811 	    "probertt",
812 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
813 	    "ProbeRTT related Controls");
814 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
815 	    SYSCTL_CHILDREN(rack_probertt),
816 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
817 	    &rack_atexit_prtt_hbp, 130,
818 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
819 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
820 	    SYSCTL_CHILDREN(rack_probertt),
821 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
822 	    &rack_atexit_prtt, 130,
823 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
824 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
825 	    SYSCTL_CHILDREN(rack_probertt),
826 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
827 	    &rack_per_of_gp_probertt, 60,
828 	    "What percentage of goodput do we pace at in probertt");
829 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
830 	    SYSCTL_CHILDREN(rack_probertt),
831 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
832 	    &rack_per_of_gp_probertt_reduce, 10,
833 	    "What percentage of goodput do we reduce every gp_srtt");
834 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
835 	    SYSCTL_CHILDREN(rack_probertt),
836 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
837 	    &rack_per_of_gp_lowthresh, 40,
838 	    "What percentage of goodput do we allow the multiplier to fall to");
839 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
840 	    SYSCTL_CHILDREN(rack_probertt),
841 	    OID_AUTO, "time_between", CTLFLAG_RW,
842 	    & rack_time_between_probertt, 96000000,
843 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
844 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
845 	    SYSCTL_CHILDREN(rack_probertt),
846 	    OID_AUTO, "safety", CTLFLAG_RW,
847 	    &rack_probe_rtt_safety_val, 2000000,
848 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
849 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
850 	    SYSCTL_CHILDREN(rack_probertt),
851 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
852 	    &rack_probe_rtt_sets_cwnd, 0,
853 	    "Do we set the cwnd too (if always_lower is on)");
854 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
855 	    SYSCTL_CHILDREN(rack_probertt),
856 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
857 	    &rack_max_drain_wait, 2,
858 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
859 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
860 	    SYSCTL_CHILDREN(rack_probertt),
861 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
862 	    &rack_must_drain, 1,
863 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
864 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
865 	    SYSCTL_CHILDREN(rack_probertt),
866 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
867 	    &rack_probertt_use_min_rtt_entry, 1,
868 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
869 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
870 	    SYSCTL_CHILDREN(rack_probertt),
871 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
872 	    &rack_probertt_use_min_rtt_exit, 0,
873 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
874 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
875 	    SYSCTL_CHILDREN(rack_probertt),
876 	    OID_AUTO, "length_div", CTLFLAG_RW,
877 	    &rack_probertt_gpsrtt_cnt_div, 0,
878 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
879 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
880 	    SYSCTL_CHILDREN(rack_probertt),
881 	    OID_AUTO, "length_mul", CTLFLAG_RW,
882 	    &rack_probertt_gpsrtt_cnt_mul, 0,
883 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
884 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
885 	    SYSCTL_CHILDREN(rack_probertt),
886 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
887 	    &rack_min_probertt_hold, 200000,
888 	    "What is the minimum time we hold probertt at target");
889 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
890 	    SYSCTL_CHILDREN(rack_probertt),
891 	    OID_AUTO, "filter_life", CTLFLAG_RW,
892 	    &rack_probertt_filter_life, 10000000,
893 	    "What is the time for the filters life in useconds");
894 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
895 	    SYSCTL_CHILDREN(rack_probertt),
896 	    OID_AUTO, "lower_within", CTLFLAG_RW,
897 	    &rack_probertt_lower_within, 10,
898 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
899 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
900 	    SYSCTL_CHILDREN(rack_probertt),
901 	    OID_AUTO, "must_move", CTLFLAG_RW,
902 	    &rack_min_rtt_movement, 250,
903 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
904 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
905 	    SYSCTL_CHILDREN(rack_probertt),
906 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
907 	    &rack_probertt_clear_is, 1,
908 	    "Do we clear I/S counts on exiting probe-rtt");
909 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
910 	    SYSCTL_CHILDREN(rack_probertt),
911 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
912 	    &rack_max_drain_hbp, 1,
913 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
914 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
915 	    SYSCTL_CHILDREN(rack_probertt),
916 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
917 	    &rack_hbp_thresh, 3,
918 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
919 	/* Pacing related sysctls */
920 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
921 	    SYSCTL_CHILDREN(rack_sysctl_root),
922 	    OID_AUTO,
923 	    "pacing",
924 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
925 	    "Pacing related Controls");
926 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
927 	    SYSCTL_CHILDREN(rack_pacing),
928 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
929 	    &rack_max_per_above, 30,
930 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
931 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
932 	    SYSCTL_CHILDREN(rack_pacing),
933 	    OID_AUTO, "pace_to_one", CTLFLAG_RW,
934 	    &rack_pace_one_seg, 0,
935 	    "Do we allow low b/w pacing of 1MSS instead of two");
936 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
937 	    SYSCTL_CHILDREN(rack_pacing),
938 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
939 	    &rack_limit_time_with_srtt, 0,
940 	    "Do we limit pacing time based on srtt");
941 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
942 	    SYSCTL_CHILDREN(rack_pacing),
943 	    OID_AUTO, "init_win", CTLFLAG_RW,
944 	    &rack_default_init_window, 0,
945 	    "Do we have a rack initial window 0 = system default");
946 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
947 	    SYSCTL_CHILDREN(rack_pacing),
948 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
949 	    &rack_per_of_gp_ss, 250,
950 	    "If non zero, what percentage of goodput to pace at in slow start");
951 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
952 	    SYSCTL_CHILDREN(rack_pacing),
953 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
954 	    &rack_per_of_gp_ca, 150,
955 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
956 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
957 	    SYSCTL_CHILDREN(rack_pacing),
958 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
959 	    &rack_per_of_gp_rec, 200,
960 	    "If non zero, what percentage of goodput to pace at in recovery");
961 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
962 	    SYSCTL_CHILDREN(rack_pacing),
963 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
964 	    &rack_hptsi_segments, 40,
965 	    "What size is the max for TSO segments in pacing and burst mitigation");
966 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
967 	    SYSCTL_CHILDREN(rack_pacing),
968 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
969 	    &rack_slot_reduction, 4,
970 	    "When doing only burst mitigation what is the reduce divisor");
971 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
972 	    SYSCTL_CHILDREN(rack_sysctl_root),
973 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
974 	    &rack_pace_every_seg, 0,
975 	    "If set we use pacing, if clear we use only the original burst mitigation");
976 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
977 	    SYSCTL_CHILDREN(rack_pacing),
978 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
979 	    &rack_bw_rate_cap, 0,
980 	    "If set we apply this value to the absolute rate cap used by pacing");
981 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
982 	    SYSCTL_CHILDREN(rack_sysctl_root),
983 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
984 	    &rack_req_measurements, 1,
985 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
986 	/* Hardware pacing */
987 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
988 	    SYSCTL_CHILDREN(rack_sysctl_root),
989 	    OID_AUTO,
990 	    "hdwr_pacing",
991 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
992 	    "Pacing related Controls");
993 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
994 	    SYSCTL_CHILDREN(rack_hw_pacing),
995 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
996 	    &rack_hw_rwnd_factor, 2,
997 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
998 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
999 	    SYSCTL_CHILDREN(rack_hw_pacing),
1000 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1001 	    &rack_enobuf_hw_boost_mult, 2,
1002 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1003 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1004 	    SYSCTL_CHILDREN(rack_hw_pacing),
1005 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1006 	    &rack_enobuf_hw_max, 2,
1007 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1008 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1009 	    SYSCTL_CHILDREN(rack_hw_pacing),
1010 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1011 	    &rack_enobuf_hw_min, 2,
1012 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1013 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1014 	    SYSCTL_CHILDREN(rack_hw_pacing),
1015 	    OID_AUTO, "enable", CTLFLAG_RW,
1016 	    &rack_enable_hw_pacing, 0,
1017 	    "Should RACK attempt to use hw pacing?");
1018 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1019 	    SYSCTL_CHILDREN(rack_hw_pacing),
1020 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1021 	    &rack_hw_rate_caps, 1,
1022 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1023 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1024 	    SYSCTL_CHILDREN(rack_hw_pacing),
1025 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1026 	    &rack_hw_rate_min, 0,
1027 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1028 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1029 	    SYSCTL_CHILDREN(rack_hw_pacing),
1030 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1031 	    &rack_hw_rate_to_low, 0,
1032 	    "If we fall below this rate, dis-engage hw pacing?");
1033 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1034 	    SYSCTL_CHILDREN(rack_hw_pacing),
1035 	    OID_AUTO, "up_only", CTLFLAG_RW,
1036 	    &rack_hw_up_only, 1,
1037 	    "Do we allow hw pacing to lower the rate selected?");
1038 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1039 	    SYSCTL_CHILDREN(rack_hw_pacing),
1040 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1041 	    &rack_hw_pace_extra_slots, 2,
1042 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1043 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1044 	    SYSCTL_CHILDREN(rack_sysctl_root),
1045 	    OID_AUTO,
1046 	    "timely",
1047 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1048 	    "Rack Timely RTT Controls");
1049 	/* Timely based GP dynmics */
1050 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1051 	    SYSCTL_CHILDREN(rack_timely),
1052 	    OID_AUTO, "upper", CTLFLAG_RW,
1053 	    &rack_gp_per_bw_mul_up, 2,
1054 	    "Rack timely upper range for equal b/w (in percentage)");
1055 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1056 	    SYSCTL_CHILDREN(rack_timely),
1057 	    OID_AUTO, "lower", CTLFLAG_RW,
1058 	    &rack_gp_per_bw_mul_down, 4,
1059 	    "Rack timely lower range for equal b/w (in percentage)");
1060 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1061 	    SYSCTL_CHILDREN(rack_timely),
1062 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1063 	    &rack_gp_rtt_maxmul, 3,
1064 	    "Rack timely multipler of lowest rtt for rtt_max");
1065 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1066 	    SYSCTL_CHILDREN(rack_timely),
1067 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1068 	    &rack_gp_rtt_mindiv, 4,
1069 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1070 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1071 	    SYSCTL_CHILDREN(rack_timely),
1072 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1073 	    &rack_gp_rtt_minmul, 1,
1074 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1075 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1076 	    SYSCTL_CHILDREN(rack_timely),
1077 	    OID_AUTO, "decrease", CTLFLAG_RW,
1078 	    &rack_gp_decrease_per, 20,
1079 	    "Rack timely decrease percentage of our GP multiplication factor");
1080 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1081 	    SYSCTL_CHILDREN(rack_timely),
1082 	    OID_AUTO, "increase", CTLFLAG_RW,
1083 	    &rack_gp_increase_per, 2,
1084 	    "Rack timely increase perentage of our GP multiplication factor");
1085 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1086 	    SYSCTL_CHILDREN(rack_timely),
1087 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1088 	    &rack_per_lower_bound, 50,
1089 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1090 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1091 	    SYSCTL_CHILDREN(rack_timely),
1092 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1093 	    &rack_per_upper_bound_ss, 0,
1094 	    "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1095 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1096 	    SYSCTL_CHILDREN(rack_timely),
1097 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1098 	    &rack_per_upper_bound_ca, 0,
1099 	    "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1100 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1101 	    SYSCTL_CHILDREN(rack_timely),
1102 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1103 	    &rack_do_dyn_mul, 0,
1104 	    "Rack timely do we enable dynmaic timely goodput by default");
1105 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1106 	    SYSCTL_CHILDREN(rack_timely),
1107 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1108 	    &rack_gp_no_rec_chg, 1,
1109 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1110 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1111 	    SYSCTL_CHILDREN(rack_timely),
1112 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1113 	    &rack_timely_dec_clear, 6,
1114 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1115 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1116 	    SYSCTL_CHILDREN(rack_timely),
1117 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1118 	    &rack_timely_max_push_rise, 3,
1119 	    "Rack timely how many times do we push up with b/w increase");
1120 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1121 	    SYSCTL_CHILDREN(rack_timely),
1122 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1123 	    &rack_timely_max_push_drop, 3,
1124 	    "Rack timely how many times do we push back on b/w decent");
1125 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1126 	    SYSCTL_CHILDREN(rack_timely),
1127 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1128 	    &rack_timely_min_segs, 4,
1129 	    "Rack timely when setting the cwnd what is the min num segments");
1130 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1131 	    SYSCTL_CHILDREN(rack_timely),
1132 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1133 	    &rack_use_max_for_nobackoff, 0,
1134 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1135 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1136 	    SYSCTL_CHILDREN(rack_timely),
1137 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1138 	    &rack_timely_int_timely_only, 0,
1139 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1140 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1141 	    SYSCTL_CHILDREN(rack_timely),
1142 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1143 	    &rack_timely_no_stopping, 0,
1144 	    "Rack timely don't stop increase");
1145 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1146 	    SYSCTL_CHILDREN(rack_timely),
1147 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1148 	    &rack_down_raise_thresh, 100,
1149 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1150 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1151 	    SYSCTL_CHILDREN(rack_timely),
1152 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1153 	    &rack_req_segs, 1,
1154 	    "Bottom dragging if not these many segments outstanding and room");
1155 
1156 	/* TLP and Rack related parameters */
1157 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1158 	    SYSCTL_CHILDREN(rack_sysctl_root),
1159 	    OID_AUTO,
1160 	    "tlp",
1161 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1162 	    "TLP and Rack related Controls");
1163 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1164 	    SYSCTL_CHILDREN(rack_tlp),
1165 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1166 	    &use_rack_rr, 1,
1167 	    "Do we use Rack Rapid Recovery");
1168 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1169 	    SYSCTL_CHILDREN(rack_tlp),
1170 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1171 	    &rack_max_abc_post_recovery, 2,
1172 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1173 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1174 	    SYSCTL_CHILDREN(rack_tlp),
1175 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1176 	    &rack_non_rxt_use_cr, 0,
1177 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1178 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1179 	    SYSCTL_CHILDREN(rack_tlp),
1180 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1181 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1182 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1183 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1184 	    SYSCTL_CHILDREN(rack_tlp),
1185 	    OID_AUTO, "limit", CTLFLAG_RW,
1186 	    &rack_tlp_limit, 2,
1187 	    "How many TLP's can be sent without sending new data");
1188 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1189 	    SYSCTL_CHILDREN(rack_tlp),
1190 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1191 	    &rack_tlp_use_greater, 1,
1192 	    "Should we use the rack_rtt time if its greater than srtt");
1193 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1194 	    SYSCTL_CHILDREN(rack_tlp),
1195 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1196 	    &rack_tlp_min, 10000,
1197 	    "TLP minimum timeout per the specification (in microseconds)");
1198 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1199 	    SYSCTL_CHILDREN(rack_tlp),
1200 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1201 	    &rack_always_send_oldest, 0,
1202 	    "Should we always send the oldest TLP and RACK-TLP");
1203 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1204 	    SYSCTL_CHILDREN(rack_tlp),
1205 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1206 	    &rack_limited_retran, 0,
1207 	    "How many times can a rack timeout drive out sends");
1208 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1209 	    SYSCTL_CHILDREN(rack_tlp),
1210 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1211 	    &rack_lower_cwnd_at_tlp, 0,
1212 	    "When a TLP completes a retran should we enter recovery");
1213 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1214 	    SYSCTL_CHILDREN(rack_tlp),
1215 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1216 	    &rack_reorder_thresh, 2,
1217 	    "What factor for rack will be added when seeing reordering (shift right)");
1218 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1219 	    SYSCTL_CHILDREN(rack_tlp),
1220 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1221 	    &rack_tlp_thresh, 1,
1222 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1223 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1224 	    SYSCTL_CHILDREN(rack_tlp),
1225 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1226 	    &rack_reorder_fade, 60000000,
1227 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1228 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229 	    SYSCTL_CHILDREN(rack_tlp),
1230 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1231 	    &rack_pkt_delay, 1000,
1232 	    "Extra RACK time (in microseconds) besides reordering thresh");
1233 
1234 	/* Timer related controls */
1235 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1236 	    SYSCTL_CHILDREN(rack_sysctl_root),
1237 	    OID_AUTO,
1238 	    "timers",
1239 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1240 	    "Timer related controls");
1241 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1242 	    SYSCTL_CHILDREN(rack_timers),
1243 	    OID_AUTO, "persmin", CTLFLAG_RW,
1244 	    &rack_persist_min, 250000,
1245 	    "What is the minimum time in microseconds between persists");
1246 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1247 	    SYSCTL_CHILDREN(rack_timers),
1248 	    OID_AUTO, "persmax", CTLFLAG_RW,
1249 	    &rack_persist_max, 2000000,
1250 	    "What is the largest delay in microseconds between persists");
1251 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1252 	    SYSCTL_CHILDREN(rack_timers),
1253 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1254 	    &rack_delayed_ack_time, 40000,
1255 	    "Delayed ack time (40ms in microseconds)");
1256 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1257 	    SYSCTL_CHILDREN(rack_timers),
1258 	    OID_AUTO, "minrto", CTLFLAG_RW,
1259 	    &rack_rto_min, 30000,
1260 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1261 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1262 	    SYSCTL_CHILDREN(rack_timers),
1263 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1264 	    &rack_rto_max, 4000000,
1265 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1266 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1267 	    SYSCTL_CHILDREN(rack_timers),
1268 	    OID_AUTO, "minto", CTLFLAG_RW,
1269 	    &rack_min_to, 1000,
1270 	    "Minimum rack timeout in microseconds");
1271 	/* Measure controls */
1272 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1273 	    SYSCTL_CHILDREN(rack_sysctl_root),
1274 	    OID_AUTO,
1275 	    "measure",
1276 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1277 	    "Measure related controls");
1278 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1279 	    SYSCTL_CHILDREN(rack_measure),
1280 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1281 	    &rack_wma_divisor, 8,
1282 	    "When doing b/w calculation what is the  divisor for the WMA");
1283 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1284 	    SYSCTL_CHILDREN(rack_measure),
1285 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1286 	    &rack_cwnd_block_ends_measure, 0,
1287 	    "Does a cwnd just-return end the measurement window (app limited)");
1288 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1289 	    SYSCTL_CHILDREN(rack_measure),
1290 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1291 	    &rack_rwnd_block_ends_measure, 0,
1292 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1293 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1294 	    SYSCTL_CHILDREN(rack_measure),
1295 	    OID_AUTO, "min_target", CTLFLAG_RW,
1296 	    &rack_def_data_window, 20,
1297 	    "What is the minimum target window (in mss) for a GP measurements");
1298 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1299 	    SYSCTL_CHILDREN(rack_measure),
1300 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1301 	    &rack_goal_bdp, 2,
1302 	    "What is the goal BDP to measure");
1303 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1304 	    SYSCTL_CHILDREN(rack_measure),
1305 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1306 	    &rack_min_srtts, 1,
1307 	    "What is the goal BDP to measure");
1308 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1309 	    SYSCTL_CHILDREN(rack_measure),
1310 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1311 	    &rack_min_measure_usec, 0,
1312 	    "What is the Minimum time time for a measurement if 0, this is off");
1313 	/* Features */
1314 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1315 	    SYSCTL_CHILDREN(rack_sysctl_root),
1316 	    OID_AUTO,
1317 	    "features",
1318 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1319 	    "Feature controls");
1320 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1321 	    SYSCTL_CHILDREN(rack_features),
1322 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1323 	    &rack_use_cmp_acks, 1,
1324 	    "Should RACK have LRO send compressed acks");
1325 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1326 	    SYSCTL_CHILDREN(rack_features),
1327 	    OID_AUTO, "fsb", CTLFLAG_RW,
1328 	    &rack_use_fsb, 1,
1329 	    "Should RACK use the fast send block?");
1330 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1331 	    SYSCTL_CHILDREN(rack_features),
1332 	    OID_AUTO, "rfo", CTLFLAG_RW,
1333 	    &rack_use_rfo, 1,
1334 	    "Should RACK use rack_fast_output()?");
1335 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1336 	    SYSCTL_CHILDREN(rack_features),
1337 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1338 	    &rack_use_rsm_rfo, 1,
1339 	    "Should RACK use rack_fast_rsm_output()?");
1340 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1341 	    SYSCTL_CHILDREN(rack_features),
1342 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1343 	    &rack_enable_mqueue_for_nonpaced, 0,
1344 	    "Should RACK use mbuf queuing for non-paced connections");
1345 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1346 	    SYSCTL_CHILDREN(rack_features),
1347 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1348 	    &rack_do_hystart, 0,
1349 	    "Should RACK enable HyStart++ on connections?");
1350 	/* Misc rack controls */
1351 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1352 	    SYSCTL_CHILDREN(rack_sysctl_root),
1353 	    OID_AUTO,
1354 	    "misc",
1355 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1356 	    "Misc related controls");
1357 #ifdef TCP_ACCOUNTING
1358 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1359 	    SYSCTL_CHILDREN(rack_misc),
1360 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1361 	    &rack_tcp_accounting, 0,
1362 	    "Should we turn on TCP accounting for all rack sessions?");
1363 #endif
1364 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1365 	    SYSCTL_CHILDREN(rack_misc),
1366 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1367 	    &rack_apply_rtt_with_reduced_conf, 0,
1368 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1369 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1370 	    SYSCTL_CHILDREN(rack_misc),
1371 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1372 	    &rack_dsack_std_based, 3,
1373 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1374 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1375 	    SYSCTL_CHILDREN(rack_misc),
1376 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1377 	    &rack_prr_addbackmax, 2,
1378 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1379 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1380 	    SYSCTL_CHILDREN(rack_misc),
1381 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1382 	    &rack_stats_gets_ms_rtt, 1,
1383 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1384 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1385 	    SYSCTL_CHILDREN(rack_misc),
1386 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1387 	    &rack_client_low_buf, 0,
1388 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1389 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1390 	    SYSCTL_CHILDREN(rack_misc),
1391 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1392 	    &rack_def_profile, 0,
1393 	    "Should RACK use a default profile (0=no, num == profile num)?");
1394 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1395 	    SYSCTL_CHILDREN(rack_misc),
1396 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1397 	    &rack_enable_shared_cwnd, 1,
1398 	    "Should RACK try to use the shared cwnd on connections where allowed");
1399 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1400 	    SYSCTL_CHILDREN(rack_misc),
1401 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1402 	    &rack_limits_scwnd, 1,
1403 	    "Should RACK place low end time limits on the shared cwnd feature");
1404 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1405 	    SYSCTL_CHILDREN(rack_misc),
1406 	    OID_AUTO, "iMac_dack", CTLFLAG_RW,
1407 	    &rack_use_imac_dack, 0,
1408 	    "Should RACK try to emulate iMac delayed ack");
1409 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1410 	    SYSCTL_CHILDREN(rack_misc),
1411 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1412 	    &rack_disable_prr, 0,
1413 	    "Should RACK not use prr and only pace (must have pacing on)");
1414 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1415 	    SYSCTL_CHILDREN(rack_misc),
1416 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1417 	    &rack_verbose_logging, 0,
1418 	    "Should RACK black box logging be verbose");
1419 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1420 	    SYSCTL_CHILDREN(rack_misc),
1421 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1422 	    &rack_ignore_data_after_close, 1,
1423 	    "Do we hold off sending a RST until all pending data is ack'd");
1424 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1425 	    SYSCTL_CHILDREN(rack_misc),
1426 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1427 	    &rack_sack_not_required, 1,
1428 	    "Do we allow rack to run on connections not supporting SACK");
1429 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1430 	    SYSCTL_CHILDREN(rack_misc),
1431 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1432 	    &rack_send_a_lot_in_prr, 1,
1433 	    "Send a lot in prr");
1434 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1435 	    SYSCTL_CHILDREN(rack_misc),
1436 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1437 	    &rack_autosndbuf_inc, 20,
1438 	    "What percentage should rack scale up its snd buffer by?");
1439 	/* Sack Attacker detection stuff */
1440 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1441 	    SYSCTL_CHILDREN(rack_attack),
1442 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1443 	    &rack_highest_sack_thresh_seen, 0,
1444 	    "Highest sack to ack ratio seen");
1445 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1446 	    SYSCTL_CHILDREN(rack_attack),
1447 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1448 	    &rack_highest_move_thresh_seen, 0,
1449 	    "Highest move to non-move ratio seen");
1450 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1451 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1452 	    SYSCTL_CHILDREN(rack_attack),
1453 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1454 	    &rack_ack_total,
1455 	    "Total number of Ack's");
1456 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1457 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1458 	    SYSCTL_CHILDREN(rack_attack),
1459 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1460 	    &rack_express_sack,
1461 	    "Total expresss number of Sack's");
1462 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1463 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1464 	    SYSCTL_CHILDREN(rack_attack),
1465 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1466 	    &rack_sack_total,
1467 	    "Total number of SACKs");
1468 	rack_move_none = counter_u64_alloc(M_WAITOK);
1469 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1470 	    SYSCTL_CHILDREN(rack_attack),
1471 	    OID_AUTO, "move_none", CTLFLAG_RD,
1472 	    &rack_move_none,
1473 	    "Total number of SACK index reuse of postions under threshold");
1474 	rack_move_some = counter_u64_alloc(M_WAITOK);
1475 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1476 	    SYSCTL_CHILDREN(rack_attack),
1477 	    OID_AUTO, "move_some", CTLFLAG_RD,
1478 	    &rack_move_some,
1479 	    "Total number of SACK index reuse of postions over threshold");
1480 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1481 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1482 	    SYSCTL_CHILDREN(rack_attack),
1483 	    OID_AUTO, "attacks", CTLFLAG_RD,
1484 	    &rack_sack_attacks_detected,
1485 	    "Total number of SACK attackers that had sack disabled");
1486 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1487 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1488 	    SYSCTL_CHILDREN(rack_attack),
1489 	    OID_AUTO, "reversed", CTLFLAG_RD,
1490 	    &rack_sack_attacks_reversed,
1491 	    "Total number of SACK attackers that were later determined false positive");
1492 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1493 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1494 	    SYSCTL_CHILDREN(rack_attack),
1495 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1496 	    &rack_sack_used_next_merge,
1497 	    "Total number of times we used the next merge");
1498 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1499 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1500 	    SYSCTL_CHILDREN(rack_attack),
1501 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1502 	    &rack_sack_used_prev_merge,
1503 	    "Total number of times we used the prev merge");
1504 	/* Counters */
1505 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1506 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1507 	    SYSCTL_CHILDREN(rack_counters),
1508 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1509 	    &rack_fto_send, "Total number of rack_fast_output sends");
1510 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1511 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1512 	    SYSCTL_CHILDREN(rack_counters),
1513 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1514 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1515 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1516 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1517 	    SYSCTL_CHILDREN(rack_counters),
1518 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1519 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1520 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1521 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1522 	    SYSCTL_CHILDREN(rack_counters),
1523 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1524 	    &rack_non_fto_send, "Total number of rack_output first sends");
1525 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1526 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1527 	    SYSCTL_CHILDREN(rack_counters),
1528 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1529 	    &rack_extended_rfo, "Total number of times we extended rfo");
1530 
1531 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1532 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1533 	    SYSCTL_CHILDREN(rack_counters),
1534 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1535 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1536 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1537 
1538 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1539 	    SYSCTL_CHILDREN(rack_counters),
1540 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1541 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1542 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1543 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1544 	    SYSCTL_CHILDREN(rack_counters),
1545 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1546 	    &rack_tlp_tot,
1547 	    "Total number of tail loss probe expirations");
1548 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1549 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1550 	    SYSCTL_CHILDREN(rack_counters),
1551 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1552 	    &rack_tlp_newdata,
1553 	    "Total number of tail loss probe sending new data");
1554 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1555 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1556 	    SYSCTL_CHILDREN(rack_counters),
1557 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1558 	    &rack_tlp_retran,
1559 	    "Total number of tail loss probe sending retransmitted data");
1560 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1561 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1562 	    SYSCTL_CHILDREN(rack_counters),
1563 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1564 	    &rack_tlp_retran_bytes,
1565 	    "Total bytes of tail loss probe sending retransmitted data");
1566 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1567 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1568 	    SYSCTL_CHILDREN(rack_counters),
1569 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1570 	    &rack_to_tot,
1571 	    "Total number of times the rack to expired");
1572 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1573 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1574 	    SYSCTL_CHILDREN(rack_counters),
1575 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1576 	    &rack_saw_enobuf,
1577 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1578 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1579 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1580 	    SYSCTL_CHILDREN(rack_counters),
1581 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1582 	    &rack_saw_enobuf_hw,
1583 	    "Total number of times a send returned enobuf for hdwr paced connections");
1584 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1585 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1586 	    SYSCTL_CHILDREN(rack_counters),
1587 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1588 	    &rack_saw_enetunreach,
1589 	    "Total number of times a send received a enetunreachable");
1590 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1591 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1592 	    SYSCTL_CHILDREN(rack_counters),
1593 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1594 	    &rack_hot_alloc,
1595 	    "Total allocations from the top of our list");
1596 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1597 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1598 	    SYSCTL_CHILDREN(rack_counters),
1599 	    OID_AUTO, "allocs", CTLFLAG_RD,
1600 	    &rack_to_alloc,
1601 	    "Total allocations of tracking structures");
1602 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1603 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1604 	    SYSCTL_CHILDREN(rack_counters),
1605 	    OID_AUTO, "allochard", CTLFLAG_RD,
1606 	    &rack_to_alloc_hard,
1607 	    "Total allocations done with sleeping the hard way");
1608 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1609 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1610 	    SYSCTL_CHILDREN(rack_counters),
1611 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1612 	    &rack_to_alloc_emerg,
1613 	    "Total allocations done from emergency cache");
1614 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1615 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1616 	    SYSCTL_CHILDREN(rack_counters),
1617 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1618 	    &rack_to_alloc_limited,
1619 	    "Total allocations dropped due to limit");
1620 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1621 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1622 	    SYSCTL_CHILDREN(rack_counters),
1623 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1624 	    &rack_alloc_limited_conns,
1625 	    "Connections with allocations dropped due to limit");
1626 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1627 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1628 	    SYSCTL_CHILDREN(rack_counters),
1629 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1630 	    &rack_split_limited,
1631 	    "Split allocations dropped due to limit");
1632 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1633 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1634 	    SYSCTL_CHILDREN(rack_counters),
1635 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1636 	    &rack_persists_sends,
1637 	    "Number of times we sent a persist probe");
1638 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1639 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1640 	    SYSCTL_CHILDREN(rack_counters),
1641 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1642 	    &rack_persists_acks,
1643 	    "Number of times a persist probe was acked");
1644 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1645 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1646 	    SYSCTL_CHILDREN(rack_counters),
1647 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1648 	    &rack_persists_loss,
1649 	    "Number of times we detected a lost persist probe (no ack)");
1650 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1651 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1652 	    SYSCTL_CHILDREN(rack_counters),
1653 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1654 	    &rack_persists_lost_ends,
1655 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1656 #ifdef INVARIANTS
1657 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1658 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1659 	    SYSCTL_CHILDREN(rack_counters),
1660 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1661 	    &rack_adjust_map_bw,
1662 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1663 #endif
1664 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1665 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1666 	    SYSCTL_CHILDREN(rack_counters),
1667 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1668 	    &rack_multi_single_eq,
1669 	    "Number of compressed acks total represented");
1670 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1671 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1672 	    SYSCTL_CHILDREN(rack_counters),
1673 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1674 	    &rack_proc_non_comp_ack,
1675 	    "Number of non compresseds acks that we processed");
1676 
1677 
1678 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1679 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680 	    SYSCTL_CHILDREN(rack_counters),
1681 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1682 	    &rack_sack_proc_all,
1683 	    "Total times we had to walk whole list for sack processing");
1684 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1685 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1686 	    SYSCTL_CHILDREN(rack_counters),
1687 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1688 	    &rack_sack_proc_restart,
1689 	    "Total times we had to walk whole list due to a restart");
1690 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1691 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1692 	    SYSCTL_CHILDREN(rack_counters),
1693 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1694 	    &rack_sack_proc_short,
1695 	    "Total times we took shortcut for sack processing");
1696 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1697 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1698 	    SYSCTL_CHILDREN(rack_attack),
1699 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1700 	    &rack_sack_skipped_acked,
1701 	    "Total number of times we skipped previously sacked");
1702 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1703 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1704 	    SYSCTL_CHILDREN(rack_attack),
1705 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1706 	    &rack_sack_splits,
1707 	    "Total number of times we did the old fashion tree split");
1708 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1709 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1710 	    SYSCTL_CHILDREN(rack_counters),
1711 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1712 	    &rack_input_idle_reduces,
1713 	    "Total number of idle reductions on input");
1714 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1715 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1716 	    SYSCTL_CHILDREN(rack_counters),
1717 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1718 	    &rack_collapsed_win,
1719 	    "Total number of collapsed windows");
1720 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1721 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1722 	    SYSCTL_CHILDREN(rack_counters),
1723 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1724 	    &rack_try_scwnd,
1725 	    "Total number of scwnd attempts");
1726 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1727 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1728 	    OID_AUTO, "outsize", CTLFLAG_RD,
1729 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1730 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1731 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1732 	    OID_AUTO, "opts", CTLFLAG_RD,
1733 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1734 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1735 	    SYSCTL_CHILDREN(rack_sysctl_root),
1736 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1737 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1738 }
1739 
1740 static __inline int
1741 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1742 {
1743 	if (SEQ_GEQ(b->r_start, a->r_start) &&
1744 	    SEQ_LT(b->r_start, a->r_end)) {
1745 		/*
1746 		 * The entry b is within the
1747 		 * block a. i.e.:
1748 		 * a --   |-------------|
1749 		 * b --   |----|
1750 		 * <or>
1751 		 * b --       |------|
1752 		 * <or>
1753 		 * b --       |-----------|
1754 		 */
1755 		return (0);
1756 	} else if (SEQ_GEQ(b->r_start, a->r_end)) {
1757 		/*
1758 		 * b falls as either the next
1759 		 * sequence block after a so a
1760 		 * is said to be smaller than b.
1761 		 * i.e:
1762 		 * a --   |------|
1763 		 * b --          |--------|
1764 		 * or
1765 		 * b --              |-----|
1766 		 */
1767 		return (1);
1768 	}
1769 	/*
1770 	 * Whats left is where a is
1771 	 * larger than b. i.e:
1772 	 * a --         |-------|
1773 	 * b --  |---|
1774 	 * or even possibly
1775 	 * b --   |--------------|
1776 	 */
1777 	return (-1);
1778 }
1779 
1780 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1781 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1782 
1783 static uint32_t
1784 rc_init_window(struct tcp_rack *rack)
1785 {
1786 	uint32_t win;
1787 
1788 	if (rack->rc_init_win == 0) {
1789 		/*
1790 		 * Nothing set by the user, use the system stack
1791 		 * default.
1792 		 */
1793 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1794 	}
1795 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1796 	return (win);
1797 }
1798 
1799 static uint64_t
1800 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1801 {
1802 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1803 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1804 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1805 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1806 	else
1807 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1808 }
1809 
1810 static uint64_t
1811 rack_get_bw(struct tcp_rack *rack)
1812 {
1813 	if (rack->use_fixed_rate) {
1814 		/* Return the fixed pacing rate */
1815 		return (rack_get_fixed_pacing_bw(rack));
1816 	}
1817 	if (rack->r_ctl.gp_bw == 0) {
1818 		/*
1819 		 * We have yet no b/w measurement,
1820 		 * if we have a user set initial bw
1821 		 * return it. If we don't have that and
1822 		 * we have an srtt, use the tcp IW (10) to
1823 		 * calculate a fictional b/w over the SRTT
1824 		 * which is more or less a guess. Note
1825 		 * we don't use our IW from rack on purpose
1826 		 * so if we have like IW=30, we are not
1827 		 * calculating a "huge" b/w.
1828 		 */
1829 		uint64_t bw, srtt;
1830 		if (rack->r_ctl.init_rate)
1831 			return (rack->r_ctl.init_rate);
1832 
1833 		/* Has the user set a max peak rate? */
1834 #ifdef NETFLIX_PEAKRATE
1835 		if (rack->rc_tp->t_maxpeakrate)
1836 			return (rack->rc_tp->t_maxpeakrate);
1837 #endif
1838 		/* Ok lets come up with the IW guess, if we have a srtt */
1839 		if (rack->rc_tp->t_srtt == 0) {
1840 			/*
1841 			 * Go with old pacing method
1842 			 * i.e. burst mitigation only.
1843 			 */
1844 			return (0);
1845 		}
1846 		/* Ok lets get the initial TCP win (not racks) */
1847 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1848 		srtt = (uint64_t)rack->rc_tp->t_srtt;
1849 		bw *= (uint64_t)USECS_IN_SECOND;
1850 		bw /= srtt;
1851 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1852 			bw = rack->r_ctl.bw_rate_cap;
1853 		return (bw);
1854 	} else {
1855 		uint64_t bw;
1856 
1857 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
1858 			/* Averaging is done, we can return the value */
1859 			bw = rack->r_ctl.gp_bw;
1860 		} else {
1861 			/* Still doing initial average must calculate */
1862 			bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
1863 		}
1864 #ifdef NETFLIX_PEAKRATE
1865 		if ((rack->rc_tp->t_maxpeakrate) &&
1866 		    (bw > rack->rc_tp->t_maxpeakrate)) {
1867 			/* The user has set a peak rate to pace at
1868 			 * don't allow us to pace faster than that.
1869 			 */
1870 			return (rack->rc_tp->t_maxpeakrate);
1871 		}
1872 #endif
1873 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1874 			bw = rack->r_ctl.bw_rate_cap;
1875 		return (bw);
1876 	}
1877 }
1878 
1879 static uint16_t
1880 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1881 {
1882 	if (rack->use_fixed_rate) {
1883 		return (100);
1884 	} else if (rack->in_probe_rtt && (rsm == NULL))
1885 		return (rack->r_ctl.rack_per_of_gp_probertt);
1886 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
1887 		  rack->r_ctl.rack_per_of_gp_rec)) {
1888 		if (rsm) {
1889 			/* a retransmission always use the recovery rate */
1890 			return (rack->r_ctl.rack_per_of_gp_rec);
1891 		} else if (rack->rack_rec_nonrxt_use_cr) {
1892 			/* Directed to use the configured rate */
1893 			goto configured_rate;
1894 		} else if (rack->rack_no_prr &&
1895 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1896 			/* No PRR, lets just use the b/w estimate only */
1897 			return (100);
1898 		} else {
1899 			/*
1900 			 * Here we may have a non-retransmit but we
1901 			 * have no overrides, so just use the recovery
1902 			 * rate (prr is in effect).
1903 			 */
1904 			return (rack->r_ctl.rack_per_of_gp_rec);
1905 		}
1906 	}
1907 configured_rate:
1908 	/* For the configured rate we look at our cwnd vs the ssthresh */
1909 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1910 		return (rack->r_ctl.rack_per_of_gp_ss);
1911 	else
1912 		return (rack->r_ctl.rack_per_of_gp_ca);
1913 }
1914 
1915 static void
1916 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
1917 {
1918 	/*
1919 	 * Types of logs (mod value)
1920 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
1921 	 * 2 = a dsack round begins, persist is reset to 16.
1922 	 * 3 = a dsack round ends
1923 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
1924 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
1925 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
1926 	 */
1927 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1928 		union tcp_log_stackspecific log;
1929 		struct timeval tv;
1930 
1931 		memset(&log, 0, sizeof(log));
1932 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
1933 		log.u_bbr.flex1 <<= 1;
1934 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
1935 		log.u_bbr.flex1 <<= 1;
1936 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
1937 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
1938 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
1939 		log.u_bbr.flex4 = flex4;
1940 		log.u_bbr.flex5 = flex5;
1941 		log.u_bbr.flex6 = flex6;
1942 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
1943 		log.u_bbr.flex8 = mod;
1944 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1945 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1946 		    &rack->rc_inp->inp_socket->so_rcv,
1947 		    &rack->rc_inp->inp_socket->so_snd,
1948 		    RACK_DSACK_HANDLING, 0,
1949 		    0, &log, false, &tv);
1950 	}
1951 }
1952 
1953 static void
1954 rack_log_hdwr_pacing(struct tcp_rack *rack,
1955 		     uint64_t rate, uint64_t hw_rate, int line,
1956 		     int error, uint16_t mod)
1957 {
1958 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1959 		union tcp_log_stackspecific log;
1960 		struct timeval tv;
1961 		const struct ifnet *ifp;
1962 
1963 		memset(&log, 0, sizeof(log));
1964 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
1965 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
1966 		if (rack->r_ctl.crte) {
1967 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
1968 		} else if (rack->rc_inp->inp_route.ro_nh &&
1969 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
1970 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
1971 		} else
1972 			ifp = NULL;
1973 		if (ifp) {
1974 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
1975 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
1976 		}
1977 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1978 		log.u_bbr.bw_inuse = rate;
1979 		log.u_bbr.flex5 = line;
1980 		log.u_bbr.flex6 = error;
1981 		log.u_bbr.flex7 = mod;
1982 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
1983 		log.u_bbr.flex8 = rack->use_fixed_rate;
1984 		log.u_bbr.flex8 <<= 1;
1985 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
1986 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
1987 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
1988 		if (rack->r_ctl.crte)
1989 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
1990 		else
1991 			log.u_bbr.cur_del_rate = 0;
1992 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
1993 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1994 		    &rack->rc_inp->inp_socket->so_rcv,
1995 		    &rack->rc_inp->inp_socket->so_snd,
1996 		    BBR_LOG_HDWR_PACE, 0,
1997 		    0, &log, false, &tv);
1998 	}
1999 }
2000 
2001 static uint64_t
2002 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2003 {
2004 	/*
2005 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2006 	 */
2007 	uint64_t bw_est, high_rate;
2008 	uint64_t gain;
2009 
2010 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2011 	bw_est = bw * gain;
2012 	bw_est /= (uint64_t)100;
2013 	/* Never fall below the minimum (def 64kbps) */
2014 	if (bw_est < RACK_MIN_BW)
2015 		bw_est = RACK_MIN_BW;
2016 	if (rack->r_rack_hw_rate_caps) {
2017 		/* Rate caps are in place */
2018 		if (rack->r_ctl.crte != NULL) {
2019 			/* We have a hdwr rate already */
2020 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2021 			if (bw_est >= high_rate) {
2022 				/* We are capping bw at the highest rate table entry */
2023 				rack_log_hdwr_pacing(rack,
2024 						     bw_est, high_rate, __LINE__,
2025 						     0, 3);
2026 				bw_est = high_rate;
2027 				if (capped)
2028 					*capped = 1;
2029 			}
2030 		} else if ((rack->rack_hdrw_pacing == 0) &&
2031 			   (rack->rack_hdw_pace_ena) &&
2032 			   (rack->rack_attempt_hdwr_pace == 0) &&
2033 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2034 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2035 			/*
2036 			 * Special case, we have not yet attempted hardware
2037 			 * pacing, and yet we may, when we do, find out if we are
2038 			 * above the highest rate. We need to know the maxbw for the interface
2039 			 * in question (if it supports ratelimiting). We get back
2040 			 * a 0, if the interface is not found in the RL lists.
2041 			 */
2042 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2043 			if (high_rate) {
2044 				/* Yep, we have a rate is it above this rate? */
2045 				if (bw_est > high_rate) {
2046 					bw_est = high_rate;
2047 					if (capped)
2048 						*capped = 1;
2049 				}
2050 			}
2051 		}
2052 	}
2053 	return (bw_est);
2054 }
2055 
2056 static void
2057 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2058 {
2059 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2060 		union tcp_log_stackspecific log;
2061 		struct timeval tv;
2062 
2063 		if ((mod != 1) && (rack_verbose_logging == 0)) {
2064 			/*
2065 			 * We get 3 values currently for mod
2066 			 * 1 - We are retransmitting and this tells the reason.
2067 			 * 2 - We are clearing a dup-ack count.
2068 			 * 3 - We are incrementing a dup-ack count.
2069 			 *
2070 			 * The clear/increment are only logged
2071 			 * if you have BBverbose on.
2072 			 */
2073 			return;
2074 		}
2075 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2076 		log.u_bbr.flex1 = tsused;
2077 		log.u_bbr.flex2 = thresh;
2078 		log.u_bbr.flex3 = rsm->r_flags;
2079 		log.u_bbr.flex4 = rsm->r_dupack;
2080 		log.u_bbr.flex5 = rsm->r_start;
2081 		log.u_bbr.flex6 = rsm->r_end;
2082 		log.u_bbr.flex8 = mod;
2083 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2084 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2085 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2086 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2087 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2088 		log.u_bbr.pacing_gain = rack->r_must_retran;
2089 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2090 		    &rack->rc_inp->inp_socket->so_rcv,
2091 		    &rack->rc_inp->inp_socket->so_snd,
2092 		    BBR_LOG_SETTINGS_CHG, 0,
2093 		    0, &log, false, &tv);
2094 	}
2095 }
2096 
2097 static void
2098 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2099 {
2100 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2101 		union tcp_log_stackspecific log;
2102 		struct timeval tv;
2103 
2104 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2105 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2106 		log.u_bbr.flex2 = to;
2107 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2108 		log.u_bbr.flex4 = slot;
2109 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2110 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2111 		log.u_bbr.flex7 = rack->rc_in_persist;
2112 		log.u_bbr.flex8 = which;
2113 		if (rack->rack_no_prr)
2114 			log.u_bbr.pkts_out = 0;
2115 		else
2116 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2117 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2118 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2119 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2120 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2121 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2122 		log.u_bbr.pacing_gain = rack->r_must_retran;
2123 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2124 		log.u_bbr.lost = rack_rto_min;
2125 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2126 		    &rack->rc_inp->inp_socket->so_rcv,
2127 		    &rack->rc_inp->inp_socket->so_snd,
2128 		    BBR_LOG_TIMERSTAR, 0,
2129 		    0, &log, false, &tv);
2130 	}
2131 }
2132 
2133 static void
2134 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2135 {
2136 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2137 		union tcp_log_stackspecific log;
2138 		struct timeval tv;
2139 
2140 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2141 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2142 		log.u_bbr.flex8 = to_num;
2143 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2144 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2145 		if (rsm == NULL)
2146 			log.u_bbr.flex3 = 0;
2147 		else
2148 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2149 		if (rack->rack_no_prr)
2150 			log.u_bbr.flex5 = 0;
2151 		else
2152 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2153 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2154 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2155 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2156 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2157 		log.u_bbr.pacing_gain = rack->r_must_retran;
2158 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2159 		    &rack->rc_inp->inp_socket->so_rcv,
2160 		    &rack->rc_inp->inp_socket->so_snd,
2161 		    BBR_LOG_RTO, 0,
2162 		    0, &log, false, &tv);
2163 	}
2164 }
2165 
2166 static void
2167 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2168 		 struct rack_sendmap *prev,
2169 		 struct rack_sendmap *rsm,
2170 		 struct rack_sendmap *next,
2171 		 int flag, uint32_t th_ack, int line)
2172 {
2173 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2174 		union tcp_log_stackspecific log;
2175 		struct timeval tv;
2176 
2177 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2178 		log.u_bbr.flex8 = flag;
2179 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2180 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2181 		log.u_bbr.delRate = (uint64_t)rsm;
2182 		log.u_bbr.rttProp = (uint64_t)next;
2183 		log.u_bbr.flex7 = 0;
2184 		if (prev) {
2185 			log.u_bbr.flex1 = prev->r_start;
2186 			log.u_bbr.flex2 = prev->r_end;
2187 			log.u_bbr.flex7 |= 0x4;
2188 		}
2189 		if (rsm) {
2190 			log.u_bbr.flex3 = rsm->r_start;
2191 			log.u_bbr.flex4 = rsm->r_end;
2192 			log.u_bbr.flex7 |= 0x2;
2193 		}
2194 		if (next) {
2195 			log.u_bbr.flex5 = next->r_start;
2196 			log.u_bbr.flex6 = next->r_end;
2197 			log.u_bbr.flex7 |= 0x1;
2198 		}
2199 		log.u_bbr.applimited = line;
2200 		log.u_bbr.pkts_out = th_ack;
2201 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2202 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2203 		if (rack->rack_no_prr)
2204 			log.u_bbr.lost = 0;
2205 		else
2206 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2207 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2208 		    &rack->rc_inp->inp_socket->so_rcv,
2209 		    &rack->rc_inp->inp_socket->so_snd,
2210 		    TCP_LOG_MAPCHG, 0,
2211 		    0, &log, false, &tv);
2212 	}
2213 }
2214 
2215 static void
2216 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2217 		 struct rack_sendmap *rsm, int conf)
2218 {
2219 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2220 		union tcp_log_stackspecific log;
2221 		struct timeval tv;
2222 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2223 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2224 		log.u_bbr.flex1 = t;
2225 		log.u_bbr.flex2 = len;
2226 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2227 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2228 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2229 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2230 		log.u_bbr.flex7 = conf;
2231 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2232 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2233 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2234 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2235 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2236 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2237 		if (rsm) {
2238 			log.u_bbr.pkt_epoch = rsm->r_start;
2239 			log.u_bbr.lost = rsm->r_end;
2240 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2241 			/* We loose any upper of the 24 bits */
2242 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2243 		} else {
2244 			/* Its a SYN */
2245 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2246 			log.u_bbr.lost = 0;
2247 			log.u_bbr.cwnd_gain = 0;
2248 			log.u_bbr.pacing_gain = 0;
2249 		}
2250 		/* Write out general bits of interest rrs here */
2251 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2252 		log.u_bbr.use_lt_bw <<= 1;
2253 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2254 		log.u_bbr.use_lt_bw <<= 1;
2255 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2256 		log.u_bbr.use_lt_bw <<= 1;
2257 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2258 		log.u_bbr.use_lt_bw <<= 1;
2259 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2260 		log.u_bbr.use_lt_bw <<= 1;
2261 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2262 		log.u_bbr.use_lt_bw <<= 1;
2263 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2264 		log.u_bbr.use_lt_bw <<= 1;
2265 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2266 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2267 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2268 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2269 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2270 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2271 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2272 		log.u_bbr.bw_inuse <<= 32;
2273 		if (rsm)
2274 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2275 		TCP_LOG_EVENTP(tp, NULL,
2276 		    &rack->rc_inp->inp_socket->so_rcv,
2277 		    &rack->rc_inp->inp_socket->so_snd,
2278 		    BBR_LOG_BBRRTT, 0,
2279 		    0, &log, false, &tv);
2280 
2281 
2282 	}
2283 }
2284 
2285 static void
2286 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2287 {
2288 	/*
2289 	 * Log the rtt sample we are
2290 	 * applying to the srtt algorithm in
2291 	 * useconds.
2292 	 */
2293 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2294 		union tcp_log_stackspecific log;
2295 		struct timeval tv;
2296 
2297 		/* Convert our ms to a microsecond */
2298 		memset(&log, 0, sizeof(log));
2299 		log.u_bbr.flex1 = rtt;
2300 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2301 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
2302 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2303 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2304 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2305 		log.u_bbr.flex7 = 1;
2306 		log.u_bbr.flex8 = rack->sack_attack_disable;
2307 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2308 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2309 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2310 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2311 		log.u_bbr.pacing_gain = rack->r_must_retran;
2312 		/*
2313 		 * We capture in delRate the upper 32 bits as
2314 		 * the confidence level we had declared, and the
2315 		 * lower 32 bits as the actual RTT using the arrival
2316 		 * timestamp.
2317 		 */
2318 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2319 		log.u_bbr.delRate <<= 32;
2320 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2321 		/* Lets capture all the things that make up t_rtxcur */
2322 		log.u_bbr.applimited = rack_rto_min;
2323 		log.u_bbr.epoch = rack_rto_max;
2324 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2325 		log.u_bbr.lost = rack_rto_min;
2326 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2327 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2328 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2329 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2330 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2331 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2332 		    &rack->rc_inp->inp_socket->so_rcv,
2333 		    &rack->rc_inp->inp_socket->so_snd,
2334 		    TCP_LOG_RTT, 0,
2335 		    0, &log, false, &tv);
2336 	}
2337 }
2338 
2339 static void
2340 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2341 {
2342 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2343 		union tcp_log_stackspecific log;
2344 		struct timeval tv;
2345 
2346 		/* Convert our ms to a microsecond */
2347 		memset(&log, 0, sizeof(log));
2348 		log.u_bbr.flex1 = rtt;
2349 		log.u_bbr.flex2 = send_time;
2350 		log.u_bbr.flex3 = ack_time;
2351 		log.u_bbr.flex4 = where;
2352 		log.u_bbr.flex7 = 2;
2353 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2354 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2355 		    &rack->rc_inp->inp_socket->so_rcv,
2356 		    &rack->rc_inp->inp_socket->so_snd,
2357 		    TCP_LOG_RTT, 0,
2358 		    0, &log, false, &tv);
2359 	}
2360 }
2361 
2362 
2363 
2364 static inline void
2365 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2366 {
2367 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2368 		union tcp_log_stackspecific log;
2369 		struct timeval tv;
2370 
2371 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2372 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2373 		log.u_bbr.flex1 = line;
2374 		log.u_bbr.flex2 = tick;
2375 		log.u_bbr.flex3 = tp->t_maxunacktime;
2376 		log.u_bbr.flex4 = tp->t_acktime;
2377 		log.u_bbr.flex8 = event;
2378 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2379 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2380 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2381 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2382 		log.u_bbr.pacing_gain = rack->r_must_retran;
2383 		TCP_LOG_EVENTP(tp, NULL,
2384 		    &rack->rc_inp->inp_socket->so_rcv,
2385 		    &rack->rc_inp->inp_socket->so_snd,
2386 		    BBR_LOG_PROGRESS, 0,
2387 		    0, &log, false, &tv);
2388 	}
2389 }
2390 
2391 static void
2392 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2393 {
2394 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2395 		union tcp_log_stackspecific log;
2396 
2397 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2398 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2399 		log.u_bbr.flex1 = slot;
2400 		if (rack->rack_no_prr)
2401 			log.u_bbr.flex2 = 0;
2402 		else
2403 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2404 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2405 		log.u_bbr.flex8 = rack->rc_in_persist;
2406 		log.u_bbr.timeStamp = cts;
2407 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2408 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2409 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2410 		log.u_bbr.pacing_gain = rack->r_must_retran;
2411 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2412 		    &rack->rc_inp->inp_socket->so_rcv,
2413 		    &rack->rc_inp->inp_socket->so_snd,
2414 		    BBR_LOG_BBRSND, 0,
2415 		    0, &log, false, tv);
2416 	}
2417 }
2418 
2419 static void
2420 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2421 {
2422 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2423 		union tcp_log_stackspecific log;
2424 		struct timeval tv;
2425 
2426 		memset(&log, 0, sizeof(log));
2427 		log.u_bbr.flex1 = did_out;
2428 		log.u_bbr.flex2 = nxt_pkt;
2429 		log.u_bbr.flex3 = way_out;
2430 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2431 		if (rack->rack_no_prr)
2432 			log.u_bbr.flex5 = 0;
2433 		else
2434 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2435 		log.u_bbr.flex6 = nsegs;
2436 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2437 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
2438 		log.u_bbr.flex7 <<= 1;
2439 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
2440 		log.u_bbr.flex7 <<= 1;
2441 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
2442 		log.u_bbr.flex8 = rack->rc_in_persist;
2443 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2444 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2445 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2446 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2447 		log.u_bbr.use_lt_bw <<= 1;
2448 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2449 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2450 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2451 		log.u_bbr.pacing_gain = rack->r_must_retran;
2452 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2453 		    &rack->rc_inp->inp_socket->so_rcv,
2454 		    &rack->rc_inp->inp_socket->so_snd,
2455 		    BBR_LOG_DOSEG_DONE, 0,
2456 		    0, &log, false, &tv);
2457 	}
2458 }
2459 
2460 static void
2461 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2462 {
2463 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2464 		union tcp_log_stackspecific log;
2465 		struct timeval tv;
2466 
2467 		memset(&log, 0, sizeof(log));
2468 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2469 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2470 		log.u_bbr.flex4 = arg1;
2471 		log.u_bbr.flex5 = arg2;
2472 		log.u_bbr.flex6 = arg3;
2473 		log.u_bbr.flex8 = frm;
2474 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2475 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2476 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2477 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2478 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2479 		log.u_bbr.pacing_gain = rack->r_must_retran;
2480 		TCP_LOG_EVENTP(tp, NULL,
2481 		    &tp->t_inpcb->inp_socket->so_rcv,
2482 		    &tp->t_inpcb->inp_socket->so_snd,
2483 		    TCP_HDWR_PACE_SIZE, 0,
2484 		    0, &log, false, &tv);
2485 	}
2486 }
2487 
2488 static void
2489 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2490 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2491 {
2492 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2493 		union tcp_log_stackspecific log;
2494 		struct timeval tv;
2495 
2496 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2497 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2498 		log.u_bbr.flex1 = slot;
2499 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2500 		log.u_bbr.flex4 = reason;
2501 		if (rack->rack_no_prr)
2502 			log.u_bbr.flex5 = 0;
2503 		else
2504 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2505 		log.u_bbr.flex7 = hpts_calling;
2506 		log.u_bbr.flex8 = rack->rc_in_persist;
2507 		log.u_bbr.lt_epoch = cwnd_to_use;
2508 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2509 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2510 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2511 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2512 		log.u_bbr.pacing_gain = rack->r_must_retran;
2513 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2514 		    &rack->rc_inp->inp_socket->so_rcv,
2515 		    &rack->rc_inp->inp_socket->so_snd,
2516 		    BBR_LOG_JUSTRET, 0,
2517 		    tlen, &log, false, &tv);
2518 	}
2519 }
2520 
2521 static void
2522 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2523 		   struct timeval *tv, uint32_t flags_on_entry)
2524 {
2525 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2526 		union tcp_log_stackspecific log;
2527 
2528 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2529 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2530 		log.u_bbr.flex1 = line;
2531 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2532 		log.u_bbr.flex3 = flags_on_entry;
2533 		log.u_bbr.flex4 = us_cts;
2534 		if (rack->rack_no_prr)
2535 			log.u_bbr.flex5 = 0;
2536 		else
2537 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2538 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2539 		log.u_bbr.flex7 = hpts_removed;
2540 		log.u_bbr.flex8 = 1;
2541 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2542 		log.u_bbr.timeStamp = us_cts;
2543 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2544 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2545 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2546 		log.u_bbr.pacing_gain = rack->r_must_retran;
2547 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2548 		    &rack->rc_inp->inp_socket->so_rcv,
2549 		    &rack->rc_inp->inp_socket->so_snd,
2550 		    BBR_LOG_TIMERCANC, 0,
2551 		    0, &log, false, tv);
2552 	}
2553 }
2554 
2555 static void
2556 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2557 			  uint32_t flex1, uint32_t flex2,
2558 			  uint32_t flex3, uint32_t flex4,
2559 			  uint32_t flex5, uint32_t flex6,
2560 			  uint16_t flex7, uint8_t mod)
2561 {
2562 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2563 		union tcp_log_stackspecific log;
2564 		struct timeval tv;
2565 
2566 		if (mod == 1) {
2567 			/* No you can't use 1, its for the real to cancel */
2568 			return;
2569 		}
2570 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2571 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2572 		log.u_bbr.flex1 = flex1;
2573 		log.u_bbr.flex2 = flex2;
2574 		log.u_bbr.flex3 = flex3;
2575 		log.u_bbr.flex4 = flex4;
2576 		log.u_bbr.flex5 = flex5;
2577 		log.u_bbr.flex6 = flex6;
2578 		log.u_bbr.flex7 = flex7;
2579 		log.u_bbr.flex8 = mod;
2580 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2581 		    &rack->rc_inp->inp_socket->so_rcv,
2582 		    &rack->rc_inp->inp_socket->so_snd,
2583 		    BBR_LOG_TIMERCANC, 0,
2584 		    0, &log, false, &tv);
2585 	}
2586 }
2587 
2588 static void
2589 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2590 {
2591 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2592 		union tcp_log_stackspecific log;
2593 		struct timeval tv;
2594 
2595 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2596 		log.u_bbr.flex1 = timers;
2597 		log.u_bbr.flex2 = ret;
2598 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2599 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2600 		log.u_bbr.flex5 = cts;
2601 		if (rack->rack_no_prr)
2602 			log.u_bbr.flex6 = 0;
2603 		else
2604 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2605 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2606 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2607 		log.u_bbr.pacing_gain = rack->r_must_retran;
2608 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2609 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2610 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2611 		    &rack->rc_inp->inp_socket->so_rcv,
2612 		    &rack->rc_inp->inp_socket->so_snd,
2613 		    BBR_LOG_TO_PROCESS, 0,
2614 		    0, &log, false, &tv);
2615 	}
2616 }
2617 
2618 static void
2619 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2620 {
2621 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2622 		union tcp_log_stackspecific log;
2623 		struct timeval tv;
2624 
2625 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2626 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2627 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2628 		if (rack->rack_no_prr)
2629 			log.u_bbr.flex3 = 0;
2630 		else
2631 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2632 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2633 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2634 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2635 		log.u_bbr.flex8 = frm;
2636 		log.u_bbr.pkts_out = orig_cwnd;
2637 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2638 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2639 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2640 		log.u_bbr.use_lt_bw <<= 1;
2641 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2642 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2643 		    &rack->rc_inp->inp_socket->so_rcv,
2644 		    &rack->rc_inp->inp_socket->so_snd,
2645 		    BBR_LOG_BBRUPD, 0,
2646 		    0, &log, false, &tv);
2647 	}
2648 }
2649 
2650 #ifdef NETFLIX_EXP_DETECTION
2651 static void
2652 rack_log_sad(struct tcp_rack *rack, int event)
2653 {
2654 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2655 		union tcp_log_stackspecific log;
2656 		struct timeval tv;
2657 
2658 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2659 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
2660 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2661 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2662 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2663 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2664 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2665 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2666 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2667 		log.u_bbr.lt_epoch |= rack->do_detection;
2668 		log.u_bbr.applimited = tcp_map_minimum;
2669 		log.u_bbr.flex7 = rack->sack_attack_disable;
2670 		log.u_bbr.flex8 = event;
2671 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2672 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2673 		log.u_bbr.delivered = tcp_sad_decay_val;
2674 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2675 		    &rack->rc_inp->inp_socket->so_rcv,
2676 		    &rack->rc_inp->inp_socket->so_snd,
2677 		    TCP_SAD_DETECTION, 0,
2678 		    0, &log, false, &tv);
2679 	}
2680 }
2681 #endif
2682 
2683 static void
2684 rack_counter_destroy(void)
2685 {
2686 	counter_u64_free(rack_fto_send);
2687 	counter_u64_free(rack_fto_rsm_send);
2688 	counter_u64_free(rack_nfto_resend);
2689 	counter_u64_free(rack_hw_pace_init_fail);
2690 	counter_u64_free(rack_hw_pace_lost);
2691 	counter_u64_free(rack_non_fto_send);
2692 	counter_u64_free(rack_extended_rfo);
2693 	counter_u64_free(rack_ack_total);
2694 	counter_u64_free(rack_express_sack);
2695 	counter_u64_free(rack_sack_total);
2696 	counter_u64_free(rack_move_none);
2697 	counter_u64_free(rack_move_some);
2698 	counter_u64_free(rack_sack_attacks_detected);
2699 	counter_u64_free(rack_sack_attacks_reversed);
2700 	counter_u64_free(rack_sack_used_next_merge);
2701 	counter_u64_free(rack_sack_used_prev_merge);
2702 	counter_u64_free(rack_tlp_tot);
2703 	counter_u64_free(rack_tlp_newdata);
2704 	counter_u64_free(rack_tlp_retran);
2705 	counter_u64_free(rack_tlp_retran_bytes);
2706 	counter_u64_free(rack_to_tot);
2707 	counter_u64_free(rack_saw_enobuf);
2708 	counter_u64_free(rack_saw_enobuf_hw);
2709 	counter_u64_free(rack_saw_enetunreach);
2710 	counter_u64_free(rack_hot_alloc);
2711 	counter_u64_free(rack_to_alloc);
2712 	counter_u64_free(rack_to_alloc_hard);
2713 	counter_u64_free(rack_to_alloc_emerg);
2714 	counter_u64_free(rack_to_alloc_limited);
2715 	counter_u64_free(rack_alloc_limited_conns);
2716 	counter_u64_free(rack_split_limited);
2717 	counter_u64_free(rack_multi_single_eq);
2718 	counter_u64_free(rack_proc_non_comp_ack);
2719 	counter_u64_free(rack_sack_proc_all);
2720 	counter_u64_free(rack_sack_proc_restart);
2721 	counter_u64_free(rack_sack_proc_short);
2722 	counter_u64_free(rack_sack_skipped_acked);
2723 	counter_u64_free(rack_sack_splits);
2724 	counter_u64_free(rack_input_idle_reduces);
2725 	counter_u64_free(rack_collapsed_win);
2726 	counter_u64_free(rack_try_scwnd);
2727 	counter_u64_free(rack_persists_sends);
2728 	counter_u64_free(rack_persists_acks);
2729 	counter_u64_free(rack_persists_loss);
2730 	counter_u64_free(rack_persists_lost_ends);
2731 #ifdef INVARIANTS
2732 	counter_u64_free(rack_adjust_map_bw);
2733 #endif
2734 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2735 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2736 }
2737 
2738 static struct rack_sendmap *
2739 rack_alloc(struct tcp_rack *rack)
2740 {
2741 	struct rack_sendmap *rsm;
2742 
2743 	/*
2744 	 * First get the top of the list it in
2745 	 * theory is the "hottest" rsm we have,
2746 	 * possibly just freed by ack processing.
2747 	 */
2748 	if (rack->rc_free_cnt > rack_free_cache) {
2749 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2750 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2751 		counter_u64_add(rack_hot_alloc, 1);
2752 		rack->rc_free_cnt--;
2753 		return (rsm);
2754 	}
2755 	/*
2756 	 * Once we get under our free cache we probably
2757 	 * no longer have a "hot" one available. Lets
2758 	 * get one from UMA.
2759 	 */
2760 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
2761 	if (rsm) {
2762 		rack->r_ctl.rc_num_maps_alloced++;
2763 		counter_u64_add(rack_to_alloc, 1);
2764 		return (rsm);
2765 	}
2766 	/*
2767 	 * Dig in to our aux rsm's (the last two) since
2768 	 * UMA failed to get us one.
2769 	 */
2770 	if (rack->rc_free_cnt) {
2771 		counter_u64_add(rack_to_alloc_emerg, 1);
2772 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2773 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2774 		rack->rc_free_cnt--;
2775 		return (rsm);
2776 	}
2777 	return (NULL);
2778 }
2779 
2780 static struct rack_sendmap *
2781 rack_alloc_full_limit(struct tcp_rack *rack)
2782 {
2783 	if ((V_tcp_map_entries_limit > 0) &&
2784 	    (rack->do_detection == 0) &&
2785 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2786 		counter_u64_add(rack_to_alloc_limited, 1);
2787 		if (!rack->alloc_limit_reported) {
2788 			rack->alloc_limit_reported = 1;
2789 			counter_u64_add(rack_alloc_limited_conns, 1);
2790 		}
2791 		return (NULL);
2792 	}
2793 	return (rack_alloc(rack));
2794 }
2795 
2796 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2797 static struct rack_sendmap *
2798 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2799 {
2800 	struct rack_sendmap *rsm;
2801 
2802 	if (limit_type) {
2803 		/* currently there is only one limit type */
2804 		if (V_tcp_map_split_limit > 0 &&
2805 		    (rack->do_detection == 0) &&
2806 		    rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2807 			counter_u64_add(rack_split_limited, 1);
2808 			if (!rack->alloc_limit_reported) {
2809 				rack->alloc_limit_reported = 1;
2810 				counter_u64_add(rack_alloc_limited_conns, 1);
2811 			}
2812 			return (NULL);
2813 		}
2814 	}
2815 
2816 	/* allocate and mark in the limit type, if set */
2817 	rsm = rack_alloc(rack);
2818 	if (rsm != NULL && limit_type) {
2819 		rsm->r_limit_type = limit_type;
2820 		rack->r_ctl.rc_num_split_allocs++;
2821 	}
2822 	return (rsm);
2823 }
2824 
2825 static void
2826 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2827 {
2828 	if (rsm->r_flags & RACK_APP_LIMITED) {
2829 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
2830 			rack->r_ctl.rc_app_limited_cnt--;
2831 		}
2832 	}
2833 	if (rsm->r_limit_type) {
2834 		/* currently there is only one limit type */
2835 		rack->r_ctl.rc_num_split_allocs--;
2836 	}
2837 	if (rsm == rack->r_ctl.rc_first_appl) {
2838 		if (rack->r_ctl.rc_app_limited_cnt == 0)
2839 			rack->r_ctl.rc_first_appl = NULL;
2840 		else {
2841 			/* Follow the next one out */
2842 			struct rack_sendmap fe;
2843 
2844 			fe.r_start = rsm->r_nseq_appl;
2845 			rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2846 		}
2847 	}
2848 	if (rsm == rack->r_ctl.rc_resend)
2849 		rack->r_ctl.rc_resend = NULL;
2850 	if (rsm == rack->r_ctl.rc_end_appl)
2851 		rack->r_ctl.rc_end_appl = NULL;
2852 	if (rack->r_ctl.rc_tlpsend == rsm)
2853 		rack->r_ctl.rc_tlpsend = NULL;
2854 	if (rack->r_ctl.rc_sacklast == rsm)
2855 		rack->r_ctl.rc_sacklast = NULL;
2856 	memset(rsm, 0, sizeof(struct rack_sendmap));
2857 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
2858 	rack->rc_free_cnt++;
2859 }
2860 
2861 static void
2862 rack_free_trim(struct tcp_rack *rack)
2863 {
2864 	struct rack_sendmap *rsm;
2865 
2866 	/*
2867 	 * Free up all the tail entries until
2868 	 * we get our list down to the limit.
2869 	 */
2870 	while (rack->rc_free_cnt > rack_free_cache) {
2871 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
2872 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2873 		rack->rc_free_cnt--;
2874 		uma_zfree(rack_zone, rsm);
2875 	}
2876 }
2877 
2878 
2879 static uint32_t
2880 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2881 {
2882 	uint64_t srtt, bw, len, tim;
2883 	uint32_t segsiz, def_len, minl;
2884 
2885 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2886 	def_len = rack_def_data_window * segsiz;
2887 	if (rack->rc_gp_filled == 0) {
2888 		/*
2889 		 * We have no measurement (IW is in flight?) so
2890 		 * we can only guess using our data_window sysctl
2891 		 * value (usually 20MSS).
2892 		 */
2893 		return (def_len);
2894 	}
2895 	/*
2896 	 * Now we have a number of factors to consider.
2897 	 *
2898 	 * 1) We have a desired BDP which is usually
2899 	 *    at least 2.
2900 	 * 2) We have a minimum number of rtt's usually 1 SRTT
2901 	 *    but we allow it too to be more.
2902 	 * 3) We want to make sure a measurement last N useconds (if
2903 	 *    we have set rack_min_measure_usec.
2904 	 *
2905 	 * We handle the first concern here by trying to create a data
2906 	 * window of max(rack_def_data_window, DesiredBDP). The
2907 	 * second concern we handle in not letting the measurement
2908 	 * window end normally until at least the required SRTT's
2909 	 * have gone by which is done further below in
2910 	 * rack_enough_for_measurement(). Finally the third concern
2911 	 * we also handle here by calculating how long that time
2912 	 * would take at the current BW and then return the
2913 	 * max of our first calculation and that length. Note
2914 	 * that if rack_min_measure_usec is 0, we don't deal
2915 	 * with concern 3. Also for both Concern 1 and 3 an
2916 	 * application limited period could end the measurement
2917 	 * earlier.
2918 	 *
2919 	 * So lets calculate the BDP with the "known" b/w using
2920 	 * the SRTT has our rtt and then multiply it by the
2921 	 * goal.
2922 	 */
2923 	bw = rack_get_bw(rack);
2924 	srtt = (uint64_t)tp->t_srtt;
2925 	len = bw * srtt;
2926 	len /= (uint64_t)HPTS_USEC_IN_SEC;
2927 	len *= max(1, rack_goal_bdp);
2928 	/* Now we need to round up to the nearest MSS */
2929 	len = roundup(len, segsiz);
2930 	if (rack_min_measure_usec) {
2931 		/* Now calculate our min length for this b/w */
2932 		tim = rack_min_measure_usec;
2933 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
2934 		if (minl == 0)
2935 			minl = 1;
2936 		minl = roundup(minl, segsiz);
2937 		if (len < minl)
2938 			len = minl;
2939 	}
2940 	/*
2941 	 * Now if we have a very small window we want
2942 	 * to attempt to get the window that is
2943 	 * as small as possible. This happens on
2944 	 * low b/w connections and we don't want to
2945 	 * span huge numbers of rtt's between measurements.
2946 	 *
2947 	 * We basically include 2 over our "MIN window" so
2948 	 * that the measurement can be shortened (possibly) by
2949 	 * an ack'ed packet.
2950 	 */
2951 	if (len < def_len)
2952 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
2953 	else
2954 		return (max((uint32_t)len, def_len));
2955 
2956 }
2957 
2958 static int
2959 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
2960 {
2961 	uint32_t tim, srtts, segsiz;
2962 
2963 	/*
2964 	 * Has enough time passed for the GP measurement to be valid?
2965 	 */
2966 	if ((tp->snd_max == tp->snd_una) ||
2967 	    (th_ack == tp->snd_max)){
2968 		/* All is acked */
2969 		*quality = RACK_QUALITY_ALLACKED;
2970 		return (1);
2971 	}
2972 	if (SEQ_LT(th_ack, tp->gput_seq)) {
2973 		/* Not enough bytes yet */
2974 		return (0);
2975 	}
2976 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2977 	if (SEQ_LT(th_ack, tp->gput_ack) &&
2978 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
2979 		/* Not enough bytes yet */
2980 		return (0);
2981 	}
2982 	if (rack->r_ctl.rc_first_appl &&
2983 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
2984 		/*
2985 		 * We are up to the app limited send point
2986 		 * we have to measure irrespective of the time..
2987 		 */
2988 		*quality = RACK_QUALITY_APPLIMITED;
2989 		return (1);
2990 	}
2991 	/* Now what about time? */
2992 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
2993 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
2994 	if (tim >= srtts) {
2995 		*quality = RACK_QUALITY_HIGH;
2996 		return (1);
2997 	}
2998 	/* Nope not even a full SRTT has passed */
2999 	return (0);
3000 }
3001 
3002 static void
3003 rack_log_timely(struct tcp_rack *rack,
3004 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3005 		uint64_t up_bnd, int line, uint8_t method)
3006 {
3007 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3008 		union tcp_log_stackspecific log;
3009 		struct timeval tv;
3010 
3011 		memset(&log, 0, sizeof(log));
3012 		log.u_bbr.flex1 = logged;
3013 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3014 		log.u_bbr.flex2 <<= 4;
3015 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3016 		log.u_bbr.flex2 <<= 4;
3017 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3018 		log.u_bbr.flex2 <<= 4;
3019 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3020 		log.u_bbr.flex3 = rack->rc_gp_incr;
3021 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3022 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3023 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3024 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3025 		log.u_bbr.flex8 = method;
3026 		log.u_bbr.cur_del_rate = cur_bw;
3027 		log.u_bbr.delRate = low_bnd;
3028 		log.u_bbr.bw_inuse = up_bnd;
3029 		log.u_bbr.rttProp = rack_get_bw(rack);
3030 		log.u_bbr.pkt_epoch = line;
3031 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3032 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3033 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3034 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3035 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3036 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3037 		log.u_bbr.cwnd_gain <<= 1;
3038 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3039 		log.u_bbr.cwnd_gain <<= 1;
3040 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3041 		log.u_bbr.cwnd_gain <<= 1;
3042 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3043 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3044 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3045 		    &rack->rc_inp->inp_socket->so_rcv,
3046 		    &rack->rc_inp->inp_socket->so_snd,
3047 		    TCP_TIMELY_WORK, 0,
3048 		    0, &log, false, &tv);
3049 	}
3050 }
3051 
3052 static int
3053 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3054 {
3055 	/*
3056 	 * Before we increase we need to know if
3057 	 * the estimate just made was less than
3058 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3059 	 *
3060 	 * If we already are pacing at a fast enough
3061 	 * rate to push us faster there is no sense of
3062 	 * increasing.
3063 	 *
3064 	 * We first caculate our actual pacing rate (ss or ca multipler
3065 	 * times our cur_bw).
3066 	 *
3067 	 * Then we take the last measured rate and multipy by our
3068 	 * maximum pacing overage to give us a max allowable rate.
3069 	 *
3070 	 * If our act_rate is smaller than our max_allowable rate
3071 	 * then we should increase. Else we should hold steady.
3072 	 *
3073 	 */
3074 	uint64_t act_rate, max_allow_rate;
3075 
3076 	if (rack_timely_no_stopping)
3077 		return (1);
3078 
3079 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3080 		/*
3081 		 * Initial startup case or
3082 		 * everything is acked case.
3083 		 */
3084 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3085 				__LINE__, 9);
3086 		return (1);
3087 	}
3088 	if (mult <= 100) {
3089 		/*
3090 		 * We can always pace at or slightly above our rate.
3091 		 */
3092 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3093 				__LINE__, 9);
3094 		return (1);
3095 	}
3096 	act_rate = cur_bw * (uint64_t)mult;
3097 	act_rate /= 100;
3098 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3099 	max_allow_rate /= 100;
3100 	if (act_rate < max_allow_rate) {
3101 		/*
3102 		 * Here the rate we are actually pacing at
3103 		 * is smaller than 10% above our last measurement.
3104 		 * This means we are pacing below what we would
3105 		 * like to try to achieve (plus some wiggle room).
3106 		 */
3107 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3108 				__LINE__, 9);
3109 		return (1);
3110 	} else {
3111 		/*
3112 		 * Here we are already pacing at least rack_max_per_above(10%)
3113 		 * what we are getting back. This indicates most likely
3114 		 * that we are being limited (cwnd/rwnd/app) and can't
3115 		 * get any more b/w. There is no sense of trying to
3116 		 * raise up the pacing rate its not speeding us up
3117 		 * and we already are pacing faster than we are getting.
3118 		 */
3119 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3120 				__LINE__, 8);
3121 		return (0);
3122 	}
3123 }
3124 
3125 static void
3126 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3127 {
3128 	/*
3129 	 * When we drag bottom, we want to assure
3130 	 * that no multiplier is below 1.0, if so
3131 	 * we want to restore it to at least that.
3132 	 */
3133 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3134 		/* This is unlikely we usually do not touch recovery */
3135 		rack->r_ctl.rack_per_of_gp_rec = 100;
3136 	}
3137 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3138 		rack->r_ctl.rack_per_of_gp_ca = 100;
3139 	}
3140 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3141 		rack->r_ctl.rack_per_of_gp_ss = 100;
3142 	}
3143 }
3144 
3145 static void
3146 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3147 {
3148 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3149 		rack->r_ctl.rack_per_of_gp_ca = 100;
3150 	}
3151 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3152 		rack->r_ctl.rack_per_of_gp_ss = 100;
3153 	}
3154 }
3155 
3156 static void
3157 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3158 {
3159 	int32_t  calc, logged, plus;
3160 
3161 	logged = 0;
3162 
3163 	if (override) {
3164 		/*
3165 		 * override is passed when we are
3166 		 * loosing b/w and making one last
3167 		 * gasp at trying to not loose out
3168 		 * to a new-reno flow.
3169 		 */
3170 		goto extra_boost;
3171 	}
3172 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3173 	if (rack->rc_gp_incr &&
3174 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3175 		/*
3176 		 * Reset and get 5 strokes more before the boost. Note
3177 		 * that the count is 0 based so we have to add one.
3178 		 */
3179 extra_boost:
3180 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3181 		rack->rc_gp_timely_inc_cnt = 0;
3182 	} else
3183 		plus = (uint32_t)rack_gp_increase_per;
3184 	/* Must be at least 1% increase for true timely increases */
3185 	if ((plus < 1) &&
3186 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3187 		plus = 1;
3188 	if (rack->rc_gp_saw_rec &&
3189 	    (rack->rc_gp_no_rec_chg == 0) &&
3190 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3191 				  rack->r_ctl.rack_per_of_gp_rec)) {
3192 		/* We have been in recovery ding it too */
3193 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3194 		if (calc > 0xffff)
3195 			calc = 0xffff;
3196 		logged |= 1;
3197 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3198 		if (rack_per_upper_bound_ss &&
3199 		    (rack->rc_dragged_bottom == 0) &&
3200 		    (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3201 			rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3202 	}
3203 	if (rack->rc_gp_saw_ca &&
3204 	    (rack->rc_gp_saw_ss == 0) &&
3205 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3206 				  rack->r_ctl.rack_per_of_gp_ca)) {
3207 		/* In CA */
3208 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3209 		if (calc > 0xffff)
3210 			calc = 0xffff;
3211 		logged |= 2;
3212 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3213 		if (rack_per_upper_bound_ca &&
3214 		    (rack->rc_dragged_bottom == 0) &&
3215 		    (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3216 			rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3217 	}
3218 	if (rack->rc_gp_saw_ss &&
3219 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3220 				  rack->r_ctl.rack_per_of_gp_ss)) {
3221 		/* In SS */
3222 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3223 		if (calc > 0xffff)
3224 			calc = 0xffff;
3225 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3226 		if (rack_per_upper_bound_ss &&
3227 		    (rack->rc_dragged_bottom == 0) &&
3228 		    (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3229 			rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3230 		logged |= 4;
3231 	}
3232 	if (logged &&
3233 	    (rack->rc_gp_incr == 0)){
3234 		/* Go into increment mode */
3235 		rack->rc_gp_incr = 1;
3236 		rack->rc_gp_timely_inc_cnt = 0;
3237 	}
3238 	if (rack->rc_gp_incr &&
3239 	    logged &&
3240 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3241 		rack->rc_gp_timely_inc_cnt++;
3242 	}
3243 	rack_log_timely(rack,  logged, plus, 0, 0,
3244 			__LINE__, 1);
3245 }
3246 
3247 static uint32_t
3248 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3249 {
3250 	/*
3251 	 * norm_grad = rtt_diff / minrtt;
3252 	 * new_per = curper * (1 - B * norm_grad)
3253 	 *
3254 	 * B = rack_gp_decrease_per (default 10%)
3255 	 * rtt_dif = input var current rtt-diff
3256 	 * curper = input var current percentage
3257 	 * minrtt = from rack filter
3258 	 *
3259 	 */
3260 	uint64_t perf;
3261 
3262 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3263 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3264 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3265 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3266 		     (uint64_t)1000000)) /
3267 		(uint64_t)1000000);
3268 	if (perf > curper) {
3269 		/* TSNH */
3270 		perf = curper - 1;
3271 	}
3272 	return ((uint32_t)perf);
3273 }
3274 
3275 static uint32_t
3276 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3277 {
3278 	/*
3279 	 *                                   highrttthresh
3280 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3281 	 *                                     gp_srtt
3282 	 *
3283 	 * B = rack_gp_decrease_per (default 10%)
3284 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3285 	 */
3286 	uint64_t perf;
3287 	uint32_t highrttthresh;
3288 
3289 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3290 
3291 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3292 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3293 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3294 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3295 	return (perf);
3296 }
3297 
3298 static void
3299 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3300 {
3301 	uint64_t logvar, logvar2, logvar3;
3302 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3303 
3304 	if (rack->rc_gp_incr) {
3305 		/* Turn off increment counting */
3306 		rack->rc_gp_incr = 0;
3307 		rack->rc_gp_timely_inc_cnt = 0;
3308 	}
3309 	ss_red = ca_red = rec_red = 0;
3310 	logged = 0;
3311 	/* Calculate the reduction value */
3312 	if (rtt_diff < 0) {
3313 		rtt_diff *= -1;
3314 	}
3315 	/* Must be at least 1% reduction */
3316 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3317 		/* We have been in recovery ding it too */
3318 		if (timely_says == 2) {
3319 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3320 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3321 			if (alt < new_per)
3322 				val = alt;
3323 			else
3324 				val = new_per;
3325 		} else
3326 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3327 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
3328 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3329 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3330 		} else {
3331 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3332 			rec_red = 0;
3333 		}
3334 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3335 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3336 		logged |= 1;
3337 	}
3338 	if (rack->rc_gp_saw_ss) {
3339 		/* Sent in SS */
3340 		if (timely_says == 2) {
3341 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3342 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3343 			if (alt < new_per)
3344 				val = alt;
3345 			else
3346 				val = new_per;
3347 		} else
3348 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3349 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3350 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3351 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3352 		} else {
3353 			ss_red = new_per;
3354 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3355 			logvar = new_per;
3356 			logvar <<= 32;
3357 			logvar |= alt;
3358 			logvar2 = (uint32_t)rtt;
3359 			logvar2 <<= 32;
3360 			logvar2 |= (uint32_t)rtt_diff;
3361 			logvar3 = rack_gp_rtt_maxmul;
3362 			logvar3 <<= 32;
3363 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3364 			rack_log_timely(rack, timely_says,
3365 					logvar2, logvar3,
3366 					logvar, __LINE__, 10);
3367 		}
3368 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3369 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3370 		logged |= 4;
3371 	} else if (rack->rc_gp_saw_ca) {
3372 		/* Sent in CA */
3373 		if (timely_says == 2) {
3374 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3375 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3376 			if (alt < new_per)
3377 				val = alt;
3378 			else
3379 				val = new_per;
3380 		} else
3381 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3382 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
3383 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3384 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3385 		} else {
3386 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3387 			ca_red = 0;
3388 			logvar = new_per;
3389 			logvar <<= 32;
3390 			logvar |= alt;
3391 			logvar2 = (uint32_t)rtt;
3392 			logvar2 <<= 32;
3393 			logvar2 |= (uint32_t)rtt_diff;
3394 			logvar3 = rack_gp_rtt_maxmul;
3395 			logvar3 <<= 32;
3396 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3397 			rack_log_timely(rack, timely_says,
3398 					logvar2, logvar3,
3399 					logvar, __LINE__, 10);
3400 		}
3401 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3402 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3403 		logged |= 2;
3404 	}
3405 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
3406 		rack->rc_gp_timely_dec_cnt++;
3407 		if (rack_timely_dec_clear &&
3408 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3409 			rack->rc_gp_timely_dec_cnt = 0;
3410 	}
3411 	logvar = ss_red;
3412 	logvar <<= 32;
3413 	logvar |= ca_red;
3414 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3415 			__LINE__, 2);
3416 }
3417 
3418 static void
3419 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3420 		     uint32_t rtt, uint32_t line, uint8_t reas)
3421 {
3422 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3423 		union tcp_log_stackspecific log;
3424 		struct timeval tv;
3425 
3426 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3427 		log.u_bbr.flex1 = line;
3428 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3429 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3430 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3431 		log.u_bbr.flex5 = rtt;
3432 		log.u_bbr.flex6 = rack->rc_highly_buffered;
3433 		log.u_bbr.flex6 <<= 1;
3434 		log.u_bbr.flex6 |= rack->forced_ack;
3435 		log.u_bbr.flex6 <<= 1;
3436 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3437 		log.u_bbr.flex6 <<= 1;
3438 		log.u_bbr.flex6 |= rack->in_probe_rtt;
3439 		log.u_bbr.flex6 <<= 1;
3440 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3441 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3442 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3443 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3444 		log.u_bbr.flex8 = reas;
3445 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3446 		log.u_bbr.delRate = rack_get_bw(rack);
3447 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3448 		log.u_bbr.cur_del_rate <<= 32;
3449 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3450 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3451 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3452 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3453 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3454 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3455 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3456 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3457 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3458 		log.u_bbr.rttProp = us_cts;
3459 		log.u_bbr.rttProp <<= 32;
3460 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3461 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3462 		    &rack->rc_inp->inp_socket->so_rcv,
3463 		    &rack->rc_inp->inp_socket->so_snd,
3464 		    BBR_LOG_RTT_SHRINKS, 0,
3465 		    0, &log, false, &rack->r_ctl.act_rcv_time);
3466 	}
3467 }
3468 
3469 static void
3470 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3471 {
3472 	uint64_t bwdp;
3473 
3474 	bwdp = rack_get_bw(rack);
3475 	bwdp *= (uint64_t)rtt;
3476 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3477 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3478 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3479 		/*
3480 		 * A window protocol must be able to have 4 packets
3481 		 * outstanding as the floor in order to function
3482 		 * (especially considering delayed ack :D).
3483 		 */
3484 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3485 	}
3486 }
3487 
3488 static void
3489 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3490 {
3491 	/**
3492 	 * ProbeRTT is a bit different in rack_pacing than in
3493 	 * BBR. It is like BBR in that it uses the lowering of
3494 	 * the RTT as a signal that we saw something new and
3495 	 * counts from there for how long between. But it is
3496 	 * different in that its quite simple. It does not
3497 	 * play with the cwnd and wait until we get down
3498 	 * to N segments outstanding and hold that for
3499 	 * 200ms. Instead it just sets the pacing reduction
3500 	 * rate to a set percentage (70 by default) and hold
3501 	 * that for a number of recent GP Srtt's.
3502 	 */
3503 	uint32_t segsiz;
3504 
3505 	if (rack->rc_gp_dyn_mul == 0)
3506 		return;
3507 
3508 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3509 		/* We are idle */
3510 		return;
3511 	}
3512 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3513 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3514 		/*
3515 		 * Stop the goodput now, the idea here is
3516 		 * that future measurements with in_probe_rtt
3517 		 * won't register if they are not greater so
3518 		 * we want to get what info (if any) is available
3519 		 * now.
3520 		 */
3521 		rack_do_goodput_measurement(rack->rc_tp, rack,
3522 					    rack->rc_tp->snd_una, __LINE__,
3523 					    RACK_QUALITY_PROBERTT);
3524 	}
3525 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3526 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3527 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3528 		     rack->r_ctl.rc_pace_min_segs);
3529 	rack->in_probe_rtt = 1;
3530 	rack->measure_saw_probe_rtt = 1;
3531 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3532 	rack->r_ctl.rc_time_probertt_starts = 0;
3533 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3534 	if (rack_probertt_use_min_rtt_entry)
3535 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3536 	else
3537 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3538 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3539 			     __LINE__, RACK_RTTS_ENTERPROBE);
3540 }
3541 
3542 static void
3543 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3544 {
3545 	struct rack_sendmap *rsm;
3546 	uint32_t segsiz;
3547 
3548 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3549 		     rack->r_ctl.rc_pace_min_segs);
3550 	rack->in_probe_rtt = 0;
3551 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3552 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3553 		/*
3554 		 * Stop the goodput now, the idea here is
3555 		 * that future measurements with in_probe_rtt
3556 		 * won't register if they are not greater so
3557 		 * we want to get what info (if any) is available
3558 		 * now.
3559 		 */
3560 		rack_do_goodput_measurement(rack->rc_tp, rack,
3561 					    rack->rc_tp->snd_una, __LINE__,
3562 					    RACK_QUALITY_PROBERTT);
3563 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3564 		/*
3565 		 * We don't have enough data to make a measurement.
3566 		 * So lets just stop and start here after exiting
3567 		 * probe-rtt. We probably are not interested in
3568 		 * the results anyway.
3569 		 */
3570 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3571 	}
3572 	/*
3573 	 * Measurements through the current snd_max are going
3574 	 * to be limited by the slower pacing rate.
3575 	 *
3576 	 * We need to mark these as app-limited so we
3577 	 * don't collapse the b/w.
3578 	 */
3579 	rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3580 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3581 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3582 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3583 		else {
3584 			/*
3585 			 * Go out to the end app limited and mark
3586 			 * this new one as next and move the end_appl up
3587 			 * to this guy.
3588 			 */
3589 			if (rack->r_ctl.rc_end_appl)
3590 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3591 			rack->r_ctl.rc_end_appl = rsm;
3592 		}
3593 		rsm->r_flags |= RACK_APP_LIMITED;
3594 		rack->r_ctl.rc_app_limited_cnt++;
3595 	}
3596 	/*
3597 	 * Now, we need to examine our pacing rate multipliers.
3598 	 * If its under 100%, we need to kick it back up to
3599 	 * 100%. We also don't let it be over our "max" above
3600 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3601 	 * Note setting clamp_atexit_prtt to 0 has the effect
3602 	 * of setting CA/SS to 100% always at exit (which is
3603 	 * the default behavior).
3604 	 */
3605 	if (rack_probertt_clear_is) {
3606 		rack->rc_gp_incr = 0;
3607 		rack->rc_gp_bwred = 0;
3608 		rack->rc_gp_timely_inc_cnt = 0;
3609 		rack->rc_gp_timely_dec_cnt = 0;
3610 	}
3611 	/* Do we do any clamping at exit? */
3612 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3613 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3614 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3615 	}
3616 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3617 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3618 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3619 	}
3620 	/*
3621 	 * Lets set rtt_diff to 0, so that we will get a "boost"
3622 	 * after exiting.
3623 	 */
3624 	rack->r_ctl.rc_rtt_diff = 0;
3625 
3626 	/* Clear all flags so we start fresh */
3627 	rack->rc_tp->t_bytes_acked = 0;
3628 	rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3629 	/*
3630 	 * If configured to, set the cwnd and ssthresh to
3631 	 * our targets.
3632 	 */
3633 	if (rack_probe_rtt_sets_cwnd) {
3634 		uint64_t ebdp;
3635 		uint32_t setto;
3636 
3637 		/* Set ssthresh so we get into CA once we hit our target */
3638 		if (rack_probertt_use_min_rtt_exit == 1) {
3639 			/* Set to min rtt */
3640 			rack_set_prtt_target(rack, segsiz,
3641 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3642 		} else if (rack_probertt_use_min_rtt_exit == 2) {
3643 			/* Set to current gp rtt */
3644 			rack_set_prtt_target(rack, segsiz,
3645 					     rack->r_ctl.rc_gp_srtt);
3646 		} else if (rack_probertt_use_min_rtt_exit == 3) {
3647 			/* Set to entry gp rtt */
3648 			rack_set_prtt_target(rack, segsiz,
3649 					     rack->r_ctl.rc_entry_gp_rtt);
3650 		} else {
3651 			uint64_t sum;
3652 			uint32_t setval;
3653 
3654 			sum = rack->r_ctl.rc_entry_gp_rtt;
3655 			sum *= 10;
3656 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3657 			if (sum >= 20) {
3658 				/*
3659 				 * A highly buffered path needs
3660 				 * cwnd space for timely to work.
3661 				 * Lets set things up as if
3662 				 * we are heading back here again.
3663 				 */
3664 				setval = rack->r_ctl.rc_entry_gp_rtt;
3665 			} else if (sum >= 15) {
3666 				/*
3667 				 * Lets take the smaller of the
3668 				 * two since we are just somewhat
3669 				 * buffered.
3670 				 */
3671 				setval = rack->r_ctl.rc_gp_srtt;
3672 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
3673 					setval = rack->r_ctl.rc_entry_gp_rtt;
3674 			} else {
3675 				/*
3676 				 * Here we are not highly buffered
3677 				 * and should pick the min we can to
3678 				 * keep from causing loss.
3679 				 */
3680 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3681 			}
3682 			rack_set_prtt_target(rack, segsiz,
3683 					     setval);
3684 		}
3685 		if (rack_probe_rtt_sets_cwnd > 1) {
3686 			/* There is a percentage here to boost */
3687 			ebdp = rack->r_ctl.rc_target_probertt_flight;
3688 			ebdp *= rack_probe_rtt_sets_cwnd;
3689 			ebdp /= 100;
3690 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3691 		} else
3692 			setto = rack->r_ctl.rc_target_probertt_flight;
3693 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3694 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3695 			/* Enforce a min */
3696 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3697 		}
3698 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
3699 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3700 	}
3701 	rack_log_rtt_shrinks(rack,  us_cts,
3702 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3703 			     __LINE__, RACK_RTTS_EXITPROBE);
3704 	/* Clear times last so log has all the info */
3705 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3706 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3707 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3708 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
3709 }
3710 
3711 static void
3712 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3713 {
3714 	/* Check in on probe-rtt */
3715 	if (rack->rc_gp_filled == 0) {
3716 		/* We do not do p-rtt unless we have gp measurements */
3717 		return;
3718 	}
3719 	if (rack->in_probe_rtt) {
3720 		uint64_t no_overflow;
3721 		uint32_t endtime, must_stay;
3722 
3723 		if (rack->r_ctl.rc_went_idle_time &&
3724 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3725 			/*
3726 			 * We went idle during prtt, just exit now.
3727 			 */
3728 			rack_exit_probertt(rack, us_cts);
3729 		} else if (rack_probe_rtt_safety_val &&
3730 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3731 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3732 			/*
3733 			 * Probe RTT safety value triggered!
3734 			 */
3735 			rack_log_rtt_shrinks(rack,  us_cts,
3736 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3737 					     __LINE__, RACK_RTTS_SAFETY);
3738 			rack_exit_probertt(rack, us_cts);
3739 		}
3740 		/* Calculate the max we will wait */
3741 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3742 		if (rack->rc_highly_buffered)
3743 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3744 		/* Calculate the min we must wait */
3745 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3746 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3747 		    TSTMP_LT(us_cts, endtime)) {
3748 			uint32_t calc;
3749 			/* Do we lower more? */
3750 no_exit:
3751 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3752 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3753 			else
3754 				calc = 0;
3755 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3756 			if (calc) {
3757 				/* Maybe */
3758 				calc *= rack_per_of_gp_probertt_reduce;
3759 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3760 				/* Limit it too */
3761 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3762 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3763 			}
3764 			/* We must reach target or the time set */
3765 			return;
3766 		}
3767 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
3768 			if ((TSTMP_LT(us_cts, must_stay) &&
3769 			     rack->rc_highly_buffered) ||
3770 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3771 			      rack->r_ctl.rc_target_probertt_flight)) {
3772 				/* We are not past the must_stay time */
3773 				goto no_exit;
3774 			}
3775 			rack_log_rtt_shrinks(rack,  us_cts,
3776 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3777 					     __LINE__, RACK_RTTS_REACHTARGET);
3778 			rack->r_ctl.rc_time_probertt_starts = us_cts;
3779 			if (rack->r_ctl.rc_time_probertt_starts == 0)
3780 				rack->r_ctl.rc_time_probertt_starts = 1;
3781 			/* Restore back to our rate we want to pace at in prtt */
3782 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3783 		}
3784 		/*
3785 		 * Setup our end time, some number of gp_srtts plus 200ms.
3786 		 */
3787 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3788 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3789 		if (rack_probertt_gpsrtt_cnt_div)
3790 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3791 		else
3792 			endtime = 0;
3793 		endtime += rack_min_probertt_hold;
3794 		endtime += rack->r_ctl.rc_time_probertt_starts;
3795 		if (TSTMP_GEQ(us_cts,  endtime)) {
3796 			/* yes, exit probertt */
3797 			rack_exit_probertt(rack, us_cts);
3798 		}
3799 
3800 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3801 		/* Go into probertt, its been too long since we went lower */
3802 		rack_enter_probertt(rack, us_cts);
3803 	}
3804 }
3805 
3806 static void
3807 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3808 		       uint32_t rtt, int32_t rtt_diff)
3809 {
3810 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
3811 	uint32_t losses;
3812 
3813 	if ((rack->rc_gp_dyn_mul == 0) ||
3814 	    (rack->use_fixed_rate) ||
3815 	    (rack->in_probe_rtt) ||
3816 	    (rack->rc_always_pace == 0)) {
3817 		/* No dynamic GP multipler in play */
3818 		return;
3819 	}
3820 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3821 	cur_bw = rack_get_bw(rack);
3822 	/* Calculate our up and down range */
3823 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3824 	up_bnd /= 100;
3825 	up_bnd += rack->r_ctl.last_gp_comp_bw;
3826 
3827 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3828 	subfr /= 100;
3829 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3830 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3831 		/*
3832 		 * This is the case where our RTT is above
3833 		 * the max target and we have been configured
3834 		 * to just do timely no bonus up stuff in that case.
3835 		 *
3836 		 * There are two configurations, set to 1, and we
3837 		 * just do timely if we are over our max. If its
3838 		 * set above 1 then we slam the multipliers down
3839 		 * to 100 and then decrement per timely.
3840 		 */
3841 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3842 				__LINE__, 3);
3843 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3844 			rack_validate_multipliers_at_or_below_100(rack);
3845 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3846 	} else if ((last_bw_est < low_bnd) && !losses) {
3847 		/*
3848 		 * We are decreasing this is a bit complicated this
3849 		 * means we are loosing ground. This could be
3850 		 * because another flow entered and we are competing
3851 		 * for b/w with it. This will push the RTT up which
3852 		 * makes timely unusable unless we want to get shoved
3853 		 * into a corner and just be backed off (the age
3854 		 * old problem with delay based CC).
3855 		 *
3856 		 * On the other hand if it was a route change we
3857 		 * would like to stay somewhat contained and not
3858 		 * blow out the buffers.
3859 		 */
3860 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3861 				__LINE__, 3);
3862 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3863 		if (rack->rc_gp_bwred == 0) {
3864 			/* Go into reduction counting */
3865 			rack->rc_gp_bwred = 1;
3866 			rack->rc_gp_timely_dec_cnt = 0;
3867 		}
3868 		if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3869 		    (timely_says == 0)) {
3870 			/*
3871 			 * Push another time with a faster pacing
3872 			 * to try to gain back (we include override to
3873 			 * get a full raise factor).
3874 			 */
3875 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3876 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3877 			    (timely_says == 0) ||
3878 			    (rack_down_raise_thresh == 0)) {
3879 				/*
3880 				 * Do an override up in b/w if we were
3881 				 * below the threshold or if the threshold
3882 				 * is zero we always do the raise.
3883 				 */
3884 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3885 			} else {
3886 				/* Log it stays the same */
3887 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3888 						__LINE__, 11);
3889 			}
3890 			rack->rc_gp_timely_dec_cnt++;
3891 			/* We are not incrementing really no-count */
3892 			rack->rc_gp_incr = 0;
3893 			rack->rc_gp_timely_inc_cnt = 0;
3894 		} else {
3895 			/*
3896 			 * Lets just use the RTT
3897 			 * information and give up
3898 			 * pushing.
3899 			 */
3900 			goto use_timely;
3901 		}
3902 	} else if ((timely_says != 2) &&
3903 		    !losses &&
3904 		    (last_bw_est > up_bnd)) {
3905 		/*
3906 		 * We are increasing b/w lets keep going, updating
3907 		 * our b/w and ignoring any timely input, unless
3908 		 * of course we are at our max raise (if there is one).
3909 		 */
3910 
3911 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3912 				__LINE__, 3);
3913 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3914 		if (rack->rc_gp_saw_ss &&
3915 		    rack_per_upper_bound_ss &&
3916 		     (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3917 			    /*
3918 			     * In cases where we can't go higher
3919 			     * we should just use timely.
3920 			     */
3921 			    goto use_timely;
3922 		}
3923 		if (rack->rc_gp_saw_ca &&
3924 		    rack_per_upper_bound_ca &&
3925 		    (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
3926 			    /*
3927 			     * In cases where we can't go higher
3928 			     * we should just use timely.
3929 			     */
3930 			    goto use_timely;
3931 		}
3932 		rack->rc_gp_bwred = 0;
3933 		rack->rc_gp_timely_dec_cnt = 0;
3934 		/* You get a set number of pushes if timely is trying to reduce */
3935 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
3936 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3937 		} else {
3938 			/* Log it stays the same */
3939 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
3940 			    __LINE__, 12);
3941 		}
3942 		return;
3943 	} else {
3944 		/*
3945 		 * We are staying between the lower and upper range bounds
3946 		 * so use timely to decide.
3947 		 */
3948 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3949 				__LINE__, 3);
3950 use_timely:
3951 		if (timely_says) {
3952 			rack->rc_gp_incr = 0;
3953 			rack->rc_gp_timely_inc_cnt = 0;
3954 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
3955 			    !losses &&
3956 			    (last_bw_est < low_bnd)) {
3957 				/* We are loosing ground */
3958 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3959 				rack->rc_gp_timely_dec_cnt++;
3960 				/* We are not incrementing really no-count */
3961 				rack->rc_gp_incr = 0;
3962 				rack->rc_gp_timely_inc_cnt = 0;
3963 			} else
3964 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3965 		} else {
3966 			rack->rc_gp_bwred = 0;
3967 			rack->rc_gp_timely_dec_cnt = 0;
3968 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3969 		}
3970 	}
3971 }
3972 
3973 static int32_t
3974 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
3975 {
3976 	int32_t timely_says;
3977 	uint64_t log_mult, log_rtt_a_diff;
3978 
3979 	log_rtt_a_diff = rtt;
3980 	log_rtt_a_diff <<= 32;
3981 	log_rtt_a_diff |= (uint32_t)rtt_diff;
3982 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
3983 		    rack_gp_rtt_maxmul)) {
3984 		/* Reduce the b/w multipler */
3985 		timely_says = 2;
3986 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3987 		log_mult <<= 32;
3988 		log_mult |= prev_rtt;
3989 		rack_log_timely(rack,  timely_says, log_mult,
3990 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3991 				log_rtt_a_diff, __LINE__, 4);
3992 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3993 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
3994 			    max(rack_gp_rtt_mindiv , 1)))) {
3995 		/* Increase the b/w multipler */
3996 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3997 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
3998 			 max(rack_gp_rtt_mindiv , 1));
3999 		log_mult <<= 32;
4000 		log_mult |= prev_rtt;
4001 		timely_says = 0;
4002 		rack_log_timely(rack,  timely_says, log_mult ,
4003 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4004 				log_rtt_a_diff, __LINE__, 5);
4005 	} else {
4006 		/*
4007 		 * Use a gradient to find it the timely gradient
4008 		 * is:
4009 		 * grad = rc_rtt_diff / min_rtt;
4010 		 *
4011 		 * anything below or equal to 0 will be
4012 		 * a increase indication. Anything above
4013 		 * zero is a decrease. Note we take care
4014 		 * of the actual gradient calculation
4015 		 * in the reduction (its not needed for
4016 		 * increase).
4017 		 */
4018 		log_mult = prev_rtt;
4019 		if (rtt_diff <= 0) {
4020 			/*
4021 			 * Rttdiff is less than zero, increase the
4022 			 * b/w multipler (its 0 or negative)
4023 			 */
4024 			timely_says = 0;
4025 			rack_log_timely(rack,  timely_says, log_mult,
4026 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4027 		} else {
4028 			/* Reduce the b/w multipler */
4029 			timely_says = 1;
4030 			rack_log_timely(rack,  timely_says, log_mult,
4031 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4032 		}
4033 	}
4034 	return (timely_says);
4035 }
4036 
4037 static void
4038 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4039 			    tcp_seq th_ack, int line, uint8_t quality)
4040 {
4041 	uint64_t tim, bytes_ps, ltim, stim, utim;
4042 	uint32_t segsiz, bytes, reqbytes, us_cts;
4043 	int32_t gput, new_rtt_diff, timely_says;
4044 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4045 	int did_add = 0;
4046 
4047 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4048 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4049 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4050 		tim = us_cts - tp->gput_ts;
4051 	else
4052 		tim = 0;
4053 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4054 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4055 	else
4056 		stim = 0;
4057 	/*
4058 	 * Use the larger of the send time or ack time. This prevents us
4059 	 * from being influenced by ack artifacts to come up with too
4060 	 * high of measurement. Note that since we are spanning over many more
4061 	 * bytes in most of our measurements hopefully that is less likely to
4062 	 * occur.
4063 	 */
4064 	if (tim > stim)
4065 		utim = max(tim, 1);
4066 	else
4067 		utim = max(stim, 1);
4068 	/* Lets get a msec time ltim too for the old stuff */
4069 	ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4070 	gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4071 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4072 	if ((tim == 0) && (stim == 0)) {
4073 		/*
4074 		 * Invalid measurement time, maybe
4075 		 * all on one ack/one send?
4076 		 */
4077 		bytes = 0;
4078 		bytes_ps = 0;
4079 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4080 					   0, 0, 0, 10, __LINE__, NULL, quality);
4081 		goto skip_measurement;
4082 	}
4083 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4084 		/* We never made a us_rtt measurement? */
4085 		bytes = 0;
4086 		bytes_ps = 0;
4087 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4088 					   0, 0, 0, 10, __LINE__, NULL, quality);
4089 		goto skip_measurement;
4090 	}
4091 	/*
4092 	 * Calculate the maximum possible b/w this connection
4093 	 * could have. We base our calculation on the lowest
4094 	 * rtt we have seen during the measurement and the
4095 	 * largest rwnd the client has given us in that time. This
4096 	 * forms a BDP that is the maximum that we could ever
4097 	 * get to the client. Anything larger is not valid.
4098 	 *
4099 	 * I originally had code here that rejected measurements
4100 	 * where the time was less than 1/2 the latest us_rtt.
4101 	 * But after thinking on that I realized its wrong since
4102 	 * say you had a 150Mbps or even 1Gbps link, and you
4103 	 * were a long way away.. example I am in Europe (100ms rtt)
4104 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4105 	 * bytes my time would be 1.2ms, and yet my rtt would say
4106 	 * the measurement was invalid the time was < 50ms. The
4107 	 * same thing is true for 150Mb (8ms of time).
4108 	 *
4109 	 * A better way I realized is to look at what the maximum
4110 	 * the connection could possibly do. This is gated on
4111 	 * the lowest RTT we have seen and the highest rwnd.
4112 	 * We should in theory never exceed that, if we are
4113 	 * then something on the path is storing up packets
4114 	 * and then feeding them all at once to our endpoint
4115 	 * messing up our measurement.
4116 	 */
4117 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4118 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4119 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4120 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4121 		/* No measurement can be made */
4122 		bytes = 0;
4123 		bytes_ps = 0;
4124 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4125 					   0, 0, 0, 10, __LINE__, NULL, quality);
4126 		goto skip_measurement;
4127 	} else
4128 		bytes = (th_ack - tp->gput_seq);
4129 	bytes_ps = (uint64_t)bytes;
4130 	/*
4131 	 * Don't measure a b/w for pacing unless we have gotten at least
4132 	 * an initial windows worth of data in this measurement interval.
4133 	 *
4134 	 * Small numbers of bytes get badly influenced by delayed ack and
4135 	 * other artifacts. Note we take the initial window or our
4136 	 * defined minimum GP (defaulting to 10 which hopefully is the
4137 	 * IW).
4138 	 */
4139 	if (rack->rc_gp_filled == 0) {
4140 		/*
4141 		 * The initial estimate is special. We
4142 		 * have blasted out an IW worth of packets
4143 		 * without a real valid ack ts results. We
4144 		 * then setup the app_limited_needs_set flag,
4145 		 * this should get the first ack in (probably 2
4146 		 * MSS worth) to be recorded as the timestamp.
4147 		 * We thus allow a smaller number of bytes i.e.
4148 		 * IW - 2MSS.
4149 		 */
4150 		reqbytes -= (2 * segsiz);
4151 		/* Also lets fill previous for our first measurement to be neutral */
4152 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4153 	}
4154 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4155 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4156 					   rack->r_ctl.rc_app_limited_cnt,
4157 					   0, 0, 10, __LINE__, NULL, quality);
4158 		goto skip_measurement;
4159 	}
4160 	/*
4161 	 * We now need to calculate the Timely like status so
4162 	 * we can update (possibly) the b/w multipliers.
4163 	 */
4164 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4165 	if (rack->rc_gp_filled == 0) {
4166 		/* No previous reading */
4167 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4168 	} else {
4169 		if (rack->measure_saw_probe_rtt == 0) {
4170 			/*
4171 			 * We don't want a probertt to be counted
4172 			 * since it will be negative incorrectly. We
4173 			 * expect to be reducing the RTT when we
4174 			 * pace at a slower rate.
4175 			 */
4176 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4177 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4178 		}
4179 	}
4180 	timely_says = rack_make_timely_judgement(rack,
4181 		rack->r_ctl.rc_gp_srtt,
4182 		rack->r_ctl.rc_rtt_diff,
4183 	        rack->r_ctl.rc_prev_gp_srtt
4184 		);
4185 	bytes_ps *= HPTS_USEC_IN_SEC;
4186 	bytes_ps /= utim;
4187 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4188 		/*
4189 		 * Something is on path playing
4190 		 * since this b/w is not possible based
4191 		 * on our BDP (highest rwnd and lowest rtt
4192 		 * we saw in the measurement window).
4193 		 *
4194 		 * Another option here would be to
4195 		 * instead skip the measurement.
4196 		 */
4197 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4198 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4199 					   11, __LINE__, NULL, quality);
4200 		bytes_ps = rack->r_ctl.last_max_bw;
4201 	}
4202 	/* We store gp for b/w in bytes per second */
4203 	if (rack->rc_gp_filled == 0) {
4204 		/* Initial measurement */
4205 		if (bytes_ps) {
4206 			rack->r_ctl.gp_bw = bytes_ps;
4207 			rack->rc_gp_filled = 1;
4208 			rack->r_ctl.num_measurements = 1;
4209 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4210 		} else {
4211 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4212 						   rack->r_ctl.rc_app_limited_cnt,
4213 						   0, 0, 10, __LINE__, NULL, quality);
4214 		}
4215 		if (tcp_in_hpts(rack->rc_inp) &&
4216 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4217 			/*
4218 			 * Ok we can't trust the pacer in this case
4219 			 * where we transition from un-paced to paced.
4220 			 * Or for that matter when the burst mitigation
4221 			 * was making a wild guess and got it wrong.
4222 			 * Stop the pacer and clear up all the aggregate
4223 			 * delays etc.
4224 			 */
4225 			tcp_hpts_remove(rack->rc_inp);
4226 			rack->r_ctl.rc_hpts_flags = 0;
4227 			rack->r_ctl.rc_last_output_to = 0;
4228 		}
4229 		did_add = 2;
4230 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4231 		/* Still a small number run an average */
4232 		rack->r_ctl.gp_bw += bytes_ps;
4233 		addpart = rack->r_ctl.num_measurements;
4234 		rack->r_ctl.num_measurements++;
4235 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4236 			/* We have collected enought to move forward */
4237 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4238 		}
4239 		did_add = 3;
4240 	} else {
4241 		/*
4242 		 * We want to take 1/wma of the goodput and add in to 7/8th
4243 		 * of the old value weighted by the srtt. So if your measurement
4244 		 * period is say 2 SRTT's long you would get 1/4 as the
4245 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
4246 		 *
4247 		 * But we must be careful not to take too much i.e. if the
4248 		 * srtt is say 20ms and the measurement is taken over
4249 		 * 400ms our weight would be 400/20 i.e. 20. On the
4250 		 * other hand if we get a measurement over 1ms with a
4251 		 * 10ms rtt we only want to take a much smaller portion.
4252 		 */
4253 		if (rack->r_ctl.num_measurements < 0xff) {
4254 			rack->r_ctl.num_measurements++;
4255 		}
4256 		srtt = (uint64_t)tp->t_srtt;
4257 		if (srtt == 0) {
4258 			/*
4259 			 * Strange why did t_srtt go back to zero?
4260 			 */
4261 			if (rack->r_ctl.rc_rack_min_rtt)
4262 				srtt = rack->r_ctl.rc_rack_min_rtt;
4263 			else
4264 				srtt = HPTS_USEC_IN_MSEC;
4265 		}
4266 		/*
4267 		 * XXXrrs: Note for reviewers, in playing with
4268 		 * dynamic pacing I discovered this GP calculation
4269 		 * as done originally leads to some undesired results.
4270 		 * Basically you can get longer measurements contributing
4271 		 * too much to the WMA. Thus I changed it if you are doing
4272 		 * dynamic adjustments to only do the aportioned adjustment
4273 		 * if we have a very small (time wise) measurement. Longer
4274 		 * measurements just get there weight (defaulting to 1/8)
4275 		 * add to the WMA. We may want to think about changing
4276 		 * this to always do that for both sides i.e. dynamic
4277 		 * and non-dynamic... but considering lots of folks
4278 		 * were playing with this I did not want to change the
4279 		 * calculation per.se. without your thoughts.. Lawerence?
4280 		 * Peter??
4281 		 */
4282 		if (rack->rc_gp_dyn_mul == 0) {
4283 			subpart = rack->r_ctl.gp_bw * utim;
4284 			subpart /= (srtt * 8);
4285 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
4286 				/*
4287 				 * The b/w update takes no more
4288 				 * away then 1/2 our running total
4289 				 * so factor it in.
4290 				 */
4291 				addpart = bytes_ps * utim;
4292 				addpart /= (srtt * 8);
4293 			} else {
4294 				/*
4295 				 * Don't allow a single measurement
4296 				 * to account for more than 1/2 of the
4297 				 * WMA. This could happen on a retransmission
4298 				 * where utim becomes huge compared to
4299 				 * srtt (multiple retransmissions when using
4300 				 * the sending rate which factors in all the
4301 				 * transmissions from the first one).
4302 				 */
4303 				subpart = rack->r_ctl.gp_bw / 2;
4304 				addpart = bytes_ps / 2;
4305 			}
4306 			resid_bw = rack->r_ctl.gp_bw - subpart;
4307 			rack->r_ctl.gp_bw = resid_bw + addpart;
4308 			did_add = 1;
4309 		} else {
4310 			if ((utim / srtt) <= 1) {
4311 				/*
4312 				 * The b/w update was over a small period
4313 				 * of time. The idea here is to prevent a small
4314 				 * measurement time period from counting
4315 				 * too much. So we scale it based on the
4316 				 * time so it attributes less than 1/rack_wma_divisor
4317 				 * of its measurement.
4318 				 */
4319 				subpart = rack->r_ctl.gp_bw * utim;
4320 				subpart /= (srtt * rack_wma_divisor);
4321 				addpart = bytes_ps * utim;
4322 				addpart /= (srtt * rack_wma_divisor);
4323 			} else {
4324 				/*
4325 				 * The scaled measurement was long
4326 				 * enough so lets just add in the
4327 				 * portion of the measurement i.e. 1/rack_wma_divisor
4328 				 */
4329 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4330 				addpart = bytes_ps / rack_wma_divisor;
4331 			}
4332 			if ((rack->measure_saw_probe_rtt == 0) ||
4333 		            (bytes_ps > rack->r_ctl.gp_bw)) {
4334 				/*
4335 				 * For probe-rtt we only add it in
4336 				 * if its larger, all others we just
4337 				 * add in.
4338 				 */
4339 				did_add = 1;
4340 				resid_bw = rack->r_ctl.gp_bw - subpart;
4341 				rack->r_ctl.gp_bw = resid_bw + addpart;
4342 			}
4343 		}
4344 	}
4345 	if ((rack->gp_ready == 0) &&
4346 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4347 		/* We have enough measurements now */
4348 		rack->gp_ready = 1;
4349 		rack_set_cc_pacing(rack);
4350 		if (rack->defer_options)
4351 			rack_apply_deferred_options(rack);
4352 	}
4353 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4354 				   rack_get_bw(rack), 22, did_add, NULL, quality);
4355 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
4356 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4357 		rack_update_multiplier(rack, timely_says, bytes_ps,
4358 				       rack->r_ctl.rc_gp_srtt,
4359 				       rack->r_ctl.rc_rtt_diff);
4360 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4361 				   rack_get_bw(rack), 3, line, NULL, quality);
4362 	/* reset the gp srtt and setup the new prev */
4363 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4364 	/* Record the lost count for the next measurement */
4365 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4366 	/*
4367 	 * We restart our diffs based on the gpsrtt in the
4368 	 * measurement window.
4369 	 */
4370 	rack->rc_gp_rtt_set = 0;
4371 	rack->rc_gp_saw_rec = 0;
4372 	rack->rc_gp_saw_ca = 0;
4373 	rack->rc_gp_saw_ss = 0;
4374 	rack->rc_dragged_bottom = 0;
4375 skip_measurement:
4376 
4377 #ifdef STATS
4378 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4379 				 gput);
4380 	/*
4381 	 * XXXLAS: This is a temporary hack, and should be
4382 	 * chained off VOI_TCP_GPUT when stats(9) grows an
4383 	 * API to deal with chained VOIs.
4384 	 */
4385 	if (tp->t_stats_gput_prev > 0)
4386 		stats_voi_update_abs_s32(tp->t_stats,
4387 					 VOI_TCP_GPUT_ND,
4388 					 ((gput - tp->t_stats_gput_prev) * 100) /
4389 					 tp->t_stats_gput_prev);
4390 #endif
4391 	tp->t_flags &= ~TF_GPUTINPROG;
4392 	tp->t_stats_gput_prev = gput;
4393 	/*
4394 	 * Now are we app limited now and there is space from where we
4395 	 * were to where we want to go?
4396 	 *
4397 	 * We don't do the other case i.e. non-applimited here since
4398 	 * the next send will trigger us picking up the missing data.
4399 	 */
4400 	if (rack->r_ctl.rc_first_appl &&
4401 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
4402 	    rack->r_ctl.rc_app_limited_cnt &&
4403 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4404 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4405 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4406 		/*
4407 		 * Yep there is enough outstanding to make a measurement here.
4408 		 */
4409 		struct rack_sendmap *rsm, fe;
4410 
4411 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4412 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4413 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4414 		rack->app_limited_needs_set = 0;
4415 		tp->gput_seq = th_ack;
4416 		if (rack->in_probe_rtt)
4417 			rack->measure_saw_probe_rtt = 1;
4418 		else if ((rack->measure_saw_probe_rtt) &&
4419 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4420 			rack->measure_saw_probe_rtt = 0;
4421 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4422 			/* There is a full window to gain info from */
4423 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4424 		} else {
4425 			/* We can only measure up to the applimited point */
4426 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4427 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4428 				/*
4429 				 * We don't have enough to make a measurement.
4430 				 */
4431 				tp->t_flags &= ~TF_GPUTINPROG;
4432 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4433 							   0, 0, 0, 6, __LINE__, NULL, quality);
4434 				return;
4435 			}
4436 		}
4437 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
4438 			/*
4439 			 * We will get no more data into the SB
4440 			 * this means we need to have the data available
4441 			 * before we start a measurement.
4442 			 */
4443 			if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4444 				/* Nope not enough data. */
4445 				return;
4446 			}
4447 		}
4448 		tp->t_flags |= TF_GPUTINPROG;
4449 		/*
4450 		 * Now we need to find the timestamp of the send at tp->gput_seq
4451 		 * for the send based measurement.
4452 		 */
4453 		fe.r_start = tp->gput_seq;
4454 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4455 		if (rsm) {
4456 			/* Ok send-based limit is set */
4457 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4458 				/*
4459 				 * Move back to include the earlier part
4460 				 * so our ack time lines up right (this may
4461 				 * make an overlapping measurement but thats
4462 				 * ok).
4463 				 */
4464 				tp->gput_seq = rsm->r_start;
4465 			}
4466 			if (rsm->r_flags & RACK_ACKED)
4467 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4468 			else
4469 				rack->app_limited_needs_set = 1;
4470 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4471 		} else {
4472 			/*
4473 			 * If we don't find the rsm due to some
4474 			 * send-limit set the current time, which
4475 			 * basically disables the send-limit.
4476 			 */
4477 			struct timeval tv;
4478 
4479 			microuptime(&tv);
4480 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4481 		}
4482 		rack_log_pacing_delay_calc(rack,
4483 					   tp->gput_seq,
4484 					   tp->gput_ack,
4485 					   (uint64_t)rsm,
4486 					   tp->gput_ts,
4487 					   rack->r_ctl.rc_app_limited_cnt,
4488 					   9,
4489 					   __LINE__, NULL, quality);
4490 	}
4491 }
4492 
4493 /*
4494  * CC wrapper hook functions
4495  */
4496 static void
4497 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4498     uint16_t type, int32_t recovery)
4499 {
4500 	uint32_t prior_cwnd, acked;
4501 	struct tcp_log_buffer *lgb = NULL;
4502 	uint8_t labc_to_use, quality;
4503 
4504 	INP_WLOCK_ASSERT(tp->t_inpcb);
4505 	tp->ccv->nsegs = nsegs;
4506 	acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4507 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4508 		uint32_t max;
4509 
4510 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4511 		if (tp->ccv->bytes_this_ack > max) {
4512 			tp->ccv->bytes_this_ack = max;
4513 		}
4514 	}
4515 #ifdef STATS
4516 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4517 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4518 #endif
4519 	quality = RACK_QUALITY_NONE;
4520 	if ((tp->t_flags & TF_GPUTINPROG) &&
4521 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4522 		/* Measure the Goodput */
4523 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4524 #ifdef NETFLIX_PEAKRATE
4525 		if ((type == CC_ACK) &&
4526 		    (tp->t_maxpeakrate)) {
4527 			/*
4528 			 * We update t_peakrate_thr. This gives us roughly
4529 			 * one update per round trip time. Note
4530 			 * it will only be used if pace_always is off i.e
4531 			 * we don't do this for paced flows.
4532 			 */
4533 			rack_update_peakrate_thr(tp);
4534 		}
4535 #endif
4536 	}
4537 	/* Which way our we limited, if not cwnd limited no advance in CA */
4538 	if (tp->snd_cwnd <= tp->snd_wnd)
4539 		tp->ccv->flags |= CCF_CWND_LIMITED;
4540 	else
4541 		tp->ccv->flags &= ~CCF_CWND_LIMITED;
4542 	if (tp->snd_cwnd > tp->snd_ssthresh) {
4543 		tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4544 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4545 		/* For the setting of a window past use the actual scwnd we are using */
4546 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4547 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4548 			tp->ccv->flags |= CCF_ABC_SENTAWND;
4549 		}
4550 	} else {
4551 		tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4552 		tp->t_bytes_acked = 0;
4553 	}
4554 	prior_cwnd = tp->snd_cwnd;
4555 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4556 	    (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4557 		labc_to_use = rack->rc_labc;
4558 	else
4559 		labc_to_use = rack_max_abc_post_recovery;
4560 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4561 		union tcp_log_stackspecific log;
4562 		struct timeval tv;
4563 
4564 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4565 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4566 		log.u_bbr.flex1 = th_ack;
4567 		log.u_bbr.flex2 = tp->ccv->flags;
4568 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4569 		log.u_bbr.flex4 = tp->ccv->nsegs;
4570 		log.u_bbr.flex5 = labc_to_use;
4571 		log.u_bbr.flex6 = prior_cwnd;
4572 		log.u_bbr.flex7 = V_tcp_do_newsack;
4573 		log.u_bbr.flex8 = 1;
4574 		lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4575 				     0, &log, false, NULL, NULL, 0, &tv);
4576 	}
4577 	if (CC_ALGO(tp)->ack_received != NULL) {
4578 		/* XXXLAS: Find a way to live without this */
4579 		tp->ccv->curack = th_ack;
4580 		tp->ccv->labc = labc_to_use;
4581 		tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4582 		CC_ALGO(tp)->ack_received(tp->ccv, type);
4583 	}
4584 	if (lgb) {
4585 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4586 	}
4587 	if (rack->r_must_retran) {
4588 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4589 			/*
4590 			 * We now are beyond the rxt point so lets disable
4591 			 * the flag.
4592 			 */
4593 			rack->r_ctl.rc_out_at_rto = 0;
4594 			rack->r_must_retran = 0;
4595 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4596 			/*
4597 			 * Only decrement the rc_out_at_rto if the cwnd advances
4598 			 * at least a whole segment. Otherwise next time the peer
4599 			 * acks, we won't be able to send this generaly happens
4600 			 * when we are in Congestion Avoidance.
4601 			 */
4602 			if (acked <= rack->r_ctl.rc_out_at_rto){
4603 				rack->r_ctl.rc_out_at_rto -= acked;
4604 			} else {
4605 				rack->r_ctl.rc_out_at_rto = 0;
4606 			}
4607 		}
4608 	}
4609 #ifdef STATS
4610 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4611 #endif
4612 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4613 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4614 	}
4615 #ifdef NETFLIX_PEAKRATE
4616 	/* we enforce max peak rate if it is set and we are not pacing */
4617 	if ((rack->rc_always_pace == 0) &&
4618 	    tp->t_peakrate_thr &&
4619 	    (tp->snd_cwnd > tp->t_peakrate_thr)) {
4620 		tp->snd_cwnd = tp->t_peakrate_thr;
4621 	}
4622 #endif
4623 }
4624 
4625 static void
4626 tcp_rack_partialack(struct tcpcb *tp)
4627 {
4628 	struct tcp_rack *rack;
4629 
4630 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4631 	INP_WLOCK_ASSERT(tp->t_inpcb);
4632 	/*
4633 	 * If we are doing PRR and have enough
4634 	 * room to send <or> we are pacing and prr
4635 	 * is disabled we will want to see if we
4636 	 * can send data (by setting r_wanted_output to
4637 	 * true).
4638 	 */
4639 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4640 	    rack->rack_no_prr)
4641 		rack->r_wanted_output = 1;
4642 }
4643 
4644 static void
4645 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4646 {
4647 	struct tcp_rack *rack;
4648 	uint32_t orig_cwnd;
4649 
4650 	orig_cwnd = tp->snd_cwnd;
4651 	INP_WLOCK_ASSERT(tp->t_inpcb);
4652 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4653 	/* only alert CC if we alerted when we entered */
4654 	if (CC_ALGO(tp)->post_recovery != NULL) {
4655 		tp->ccv->curack = th_ack;
4656 		CC_ALGO(tp)->post_recovery(tp->ccv);
4657 		if (tp->snd_cwnd < tp->snd_ssthresh) {
4658 			/*
4659 			 * Rack has burst control and pacing
4660 			 * so lets not set this any lower than
4661 			 * snd_ssthresh per RFC-6582 (option 2).
4662 			 */
4663 			tp->snd_cwnd = tp->snd_ssthresh;
4664 		}
4665 	}
4666 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4667 		union tcp_log_stackspecific log;
4668 		struct timeval tv;
4669 
4670 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4671 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4672 		log.u_bbr.flex1 = th_ack;
4673 		log.u_bbr.flex2 = tp->ccv->flags;
4674 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4675 		log.u_bbr.flex4 = tp->ccv->nsegs;
4676 		log.u_bbr.flex5 = V_tcp_abc_l_var;
4677 		log.u_bbr.flex6 = orig_cwnd;
4678 		log.u_bbr.flex7 = V_tcp_do_newsack;
4679 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4680 		log.u_bbr.flex8 = 2;
4681 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4682 			       0, &log, false, NULL, NULL, 0, &tv);
4683 	}
4684 	if ((rack->rack_no_prr == 0) &&
4685 	    (rack->no_prr_addback == 0) &&
4686 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
4687 		/*
4688 		 * Suck the next prr cnt back into cwnd, but
4689 		 * only do that if we are not application limited.
4690 		 */
4691 		if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4692 			/*
4693 			 * We are allowed to add back to the cwnd the amount we did
4694 			 * not get out if:
4695 			 * a) no_prr_addback is off.
4696 			 * b) we are not app limited
4697 			 * c) we are doing prr
4698 			 * <and>
4699 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4700 			 */
4701 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4702 					    rack->r_ctl.rc_prr_sndcnt);
4703 		}
4704 		rack->r_ctl.rc_prr_sndcnt = 0;
4705 		rack_log_to_prr(rack, 1, 0);
4706 	}
4707 	rack_log_to_prr(rack, 14, orig_cwnd);
4708 	tp->snd_recover = tp->snd_una;
4709 	if (rack->r_ctl.dsack_persist) {
4710 		rack->r_ctl.dsack_persist--;
4711 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4712 			rack->r_ctl.num_dsack = 0;
4713 		}
4714 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4715 	}
4716 	EXIT_RECOVERY(tp->t_flags);
4717 }
4718 
4719 static void
4720 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack)
4721 {
4722 	struct tcp_rack *rack;
4723 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4724 
4725 	INP_WLOCK_ASSERT(tp->t_inpcb);
4726 #ifdef STATS
4727 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4728 #endif
4729 	if (IN_RECOVERY(tp->t_flags) == 0) {
4730 		in_rec_at_entry = 0;
4731 		ssthresh_enter = tp->snd_ssthresh;
4732 		cwnd_enter = tp->snd_cwnd;
4733 	} else
4734 		in_rec_at_entry = 1;
4735 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4736 	switch (type) {
4737 	case CC_NDUPACK:
4738 		tp->t_flags &= ~TF_WASFRECOVERY;
4739 		tp->t_flags &= ~TF_WASCRECOVERY;
4740 		if (!IN_FASTRECOVERY(tp->t_flags)) {
4741 			rack->r_ctl.rc_prr_delivered = 0;
4742 			rack->r_ctl.rc_prr_out = 0;
4743 			if (rack->rack_no_prr == 0) {
4744 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4745 				rack_log_to_prr(rack, 2, in_rec_at_entry);
4746 			}
4747 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4748 			tp->snd_recover = tp->snd_max;
4749 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4750 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4751 		}
4752 		break;
4753 	case CC_ECN:
4754 		if (!IN_CONGRECOVERY(tp->t_flags) ||
4755 		    /*
4756 		     * Allow ECN reaction on ACK to CWR, if
4757 		     * that data segment was also CE marked.
4758 		     */
4759 		    SEQ_GEQ(ack, tp->snd_recover)) {
4760 			EXIT_CONGRECOVERY(tp->t_flags);
4761 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4762 			tp->snd_recover = tp->snd_max + 1;
4763 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4764 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4765 		}
4766 		break;
4767 	case CC_RTO:
4768 		tp->t_dupacks = 0;
4769 		tp->t_bytes_acked = 0;
4770 		EXIT_RECOVERY(tp->t_flags);
4771 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4772 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4773 		orig_cwnd = tp->snd_cwnd;
4774 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
4775 		rack_log_to_prr(rack, 16, orig_cwnd);
4776 		if (tp->t_flags2 & TF2_ECN_PERMIT)
4777 			tp->t_flags2 |= TF2_ECN_SND_CWR;
4778 		break;
4779 	case CC_RTO_ERR:
4780 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4781 		/* RTO was unnecessary, so reset everything. */
4782 		tp->snd_cwnd = tp->snd_cwnd_prev;
4783 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
4784 		tp->snd_recover = tp->snd_recover_prev;
4785 		if (tp->t_flags & TF_WASFRECOVERY) {
4786 			ENTER_FASTRECOVERY(tp->t_flags);
4787 			tp->t_flags &= ~TF_WASFRECOVERY;
4788 		}
4789 		if (tp->t_flags & TF_WASCRECOVERY) {
4790 			ENTER_CONGRECOVERY(tp->t_flags);
4791 			tp->t_flags &= ~TF_WASCRECOVERY;
4792 		}
4793 		tp->snd_nxt = tp->snd_max;
4794 		tp->t_badrxtwin = 0;
4795 		break;
4796 	}
4797 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4798 	    (type != CC_RTO)){
4799 		tp->ccv->curack = ack;
4800 		CC_ALGO(tp)->cong_signal(tp->ccv, type);
4801 	}
4802 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4803 		rack_log_to_prr(rack, 15, cwnd_enter);
4804 		rack->r_ctl.dsack_byte_cnt = 0;
4805 		rack->r_ctl.retran_during_recovery = 0;
4806 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4807 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4808 		rack->r_ent_rec_ns = 1;
4809 	}
4810 }
4811 
4812 static inline void
4813 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4814 {
4815 	uint32_t i_cwnd;
4816 
4817 	INP_WLOCK_ASSERT(tp->t_inpcb);
4818 
4819 #ifdef NETFLIX_STATS
4820 	KMOD_TCPSTAT_INC(tcps_idle_restarts);
4821 	if (tp->t_state == TCPS_ESTABLISHED)
4822 		KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4823 #endif
4824 	if (CC_ALGO(tp)->after_idle != NULL)
4825 		CC_ALGO(tp)->after_idle(tp->ccv);
4826 
4827 	if (tp->snd_cwnd == 1)
4828 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
4829 	else
4830 		i_cwnd = rc_init_window(rack);
4831 
4832 	/*
4833 	 * Being idle is no different than the initial window. If the cc
4834 	 * clamps it down below the initial window raise it to the initial
4835 	 * window.
4836 	 */
4837 	if (tp->snd_cwnd < i_cwnd) {
4838 		tp->snd_cwnd = i_cwnd;
4839 	}
4840 }
4841 
4842 /*
4843  * Indicate whether this ack should be delayed.  We can delay the ack if
4844  * following conditions are met:
4845  *	- There is no delayed ack timer in progress.
4846  *	- Our last ack wasn't a 0-sized window. We never want to delay
4847  *	  the ack that opens up a 0-sized window.
4848  *	- LRO wasn't used for this segment. We make sure by checking that the
4849  *	  segment size is not larger than the MSS.
4850  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
4851  *	  connection.
4852  */
4853 #define DELAY_ACK(tp, tlen)			 \
4854 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4855 	((tp->t_flags & TF_DELACK) == 0) &&	 \
4856 	(tlen <= tp->t_maxseg) &&		 \
4857 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4858 
4859 static struct rack_sendmap *
4860 rack_find_lowest_rsm(struct tcp_rack *rack)
4861 {
4862 	struct rack_sendmap *rsm;
4863 
4864 	/*
4865 	 * Walk the time-order transmitted list looking for an rsm that is
4866 	 * not acked. This will be the one that was sent the longest time
4867 	 * ago that is still outstanding.
4868 	 */
4869 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4870 		if (rsm->r_flags & RACK_ACKED) {
4871 			continue;
4872 		}
4873 		goto finish;
4874 	}
4875 finish:
4876 	return (rsm);
4877 }
4878 
4879 static struct rack_sendmap *
4880 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4881 {
4882 	struct rack_sendmap *prsm;
4883 
4884 	/*
4885 	 * Walk the sequence order list backward until we hit and arrive at
4886 	 * the highest seq not acked. In theory when this is called it
4887 	 * should be the last segment (which it was not).
4888 	 */
4889 	prsm = rsm;
4890 	RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4891 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4892 			continue;
4893 		}
4894 		return (prsm);
4895 	}
4896 	return (NULL);
4897 }
4898 
4899 static uint32_t
4900 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4901 {
4902 	int32_t lro;
4903 	uint32_t thresh;
4904 
4905 	/*
4906 	 * lro is the flag we use to determine if we have seen reordering.
4907 	 * If it gets set we have seen reordering. The reorder logic either
4908 	 * works in one of two ways:
4909 	 *
4910 	 * If reorder-fade is configured, then we track the last time we saw
4911 	 * re-ordering occur. If we reach the point where enough time as
4912 	 * passed we no longer consider reordering has occuring.
4913 	 *
4914 	 * Or if reorder-face is 0, then once we see reordering we consider
4915 	 * the connection to alway be subject to reordering and just set lro
4916 	 * to 1.
4917 	 *
4918 	 * In the end if lro is non-zero we add the extra time for
4919 	 * reordering in.
4920 	 */
4921 	if (srtt == 0)
4922 		srtt = 1;
4923 	if (rack->r_ctl.rc_reorder_ts) {
4924 		if (rack->r_ctl.rc_reorder_fade) {
4925 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
4926 				lro = cts - rack->r_ctl.rc_reorder_ts;
4927 				if (lro == 0) {
4928 					/*
4929 					 * No time as passed since the last
4930 					 * reorder, mark it as reordering.
4931 					 */
4932 					lro = 1;
4933 				}
4934 			} else {
4935 				/* Negative time? */
4936 				lro = 0;
4937 			}
4938 			if (lro > rack->r_ctl.rc_reorder_fade) {
4939 				/* Turn off reordering seen too */
4940 				rack->r_ctl.rc_reorder_ts = 0;
4941 				lro = 0;
4942 			}
4943 		} else {
4944 			/* Reodering does not fade */
4945 			lro = 1;
4946 		}
4947 	} else {
4948 		lro = 0;
4949 	}
4950 	if (rack->rc_rack_tmr_std_based == 0) {
4951 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
4952 	} else {
4953 		/* Standards based pkt-delay is 1/4 srtt */
4954 		thresh = srtt +  (srtt >> 2);
4955 	}
4956 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
4957 		/* It must be set, if not you get 1/4 rtt */
4958 		if (rack->r_ctl.rc_reorder_shift)
4959 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
4960 		else
4961 			thresh += (srtt >> 2);
4962 	}
4963 	if (rack->rc_rack_use_dsack &&
4964 	    lro &&
4965 	    (rack->r_ctl.num_dsack > 0)) {
4966 		/*
4967 		 * We only increase the reordering window if we
4968 		 * have seen reordering <and> we have a DSACK count.
4969 		 */
4970 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
4971 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
4972 	}
4973 	/* SRTT * 2 is the ceiling */
4974 	if (thresh > (srtt * 2)) {
4975 		thresh = srtt * 2;
4976 	}
4977 	/* And we don't want it above the RTO max either */
4978 	if (thresh > rack_rto_max) {
4979 		thresh = rack_rto_max;
4980 	}
4981 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
4982 	return (thresh);
4983 }
4984 
4985 static uint32_t
4986 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
4987 		     struct rack_sendmap *rsm, uint32_t srtt)
4988 {
4989 	struct rack_sendmap *prsm;
4990 	uint32_t thresh, len;
4991 	int segsiz;
4992 
4993 	if (srtt == 0)
4994 		srtt = 1;
4995 	if (rack->r_ctl.rc_tlp_threshold)
4996 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
4997 	else
4998 		thresh = (srtt * 2);
4999 
5000 	/* Get the previous sent packet, if any */
5001 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5002 	len = rsm->r_end - rsm->r_start;
5003 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5004 		/* Exactly like the ID */
5005 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5006 			uint32_t alt_thresh;
5007 			/*
5008 			 * Compensate for delayed-ack with the d-ack time.
5009 			 */
5010 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5011 			if (alt_thresh > thresh)
5012 				thresh = alt_thresh;
5013 		}
5014 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5015 		/* 2.1 behavior */
5016 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5017 		if (prsm && (len <= segsiz)) {
5018 			/*
5019 			 * Two packets outstanding, thresh should be (2*srtt) +
5020 			 * possible inter-packet delay (if any).
5021 			 */
5022 			uint32_t inter_gap = 0;
5023 			int idx, nidx;
5024 
5025 			idx = rsm->r_rtr_cnt - 1;
5026 			nidx = prsm->r_rtr_cnt - 1;
5027 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5028 				/* Yes it was sent later (or at the same time) */
5029 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5030 			}
5031 			thresh += inter_gap;
5032 		} else if (len <= segsiz) {
5033 			/*
5034 			 * Possibly compensate for delayed-ack.
5035 			 */
5036 			uint32_t alt_thresh;
5037 
5038 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5039 			if (alt_thresh > thresh)
5040 				thresh = alt_thresh;
5041 		}
5042 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5043 		/* 2.2 behavior */
5044 		if (len <= segsiz) {
5045 			uint32_t alt_thresh;
5046 			/*
5047 			 * Compensate for delayed-ack with the d-ack time.
5048 			 */
5049 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5050 			if (alt_thresh > thresh)
5051 				thresh = alt_thresh;
5052 		}
5053 	}
5054 	/* Not above an RTO */
5055 	if (thresh > tp->t_rxtcur) {
5056 		thresh = tp->t_rxtcur;
5057 	}
5058 	/* Not above a RTO max */
5059 	if (thresh > rack_rto_max) {
5060 		thresh = rack_rto_max;
5061 	}
5062 	/* Apply user supplied min TLP */
5063 	if (thresh < rack_tlp_min) {
5064 		thresh = rack_tlp_min;
5065 	}
5066 	return (thresh);
5067 }
5068 
5069 static uint32_t
5070 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5071 {
5072 	/*
5073 	 * We want the rack_rtt which is the
5074 	 * last rtt we measured. However if that
5075 	 * does not exist we fallback to the srtt (which
5076 	 * we probably will never do) and then as a last
5077 	 * resort we use RACK_INITIAL_RTO if no srtt is
5078 	 * yet set.
5079 	 */
5080 	if (rack->rc_rack_rtt)
5081 		return (rack->rc_rack_rtt);
5082 	else if (tp->t_srtt == 0)
5083 		return (RACK_INITIAL_RTO);
5084 	return (tp->t_srtt);
5085 }
5086 
5087 static struct rack_sendmap *
5088 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5089 {
5090 	/*
5091 	 * Check to see that we don't need to fall into recovery. We will
5092 	 * need to do so if our oldest transmit is past the time we should
5093 	 * have had an ack.
5094 	 */
5095 	struct tcp_rack *rack;
5096 	struct rack_sendmap *rsm;
5097 	int32_t idx;
5098 	uint32_t srtt, thresh;
5099 
5100 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5101 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5102 		return (NULL);
5103 	}
5104 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5105 	if (rsm == NULL)
5106 		return (NULL);
5107 
5108 	if (rsm->r_flags & RACK_ACKED) {
5109 		rsm = rack_find_lowest_rsm(rack);
5110 		if (rsm == NULL)
5111 			return (NULL);
5112 	}
5113 	idx = rsm->r_rtr_cnt - 1;
5114 	srtt = rack_grab_rtt(tp, rack);
5115 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5116 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5117 		return (NULL);
5118 	}
5119 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5120 		return (NULL);
5121 	}
5122 	/* Ok if we reach here we are over-due and this guy can be sent */
5123 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
5124 	return (rsm);
5125 }
5126 
5127 static uint32_t
5128 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5129 {
5130 	int32_t t;
5131 	int32_t tt;
5132 	uint32_t ret_val;
5133 
5134 	t = (tp->t_srtt + (tp->t_rttvar << 2));
5135 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5136  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5137 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5138 	ret_val = (uint32_t)tt;
5139 	return (ret_val);
5140 }
5141 
5142 static uint32_t
5143 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5144 {
5145 	/*
5146 	 * Start the FR timer, we do this based on getting the first one in
5147 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
5148 	 * events we need to stop the running timer (if its running) before
5149 	 * starting the new one.
5150 	 */
5151 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5152 	uint32_t srtt_cur;
5153 	int32_t idx;
5154 	int32_t is_tlp_timer = 0;
5155 	struct rack_sendmap *rsm;
5156 
5157 	if (rack->t_timers_stopped) {
5158 		/* All timers have been stopped none are to run */
5159 		return (0);
5160 	}
5161 	if (rack->rc_in_persist) {
5162 		/* We can't start any timer in persists */
5163 		return (rack_get_persists_timer_val(tp, rack));
5164 	}
5165 	rack->rc_on_min_to = 0;
5166 	if ((tp->t_state < TCPS_ESTABLISHED) ||
5167 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5168 		goto activate_rxt;
5169 	}
5170 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5171 	if ((rsm == NULL) || sup_rack) {
5172 		/* Nothing on the send map or no rack */
5173 activate_rxt:
5174 		time_since_sent = 0;
5175 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5176 		if (rsm) {
5177 			/*
5178 			 * Should we discount the RTX timer any?
5179 			 *
5180 			 * We want to discount it the smallest amount.
5181 			 * If a timer (Rack/TLP or RXT) has gone off more
5182 			 * recently thats the discount we want to use (now - timer time).
5183 			 * If the retransmit of the oldest packet was more recent then
5184 			 * we want to use that (now - oldest-packet-last_transmit_time).
5185 			 *
5186 			 */
5187 			idx = rsm->r_rtr_cnt - 1;
5188 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5189 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5190 			else
5191 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5192 			if (TSTMP_GT(cts, tstmp_touse))
5193 			    time_since_sent = cts - tstmp_touse;
5194 		}
5195 		if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5196 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5197 			to = tp->t_rxtcur;
5198 			if (to > time_since_sent)
5199 				to -= time_since_sent;
5200 			else
5201 				to = rack->r_ctl.rc_min_to;
5202 			if (to == 0)
5203 				to = 1;
5204 			/* Special case for KEEPINIT */
5205 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5206 			    (TP_KEEPINIT(tp) != 0) &&
5207 			    rsm) {
5208 				/*
5209 				 * We have to put a ceiling on the rxt timer
5210 				 * of the keep-init timeout.
5211 				 */
5212 				uint32_t max_time, red;
5213 
5214 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5215 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5216 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5217 					if (red < max_time)
5218 						max_time -= red;
5219 					else
5220 						max_time = 1;
5221 				}
5222 				/* Reduce timeout to the keep value if needed */
5223 				if (max_time < to)
5224 					to = max_time;
5225 			}
5226 			return (to);
5227 		}
5228 		return (0);
5229 	}
5230 	if (rsm->r_flags & RACK_ACKED) {
5231 		rsm = rack_find_lowest_rsm(rack);
5232 		if (rsm == NULL) {
5233 			/* No lowest? */
5234 			goto activate_rxt;
5235 		}
5236 	}
5237 	if (rack->sack_attack_disable) {
5238 		/*
5239 		 * We don't want to do
5240 		 * any TLP's if you are an attacker.
5241 		 * Though if you are doing what
5242 		 * is expected you may still have
5243 		 * SACK-PASSED marks.
5244 		 */
5245 		goto activate_rxt;
5246 	}
5247 	/* Convert from ms to usecs */
5248 	if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5249 		if ((tp->t_flags & TF_SENTFIN) &&
5250 		    ((tp->snd_max - tp->snd_una) == 1) &&
5251 		    (rsm->r_flags & RACK_HAS_FIN)) {
5252 			/*
5253 			 * We don't start a rack timer if all we have is a
5254 			 * FIN outstanding.
5255 			 */
5256 			goto activate_rxt;
5257 		}
5258 		if ((rack->use_rack_rr == 0) &&
5259 		    (IN_FASTRECOVERY(tp->t_flags)) &&
5260 		    (rack->rack_no_prr == 0) &&
5261 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5262 			/*
5263 			 * We are not cheating, in recovery  and
5264 			 * not enough ack's to yet get our next
5265 			 * retransmission out.
5266 			 *
5267 			 * Note that classified attackers do not
5268 			 * get to use the rack-cheat.
5269 			 */
5270 			goto activate_tlp;
5271 		}
5272 		srtt = rack_grab_rtt(tp, rack);
5273 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
5274 		idx = rsm->r_rtr_cnt - 1;
5275 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5276 		if (SEQ_GEQ(exp, cts)) {
5277 			to = exp - cts;
5278 			if (to < rack->r_ctl.rc_min_to) {
5279 				to = rack->r_ctl.rc_min_to;
5280 				if (rack->r_rr_config == 3)
5281 					rack->rc_on_min_to = 1;
5282 			}
5283 		} else {
5284 			to = rack->r_ctl.rc_min_to;
5285 			if (rack->r_rr_config == 3)
5286 				rack->rc_on_min_to = 1;
5287 		}
5288 	} else {
5289 		/* Ok we need to do a TLP not RACK */
5290 activate_tlp:
5291 		if ((rack->rc_tlp_in_progress != 0) &&
5292 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5293 			/*
5294 			 * The previous send was a TLP and we have sent
5295 			 * N TLP's without sending new data.
5296 			 */
5297 			goto activate_rxt;
5298 		}
5299 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5300 		if (rsm == NULL) {
5301 			/* We found no rsm to TLP with. */
5302 			goto activate_rxt;
5303 		}
5304 		if (rsm->r_flags & RACK_HAS_FIN) {
5305 			/* If its a FIN we dont do TLP */
5306 			rsm = NULL;
5307 			goto activate_rxt;
5308 		}
5309 		idx = rsm->r_rtr_cnt - 1;
5310 		time_since_sent = 0;
5311 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5312 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5313 		else
5314 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5315 		if (TSTMP_GT(cts, tstmp_touse))
5316 		    time_since_sent = cts - tstmp_touse;
5317 		is_tlp_timer = 1;
5318 		if (tp->t_srtt) {
5319 			if ((rack->rc_srtt_measure_made == 0) &&
5320 			    (tp->t_srtt == 1)) {
5321 				/*
5322 				 * If another stack as run and set srtt to 1,
5323 				 * then the srtt was 0, so lets use the initial.
5324 				 */
5325 				srtt = RACK_INITIAL_RTO;
5326 			} else {
5327 				srtt_cur = tp->t_srtt;
5328 				srtt = srtt_cur;
5329 			}
5330 		} else
5331 			srtt = RACK_INITIAL_RTO;
5332 		/*
5333 		 * If the SRTT is not keeping up and the
5334 		 * rack RTT has spiked we want to use
5335 		 * the last RTT not the smoothed one.
5336 		 */
5337 		if (rack_tlp_use_greater &&
5338 		    tp->t_srtt &&
5339 		    (srtt < rack_grab_rtt(tp, rack))) {
5340 			srtt = rack_grab_rtt(tp, rack);
5341 		}
5342 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5343 		if (thresh > time_since_sent) {
5344 			to = thresh - time_since_sent;
5345 		} else {
5346 			to = rack->r_ctl.rc_min_to;
5347 			rack_log_alt_to_to_cancel(rack,
5348 						  thresh,		/* flex1 */
5349 						  time_since_sent,	/* flex2 */
5350 						  tstmp_touse,		/* flex3 */
5351 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5352 						  (uint32_t)rsm->r_tim_lastsent[idx],
5353 						  srtt,
5354 						  idx, 99);
5355 		}
5356 		if (to < rack_tlp_min) {
5357 			to = rack_tlp_min;
5358 		}
5359 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5360 			/*
5361 			 * If the TLP time works out to larger than the max
5362 			 * RTO lets not do TLP.. just RTO.
5363 			 */
5364 			goto activate_rxt;
5365 		}
5366 	}
5367 	if (is_tlp_timer == 0) {
5368 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5369 	} else {
5370 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5371 	}
5372 	if (to == 0)
5373 		to = 1;
5374 	return (to);
5375 }
5376 
5377 static void
5378 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5379 {
5380 	if (rack->rc_in_persist == 0) {
5381 		if (tp->t_flags & TF_GPUTINPROG) {
5382 			/*
5383 			 * Stop the goodput now, the calling of the
5384 			 * measurement function clears the flag.
5385 			 */
5386 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5387 						    RACK_QUALITY_PERSIST);
5388 		}
5389 #ifdef NETFLIX_SHARED_CWND
5390 		if (rack->r_ctl.rc_scw) {
5391 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5392 			rack->rack_scwnd_is_idle = 1;
5393 		}
5394 #endif
5395 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5396 		if (rack->r_ctl.rc_went_idle_time == 0)
5397 			rack->r_ctl.rc_went_idle_time = 1;
5398 		rack_timer_cancel(tp, rack, cts, __LINE__);
5399 		rack->r_ctl.persist_lost_ends = 0;
5400 		rack->probe_not_answered = 0;
5401 		rack->forced_ack = 0;
5402 		tp->t_rxtshift = 0;
5403 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5404 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5405 		rack->rc_in_persist = 1;
5406 	}
5407 }
5408 
5409 static void
5410 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5411 {
5412 	if (tcp_in_hpts(rack->rc_inp)) {
5413 		tcp_hpts_remove(rack->rc_inp);
5414 		rack->r_ctl.rc_hpts_flags = 0;
5415 	}
5416 #ifdef NETFLIX_SHARED_CWND
5417 	if (rack->r_ctl.rc_scw) {
5418 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5419 		rack->rack_scwnd_is_idle = 0;
5420 	}
5421 #endif
5422 	if (rack->rc_gp_dyn_mul &&
5423 	    (rack->use_fixed_rate == 0) &&
5424 	    (rack->rc_always_pace)) {
5425 		/*
5426 		 * Do we count this as if a probe-rtt just
5427 		 * finished?
5428 		 */
5429 		uint32_t time_idle, idle_min;
5430 
5431 		time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5432 		idle_min = rack_min_probertt_hold;
5433 		if (rack_probertt_gpsrtt_cnt_div) {
5434 			uint64_t extra;
5435 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5436 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
5437 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5438 			idle_min += (uint32_t)extra;
5439 		}
5440 		if (time_idle >= idle_min) {
5441 			/* Yes, we count it as a probe-rtt. */
5442 			uint32_t us_cts;
5443 
5444 			us_cts = tcp_get_usecs(NULL);
5445 			if (rack->in_probe_rtt == 0) {
5446 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5447 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5448 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5449 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5450 			} else {
5451 				rack_exit_probertt(rack, us_cts);
5452 			}
5453 		}
5454 	}
5455 	rack->rc_in_persist = 0;
5456 	rack->r_ctl.rc_went_idle_time = 0;
5457 	tp->t_rxtshift = 0;
5458 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5459 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5460 	rack->r_ctl.rc_agg_delayed = 0;
5461 	rack->r_early = 0;
5462 	rack->r_late = 0;
5463 	rack->r_ctl.rc_agg_early = 0;
5464 }
5465 
5466 static void
5467 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5468 		   struct hpts_diag *diag, struct timeval *tv)
5469 {
5470 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5471 		union tcp_log_stackspecific log;
5472 
5473 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5474 		log.u_bbr.flex1 = diag->p_nxt_slot;
5475 		log.u_bbr.flex2 = diag->p_cur_slot;
5476 		log.u_bbr.flex3 = diag->slot_req;
5477 		log.u_bbr.flex4 = diag->inp_hptsslot;
5478 		log.u_bbr.flex5 = diag->slot_remaining;
5479 		log.u_bbr.flex6 = diag->need_new_to;
5480 		log.u_bbr.flex7 = diag->p_hpts_active;
5481 		log.u_bbr.flex8 = diag->p_on_min_sleep;
5482 		/* Hijack other fields as needed */
5483 		log.u_bbr.epoch = diag->have_slept;
5484 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
5485 		log.u_bbr.pkts_out = diag->co_ret;
5486 		log.u_bbr.applimited = diag->hpts_sleep_time;
5487 		log.u_bbr.delivered = diag->p_prev_slot;
5488 		log.u_bbr.inflight = diag->p_runningslot;
5489 		log.u_bbr.bw_inuse = diag->wheel_slot;
5490 		log.u_bbr.rttProp = diag->wheel_cts;
5491 		log.u_bbr.timeStamp = cts;
5492 		log.u_bbr.delRate = diag->maxslots;
5493 		log.u_bbr.cur_del_rate = diag->p_curtick;
5494 		log.u_bbr.cur_del_rate <<= 32;
5495 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
5496 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5497 		    &rack->rc_inp->inp_socket->so_rcv,
5498 		    &rack->rc_inp->inp_socket->so_snd,
5499 		    BBR_LOG_HPTSDIAG, 0,
5500 		    0, &log, false, tv);
5501 	}
5502 
5503 }
5504 
5505 static void
5506 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5507 {
5508 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5509 		union tcp_log_stackspecific log;
5510 		struct timeval tv;
5511 
5512 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5513 		log.u_bbr.flex1 = sb->sb_flags;
5514 		log.u_bbr.flex2 = len;
5515 		log.u_bbr.flex3 = sb->sb_state;
5516 		log.u_bbr.flex8 = type;
5517 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5518 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5519 		    &rack->rc_inp->inp_socket->so_rcv,
5520 		    &rack->rc_inp->inp_socket->so_snd,
5521 		    TCP_LOG_SB_WAKE, 0,
5522 		    len, &log, false, &tv);
5523 	}
5524 }
5525 
5526 static void
5527 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5528       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5529 {
5530 	struct hpts_diag diag;
5531 	struct inpcb *inp;
5532 	struct timeval tv;
5533 	uint32_t delayed_ack = 0;
5534 	uint32_t hpts_timeout;
5535 	uint32_t entry_slot = slot;
5536 	uint8_t stopped;
5537 	uint32_t left = 0;
5538 	uint32_t us_cts;
5539 
5540 	inp = tp->t_inpcb;
5541 	if ((tp->t_state == TCPS_CLOSED) ||
5542 	    (tp->t_state == TCPS_LISTEN)) {
5543 		return;
5544 	}
5545 	if (tcp_in_hpts(inp)) {
5546 		/* Already on the pacer */
5547 		return;
5548 	}
5549 	stopped = rack->rc_tmr_stopped;
5550 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5551 		left = rack->r_ctl.rc_timer_exp - cts;
5552 	}
5553 	rack->r_ctl.rc_timer_exp = 0;
5554 	rack->r_ctl.rc_hpts_flags = 0;
5555 	us_cts = tcp_get_usecs(&tv);
5556 	/* Now early/late accounting */
5557 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5558 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5559 		/*
5560 		 * We have a early carry over set,
5561 		 * we can always add more time so we
5562 		 * can always make this compensation.
5563 		 *
5564 		 * Note if ack's are allowed to wake us do not
5565 		 * penalize the next timer for being awoke
5566 		 * by an ack aka the rc_agg_early (non-paced mode).
5567 		 */
5568 		slot += rack->r_ctl.rc_agg_early;
5569 		rack->r_early = 0;
5570 		rack->r_ctl.rc_agg_early = 0;
5571 	}
5572 	if (rack->r_late) {
5573 		/*
5574 		 * This is harder, we can
5575 		 * compensate some but it
5576 		 * really depends on what
5577 		 * the current pacing time is.
5578 		 */
5579 		if (rack->r_ctl.rc_agg_delayed >= slot) {
5580 			/*
5581 			 * We can't compensate for it all.
5582 			 * And we have to have some time
5583 			 * on the clock. We always have a min
5584 			 * 10 slots (10 x 10 i.e. 100 usecs).
5585 			 */
5586 			if (slot <= HPTS_TICKS_PER_SLOT) {
5587 				/* We gain delay */
5588 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5589 				slot = HPTS_TICKS_PER_SLOT;
5590 			} else {
5591 				/* We take off some */
5592 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5593 				slot = HPTS_TICKS_PER_SLOT;
5594 			}
5595 		} else {
5596 			slot -= rack->r_ctl.rc_agg_delayed;
5597 			rack->r_ctl.rc_agg_delayed = 0;
5598 			/* Make sure we have 100 useconds at minimum */
5599 			if (slot < HPTS_TICKS_PER_SLOT) {
5600 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5601 				slot = HPTS_TICKS_PER_SLOT;
5602 			}
5603 			if (rack->r_ctl.rc_agg_delayed == 0)
5604 				rack->r_late = 0;
5605 		}
5606 	}
5607 	if (slot) {
5608 		/* We are pacing too */
5609 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5610 	}
5611 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5612 #ifdef NETFLIX_EXP_DETECTION
5613 	if (rack->sack_attack_disable &&
5614 	    (slot < tcp_sad_pacing_interval)) {
5615 		/*
5616 		 * We have a potential attacker on
5617 		 * the line. We have possibly some
5618 		 * (or now) pacing time set. We want to
5619 		 * slow down the processing of sacks by some
5620 		 * amount (if it is an attacker). Set the default
5621 		 * slot for attackers in place (unless the orginal
5622 		 * interval is longer). Its stored in
5623 		 * micro-seconds, so lets convert to msecs.
5624 		 */
5625 		slot = tcp_sad_pacing_interval;
5626 	}
5627 #endif
5628 	if (tp->t_flags & TF_DELACK) {
5629 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
5630 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5631 	}
5632 	if (delayed_ack && ((hpts_timeout == 0) ||
5633 			    (delayed_ack < hpts_timeout)))
5634 		hpts_timeout = delayed_ack;
5635 	else
5636 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5637 	/*
5638 	 * If no timers are going to run and we will fall off the hptsi
5639 	 * wheel, we resort to a keep-alive timer if its configured.
5640 	 */
5641 	if ((hpts_timeout == 0) &&
5642 	    (slot == 0)) {
5643 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5644 		    (tp->t_state <= TCPS_CLOSING)) {
5645 			/*
5646 			 * Ok we have no timer (persists, rack, tlp, rxt  or
5647 			 * del-ack), we don't have segments being paced. So
5648 			 * all that is left is the keepalive timer.
5649 			 */
5650 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5651 				/* Get the established keep-alive time */
5652 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5653 			} else {
5654 				/*
5655 				 * Get the initial setup keep-alive time,
5656 				 * note that this is probably not going to
5657 				 * happen, since rack will be running a rxt timer
5658 				 * if a SYN of some sort is outstanding. It is
5659 				 * actually handled in rack_timeout_rxt().
5660 				 */
5661 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5662 			}
5663 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5664 			if (rack->in_probe_rtt) {
5665 				/*
5666 				 * We want to instead not wake up a long time from
5667 				 * now but to wake up about the time we would
5668 				 * exit probe-rtt and initiate a keep-alive ack.
5669 				 * This will get us out of probe-rtt and update
5670 				 * our min-rtt.
5671 				 */
5672 				hpts_timeout = rack_min_probertt_hold;
5673 			}
5674 		}
5675 	}
5676 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5677 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5678 		/*
5679 		 * RACK, TLP, persists and RXT timers all are restartable
5680 		 * based on actions input .. i.e we received a packet (ack
5681 		 * or sack) and that changes things (rw, or snd_una etc).
5682 		 * Thus we can restart them with a new value. For
5683 		 * keep-alive, delayed_ack we keep track of what was left
5684 		 * and restart the timer with a smaller value.
5685 		 */
5686 		if (left < hpts_timeout)
5687 			hpts_timeout = left;
5688 	}
5689 	if (hpts_timeout) {
5690 		/*
5691 		 * Hack alert for now we can't time-out over 2,147,483
5692 		 * seconds (a bit more than 596 hours), which is probably ok
5693 		 * :).
5694 		 */
5695 		if (hpts_timeout > 0x7ffffffe)
5696 			hpts_timeout = 0x7ffffffe;
5697 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5698 	}
5699 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5700 	if ((rack->gp_ready == 0) &&
5701 	    (rack->use_fixed_rate == 0) &&
5702 	    (hpts_timeout < slot) &&
5703 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5704 		/*
5705 		 * We have no good estimate yet for the
5706 		 * old clunky burst mitigation or the
5707 		 * real pacing. And the tlp or rxt is smaller
5708 		 * than the pacing calculation. Lets not
5709 		 * pace that long since we know the calculation
5710 		 * so far is not accurate.
5711 		 */
5712 		slot = hpts_timeout;
5713 	}
5714 	/**
5715 	 * Turn off all the flags for queuing by default. The
5716 	 * flags have important meanings to what happens when
5717 	 * LRO interacts with the transport. Most likely (by default now)
5718 	 * mbuf_queueing and ack compression are on. So the transport
5719 	 * has a couple of flags that control what happens (if those
5720 	 * are not on then these flags won't have any effect since it
5721 	 * won't go through the queuing LRO path).
5722 	 *
5723 	 * INP_MBUF_QUEUE_READY - This flags says that I am busy
5724 	 *                        pacing output, so don't disturb. But
5725 	 *                        it also means LRO can wake me if there
5726 	 *                        is a SACK arrival.
5727 	 *
5728 	 * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5729 	 *                       with the above flag (QUEUE_READY) and
5730 	 *                       when present it says don't even wake me
5731 	 *                       if a SACK arrives.
5732 	 *
5733 	 * The idea behind these flags is that if we are pacing we
5734 	 * set the MBUF_QUEUE_READY and only get woken up if
5735 	 * a SACK arrives (which could change things) or if
5736 	 * our pacing timer expires. If, however, we have a rack
5737 	 * timer running, then we don't even want a sack to wake
5738 	 * us since the rack timer has to expire before we can send.
5739 	 *
5740 	 * Other cases should usually have none of the flags set
5741 	 * so LRO can call into us.
5742 	 */
5743 	inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5744 	if (slot) {
5745 		rack->r_ctl.rc_last_output_to = us_cts + slot;
5746 		/*
5747 		 * A pacing timer (slot) is being set, in
5748 		 * such a case we cannot send (we are blocked by
5749 		 * the timer). So lets tell LRO that it should not
5750 		 * wake us unless there is a SACK. Note this only
5751 		 * will be effective if mbuf queueing is on or
5752 		 * compressed acks are being processed.
5753 		 */
5754 		inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5755 		/*
5756 		 * But wait if we have a Rack timer running
5757 		 * even a SACK should not disturb us (with
5758 		 * the exception of r_rr_config 3).
5759 		 */
5760 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5761 		    (rack->r_rr_config != 3))
5762 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5763 		if (rack->rc_ack_can_sendout_data) {
5764 			/*
5765 			 * Ahh but wait, this is that special case
5766 			 * where the pacing timer can be disturbed
5767 			 * backout the changes (used for non-paced
5768 			 * burst limiting).
5769 			 */
5770 			inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5771 		}
5772 		if ((rack->use_rack_rr) &&
5773 		    (rack->r_rr_config < 2) &&
5774 		    ((hpts_timeout) && (hpts_timeout < slot))) {
5775 			/*
5776 			 * Arrange for the hpts to kick back in after the
5777 			 * t-o if the t-o does not cause a send.
5778 			 */
5779 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5780 						   __LINE__, &diag);
5781 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5782 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5783 		} else {
5784 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
5785 						   __LINE__, &diag);
5786 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5787 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5788 		}
5789 	} else if (hpts_timeout) {
5790 		/*
5791 		 * With respect to inp_flags2 here, lets let any new acks wake
5792 		 * us up here. Since we are not pacing (no pacing timer), output
5793 		 * can happen so we should let it. If its a Rack timer, then any inbound
5794 		 * packet probably won't change the sending (we will be blocked)
5795 		 * but it may change the prr stats so letting it in (the set defaults
5796 		 * at the start of this block) are good enough.
5797 		 */
5798 		(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5799 					   __LINE__, &diag);
5800 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5801 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5802 	} else {
5803 		/* No timer starting */
5804 #ifdef INVARIANTS
5805 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5806 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5807 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5808 		}
5809 #endif
5810 	}
5811 	rack->rc_tmr_stopped = 0;
5812 	if (slot)
5813 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5814 }
5815 
5816 /*
5817  * RACK Timer, here we simply do logging and house keeping.
5818  * the normal rack_output() function will call the
5819  * appropriate thing to check if we need to do a RACK retransmit.
5820  * We return 1, saying don't proceed with rack_output only
5821  * when all timers have been stopped (destroyed PCB?).
5822  */
5823 static int
5824 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5825 {
5826 	/*
5827 	 * This timer simply provides an internal trigger to send out data.
5828 	 * The check_recovery_mode call will see if there are needed
5829 	 * retransmissions, if so we will enter fast-recovery. The output
5830 	 * call may or may not do the same thing depending on sysctl
5831 	 * settings.
5832 	 */
5833 	struct rack_sendmap *rsm;
5834 
5835 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5836 		return (1);
5837 	}
5838 	counter_u64_add(rack_to_tot, 1);
5839 	if (rack->r_state && (rack->r_state != tp->t_state))
5840 		rack_set_state(tp, rack);
5841 	rack->rc_on_min_to = 0;
5842 	rsm = rack_check_recovery_mode(tp, cts);
5843 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5844 	if (rsm) {
5845 		rack->r_ctl.rc_resend = rsm;
5846 		rack->r_timer_override = 1;
5847 		if (rack->use_rack_rr) {
5848 			/*
5849 			 * Don't accumulate extra pacing delay
5850 			 * we are allowing the rack timer to
5851 			 * over-ride pacing i.e. rrr takes precedence
5852 			 * if the pacing interval is longer than the rrr
5853 			 * time (in other words we get the min pacing
5854 			 * time versus rrr pacing time).
5855 			 */
5856 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5857 		}
5858 	}
5859 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5860 	if (rsm == NULL) {
5861 		/* restart a timer and return 1 */
5862 		rack_start_hpts_timer(rack, tp, cts,
5863 				      0, 0, 0);
5864 		return (1);
5865 	}
5866 	return (0);
5867 }
5868 
5869 static void
5870 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5871 {
5872 	if (rsm->m->m_len > rsm->orig_m_len) {
5873 		/*
5874 		 * Mbuf grew, caused by sbcompress, our offset does
5875 		 * not change.
5876 		 */
5877 		rsm->orig_m_len = rsm->m->m_len;
5878 	} else if (rsm->m->m_len < rsm->orig_m_len) {
5879 		/*
5880 		 * Mbuf shrank, trimmed off the top by an ack, our
5881 		 * offset changes.
5882 		 */
5883 		rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
5884 		rsm->orig_m_len = rsm->m->m_len;
5885 	}
5886 }
5887 
5888 static void
5889 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
5890 {
5891 	struct mbuf *m;
5892 	uint32_t soff;
5893 
5894 	if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
5895 		/* Fix up the orig_m_len and possibly the mbuf offset */
5896 		rack_adjust_orig_mlen(src_rsm);
5897 	}
5898 	m = src_rsm->m;
5899 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
5900 	while (soff >= m->m_len) {
5901 		/* Move out past this mbuf */
5902 		soff -= m->m_len;
5903 		m = m->m_next;
5904 		KASSERT((m != NULL),
5905 			("rsm:%p nrsm:%p hit at soff:%u null m",
5906 			 src_rsm, rsm, soff));
5907 	}
5908 	rsm->m = m;
5909 	rsm->soff = soff;
5910 	rsm->orig_m_len = m->m_len;
5911 }
5912 
5913 static __inline void
5914 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5915 	       struct rack_sendmap *rsm, uint32_t start)
5916 {
5917 	int idx;
5918 
5919 	nrsm->r_start = start;
5920 	nrsm->r_end = rsm->r_end;
5921 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
5922 	nrsm->r_flags = rsm->r_flags;
5923 	nrsm->r_dupack = rsm->r_dupack;
5924 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
5925 	nrsm->r_rtr_bytes = 0;
5926 	nrsm->r_fas = rsm->r_fas;
5927 	rsm->r_end = nrsm->r_start;
5928 	nrsm->r_just_ret = rsm->r_just_ret;
5929 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
5930 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
5931 	}
5932 	/* Now if we have SYN flag we keep it on the left edge */
5933 	if (nrsm->r_flags & RACK_HAS_SYN)
5934 		nrsm->r_flags &= ~RACK_HAS_SYN;
5935 	/* Now if we have a FIN flag we keep it on the right edge */
5936 	if (rsm->r_flags & RACK_HAS_FIN)
5937 		rsm->r_flags &= ~RACK_HAS_FIN;
5938 	/* Push bit must go to the right edge as well */
5939 	if (rsm->r_flags & RACK_HAD_PUSH)
5940 		rsm->r_flags &= ~RACK_HAD_PUSH;
5941 	/* Clone over the state of the hw_tls flag */
5942 	nrsm->r_hw_tls = rsm->r_hw_tls;
5943 	/*
5944 	 * Now we need to find nrsm's new location in the mbuf chain
5945 	 * we basically calculate a new offset, which is soff +
5946 	 * how much is left in original rsm. Then we walk out the mbuf
5947 	 * chain to find the righ postion, it may be the same mbuf
5948 	 * or maybe not.
5949 	 */
5950 	KASSERT(((rsm->m != NULL) ||
5951 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
5952 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
5953 	if (rsm->m)
5954 		rack_setup_offset_for_rsm(rsm, nrsm);
5955 }
5956 
5957 static struct rack_sendmap *
5958 rack_merge_rsm(struct tcp_rack *rack,
5959 	       struct rack_sendmap *l_rsm,
5960 	       struct rack_sendmap *r_rsm)
5961 {
5962 	/*
5963 	 * We are merging two ack'd RSM's,
5964 	 * the l_rsm is on the left (lower seq
5965 	 * values) and the r_rsm is on the right
5966 	 * (higher seq value). The simplest way
5967 	 * to merge these is to move the right
5968 	 * one into the left. I don't think there
5969 	 * is any reason we need to try to find
5970 	 * the oldest (or last oldest retransmitted).
5971 	 */
5972 #ifdef INVARIANTS
5973 	struct rack_sendmap *rm;
5974 #endif
5975 	rack_log_map_chg(rack->rc_tp, rack, NULL,
5976 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
5977 	l_rsm->r_end = r_rsm->r_end;
5978 	if (l_rsm->r_dupack < r_rsm->r_dupack)
5979 		l_rsm->r_dupack = r_rsm->r_dupack;
5980 	if (r_rsm->r_rtr_bytes)
5981 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
5982 	if (r_rsm->r_in_tmap) {
5983 		/* This really should not happen */
5984 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
5985 		r_rsm->r_in_tmap = 0;
5986 	}
5987 
5988 	/* Now the flags */
5989 	if (r_rsm->r_flags & RACK_HAS_FIN)
5990 		l_rsm->r_flags |= RACK_HAS_FIN;
5991 	if (r_rsm->r_flags & RACK_TLP)
5992 		l_rsm->r_flags |= RACK_TLP;
5993 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
5994 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
5995 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
5996 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
5997 		/*
5998 		 * If both are app-limited then let the
5999 		 * free lower the count. If right is app
6000 		 * limited and left is not, transfer.
6001 		 */
6002 		l_rsm->r_flags |= RACK_APP_LIMITED;
6003 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
6004 		if (r_rsm == rack->r_ctl.rc_first_appl)
6005 			rack->r_ctl.rc_first_appl = l_rsm;
6006 	}
6007 #ifndef INVARIANTS
6008 	(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6009 #else
6010 	rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6011 	if (rm != r_rsm) {
6012 		panic("removing head in rack:%p rsm:%p rm:%p",
6013 		      rack, r_rsm, rm);
6014 	}
6015 #endif
6016 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6017 		/* Transfer the split limit to the map we free */
6018 		r_rsm->r_limit_type = l_rsm->r_limit_type;
6019 		l_rsm->r_limit_type = 0;
6020 	}
6021 	rack_free(rack, r_rsm);
6022 	return (l_rsm);
6023 }
6024 
6025 /*
6026  * TLP Timer, here we simply setup what segment we want to
6027  * have the TLP expire on, the normal rack_output() will then
6028  * send it out.
6029  *
6030  * We return 1, saying don't proceed with rack_output only
6031  * when all timers have been stopped (destroyed PCB?).
6032  */
6033 static int
6034 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6035 {
6036 	/*
6037 	 * Tail Loss Probe.
6038 	 */
6039 	struct rack_sendmap *rsm = NULL;
6040 #ifdef INVARIANTS
6041 	struct rack_sendmap *insret;
6042 #endif
6043 	struct socket *so;
6044 	uint32_t amm;
6045 	uint32_t out, avail;
6046 	int collapsed_win = 0;
6047 
6048 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6049 		return (1);
6050 	}
6051 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6052 		/* Its not time yet */
6053 		return (0);
6054 	}
6055 	if (ctf_progress_timeout_check(tp, true)) {
6056 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6057 		return (-ETIMEDOUT);	/* tcp_drop() */
6058 	}
6059 	/*
6060 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
6061 	 * need to figure out how to force a full MSS segment out.
6062 	 */
6063 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6064 	rack->r_ctl.retran_during_recovery = 0;
6065 	rack->r_ctl.dsack_byte_cnt = 0;
6066 	counter_u64_add(rack_tlp_tot, 1);
6067 	if (rack->r_state && (rack->r_state != tp->t_state))
6068 		rack_set_state(tp, rack);
6069 	so = tp->t_inpcb->inp_socket;
6070 	avail = sbavail(&so->so_snd);
6071 	out = tp->snd_max - tp->snd_una;
6072 	if (out > tp->snd_wnd) {
6073 		/* special case, we need a retransmission */
6074 		collapsed_win = 1;
6075 		goto need_retran;
6076 	}
6077 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6078 		rack->r_ctl.dsack_persist--;
6079 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6080 			rack->r_ctl.num_dsack = 0;
6081 		}
6082 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6083 	}
6084 	if ((tp->t_flags & TF_GPUTINPROG) &&
6085 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6086 		/*
6087 		 * If this is the second in a row
6088 		 * TLP and we are doing a measurement
6089 		 * its time to abandon the measurement.
6090 		 * Something is likely broken on
6091 		 * the clients network and measuring a
6092 		 * broken network does us no good.
6093 		 */
6094 		tp->t_flags &= ~TF_GPUTINPROG;
6095 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6096 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6097 					   tp->gput_seq,
6098 					   0, 0, 18, __LINE__, NULL, 0);
6099 	}
6100 	/*
6101 	 * Check our send oldest always settings, and if
6102 	 * there is an oldest to send jump to the need_retran.
6103 	 */
6104 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6105 		goto need_retran;
6106 
6107 	if (avail > out) {
6108 		/* New data is available */
6109 		amm = avail - out;
6110 		if (amm > ctf_fixed_maxseg(tp)) {
6111 			amm = ctf_fixed_maxseg(tp);
6112 			if ((amm + out) > tp->snd_wnd) {
6113 				/* We are rwnd limited */
6114 				goto need_retran;
6115 			}
6116 		} else if (amm < ctf_fixed_maxseg(tp)) {
6117 			/* not enough to fill a MTU */
6118 			goto need_retran;
6119 		}
6120 		if (IN_FASTRECOVERY(tp->t_flags)) {
6121 			/* Unlikely */
6122 			if (rack->rack_no_prr == 0) {
6123 				if (out + amm <= tp->snd_wnd) {
6124 					rack->r_ctl.rc_prr_sndcnt = amm;
6125 					rack->r_ctl.rc_tlp_new_data = amm;
6126 					rack_log_to_prr(rack, 4, 0);
6127 				}
6128 			} else
6129 				goto need_retran;
6130 		} else {
6131 			/* Set the send-new override */
6132 			if (out + amm <= tp->snd_wnd)
6133 				rack->r_ctl.rc_tlp_new_data = amm;
6134 			else
6135 				goto need_retran;
6136 		}
6137 		rack->r_ctl.rc_tlpsend = NULL;
6138 		counter_u64_add(rack_tlp_newdata, 1);
6139 		goto send;
6140 	}
6141 need_retran:
6142 	/*
6143 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
6144 	 * optionally the first un-acked segment.
6145 	 */
6146 	if (collapsed_win == 0) {
6147 		if (rack_always_send_oldest)
6148 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6149 		else {
6150 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6151 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6152 				rsm = rack_find_high_nonack(rack, rsm);
6153 			}
6154 		}
6155 		if (rsm == NULL) {
6156 #ifdef TCP_BLACKBOX
6157 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6158 #endif
6159 			goto out;
6160 		}
6161 	} else {
6162 		/*
6163 		 * We must find the last segment
6164 		 * that was acceptable by the client.
6165 		 */
6166 		RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6167 			if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6168 				/* Found one */
6169 				break;
6170 			}
6171 		}
6172 		if (rsm == NULL) {
6173 			/* None? if so send the first */
6174 			rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6175 			if (rsm == NULL) {
6176 #ifdef TCP_BLACKBOX
6177 				tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6178 #endif
6179 				goto out;
6180 			}
6181 		}
6182 	}
6183 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6184 		/*
6185 		 * We need to split this the last segment in two.
6186 		 */
6187 		struct rack_sendmap *nrsm;
6188 
6189 		nrsm = rack_alloc_full_limit(rack);
6190 		if (nrsm == NULL) {
6191 			/*
6192 			 * No memory to split, we will just exit and punt
6193 			 * off to the RXT timer.
6194 			 */
6195 			goto out;
6196 		}
6197 		rack_clone_rsm(rack, nrsm, rsm,
6198 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
6199 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6200 #ifndef INVARIANTS
6201 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6202 #else
6203 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6204 		if (insret != NULL) {
6205 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6206 			      nrsm, insret, rack, rsm);
6207 		}
6208 #endif
6209 		if (rsm->r_in_tmap) {
6210 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6211 			nrsm->r_in_tmap = 1;
6212 		}
6213 		rsm = nrsm;
6214 	}
6215 	rack->r_ctl.rc_tlpsend = rsm;
6216 send:
6217 	/* Make sure output path knows we are doing a TLP */
6218 	*doing_tlp = 1;
6219 	rack->r_timer_override = 1;
6220 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6221 	return (0);
6222 out:
6223 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6224 	return (0);
6225 }
6226 
6227 /*
6228  * Delayed ack Timer, here we simply need to setup the
6229  * ACK_NOW flag and remove the DELACK flag. From there
6230  * the output routine will send the ack out.
6231  *
6232  * We only return 1, saying don't proceed, if all timers
6233  * are stopped (destroyed PCB?).
6234  */
6235 static int
6236 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6237 {
6238 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6239 		return (1);
6240 	}
6241 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6242 	tp->t_flags &= ~TF_DELACK;
6243 	tp->t_flags |= TF_ACKNOW;
6244 	KMOD_TCPSTAT_INC(tcps_delack);
6245 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6246 	return (0);
6247 }
6248 
6249 /*
6250  * Persists timer, here we simply send the
6251  * same thing as a keepalive will.
6252  * the one byte send.
6253  *
6254  * We only return 1, saying don't proceed, if all timers
6255  * are stopped (destroyed PCB?).
6256  */
6257 static int
6258 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6259 {
6260 	struct tcptemp *t_template;
6261 #ifdef INVARIANTS
6262 	struct inpcb *inp = tp->t_inpcb;
6263 #endif
6264 	int32_t retval = 1;
6265 
6266 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6267 		return (1);
6268 	}
6269 	if (rack->rc_in_persist == 0)
6270 		return (0);
6271 	if (ctf_progress_timeout_check(tp, false)) {
6272 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6273 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6274 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6275 		return (-ETIMEDOUT);	/* tcp_drop() */
6276 	}
6277 	KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6278 	/*
6279 	 * Persistence timer into zero window. Force a byte to be output, if
6280 	 * possible.
6281 	 */
6282 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
6283 	/*
6284 	 * Hack: if the peer is dead/unreachable, we do not time out if the
6285 	 * window is closed.  After a full backoff, drop the connection if
6286 	 * the idle time (no responses to probes) reaches the maximum
6287 	 * backoff that we would use if retransmitting.
6288 	 */
6289 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6290 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6291 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6292 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6293 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6294 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6295 		retval = -ETIMEDOUT;	/* tcp_drop() */
6296 		goto out;
6297 	}
6298 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6299 	    tp->snd_una == tp->snd_max)
6300 		rack_exit_persist(tp, rack, cts);
6301 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6302 	/*
6303 	 * If the user has closed the socket then drop a persisting
6304 	 * connection after a much reduced timeout.
6305 	 */
6306 	if (tp->t_state > TCPS_CLOSE_WAIT &&
6307 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6308 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6309 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6310 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6311 		retval = -ETIMEDOUT;	/* tcp_drop() */
6312 		goto out;
6313 	}
6314 	t_template = tcpip_maketemplate(rack->rc_inp);
6315 	if (t_template) {
6316 		/* only set it if we were answered */
6317 		if (rack->forced_ack == 0) {
6318 			rack->forced_ack = 1;
6319 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6320 		} else {
6321 			rack->probe_not_answered = 1;
6322 			counter_u64_add(rack_persists_loss, 1);
6323 			rack->r_ctl.persist_lost_ends++;
6324 		}
6325 		counter_u64_add(rack_persists_sends, 1);
6326 		tcp_respond(tp, t_template->tt_ipgen,
6327 			    &t_template->tt_t, (struct mbuf *)NULL,
6328 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6329 		/* This sends an ack */
6330 		if (tp->t_flags & TF_DELACK)
6331 			tp->t_flags &= ~TF_DELACK;
6332 		free(t_template, M_TEMP);
6333 	}
6334 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6335 		tp->t_rxtshift++;
6336 out:
6337 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6338 	rack_start_hpts_timer(rack, tp, cts,
6339 			      0, 0, 0);
6340 	return (retval);
6341 }
6342 
6343 /*
6344  * If a keepalive goes off, we had no other timers
6345  * happening. We always return 1 here since this
6346  * routine either drops the connection or sends
6347  * out a segment with respond.
6348  */
6349 static int
6350 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6351 {
6352 	struct tcptemp *t_template;
6353 	struct inpcb *inp;
6354 
6355 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6356 		return (1);
6357 	}
6358 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6359 	inp = tp->t_inpcb;
6360 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6361 	/*
6362 	 * Keep-alive timer went off; send something or drop connection if
6363 	 * idle for too long.
6364 	 */
6365 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
6366 	if (tp->t_state < TCPS_ESTABLISHED)
6367 		goto dropit;
6368 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6369 	    tp->t_state <= TCPS_CLOSING) {
6370 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6371 			goto dropit;
6372 		/*
6373 		 * Send a packet designed to force a response if the peer is
6374 		 * up and reachable: either an ACK if the connection is
6375 		 * still alive, or an RST if the peer has closed the
6376 		 * connection due to timeout or reboot. Using sequence
6377 		 * number tp->snd_una-1 causes the transmitted zero-length
6378 		 * segment to lie outside the receive window; by the
6379 		 * protocol spec, this requires the correspondent TCP to
6380 		 * respond.
6381 		 */
6382 		KMOD_TCPSTAT_INC(tcps_keepprobe);
6383 		t_template = tcpip_maketemplate(inp);
6384 		if (t_template) {
6385 			if (rack->forced_ack == 0) {
6386 				rack->forced_ack = 1;
6387 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6388 			} else {
6389 				rack->probe_not_answered = 1;
6390 			}
6391 			tcp_respond(tp, t_template->tt_ipgen,
6392 			    &t_template->tt_t, (struct mbuf *)NULL,
6393 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6394 			free(t_template, M_TEMP);
6395 		}
6396 	}
6397 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6398 	return (1);
6399 dropit:
6400 	KMOD_TCPSTAT_INC(tcps_keepdrops);
6401 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6402 	return (-ETIMEDOUT);	/* tcp_drop() */
6403 }
6404 
6405 /*
6406  * Retransmit helper function, clear up all the ack
6407  * flags and take care of important book keeping.
6408  */
6409 static void
6410 rack_remxt_tmr(struct tcpcb *tp)
6411 {
6412 	/*
6413 	 * The retransmit timer went off, all sack'd blocks must be
6414 	 * un-acked.
6415 	 */
6416 	struct rack_sendmap *rsm, *trsm = NULL;
6417 	struct tcp_rack *rack;
6418 
6419 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6420 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6421 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6422 	if (rack->r_state && (rack->r_state != tp->t_state))
6423 		rack_set_state(tp, rack);
6424 	/*
6425 	 * Ideally we would like to be able to
6426 	 * mark SACK-PASS on anything not acked here.
6427 	 *
6428 	 * However, if we do that we would burst out
6429 	 * all that data 1ms apart. This would be unwise,
6430 	 * so for now we will just let the normal rxt timer
6431 	 * and tlp timer take care of it.
6432 	 *
6433 	 * Also we really need to stick them back in sequence
6434 	 * order. This way we send in the proper order and any
6435 	 * sacks that come floating in will "re-ack" the data.
6436 	 * To do this we zap the tmap with an INIT and then
6437 	 * walk through and place every rsm in the RB tree
6438 	 * back in its seq ordered place.
6439 	 */
6440 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
6441 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6442 		rsm->r_dupack = 0;
6443 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6444 		/* We must re-add it back to the tlist */
6445 		if (trsm == NULL) {
6446 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6447 		} else {
6448 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6449 		}
6450 		rsm->r_in_tmap = 1;
6451 		trsm = rsm;
6452 		if (rsm->r_flags & RACK_ACKED)
6453 			rsm->r_flags |= RACK_WAS_ACKED;
6454 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6455 		rsm->r_flags |= RACK_MUST_RXT;
6456 	}
6457 	/* Clear the count (we just un-acked them) */
6458 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6459 	rack->r_ctl.rc_sacked = 0;
6460 	rack->r_ctl.rc_sacklast = NULL;
6461 	rack->r_ctl.rc_agg_delayed = 0;
6462 	rack->r_early = 0;
6463 	rack->r_ctl.rc_agg_early = 0;
6464 	rack->r_late = 0;
6465 	/* Clear the tlp rtx mark */
6466 	rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6467 	if (rack->r_ctl.rc_resend != NULL)
6468 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6469 	rack->r_ctl.rc_prr_sndcnt = 0;
6470 	rack_log_to_prr(rack, 6, 0);
6471 	rack->r_timer_override = 1;
6472 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6473 #ifdef NETFLIX_EXP_DETECTION
6474 	    || (rack->sack_attack_disable != 0)
6475 #endif
6476 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6477 		/*
6478 		 * For non-sack customers new data
6479 		 * needs to go out as retransmits until
6480 		 * we retransmit up to snd_max.
6481 		 */
6482 		rack->r_must_retran = 1;
6483 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6484 						rack->r_ctl.rc_sacked);
6485 	}
6486 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6487 }
6488 
6489 static void
6490 rack_convert_rtts(struct tcpcb *tp)
6491 {
6492 	if (tp->t_srtt > 1) {
6493 		uint32_t val, frac;
6494 
6495 		val = tp->t_srtt >> TCP_RTT_SHIFT;
6496 		frac = tp->t_srtt & 0x1f;
6497 		tp->t_srtt = TICKS_2_USEC(val);
6498 		/*
6499 		 * frac is the fractional part of the srtt (if any)
6500 		 * but its in ticks and every bit represents
6501 		 * 1/32nd of a hz.
6502 		 */
6503 		if (frac) {
6504 			if (hz == 1000) {
6505 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6506 			} else {
6507 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6508 			}
6509 			tp->t_srtt += frac;
6510 		}
6511 	}
6512 	if (tp->t_rttvar) {
6513 		uint32_t val, frac;
6514 
6515 		val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6516 		frac = tp->t_rttvar & 0x1f;
6517 		tp->t_rttvar = TICKS_2_USEC(val);
6518 		/*
6519 		 * frac is the fractional part of the srtt (if any)
6520 		 * but its in ticks and every bit represents
6521 		 * 1/32nd of a hz.
6522 		 */
6523 		if (frac) {
6524 			if (hz == 1000) {
6525 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6526 			} else {
6527 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6528 			}
6529 			tp->t_rttvar += frac;
6530 		}
6531 	}
6532 	tp->t_rxtcur = RACK_REXMTVAL(tp);
6533 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6534 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6535 	}
6536 	if (tp->t_rxtcur > rack_rto_max) {
6537 		tp->t_rxtcur = rack_rto_max;
6538 	}
6539 }
6540 
6541 static void
6542 rack_cc_conn_init(struct tcpcb *tp)
6543 {
6544 	struct tcp_rack *rack;
6545 	uint32_t srtt;
6546 
6547 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6548 	srtt = tp->t_srtt;
6549 	cc_conn_init(tp);
6550 	/*
6551 	 * Now convert to rack's internal format,
6552 	 * if required.
6553 	 */
6554 	if ((srtt == 0) && (tp->t_srtt != 0))
6555 		rack_convert_rtts(tp);
6556 	/*
6557 	 * We want a chance to stay in slowstart as
6558 	 * we create a connection. TCP spec says that
6559 	 * initially ssthresh is infinite. For our
6560 	 * purposes that is the snd_wnd.
6561 	 */
6562 	if (tp->snd_ssthresh < tp->snd_wnd) {
6563 		tp->snd_ssthresh = tp->snd_wnd;
6564 	}
6565 	/*
6566 	 * We also want to assure a IW worth of
6567 	 * data can get inflight.
6568 	 */
6569 	if (rc_init_window(rack) < tp->snd_cwnd)
6570 		tp->snd_cwnd = rc_init_window(rack);
6571 }
6572 
6573 /*
6574  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6575  * we will setup to retransmit the lowest seq number outstanding.
6576  */
6577 static int
6578 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6579 {
6580 	int32_t rexmt;
6581 	int32_t retval = 0;
6582 	bool isipv6;
6583 
6584 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6585 		return (1);
6586 	}
6587 	if ((tp->t_flags & TF_GPUTINPROG) &&
6588 	    (tp->t_rxtshift)) {
6589 		/*
6590 		 * We have had a second timeout
6591 		 * measurements on successive rxt's are not profitable.
6592 		 * It is unlikely to be of any use (the network is
6593 		 * broken or the client went away).
6594 		 */
6595 		tp->t_flags &= ~TF_GPUTINPROG;
6596 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6597 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6598 					   tp->gput_seq,
6599 					   0, 0, 18, __LINE__, NULL, 0);
6600 	}
6601 	if (ctf_progress_timeout_check(tp, false)) {
6602 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6603 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6604 		return (-ETIMEDOUT);	/* tcp_drop() */
6605 	}
6606 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6607 	rack->r_ctl.retran_during_recovery = 0;
6608 	rack->r_ctl.dsack_byte_cnt = 0;
6609 	if (IN_FASTRECOVERY(tp->t_flags))
6610 		tp->t_flags |= TF_WASFRECOVERY;
6611 	else
6612 		tp->t_flags &= ~TF_WASFRECOVERY;
6613 	if (IN_CONGRECOVERY(tp->t_flags))
6614 		tp->t_flags |= TF_WASCRECOVERY;
6615 	else
6616 		tp->t_flags &= ~TF_WASCRECOVERY;
6617 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6618 	    (tp->snd_una == tp->snd_max)) {
6619 		/* Nothing outstanding .. nothing to do */
6620 		return (0);
6621 	}
6622 	if (rack->r_ctl.dsack_persist) {
6623 		rack->r_ctl.dsack_persist--;
6624 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6625 			rack->r_ctl.num_dsack = 0;
6626 		}
6627 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6628 	}
6629 	/*
6630 	 * Rack can only run one timer  at a time, so we cannot
6631 	 * run a KEEPINIT (gating SYN sending) and a retransmit
6632 	 * timer for the SYN. So if we are in a front state and
6633 	 * have a KEEPINIT timer we need to check the first transmit
6634 	 * against now to see if we have exceeded the KEEPINIT time
6635 	 * (if one is set).
6636 	 */
6637 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6638 	    (TP_KEEPINIT(tp) != 0)) {
6639 		struct rack_sendmap *rsm;
6640 
6641 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6642 		if (rsm) {
6643 			/* Ok we have something outstanding to test keepinit with */
6644 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6645 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6646 				/* We have exceeded the KEEPINIT time */
6647 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6648 				goto drop_it;
6649 			}
6650 		}
6651 	}
6652 	/*
6653 	 * Retransmission timer went off.  Message has not been acked within
6654 	 * retransmit interval.  Back off to a longer retransmit interval
6655 	 * and retransmit one segment.
6656 	 */
6657 	rack_remxt_tmr(tp);
6658 	if ((rack->r_ctl.rc_resend == NULL) ||
6659 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6660 		/*
6661 		 * If the rwnd collapsed on
6662 		 * the one we are retransmitting
6663 		 * it does not count against the
6664 		 * rxt count.
6665 		 */
6666 		tp->t_rxtshift++;
6667 	}
6668 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6669 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6670 drop_it:
6671 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
6672 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6673 		/* XXXGL: previously t_softerror was casted to uint16_t */
6674 		MPASS(tp->t_softerror >= 0);
6675 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6676 		goto out;	/* tcp_drop() */
6677 	}
6678 	if (tp->t_state == TCPS_SYN_SENT) {
6679 		/*
6680 		 * If the SYN was retransmitted, indicate CWND to be limited
6681 		 * to 1 segment in cc_conn_init().
6682 		 */
6683 		tp->snd_cwnd = 1;
6684 	} else if (tp->t_rxtshift == 1) {
6685 		/*
6686 		 * first retransmit; record ssthresh and cwnd so they can be
6687 		 * recovered if this turns out to be a "bad" retransmit. A
6688 		 * retransmit is considered "bad" if an ACK for this segment
6689 		 * is received within RTT/2 interval; the assumption here is
6690 		 * that the ACK was already in flight.  See "On Estimating
6691 		 * End-to-End Network Path Properties" by Allman and Paxson
6692 		 * for more details.
6693 		 */
6694 		tp->snd_cwnd_prev = tp->snd_cwnd;
6695 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
6696 		tp->snd_recover_prev = tp->snd_recover;
6697 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6698 		tp->t_flags |= TF_PREVVALID;
6699 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6700 		tp->t_flags &= ~TF_PREVVALID;
6701 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6702 	if ((tp->t_state == TCPS_SYN_SENT) ||
6703 	    (tp->t_state == TCPS_SYN_RECEIVED))
6704 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6705 	else
6706 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6707 
6708 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6709 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6710 	/*
6711 	 * We enter the path for PLMTUD if connection is established or, if
6712 	 * connection is FIN_WAIT_1 status, reason for the last is that if
6713 	 * amount of data we send is very small, we could send it in couple
6714 	 * of packets and process straight to FIN. In that case we won't
6715 	 * catch ESTABLISHED state.
6716 	 */
6717 #ifdef INET6
6718 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6719 #else
6720 	isipv6 = false;
6721 #endif
6722 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
6723 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6724 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6725 	    ((tp->t_state == TCPS_ESTABLISHED) ||
6726 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
6727 		/*
6728 		 * Idea here is that at each stage of mtu probe (usually,
6729 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
6730 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6731 		 * should take care of that.
6732 		 */
6733 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6734 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6735 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6736 		    tp->t_rxtshift % 2 == 0)) {
6737 			/*
6738 			 * Enter Path MTU Black-hole Detection mechanism: -
6739 			 * Disable Path MTU Discovery (IP "DF" bit). -
6740 			 * Reduce MTU to lower value than what we negotiated
6741 			 * with peer.
6742 			 */
6743 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6744 				/* Record that we may have found a black hole. */
6745 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6746 				/* Keep track of previous MSS. */
6747 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6748 			}
6749 
6750 			/*
6751 			 * Reduce the MSS to blackhole value or to the
6752 			 * default in an attempt to retransmit.
6753 			 */
6754 #ifdef INET6
6755 			if (isipv6 &&
6756 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6757 				/* Use the sysctl tuneable blackhole MSS. */
6758 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6759 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6760 			} else if (isipv6) {
6761 				/* Use the default MSS. */
6762 				tp->t_maxseg = V_tcp_v6mssdflt;
6763 				/*
6764 				 * Disable Path MTU Discovery when we switch
6765 				 * to minmss.
6766 				 */
6767 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6768 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6769 			}
6770 #endif
6771 #if defined(INET6) && defined(INET)
6772 			else
6773 #endif
6774 #ifdef INET
6775 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6776 				/* Use the sysctl tuneable blackhole MSS. */
6777 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6778 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6779 			} else {
6780 				/* Use the default MSS. */
6781 				tp->t_maxseg = V_tcp_mssdflt;
6782 				/*
6783 				 * Disable Path MTU Discovery when we switch
6784 				 * to minmss.
6785 				 */
6786 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6787 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6788 			}
6789 #endif
6790 		} else {
6791 			/*
6792 			 * If further retransmissions are still unsuccessful
6793 			 * with a lowered MTU, maybe this isn't a blackhole
6794 			 * and we restore the previous MSS and blackhole
6795 			 * detection flags. The limit '6' is determined by
6796 			 * giving each probe stage (1448, 1188, 524) 2
6797 			 * chances to recover.
6798 			 */
6799 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6800 			    (tp->t_rxtshift >= 6)) {
6801 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6802 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6803 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6804 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6805 			}
6806 		}
6807 	}
6808 	/*
6809 	 * Disable RFC1323 and SACK if we haven't got any response to
6810 	 * our third SYN to work-around some broken terminal servers
6811 	 * (most of which have hopefully been retired) that have bad VJ
6812 	 * header compression code which trashes TCP segments containing
6813 	 * unknown-to-them TCP options.
6814 	 */
6815 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6816 	    (tp->t_rxtshift == 3))
6817 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6818 	/*
6819 	 * If we backed off this far, our srtt estimate is probably bogus.
6820 	 * Clobber it so we'll take the next rtt measurement as our srtt;
6821 	 * move the current srtt into rttvar to keep the current retransmit
6822 	 * times until then.
6823 	 */
6824 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6825 #ifdef INET6
6826 		if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6827 			in6_losing(tp->t_inpcb);
6828 		else
6829 #endif
6830 			in_losing(tp->t_inpcb);
6831 		tp->t_rttvar += tp->t_srtt;
6832 		tp->t_srtt = 0;
6833 	}
6834 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6835 	tp->snd_recover = tp->snd_max;
6836 	tp->t_flags |= TF_ACKNOW;
6837 	tp->t_rtttime = 0;
6838 	rack_cong_signal(tp, CC_RTO, tp->snd_una);
6839 out:
6840 	return (retval);
6841 }
6842 
6843 static int
6844 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6845 {
6846 	int32_t ret = 0;
6847 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6848 
6849 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6850 	    (tp->t_flags & TF_GPUTINPROG)) {
6851 		/*
6852 		 * We have a goodput in progress
6853 		 * and we have entered a late state.
6854 		 * Do we have enough data in the sb
6855 		 * to handle the GPUT request?
6856 		 */
6857 		uint32_t bytes;
6858 
6859 		bytes = tp->gput_ack - tp->gput_seq;
6860 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
6861 			bytes += tp->gput_seq - tp->snd_una;
6862 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
6863 			/*
6864 			 * There are not enough bytes in the socket
6865 			 * buffer that have been sent to cover this
6866 			 * measurement. Cancel it.
6867 			 */
6868 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6869 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
6870 						   tp->gput_seq,
6871 						   0, 0, 18, __LINE__, NULL, 0);
6872 			tp->t_flags &= ~TF_GPUTINPROG;
6873 		}
6874 	}
6875 	if (timers == 0) {
6876 		return (0);
6877 	}
6878 	if (tp->t_state == TCPS_LISTEN) {
6879 		/* no timers on listen sockets */
6880 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6881 			return (0);
6882 		return (1);
6883 	}
6884 	if ((timers & PACE_TMR_RACK) &&
6885 	    rack->rc_on_min_to) {
6886 		/*
6887 		 * For the rack timer when we
6888 		 * are on a min-timeout (which means rrr_conf = 3)
6889 		 * we don't want to check the timer. It may
6890 		 * be going off for a pace and thats ok we
6891 		 * want to send the retransmit (if its ready).
6892 		 *
6893 		 * If its on a normal rack timer (non-min) then
6894 		 * we will check if its expired.
6895 		 */
6896 		goto skip_time_check;
6897 	}
6898 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6899 		uint32_t left;
6900 
6901 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6902 			ret = -1;
6903 			rack_log_to_processing(rack, cts, ret, 0);
6904 			return (0);
6905 		}
6906 		if (hpts_calling == 0) {
6907 			/*
6908 			 * A user send or queued mbuf (sack) has called us? We
6909 			 * return 0 and let the pacing guards
6910 			 * deal with it if they should or
6911 			 * should not cause a send.
6912 			 */
6913 			ret = -2;
6914 			rack_log_to_processing(rack, cts, ret, 0);
6915 			return (0);
6916 		}
6917 		/*
6918 		 * Ok our timer went off early and we are not paced false
6919 		 * alarm, go back to sleep.
6920 		 */
6921 		ret = -3;
6922 		left = rack->r_ctl.rc_timer_exp - cts;
6923 		tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
6924 		rack_log_to_processing(rack, cts, ret, left);
6925 		return (1);
6926 	}
6927 skip_time_check:
6928 	rack->rc_tmr_stopped = 0;
6929 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6930 	if (timers & PACE_TMR_DELACK) {
6931 		ret = rack_timeout_delack(tp, rack, cts);
6932 	} else if (timers & PACE_TMR_RACK) {
6933 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6934 		rack->r_fast_output = 0;
6935 		ret = rack_timeout_rack(tp, rack, cts);
6936 	} else if (timers & PACE_TMR_TLP) {
6937 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6938 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
6939 	} else if (timers & PACE_TMR_RXT) {
6940 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6941 		rack->r_fast_output = 0;
6942 		ret = rack_timeout_rxt(tp, rack, cts);
6943 	} else if (timers & PACE_TMR_PERSIT) {
6944 		ret = rack_timeout_persist(tp, rack, cts);
6945 	} else if (timers & PACE_TMR_KEEP) {
6946 		ret = rack_timeout_keepalive(tp, rack, cts);
6947 	}
6948 	rack_log_to_processing(rack, cts, ret, timers);
6949 	return (ret);
6950 }
6951 
6952 static void
6953 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6954 {
6955 	struct timeval tv;
6956 	uint32_t us_cts, flags_on_entry;
6957 	uint8_t hpts_removed = 0;
6958 
6959 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
6960 	us_cts = tcp_get_usecs(&tv);
6961 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
6962 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
6963 	     ((tp->snd_max - tp->snd_una) == 0))) {
6964 		tcp_hpts_remove(rack->rc_inp);
6965 		hpts_removed = 1;
6966 		/* If we were not delayed cancel out the flag. */
6967 		if ((tp->snd_max - tp->snd_una) == 0)
6968 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6969 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6970 	}
6971 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
6972 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
6973 		if (tcp_in_hpts(rack->rc_inp) &&
6974 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
6975 			/*
6976 			 * Canceling timer's when we have no output being
6977 			 * paced. We also must remove ourselves from the
6978 			 * hpts.
6979 			 */
6980 			tcp_hpts_remove(rack->rc_inp);
6981 			hpts_removed = 1;
6982 		}
6983 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
6984 	}
6985 	if (hpts_removed == 0)
6986 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6987 }
6988 
6989 static void
6990 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
6991 {
6992 	return;
6993 }
6994 
6995 static int
6996 rack_stopall(struct tcpcb *tp)
6997 {
6998 	struct tcp_rack *rack;
6999 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7000 	rack->t_timers_stopped = 1;
7001 	return (0);
7002 }
7003 
7004 static void
7005 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7006 {
7007 	return;
7008 }
7009 
7010 static int
7011 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7012 {
7013 	return (0);
7014 }
7015 
7016 static void
7017 rack_stop_all_timers(struct tcpcb *tp)
7018 {
7019 	struct tcp_rack *rack;
7020 
7021 	/*
7022 	 * Assure no timers are running.
7023 	 */
7024 	if (tcp_timer_active(tp, TT_PERSIST)) {
7025 		/* We enter in persists, set the flag appropriately */
7026 		rack = (struct tcp_rack *)tp->t_fb_ptr;
7027 		rack->rc_in_persist = 1;
7028 	}
7029 	tcp_timer_suspend(tp, TT_PERSIST);
7030 	tcp_timer_suspend(tp, TT_REXMT);
7031 	tcp_timer_suspend(tp, TT_KEEP);
7032 	tcp_timer_suspend(tp, TT_DELACK);
7033 }
7034 
7035 static void
7036 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7037     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7038 {
7039 	int32_t idx;
7040 
7041 	rsm->r_rtr_cnt++;
7042 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7043 	rsm->r_dupack = 0;
7044 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7045 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7046 		rsm->r_flags |= RACK_OVERMAX;
7047 	}
7048 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7049 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7050 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7051 	}
7052 	idx = rsm->r_rtr_cnt - 1;
7053 	rsm->r_tim_lastsent[idx] = ts;
7054 	/*
7055 	 * Here we don't add in the len of send, since its already
7056 	 * in snduna <->snd_max.
7057 	 */
7058 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
7059 				     rack->r_ctl.rc_sacked);
7060 	if (rsm->r_flags & RACK_ACKED) {
7061 		/* Problably MTU discovery messing with us */
7062 		rsm->r_flags &= ~RACK_ACKED;
7063 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7064 	}
7065 	if (rsm->r_in_tmap) {
7066 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7067 		rsm->r_in_tmap = 0;
7068 	}
7069 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7070 	rsm->r_in_tmap = 1;
7071 	if (rsm->r_flags & RACK_SACK_PASSED) {
7072 		/* We have retransmitted due to the SACK pass */
7073 		rsm->r_flags &= ~RACK_SACK_PASSED;
7074 		rsm->r_flags |= RACK_WAS_SACKPASS;
7075 	}
7076 }
7077 
7078 static uint32_t
7079 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7080     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7081 {
7082 	/*
7083 	 * We (re-)transmitted starting at rsm->r_start for some length
7084 	 * (possibly less than r_end.
7085 	 */
7086 	struct rack_sendmap *nrsm;
7087 #ifdef INVARIANTS
7088 	struct rack_sendmap *insret;
7089 #endif
7090 	uint32_t c_end;
7091 	int32_t len;
7092 
7093 	len = *lenp;
7094 	c_end = rsm->r_start + len;
7095 	if (SEQ_GEQ(c_end, rsm->r_end)) {
7096 		/*
7097 		 * We retransmitted the whole piece or more than the whole
7098 		 * slopping into the next rsm.
7099 		 */
7100 		rack_update_rsm(tp, rack, rsm, ts, add_flag);
7101 		if (c_end == rsm->r_end) {
7102 			*lenp = 0;
7103 			return (0);
7104 		} else {
7105 			int32_t act_len;
7106 
7107 			/* Hangs over the end return whats left */
7108 			act_len = rsm->r_end - rsm->r_start;
7109 			*lenp = (len - act_len);
7110 			return (rsm->r_end);
7111 		}
7112 		/* We don't get out of this block. */
7113 	}
7114 	/*
7115 	 * Here we retransmitted less than the whole thing which means we
7116 	 * have to split this into what was transmitted and what was not.
7117 	 */
7118 	nrsm = rack_alloc_full_limit(rack);
7119 	if (nrsm == NULL) {
7120 		/*
7121 		 * We can't get memory, so lets not proceed.
7122 		 */
7123 		*lenp = 0;
7124 		return (0);
7125 	}
7126 	/*
7127 	 * So here we are going to take the original rsm and make it what we
7128 	 * retransmitted. nrsm will be the tail portion we did not
7129 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7130 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7131 	 * 1, 6 and the new piece will be 6, 11.
7132 	 */
7133 	rack_clone_rsm(rack, nrsm, rsm, c_end);
7134 	nrsm->r_dupack = 0;
7135 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7136 #ifndef INVARIANTS
7137 	(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7138 #else
7139 	insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7140 	if (insret != NULL) {
7141 		panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7142 		      nrsm, insret, rack, rsm);
7143 	}
7144 #endif
7145 	if (rsm->r_in_tmap) {
7146 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7147 		nrsm->r_in_tmap = 1;
7148 	}
7149 	rsm->r_flags &= (~RACK_HAS_FIN);
7150 	rack_update_rsm(tp, rack, rsm, ts, add_flag);
7151 	/* Log a split of rsm into rsm and nrsm */
7152 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7153 	*lenp = 0;
7154 	return (0);
7155 }
7156 
7157 static void
7158 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7159 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
7160 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7161 {
7162 	struct tcp_rack *rack;
7163 	struct rack_sendmap *rsm, *nrsm, fe;
7164 #ifdef INVARIANTS
7165 	struct rack_sendmap *insret;
7166 #endif
7167 	register uint32_t snd_max, snd_una;
7168 
7169 	/*
7170 	 * Add to the RACK log of packets in flight or retransmitted. If
7171 	 * there is a TS option we will use the TS echoed, if not we will
7172 	 * grab a TS.
7173 	 *
7174 	 * Retransmissions will increment the count and move the ts to its
7175 	 * proper place. Note that if options do not include TS's then we
7176 	 * won't be able to effectively use the ACK for an RTT on a retran.
7177 	 *
7178 	 * Notes about r_start and r_end. Lets consider a send starting at
7179 	 * sequence 1 for 10 bytes. In such an example the r_start would be
7180 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7181 	 * This means that r_end is actually the first sequence for the next
7182 	 * slot (11).
7183 	 *
7184 	 */
7185 	/*
7186 	 * If err is set what do we do XXXrrs? should we not add the thing?
7187 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
7188 	 * i.e. proceed with add ** do this for now.
7189 	 */
7190 	INP_WLOCK_ASSERT(tp->t_inpcb);
7191 	if (err)
7192 		/*
7193 		 * We don't log errors -- we could but snd_max does not
7194 		 * advance in this case either.
7195 		 */
7196 		return;
7197 
7198 	if (th_flags & TH_RST) {
7199 		/*
7200 		 * We don't log resets and we return immediately from
7201 		 * sending
7202 		 */
7203 		return;
7204 	}
7205 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7206 	snd_una = tp->snd_una;
7207 	snd_max = tp->snd_max;
7208 	if (th_flags & (TH_SYN | TH_FIN)) {
7209 		/*
7210 		 * The call to rack_log_output is made before bumping
7211 		 * snd_max. This means we can record one extra byte on a SYN
7212 		 * or FIN if seq_out is adding more on and a FIN is present
7213 		 * (and we are not resending).
7214 		 */
7215 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7216 			len++;
7217 		if (th_flags & TH_FIN)
7218 			len++;
7219 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
7220 			/*
7221 			 * The add/update as not been done for the FIN/SYN
7222 			 * yet.
7223 			 */
7224 			snd_max = tp->snd_nxt;
7225 		}
7226 	}
7227 	if (SEQ_LEQ((seq_out + len), snd_una)) {
7228 		/* Are sending an old segment to induce an ack (keep-alive)? */
7229 		return;
7230 	}
7231 	if (SEQ_LT(seq_out, snd_una)) {
7232 		/* huh? should we panic? */
7233 		uint32_t end;
7234 
7235 		end = seq_out + len;
7236 		seq_out = snd_una;
7237 		if (SEQ_GEQ(end, seq_out))
7238 			len = end - seq_out;
7239 		else
7240 			len = 0;
7241 	}
7242 	if (len == 0) {
7243 		/* We don't log zero window probes */
7244 		return;
7245 	}
7246 	if (IN_FASTRECOVERY(tp->t_flags)) {
7247 		rack->r_ctl.rc_prr_out += len;
7248 	}
7249 	/* First question is it a retransmission or new? */
7250 	if (seq_out == snd_max) {
7251 		/* Its new */
7252 again:
7253 		rsm = rack_alloc(rack);
7254 		if (rsm == NULL) {
7255 			/*
7256 			 * Hmm out of memory and the tcb got destroyed while
7257 			 * we tried to wait.
7258 			 */
7259 			return;
7260 		}
7261 		if (th_flags & TH_FIN) {
7262 			rsm->r_flags = RACK_HAS_FIN|add_flag;
7263 		} else {
7264 			rsm->r_flags = add_flag;
7265 		}
7266 		if (hw_tls)
7267 			rsm->r_hw_tls = 1;
7268 		rsm->r_tim_lastsent[0] = cts;
7269 		rsm->r_rtr_cnt = 1;
7270 		rsm->r_rtr_bytes = 0;
7271 		if (th_flags & TH_SYN) {
7272 			/* The data space is one beyond snd_una */
7273 			rsm->r_flags |= RACK_HAS_SYN;
7274 		}
7275 		rsm->r_start = seq_out;
7276 		rsm->r_end = rsm->r_start + len;
7277 		rsm->r_dupack = 0;
7278 		/*
7279 		 * save off the mbuf location that
7280 		 * sndmbuf_noadv returned (which is
7281 		 * where we started copying from)..
7282 		 */
7283 		rsm->m = s_mb;
7284 		rsm->soff = s_moff;
7285 		/*
7286 		 * Here we do add in the len of send, since its not yet
7287 		 * reflected in in snduna <->snd_max
7288 		 */
7289 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7290 					      rack->r_ctl.rc_sacked) +
7291 			      (rsm->r_end - rsm->r_start));
7292 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7293 		if (rsm->m) {
7294 			if (rsm->m->m_len <= rsm->soff) {
7295 				/*
7296 				 * XXXrrs Question, will this happen?
7297 				 *
7298 				 * If sbsndptr is set at the correct place
7299 				 * then s_moff should always be somewhere
7300 				 * within rsm->m. But if the sbsndptr was
7301 				 * off then that won't be true. If it occurs
7302 				 * we need to walkout to the correct location.
7303 				 */
7304 				struct mbuf *lm;
7305 
7306 				lm = rsm->m;
7307 				while (lm->m_len <= rsm->soff) {
7308 					rsm->soff -= lm->m_len;
7309 					lm = lm->m_next;
7310 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7311 							     __func__, rack, s_moff, s_mb, rsm->soff));
7312 				}
7313 				rsm->m = lm;
7314 			}
7315 			rsm->orig_m_len = rsm->m->m_len;
7316 		} else
7317 			rsm->orig_m_len = 0;
7318 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7319 		/* Log a new rsm */
7320 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7321 #ifndef INVARIANTS
7322 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7323 #else
7324 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7325 		if (insret != NULL) {
7326 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7327 			      nrsm, insret, rack, rsm);
7328 		}
7329 #endif
7330 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7331 		rsm->r_in_tmap = 1;
7332 		/*
7333 		 * Special case detection, is there just a single
7334 		 * packet outstanding when we are not in recovery?
7335 		 *
7336 		 * If this is true mark it so.
7337 		 */
7338 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7339 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7340 			struct rack_sendmap *prsm;
7341 
7342 			prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7343 			if (prsm)
7344 				prsm->r_one_out_nr = 1;
7345 		}
7346 		return;
7347 	}
7348 	/*
7349 	 * If we reach here its a retransmission and we need to find it.
7350 	 */
7351 	memset(&fe, 0, sizeof(fe));
7352 more:
7353 	if (hintrsm && (hintrsm->r_start == seq_out)) {
7354 		rsm = hintrsm;
7355 		hintrsm = NULL;
7356 	} else {
7357 		/* No hints sorry */
7358 		rsm = NULL;
7359 	}
7360 	if ((rsm) && (rsm->r_start == seq_out)) {
7361 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7362 		if (len == 0) {
7363 			return;
7364 		} else {
7365 			goto more;
7366 		}
7367 	}
7368 	/* Ok it was not the last pointer go through it the hard way. */
7369 refind:
7370 	fe.r_start = seq_out;
7371 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7372 	if (rsm) {
7373 		if (rsm->r_start == seq_out) {
7374 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7375 			if (len == 0) {
7376 				return;
7377 			} else {
7378 				goto refind;
7379 			}
7380 		}
7381 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7382 			/* Transmitted within this piece */
7383 			/*
7384 			 * Ok we must split off the front and then let the
7385 			 * update do the rest
7386 			 */
7387 			nrsm = rack_alloc_full_limit(rack);
7388 			if (nrsm == NULL) {
7389 				rack_update_rsm(tp, rack, rsm, cts, add_flag);
7390 				return;
7391 			}
7392 			/*
7393 			 * copy rsm to nrsm and then trim the front of rsm
7394 			 * to not include this part.
7395 			 */
7396 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
7397 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7398 #ifndef INVARIANTS
7399 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7400 #else
7401 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7402 			if (insret != NULL) {
7403 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7404 				      nrsm, insret, rack, rsm);
7405 			}
7406 #endif
7407 			if (rsm->r_in_tmap) {
7408 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7409 				nrsm->r_in_tmap = 1;
7410 			}
7411 			rsm->r_flags &= (~RACK_HAS_FIN);
7412 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7413 			if (len == 0) {
7414 				return;
7415 			} else if (len > 0)
7416 				goto refind;
7417 		}
7418 	}
7419 	/*
7420 	 * Hmm not found in map did they retransmit both old and on into the
7421 	 * new?
7422 	 */
7423 	if (seq_out == tp->snd_max) {
7424 		goto again;
7425 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
7426 #ifdef INVARIANTS
7427 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7428 		       seq_out, len, tp->snd_una, tp->snd_max);
7429 		printf("Starting Dump of all rack entries\n");
7430 		RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7431 			printf("rsm:%p start:%u end:%u\n",
7432 			       rsm, rsm->r_start, rsm->r_end);
7433 		}
7434 		printf("Dump complete\n");
7435 		panic("seq_out not found rack:%p tp:%p",
7436 		      rack, tp);
7437 #endif
7438 	} else {
7439 #ifdef INVARIANTS
7440 		/*
7441 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
7442 		 * flag)
7443 		 */
7444 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7445 		      seq_out, len, tp->snd_max, tp);
7446 #endif
7447 	}
7448 }
7449 
7450 /*
7451  * Record one of the RTT updates from an ack into
7452  * our sample structure.
7453  */
7454 
7455 static void
7456 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7457 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7458 {
7459 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7460 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7461 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7462 	}
7463 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7464 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7465 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7466 	}
7467 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7468 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7469 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
7470 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7471 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7472 	}
7473 	if ((confidence == 1) &&
7474 	    ((rsm == NULL) ||
7475 	     (rsm->r_just_ret) ||
7476 	     (rsm->r_one_out_nr &&
7477 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7478 		/*
7479 		 * If the rsm had a just return
7480 		 * hit it then we can't trust the
7481 		 * rtt measurement for buffer deterimination
7482 		 * Note that a confidence of 2, indicates
7483 		 * SACK'd which overrides the r_just_ret or
7484 		 * the r_one_out_nr. If it was a CUM-ACK and
7485 		 * we had only two outstanding, but get an
7486 		 * ack for only 1. Then that also lowers our
7487 		 * confidence.
7488 		 */
7489 		confidence = 0;
7490 	}
7491 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7492 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7493 		if (rack->r_ctl.rack_rs.confidence == 0) {
7494 			/*
7495 			 * We take anything with no current confidence
7496 			 * saved.
7497 			 */
7498 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7499 			rack->r_ctl.rack_rs.confidence = confidence;
7500 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7501 		} else if (confidence || rack->r_ctl.rack_rs.confidence) {
7502 			/*
7503 			 * Once we have a confident number,
7504 			 * we can update it with a smaller
7505 			 * value since this confident number
7506 			 * may include the DSACK time until
7507 			 * the next segment (the second one) arrived.
7508 			 */
7509 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7510 			rack->r_ctl.rack_rs.confidence = confidence;
7511 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7512 		}
7513 	}
7514 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7515 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7516 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7517 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
7518 }
7519 
7520 /*
7521  * Collect new round-trip time estimate
7522  * and update averages and current timeout.
7523  */
7524 static void
7525 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7526 {
7527 	int32_t delta;
7528 	int32_t rtt;
7529 
7530 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7531 		/* No valid sample */
7532 		return;
7533 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7534 		/* We are to use the lowest RTT seen in a single ack */
7535 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7536 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7537 		/* We are to use the highest RTT seen in a single ack */
7538 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7539 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7540 		/* We are to use the average RTT seen in a single ack */
7541 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7542 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7543 	} else {
7544 #ifdef INVARIANTS
7545 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7546 #endif
7547 		return;
7548 	}
7549 	if (rtt == 0)
7550 		rtt = 1;
7551 	if (rack->rc_gp_rtt_set == 0) {
7552 		/*
7553 		 * With no RTT we have to accept
7554 		 * even one we are not confident of.
7555 		 */
7556 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7557 		rack->rc_gp_rtt_set = 1;
7558 	} else if (rack->r_ctl.rack_rs.confidence) {
7559 		/* update the running gp srtt */
7560 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7561 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7562 	}
7563 	if (rack->r_ctl.rack_rs.confidence) {
7564 		/*
7565 		 * record the low and high for highly buffered path computation,
7566 		 * we only do this if we are confident (not a retransmission).
7567 		 */
7568 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7569 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7570 		}
7571 		if (rack->rc_highly_buffered == 0) {
7572 			/*
7573 			 * Currently once we declare a path has
7574 			 * highly buffered there is no going
7575 			 * back, which may be a problem...
7576 			 */
7577 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7578 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7579 						     rack->r_ctl.rc_highest_us_rtt,
7580 						     rack->r_ctl.rc_lowest_us_rtt,
7581 						     RACK_RTTS_SEEHBP);
7582 				rack->rc_highly_buffered = 1;
7583 			}
7584 		}
7585 	}
7586 	if ((rack->r_ctl.rack_rs.confidence) ||
7587 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7588 		/*
7589 		 * If we are highly confident of it <or> it was
7590 		 * never retransmitted we accept it as the last us_rtt.
7591 		 */
7592 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7593 		/* The lowest rtt can be set if its was not retransmited */
7594 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7595 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7596 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
7597 				rack->r_ctl.rc_lowest_us_rtt = 1;
7598 		}
7599 	}
7600 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7601 	if (tp->t_srtt != 0) {
7602 		/*
7603 		 * We keep a simple srtt in microseconds, like our rtt
7604 		 * measurement. We don't need to do any tricks with shifting
7605 		 * etc. Instead we just add in 1/8th of the new measurement
7606 		 * and subtract out 1/8 of the old srtt. We do the same with
7607 		 * the variance after finding the absolute value of the
7608 		 * difference between this sample and the current srtt.
7609 		 */
7610 		delta = tp->t_srtt - rtt;
7611 		/* Take off 1/8th of the current sRTT */
7612 		tp->t_srtt -= (tp->t_srtt >> 3);
7613 		/* Add in 1/8th of the new RTT just measured */
7614 		tp->t_srtt += (rtt >> 3);
7615 		if (tp->t_srtt <= 0)
7616 			tp->t_srtt = 1;
7617 		/* Now lets make the absolute value of the variance */
7618 		if (delta < 0)
7619 			delta = -delta;
7620 		/* Subtract out 1/8th */
7621 		tp->t_rttvar -= (tp->t_rttvar >> 3);
7622 		/* Add in 1/8th of the new variance we just saw */
7623 		tp->t_rttvar += (delta >> 3);
7624 		if (tp->t_rttvar <= 0)
7625 			tp->t_rttvar = 1;
7626 		if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7627 			tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7628 	} else {
7629 		/*
7630 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
7631 		 * variance to half the rtt (so our first retransmit happens
7632 		 * at 3*rtt).
7633 		 */
7634 		tp->t_srtt = rtt;
7635 		tp->t_rttvar = rtt >> 1;
7636 		tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7637 	}
7638 	rack->rc_srtt_measure_made = 1;
7639 	KMOD_TCPSTAT_INC(tcps_rttupdated);
7640 	tp->t_rttupdated++;
7641 #ifdef STATS
7642 	if (rack_stats_gets_ms_rtt == 0) {
7643 		/* Send in the microsecond rtt used for rxt timeout purposes */
7644 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7645 	} else if (rack_stats_gets_ms_rtt == 1) {
7646 		/* Send in the millisecond rtt used for rxt timeout purposes */
7647 		int32_t ms_rtt;
7648 
7649 		/* Round up */
7650 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7651 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7652 	} else if (rack_stats_gets_ms_rtt == 2) {
7653 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
7654 		int32_t ms_rtt;
7655 
7656 		/* Round up */
7657 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7658 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7659 	}  else {
7660 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
7661 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7662 	}
7663 
7664 #endif
7665 	/*
7666 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
7667 	 * way we do the smoothing, srtt and rttvar will each average +1/2
7668 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
7669 	 * tick of rounding and 1 extra tick because of +-1/2 tick
7670 	 * uncertainty in the firing of the timer.  The bias will give us
7671 	 * exactly the 1.5 tick we need.  But, because the bias is
7672 	 * statistical, we have to test that we don't drop below the minimum
7673 	 * feasible timer (which is 2 ticks).
7674 	 */
7675 	tp->t_rxtshift = 0;
7676 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7677 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7678 	rack_log_rtt_sample(rack, rtt);
7679 	tp->t_softerror = 0;
7680 }
7681 
7682 
7683 static void
7684 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7685 {
7686 	/*
7687 	 * Apply to filter the inbound us-rtt at us_cts.
7688 	 */
7689 	uint32_t old_rtt;
7690 
7691 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7692 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7693 			       us_rtt, us_cts);
7694 	if (old_rtt > us_rtt) {
7695 		/* We just hit a new lower rtt time */
7696 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7697 				     __LINE__, RACK_RTTS_NEWRTT);
7698 		/*
7699 		 * Only count it if its lower than what we saw within our
7700 		 * calculated range.
7701 		 */
7702 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7703 			if (rack_probertt_lower_within &&
7704 			    rack->rc_gp_dyn_mul &&
7705 			    (rack->use_fixed_rate == 0) &&
7706 			    (rack->rc_always_pace)) {
7707 				/*
7708 				 * We are seeing a new lower rtt very close
7709 				 * to the time that we would have entered probe-rtt.
7710 				 * This is probably due to the fact that a peer flow
7711 				 * has entered probe-rtt. Lets go in now too.
7712 				 */
7713 				uint32_t val;
7714 
7715 				val = rack_probertt_lower_within * rack_time_between_probertt;
7716 				val /= 100;
7717 				if ((rack->in_probe_rtt == 0)  &&
7718 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
7719 					rack_enter_probertt(rack, us_cts);
7720 				}
7721 			}
7722 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7723 		}
7724 	}
7725 }
7726 
7727 static int
7728 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7729     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7730 {
7731 	uint32_t us_rtt;
7732 	int32_t i, all;
7733 	uint32_t t, len_acked;
7734 
7735 	if ((rsm->r_flags & RACK_ACKED) ||
7736 	    (rsm->r_flags & RACK_WAS_ACKED))
7737 		/* Already done */
7738 		return (0);
7739 	if (rsm->r_no_rtt_allowed) {
7740 		/* Not allowed */
7741 		return (0);
7742 	}
7743 	if (ack_type == CUM_ACKED) {
7744 		if (SEQ_GT(th_ack, rsm->r_end)) {
7745 			len_acked = rsm->r_end - rsm->r_start;
7746 			all = 1;
7747 		} else {
7748 			len_acked = th_ack - rsm->r_start;
7749 			all = 0;
7750 		}
7751 	} else {
7752 		len_acked = rsm->r_end - rsm->r_start;
7753 		all = 0;
7754 	}
7755 	if (rsm->r_rtr_cnt == 1) {
7756 
7757 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7758 		if ((int)t <= 0)
7759 			t = 1;
7760 		if (!tp->t_rttlow || tp->t_rttlow > t)
7761 			tp->t_rttlow = t;
7762 		if (!rack->r_ctl.rc_rack_min_rtt ||
7763 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7764 			rack->r_ctl.rc_rack_min_rtt = t;
7765 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
7766 				rack->r_ctl.rc_rack_min_rtt = 1;
7767 			}
7768 		}
7769 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7770 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7771 		else
7772 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7773 		if (us_rtt == 0)
7774 			us_rtt = 1;
7775 		if (CC_ALGO(tp)->rttsample != NULL) {
7776 			/* Kick the RTT to the CC */
7777 			CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
7778 		}
7779 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7780 		if (ack_type == SACKED) {
7781 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7782 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7783 		} else {
7784 			/*
7785 			 * We need to setup what our confidence
7786 			 * is in this ack.
7787 			 *
7788 			 * If the rsm was app limited and it is
7789 			 * less than a mss in length (the end
7790 			 * of the send) then we have a gap. If we
7791 			 * were app limited but say we were sending
7792 			 * multiple MSS's then we are more confident
7793 			 * int it.
7794 			 *
7795 			 * When we are not app-limited then we see if
7796 			 * the rsm is being included in the current
7797 			 * measurement, we tell this by the app_limited_needs_set
7798 			 * flag.
7799 			 *
7800 			 * Note that being cwnd blocked is not applimited
7801 			 * as well as the pacing delay between packets which
7802 			 * are sending only 1 or 2 MSS's also will show up
7803 			 * in the RTT. We probably need to examine this algorithm
7804 			 * a bit more and enhance it to account for the delay
7805 			 * between rsm's. We could do that by saving off the
7806 			 * pacing delay of each rsm (in an rsm) and then
7807 			 * factoring that in somehow though for now I am
7808 			 * not sure how :)
7809 			 */
7810 			int calc_conf = 0;
7811 
7812 			if (rsm->r_flags & RACK_APP_LIMITED) {
7813 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7814 					calc_conf = 0;
7815 				else
7816 					calc_conf = 1;
7817 			} else if (rack->app_limited_needs_set == 0) {
7818 				calc_conf = 1;
7819 			} else {
7820 				calc_conf = 0;
7821 			}
7822 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7823 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7824 					    calc_conf, rsm, rsm->r_rtr_cnt);
7825 		}
7826 		if ((rsm->r_flags & RACK_TLP) &&
7827 		    (!IN_FASTRECOVERY(tp->t_flags))) {
7828 			/* Segment was a TLP and our retrans matched */
7829 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7830 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
7831 			}
7832 		}
7833 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7834 			/* New more recent rack_tmit_time */
7835 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7836 			rack->rc_rack_rtt = t;
7837 		}
7838 		return (1);
7839 	}
7840 	/*
7841 	 * We clear the soft/rxtshift since we got an ack.
7842 	 * There is no assurance we will call the commit() function
7843 	 * so we need to clear these to avoid incorrect handling.
7844 	 */
7845 	tp->t_rxtshift = 0;
7846 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7847 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7848 	tp->t_softerror = 0;
7849 	if (to && (to->to_flags & TOF_TS) &&
7850 	    (ack_type == CUM_ACKED) &&
7851 	    (to->to_tsecr) &&
7852 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7853 		/*
7854 		 * Now which timestamp does it match? In this block the ACK
7855 		 * must be coming from a previous transmission.
7856 		 */
7857 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
7858 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7859 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7860 				if ((int)t <= 0)
7861 					t = 1;
7862 				if (CC_ALGO(tp)->rttsample != NULL) {
7863 					/*
7864 					 * Kick the RTT to the CC, here
7865 					 * we lie a bit in that we know the
7866 					 * retransmission is correct even though
7867 					 * we retransmitted. This is because
7868 					 * we match the timestamps.
7869 					 */
7870 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
7871 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
7872 					else
7873 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
7874 					CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
7875 				}
7876 				if ((i + 1) < rsm->r_rtr_cnt) {
7877 					/*
7878 					 * The peer ack'd from our previous
7879 					 * transmission. We have a spurious
7880 					 * retransmission and thus we dont
7881 					 * want to update our rack_rtt.
7882 					 *
7883 					 * Hmm should there be a CC revert here?
7884 					 *
7885 					 */
7886 					return (0);
7887 				}
7888 				if (!tp->t_rttlow || tp->t_rttlow > t)
7889 					tp->t_rttlow = t;
7890 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7891 					rack->r_ctl.rc_rack_min_rtt = t;
7892 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
7893 						rack->r_ctl.rc_rack_min_rtt = 1;
7894 					}
7895 				}
7896 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7897 					   (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7898 					/* New more recent rack_tmit_time */
7899 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7900 					rack->rc_rack_rtt = t;
7901 				}
7902 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7903 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7904 						    rsm->r_rtr_cnt);
7905 				return (1);
7906 			}
7907 		}
7908 		goto ts_not_found;
7909 	} else {
7910 		/*
7911 		 * Ok its a SACK block that we retransmitted. or a windows
7912 		 * machine without timestamps. We can tell nothing from the
7913 		 * time-stamp since its not there or the time the peer last
7914 		 * recieved a segment that moved forward its cum-ack point.
7915 		 */
7916 ts_not_found:
7917 		i = rsm->r_rtr_cnt - 1;
7918 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7919 		if ((int)t <= 0)
7920 			t = 1;
7921 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7922 			/*
7923 			 * We retransmitted and the ack came back in less
7924 			 * than the smallest rtt we have observed. We most
7925 			 * likely did an improper retransmit as outlined in
7926 			 * 6.2 Step 2 point 2 in the rack-draft so we
7927 			 * don't want to update our rack_rtt. We in
7928 			 * theory (in future) might want to think about reverting our
7929 			 * cwnd state but we won't for now.
7930 			 */
7931 			return (0);
7932 		} else if (rack->r_ctl.rc_rack_min_rtt) {
7933 			/*
7934 			 * We retransmitted it and the retransmit did the
7935 			 * job.
7936 			 */
7937 			if (!rack->r_ctl.rc_rack_min_rtt ||
7938 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7939 				rack->r_ctl.rc_rack_min_rtt = t;
7940 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
7941 					rack->r_ctl.rc_rack_min_rtt = 1;
7942 				}
7943 			}
7944 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7945 				/* New more recent rack_tmit_time */
7946 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7947 				rack->rc_rack_rtt = t;
7948 			}
7949 			return (1);
7950 		}
7951 	}
7952 	return (0);
7953 }
7954 
7955 /*
7956  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7957  */
7958 static void
7959 rack_log_sack_passed(struct tcpcb *tp,
7960     struct tcp_rack *rack, struct rack_sendmap *rsm)
7961 {
7962 	struct rack_sendmap *nrsm;
7963 
7964 	nrsm = rsm;
7965 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7966 	    rack_head, r_tnext) {
7967 		if (nrsm == rsm) {
7968 			/* Skip orginal segment he is acked */
7969 			continue;
7970 		}
7971 		if (nrsm->r_flags & RACK_ACKED) {
7972 			/*
7973 			 * Skip ack'd segments, though we
7974 			 * should not see these, since tmap
7975 			 * should not have ack'd segments.
7976 			 */
7977 			continue;
7978 		}
7979 		if (nrsm->r_flags & RACK_SACK_PASSED) {
7980 			/*
7981 			 * We found one that is already marked
7982 			 * passed, we have been here before and
7983 			 * so all others below this are marked.
7984 			 */
7985 			break;
7986 		}
7987 		nrsm->r_flags |= RACK_SACK_PASSED;
7988 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
7989 	}
7990 }
7991 
7992 static void
7993 rack_need_set_test(struct tcpcb *tp,
7994 		   struct tcp_rack *rack,
7995 		   struct rack_sendmap *rsm,
7996 		   tcp_seq th_ack,
7997 		   int line,
7998 		   int use_which)
7999 {
8000 
8001 	if ((tp->t_flags & TF_GPUTINPROG) &&
8002 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8003 		/*
8004 		 * We were app limited, and this ack
8005 		 * butts up or goes beyond the point where we want
8006 		 * to start our next measurement. We need
8007 		 * to record the new gput_ts as here and
8008 		 * possibly update the start sequence.
8009 		 */
8010 		uint32_t seq, ts;
8011 
8012 		if (rsm->r_rtr_cnt > 1) {
8013 			/*
8014 			 * This is a retransmit, can we
8015 			 * really make any assessment at this
8016 			 * point?  We are not really sure of
8017 			 * the timestamp, is it this or the
8018 			 * previous transmission?
8019 			 *
8020 			 * Lets wait for something better that
8021 			 * is not retransmitted.
8022 			 */
8023 			return;
8024 		}
8025 		seq = tp->gput_seq;
8026 		ts = tp->gput_ts;
8027 		rack->app_limited_needs_set = 0;
8028 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8029 		/* Do we start at a new end? */
8030 		if ((use_which == RACK_USE_BEG) &&
8031 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8032 			/*
8033 			 * When we get an ACK that just eats
8034 			 * up some of the rsm, we set RACK_USE_BEG
8035 			 * since whats at r_start (i.e. th_ack)
8036 			 * is left unacked and thats where the
8037 			 * measurement not starts.
8038 			 */
8039 			tp->gput_seq = rsm->r_start;
8040 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8041 		}
8042 		if ((use_which == RACK_USE_END) &&
8043 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8044 			    /*
8045 			     * We use the end when the cumack
8046 			     * is moving forward and completely
8047 			     * deleting the rsm passed so basically
8048 			     * r_end holds th_ack.
8049 			     *
8050 			     * For SACK's we also want to use the end
8051 			     * since this piece just got sacked and
8052 			     * we want to target anything after that
8053 			     * in our measurement.
8054 			     */
8055 			    tp->gput_seq = rsm->r_end;
8056 			    rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8057 		}
8058 		if (use_which == RACK_USE_END_OR_THACK) {
8059 			/*
8060 			 * special case for ack moving forward,
8061 			 * not a sack, we need to move all the
8062 			 * way up to where this ack cum-ack moves
8063 			 * to.
8064 			 */
8065 			if (SEQ_GT(th_ack, rsm->r_end))
8066 				tp->gput_seq = th_ack;
8067 			else
8068 				tp->gput_seq = rsm->r_end;
8069 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8070 		}
8071 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8072 			/*
8073 			 * We moved beyond this guy's range, re-calculate
8074 			 * the new end point.
8075 			 */
8076 			if (rack->rc_gp_filled == 0) {
8077 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8078 			} else {
8079 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8080 			}
8081 		}
8082 		/*
8083 		 * We are moving the goal post, we may be able to clear the
8084 		 * measure_saw_probe_rtt flag.
8085 		 */
8086 		if ((rack->in_probe_rtt == 0) &&
8087 		    (rack->measure_saw_probe_rtt) &&
8088 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8089 			rack->measure_saw_probe_rtt = 0;
8090 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8091 					   seq, tp->gput_seq, 0, 5, line, NULL, 0);
8092 		if (rack->rc_gp_filled &&
8093 		    ((tp->gput_ack - tp->gput_seq) <
8094 		     max(rc_init_window(rack), (MIN_GP_WIN *
8095 						ctf_fixed_maxseg(tp))))) {
8096 			uint32_t ideal_amount;
8097 
8098 			ideal_amount = rack_get_measure_window(tp, rack);
8099 			if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8100 				/*
8101 				 * There is no sense of continuing this measurement
8102 				 * because its too small to gain us anything we
8103 				 * trust. Skip it and that way we can start a new
8104 				 * measurement quicker.
8105 				 */
8106 				tp->t_flags &= ~TF_GPUTINPROG;
8107 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8108 							   0, 0, 0, 6, __LINE__, NULL, 0);
8109 			} else {
8110 				/*
8111 				 * Reset the window further out.
8112 				 */
8113 				tp->gput_ack = tp->gput_seq + ideal_amount;
8114 			}
8115 		}
8116 	}
8117 }
8118 
8119 static inline int
8120 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8121 {
8122 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8123 		/* Behind our TLP definition or right at */
8124 		return (0);
8125 	}
8126 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8127 		/* The start is beyond or right at our end of TLP definition */
8128 		return (0);
8129 	}
8130 	/* It has to be a sub-part of the original TLP recorded */
8131 	return (1);
8132 }
8133 
8134 
8135 static uint32_t
8136 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8137 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8138 {
8139 	uint32_t start, end, changed = 0;
8140 	struct rack_sendmap stack_map;
8141 	struct rack_sendmap *rsm, *nrsm, fe, *prev, *next;
8142 #ifdef INVARIANTS
8143 	struct rack_sendmap *insret;
8144 #endif
8145 	int32_t used_ref = 1;
8146 	int moved = 0;
8147 
8148 	start = sack->start;
8149 	end = sack->end;
8150 	rsm = *prsm;
8151 	memset(&fe, 0, sizeof(fe));
8152 do_rest_ofb:
8153 	if ((rsm == NULL) ||
8154 	    (SEQ_LT(end, rsm->r_start)) ||
8155 	    (SEQ_GEQ(start, rsm->r_end)) ||
8156 	    (SEQ_LT(start, rsm->r_start))) {
8157 		/*
8158 		 * We are not in the right spot,
8159 		 * find the correct spot in the tree.
8160 		 */
8161 		used_ref = 0;
8162 		fe.r_start = start;
8163 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8164 		moved++;
8165 	}
8166 	if (rsm == NULL) {
8167 		/* TSNH */
8168 		goto out;
8169 	}
8170 	/* Ok we have an ACK for some piece of this rsm */
8171 	if (rsm->r_start != start) {
8172 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8173 			/*
8174 			 * Before any splitting or hookery is
8175 			 * done is it a TLP of interest i.e. rxt?
8176 			 */
8177 			if ((rsm->r_flags & RACK_TLP) &&
8178 			    (rsm->r_rtr_cnt > 1)) {
8179 				/*
8180 				 * We are splitting a rxt TLP, check
8181 				 * if we need to save off the start/end
8182 				 */
8183 				if (rack->rc_last_tlp_acked_set &&
8184 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8185 					/*
8186 					 * We already turned this on since we are inside
8187 					 * the previous one was a partially sack now we
8188 					 * are getting another one (maybe all of it).
8189 					 *
8190 					 */
8191 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8192 					/*
8193 					 * Lets make sure we have all of it though.
8194 					 */
8195 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8196 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8197 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8198 								     rack->r_ctl.last_tlp_acked_end);
8199 					}
8200 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8201 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8202 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8203 								     rack->r_ctl.last_tlp_acked_end);
8204 					}
8205 				} else {
8206 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8207 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8208 					rack->rc_last_tlp_past_cumack = 0;
8209 					rack->rc_last_tlp_acked_set = 1;
8210 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8211 				}
8212 			}
8213 			/**
8214 			 * Need to split this in two pieces the before and after,
8215 			 * the before remains in the map, the after must be
8216 			 * added. In other words we have:
8217 			 * rsm        |--------------|
8218 			 * sackblk        |------->
8219 			 * rsm will become
8220 			 *     rsm    |---|
8221 			 * and nrsm will be  the sacked piece
8222 			 *     nrsm       |----------|
8223 			 *
8224 			 * But before we start down that path lets
8225 			 * see if the sack spans over on top of
8226 			 * the next guy and it is already sacked.
8227 			 *
8228 			 */
8229 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8230 			if (next && (next->r_flags & RACK_ACKED) &&
8231 			    SEQ_GEQ(end, next->r_start)) {
8232 				/**
8233 				 * So the next one is already acked, and
8234 				 * we can thus by hookery use our stack_map
8235 				 * to reflect the piece being sacked and
8236 				 * then adjust the two tree entries moving
8237 				 * the start and ends around. So we start like:
8238 				 *  rsm     |------------|             (not-acked)
8239 				 *  next                 |-----------| (acked)
8240 				 *  sackblk        |-------->
8241 				 *  We want to end like so:
8242 				 *  rsm     |------|                   (not-acked)
8243 				 *  next           |-----------------| (acked)
8244 				 *  nrsm           |-----|
8245 				 * Where nrsm is a temporary stack piece we
8246 				 * use to update all the gizmos.
8247 				 */
8248 				/* Copy up our fudge block */
8249 				nrsm = &stack_map;
8250 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8251 				/* Now adjust our tree blocks */
8252 				rsm->r_end = start;
8253 				next->r_start = start;
8254 				/* Now we must adjust back where next->m is */
8255 				rack_setup_offset_for_rsm(rsm, next);
8256 
8257 				/* We don't need to adjust rsm, it did not change */
8258 				/* Clear out the dup ack count of the remainder */
8259 				rsm->r_dupack = 0;
8260 				rsm->r_just_ret = 0;
8261 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8262 				/* Now lets make sure our fudge block is right */
8263 				nrsm->r_start = start;
8264 				/* Now lets update all the stats and such */
8265 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8266 				if (rack->app_limited_needs_set)
8267 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8268 				changed += (nrsm->r_end - nrsm->r_start);
8269 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8270 				if (nrsm->r_flags & RACK_SACK_PASSED) {
8271 					rack->r_ctl.rc_reorder_ts = cts;
8272 				}
8273 				/*
8274 				 * Now we want to go up from rsm (the
8275 				 * one left un-acked) to the next one
8276 				 * in the tmap. We do this so when
8277 				 * we walk backwards we include marking
8278 				 * sack-passed on rsm (The one passed in
8279 				 * is skipped since it is generally called
8280 				 * on something sacked before removing it
8281 				 * from the tmap).
8282 				 */
8283 				if (rsm->r_in_tmap) {
8284 					nrsm = TAILQ_NEXT(rsm, r_tnext);
8285 					/*
8286 					 * Now that we have the next
8287 					 * one walk backwards from there.
8288 					 */
8289 					if (nrsm && nrsm->r_in_tmap)
8290 						rack_log_sack_passed(tp, rack, nrsm);
8291 				}
8292 				/* Now are we done? */
8293 				if (SEQ_LT(end, next->r_end) ||
8294 				    (end == next->r_end)) {
8295 					/* Done with block */
8296 					goto out;
8297 				}
8298 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8299 				counter_u64_add(rack_sack_used_next_merge, 1);
8300 				/* Postion for the next block */
8301 				start = next->r_end;
8302 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8303 				if (rsm == NULL)
8304 					goto out;
8305 			} else {
8306 				/**
8307 				 * We can't use any hookery here, so we
8308 				 * need to split the map. We enter like
8309 				 * so:
8310 				 *  rsm      |--------|
8311 				 *  sackblk       |----->
8312 				 * We will add the new block nrsm and
8313 				 * that will be the new portion, and then
8314 				 * fall through after reseting rsm. So we
8315 				 * split and look like this:
8316 				 *  rsm      |----|
8317 				 *  sackblk       |----->
8318 				 *  nrsm          |---|
8319 				 * We then fall through reseting
8320 				 * rsm to nrsm, so the next block
8321 				 * picks it up.
8322 				 */
8323 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8324 				if (nrsm == NULL) {
8325 					/*
8326 					 * failed XXXrrs what can we do but loose the sack
8327 					 * info?
8328 					 */
8329 					goto out;
8330 				}
8331 				counter_u64_add(rack_sack_splits, 1);
8332 				rack_clone_rsm(rack, nrsm, rsm, start);
8333 				rsm->r_just_ret = 0;
8334 #ifndef INVARIANTS
8335 				(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8336 #else
8337 				insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8338 				if (insret != NULL) {
8339 					panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8340 					      nrsm, insret, rack, rsm);
8341 				}
8342 #endif
8343 				if (rsm->r_in_tmap) {
8344 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8345 					nrsm->r_in_tmap = 1;
8346 				}
8347 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8348 				rsm->r_flags &= (~RACK_HAS_FIN);
8349 				/* Position us to point to the new nrsm that starts the sack blk */
8350 				rsm = nrsm;
8351 			}
8352 		} else {
8353 			/* Already sacked this piece */
8354 			counter_u64_add(rack_sack_skipped_acked, 1);
8355 			moved++;
8356 			if (end == rsm->r_end) {
8357 				/* Done with block */
8358 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8359 				goto out;
8360 			} else if (SEQ_LT(end, rsm->r_end)) {
8361 				/* A partial sack to a already sacked block */
8362 				moved++;
8363 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8364 				goto out;
8365 			} else {
8366 				/*
8367 				 * The end goes beyond this guy
8368 				 * repostion the start to the
8369 				 * next block.
8370 				 */
8371 				start = rsm->r_end;
8372 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8373 				if (rsm == NULL)
8374 					goto out;
8375 			}
8376 		}
8377 	}
8378 	if (SEQ_GEQ(end, rsm->r_end)) {
8379 		/**
8380 		 * The end of this block is either beyond this guy or right
8381 		 * at this guy. I.e.:
8382 		 *  rsm ---                 |-----|
8383 		 *  end                     |-----|
8384 		 *  <or>
8385 		 *  end                     |---------|
8386 		 */
8387 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8388 			/*
8389 			 * Is it a TLP of interest?
8390 			 */
8391 			if ((rsm->r_flags & RACK_TLP) &&
8392 			    (rsm->r_rtr_cnt > 1)) {
8393 				/*
8394 				 * We are splitting a rxt TLP, check
8395 				 * if we need to save off the start/end
8396 				 */
8397 				if (rack->rc_last_tlp_acked_set &&
8398 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8399 					/*
8400 					 * We already turned this on since we are inside
8401 					 * the previous one was a partially sack now we
8402 					 * are getting another one (maybe all of it).
8403 					 */
8404 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8405 					/*
8406 					 * Lets make sure we have all of it though.
8407 					 */
8408 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8409 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8410 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8411 								     rack->r_ctl.last_tlp_acked_end);
8412 					}
8413 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8414 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8415 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8416 								     rack->r_ctl.last_tlp_acked_end);
8417 					}
8418 				} else {
8419 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8420 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8421 					rack->rc_last_tlp_past_cumack = 0;
8422 					rack->rc_last_tlp_acked_set = 1;
8423 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8424 				}
8425 			}
8426 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8427 			changed += (rsm->r_end - rsm->r_start);
8428 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8429 			if (rsm->r_in_tmap) /* should be true */
8430 				rack_log_sack_passed(tp, rack, rsm);
8431 			/* Is Reordering occuring? */
8432 			if (rsm->r_flags & RACK_SACK_PASSED) {
8433 				rsm->r_flags &= ~RACK_SACK_PASSED;
8434 				rack->r_ctl.rc_reorder_ts = cts;
8435 			}
8436 			if (rack->app_limited_needs_set)
8437 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8438 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8439 			rsm->r_flags |= RACK_ACKED;
8440 			if (rsm->r_in_tmap) {
8441 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8442 				rsm->r_in_tmap = 0;
8443 			}
8444 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8445 		} else {
8446 			counter_u64_add(rack_sack_skipped_acked, 1);
8447 			moved++;
8448 		}
8449 		if (end == rsm->r_end) {
8450 			/* This block only - done, setup for next */
8451 			goto out;
8452 		}
8453 		/*
8454 		 * There is more not coverend by this rsm move on
8455 		 * to the next block in the RB tree.
8456 		 */
8457 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8458 		start = rsm->r_end;
8459 		rsm = nrsm;
8460 		if (rsm == NULL)
8461 			goto out;
8462 		goto do_rest_ofb;
8463 	}
8464 	/**
8465 	 * The end of this sack block is smaller than
8466 	 * our rsm i.e.:
8467 	 *  rsm ---                 |-----|
8468 	 *  end                     |--|
8469 	 */
8470 	if ((rsm->r_flags & RACK_ACKED) == 0) {
8471 		/*
8472 		 * Is it a TLP of interest?
8473 		 */
8474 		if ((rsm->r_flags & RACK_TLP) &&
8475 		    (rsm->r_rtr_cnt > 1)) {
8476 			/*
8477 			 * We are splitting a rxt TLP, check
8478 			 * if we need to save off the start/end
8479 			 */
8480 			if (rack->rc_last_tlp_acked_set &&
8481 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8482 				/*
8483 				 * We already turned this on since we are inside
8484 				 * the previous one was a partially sack now we
8485 				 * are getting another one (maybe all of it).
8486 				 */
8487 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8488 				/*
8489 				 * Lets make sure we have all of it though.
8490 				 */
8491 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8492 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8493 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8494 							     rack->r_ctl.last_tlp_acked_end);
8495 				}
8496 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8497 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8498 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8499 							     rack->r_ctl.last_tlp_acked_end);
8500 				}
8501 			} else {
8502 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8503 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8504 				rack->rc_last_tlp_past_cumack = 0;
8505 				rack->rc_last_tlp_acked_set = 1;
8506 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8507 			}
8508 		}
8509 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8510 		if (prev &&
8511 		    (prev->r_flags & RACK_ACKED)) {
8512 			/**
8513 			 * Goal, we want the right remainder of rsm to shrink
8514 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
8515 			 * We want to expand prev to go all the way
8516 			 * to prev->r_end <- end.
8517 			 * so in the tree we have before:
8518 			 *   prev     |--------|         (acked)
8519 			 *   rsm               |-------| (non-acked)
8520 			 *   sackblk           |-|
8521 			 * We churn it so we end up with
8522 			 *   prev     |----------|       (acked)
8523 			 *   rsm                 |-----| (non-acked)
8524 			 *   nrsm              |-| (temporary)
8525 			 *
8526 			 * Note if either prev/rsm is a TLP we don't
8527 			 * do this.
8528 			 */
8529 			nrsm = &stack_map;
8530 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8531 			prev->r_end = end;
8532 			rsm->r_start = end;
8533 			/* Now adjust nrsm (stack copy) to be
8534 			 * the one that is the small
8535 			 * piece that was "sacked".
8536 			 */
8537 			nrsm->r_end = end;
8538 			rsm->r_dupack = 0;
8539 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8540 			/*
8541 			 * Now that the rsm has had its start moved forward
8542 			 * lets go ahead and get its new place in the world.
8543 			 */
8544 			rack_setup_offset_for_rsm(prev, rsm);
8545 			/*
8546 			 * Now nrsm is our new little piece
8547 			 * that is acked (which was merged
8548 			 * to prev). Update the rtt and changed
8549 			 * based on that. Also check for reordering.
8550 			 */
8551 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8552 			if (rack->app_limited_needs_set)
8553 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8554 			changed += (nrsm->r_end - nrsm->r_start);
8555 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8556 			if (nrsm->r_flags & RACK_SACK_PASSED) {
8557 				rack->r_ctl.rc_reorder_ts = cts;
8558 			}
8559 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8560 			rsm = prev;
8561 			counter_u64_add(rack_sack_used_prev_merge, 1);
8562 		} else {
8563 			/**
8564 			 * This is the case where our previous
8565 			 * block is not acked either, so we must
8566 			 * split the block in two.
8567 			 */
8568 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8569 			if (nrsm == NULL) {
8570 				/* failed rrs what can we do but loose the sack info? */
8571 				goto out;
8572 			}
8573 			if ((rsm->r_flags & RACK_TLP) &&
8574 			    (rsm->r_rtr_cnt > 1)) {
8575 				/*
8576 				 * We are splitting a rxt TLP, check
8577 				 * if we need to save off the start/end
8578 				 */
8579 				if (rack->rc_last_tlp_acked_set &&
8580 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8581 					    /*
8582 					     * We already turned this on since this block is inside
8583 					     * the previous one was a partially sack now we
8584 					     * are getting another one (maybe all of it).
8585 					     */
8586 					    rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8587 					    /*
8588 					     * Lets make sure we have all of it though.
8589 					     */
8590 					    if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8591 						    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8592 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8593 									 rack->r_ctl.last_tlp_acked_end);
8594 					    }
8595 					    if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8596 						    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8597 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8598 									 rack->r_ctl.last_tlp_acked_end);
8599 					    }
8600 				    } else {
8601 					    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8602 					    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8603 					    rack->rc_last_tlp_acked_set = 1;
8604 					    rack->rc_last_tlp_past_cumack = 0;
8605 					    rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8606 				    }
8607 			}
8608 			/**
8609 			 * In this case nrsm becomes
8610 			 * nrsm->r_start = end;
8611 			 * nrsm->r_end = rsm->r_end;
8612 			 * which is un-acked.
8613 			 * <and>
8614 			 * rsm->r_end = nrsm->r_start;
8615 			 * i.e. the remaining un-acked
8616 			 * piece is left on the left
8617 			 * hand side.
8618 			 *
8619 			 * So we start like this
8620 			 * rsm      |----------| (not acked)
8621 			 * sackblk  |---|
8622 			 * build it so we have
8623 			 * rsm      |---|         (acked)
8624 			 * nrsm         |------|  (not acked)
8625 			 */
8626 			counter_u64_add(rack_sack_splits, 1);
8627 			rack_clone_rsm(rack, nrsm, rsm, end);
8628 			rsm->r_flags &= (~RACK_HAS_FIN);
8629 			rsm->r_just_ret = 0;
8630 #ifndef INVARIANTS
8631 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8632 #else
8633 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8634 			if (insret != NULL) {
8635 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8636 				      nrsm, insret, rack, rsm);
8637 			}
8638 #endif
8639 			if (rsm->r_in_tmap) {
8640 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8641 				nrsm->r_in_tmap = 1;
8642 			}
8643 			nrsm->r_dupack = 0;
8644 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8645 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8646 			changed += (rsm->r_end - rsm->r_start);
8647 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8648 			if (rsm->r_in_tmap) /* should be true */
8649 				rack_log_sack_passed(tp, rack, rsm);
8650 			/* Is Reordering occuring? */
8651 			if (rsm->r_flags & RACK_SACK_PASSED) {
8652 				rsm->r_flags &= ~RACK_SACK_PASSED;
8653 				rack->r_ctl.rc_reorder_ts = cts;
8654 			}
8655 			if (rack->app_limited_needs_set)
8656 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8657 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8658 			rsm->r_flags |= RACK_ACKED;
8659 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8660 			if (rsm->r_in_tmap) {
8661 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8662 				rsm->r_in_tmap = 0;
8663 			}
8664 		}
8665 	} else if (start != end){
8666 		/*
8667 		 * The block was already acked.
8668 		 */
8669 		counter_u64_add(rack_sack_skipped_acked, 1);
8670 		moved++;
8671 	}
8672 out:
8673 	if (rsm &&
8674 	    ((rsm->r_flags & RACK_TLP) == 0) &&
8675 	    (rsm->r_flags & RACK_ACKED)) {
8676 		/*
8677 		 * Now can we merge where we worked
8678 		 * with either the previous or
8679 		 * next block?
8680 		 */
8681 		next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8682 		while (next) {
8683 			if (next->r_flags & RACK_TLP)
8684 				break;
8685 			if (next->r_flags & RACK_ACKED) {
8686 			/* yep this and next can be merged */
8687 				rsm = rack_merge_rsm(rack, rsm, next);
8688 				next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8689 			} else
8690 				break;
8691 		}
8692 		/* Now what about the previous? */
8693 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8694 		while (prev) {
8695 			if (prev->r_flags & RACK_TLP)
8696 				break;
8697 			if (prev->r_flags & RACK_ACKED) {
8698 				/* yep the previous and this can be merged */
8699 				rsm = rack_merge_rsm(rack, prev, rsm);
8700 				prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8701 			} else
8702 				break;
8703 		}
8704 	}
8705 	if (used_ref == 0) {
8706 		counter_u64_add(rack_sack_proc_all, 1);
8707 	} else {
8708 		counter_u64_add(rack_sack_proc_short, 1);
8709 	}
8710 	/* Save off the next one for quick reference. */
8711 	if (rsm)
8712 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8713 	else
8714 		nrsm = NULL;
8715 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
8716 	/* Pass back the moved. */
8717 	*moved_two = moved;
8718 	return (changed);
8719 }
8720 
8721 static void inline
8722 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8723 {
8724 	struct rack_sendmap *tmap;
8725 
8726 	tmap = NULL;
8727 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
8728 		/* Its no longer sacked, mark it so */
8729 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8730 #ifdef INVARIANTS
8731 		if (rsm->r_in_tmap) {
8732 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
8733 			      rack, rsm, rsm->r_flags);
8734 		}
8735 #endif
8736 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8737 		/* Rebuild it into our tmap */
8738 		if (tmap == NULL) {
8739 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8740 			tmap = rsm;
8741 		} else {
8742 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8743 			tmap = rsm;
8744 		}
8745 		tmap->r_in_tmap = 1;
8746 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8747 	}
8748 	/*
8749 	 * Now lets possibly clear the sack filter so we start
8750 	 * recognizing sacks that cover this area.
8751 	 */
8752 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8753 
8754 }
8755 
8756 static void
8757 rack_do_decay(struct tcp_rack *rack)
8758 {
8759 	struct timeval res;
8760 
8761 #define	timersub(tvp, uvp, vvp)						\
8762 	do {								\
8763 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
8764 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
8765 		if ((vvp)->tv_usec < 0) {				\
8766 			(vvp)->tv_sec--;				\
8767 			(vvp)->tv_usec += 1000000;			\
8768 		}							\
8769 	} while (0)
8770 
8771 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8772 #undef timersub
8773 
8774 	rack->r_ctl.input_pkt++;
8775 	if ((rack->rc_in_persist) ||
8776 	    (res.tv_sec >= 1) ||
8777 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8778 		/*
8779 		 * Check for decay of non-SAD,
8780 		 * we want all SAD detection metrics to
8781 		 * decay 1/4 per second (or more) passed.
8782 		 */
8783 #ifdef NETFLIX_EXP_DETECTION
8784 		uint32_t pkt_delta;
8785 
8786 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8787 #endif
8788 		/* Update our saved tracking values */
8789 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8790 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8791 		/* Now do we escape without decay? */
8792 #ifdef NETFLIX_EXP_DETECTION
8793 		if (rack->rc_in_persist ||
8794 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8795 		    (pkt_delta < tcp_sad_low_pps)){
8796 			/*
8797 			 * We don't decay idle connections
8798 			 * or ones that have a low input pps.
8799 			 */
8800 			return;
8801 		}
8802 		/* Decay the counters */
8803 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8804 							tcp_sad_decay_val);
8805 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8806 							 tcp_sad_decay_val);
8807 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8808 							       tcp_sad_decay_val);
8809 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8810 								tcp_sad_decay_val);
8811 #endif
8812 	}
8813 }
8814 
8815 static void
8816 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8817 {
8818 	struct rack_sendmap *rsm;
8819 #ifdef INVARIANTS
8820 	struct rack_sendmap *rm;
8821 #endif
8822 
8823 	/*
8824 	 * The ACK point is advancing to th_ack, we must drop off
8825 	 * the packets in the rack log and calculate any eligble
8826 	 * RTT's.
8827 	 */
8828 	rack->r_wanted_output = 1;
8829 
8830 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
8831 	if ((rack->rc_last_tlp_acked_set == 1)&&
8832 	    (rack->rc_last_tlp_past_cumack == 1) &&
8833 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
8834 		/*
8835 		 * We have reached the point where our last rack
8836 		 * tlp retransmit sequence is ahead of the cum-ack.
8837 		 * This can only happen when the cum-ack moves all
8838 		 * the way around (its been a full 2^^31+1 bytes
8839 		 * or more since we sent a retransmitted TLP). Lets
8840 		 * turn off the valid flag since its not really valid.
8841 		 *
8842 		 * Note since sack's also turn on this event we have
8843 		 * a complication, we have to wait to age it out until
8844 		 * the cum-ack is by the TLP before checking which is
8845 		 * what the next else clause does.
8846 		 */
8847 		rack_log_dsack_event(rack, 9, __LINE__,
8848 				     rack->r_ctl.last_tlp_acked_start,
8849 				     rack->r_ctl.last_tlp_acked_end);
8850 		rack->rc_last_tlp_acked_set = 0;
8851 		rack->rc_last_tlp_past_cumack = 0;
8852 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
8853 		   (rack->rc_last_tlp_past_cumack == 0) &&
8854 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
8855 		/*
8856 		 * It is safe to start aging TLP's out.
8857 		 */
8858 		rack->rc_last_tlp_past_cumack = 1;
8859 	}
8860 	/* We do the same for the tlp send seq as well */
8861 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8862 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
8863 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
8864 		rack_log_dsack_event(rack, 9, __LINE__,
8865 				     rack->r_ctl.last_sent_tlp_seq,
8866 				     (rack->r_ctl.last_sent_tlp_seq +
8867 				      rack->r_ctl.last_sent_tlp_len));
8868 		rack->rc_last_sent_tlp_seq_valid = 0;
8869 		rack->rc_last_sent_tlp_past_cumack = 0;
8870 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8871 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
8872 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
8873 		/*
8874 		 * It is safe to start aging TLP's send.
8875 		 */
8876 		rack->rc_last_sent_tlp_past_cumack = 1;
8877 	}
8878 more:
8879 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8880 	if (rsm == NULL) {
8881 		if ((th_ack - 1) == tp->iss) {
8882 			/*
8883 			 * For the SYN incoming case we will not
8884 			 * have called tcp_output for the sending of
8885 			 * the SYN, so there will be no map. All
8886 			 * other cases should probably be a panic.
8887 			 */
8888 			return;
8889 		}
8890 		if (tp->t_flags & TF_SENTFIN) {
8891 			/* if we sent a FIN we often will not have map */
8892 			return;
8893 		}
8894 #ifdef INVARIANTS
8895 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8896 		      tp,
8897 		      tp->t_state, th_ack, rack,
8898 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
8899 #endif
8900 		return;
8901 	}
8902 	if (SEQ_LT(th_ack, rsm->r_start)) {
8903 		/* Huh map is missing this */
8904 #ifdef INVARIANTS
8905 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8906 		       rsm->r_start,
8907 		       th_ack, tp->t_state, rack->r_state);
8908 #endif
8909 		return;
8910 	}
8911 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8912 
8913 	/* Now was it a retransmitted TLP? */
8914 	if ((rsm->r_flags & RACK_TLP) &&
8915 	    (rsm->r_rtr_cnt > 1)) {
8916 		/*
8917 		 * Yes, this rsm was a TLP and retransmitted, remember that
8918 		 * since if a DSACK comes back on this we don't want
8919 		 * to think of it as a reordered segment. This may
8920 		 * get updated again with possibly even other TLPs
8921 		 * in flight, but thats ok. Only when we don't send
8922 		 * a retransmitted TLP for 1/2 the sequences space
8923 		 * will it get turned off (above).
8924 		 */
8925 		if (rack->rc_last_tlp_acked_set &&
8926 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8927 			/*
8928 			 * We already turned this on since the end matches,
8929 			 * the previous one was a partially ack now we
8930 			 * are getting another one (maybe all of it).
8931 			 */
8932 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8933 			/*
8934 			 * Lets make sure we have all of it though.
8935 			 */
8936 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8937 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8938 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8939 						     rack->r_ctl.last_tlp_acked_end);
8940 			}
8941 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8942 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8943 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8944 						     rack->r_ctl.last_tlp_acked_end);
8945 			}
8946 		} else {
8947 			rack->rc_last_tlp_past_cumack = 1;
8948 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8949 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8950 			rack->rc_last_tlp_acked_set = 1;
8951 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8952 		}
8953 	}
8954 	/* Now do we consume the whole thing? */
8955 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
8956 		/* Its all consumed. */
8957 		uint32_t left;
8958 		uint8_t newly_acked;
8959 
8960 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8961 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8962 		rsm->r_rtr_bytes = 0;
8963 		/* Record the time of highest cumack sent */
8964 		rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8965 #ifndef INVARIANTS
8966 		(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8967 #else
8968 		rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8969 		if (rm != rsm) {
8970 			panic("removing head in rack:%p rsm:%p rm:%p",
8971 			      rack, rsm, rm);
8972 		}
8973 #endif
8974 		if (rsm->r_in_tmap) {
8975 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8976 			rsm->r_in_tmap = 0;
8977 		}
8978 		newly_acked = 1;
8979 		if (rsm->r_flags & RACK_ACKED) {
8980 			/*
8981 			 * It was acked on the scoreboard -- remove
8982 			 * it from total
8983 			 */
8984 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8985 			newly_acked = 0;
8986 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
8987 			/*
8988 			 * There are segments ACKED on the
8989 			 * scoreboard further up. We are seeing
8990 			 * reordering.
8991 			 */
8992 			rsm->r_flags &= ~RACK_SACK_PASSED;
8993 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8994 			rsm->r_flags |= RACK_ACKED;
8995 			rack->r_ctl.rc_reorder_ts = cts;
8996 			if (rack->r_ent_rec_ns) {
8997 				/*
8998 				 * We have sent no more, and we saw an sack
8999 				 * then ack arrive.
9000 				 */
9001 				rack->r_might_revert = 1;
9002 			}
9003 		}
9004 		if ((rsm->r_flags & RACK_TO_REXT) &&
9005 		    (tp->t_flags & TF_RCVD_TSTMP) &&
9006 		    (to->to_flags & TOF_TS) &&
9007 		    (to->to_tsecr != 0) &&
9008 		    (tp->t_flags & TF_PREVVALID)) {
9009 			/*
9010 			 * We can use the timestamp to see
9011 			 * if this retransmission was from the
9012 			 * first transmit. If so we made a mistake.
9013 			 */
9014 			tp->t_flags &= ~TF_PREVVALID;
9015 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9016 				/* The first transmit is what this ack is for */
9017 				rack_cong_signal(tp, CC_RTO_ERR, th_ack);
9018 			}
9019 		}
9020 		left = th_ack - rsm->r_end;
9021 		if (rack->app_limited_needs_set && newly_acked)
9022 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9023 		/* Free back to zone */
9024 		rack_free(rack, rsm);
9025 		if (left) {
9026 			goto more;
9027 		}
9028 		/* Check for reneging */
9029 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9030 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9031 			/*
9032 			 * The peer has moved snd_una up to
9033 			 * the edge of this send, i.e. one
9034 			 * that it had previously acked. The only
9035 			 * way that can be true if the peer threw
9036 			 * away data (space issues) that it had
9037 			 * previously sacked (else it would have
9038 			 * given us snd_una up to (rsm->r_end).
9039 			 * We need to undo the acked markings here.
9040 			 *
9041 			 * Note we have to look to make sure th_ack is
9042 			 * our rsm->r_start in case we get an old ack
9043 			 * where th_ack is behind snd_una.
9044 			 */
9045 			rack_peer_reneges(rack, rsm, th_ack);
9046 		}
9047 		return;
9048 	}
9049 	if (rsm->r_flags & RACK_ACKED) {
9050 		/*
9051 		 * It was acked on the scoreboard -- remove it from
9052 		 * total for the part being cum-acked.
9053 		 */
9054 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9055 	}
9056 	/*
9057 	 * Clear the dup ack count for
9058 	 * the piece that remains.
9059 	 */
9060 	rsm->r_dupack = 0;
9061 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9062 	if (rsm->r_rtr_bytes) {
9063 		/*
9064 		 * It was retransmitted adjust the
9065 		 * sack holes for what was acked.
9066 		 */
9067 		int ack_am;
9068 
9069 		ack_am = (th_ack - rsm->r_start);
9070 		if (ack_am >= rsm->r_rtr_bytes) {
9071 			rack->r_ctl.rc_holes_rxt -= ack_am;
9072 			rsm->r_rtr_bytes -= ack_am;
9073 		}
9074 	}
9075 	/*
9076 	 * Update where the piece starts and record
9077 	 * the time of send of highest cumack sent.
9078 	 */
9079 	rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9080 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9081 	/* Now we need to move our offset forward too */
9082 	if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9083 		/* Fix up the orig_m_len and possibly the mbuf offset */
9084 		rack_adjust_orig_mlen(rsm);
9085 	}
9086 	rsm->soff += (th_ack - rsm->r_start);
9087 	rsm->r_start = th_ack;
9088 	/* Now do we need to move the mbuf fwd too? */
9089 	if (rsm->m) {
9090 		while (rsm->soff >= rsm->m->m_len) {
9091 			rsm->soff -= rsm->m->m_len;
9092 			rsm->m = rsm->m->m_next;
9093 			KASSERT((rsm->m != NULL),
9094 				(" nrsm:%p hit at soff:%u null m",
9095 				 rsm, rsm->soff));
9096 		}
9097 		rsm->orig_m_len = rsm->m->m_len;
9098 	}
9099 	if (rack->app_limited_needs_set)
9100 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9101 }
9102 
9103 static void
9104 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9105 {
9106 	struct rack_sendmap *rsm;
9107 	int sack_pass_fnd = 0;
9108 
9109 	if (rack->r_might_revert) {
9110 		/*
9111 		 * Ok we have reordering, have not sent anything, we
9112 		 * might want to revert the congestion state if nothing
9113 		 * further has SACK_PASSED on it. Lets check.
9114 		 *
9115 		 * We also get here when we have DSACKs come in for
9116 		 * all the data that we FR'd. Note that a rxt or tlp
9117 		 * timer clears this from happening.
9118 		 */
9119 
9120 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9121 			if (rsm->r_flags & RACK_SACK_PASSED) {
9122 				sack_pass_fnd = 1;
9123 				break;
9124 			}
9125 		}
9126 		if (sack_pass_fnd == 0) {
9127 			/*
9128 			 * We went into recovery
9129 			 * incorrectly due to reordering!
9130 			 */
9131 			int orig_cwnd;
9132 
9133 			rack->r_ent_rec_ns = 0;
9134 			orig_cwnd = tp->snd_cwnd;
9135 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9136 			tp->snd_recover = tp->snd_una;
9137 			rack_log_to_prr(rack, 14, orig_cwnd);
9138 			EXIT_RECOVERY(tp->t_flags);
9139 		}
9140 		rack->r_might_revert = 0;
9141 	}
9142 }
9143 
9144 #ifdef NETFLIX_EXP_DETECTION
9145 static void
9146 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9147 {
9148 	if ((rack->do_detection || tcp_force_detection) &&
9149 	    tcp_sack_to_ack_thresh &&
9150 	    tcp_sack_to_move_thresh &&
9151 	    ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9152 		/*
9153 		 * We have thresholds set to find
9154 		 * possible attackers and disable sack.
9155 		 * Check them.
9156 		 */
9157 		uint64_t ackratio, moveratio, movetotal;
9158 
9159 		/* Log detecting */
9160 		rack_log_sad(rack, 1);
9161 		ackratio = (uint64_t)(rack->r_ctl.sack_count);
9162 		ackratio *= (uint64_t)(1000);
9163 		if (rack->r_ctl.ack_count)
9164 			ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9165 		else {
9166 			/* We really should not hit here */
9167 			ackratio = 1000;
9168 		}
9169 		if ((rack->sack_attack_disable == 0) &&
9170 		    (ackratio > rack_highest_sack_thresh_seen))
9171 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9172 		movetotal = rack->r_ctl.sack_moved_extra;
9173 		movetotal += rack->r_ctl.sack_noextra_move;
9174 		moveratio = rack->r_ctl.sack_moved_extra;
9175 		moveratio *= (uint64_t)1000;
9176 		if (movetotal)
9177 			moveratio /= movetotal;
9178 		else {
9179 			/* No moves, thats pretty good */
9180 			moveratio = 0;
9181 		}
9182 		if ((rack->sack_attack_disable == 0) &&
9183 		    (moveratio > rack_highest_move_thresh_seen))
9184 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
9185 		if (rack->sack_attack_disable == 0) {
9186 			if ((ackratio > tcp_sack_to_ack_thresh) &&
9187 			    (moveratio > tcp_sack_to_move_thresh)) {
9188 				/* Disable sack processing */
9189 				rack->sack_attack_disable = 1;
9190 				if (rack->r_rep_attack == 0) {
9191 					rack->r_rep_attack = 1;
9192 					counter_u64_add(rack_sack_attacks_detected, 1);
9193 				}
9194 				if (tcp_attack_on_turns_on_logging) {
9195 					/*
9196 					 * Turn on logging, used for debugging
9197 					 * false positives.
9198 					 */
9199 					rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9200 				}
9201 				/* Clamp the cwnd at flight size */
9202 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9203 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9204 				rack_log_sad(rack, 2);
9205 			}
9206 		} else {
9207 			/* We are sack-disabled check for false positives */
9208 			if ((ackratio <= tcp_restoral_thresh) ||
9209 			    (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9210 				rack->sack_attack_disable = 0;
9211 				rack_log_sad(rack, 3);
9212 				/* Restart counting */
9213 				rack->r_ctl.sack_count = 0;
9214 				rack->r_ctl.sack_moved_extra = 0;
9215 				rack->r_ctl.sack_noextra_move = 1;
9216 				rack->r_ctl.ack_count = max(1,
9217 				      (bytes_this_ack / segsiz));
9218 
9219 				if (rack->r_rep_reverse == 0) {
9220 					rack->r_rep_reverse = 1;
9221 					counter_u64_add(rack_sack_attacks_reversed, 1);
9222 				}
9223 				/* Restore the cwnd */
9224 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9225 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9226 			}
9227 		}
9228 	}
9229 }
9230 #endif
9231 
9232 static int
9233 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9234 {
9235 
9236 	uint32_t am, l_end;
9237 	int was_tlp = 0;
9238 
9239 	if (SEQ_GT(end, start))
9240 		am = end - start;
9241 	else
9242 		am = 0;
9243 	if ((rack->rc_last_tlp_acked_set ) &&
9244 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9245 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9246 		/*
9247 		 * The DSACK is because of a TLP which we don't
9248 		 * do anything with the reordering window over since
9249 		 * it was not reordering that caused the DSACK but
9250 		 * our previous retransmit TLP.
9251 		 */
9252 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
9253 		was_tlp = 1;
9254 		goto skip_dsack_round;
9255 	}
9256 	if (rack->rc_last_sent_tlp_seq_valid) {
9257 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9258 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9259 		    (SEQ_LEQ(end, l_end))) {
9260 			/*
9261 			 * This dsack is from the last sent TLP, ignore it
9262 			 * for reordering purposes.
9263 			 */
9264 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
9265 			was_tlp = 1;
9266 			goto skip_dsack_round;
9267 		}
9268 	}
9269 	if (rack->rc_dsack_round_seen == 0) {
9270 		rack->rc_dsack_round_seen = 1;
9271 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9272 		rack->r_ctl.num_dsack++;
9273 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
9274 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9275 	}
9276 skip_dsack_round:
9277 	/*
9278 	 * We keep track of how many DSACK blocks we get
9279 	 * after a recovery incident.
9280 	 */
9281 	rack->r_ctl.dsack_byte_cnt += am;
9282 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9283 	    rack->r_ctl.retran_during_recovery &&
9284 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9285 		/*
9286 		 * False recovery most likely culprit is reordering. If
9287 		 * nothing else is missing we need to revert.
9288 		 */
9289 		rack->r_might_revert = 1;
9290 		rack_handle_might_revert(rack->rc_tp, rack);
9291 		rack->r_might_revert = 0;
9292 		rack->r_ctl.retran_during_recovery = 0;
9293 		rack->r_ctl.dsack_byte_cnt = 0;
9294 	}
9295 	return (was_tlp);
9296 }
9297 
9298 static void
9299 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9300 {
9301 	/* Deal with changed and PRR here (in recovery only) */
9302 	uint32_t pipe, snd_una;
9303 
9304 	rack->r_ctl.rc_prr_delivered += changed;
9305 
9306 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9307 		/*
9308 		 * It is all outstanding, we are application limited
9309 		 * and thus we don't need more room to send anything.
9310 		 * Note we use tp->snd_una here and not th_ack because
9311 		 * the data as yet not been cut from the sb.
9312 		 */
9313 		rack->r_ctl.rc_prr_sndcnt = 0;
9314 		return;
9315 	}
9316 	/* Compute prr_sndcnt */
9317 	if (SEQ_GT(tp->snd_una, th_ack)) {
9318 		snd_una = tp->snd_una;
9319 	} else {
9320 		snd_una = th_ack;
9321 	}
9322 	pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9323 	if (pipe > tp->snd_ssthresh) {
9324 		long sndcnt;
9325 
9326 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9327 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
9328 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9329 		else {
9330 			rack->r_ctl.rc_prr_sndcnt = 0;
9331 			rack_log_to_prr(rack, 9, 0);
9332 			sndcnt = 0;
9333 		}
9334 		sndcnt++;
9335 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9336 			sndcnt -= rack->r_ctl.rc_prr_out;
9337 		else
9338 			sndcnt = 0;
9339 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
9340 		rack_log_to_prr(rack, 10, 0);
9341 	} else {
9342 		uint32_t limit;
9343 
9344 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9345 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9346 		else
9347 			limit = 0;
9348 		if (changed > limit)
9349 			limit = changed;
9350 		limit += ctf_fixed_maxseg(tp);
9351 		if (tp->snd_ssthresh > pipe) {
9352 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9353 			rack_log_to_prr(rack, 11, 0);
9354 		} else {
9355 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9356 			rack_log_to_prr(rack, 12, 0);
9357 		}
9358 	}
9359 }
9360 
9361 static void
9362 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9363 {
9364 	uint32_t changed;
9365 	struct tcp_rack *rack;
9366 	struct rack_sendmap *rsm;
9367 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9368 	register uint32_t th_ack;
9369 	int32_t i, j, k, num_sack_blks = 0;
9370 	uint32_t cts, acked, ack_point;
9371 	int loop_start = 0, moved_two = 0;
9372 	uint32_t tsused;
9373 
9374 
9375 	INP_WLOCK_ASSERT(tp->t_inpcb);
9376 	if (tcp_get_flags(th) & TH_RST) {
9377 		/* We don't log resets */
9378 		return;
9379 	}
9380 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9381 	cts = tcp_get_usecs(NULL);
9382 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9383 	changed = 0;
9384 	th_ack = th->th_ack;
9385 	if (rack->sack_attack_disable == 0)
9386 		rack_do_decay(rack);
9387 	if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9388 		/*
9389 		 * You only get credit for
9390 		 * MSS and greater (and you get extra
9391 		 * credit for larger cum-ack moves).
9392 		 */
9393 		int ac;
9394 
9395 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9396 		rack->r_ctl.ack_count += ac;
9397 		counter_u64_add(rack_ack_total, ac);
9398 	}
9399 	if (rack->r_ctl.ack_count > 0xfff00000) {
9400 		/*
9401 		 * reduce the number to keep us under
9402 		 * a uint32_t.
9403 		 */
9404 		rack->r_ctl.ack_count /= 2;
9405 		rack->r_ctl.sack_count /= 2;
9406 	}
9407 	if (SEQ_GT(th_ack, tp->snd_una)) {
9408 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9409 		tp->t_acktime = ticks;
9410 	}
9411 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
9412 		changed = th_ack - rsm->r_start;
9413 	if (changed) {
9414 		rack_process_to_cumack(tp, rack, th_ack, cts, to);
9415 	}
9416 	if ((to->to_flags & TOF_SACK) == 0) {
9417 		/* We are done nothing left and no sack. */
9418 		rack_handle_might_revert(tp, rack);
9419 		/*
9420 		 * For cases where we struck a dup-ack
9421 		 * with no SACK, add to the changes so
9422 		 * PRR will work right.
9423 		 */
9424 		if (dup_ack_struck && (changed == 0)) {
9425 			changed += ctf_fixed_maxseg(rack->rc_tp);
9426 		}
9427 		goto out;
9428 	}
9429 	/* Sack block processing */
9430 	if (SEQ_GT(th_ack, tp->snd_una))
9431 		ack_point = th_ack;
9432 	else
9433 		ack_point = tp->snd_una;
9434 	for (i = 0; i < to->to_nsacks; i++) {
9435 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
9436 		      &sack, sizeof(sack));
9437 		sack.start = ntohl(sack.start);
9438 		sack.end = ntohl(sack.end);
9439 		if (SEQ_GT(sack.end, sack.start) &&
9440 		    SEQ_GT(sack.start, ack_point) &&
9441 		    SEQ_LT(sack.start, tp->snd_max) &&
9442 		    SEQ_GT(sack.end, ack_point) &&
9443 		    SEQ_LEQ(sack.end, tp->snd_max)) {
9444 			sack_blocks[num_sack_blks] = sack;
9445 			num_sack_blks++;
9446 		} else if (SEQ_LEQ(sack.start, th_ack) &&
9447 			   SEQ_LEQ(sack.end, th_ack)) {
9448 			int was_tlp;
9449 
9450 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9451 			/*
9452 			 * Its a D-SACK block.
9453 			 */
9454 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9455 		}
9456 	}
9457 	if (rack->rc_dsack_round_seen) {
9458 		/* Is the dsack roound over? */
9459 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9460 			/* Yes it is */
9461 			rack->rc_dsack_round_seen = 0;
9462 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9463 		}
9464 	}
9465 	/*
9466 	 * Sort the SACK blocks so we can update the rack scoreboard with
9467 	 * just one pass.
9468 	 */
9469 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9470 					 num_sack_blks, th->th_ack);
9471 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9472 	if (num_sack_blks == 0) {
9473 		/* Nothing to sack (DSACKs?) */
9474 		goto out_with_totals;
9475 	}
9476 	if (num_sack_blks < 2) {
9477 		/* Only one, we don't need to sort */
9478 		goto do_sack_work;
9479 	}
9480 	/* Sort the sacks */
9481 	for (i = 0; i < num_sack_blks; i++) {
9482 		for (j = i + 1; j < num_sack_blks; j++) {
9483 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9484 				sack = sack_blocks[i];
9485 				sack_blocks[i] = sack_blocks[j];
9486 				sack_blocks[j] = sack;
9487 			}
9488 		}
9489 	}
9490 	/*
9491 	 * Now are any of the sack block ends the same (yes some
9492 	 * implementations send these)?
9493 	 */
9494 again:
9495 	if (num_sack_blks == 0)
9496 		goto out_with_totals;
9497 	if (num_sack_blks > 1) {
9498 		for (i = 0; i < num_sack_blks; i++) {
9499 			for (j = i + 1; j < num_sack_blks; j++) {
9500 				if (sack_blocks[i].end == sack_blocks[j].end) {
9501 					/*
9502 					 * Ok these two have the same end we
9503 					 * want the smallest end and then
9504 					 * throw away the larger and start
9505 					 * again.
9506 					 */
9507 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9508 						/*
9509 						 * The second block covers
9510 						 * more area use that
9511 						 */
9512 						sack_blocks[i].start = sack_blocks[j].start;
9513 					}
9514 					/*
9515 					 * Now collapse out the dup-sack and
9516 					 * lower the count
9517 					 */
9518 					for (k = (j + 1); k < num_sack_blks; k++) {
9519 						sack_blocks[j].start = sack_blocks[k].start;
9520 						sack_blocks[j].end = sack_blocks[k].end;
9521 						j++;
9522 					}
9523 					num_sack_blks--;
9524 					goto again;
9525 				}
9526 			}
9527 		}
9528 	}
9529 do_sack_work:
9530 	/*
9531 	 * First lets look to see if
9532 	 * we have retransmitted and
9533 	 * can use the transmit next?
9534 	 */
9535 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9536 	if (rsm &&
9537 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9538 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9539 		/*
9540 		 * We probably did the FR and the next
9541 		 * SACK in continues as we would expect.
9542 		 */
9543 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9544 		if (acked) {
9545 			rack->r_wanted_output = 1;
9546 			changed += acked;
9547 		}
9548 		if (num_sack_blks == 1) {
9549 			/*
9550 			 * This is what we would expect from
9551 			 * a normal implementation to happen
9552 			 * after we have retransmitted the FR,
9553 			 * i.e the sack-filter pushes down
9554 			 * to 1 block and the next to be retransmitted
9555 			 * is the sequence in the sack block (has more
9556 			 * are acked). Count this as ACK'd data to boost
9557 			 * up the chances of recovering any false positives.
9558 			 */
9559 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9560 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9561 			counter_u64_add(rack_express_sack, 1);
9562 			if (rack->r_ctl.ack_count > 0xfff00000) {
9563 				/*
9564 				 * reduce the number to keep us under
9565 				 * a uint32_t.
9566 				 */
9567 				rack->r_ctl.ack_count /= 2;
9568 				rack->r_ctl.sack_count /= 2;
9569 			}
9570 			goto out_with_totals;
9571 		} else {
9572 			/*
9573 			 * Start the loop through the
9574 			 * rest of blocks, past the first block.
9575 			 */
9576 			moved_two = 0;
9577 			loop_start = 1;
9578 		}
9579 	}
9580 	/* Its a sack of some sort */
9581 	rack->r_ctl.sack_count++;
9582 	if (rack->r_ctl.sack_count > 0xfff00000) {
9583 		/*
9584 		 * reduce the number to keep us under
9585 		 * a uint32_t.
9586 		 */
9587 		rack->r_ctl.ack_count /= 2;
9588 		rack->r_ctl.sack_count /= 2;
9589 	}
9590 	counter_u64_add(rack_sack_total, 1);
9591 	if (rack->sack_attack_disable) {
9592 		/* An attacker disablement is in place */
9593 		if (num_sack_blks > 1) {
9594 			rack->r_ctl.sack_count += (num_sack_blks - 1);
9595 			rack->r_ctl.sack_moved_extra++;
9596 			counter_u64_add(rack_move_some, 1);
9597 			if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9598 				rack->r_ctl.sack_moved_extra /= 2;
9599 				rack->r_ctl.sack_noextra_move /= 2;
9600 			}
9601 		}
9602 		goto out;
9603 	}
9604 	rsm = rack->r_ctl.rc_sacklast;
9605 	for (i = loop_start; i < num_sack_blks; i++) {
9606 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9607 		if (acked) {
9608 			rack->r_wanted_output = 1;
9609 			changed += acked;
9610 		}
9611 		if (moved_two) {
9612 			/*
9613 			 * If we did not get a SACK for at least a MSS and
9614 			 * had to move at all, or if we moved more than our
9615 			 * threshold, it counts against the "extra" move.
9616 			 */
9617 			rack->r_ctl.sack_moved_extra += moved_two;
9618 			counter_u64_add(rack_move_some, 1);
9619 		} else {
9620 			/*
9621 			 * else we did not have to move
9622 			 * any more than we would expect.
9623 			 */
9624 			rack->r_ctl.sack_noextra_move++;
9625 			counter_u64_add(rack_move_none, 1);
9626 		}
9627 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9628 			/*
9629 			 * If the SACK was not a full MSS then
9630 			 * we add to sack_count the number of
9631 			 * MSS's (or possibly more than
9632 			 * a MSS if its a TSO send) we had to skip by.
9633 			 */
9634 			rack->r_ctl.sack_count += moved_two;
9635 			counter_u64_add(rack_sack_total, moved_two);
9636 		}
9637 		/*
9638 		 * Now we need to setup for the next
9639 		 * round. First we make sure we won't
9640 		 * exceed the size of our uint32_t on
9641 		 * the various counts, and then clear out
9642 		 * moved_two.
9643 		 */
9644 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9645 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9646 			rack->r_ctl.sack_moved_extra /= 2;
9647 			rack->r_ctl.sack_noextra_move /= 2;
9648 		}
9649 		if (rack->r_ctl.sack_count > 0xfff00000) {
9650 			rack->r_ctl.ack_count /= 2;
9651 			rack->r_ctl.sack_count /= 2;
9652 		}
9653 		moved_two = 0;
9654 	}
9655 out_with_totals:
9656 	if (num_sack_blks > 1) {
9657 		/*
9658 		 * You get an extra stroke if
9659 		 * you have more than one sack-blk, this
9660 		 * could be where we are skipping forward
9661 		 * and the sack-filter is still working, or
9662 		 * it could be an attacker constantly
9663 		 * moving us.
9664 		 */
9665 		rack->r_ctl.sack_moved_extra++;
9666 		counter_u64_add(rack_move_some, 1);
9667 	}
9668 out:
9669 #ifdef NETFLIX_EXP_DETECTION
9670 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9671 #endif
9672 	if (changed) {
9673 		/* Something changed cancel the rack timer */
9674 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9675 	}
9676 	tsused = tcp_get_usecs(NULL);
9677 	rsm = tcp_rack_output(tp, rack, tsused);
9678 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9679 	    rsm) {
9680 		/* Enter recovery */
9681 		entered_recovery = 1;
9682 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9683 		/*
9684 		 * When we enter recovery we need to assure we send
9685 		 * one packet.
9686 		 */
9687 		if (rack->rack_no_prr == 0) {
9688 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9689 			rack_log_to_prr(rack, 8, 0);
9690 		}
9691 		rack->r_timer_override = 1;
9692 		rack->r_early = 0;
9693 		rack->r_ctl.rc_agg_early = 0;
9694 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
9695 		   rsm &&
9696 		   (rack->r_rr_config == 3)) {
9697 		/*
9698 		 * Assure we can output and we get no
9699 		 * remembered pace time except the retransmit.
9700 		 */
9701 		rack->r_timer_override = 1;
9702 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9703 		rack->r_ctl.rc_resend = rsm;
9704 	}
9705 	if (IN_FASTRECOVERY(tp->t_flags) &&
9706 	    (rack->rack_no_prr == 0) &&
9707 	    (entered_recovery == 0)) {
9708 		rack_update_prr(tp, rack, changed, th_ack);
9709 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9710 		     ((tcp_in_hpts(rack->rc_inp) == 0) &&
9711 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9712 			/*
9713 			 * If you are pacing output you don't want
9714 			 * to override.
9715 			 */
9716 			rack->r_early = 0;
9717 			rack->r_ctl.rc_agg_early = 0;
9718 			rack->r_timer_override = 1;
9719 		}
9720 	}
9721 }
9722 
9723 static void
9724 rack_strike_dupack(struct tcp_rack *rack)
9725 {
9726 	struct rack_sendmap *rsm;
9727 
9728 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9729 	while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9730 		rsm = TAILQ_NEXT(rsm, r_tnext);
9731 	}
9732 	if (rsm && (rsm->r_dupack < 0xff)) {
9733 		rsm->r_dupack++;
9734 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9735 			struct timeval tv;
9736 			uint32_t cts;
9737 			/*
9738 			 * Here we see if we need to retransmit. For
9739 			 * a SACK type connection if enough time has passed
9740 			 * we will get a return of the rsm. For a non-sack
9741 			 * connection we will get the rsm returned if the
9742 			 * dupack value is 3 or more.
9743 			 */
9744 			cts = tcp_get_usecs(&tv);
9745 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9746 			if (rack->r_ctl.rc_resend != NULL) {
9747 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9748 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9749 							 rack->rc_tp->snd_una);
9750 				}
9751 				rack->r_wanted_output = 1;
9752 				rack->r_timer_override = 1;
9753 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9754 			}
9755 		} else {
9756 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9757 		}
9758 	}
9759 }
9760 
9761 static void
9762 rack_check_bottom_drag(struct tcpcb *tp,
9763 		       struct tcp_rack *rack,
9764 		       struct socket *so, int32_t acked)
9765 {
9766 	uint32_t segsiz, minseg;
9767 
9768 	segsiz = ctf_fixed_maxseg(tp);
9769 	minseg = segsiz;
9770 
9771 	if (tp->snd_max == tp->snd_una) {
9772 		/*
9773 		 * We are doing dynamic pacing and we are way
9774 		 * under. Basically everything got acked while
9775 		 * we were still waiting on the pacer to expire.
9776 		 *
9777 		 * This means we need to boost the b/w in
9778 		 * addition to any earlier boosting of
9779 		 * the multipler.
9780 		 */
9781 		rack->rc_dragged_bottom = 1;
9782 		rack_validate_multipliers_at_or_above100(rack);
9783 		/*
9784 		 * Lets use the segment bytes acked plus
9785 		 * the lowest RTT seen as the basis to
9786 		 * form a b/w estimate. This will be off
9787 		 * due to the fact that the true estimate
9788 		 * should be around 1/2 the time of the RTT
9789 		 * but we can settle for that.
9790 		 */
9791 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9792 		    acked) {
9793 			uint64_t bw, calc_bw, rtt;
9794 
9795 			rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9796 			if (rtt == 0) {
9797 				/* no us sample is there a ms one? */
9798 				if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9799 					rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9800 				} else {
9801 					goto no_measurement;
9802 				}
9803 			}
9804 			bw = acked;
9805 			calc_bw = bw * 1000000;
9806 			calc_bw /= rtt;
9807 			if (rack->r_ctl.last_max_bw &&
9808 			    (rack->r_ctl.last_max_bw < calc_bw)) {
9809 				/*
9810 				 * If we have a last calculated max bw
9811 				 * enforce it.
9812 				 */
9813 				calc_bw = rack->r_ctl.last_max_bw;
9814 			}
9815 			/* now plop it in */
9816 			if (rack->rc_gp_filled == 0) {
9817 				if (calc_bw > ONE_POINT_TWO_MEG) {
9818 					/*
9819 					 * If we have no measurement
9820 					 * don't let us set in more than
9821 					 * 1.2Mbps. If we are still too
9822 					 * low after pacing with this we
9823 					 * will hopefully have a max b/w
9824 					 * available to sanity check things.
9825 					 */
9826 					calc_bw = ONE_POINT_TWO_MEG;
9827 				}
9828 				rack->r_ctl.rc_rtt_diff = 0;
9829 				rack->r_ctl.gp_bw = calc_bw;
9830 				rack->rc_gp_filled = 1;
9831 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9832 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9833 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9834 			} else if (calc_bw > rack->r_ctl.gp_bw) {
9835 				rack->r_ctl.rc_rtt_diff = 0;
9836 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9837 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9838 				rack->r_ctl.gp_bw = calc_bw;
9839 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9840 			} else
9841 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9842 			if ((rack->gp_ready == 0) &&
9843 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9844 				/* We have enough measurements now */
9845 				rack->gp_ready = 1;
9846 				rack_set_cc_pacing(rack);
9847 				if (rack->defer_options)
9848 					rack_apply_deferred_options(rack);
9849 			}
9850 			/*
9851 			 * For acks over 1mss we do a extra boost to simulate
9852 			 * where we would get 2 acks (we want 110 for the mul).
9853 			 */
9854 			if (acked > segsiz)
9855 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9856 		} else {
9857 			/*
9858 			 * zero rtt possibly?, settle for just an old increase.
9859 			 */
9860 no_measurement:
9861 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
9862 		}
9863 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9864 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9865 					       minseg)) &&
9866 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9867 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9868 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9869 		    (segsiz * rack_req_segs))) {
9870 		/*
9871 		 * We are doing dynamic GP pacing and
9872 		 * we have everything except 1MSS or less
9873 		 * bytes left out. We are still pacing away.
9874 		 * And there is data that could be sent, This
9875 		 * means we are inserting delayed ack time in
9876 		 * our measurements because we are pacing too slow.
9877 		 */
9878 		rack_validate_multipliers_at_or_above100(rack);
9879 		rack->rc_dragged_bottom = 1;
9880 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
9881 	}
9882 }
9883 
9884 
9885 
9886 static void
9887 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9888 {
9889 	/*
9890 	 * The fast output path is enabled and we
9891 	 * have moved the cumack forward. Lets see if
9892 	 * we can expand forward the fast path length by
9893 	 * that amount. What we would ideally like to
9894 	 * do is increase the number of bytes in the
9895 	 * fast path block (left_to_send) by the
9896 	 * acked amount. However we have to gate that
9897 	 * by two factors:
9898 	 * 1) The amount outstanding and the rwnd of the peer
9899 	 *    (i.e. we don't want to exceed the rwnd of the peer).
9900 	 *    <and>
9901 	 * 2) The amount of data left in the socket buffer (i.e.
9902 	 *    we can't send beyond what is in the buffer).
9903 	 *
9904 	 * Note that this does not take into account any increase
9905 	 * in the cwnd. We will only extend the fast path by
9906 	 * what was acked.
9907 	 */
9908 	uint32_t new_total, gating_val;
9909 
9910 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9911 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9912 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9913 	if (new_total <= gating_val) {
9914 		/* We can increase left_to_send by the acked amount */
9915 		counter_u64_add(rack_extended_rfo, 1);
9916 		rack->r_ctl.fsb.left_to_send = new_total;
9917 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9918 			("rack:%p left_to_send:%u sbavail:%u out:%u",
9919 			 rack, rack->r_ctl.fsb.left_to_send,
9920 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
9921 			 (tp->snd_max - tp->snd_una)));
9922 
9923 	}
9924 }
9925 
9926 static void
9927 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9928 {
9929 	/*
9930 	 * Here any sendmap entry that points to the
9931 	 * beginning mbuf must be adjusted to the correct
9932 	 * offset. This must be called with:
9933 	 * 1) The socket buffer locked
9934 	 * 2) snd_una adjusted to its new postion.
9935 	 *
9936 	 * Note that (2) implies rack_ack_received has also
9937 	 * been called.
9938 	 *
9939 	 * We grab the first mbuf in the socket buffer and
9940 	 * then go through the front of the sendmap, recalculating
9941 	 * the stored offset for any sendmap entry that has
9942 	 * that mbuf. We must use the sb functions to do this
9943 	 * since its possible an add was done has well as
9944 	 * the subtraction we may have just completed. This should
9945 	 * not be a penalty though, since we just referenced the sb
9946 	 * to go in and trim off the mbufs that we freed (of course
9947 	 * there will be a penalty for the sendmap references though).
9948 	 */
9949 	struct mbuf *m;
9950 	struct rack_sendmap *rsm;
9951 
9952 	SOCKBUF_LOCK_ASSERT(sb);
9953 	m = sb->sb_mb;
9954 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9955 	if ((rsm == NULL) || (m == NULL)) {
9956 		/* Nothing outstanding */
9957 		return;
9958 	}
9959 	while (rsm->m && (rsm->m == m)) {
9960 		/* one to adjust */
9961 #ifdef INVARIANTS
9962 		struct mbuf *tm;
9963 		uint32_t soff;
9964 
9965 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9966 		if (rsm->orig_m_len != m->m_len) {
9967 			rack_adjust_orig_mlen(rsm);
9968 		}
9969 		if (rsm->soff != soff) {
9970 			/*
9971 			 * This is not a fatal error, we anticipate it
9972 			 * might happen (the else code), so we count it here
9973 			 * so that under invariant we can see that it really
9974 			 * does happen.
9975 			 */
9976 			counter_u64_add(rack_adjust_map_bw, 1);
9977 		}
9978 		rsm->m = tm;
9979 		rsm->soff = soff;
9980 		if (tm)
9981 			rsm->orig_m_len = rsm->m->m_len;
9982 		else
9983 			rsm->orig_m_len = 0;
9984 #else
9985 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
9986 		if (rsm->m)
9987 			rsm->orig_m_len = rsm->m->m_len;
9988 		else
9989 			rsm->orig_m_len = 0;
9990 #endif
9991 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
9992 			      rsm);
9993 		if (rsm == NULL)
9994 			break;
9995 	}
9996 }
9997 
9998 /*
9999  * Return value of 1, we do not need to call rack_process_data().
10000  * return value of 0, rack_process_data can be called.
10001  * For ret_val if its 0 the TCP is locked, if its non-zero
10002  * its unlocked and probably unsafe to touch the TCB.
10003  */
10004 static int
10005 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10006     struct tcpcb *tp, struct tcpopt *to,
10007     uint32_t tiwin, int32_t tlen,
10008     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10009 {
10010 	int32_t ourfinisacked = 0;
10011 	int32_t nsegs, acked_amount;
10012 	int32_t acked;
10013 	struct mbuf *mfree;
10014 	struct tcp_rack *rack;
10015 	int32_t under_pacing = 0;
10016 	int32_t recovery = 0;
10017 
10018 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10019 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
10020 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10021 				      &rack->r_ctl.challenge_ack_ts,
10022 				      &rack->r_ctl.challenge_ack_cnt);
10023 		rack->r_wanted_output = 1;
10024 		return (1);
10025 	}
10026 	if (rack->gp_ready &&
10027 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10028 		under_pacing = 1;
10029 	}
10030 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10031 		int in_rec, dup_ack_struck = 0;
10032 
10033 		in_rec = IN_FASTRECOVERY(tp->t_flags);
10034 		if (rack->rc_in_persist) {
10035 			tp->t_rxtshift = 0;
10036 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10037 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10038 		}
10039 		if ((th->th_ack == tp->snd_una) &&
10040 		    (tiwin == tp->snd_wnd) &&
10041 		    ((to->to_flags & TOF_SACK) == 0)) {
10042 			rack_strike_dupack(rack);
10043 			dup_ack_struck = 1;
10044 		}
10045 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10046 	}
10047 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10048 		/*
10049 		 * Old ack, behind (or duplicate to) the last one rcv'd
10050 		 * Note: We mark reordering is occuring if its
10051 		 * less than and we have not closed our window.
10052 		 */
10053 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10054 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10055 		}
10056 		return (0);
10057 	}
10058 	/*
10059 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10060 	 * something we sent.
10061 	 */
10062 	if (tp->t_flags & TF_NEEDSYN) {
10063 		/*
10064 		 * T/TCP: Connection was half-synchronized, and our SYN has
10065 		 * been ACK'd (so connection is now fully synchronized).  Go
10066 		 * to non-starred state, increment snd_una for ACK of SYN,
10067 		 * and check if we can do window scaling.
10068 		 */
10069 		tp->t_flags &= ~TF_NEEDSYN;
10070 		tp->snd_una++;
10071 		/* Do window scaling? */
10072 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10073 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10074 			tp->rcv_scale = tp->request_r_scale;
10075 			/* Send window already scaled. */
10076 		}
10077 	}
10078 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10079 	INP_WLOCK_ASSERT(tp->t_inpcb);
10080 
10081 	acked = BYTES_THIS_ACK(tp, th);
10082 	if (acked) {
10083 		/*
10084 		 * Any time we move the cum-ack forward clear
10085 		 * keep-alive tied probe-not-answered. The
10086 		 * persists clears its own on entry.
10087 		 */
10088 		rack->probe_not_answered = 0;
10089 	}
10090 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10091 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10092 	/*
10093 	 * If we just performed our first retransmit, and the ACK arrives
10094 	 * within our recovery window, then it was a mistake to do the
10095 	 * retransmit in the first place.  Recover our original cwnd and
10096 	 * ssthresh, and proceed to transmit where we left off.
10097 	 */
10098 	if ((tp->t_flags & TF_PREVVALID) &&
10099 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10100 		tp->t_flags &= ~TF_PREVVALID;
10101 		if (tp->t_rxtshift == 1 &&
10102 		    (int)(ticks - tp->t_badrxtwin) < 0)
10103 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10104 	}
10105 	if (acked) {
10106 		/* assure we are not backed off */
10107 		tp->t_rxtshift = 0;
10108 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10109 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10110 		rack->rc_tlp_in_progress = 0;
10111 		rack->r_ctl.rc_tlp_cnt_out = 0;
10112 		/*
10113 		 * If it is the RXT timer we want to
10114 		 * stop it, so we can restart a TLP.
10115 		 */
10116 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10117 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10118 #ifdef NETFLIX_HTTP_LOGGING
10119 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10120 #endif
10121 	}
10122 	/*
10123 	 * If we have a timestamp reply, update smoothed round trip time. If
10124 	 * no timestamp is present but transmit timer is running and timed
10125 	 * sequence number was acked, update smoothed round trip time. Since
10126 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
10127 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
10128 	 * timer.
10129 	 *
10130 	 * Some boxes send broken timestamp replies during the SYN+ACK
10131 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10132 	 * and blow up the retransmit timer.
10133 	 */
10134 	/*
10135 	 * If all outstanding data is acked, stop retransmit timer and
10136 	 * remember to restart (more output or persist). If there is more
10137 	 * data to be acked, restart retransmit timer, using current
10138 	 * (possibly backed-off) value.
10139 	 */
10140 	if (acked == 0) {
10141 		if (ofia)
10142 			*ofia = ourfinisacked;
10143 		return (0);
10144 	}
10145 	if (IN_RECOVERY(tp->t_flags)) {
10146 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10147 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
10148 			tcp_rack_partialack(tp);
10149 		} else {
10150 			rack_post_recovery(tp, th->th_ack);
10151 			recovery = 1;
10152 		}
10153 	}
10154 	/*
10155 	 * Let the congestion control algorithm update congestion control
10156 	 * related information. This typically means increasing the
10157 	 * congestion window.
10158 	 */
10159 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10160 	SOCKBUF_LOCK(&so->so_snd);
10161 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
10162 	tp->snd_wnd -= acked_amount;
10163 	mfree = sbcut_locked(&so->so_snd, acked_amount);
10164 	if ((sbused(&so->so_snd) == 0) &&
10165 	    (acked > acked_amount) &&
10166 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
10167 	    (tp->t_flags & TF_SENTFIN)) {
10168 		/*
10169 		 * We must be sure our fin
10170 		 * was sent and acked (we can be
10171 		 * in FIN_WAIT_1 without having
10172 		 * sent the fin).
10173 		 */
10174 		ourfinisacked = 1;
10175 	}
10176 	tp->snd_una = th->th_ack;
10177 	if (acked_amount && sbavail(&so->so_snd))
10178 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10179 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10180 	/* NB: sowwakeup_locked() does an implicit unlock. */
10181 	sowwakeup_locked(so);
10182 	m_freem(mfree);
10183 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
10184 		tp->snd_recover = tp->snd_una;
10185 
10186 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10187 		tp->snd_nxt = tp->snd_una;
10188 	}
10189 	if (under_pacing &&
10190 	    (rack->use_fixed_rate == 0) &&
10191 	    (rack->in_probe_rtt == 0) &&
10192 	    rack->rc_gp_dyn_mul &&
10193 	    rack->rc_always_pace) {
10194 		/* Check if we are dragging bottom */
10195 		rack_check_bottom_drag(tp, rack, so, acked);
10196 	}
10197 	if (tp->snd_una == tp->snd_max) {
10198 		/* Nothing left outstanding */
10199 		tp->t_flags &= ~TF_PREVVALID;
10200 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10201 		rack->r_ctl.retran_during_recovery = 0;
10202 		rack->r_ctl.dsack_byte_cnt = 0;
10203 		if (rack->r_ctl.rc_went_idle_time == 0)
10204 			rack->r_ctl.rc_went_idle_time = 1;
10205 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10206 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10207 			tp->t_acktime = 0;
10208 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10209 		/* Set need output so persist might get set */
10210 		rack->r_wanted_output = 1;
10211 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10212 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10213 		    (sbavail(&so->so_snd) == 0) &&
10214 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10215 			/*
10216 			 * The socket was gone and the
10217 			 * peer sent data (now or in the past), time to
10218 			 * reset him.
10219 			 */
10220 			*ret_val = 1;
10221 			/* tcp_close will kill the inp pre-log the Reset */
10222 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10223 			tp = tcp_close(tp);
10224 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10225 			return (1);
10226 		}
10227 	}
10228 	if (ofia)
10229 		*ofia = ourfinisacked;
10230 	return (0);
10231 }
10232 
10233 static void
10234 rack_collapsed_window(struct tcp_rack *rack)
10235 {
10236 	/*
10237 	 * Now we must walk the
10238 	 * send map and divide the
10239 	 * ones left stranded. These
10240 	 * guys can't cause us to abort
10241 	 * the connection and are really
10242 	 * "unsent". However if a buggy
10243 	 * client actually did keep some
10244 	 * of the data i.e. collapsed the win
10245 	 * and refused to ack and then opened
10246 	 * the win and acked that data. We would
10247 	 * get into an ack war, the simplier
10248 	 * method then of just pretending we
10249 	 * did not send those segments something
10250 	 * won't work.
10251 	 */
10252 	struct rack_sendmap *rsm, *nrsm, fe;
10253 #ifdef INVARIANTS
10254 	struct rack_sendmap *insret;
10255 #endif
10256 	tcp_seq max_seq;
10257 
10258 	max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10259 	memset(&fe, 0, sizeof(fe));
10260 	fe.r_start = max_seq;
10261 	/* Find the first seq past or at maxseq */
10262 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10263 	if (rsm == NULL) {
10264 		/* Nothing to do strange */
10265 		rack->rc_has_collapsed = 0;
10266 		return;
10267 	}
10268 	/*
10269 	 * Now do we need to split at
10270 	 * the collapse point?
10271 	 */
10272 	if (SEQ_GT(max_seq, rsm->r_start)) {
10273 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10274 		if (nrsm == NULL) {
10275 			/* We can't get a rsm, mark all? */
10276 			nrsm = rsm;
10277 			goto no_split;
10278 		}
10279 		/* Clone it */
10280 		rack_clone_rsm(rack, nrsm, rsm, max_seq);
10281 #ifndef INVARIANTS
10282 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10283 #else
10284 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10285 		if (insret != NULL) {
10286 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10287 			      nrsm, insret, rack, rsm);
10288 		}
10289 #endif
10290 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
10291 		if (rsm->r_in_tmap) {
10292 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10293 			nrsm->r_in_tmap = 1;
10294 		}
10295 		/*
10296 		 * Set in the new RSM as the
10297 		 * collapsed starting point
10298 		 */
10299 		rsm = nrsm;
10300 	}
10301 no_split:
10302 	counter_u64_add(rack_collapsed_win, 1);
10303 	RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10304 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
10305 	}
10306 	rack->rc_has_collapsed = 1;
10307 }
10308 
10309 static void
10310 rack_un_collapse_window(struct tcp_rack *rack)
10311 {
10312 	struct rack_sendmap *rsm;
10313 
10314 	RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
10315 		if (rsm->r_flags & RACK_RWND_COLLAPSED)
10316 			rsm->r_flags &= ~RACK_RWND_COLLAPSED;
10317 		else
10318 			break;
10319 	}
10320 	rack->rc_has_collapsed = 0;
10321 }
10322 
10323 static void
10324 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10325 			int32_t tlen, int32_t tfo_syn)
10326 {
10327 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
10328 		if (rack->rc_dack_mode &&
10329 		    (tlen > 500) &&
10330 		    (rack->rc_dack_toggle == 1)) {
10331 			goto no_delayed_ack;
10332 		}
10333 		rack_timer_cancel(tp, rack,
10334 				  rack->r_ctl.rc_rcvtime, __LINE__);
10335 		tp->t_flags |= TF_DELACK;
10336 	} else {
10337 no_delayed_ack:
10338 		rack->r_wanted_output = 1;
10339 		tp->t_flags |= TF_ACKNOW;
10340 		if (rack->rc_dack_mode) {
10341 			if (tp->t_flags & TF_DELACK)
10342 				rack->rc_dack_toggle = 1;
10343 			else
10344 				rack->rc_dack_toggle = 0;
10345 		}
10346 	}
10347 }
10348 
10349 static void
10350 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10351 {
10352 	/*
10353 	 * If fast output is in progress, lets validate that
10354 	 * the new window did not shrink on us and make it
10355 	 * so fast output should end.
10356 	 */
10357 	if (rack->r_fast_output) {
10358 		uint32_t out;
10359 
10360 		/*
10361 		 * Calculate what we will send if left as is
10362 		 * and compare that to our send window.
10363 		 */
10364 		out = ctf_outstanding(tp);
10365 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10366 			/* ok we have an issue */
10367 			if (out >= tp->snd_wnd) {
10368 				/* Turn off fast output the window is met or collapsed */
10369 				rack->r_fast_output = 0;
10370 			} else {
10371 				/* we have some room left */
10372 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10373 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10374 					/* If not at least 1 full segment never mind */
10375 					rack->r_fast_output = 0;
10376 				}
10377 			}
10378 		}
10379 	}
10380 }
10381 
10382 
10383 /*
10384  * Return value of 1, the TCB is unlocked and most
10385  * likely gone, return value of 0, the TCP is still
10386  * locked.
10387  */
10388 static int
10389 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10390     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10391     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10392 {
10393 	/*
10394 	 * Update window information. Don't look at window if no ACK: TAC's
10395 	 * send garbage on first SYN.
10396 	 */
10397 	int32_t nsegs;
10398 	int32_t tfo_syn;
10399 	struct tcp_rack *rack;
10400 
10401 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10402 	INP_WLOCK_ASSERT(tp->t_inpcb);
10403 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10404 	if ((thflags & TH_ACK) &&
10405 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10406 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10407 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10408 		/* keep track of pure window updates */
10409 		if (tlen == 0 &&
10410 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10411 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10412 		tp->snd_wnd = tiwin;
10413 		rack_validate_fo_sendwin_up(tp, rack);
10414 		tp->snd_wl1 = th->th_seq;
10415 		tp->snd_wl2 = th->th_ack;
10416 		if (tp->snd_wnd > tp->max_sndwnd)
10417 			tp->max_sndwnd = tp->snd_wnd;
10418 		rack->r_wanted_output = 1;
10419 	} else if (thflags & TH_ACK) {
10420 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10421 			tp->snd_wnd = tiwin;
10422 			rack_validate_fo_sendwin_up(tp, rack);
10423 			tp->snd_wl1 = th->th_seq;
10424 			tp->snd_wl2 = th->th_ack;
10425 		}
10426 	}
10427 	if (tp->snd_wnd < ctf_outstanding(tp))
10428 		/* The peer collapsed the window */
10429 		rack_collapsed_window(rack);
10430 	else if (rack->rc_has_collapsed)
10431 		rack_un_collapse_window(rack);
10432 	/* Was persist timer active and now we have window space? */
10433 	if ((rack->rc_in_persist != 0) &&
10434 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10435 				rack->r_ctl.rc_pace_min_segs))) {
10436 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10437 		tp->snd_nxt = tp->snd_max;
10438 		/* Make sure we output to start the timer */
10439 		rack->r_wanted_output = 1;
10440 	}
10441 	/* Do we enter persists? */
10442 	if ((rack->rc_in_persist == 0) &&
10443 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10444 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10445 	    (tp->snd_max == tp->snd_una) &&
10446 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10447 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10448 		/*
10449 		 * Here the rwnd is less than
10450 		 * the pacing size, we are established,
10451 		 * nothing is outstanding, and there is
10452 		 * data to send. Enter persists.
10453 		 */
10454 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10455 	}
10456 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10457 		m_freem(m);
10458 		return (0);
10459 	}
10460 	/*
10461 	 * don't process the URG bit, ignore them drag
10462 	 * along the up.
10463 	 */
10464 	tp->rcv_up = tp->rcv_nxt;
10465 	INP_WLOCK_ASSERT(tp->t_inpcb);
10466 
10467 	/*
10468 	 * Process the segment text, merging it into the TCP sequencing
10469 	 * queue, and arranging for acknowledgment of receipt if necessary.
10470 	 * This process logically involves adjusting tp->rcv_wnd as data is
10471 	 * presented to the user (this happens in tcp_usrreq.c, case
10472 	 * PRU_RCVD).  If a FIN has already been received on this connection
10473 	 * then we just ignore the text.
10474 	 */
10475 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10476 		   IS_FASTOPEN(tp->t_flags));
10477 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10478 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10479 		tcp_seq save_start = th->th_seq;
10480 		tcp_seq save_rnxt  = tp->rcv_nxt;
10481 		int     save_tlen  = tlen;
10482 
10483 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10484 		/*
10485 		 * Insert segment which includes th into TCP reassembly
10486 		 * queue with control block tp.  Set thflags to whether
10487 		 * reassembly now includes a segment with FIN.  This handles
10488 		 * the common case inline (segment is the next to be
10489 		 * received on an established connection, and the queue is
10490 		 * empty), avoiding linkage into and removal from the queue
10491 		 * and repetition of various conversions. Set DELACK for
10492 		 * segments received in order, but ack immediately when
10493 		 * segments are out of order (so fast retransmit can work).
10494 		 */
10495 		if (th->th_seq == tp->rcv_nxt &&
10496 		    SEGQ_EMPTY(tp) &&
10497 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
10498 		    tfo_syn)) {
10499 #ifdef NETFLIX_SB_LIMITS
10500 			u_int mcnt, appended;
10501 
10502 			if (so->so_rcv.sb_shlim) {
10503 				mcnt = m_memcnt(m);
10504 				appended = 0;
10505 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10506 				    CFO_NOSLEEP, NULL) == false) {
10507 					counter_u64_add(tcp_sb_shlim_fails, 1);
10508 					m_freem(m);
10509 					return (0);
10510 				}
10511 			}
10512 #endif
10513 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10514 			tp->rcv_nxt += tlen;
10515 			if (tlen &&
10516 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10517 			    (tp->t_fbyte_in == 0)) {
10518 				tp->t_fbyte_in = ticks;
10519 				if (tp->t_fbyte_in == 0)
10520 					tp->t_fbyte_in = 1;
10521 				if (tp->t_fbyte_out && tp->t_fbyte_in)
10522 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10523 			}
10524 			thflags = tcp_get_flags(th) & TH_FIN;
10525 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10526 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10527 			SOCKBUF_LOCK(&so->so_rcv);
10528 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10529 				m_freem(m);
10530 			} else
10531 #ifdef NETFLIX_SB_LIMITS
10532 				appended =
10533 #endif
10534 					sbappendstream_locked(&so->so_rcv, m, 0);
10535 
10536 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10537 			/* NB: sorwakeup_locked() does an implicit unlock. */
10538 			sorwakeup_locked(so);
10539 #ifdef NETFLIX_SB_LIMITS
10540 			if (so->so_rcv.sb_shlim && appended != mcnt)
10541 				counter_fo_release(so->so_rcv.sb_shlim,
10542 				    mcnt - appended);
10543 #endif
10544 		} else {
10545 			/*
10546 			 * XXX: Due to the header drop above "th" is
10547 			 * theoretically invalid by now.  Fortunately
10548 			 * m_adj() doesn't actually frees any mbufs when
10549 			 * trimming from the head.
10550 			 */
10551 			tcp_seq temp = save_start;
10552 
10553 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
10554 			tp->t_flags |= TF_ACKNOW;
10555 			if (tp->t_flags & TF_WAKESOR) {
10556 				tp->t_flags &= ~TF_WAKESOR;
10557 				/* NB: sorwakeup_locked() does an implicit unlock. */
10558 				sorwakeup_locked(so);
10559 			}
10560 		}
10561 		if ((tp->t_flags & TF_SACK_PERMIT) &&
10562 		    (save_tlen > 0) &&
10563 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
10564 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10565 				/*
10566 				 * DSACK actually handled in the fastpath
10567 				 * above.
10568 				 */
10569 				RACK_OPTS_INC(tcp_sack_path_1);
10570 				tcp_update_sack_list(tp, save_start,
10571 				    save_start + save_tlen);
10572 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10573 				if ((tp->rcv_numsacks >= 1) &&
10574 				    (tp->sackblks[0].end == save_start)) {
10575 					/*
10576 					 * Partial overlap, recorded at todrop
10577 					 * above.
10578 					 */
10579 					RACK_OPTS_INC(tcp_sack_path_2a);
10580 					tcp_update_sack_list(tp,
10581 					    tp->sackblks[0].start,
10582 					    tp->sackblks[0].end);
10583 				} else {
10584 					RACK_OPTS_INC(tcp_sack_path_2b);
10585 					tcp_update_dsack_list(tp, save_start,
10586 					    save_start + save_tlen);
10587 				}
10588 			} else if (tlen >= save_tlen) {
10589 				/* Update of sackblks. */
10590 				RACK_OPTS_INC(tcp_sack_path_3);
10591 				tcp_update_dsack_list(tp, save_start,
10592 				    save_start + save_tlen);
10593 			} else if (tlen > 0) {
10594 				RACK_OPTS_INC(tcp_sack_path_4);
10595 				tcp_update_dsack_list(tp, save_start,
10596 				    save_start + tlen);
10597 			}
10598 		}
10599 	} else {
10600 		m_freem(m);
10601 		thflags &= ~TH_FIN;
10602 	}
10603 
10604 	/*
10605 	 * If FIN is received ACK the FIN and let the user know that the
10606 	 * connection is closing.
10607 	 */
10608 	if (thflags & TH_FIN) {
10609 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10610 			/* The socket upcall is handled by socantrcvmore. */
10611 			socantrcvmore(so);
10612 			/*
10613 			 * If connection is half-synchronized (ie NEEDSYN
10614 			 * flag on) then delay ACK, so it may be piggybacked
10615 			 * when SYN is sent. Otherwise, since we received a
10616 			 * FIN then no more input can be expected, send ACK
10617 			 * now.
10618 			 */
10619 			if (tp->t_flags & TF_NEEDSYN) {
10620 				rack_timer_cancel(tp, rack,
10621 				    rack->r_ctl.rc_rcvtime, __LINE__);
10622 				tp->t_flags |= TF_DELACK;
10623 			} else {
10624 				tp->t_flags |= TF_ACKNOW;
10625 			}
10626 			tp->rcv_nxt++;
10627 		}
10628 		switch (tp->t_state) {
10629 			/*
10630 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
10631 			 * CLOSE_WAIT state.
10632 			 */
10633 		case TCPS_SYN_RECEIVED:
10634 			tp->t_starttime = ticks;
10635 			/* FALLTHROUGH */
10636 		case TCPS_ESTABLISHED:
10637 			rack_timer_cancel(tp, rack,
10638 			    rack->r_ctl.rc_rcvtime, __LINE__);
10639 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
10640 			break;
10641 
10642 			/*
10643 			 * If still in FIN_WAIT_1 STATE FIN has not been
10644 			 * acked so enter the CLOSING state.
10645 			 */
10646 		case TCPS_FIN_WAIT_1:
10647 			rack_timer_cancel(tp, rack,
10648 			    rack->r_ctl.rc_rcvtime, __LINE__);
10649 			tcp_state_change(tp, TCPS_CLOSING);
10650 			break;
10651 
10652 			/*
10653 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
10654 			 * starting the time-wait timer, turning off the
10655 			 * other standard timers.
10656 			 */
10657 		case TCPS_FIN_WAIT_2:
10658 			rack_timer_cancel(tp, rack,
10659 			    rack->r_ctl.rc_rcvtime, __LINE__);
10660 			tcp_twstart(tp);
10661 			return (1);
10662 		}
10663 	}
10664 	/*
10665 	 * Return any desired output.
10666 	 */
10667 	if ((tp->t_flags & TF_ACKNOW) ||
10668 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10669 		rack->r_wanted_output = 1;
10670 	}
10671 	INP_WLOCK_ASSERT(tp->t_inpcb);
10672 	return (0);
10673 }
10674 
10675 /*
10676  * Here nothing is really faster, its just that we
10677  * have broken out the fast-data path also just like
10678  * the fast-ack.
10679  */
10680 static int
10681 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10682     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10683     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10684 {
10685 	int32_t nsegs;
10686 	int32_t newsize = 0;	/* automatic sockbuf scaling */
10687 	struct tcp_rack *rack;
10688 #ifdef NETFLIX_SB_LIMITS
10689 	u_int mcnt, appended;
10690 #endif
10691 #ifdef TCPDEBUG
10692 	/*
10693 	 * The size of tcp_saveipgen must be the size of the max ip header,
10694 	 * now IPv6.
10695 	 */
10696 	u_char tcp_saveipgen[IP6_HDR_LEN];
10697 	struct tcphdr tcp_savetcp;
10698 	short ostate = 0;
10699 
10700 #endif
10701 	/*
10702 	 * If last ACK falls within this segment's sequence numbers, record
10703 	 * the timestamp. NOTE that the test is modified according to the
10704 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10705 	 */
10706 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10707 		return (0);
10708 	}
10709 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10710 		return (0);
10711 	}
10712 	if (tiwin && tiwin != tp->snd_wnd) {
10713 		return (0);
10714 	}
10715 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10716 		return (0);
10717 	}
10718 	if (__predict_false((to->to_flags & TOF_TS) &&
10719 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10720 		return (0);
10721 	}
10722 	if (__predict_false((th->th_ack != tp->snd_una))) {
10723 		return (0);
10724 	}
10725 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10726 		return (0);
10727 	}
10728 	if ((to->to_flags & TOF_TS) != 0 &&
10729 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10730 		tp->ts_recent_age = tcp_ts_getticks();
10731 		tp->ts_recent = to->to_tsval;
10732 	}
10733 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10734 	/*
10735 	 * This is a pure, in-sequence data packet with nothing on the
10736 	 * reassembly queue and we have enough buffer space to take it.
10737 	 */
10738 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10739 
10740 #ifdef NETFLIX_SB_LIMITS
10741 	if (so->so_rcv.sb_shlim) {
10742 		mcnt = m_memcnt(m);
10743 		appended = 0;
10744 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10745 		    CFO_NOSLEEP, NULL) == false) {
10746 			counter_u64_add(tcp_sb_shlim_fails, 1);
10747 			m_freem(m);
10748 			return (1);
10749 		}
10750 	}
10751 #endif
10752 	/* Clean receiver SACK report if present */
10753 	if (tp->rcv_numsacks)
10754 		tcp_clean_sackreport(tp);
10755 	KMOD_TCPSTAT_INC(tcps_preddat);
10756 	tp->rcv_nxt += tlen;
10757 	if (tlen &&
10758 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10759 	    (tp->t_fbyte_in == 0)) {
10760 		tp->t_fbyte_in = ticks;
10761 		if (tp->t_fbyte_in == 0)
10762 			tp->t_fbyte_in = 1;
10763 		if (tp->t_fbyte_out && tp->t_fbyte_in)
10764 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10765 	}
10766 	/*
10767 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10768 	 */
10769 	tp->snd_wl1 = th->th_seq;
10770 	/*
10771 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10772 	 */
10773 	tp->rcv_up = tp->rcv_nxt;
10774 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10775 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10776 #ifdef TCPDEBUG
10777 	if (so->so_options & SO_DEBUG)
10778 		tcp_trace(TA_INPUT, ostate, tp,
10779 		    (void *)tcp_saveipgen, &tcp_savetcp, 0);
10780 #endif
10781 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10782 
10783 	/* Add data to socket buffer. */
10784 	SOCKBUF_LOCK(&so->so_rcv);
10785 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10786 		m_freem(m);
10787 	} else {
10788 		/*
10789 		 * Set new socket buffer size. Give up when limit is
10790 		 * reached.
10791 		 */
10792 		if (newsize)
10793 			if (!sbreserve_locked(&so->so_rcv,
10794 			    newsize, so, NULL))
10795 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10796 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10797 #ifdef NETFLIX_SB_LIMITS
10798 		appended =
10799 #endif
10800 			sbappendstream_locked(&so->so_rcv, m, 0);
10801 		ctf_calc_rwin(so, tp);
10802 	}
10803 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10804 	/* NB: sorwakeup_locked() does an implicit unlock. */
10805 	sorwakeup_locked(so);
10806 #ifdef NETFLIX_SB_LIMITS
10807 	if (so->so_rcv.sb_shlim && mcnt != appended)
10808 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10809 #endif
10810 	rack_handle_delayed_ack(tp, rack, tlen, 0);
10811 	if (tp->snd_una == tp->snd_max)
10812 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10813 	return (1);
10814 }
10815 
10816 /*
10817  * This subfunction is used to try to highly optimize the
10818  * fast path. We again allow window updates that are
10819  * in sequence to remain in the fast-path. We also add
10820  * in the __predict's to attempt to help the compiler.
10821  * Note that if we return a 0, then we can *not* process
10822  * it and the caller should push the packet into the
10823  * slow-path.
10824  */
10825 static int
10826 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10827     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10828     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10829 {
10830 	int32_t acked;
10831 	int32_t nsegs;
10832 #ifdef TCPDEBUG
10833 	/*
10834 	 * The size of tcp_saveipgen must be the size of the max ip header,
10835 	 * now IPv6.
10836 	 */
10837 	u_char tcp_saveipgen[IP6_HDR_LEN];
10838 	struct tcphdr tcp_savetcp;
10839 	short ostate = 0;
10840 #endif
10841 	int32_t under_pacing = 0;
10842 	struct tcp_rack *rack;
10843 
10844 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10845 		/* Old ack, behind (or duplicate to) the last one rcv'd */
10846 		return (0);
10847 	}
10848 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10849 		/* Above what we have sent? */
10850 		return (0);
10851 	}
10852 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10853 		/* We are retransmitting */
10854 		return (0);
10855 	}
10856 	if (__predict_false(tiwin == 0)) {
10857 		/* zero window */
10858 		return (0);
10859 	}
10860 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10861 		/* We need a SYN or a FIN, unlikely.. */
10862 		return (0);
10863 	}
10864 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10865 		/* Timestamp is behind .. old ack with seq wrap? */
10866 		return (0);
10867 	}
10868 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10869 		/* Still recovering */
10870 		return (0);
10871 	}
10872 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10873 	if (rack->r_ctl.rc_sacked) {
10874 		/* We have sack holes on our scoreboard */
10875 		return (0);
10876 	}
10877 	/* Ok if we reach here, we can process a fast-ack */
10878 	if (rack->gp_ready &&
10879 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10880 		under_pacing = 1;
10881 	}
10882 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10883 	rack_log_ack(tp, to, th, 0, 0);
10884 	/* Did the window get updated? */
10885 	if (tiwin != tp->snd_wnd) {
10886 		tp->snd_wnd = tiwin;
10887 		rack_validate_fo_sendwin_up(tp, rack);
10888 		tp->snd_wl1 = th->th_seq;
10889 		if (tp->snd_wnd > tp->max_sndwnd)
10890 			tp->max_sndwnd = tp->snd_wnd;
10891 	}
10892 	/* Do we exit persists? */
10893 	if ((rack->rc_in_persist != 0) &&
10894 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10895 			       rack->r_ctl.rc_pace_min_segs))) {
10896 		rack_exit_persist(tp, rack, cts);
10897 	}
10898 	/* Do we enter persists? */
10899 	if ((rack->rc_in_persist == 0) &&
10900 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10901 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10902 	    (tp->snd_max == tp->snd_una) &&
10903 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10904 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10905 		/*
10906 		 * Here the rwnd is less than
10907 		 * the pacing size, we are established,
10908 		 * nothing is outstanding, and there is
10909 		 * data to send. Enter persists.
10910 		 */
10911 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10912 	}
10913 	/*
10914 	 * If last ACK falls within this segment's sequence numbers, record
10915 	 * the timestamp. NOTE that the test is modified according to the
10916 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10917 	 */
10918 	if ((to->to_flags & TOF_TS) != 0 &&
10919 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10920 		tp->ts_recent_age = tcp_ts_getticks();
10921 		tp->ts_recent = to->to_tsval;
10922 	}
10923 	/*
10924 	 * This is a pure ack for outstanding data.
10925 	 */
10926 	KMOD_TCPSTAT_INC(tcps_predack);
10927 
10928 	/*
10929 	 * "bad retransmit" recovery.
10930 	 */
10931 	if ((tp->t_flags & TF_PREVVALID) &&
10932 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10933 		tp->t_flags &= ~TF_PREVVALID;
10934 		if (tp->t_rxtshift == 1 &&
10935 		    (int)(ticks - tp->t_badrxtwin) < 0)
10936 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10937 	}
10938 	/*
10939 	 * Recalculate the transmit timer / rtt.
10940 	 *
10941 	 * Some boxes send broken timestamp replies during the SYN+ACK
10942 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10943 	 * and blow up the retransmit timer.
10944 	 */
10945 	acked = BYTES_THIS_ACK(tp, th);
10946 
10947 #ifdef TCP_HHOOK
10948 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10949 	hhook_run_tcp_est_in(tp, th, to);
10950 #endif
10951 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10952 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10953 	if (acked) {
10954 		struct mbuf *mfree;
10955 
10956 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
10957 		SOCKBUF_LOCK(&so->so_snd);
10958 		mfree = sbcut_locked(&so->so_snd, acked);
10959 		tp->snd_una = th->th_ack;
10960 		/* Note we want to hold the sb lock through the sendmap adjust */
10961 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10962 		/* Wake up the socket if we have room to write more */
10963 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10964 		sowwakeup_locked(so);
10965 		m_freem(mfree);
10966 		tp->t_rxtshift = 0;
10967 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10968 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10969 		rack->rc_tlp_in_progress = 0;
10970 		rack->r_ctl.rc_tlp_cnt_out = 0;
10971 		/*
10972 		 * If it is the RXT timer we want to
10973 		 * stop it, so we can restart a TLP.
10974 		 */
10975 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10976 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10977 #ifdef NETFLIX_HTTP_LOGGING
10978 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10979 #endif
10980 	}
10981 	/*
10982 	 * Let the congestion control algorithm update congestion control
10983 	 * related information. This typically means increasing the
10984 	 * congestion window.
10985 	 */
10986 	if (tp->snd_wnd < ctf_outstanding(tp)) {
10987 		/* The peer collapsed the window */
10988 		rack_collapsed_window(rack);
10989 	} else if (rack->rc_has_collapsed)
10990 		rack_un_collapse_window(rack);
10991 
10992 	/*
10993 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
10994 	 */
10995 	tp->snd_wl2 = th->th_ack;
10996 	tp->t_dupacks = 0;
10997 	m_freem(m);
10998 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
10999 
11000 	/*
11001 	 * If all outstanding data are acked, stop retransmit timer,
11002 	 * otherwise restart timer using current (possibly backed-off)
11003 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
11004 	 * If data are ready to send, let tcp_output decide between more
11005 	 * output or persist.
11006 	 */
11007 #ifdef TCPDEBUG
11008 	if (so->so_options & SO_DEBUG)
11009 		tcp_trace(TA_INPUT, ostate, tp,
11010 		    (void *)tcp_saveipgen,
11011 		    &tcp_savetcp, 0);
11012 #endif
11013 	if (under_pacing &&
11014 	    (rack->use_fixed_rate == 0) &&
11015 	    (rack->in_probe_rtt == 0) &&
11016 	    rack->rc_gp_dyn_mul &&
11017 	    rack->rc_always_pace) {
11018 		/* Check if we are dragging bottom */
11019 		rack_check_bottom_drag(tp, rack, so, acked);
11020 	}
11021 	if (tp->snd_una == tp->snd_max) {
11022 		tp->t_flags &= ~TF_PREVVALID;
11023 		rack->r_ctl.retran_during_recovery = 0;
11024 		rack->r_ctl.dsack_byte_cnt = 0;
11025 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11026 		if (rack->r_ctl.rc_went_idle_time == 0)
11027 			rack->r_ctl.rc_went_idle_time = 1;
11028 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11029 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
11030 			tp->t_acktime = 0;
11031 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11032 	}
11033 	if (acked && rack->r_fast_output)
11034 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11035 	if (sbavail(&so->so_snd)) {
11036 		rack->r_wanted_output = 1;
11037 	}
11038 	return (1);
11039 }
11040 
11041 /*
11042  * Return value of 1, the TCB is unlocked and most
11043  * likely gone, return value of 0, the TCP is still
11044  * locked.
11045  */
11046 static int
11047 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11048     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11049     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11050 {
11051 	int32_t ret_val = 0;
11052 	int32_t todrop;
11053 	int32_t ourfinisacked = 0;
11054 	struct tcp_rack *rack;
11055 
11056 	ctf_calc_rwin(so, tp);
11057 	/*
11058 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
11059 	 * SYN, drop the input. if seg contains a RST, then drop the
11060 	 * connection. if seg does not contain SYN, then drop it. Otherwise
11061 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
11062 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
11063 	 * contains an ECE and ECN support is enabled, the stream is ECN
11064 	 * capable. if SYN has been acked change to ESTABLISHED else
11065 	 * SYN_RCVD state arrange for segment to be acked (eventually)
11066 	 * continue processing rest of data/controls.
11067 	 */
11068 	if ((thflags & TH_ACK) &&
11069 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
11070 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11071 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11072 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11073 		return (1);
11074 	}
11075 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11076 		TCP_PROBE5(connect__refused, NULL, tp,
11077 		    mtod(m, const char *), tp, th);
11078 		tp = tcp_drop(tp, ECONNREFUSED);
11079 		ctf_do_drop(m, tp);
11080 		return (1);
11081 	}
11082 	if (thflags & TH_RST) {
11083 		ctf_do_drop(m, tp);
11084 		return (1);
11085 	}
11086 	if (!(thflags & TH_SYN)) {
11087 		ctf_do_drop(m, tp);
11088 		return (1);
11089 	}
11090 	tp->irs = th->th_seq;
11091 	tcp_rcvseqinit(tp);
11092 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11093 	if (thflags & TH_ACK) {
11094 		int tfo_partial = 0;
11095 
11096 		KMOD_TCPSTAT_INC(tcps_connects);
11097 		soisconnected(so);
11098 #ifdef MAC
11099 		mac_socketpeer_set_from_mbuf(m, so);
11100 #endif
11101 		/* Do window scaling on this connection? */
11102 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11103 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11104 			tp->rcv_scale = tp->request_r_scale;
11105 		}
11106 		tp->rcv_adv += min(tp->rcv_wnd,
11107 		    TCP_MAXWIN << tp->rcv_scale);
11108 		/*
11109 		 * If not all the data that was sent in the TFO SYN
11110 		 * has been acked, resend the remainder right away.
11111 		 */
11112 		if (IS_FASTOPEN(tp->t_flags) &&
11113 		    (tp->snd_una != tp->snd_max)) {
11114 			tp->snd_nxt = th->th_ack;
11115 			tfo_partial = 1;
11116 		}
11117 		/*
11118 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
11119 		 * will be turned on later.
11120 		 */
11121 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11122 			rack_timer_cancel(tp, rack,
11123 					  rack->r_ctl.rc_rcvtime, __LINE__);
11124 			tp->t_flags |= TF_DELACK;
11125 		} else {
11126 			rack->r_wanted_output = 1;
11127 			tp->t_flags |= TF_ACKNOW;
11128 			rack->rc_dack_toggle = 0;
11129 		}
11130 
11131 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
11132 
11133 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
11134 			/*
11135 			 * We advance snd_una for the
11136 			 * fast open case. If th_ack is
11137 			 * acknowledging data beyond
11138 			 * snd_una we can't just call
11139 			 * ack-processing since the
11140 			 * data stream in our send-map
11141 			 * will start at snd_una + 1 (one
11142 			 * beyond the SYN). If its just
11143 			 * equal we don't need to do that
11144 			 * and there is no send_map.
11145 			 */
11146 			tp->snd_una++;
11147 		}
11148 		/*
11149 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11150 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11151 		 */
11152 		tp->t_starttime = ticks;
11153 		if (tp->t_flags & TF_NEEDFIN) {
11154 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
11155 			tp->t_flags &= ~TF_NEEDFIN;
11156 			thflags &= ~TH_SYN;
11157 		} else {
11158 			tcp_state_change(tp, TCPS_ESTABLISHED);
11159 			TCP_PROBE5(connect__established, NULL, tp,
11160 			    mtod(m, const char *), tp, th);
11161 			rack_cc_conn_init(tp);
11162 		}
11163 	} else {
11164 		/*
11165 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
11166 		 * open.  If segment contains CC option and there is a
11167 		 * cached CC, apply TAO test. If it succeeds, connection is *
11168 		 * half-synchronized. Otherwise, do 3-way handshake:
11169 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11170 		 * there was no CC option, clear cached CC value.
11171 		 */
11172 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
11173 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
11174 	}
11175 	INP_WLOCK_ASSERT(tp->t_inpcb);
11176 	/*
11177 	 * Advance th->th_seq to correspond to first data byte. If data,
11178 	 * trim to stay within window, dropping FIN if necessary.
11179 	 */
11180 	th->th_seq++;
11181 	if (tlen > tp->rcv_wnd) {
11182 		todrop = tlen - tp->rcv_wnd;
11183 		m_adj(m, -todrop);
11184 		tlen = tp->rcv_wnd;
11185 		thflags &= ~TH_FIN;
11186 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11187 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11188 	}
11189 	tp->snd_wl1 = th->th_seq - 1;
11190 	tp->rcv_up = th->th_seq;
11191 	/*
11192 	 * Client side of transaction: already sent SYN and data. If the
11193 	 * remote host used T/TCP to validate the SYN, our data will be
11194 	 * ACK'd; if so, enter normal data segment processing in the middle
11195 	 * of step 5, ack processing. Otherwise, goto step 6.
11196 	 */
11197 	if (thflags & TH_ACK) {
11198 		/* For syn-sent we need to possibly update the rtt */
11199 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11200 			uint32_t t, mcts;
11201 
11202 			mcts = tcp_ts_getticks();
11203 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11204 			if (!tp->t_rttlow || tp->t_rttlow > t)
11205 				tp->t_rttlow = t;
11206 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11207 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11208 			tcp_rack_xmit_timer_commit(rack, tp);
11209 		}
11210 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11211 			return (ret_val);
11212 		/* We may have changed to FIN_WAIT_1 above */
11213 		if (tp->t_state == TCPS_FIN_WAIT_1) {
11214 			/*
11215 			 * In FIN_WAIT_1 STATE in addition to the processing
11216 			 * for the ESTABLISHED state if our FIN is now
11217 			 * acknowledged then enter FIN_WAIT_2.
11218 			 */
11219 			if (ourfinisacked) {
11220 				/*
11221 				 * If we can't receive any more data, then
11222 				 * closing user can proceed. Starting the
11223 				 * timer is contrary to the specification,
11224 				 * but if we don't get a FIN we'll hang
11225 				 * forever.
11226 				 *
11227 				 * XXXjl: we should release the tp also, and
11228 				 * use a compressed state.
11229 				 */
11230 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11231 					soisdisconnected(so);
11232 					tcp_timer_activate(tp, TT_2MSL,
11233 					    (tcp_fast_finwait2_recycle ?
11234 					    tcp_finwait2_timeout :
11235 					    TP_MAXIDLE(tp)));
11236 				}
11237 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
11238 			}
11239 		}
11240 	}
11241 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11242 	   tiwin, thflags, nxt_pkt));
11243 }
11244 
11245 /*
11246  * Return value of 1, the TCB is unlocked and most
11247  * likely gone, return value of 0, the TCP is still
11248  * locked.
11249  */
11250 static int
11251 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11252     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11253     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11254 {
11255 	struct tcp_rack *rack;
11256 	int32_t ret_val = 0;
11257 	int32_t ourfinisacked = 0;
11258 
11259 	ctf_calc_rwin(so, tp);
11260 	if ((thflags & TH_ACK) &&
11261 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11262 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11263 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11264 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11265 		return (1);
11266 	}
11267 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11268 	if (IS_FASTOPEN(tp->t_flags)) {
11269 		/*
11270 		 * When a TFO connection is in SYN_RECEIVED, the
11271 		 * only valid packets are the initial SYN, a
11272 		 * retransmit/copy of the initial SYN (possibly with
11273 		 * a subset of the original data), a valid ACK, a
11274 		 * FIN, or a RST.
11275 		 */
11276 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11277 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11278 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11279 			return (1);
11280 		} else if (thflags & TH_SYN) {
11281 			/* non-initial SYN is ignored */
11282 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11283 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11284 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11285 				ctf_do_drop(m, NULL);
11286 				return (0);
11287 			}
11288 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11289 			ctf_do_drop(m, NULL);
11290 			return (0);
11291 		}
11292 	}
11293 
11294 	if ((thflags & TH_RST) ||
11295 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11296 		return (__ctf_process_rst(m, th, so, tp,
11297 					  &rack->r_ctl.challenge_ack_ts,
11298 					  &rack->r_ctl.challenge_ack_cnt));
11299 	/*
11300 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11301 	 * it's less than ts_recent, drop it.
11302 	 */
11303 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11304 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11305 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11306 			return (ret_val);
11307 	}
11308 	/*
11309 	 * In the SYN-RECEIVED state, validate that the packet belongs to
11310 	 * this connection before trimming the data to fit the receive
11311 	 * window.  Check the sequence number versus IRS since we know the
11312 	 * sequence numbers haven't wrapped.  This is a partial fix for the
11313 	 * "LAND" DoS attack.
11314 	 */
11315 	if (SEQ_LT(th->th_seq, tp->irs)) {
11316 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11317 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11318 		return (1);
11319 	}
11320 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11321 			      &rack->r_ctl.challenge_ack_ts,
11322 			      &rack->r_ctl.challenge_ack_cnt)) {
11323 		return (ret_val);
11324 	}
11325 	/*
11326 	 * If last ACK falls within this segment's sequence numbers, record
11327 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11328 	 * from the latest proposal of the tcplw@cray.com list (Braden
11329 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11330 	 * with our earlier PAWS tests, so this check should be solely
11331 	 * predicated on the sequence space of this segment. 3) That we
11332 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11333 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11334 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11335 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11336 	 * p.869. In such cases, we can still calculate the RTT correctly
11337 	 * when RCV.NXT == Last.ACK.Sent.
11338 	 */
11339 	if ((to->to_flags & TOF_TS) != 0 &&
11340 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11341 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11342 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11343 		tp->ts_recent_age = tcp_ts_getticks();
11344 		tp->ts_recent = to->to_tsval;
11345 	}
11346 	tp->snd_wnd = tiwin;
11347 	rack_validate_fo_sendwin_up(tp, rack);
11348 	/*
11349 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11350 	 * is on (half-synchronized state), then queue data for later
11351 	 * processing; else drop segment and return.
11352 	 */
11353 	if ((thflags & TH_ACK) == 0) {
11354 		if (IS_FASTOPEN(tp->t_flags)) {
11355 			rack_cc_conn_init(tp);
11356 		}
11357 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11358 		    tiwin, thflags, nxt_pkt));
11359 	}
11360 	KMOD_TCPSTAT_INC(tcps_connects);
11361 	soisconnected(so);
11362 	/* Do window scaling? */
11363 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11364 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11365 		tp->rcv_scale = tp->request_r_scale;
11366 	}
11367 	/*
11368 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11369 	 * FIN-WAIT-1
11370 	 */
11371 	tp->t_starttime = ticks;
11372 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11373 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11374 		tp->t_tfo_pending = NULL;
11375 	}
11376 	if (tp->t_flags & TF_NEEDFIN) {
11377 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
11378 		tp->t_flags &= ~TF_NEEDFIN;
11379 	} else {
11380 		tcp_state_change(tp, TCPS_ESTABLISHED);
11381 		TCP_PROBE5(accept__established, NULL, tp,
11382 		    mtod(m, const char *), tp, th);
11383 		/*
11384 		 * TFO connections call cc_conn_init() during SYN
11385 		 * processing.  Calling it again here for such connections
11386 		 * is not harmless as it would undo the snd_cwnd reduction
11387 		 * that occurs when a TFO SYN|ACK is retransmitted.
11388 		 */
11389 		if (!IS_FASTOPEN(tp->t_flags))
11390 			rack_cc_conn_init(tp);
11391 	}
11392 	/*
11393 	 * Account for the ACK of our SYN prior to
11394 	 * regular ACK processing below, except for
11395 	 * simultaneous SYN, which is handled later.
11396 	 */
11397 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11398 		tp->snd_una++;
11399 	/*
11400 	 * If segment contains data or ACK, will call tcp_reass() later; if
11401 	 * not, do so now to pass queued data to user.
11402 	 */
11403 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
11404 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11405 		    (struct mbuf *)0);
11406 		if (tp->t_flags & TF_WAKESOR) {
11407 			tp->t_flags &= ~TF_WAKESOR;
11408 			/* NB: sorwakeup_locked() does an implicit unlock. */
11409 			sorwakeup_locked(so);
11410 		}
11411 	}
11412 	tp->snd_wl1 = th->th_seq - 1;
11413 	/* For syn-recv we need to possibly update the rtt */
11414 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11415 		uint32_t t, mcts;
11416 
11417 		mcts = tcp_ts_getticks();
11418 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11419 		if (!tp->t_rttlow || tp->t_rttlow > t)
11420 			tp->t_rttlow = t;
11421 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11422 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11423 		tcp_rack_xmit_timer_commit(rack, tp);
11424 	}
11425 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11426 		return (ret_val);
11427 	}
11428 	if (tp->t_state == TCPS_FIN_WAIT_1) {
11429 		/* We could have went to FIN_WAIT_1 (or EST) above */
11430 		/*
11431 		 * In FIN_WAIT_1 STATE in addition to the processing for the
11432 		 * ESTABLISHED state if our FIN is now acknowledged then
11433 		 * enter FIN_WAIT_2.
11434 		 */
11435 		if (ourfinisacked) {
11436 			/*
11437 			 * If we can't receive any more data, then closing
11438 			 * user can proceed. Starting the timer is contrary
11439 			 * to the specification, but if we don't get a FIN
11440 			 * we'll hang forever.
11441 			 *
11442 			 * XXXjl: we should release the tp also, and use a
11443 			 * compressed state.
11444 			 */
11445 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11446 				soisdisconnected(so);
11447 				tcp_timer_activate(tp, TT_2MSL,
11448 				    (tcp_fast_finwait2_recycle ?
11449 				    tcp_finwait2_timeout :
11450 				    TP_MAXIDLE(tp)));
11451 			}
11452 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
11453 		}
11454 	}
11455 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11456 	    tiwin, thflags, nxt_pkt));
11457 }
11458 
11459 /*
11460  * Return value of 1, the TCB is unlocked and most
11461  * likely gone, return value of 0, the TCP is still
11462  * locked.
11463  */
11464 static int
11465 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11466     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11467     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11468 {
11469 	int32_t ret_val = 0;
11470 	struct tcp_rack *rack;
11471 
11472 	/*
11473 	 * Header prediction: check for the two common cases of a
11474 	 * uni-directional data xfer.  If the packet has no control flags,
11475 	 * is in-sequence, the window didn't change and we're not
11476 	 * retransmitting, it's a candidate.  If the length is zero and the
11477 	 * ack moved forward, we're the sender side of the xfer.  Just free
11478 	 * the data acked & wake any higher level process that was blocked
11479 	 * waiting for space.  If the length is non-zero and the ack didn't
11480 	 * move, we're the receiver side.  If we're getting packets in-order
11481 	 * (the reassembly queue is empty), add the data toc The socket
11482 	 * buffer and note that we need a delayed ack. Make sure that the
11483 	 * hidden state-flags are also off. Since we check for
11484 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11485 	 */
11486 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11487 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11488 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11489 	    __predict_true(SEGQ_EMPTY(tp)) &&
11490 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
11491 		if (tlen == 0) {
11492 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11493 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11494 				return (0);
11495 			}
11496 		} else {
11497 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11498 			    tiwin, nxt_pkt, iptos)) {
11499 				return (0);
11500 			}
11501 		}
11502 	}
11503 	ctf_calc_rwin(so, tp);
11504 
11505 	if ((thflags & TH_RST) ||
11506 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11507 		return (__ctf_process_rst(m, th, so, tp,
11508 					  &rack->r_ctl.challenge_ack_ts,
11509 					  &rack->r_ctl.challenge_ack_cnt));
11510 
11511 	/*
11512 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11513 	 * synchronized state.
11514 	 */
11515 	if (thflags & TH_SYN) {
11516 		ctf_challenge_ack(m, th, tp, &ret_val);
11517 		return (ret_val);
11518 	}
11519 	/*
11520 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11521 	 * it's less than ts_recent, drop it.
11522 	 */
11523 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11524 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11525 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11526 			return (ret_val);
11527 	}
11528 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11529 			      &rack->r_ctl.challenge_ack_ts,
11530 			      &rack->r_ctl.challenge_ack_cnt)) {
11531 		return (ret_val);
11532 	}
11533 	/*
11534 	 * If last ACK falls within this segment's sequence numbers, record
11535 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11536 	 * from the latest proposal of the tcplw@cray.com list (Braden
11537 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11538 	 * with our earlier PAWS tests, so this check should be solely
11539 	 * predicated on the sequence space of this segment. 3) That we
11540 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11541 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11542 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11543 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11544 	 * p.869. In such cases, we can still calculate the RTT correctly
11545 	 * when RCV.NXT == Last.ACK.Sent.
11546 	 */
11547 	if ((to->to_flags & TOF_TS) != 0 &&
11548 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11549 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11550 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11551 		tp->ts_recent_age = tcp_ts_getticks();
11552 		tp->ts_recent = to->to_tsval;
11553 	}
11554 	/*
11555 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11556 	 * is on (half-synchronized state), then queue data for later
11557 	 * processing; else drop segment and return.
11558 	 */
11559 	if ((thflags & TH_ACK) == 0) {
11560 		if (tp->t_flags & TF_NEEDSYN) {
11561 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11562 			    tiwin, thflags, nxt_pkt));
11563 
11564 		} else if (tp->t_flags & TF_ACKNOW) {
11565 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11566 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11567 			return (ret_val);
11568 		} else {
11569 			ctf_do_drop(m, NULL);
11570 			return (0);
11571 		}
11572 	}
11573 	/*
11574 	 * Ack processing.
11575 	 */
11576 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11577 		return (ret_val);
11578 	}
11579 	if (sbavail(&so->so_snd)) {
11580 		if (ctf_progress_timeout_check(tp, true)) {
11581 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11582 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11583 			return (1);
11584 		}
11585 	}
11586 	/* State changes only happen in rack_process_data() */
11587 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11588 	    tiwin, thflags, nxt_pkt));
11589 }
11590 
11591 /*
11592  * Return value of 1, the TCB is unlocked and most
11593  * likely gone, return value of 0, the TCP is still
11594  * locked.
11595  */
11596 static int
11597 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11598     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11599     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11600 {
11601 	int32_t ret_val = 0;
11602 	struct tcp_rack *rack;
11603 
11604 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11605 	ctf_calc_rwin(so, tp);
11606 	if ((thflags & TH_RST) ||
11607 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11608 		return (__ctf_process_rst(m, th, so, tp,
11609 					  &rack->r_ctl.challenge_ack_ts,
11610 					  &rack->r_ctl.challenge_ack_cnt));
11611 	/*
11612 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11613 	 * synchronized state.
11614 	 */
11615 	if (thflags & TH_SYN) {
11616 		ctf_challenge_ack(m, th, tp, &ret_val);
11617 		return (ret_val);
11618 	}
11619 	/*
11620 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11621 	 * it's less than ts_recent, drop it.
11622 	 */
11623 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11624 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11625 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11626 			return (ret_val);
11627 	}
11628 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11629 			      &rack->r_ctl.challenge_ack_ts,
11630 			      &rack->r_ctl.challenge_ack_cnt)) {
11631 		return (ret_val);
11632 	}
11633 	/*
11634 	 * If last ACK falls within this segment's sequence numbers, record
11635 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11636 	 * from the latest proposal of the tcplw@cray.com list (Braden
11637 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11638 	 * with our earlier PAWS tests, so this check should be solely
11639 	 * predicated on the sequence space of this segment. 3) That we
11640 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11641 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11642 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11643 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11644 	 * p.869. In such cases, we can still calculate the RTT correctly
11645 	 * when RCV.NXT == Last.ACK.Sent.
11646 	 */
11647 	if ((to->to_flags & TOF_TS) != 0 &&
11648 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11649 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11650 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11651 		tp->ts_recent_age = tcp_ts_getticks();
11652 		tp->ts_recent = to->to_tsval;
11653 	}
11654 	/*
11655 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11656 	 * is on (half-synchronized state), then queue data for later
11657 	 * processing; else drop segment and return.
11658 	 */
11659 	if ((thflags & TH_ACK) == 0) {
11660 		if (tp->t_flags & TF_NEEDSYN) {
11661 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11662 			    tiwin, thflags, nxt_pkt));
11663 
11664 		} else if (tp->t_flags & TF_ACKNOW) {
11665 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11666 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11667 			return (ret_val);
11668 		} else {
11669 			ctf_do_drop(m, NULL);
11670 			return (0);
11671 		}
11672 	}
11673 	/*
11674 	 * Ack processing.
11675 	 */
11676 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11677 		return (ret_val);
11678 	}
11679 	if (sbavail(&so->so_snd)) {
11680 		if (ctf_progress_timeout_check(tp, true)) {
11681 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11682 						tp, tick, PROGRESS_DROP, __LINE__);
11683 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11684 			return (1);
11685 		}
11686 	}
11687 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11688 	    tiwin, thflags, nxt_pkt));
11689 }
11690 
11691 static int
11692 rack_check_data_after_close(struct mbuf *m,
11693     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11694 {
11695 	struct tcp_rack *rack;
11696 
11697 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11698 	if (rack->rc_allow_data_af_clo == 0) {
11699 	close_now:
11700 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11701 		/* tcp_close will kill the inp pre-log the Reset */
11702 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11703 		tp = tcp_close(tp);
11704 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11705 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11706 		return (1);
11707 	}
11708 	if (sbavail(&so->so_snd) == 0)
11709 		goto close_now;
11710 	/* Ok we allow data that is ignored and a followup reset */
11711 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11712 	tp->rcv_nxt = th->th_seq + *tlen;
11713 	tp->t_flags2 |= TF2_DROP_AF_DATA;
11714 	rack->r_wanted_output = 1;
11715 	*tlen = 0;
11716 	return (0);
11717 }
11718 
11719 /*
11720  * Return value of 1, the TCB is unlocked and most
11721  * likely gone, return value of 0, the TCP is still
11722  * locked.
11723  */
11724 static int
11725 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11726     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11727     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11728 {
11729 	int32_t ret_val = 0;
11730 	int32_t ourfinisacked = 0;
11731 	struct tcp_rack *rack;
11732 
11733 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11734 	ctf_calc_rwin(so, tp);
11735 
11736 	if ((thflags & TH_RST) ||
11737 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11738 		return (__ctf_process_rst(m, th, so, tp,
11739 					  &rack->r_ctl.challenge_ack_ts,
11740 					  &rack->r_ctl.challenge_ack_cnt));
11741 	/*
11742 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11743 	 * synchronized state.
11744 	 */
11745 	if (thflags & TH_SYN) {
11746 		ctf_challenge_ack(m, th, tp, &ret_val);
11747 		return (ret_val);
11748 	}
11749 	/*
11750 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11751 	 * it's less than ts_recent, drop it.
11752 	 */
11753 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11754 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11755 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11756 			return (ret_val);
11757 	}
11758 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11759 			      &rack->r_ctl.challenge_ack_ts,
11760 			      &rack->r_ctl.challenge_ack_cnt)) {
11761 		return (ret_val);
11762 	}
11763 	/*
11764 	 * If new data are received on a connection after the user processes
11765 	 * are gone, then RST the other end.
11766 	 */
11767 	if ((so->so_state & SS_NOFDREF) && tlen) {
11768 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
11769 			return (1);
11770 	}
11771 	/*
11772 	 * If last ACK falls within this segment's sequence numbers, record
11773 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11774 	 * from the latest proposal of the tcplw@cray.com list (Braden
11775 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11776 	 * with our earlier PAWS tests, so this check should be solely
11777 	 * predicated on the sequence space of this segment. 3) That we
11778 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11779 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11780 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11781 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11782 	 * p.869. In such cases, we can still calculate the RTT correctly
11783 	 * when RCV.NXT == Last.ACK.Sent.
11784 	 */
11785 	if ((to->to_flags & TOF_TS) != 0 &&
11786 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11787 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11788 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11789 		tp->ts_recent_age = tcp_ts_getticks();
11790 		tp->ts_recent = to->to_tsval;
11791 	}
11792 	/*
11793 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11794 	 * is on (half-synchronized state), then queue data for later
11795 	 * processing; else drop segment and return.
11796 	 */
11797 	if ((thflags & TH_ACK) == 0) {
11798 		if (tp->t_flags & TF_NEEDSYN) {
11799 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11800 			    tiwin, thflags, nxt_pkt));
11801 		} else if (tp->t_flags & TF_ACKNOW) {
11802 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11803 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11804 			return (ret_val);
11805 		} else {
11806 			ctf_do_drop(m, NULL);
11807 			return (0);
11808 		}
11809 	}
11810 	/*
11811 	 * Ack processing.
11812 	 */
11813 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11814 		return (ret_val);
11815 	}
11816 	if (ourfinisacked) {
11817 		/*
11818 		 * If we can't receive any more data, then closing user can
11819 		 * proceed. Starting the timer is contrary to the
11820 		 * specification, but if we don't get a FIN we'll hang
11821 		 * forever.
11822 		 *
11823 		 * XXXjl: we should release the tp also, and use a
11824 		 * compressed state.
11825 		 */
11826 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11827 			soisdisconnected(so);
11828 			tcp_timer_activate(tp, TT_2MSL,
11829 			    (tcp_fast_finwait2_recycle ?
11830 			    tcp_finwait2_timeout :
11831 			    TP_MAXIDLE(tp)));
11832 		}
11833 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
11834 	}
11835 	if (sbavail(&so->so_snd)) {
11836 		if (ctf_progress_timeout_check(tp, true)) {
11837 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11838 						tp, tick, PROGRESS_DROP, __LINE__);
11839 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11840 			return (1);
11841 		}
11842 	}
11843 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11844 	    tiwin, thflags, nxt_pkt));
11845 }
11846 
11847 /*
11848  * Return value of 1, the TCB is unlocked and most
11849  * likely gone, return value of 0, the TCP is still
11850  * locked.
11851  */
11852 static int
11853 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11854     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11855     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11856 {
11857 	int32_t ret_val = 0;
11858 	int32_t ourfinisacked = 0;
11859 	struct tcp_rack *rack;
11860 
11861 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11862 	ctf_calc_rwin(so, tp);
11863 
11864 	if ((thflags & TH_RST) ||
11865 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11866 		return (__ctf_process_rst(m, th, so, tp,
11867 					  &rack->r_ctl.challenge_ack_ts,
11868 					  &rack->r_ctl.challenge_ack_cnt));
11869 	/*
11870 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11871 	 * synchronized state.
11872 	 */
11873 	if (thflags & TH_SYN) {
11874 		ctf_challenge_ack(m, th, tp, &ret_val);
11875 		return (ret_val);
11876 	}
11877 	/*
11878 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11879 	 * it's less than ts_recent, drop it.
11880 	 */
11881 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11882 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11883 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11884 			return (ret_val);
11885 	}
11886 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11887 			      &rack->r_ctl.challenge_ack_ts,
11888 			      &rack->r_ctl.challenge_ack_cnt)) {
11889 		return (ret_val);
11890 	}
11891 	/*
11892 	 * If new data are received on a connection after the user processes
11893 	 * are gone, then RST the other end.
11894 	 */
11895 	if ((so->so_state & SS_NOFDREF) && tlen) {
11896 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
11897 			return (1);
11898 	}
11899 	/*
11900 	 * If last ACK falls within this segment's sequence numbers, record
11901 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11902 	 * from the latest proposal of the tcplw@cray.com list (Braden
11903 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11904 	 * with our earlier PAWS tests, so this check should be solely
11905 	 * predicated on the sequence space of this segment. 3) That we
11906 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11907 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11908 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11909 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11910 	 * p.869. In such cases, we can still calculate the RTT correctly
11911 	 * when RCV.NXT == Last.ACK.Sent.
11912 	 */
11913 	if ((to->to_flags & TOF_TS) != 0 &&
11914 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11915 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11916 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11917 		tp->ts_recent_age = tcp_ts_getticks();
11918 		tp->ts_recent = to->to_tsval;
11919 	}
11920 	/*
11921 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11922 	 * is on (half-synchronized state), then queue data for later
11923 	 * processing; else drop segment and return.
11924 	 */
11925 	if ((thflags & TH_ACK) == 0) {
11926 		if (tp->t_flags & TF_NEEDSYN) {
11927 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11928 			    tiwin, thflags, nxt_pkt));
11929 		} else if (tp->t_flags & TF_ACKNOW) {
11930 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11931 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11932 			return (ret_val);
11933 		} else {
11934 			ctf_do_drop(m, NULL);
11935 			return (0);
11936 		}
11937 	}
11938 	/*
11939 	 * Ack processing.
11940 	 */
11941 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11942 		return (ret_val);
11943 	}
11944 	if (ourfinisacked) {
11945 		tcp_twstart(tp);
11946 		m_freem(m);
11947 		return (1);
11948 	}
11949 	if (sbavail(&so->so_snd)) {
11950 		if (ctf_progress_timeout_check(tp, true)) {
11951 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11952 						tp, tick, PROGRESS_DROP, __LINE__);
11953 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11954 			return (1);
11955 		}
11956 	}
11957 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11958 	    tiwin, thflags, nxt_pkt));
11959 }
11960 
11961 /*
11962  * Return value of 1, the TCB is unlocked and most
11963  * likely gone, return value of 0, the TCP is still
11964  * locked.
11965  */
11966 static int
11967 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11968     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11969     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11970 {
11971 	int32_t ret_val = 0;
11972 	int32_t ourfinisacked = 0;
11973 	struct tcp_rack *rack;
11974 
11975 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11976 	ctf_calc_rwin(so, tp);
11977 
11978 	if ((thflags & TH_RST) ||
11979 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11980 		return (__ctf_process_rst(m, th, so, tp,
11981 					  &rack->r_ctl.challenge_ack_ts,
11982 					  &rack->r_ctl.challenge_ack_cnt));
11983 	/*
11984 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11985 	 * synchronized state.
11986 	 */
11987 	if (thflags & TH_SYN) {
11988 		ctf_challenge_ack(m, th, tp, &ret_val);
11989 		return (ret_val);
11990 	}
11991 	/*
11992 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11993 	 * it's less than ts_recent, drop it.
11994 	 */
11995 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11996 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11997 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11998 			return (ret_val);
11999 	}
12000 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12001 			      &rack->r_ctl.challenge_ack_ts,
12002 			      &rack->r_ctl.challenge_ack_cnt)) {
12003 		return (ret_val);
12004 	}
12005 	/*
12006 	 * If new data are received on a connection after the user processes
12007 	 * are gone, then RST the other end.
12008 	 */
12009 	if ((so->so_state & SS_NOFDREF) && tlen) {
12010 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12011 			return (1);
12012 	}
12013 	/*
12014 	 * If last ACK falls within this segment's sequence numbers, record
12015 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12016 	 * from the latest proposal of the tcplw@cray.com list (Braden
12017 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12018 	 * with our earlier PAWS tests, so this check should be solely
12019 	 * predicated on the sequence space of this segment. 3) That we
12020 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12021 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12022 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12023 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12024 	 * p.869. In such cases, we can still calculate the RTT correctly
12025 	 * when RCV.NXT == Last.ACK.Sent.
12026 	 */
12027 	if ((to->to_flags & TOF_TS) != 0 &&
12028 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12029 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12030 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12031 		tp->ts_recent_age = tcp_ts_getticks();
12032 		tp->ts_recent = to->to_tsval;
12033 	}
12034 	/*
12035 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12036 	 * is on (half-synchronized state), then queue data for later
12037 	 * processing; else drop segment and return.
12038 	 */
12039 	if ((thflags & TH_ACK) == 0) {
12040 		if (tp->t_flags & TF_NEEDSYN) {
12041 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12042 			    tiwin, thflags, nxt_pkt));
12043 		} else if (tp->t_flags & TF_ACKNOW) {
12044 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12045 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12046 			return (ret_val);
12047 		} else {
12048 			ctf_do_drop(m, NULL);
12049 			return (0);
12050 		}
12051 	}
12052 	/*
12053 	 * case TCPS_LAST_ACK: Ack processing.
12054 	 */
12055 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12056 		return (ret_val);
12057 	}
12058 	if (ourfinisacked) {
12059 		tp = tcp_close(tp);
12060 		ctf_do_drop(m, tp);
12061 		return (1);
12062 	}
12063 	if (sbavail(&so->so_snd)) {
12064 		if (ctf_progress_timeout_check(tp, true)) {
12065 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12066 						tp, tick, PROGRESS_DROP, __LINE__);
12067 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12068 			return (1);
12069 		}
12070 	}
12071 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12072 	    tiwin, thflags, nxt_pkt));
12073 }
12074 
12075 /*
12076  * Return value of 1, the TCB is unlocked and most
12077  * likely gone, return value of 0, the TCP is still
12078  * locked.
12079  */
12080 static int
12081 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12082     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12083     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12084 {
12085 	int32_t ret_val = 0;
12086 	int32_t ourfinisacked = 0;
12087 	struct tcp_rack *rack;
12088 
12089 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12090 	ctf_calc_rwin(so, tp);
12091 
12092 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
12093 	if ((thflags & TH_RST) ||
12094 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12095 		return (__ctf_process_rst(m, th, so, tp,
12096 					  &rack->r_ctl.challenge_ack_ts,
12097 					  &rack->r_ctl.challenge_ack_cnt));
12098 	/*
12099 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12100 	 * synchronized state.
12101 	 */
12102 	if (thflags & TH_SYN) {
12103 		ctf_challenge_ack(m, th, tp, &ret_val);
12104 		return (ret_val);
12105 	}
12106 	/*
12107 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12108 	 * it's less than ts_recent, drop it.
12109 	 */
12110 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12111 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12112 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12113 			return (ret_val);
12114 	}
12115 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12116 			      &rack->r_ctl.challenge_ack_ts,
12117 			      &rack->r_ctl.challenge_ack_cnt)) {
12118 		return (ret_val);
12119 	}
12120 	/*
12121 	 * If new data are received on a connection after the user processes
12122 	 * are gone, then RST the other end.
12123 	 */
12124 	if ((so->so_state & SS_NOFDREF) &&
12125 	    tlen) {
12126 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12127 			return (1);
12128 	}
12129 	/*
12130 	 * If last ACK falls within this segment's sequence numbers, record
12131 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12132 	 * from the latest proposal of the tcplw@cray.com list (Braden
12133 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12134 	 * with our earlier PAWS tests, so this check should be solely
12135 	 * predicated on the sequence space of this segment. 3) That we
12136 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12137 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12138 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12139 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12140 	 * p.869. In such cases, we can still calculate the RTT correctly
12141 	 * when RCV.NXT == Last.ACK.Sent.
12142 	 */
12143 	if ((to->to_flags & TOF_TS) != 0 &&
12144 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12145 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12146 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12147 		tp->ts_recent_age = tcp_ts_getticks();
12148 		tp->ts_recent = to->to_tsval;
12149 	}
12150 	/*
12151 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12152 	 * is on (half-synchronized state), then queue data for later
12153 	 * processing; else drop segment and return.
12154 	 */
12155 	if ((thflags & TH_ACK) == 0) {
12156 		if (tp->t_flags & TF_NEEDSYN) {
12157 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12158 			    tiwin, thflags, nxt_pkt));
12159 		} else if (tp->t_flags & TF_ACKNOW) {
12160 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12161 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12162 			return (ret_val);
12163 		} else {
12164 			ctf_do_drop(m, NULL);
12165 			return (0);
12166 		}
12167 	}
12168 	/*
12169 	 * Ack processing.
12170 	 */
12171 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12172 		return (ret_val);
12173 	}
12174 	if (sbavail(&so->so_snd)) {
12175 		if (ctf_progress_timeout_check(tp, true)) {
12176 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12177 						tp, tick, PROGRESS_DROP, __LINE__);
12178 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12179 			return (1);
12180 		}
12181 	}
12182 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12183 	    tiwin, thflags, nxt_pkt));
12184 }
12185 
12186 static void inline
12187 rack_clear_rate_sample(struct tcp_rack *rack)
12188 {
12189 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12190 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12191 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12192 }
12193 
12194 static void
12195 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12196 {
12197 	uint64_t bw_est, rate_wanted;
12198 	int chged = 0;
12199 	uint32_t user_max, orig_min, orig_max;
12200 
12201 	orig_min = rack->r_ctl.rc_pace_min_segs;
12202 	orig_max = rack->r_ctl.rc_pace_max_segs;
12203 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12204 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12205 		chged = 1;
12206 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12207 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12208 		if (user_max != rack->r_ctl.rc_pace_max_segs)
12209 			chged = 1;
12210 	}
12211 	if (rack->rc_force_max_seg) {
12212 		rack->r_ctl.rc_pace_max_segs = user_max;
12213 	} else if (rack->use_fixed_rate) {
12214 		bw_est = rack_get_bw(rack);
12215 		if ((rack->r_ctl.crte == NULL) ||
12216 		    (bw_est != rack->r_ctl.crte->rate)) {
12217 			rack->r_ctl.rc_pace_max_segs = user_max;
12218 		} else {
12219 			/* We are pacing right at the hardware rate */
12220 			uint32_t segsiz;
12221 
12222 			segsiz = min(ctf_fixed_maxseg(tp),
12223 				     rack->r_ctl.rc_pace_min_segs);
12224 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12225 				                           tp, bw_est, segsiz, 0,
12226 							   rack->r_ctl.crte, NULL);
12227 		}
12228 	} else if (rack->rc_always_pace) {
12229 		if (rack->r_ctl.gp_bw ||
12230 #ifdef NETFLIX_PEAKRATE
12231 		    rack->rc_tp->t_maxpeakrate ||
12232 #endif
12233 		    rack->r_ctl.init_rate) {
12234 			/* We have a rate of some sort set */
12235 			uint32_t  orig;
12236 
12237 			bw_est = rack_get_bw(rack);
12238 			orig = rack->r_ctl.rc_pace_max_segs;
12239 			if (fill_override)
12240 				rate_wanted = *fill_override;
12241 			else
12242 				rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12243 			if (rate_wanted) {
12244 				/* We have something */
12245 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12246 										   rate_wanted,
12247 										   ctf_fixed_maxseg(rack->rc_tp));
12248 			} else
12249 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12250 			if (orig != rack->r_ctl.rc_pace_max_segs)
12251 				chged = 1;
12252 		} else if ((rack->r_ctl.gp_bw == 0) &&
12253 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
12254 			/*
12255 			 * If we have nothing limit us to bursting
12256 			 * out IW sized pieces.
12257 			 */
12258 			chged = 1;
12259 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12260 		}
12261 	}
12262 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12263 		chged = 1;
12264 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12265 	}
12266 	if (chged)
12267 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12268 }
12269 
12270 
12271 static void
12272 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12273 {
12274 #ifdef INET6
12275 	struct ip6_hdr *ip6 = NULL;
12276 #endif
12277 #ifdef INET
12278 	struct ip *ip = NULL;
12279 #endif
12280 	struct udphdr *udp = NULL;
12281 
12282 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
12283 #ifdef INET6
12284 	if (rack->r_is_v6) {
12285 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12286 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12287 		if (tp->t_port) {
12288 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12289 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12290 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12291 			udp->uh_dport = tp->t_port;
12292 			rack->r_ctl.fsb.udp = udp;
12293 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12294 		} else
12295 		{
12296 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12297 			rack->r_ctl.fsb.udp = NULL;
12298 		}
12299 		tcpip_fillheaders(rack->rc_inp,
12300 				  tp->t_port,
12301 				  ip6, rack->r_ctl.fsb.th);
12302 	} else
12303 #endif				/* INET6 */
12304 	{
12305 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12306 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12307 		if (tp->t_port) {
12308 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12309 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12310 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12311 			udp->uh_dport = tp->t_port;
12312 			rack->r_ctl.fsb.udp = udp;
12313 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12314 		} else
12315 		{
12316 			rack->r_ctl.fsb.udp = NULL;
12317 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12318 		}
12319 		tcpip_fillheaders(rack->rc_inp,
12320 				  tp->t_port,
12321 				  ip, rack->r_ctl.fsb.th);
12322 	}
12323 	rack->r_fsb_inited = 1;
12324 }
12325 
12326 static int
12327 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12328 {
12329 	/*
12330 	 * Allocate the larger of spaces V6 if available else just
12331 	 * V4 and include udphdr (overbook)
12332 	 */
12333 #ifdef INET6
12334 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12335 #else
12336 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12337 #endif
12338 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12339 					    M_TCPFSB, M_NOWAIT|M_ZERO);
12340 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12341 		return (ENOMEM);
12342 	}
12343 	rack->r_fsb_inited = 0;
12344 	return (0);
12345 }
12346 
12347 static int
12348 rack_init(struct tcpcb *tp)
12349 {
12350 	struct tcp_rack *rack = NULL;
12351 #ifdef INVARIANTS
12352 	struct rack_sendmap *insret;
12353 #endif
12354 	uint32_t iwin, snt, us_cts;
12355 	int err;
12356 
12357 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12358 	if (tp->t_fb_ptr == NULL) {
12359 		/*
12360 		 * We need to allocate memory but cant. The INP and INP_INFO
12361 		 * locks and they are recursive (happens during setup. So a
12362 		 * scheme to drop the locks fails :(
12363 		 *
12364 		 */
12365 		return (ENOMEM);
12366 	}
12367 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12368 
12369 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12370 	RB_INIT(&rack->r_ctl.rc_mtree);
12371 	TAILQ_INIT(&rack->r_ctl.rc_free);
12372 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
12373 	rack->rc_tp = tp;
12374 	rack->rc_inp = tp->t_inpcb;
12375 	/* Set the flag */
12376 	rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12377 	/* Probably not needed but lets be sure */
12378 	rack_clear_rate_sample(rack);
12379 	/*
12380 	 * Save off the default values, socket options will poke
12381 	 * at these if pacing is not on or we have not yet
12382 	 * reached where pacing is on (gp_ready/fixed enabled).
12383 	 * When they get set into the CC module (when gp_ready
12384 	 * is enabled or we enable fixed) then we will set these
12385 	 * values into the CC and place in here the old values
12386 	 * so we have a restoral. Then we will set the flag
12387 	 * rc_pacing_cc_set. That way whenever we turn off pacing
12388 	 * or switch off this stack, we will know to go restore
12389 	 * the saved values.
12390 	 */
12391 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12392 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12393 	/* We want abe like behavior as well */
12394 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12395 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12396 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12397 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12398 	rack->r_ctl.roundends = tp->snd_max;
12399 	if (use_rack_rr)
12400 		rack->use_rack_rr = 1;
12401 	if (V_tcp_delack_enabled)
12402 		tp->t_delayed_ack = 1;
12403 	else
12404 		tp->t_delayed_ack = 0;
12405 #ifdef TCP_ACCOUNTING
12406 	if (rack_tcp_accounting) {
12407 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12408 	}
12409 #endif
12410 	if (rack_enable_shared_cwnd)
12411 		rack->rack_enable_scwnd = 1;
12412 	rack->rc_user_set_max_segs = rack_hptsi_segments;
12413 	rack->rc_force_max_seg = 0;
12414 	if (rack_use_imac_dack)
12415 		rack->rc_dack_mode = 1;
12416 	TAILQ_INIT(&rack->r_ctl.opt_list);
12417 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12418 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12419 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12420 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12421 	rack->r_ctl.rc_highest_us_rtt = 0;
12422 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12423 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12424 	if (rack_use_cmp_acks)
12425 		rack->r_use_cmp_ack = 1;
12426 	if (rack_disable_prr)
12427 		rack->rack_no_prr = 1;
12428 	if (rack_gp_no_rec_chg)
12429 		rack->rc_gp_no_rec_chg = 1;
12430 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12431 		rack->rc_always_pace = 1;
12432 		if (rack->use_fixed_rate || rack->gp_ready)
12433 			rack_set_cc_pacing(rack);
12434 	} else
12435 		rack->rc_always_pace = 0;
12436 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12437 		rack->r_mbuf_queue = 1;
12438 	else
12439 		rack->r_mbuf_queue = 0;
12440 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12441 		tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12442 	else
12443 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12444 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
12445 	if (rack_limits_scwnd)
12446 		rack->r_limit_scw = 1;
12447 	else
12448 		rack->r_limit_scw = 0;
12449 	rack->rc_labc = V_tcp_abc_l_var;
12450 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12451 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12452 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12453 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12454 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12455 	rack->r_ctl.rc_min_to = rack_min_to;
12456 	microuptime(&rack->r_ctl.act_rcv_time);
12457 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12458 	rack->rc_init_win = rack_default_init_window;
12459 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12460 	if (rack_hw_up_only)
12461 		rack->r_up_only = 1;
12462 	if (rack_do_dyn_mul) {
12463 		/* When dynamic adjustment is on CA needs to start at 100% */
12464 		rack->rc_gp_dyn_mul = 1;
12465 		if (rack_do_dyn_mul >= 100)
12466 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12467 	} else
12468 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12469 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12470 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12471 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12472 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12473 				rack_probertt_filter_life);
12474 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12475 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12476 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
12477 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12478 	rack->r_ctl.rc_time_probertt_starts = 0;
12479 	if (rack_dsack_std_based & 0x1) {
12480 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12481 		rack->rc_rack_tmr_std_based = 1;
12482 	}
12483 	if (rack_dsack_std_based & 0x2) {
12484 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12485 		rack->rc_rack_use_dsack = 1;
12486 	}
12487 	/* We require at least one measurement, even if the sysctl is 0 */
12488 	if (rack_req_measurements)
12489 		rack->r_ctl.req_measurements = rack_req_measurements;
12490 	else
12491 		rack->r_ctl.req_measurements = 1;
12492 	if (rack_enable_hw_pacing)
12493 		rack->rack_hdw_pace_ena = 1;
12494 	if (rack_hw_rate_caps)
12495 		rack->r_rack_hw_rate_caps = 1;
12496 	/* Do we force on detection? */
12497 #ifdef NETFLIX_EXP_DETECTION
12498 	if (tcp_force_detection)
12499 		rack->do_detection = 1;
12500 	else
12501 #endif
12502 		rack->do_detection = 0;
12503 	if (rack_non_rxt_use_cr)
12504 		rack->rack_rec_nonrxt_use_cr = 1;
12505 	err = rack_init_fsb(tp, rack);
12506 	if (err) {
12507 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12508 		tp->t_fb_ptr = NULL;
12509 		return (err);
12510 	}
12511 	if (tp->snd_una != tp->snd_max) {
12512 		/* Create a send map for the current outstanding data */
12513 		struct rack_sendmap *rsm;
12514 
12515 		rsm = rack_alloc(rack);
12516 		if (rsm == NULL) {
12517 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12518 			tp->t_fb_ptr = NULL;
12519 			return (ENOMEM);
12520 		}
12521 		rsm->r_no_rtt_allowed = 1;
12522 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12523 		rsm->r_rtr_cnt = 1;
12524 		rsm->r_rtr_bytes = 0;
12525 		if (tp->t_flags & TF_SENTFIN) {
12526 			rsm->r_end = tp->snd_max - 1;
12527 			rsm->r_flags |= RACK_HAS_FIN;
12528 		} else {
12529 			rsm->r_end = tp->snd_max;
12530 		}
12531 		if (tp->snd_una == tp->iss) {
12532 			/* The data space is one beyond snd_una */
12533 			rsm->r_flags |= RACK_HAS_SYN;
12534 			rsm->r_start = tp->iss;
12535 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12536 		} else
12537 			rsm->r_start = tp->snd_una;
12538 		rsm->r_dupack = 0;
12539 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12540 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12541 			if (rsm->m)
12542 				rsm->orig_m_len = rsm->m->m_len;
12543 			else
12544 				rsm->orig_m_len = 0;
12545 		} else {
12546 			/*
12547 			 * This can happen if we have a stand-alone FIN or
12548 			 *  SYN.
12549 			 */
12550 			rsm->m = NULL;
12551 			rsm->orig_m_len = 0;
12552 			rsm->soff = 0;
12553 		}
12554 #ifndef INVARIANTS
12555 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12556 #else
12557 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12558 		if (insret != NULL) {
12559 			panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12560 			      insret, rack, rsm);
12561 		}
12562 #endif
12563 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12564 		rsm->r_in_tmap = 1;
12565 	}
12566 	/*
12567 	 * Timers in Rack are kept in microseconds so lets
12568 	 * convert any initial incoming variables
12569 	 * from ticks into usecs. Note that we
12570 	 * also change the values of t_srtt and t_rttvar, if
12571 	 * they are non-zero. They are kept with a 5
12572 	 * bit decimal so we have to carefully convert
12573 	 * these to get the full precision.
12574 	 */
12575 	rack_convert_rtts(tp);
12576 	tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12577 	if (rack_do_hystart) {
12578 		tp->ccv->flags |= CCF_HYSTART_ALLOWED;
12579 		if (rack_do_hystart > 1)
12580 			tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND;
12581 		if (rack_do_hystart > 2)
12582 			tp->ccv->flags |= CCF_HYSTART_CONS_SSTH;
12583 	}
12584 	if (rack_def_profile)
12585 		rack_set_profile(rack, rack_def_profile);
12586 	/* Cancel the GP measurement in progress */
12587 	tp->t_flags &= ~TF_GPUTINPROG;
12588 	if (SEQ_GT(tp->snd_max, tp->iss))
12589 		snt = tp->snd_max - tp->iss;
12590 	else
12591 		snt = 0;
12592 	iwin = rc_init_window(rack);
12593 	if (snt < iwin) {
12594 		/* We are not past the initial window
12595 		 * so we need to make sure cwnd is
12596 		 * correct.
12597 		 */
12598 		if (tp->snd_cwnd < iwin)
12599 			tp->snd_cwnd = iwin;
12600 		/*
12601 		 * If we are within the initial window
12602 		 * we want ssthresh to be unlimited. Setting
12603 		 * it to the rwnd (which the default stack does
12604 		 * and older racks) is not really a good idea
12605 		 * since we want to be in SS and grow both the
12606 		 * cwnd and the rwnd (via dynamic rwnd growth). If
12607 		 * we set it to the rwnd then as the peer grows its
12608 		 * rwnd we will be stuck in CA and never hit SS.
12609 		 *
12610 		 * Its far better to raise it up high (this takes the
12611 		 * risk that there as been a loss already, probably
12612 		 * we should have an indicator in all stacks of loss
12613 		 * but we don't), but considering the normal use this
12614 		 * is a risk worth taking. The consequences of not
12615 		 * hitting SS are far worse than going one more time
12616 		 * into it early on (before we have sent even a IW).
12617 		 * It is highly unlikely that we will have had a loss
12618 		 * before getting the IW out.
12619 		 */
12620 		tp->snd_ssthresh = 0xffffffff;
12621 	}
12622 	rack_stop_all_timers(tp);
12623 	/* Lets setup the fsb block */
12624 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12625 	rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12626 			     __LINE__, RACK_RTTS_INIT);
12627 	return (0);
12628 }
12629 
12630 static int
12631 rack_handoff_ok(struct tcpcb *tp)
12632 {
12633 	if ((tp->t_state == TCPS_CLOSED) ||
12634 	    (tp->t_state == TCPS_LISTEN)) {
12635 		/* Sure no problem though it may not stick */
12636 		return (0);
12637 	}
12638 	if ((tp->t_state == TCPS_SYN_SENT) ||
12639 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
12640 		/*
12641 		 * We really don't know if you support sack,
12642 		 * you have to get to ESTAB or beyond to tell.
12643 		 */
12644 		return (EAGAIN);
12645 	}
12646 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12647 		/*
12648 		 * Rack will only send a FIN after all data is acknowledged.
12649 		 * So in this case we have more data outstanding. We can't
12650 		 * switch stacks until either all data and only the FIN
12651 		 * is left (in which case rack_init() now knows how
12652 		 * to deal with that) <or> all is acknowledged and we
12653 		 * are only left with incoming data, though why you
12654 		 * would want to switch to rack after all data is acknowledged
12655 		 * I have no idea (rrs)!
12656 		 */
12657 		return (EAGAIN);
12658 	}
12659 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12660 		return (0);
12661 	}
12662 	/*
12663 	 * If we reach here we don't do SACK on this connection so we can
12664 	 * never do rack.
12665 	 */
12666 	return (EINVAL);
12667 }
12668 
12669 
12670 static void
12671 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12672 {
12673 	if (tp->t_fb_ptr) {
12674 		struct tcp_rack *rack;
12675 		struct rack_sendmap *rsm, *nrsm;
12676 #ifdef INVARIANTS
12677 		struct rack_sendmap *rm;
12678 #endif
12679 
12680 		rack = (struct tcp_rack *)tp->t_fb_ptr;
12681 		if (tp->t_in_pkt) {
12682 			/*
12683 			 * It is unsafe to process the packets since a
12684 			 * reset may be lurking in them (its rare but it
12685 			 * can occur). If we were to find a RST, then we
12686 			 * would end up dropping the connection and the
12687 			 * INP lock, so when we return the caller (tcp_usrreq)
12688 			 * will blow up when it trys to unlock the inp.
12689 			 */
12690 			struct mbuf *save, *m;
12691 
12692 			m = tp->t_in_pkt;
12693 			tp->t_in_pkt = NULL;
12694 			tp->t_tail_pkt = NULL;
12695 			while (m) {
12696 				save = m->m_nextpkt;
12697 				m->m_nextpkt = NULL;
12698 				m_freem(m);
12699 				m = save;
12700 			}
12701 		}
12702 		tp->t_flags &= ~TF_FORCEDATA;
12703 #ifdef NETFLIX_SHARED_CWND
12704 		if (rack->r_ctl.rc_scw) {
12705 			uint32_t limit;
12706 
12707 			if (rack->r_limit_scw)
12708 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12709 			else
12710 				limit = 0;
12711 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12712 						  rack->r_ctl.rc_scw_index,
12713 						  limit);
12714 			rack->r_ctl.rc_scw = NULL;
12715 		}
12716 #endif
12717 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
12718 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12719 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12720 			rack->r_ctl.fsb.th = NULL;
12721 		}
12722 		/* Convert back to ticks, with  */
12723 		if (tp->t_srtt > 1) {
12724 			uint32_t val, frac;
12725 
12726 			val = USEC_2_TICKS(tp->t_srtt);
12727 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12728 			tp->t_srtt = val << TCP_RTT_SHIFT;
12729 			/*
12730 			 * frac is the fractional part here is left
12731 			 * over from converting to hz and shifting.
12732 			 * We need to convert this to the 5 bit
12733 			 * remainder.
12734 			 */
12735 			if (frac) {
12736 				if (hz == 1000) {
12737 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12738 				} else {
12739 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12740 				}
12741 				tp->t_srtt += frac;
12742 			}
12743 		}
12744 		if (tp->t_rttvar) {
12745 			uint32_t val, frac;
12746 
12747 			val = USEC_2_TICKS(tp->t_rttvar);
12748 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12749 			tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12750 			/*
12751 			 * frac is the fractional part here is left
12752 			 * over from converting to hz and shifting.
12753 			 * We need to convert this to the 5 bit
12754 			 * remainder.
12755 			 */
12756 			if (frac) {
12757 				if (hz == 1000) {
12758 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12759 				} else {
12760 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12761 				}
12762 				tp->t_rttvar += frac;
12763 			}
12764 		}
12765 		tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12766 		tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12767 		if (rack->rc_always_pace) {
12768 			tcp_decrement_paced_conn();
12769 			rack_undo_cc_pacing(rack);
12770 			rack->rc_always_pace = 0;
12771 		}
12772 		/* Clean up any options if they were not applied */
12773 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12774 			struct deferred_opt_list *dol;
12775 
12776 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12777 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12778 			free(dol, M_TCPDO);
12779 		}
12780 		/* rack does not use force data but other stacks may clear it */
12781 		if (rack->r_ctl.crte != NULL) {
12782 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12783 			rack->rack_hdrw_pacing = 0;
12784 			rack->r_ctl.crte = NULL;
12785 		}
12786 #ifdef TCP_BLACKBOX
12787 		tcp_log_flowend(tp);
12788 #endif
12789 		RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12790 #ifndef INVARIANTS
12791 			(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12792 #else
12793 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12794 			if (rm != rsm) {
12795 				panic("At fini, rack:%p rsm:%p rm:%p",
12796 				      rack, rsm, rm);
12797 			}
12798 #endif
12799 			uma_zfree(rack_zone, rsm);
12800 		}
12801 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12802 		while (rsm) {
12803 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12804 			uma_zfree(rack_zone, rsm);
12805 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12806 		}
12807 		rack->rc_free_cnt = 0;
12808 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12809 		tp->t_fb_ptr = NULL;
12810 	}
12811 	if (tp->t_inpcb) {
12812 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12813 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12814 		tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12815 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
12816 		/* Cancel the GP measurement in progress */
12817 		tp->t_flags &= ~TF_GPUTINPROG;
12818 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
12819 	}
12820 	/* Make sure snd_nxt is correctly set */
12821 	tp->snd_nxt = tp->snd_max;
12822 }
12823 
12824 static void
12825 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12826 {
12827 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12828 		rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12829 	}
12830 	switch (tp->t_state) {
12831 	case TCPS_SYN_SENT:
12832 		rack->r_state = TCPS_SYN_SENT;
12833 		rack->r_substate = rack_do_syn_sent;
12834 		break;
12835 	case TCPS_SYN_RECEIVED:
12836 		rack->r_state = TCPS_SYN_RECEIVED;
12837 		rack->r_substate = rack_do_syn_recv;
12838 		break;
12839 	case TCPS_ESTABLISHED:
12840 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12841 		rack->r_state = TCPS_ESTABLISHED;
12842 		rack->r_substate = rack_do_established;
12843 		break;
12844 	case TCPS_CLOSE_WAIT:
12845 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12846 		rack->r_state = TCPS_CLOSE_WAIT;
12847 		rack->r_substate = rack_do_close_wait;
12848 		break;
12849 	case TCPS_FIN_WAIT_1:
12850 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12851 		rack->r_state = TCPS_FIN_WAIT_1;
12852 		rack->r_substate = rack_do_fin_wait_1;
12853 		break;
12854 	case TCPS_CLOSING:
12855 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12856 		rack->r_state = TCPS_CLOSING;
12857 		rack->r_substate = rack_do_closing;
12858 		break;
12859 	case TCPS_LAST_ACK:
12860 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12861 		rack->r_state = TCPS_LAST_ACK;
12862 		rack->r_substate = rack_do_lastack;
12863 		break;
12864 	case TCPS_FIN_WAIT_2:
12865 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12866 		rack->r_state = TCPS_FIN_WAIT_2;
12867 		rack->r_substate = rack_do_fin_wait_2;
12868 		break;
12869 	case TCPS_LISTEN:
12870 	case TCPS_CLOSED:
12871 	case TCPS_TIME_WAIT:
12872 	default:
12873 		break;
12874 	};
12875 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12876 		rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12877 
12878 }
12879 
12880 static void
12881 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12882 {
12883 	/*
12884 	 * We received an ack, and then did not
12885 	 * call send or were bounced out due to the
12886 	 * hpts was running. Now a timer is up as well, is
12887 	 * it the right timer?
12888 	 */
12889 	struct rack_sendmap *rsm;
12890 	int tmr_up;
12891 
12892 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12893 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12894 		return;
12895 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12896 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12897 	    (tmr_up == PACE_TMR_RXT)) {
12898 		/* Should be an RXT */
12899 		return;
12900 	}
12901 	if (rsm == NULL) {
12902 		/* Nothing outstanding? */
12903 		if (tp->t_flags & TF_DELACK) {
12904 			if (tmr_up == PACE_TMR_DELACK)
12905 				/* We are supposed to have delayed ack up and we do */
12906 				return;
12907 		} else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12908 			/*
12909 			 * if we hit enobufs then we would expect the possiblity
12910 			 * of nothing outstanding and the RXT up (and the hptsi timer).
12911 			 */
12912 			return;
12913 		} else if (((V_tcp_always_keepalive ||
12914 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12915 			    (tp->t_state <= TCPS_CLOSING)) &&
12916 			   (tmr_up == PACE_TMR_KEEP) &&
12917 			   (tp->snd_max == tp->snd_una)) {
12918 			/* We should have keep alive up and we do */
12919 			return;
12920 		}
12921 	}
12922 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12923 		   ((tmr_up == PACE_TMR_TLP) ||
12924 		    (tmr_up == PACE_TMR_RACK) ||
12925 		    (tmr_up == PACE_TMR_RXT))) {
12926 		/*
12927 		 * Either a Rack, TLP or RXT is fine if  we
12928 		 * have outstanding data.
12929 		 */
12930 		return;
12931 	} else if (tmr_up == PACE_TMR_DELACK) {
12932 		/*
12933 		 * If the delayed ack was going to go off
12934 		 * before the rtx/tlp/rack timer were going to
12935 		 * expire, then that would be the timer in control.
12936 		 * Note we don't check the time here trusting the
12937 		 * code is correct.
12938 		 */
12939 		return;
12940 	}
12941 	/*
12942 	 * Ok the timer originally started is not what we want now.
12943 	 * We will force the hpts to be stopped if any, and restart
12944 	 * with the slot set to what was in the saved slot.
12945 	 */
12946 	if (tcp_in_hpts(rack->rc_inp)) {
12947 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12948 			uint32_t us_cts;
12949 
12950 			us_cts = tcp_get_usecs(NULL);
12951 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12952 				rack->r_early = 1;
12953 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12954 			}
12955 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12956 		}
12957 		tcp_hpts_remove(tp->t_inpcb);
12958 	}
12959 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12960 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12961 }
12962 
12963 
12964 static void
12965 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts, uint32_t high_seq)
12966 {
12967 	if ((SEQ_LT(tp->snd_wl1, seq) ||
12968 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
12969 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
12970 		/* keep track of pure window updates */
12971 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
12972 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12973 		tp->snd_wnd = tiwin;
12974 		rack_validate_fo_sendwin_up(tp, rack);
12975 		tp->snd_wl1 = seq;
12976 		tp->snd_wl2 = ack;
12977 		if (tp->snd_wnd > tp->max_sndwnd)
12978 			tp->max_sndwnd = tp->snd_wnd;
12979 	    rack->r_wanted_output = 1;
12980 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
12981 		tp->snd_wnd = tiwin;
12982 		rack_validate_fo_sendwin_up(tp, rack);
12983 		tp->snd_wl1 = seq;
12984 		tp->snd_wl2 = ack;
12985 	} else {
12986 		/* Not a valid win update */
12987 		return;
12988 	}
12989 	if (tp->snd_wnd > tp->max_sndwnd)
12990 		tp->max_sndwnd = tp->snd_wnd;
12991 	if (tp->snd_wnd < (tp->snd_max - high_seq)) {
12992 		/* The peer collapsed the window */
12993 		rack_collapsed_window(rack);
12994 	} else if (rack->rc_has_collapsed)
12995 		rack_un_collapse_window(rack);
12996 	/* Do we exit persists? */
12997 	if ((rack->rc_in_persist != 0) &&
12998 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12999 				rack->r_ctl.rc_pace_min_segs))) {
13000 		rack_exit_persist(tp, rack, cts);
13001 	}
13002 	/* Do we enter persists? */
13003 	if ((rack->rc_in_persist == 0) &&
13004 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13005 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13006 	    (tp->snd_max == tp->snd_una) &&
13007 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
13008 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
13009 		/*
13010 		 * Here the rwnd is less than
13011 		 * the pacing size, we are established,
13012 		 * nothing is outstanding, and there is
13013 		 * data to send. Enter persists.
13014 		 */
13015 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13016 	}
13017 }
13018 
13019 static void
13020 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13021 {
13022 
13023 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13024 		union tcp_log_stackspecific log;
13025 		struct timeval ltv;
13026 		char tcp_hdr_buf[60];
13027 		struct tcphdr *th;
13028 		struct timespec ts;
13029 		uint32_t orig_snd_una;
13030 		uint8_t xx = 0;
13031 
13032 #ifdef NETFLIX_HTTP_LOGGING
13033 		struct http_sendfile_track *http_req;
13034 
13035 		if (SEQ_GT(ae->ack, tp->snd_una)) {
13036 			http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13037 		} else {
13038 			http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13039 		}
13040 #endif
13041 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13042 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13043 		if (rack->rack_no_prr == 0)
13044 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13045 		else
13046 			log.u_bbr.flex1 = 0;
13047 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13048 		log.u_bbr.use_lt_bw <<= 1;
13049 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
13050 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13051 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13052 		log.u_bbr.pkts_out = tp->t_maxseg;
13053 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13054 		log.u_bbr.flex7 = 1;
13055 		log.u_bbr.lost = ae->flags;
13056 		log.u_bbr.cwnd_gain = ackval;
13057 		log.u_bbr.pacing_gain = 0x2;
13058 		if (ae->flags & TSTMP_HDWR) {
13059 			/* Record the hardware timestamp if present */
13060 			log.u_bbr.flex3 = M_TSTMP;
13061 			ts.tv_sec = ae->timestamp / 1000000000;
13062 			ts.tv_nsec = ae->timestamp % 1000000000;
13063 			ltv.tv_sec = ts.tv_sec;
13064 			ltv.tv_usec = ts.tv_nsec / 1000;
13065 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13066 		} else if (ae->flags & TSTMP_LRO) {
13067 			/* Record the LRO the arrival timestamp */
13068 			log.u_bbr.flex3 = M_TSTMP_LRO;
13069 			ts.tv_sec = ae->timestamp / 1000000000;
13070 			ts.tv_nsec = ae->timestamp % 1000000000;
13071 			ltv.tv_sec = ts.tv_sec;
13072 			ltv.tv_usec = ts.tv_nsec / 1000;
13073 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13074 		}
13075 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13076 		/* Log the rcv time */
13077 		log.u_bbr.delRate = ae->timestamp;
13078 #ifdef NETFLIX_HTTP_LOGGING
13079 		log.u_bbr.applimited = tp->t_http_closed;
13080 		log.u_bbr.applimited <<= 8;
13081 		log.u_bbr.applimited |= tp->t_http_open;
13082 		log.u_bbr.applimited <<= 8;
13083 		log.u_bbr.applimited |= tp->t_http_req;
13084 		if (http_req) {
13085 			/* Copy out any client req info */
13086 			/* seconds */
13087 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13088 			/* useconds */
13089 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13090 			log.u_bbr.rttProp = http_req->timestamp;
13091 			log.u_bbr.cur_del_rate = http_req->start;
13092 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13093 				log.u_bbr.flex8 |= 1;
13094 			} else {
13095 				log.u_bbr.flex8 |= 2;
13096 				log.u_bbr.bw_inuse = http_req->end;
13097 			}
13098 			log.u_bbr.flex6 = http_req->start_seq;
13099 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13100 				log.u_bbr.flex8 |= 4;
13101 				log.u_bbr.epoch = http_req->end_seq;
13102 			}
13103 		}
13104 #endif
13105 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13106 		th = (struct tcphdr *)tcp_hdr_buf;
13107 		th->th_seq = ae->seq;
13108 		th->th_ack = ae->ack;
13109 		th->th_win = ae->win;
13110 		/* Now fill in the ports */
13111 		th->th_sport = tp->t_inpcb->inp_fport;
13112 		th->th_dport = tp->t_inpcb->inp_lport;
13113 		tcp_set_flags(th, ae->flags);
13114 		/* Now do we have a timestamp option? */
13115 		if (ae->flags & HAS_TSTMP) {
13116 			u_char *cp;
13117 			uint32_t val;
13118 
13119 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13120 			cp = (u_char *)(th + 1);
13121 			*cp = TCPOPT_NOP;
13122 			cp++;
13123 			*cp = TCPOPT_NOP;
13124 			cp++;
13125 			*cp = TCPOPT_TIMESTAMP;
13126 			cp++;
13127 			*cp = TCPOLEN_TIMESTAMP;
13128 			cp++;
13129 			val = htonl(ae->ts_value);
13130 			bcopy((char *)&val,
13131 			      (char *)cp, sizeof(uint32_t));
13132 			val = htonl(ae->ts_echo);
13133 			bcopy((char *)&val,
13134 			      (char *)(cp + 4), sizeof(uint32_t));
13135 		} else
13136 			th->th_off = (sizeof(struct tcphdr) >> 2);
13137 
13138 		/*
13139 		 * For sane logging we need to play a little trick.
13140 		 * If the ack were fully processed we would have moved
13141 		 * snd_una to high_seq, but since compressed acks are
13142 		 * processed in two phases, at this point (logging) snd_una
13143 		 * won't be advanced. So we would see multiple acks showing
13144 		 * the advancement. We can prevent that by "pretending" that
13145 		 * snd_una was advanced and then un-advancing it so that the
13146 		 * logging code has the right value for tlb_snd_una.
13147 		 */
13148 		if (tp->snd_una != high_seq) {
13149 			orig_snd_una = tp->snd_una;
13150 			tp->snd_una = high_seq;
13151 			xx = 1;
13152 		} else
13153 			xx = 0;
13154 		TCP_LOG_EVENTP(tp, th,
13155 			       &tp->t_inpcb->inp_socket->so_rcv,
13156 			       &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
13157 			       0, &log, true, &ltv);
13158 		if (xx) {
13159 			tp->snd_una = orig_snd_una;
13160 		}
13161 	}
13162 
13163 }
13164 
13165 static void
13166 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13167 {
13168 	uint32_t us_rtt;
13169 	/*
13170 	 * A persist or keep-alive was forced out, update our
13171 	 * min rtt time. Note now worry about lost responses.
13172 	 * When a subsequent keep-alive or persist times out
13173 	 * and forced_ack is still on, then the last probe
13174 	 * was not responded to. In such cases we have a
13175 	 * sysctl that controls the behavior. Either we apply
13176 	 * the rtt but with reduced confidence (0). Or we just
13177 	 * plain don't apply the rtt estimate. Having data flow
13178 	 * will clear the probe_not_answered flag i.e. cum-ack
13179 	 * move forward <or> exiting and reentering persists.
13180 	 */
13181 
13182 	rack->forced_ack = 0;
13183 	rack->rc_tp->t_rxtshift = 0;
13184 	if ((rack->rc_in_persist &&
13185 	     (tiwin == rack->rc_tp->snd_wnd)) ||
13186 	    (rack->rc_in_persist == 0)) {
13187 		/*
13188 		 * In persists only apply the RTT update if this is
13189 		 * a response to our window probe. And that
13190 		 * means the rwnd sent must match the current
13191 		 * snd_wnd. If it does not, then we got a
13192 		 * window update ack instead. For keepalive
13193 		 * we allow the answer no matter what the window.
13194 		 *
13195 		 * Note that if the probe_not_answered is set then
13196 		 * the forced_ack_ts is the oldest one i.e. the first
13197 		 * probe sent that might have been lost. This assures
13198 		 * us that if we do calculate an RTT it is longer not
13199 		 * some short thing.
13200 		 */
13201 		if (rack->rc_in_persist)
13202 			counter_u64_add(rack_persists_acks, 1);
13203 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13204 		if (us_rtt == 0)
13205 			us_rtt = 1;
13206 		if (rack->probe_not_answered == 0) {
13207 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13208 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13209 		} else {
13210 			/* We have a retransmitted probe here too */
13211 			if (rack_apply_rtt_with_reduced_conf) {
13212 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13213 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13214 			}
13215 		}
13216 	}
13217 }
13218 
13219 static int
13220 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13221 {
13222 	/*
13223 	 * Handle a "special" compressed ack mbuf. Each incoming
13224 	 * ack has only four possible dispositions:
13225 	 *
13226 	 * A) It moves the cum-ack forward
13227 	 * B) It is behind the cum-ack.
13228 	 * C) It is a window-update ack.
13229 	 * D) It is a dup-ack.
13230 	 *
13231 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13232 	 * in the incoming mbuf. We also need to still pay attention
13233 	 * to nxt_pkt since there may be another packet after this
13234 	 * one.
13235 	 */
13236 #ifdef TCP_ACCOUNTING
13237 	uint64_t ts_val;
13238 	uint64_t rdstc;
13239 #endif
13240 	int segsiz;
13241 	struct timespec ts;
13242 	struct tcp_rack *rack;
13243 	struct tcp_ackent *ae;
13244 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13245 	int cnt, i, did_out, ourfinisacked = 0;
13246 	struct tcpopt to_holder, *to = NULL;
13247 #ifdef TCP_ACCOUNTING
13248 	int win_up_req = 0;
13249 #endif
13250 	int nsegs = 0;
13251 	int under_pacing = 1;
13252 	int recovery = 0;
13253 #ifdef TCP_ACCOUNTING
13254 	sched_pin();
13255 #endif
13256 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13257 	if (rack->gp_ready &&
13258 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13259 		under_pacing = 0;
13260 	else
13261 		under_pacing = 1;
13262 
13263 	if (rack->r_state != tp->t_state)
13264 		rack_set_state(tp, rack);
13265 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13266 	    (tp->t_flags & TF_GPUTINPROG)) {
13267 		/*
13268 		 * We have a goodput in progress
13269 		 * and we have entered a late state.
13270 		 * Do we have enough data in the sb
13271 		 * to handle the GPUT request?
13272 		 */
13273 		uint32_t bytes;
13274 
13275 		bytes = tp->gput_ack - tp->gput_seq;
13276 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
13277 			bytes += tp->gput_seq - tp->snd_una;
13278 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
13279 			/*
13280 			 * There are not enough bytes in the socket
13281 			 * buffer that have been sent to cover this
13282 			 * measurement. Cancel it.
13283 			 */
13284 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13285 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
13286 						   tp->gput_seq,
13287 						   0, 0, 18, __LINE__, NULL, 0);
13288 			tp->t_flags &= ~TF_GPUTINPROG;
13289 		}
13290 	}
13291 	to = &to_holder;
13292 	to->to_flags = 0;
13293 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13294 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13295 	cnt = m->m_len / sizeof(struct tcp_ackent);
13296 	counter_u64_add(rack_multi_single_eq, cnt);
13297 	high_seq = tp->snd_una;
13298 	the_win = tp->snd_wnd;
13299 	win_seq = tp->snd_wl1;
13300 	win_upd_ack = tp->snd_wl2;
13301 	cts = tcp_tv_to_usectick(tv);
13302 	ms_cts = tcp_tv_to_mssectick(tv);
13303 	rack->r_ctl.rc_rcvtime = cts;
13304 	segsiz = ctf_fixed_maxseg(tp);
13305 	if ((rack->rc_gp_dyn_mul) &&
13306 	    (rack->use_fixed_rate == 0) &&
13307 	    (rack->rc_always_pace)) {
13308 		/* Check in on probertt */
13309 		rack_check_probe_rtt(rack, cts);
13310 	}
13311 	for (i = 0; i < cnt; i++) {
13312 #ifdef TCP_ACCOUNTING
13313 		ts_val = get_cyclecount();
13314 #endif
13315 		rack_clear_rate_sample(rack);
13316 		ae = ((mtod(m, struct tcp_ackent *)) + i);
13317 		/* Setup the window */
13318 		tiwin = ae->win << tp->snd_scale;
13319 		if (tiwin > rack->r_ctl.rc_high_rwnd)
13320 			rack->r_ctl.rc_high_rwnd = tiwin;
13321 		/* figure out the type of ack */
13322 		if (SEQ_LT(ae->ack, high_seq)) {
13323 			/* Case B*/
13324 			ae->ack_val_set = ACK_BEHIND;
13325 		} else if (SEQ_GT(ae->ack, high_seq)) {
13326 			/* Case A */
13327 			ae->ack_val_set = ACK_CUMACK;
13328 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13329 			/* Case D */
13330 			ae->ack_val_set = ACK_DUPACK;
13331 		} else {
13332 			/* Case C */
13333 			ae->ack_val_set = ACK_RWND;
13334 		}
13335 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13336 		/* Validate timestamp */
13337 		if (ae->flags & HAS_TSTMP) {
13338 			/* Setup for a timestamp */
13339 			to->to_flags = TOF_TS;
13340 			ae->ts_echo -= tp->ts_offset;
13341 			to->to_tsecr = ae->ts_echo;
13342 			to->to_tsval = ae->ts_value;
13343 			/*
13344 			 * If echoed timestamp is later than the current time, fall back to
13345 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13346 			 * were used when this connection was established.
13347 			 */
13348 			if (TSTMP_GT(ae->ts_echo, ms_cts))
13349 				to->to_tsecr = 0;
13350 			if (tp->ts_recent &&
13351 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13352 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13353 #ifdef TCP_ACCOUNTING
13354 					rdstc = get_cyclecount();
13355 					if (rdstc > ts_val) {
13356 						counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13357 								(rdstc - ts_val));
13358 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13359 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13360 						}
13361 					}
13362 #endif
13363 					continue;
13364 				}
13365 			}
13366 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13367 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13368 				tp->ts_recent_age = tcp_ts_getticks();
13369 				tp->ts_recent = ae->ts_value;
13370 			}
13371 		} else {
13372 			/* Setup for a no options */
13373 			to->to_flags = 0;
13374 		}
13375 		/* Update the rcv time and perform idle reduction possibly */
13376 		if  (tp->t_idle_reduce &&
13377 		     (tp->snd_max == tp->snd_una) &&
13378 		     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13379 			counter_u64_add(rack_input_idle_reduces, 1);
13380 			rack_cc_after_idle(rack, tp);
13381 		}
13382 		tp->t_rcvtime = ticks;
13383 		/* Now what about ECN? */
13384 		if (tcp_ecn_input_segment(tp, ae->flags, ae->codepoint))
13385 			rack_cong_signal(tp, CC_ECN, ae->ack);
13386 #ifdef TCP_ACCOUNTING
13387 		/* Count for the specific type of ack in */
13388 		counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13389 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13390 			tp->tcp_cnt_counters[ae->ack_val_set]++;
13391 		}
13392 #endif
13393 		/*
13394 		 * Note how we could move up these in the determination
13395 		 * above, but we don't so that way the timestamp checks (and ECN)
13396 		 * is done first before we do any processing on the ACK.
13397 		 * The non-compressed path through the code has this
13398 		 * weakness (noted by @jtl) that it actually does some
13399 		 * processing before verifying the timestamp information.
13400 		 * We don't take that path here which is why we set
13401 		 * the ack_val_set first, do the timestamp and ecn
13402 		 * processing, and then look at what we have setup.
13403 		 */
13404 		if (ae->ack_val_set == ACK_BEHIND) {
13405 			/*
13406 			 * Case B flag reordering, if window is not closed
13407 			 * or it could be a keep-alive or persists
13408 			 */
13409 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13410 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13411 			}
13412 		} else if (ae->ack_val_set == ACK_DUPACK) {
13413 			/* Case D */
13414 			rack_strike_dupack(rack);
13415 		} else if (ae->ack_val_set == ACK_RWND) {
13416 			/* Case C */
13417 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13418 				ts.tv_sec = ae->timestamp / 1000000000;
13419 				ts.tv_nsec = ae->timestamp % 1000000000;
13420 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13421 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13422 			} else {
13423 				rack->r_ctl.act_rcv_time = *tv;
13424 			}
13425 			if (rack->forced_ack) {
13426 				rack_handle_probe_response(rack, tiwin,
13427 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13428 			}
13429 #ifdef TCP_ACCOUNTING
13430 			win_up_req = 1;
13431 #endif
13432 			win_upd_ack = ae->ack;
13433 			win_seq = ae->seq;
13434 			the_win = tiwin;
13435 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13436 		} else {
13437 			/* Case A */
13438 			if (SEQ_GT(ae->ack, tp->snd_max)) {
13439 				/*
13440 				 * We just send an ack since the incoming
13441 				 * ack is beyond the largest seq we sent.
13442 				 */
13443 				if ((tp->t_flags & TF_ACKNOW) == 0) {
13444 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13445 					if (tp->t_flags && TF_ACKNOW)
13446 						rack->r_wanted_output = 1;
13447 				}
13448 			} else {
13449 				nsegs++;
13450 				/* If the window changed setup to update */
13451 				if (tiwin != tp->snd_wnd) {
13452 					win_upd_ack = ae->ack;
13453 					win_seq = ae->seq;
13454 					the_win = tiwin;
13455 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13456 				}
13457 #ifdef TCP_ACCOUNTING
13458 				/* Account for the acks */
13459 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13460 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13461 				}
13462 				counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13463 						(((ae->ack - high_seq) + segsiz - 1) / segsiz));
13464 #endif
13465 				high_seq = ae->ack;
13466 				if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
13467 					union tcp_log_stackspecific log;
13468 					struct timeval tv;
13469 
13470 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13471 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13472 					log.u_bbr.flex1 = high_seq;
13473 					log.u_bbr.flex2 = rack->r_ctl.roundends;
13474 					log.u_bbr.flex3 = rack->r_ctl.current_round;
13475 					log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
13476 					log.u_bbr.flex8 = 8;
13477 					tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
13478 						       0, &log, false, NULL, NULL, 0, &tv);
13479 				}
13480 				/*
13481 				 * The draft (v3) calls for us to use SEQ_GEQ, but that
13482 				 * causes issues when we are just going app limited. Lets
13483 				 * instead use SEQ_GT <or> where its equal but more data
13484 				 * is outstanding.
13485 				 */
13486 				if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) ||
13487 				    ((high_seq == rack->r_ctl.roundends) &&
13488 				     SEQ_GT(tp->snd_max, tp->snd_una))) {
13489 					rack->r_ctl.current_round++;
13490 					rack->r_ctl.roundends = tp->snd_max;
13491 					if (CC_ALGO(tp)->newround != NULL) {
13492 						CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
13493 					}
13494 				}
13495 				/* Setup our act_rcv_time */
13496 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13497 					ts.tv_sec = ae->timestamp / 1000000000;
13498 					ts.tv_nsec = ae->timestamp % 1000000000;
13499 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13500 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13501 				} else {
13502 					rack->r_ctl.act_rcv_time = *tv;
13503 				}
13504 				rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13505 				if (rack->rc_dsack_round_seen) {
13506 					/* Is the dsack round over? */
13507 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13508 						/* Yes it is */
13509 						rack->rc_dsack_round_seen = 0;
13510 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13511 					}
13512 				}
13513 			}
13514 		}
13515 		/* And lets be sure to commit the rtt measurements for this ack */
13516 		tcp_rack_xmit_timer_commit(rack, tp);
13517 #ifdef TCP_ACCOUNTING
13518 		rdstc = get_cyclecount();
13519 		if (rdstc > ts_val) {
13520 			counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13521 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13522 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13523 				if (ae->ack_val_set == ACK_CUMACK)
13524 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13525 			}
13526 		}
13527 #endif
13528 	}
13529 #ifdef TCP_ACCOUNTING
13530 	ts_val = get_cyclecount();
13531 #endif
13532 	acked_amount = acked = (high_seq - tp->snd_una);
13533 	if (acked) {
13534 		/*
13535 		 * Clear the probe not answered flag
13536 		 * since cum-ack moved forward.
13537 		 */
13538 		rack->probe_not_answered = 0;
13539 		if (rack->sack_attack_disable == 0)
13540 			rack_do_decay(rack);
13541 		if (acked >= segsiz) {
13542 			/*
13543 			 * You only get credit for
13544 			 * MSS and greater (and you get extra
13545 			 * credit for larger cum-ack moves).
13546 			 */
13547 			int ac;
13548 
13549 			ac = acked / segsiz;
13550 			rack->r_ctl.ack_count += ac;
13551 			counter_u64_add(rack_ack_total, ac);
13552 		}
13553 		if (rack->r_ctl.ack_count > 0xfff00000) {
13554 			/*
13555 			 * reduce the number to keep us under
13556 			 * a uint32_t.
13557 			 */
13558 			rack->r_ctl.ack_count /= 2;
13559 			rack->r_ctl.sack_count /= 2;
13560 		}
13561 		if (tp->t_flags & TF_NEEDSYN) {
13562 			/*
13563 			 * T/TCP: Connection was half-synchronized, and our SYN has
13564 			 * been ACK'd (so connection is now fully synchronized).  Go
13565 			 * to non-starred state, increment snd_una for ACK of SYN,
13566 			 * and check if we can do window scaling.
13567 			 */
13568 			tp->t_flags &= ~TF_NEEDSYN;
13569 			tp->snd_una++;
13570 			acked_amount = acked = (high_seq - tp->snd_una);
13571 		}
13572 		if (acked > sbavail(&so->so_snd))
13573 			acked_amount = sbavail(&so->so_snd);
13574 #ifdef NETFLIX_EXP_DETECTION
13575 		/*
13576 		 * We only care on a cum-ack move if we are in a sack-disabled
13577 		 * state. We have already added in to the ack_count, and we never
13578 		 * would disable on a cum-ack move, so we only care to do the
13579 		 * detection if it may "undo" it, i.e. we were in disabled already.
13580 		 */
13581 		if (rack->sack_attack_disable)
13582 			rack_do_detection(tp, rack, acked_amount, segsiz);
13583 #endif
13584 		if (IN_FASTRECOVERY(tp->t_flags) &&
13585 		    (rack->rack_no_prr == 0))
13586 			rack_update_prr(tp, rack, acked_amount, high_seq);
13587 		if (IN_RECOVERY(tp->t_flags)) {
13588 			if (SEQ_LT(high_seq, tp->snd_recover) &&
13589 			    (SEQ_LT(high_seq, tp->snd_max))) {
13590 				tcp_rack_partialack(tp);
13591 			} else {
13592 				rack_post_recovery(tp, high_seq);
13593 				recovery = 1;
13594 			}
13595 		}
13596 		/* Handle the rack-log-ack part (sendmap) */
13597 		if ((sbused(&so->so_snd) == 0) &&
13598 		    (acked > acked_amount) &&
13599 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13600 		    (tp->t_flags & TF_SENTFIN)) {
13601 			/*
13602 			 * We must be sure our fin
13603 			 * was sent and acked (we can be
13604 			 * in FIN_WAIT_1 without having
13605 			 * sent the fin).
13606 			 */
13607 			ourfinisacked = 1;
13608 			/*
13609 			 * Lets make sure snd_una is updated
13610 			 * since most likely acked_amount = 0 (it
13611 			 * should be).
13612 			 */
13613 			tp->snd_una = high_seq;
13614 		}
13615 		/* Did we make a RTO error? */
13616 		if ((tp->t_flags & TF_PREVVALID) &&
13617 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13618 			tp->t_flags &= ~TF_PREVVALID;
13619 			if (tp->t_rxtshift == 1 &&
13620 			    (int)(ticks - tp->t_badrxtwin) < 0)
13621 				rack_cong_signal(tp, CC_RTO_ERR, high_seq);
13622 		}
13623 		/* Handle the data in the socket buffer */
13624 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13625 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13626 		if (acked_amount > 0) {
13627 			struct mbuf *mfree;
13628 
13629 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13630 			SOCKBUF_LOCK(&so->so_snd);
13631 			mfree = sbcut_locked(&so->so_snd, acked_amount);
13632 			tp->snd_una = high_seq;
13633 			/* Note we want to hold the sb lock through the sendmap adjust */
13634 			rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13635 			/* Wake up the socket if we have room to write more */
13636 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13637 			sowwakeup_locked(so);
13638 			m_freem(mfree);
13639 		}
13640 		/* update progress */
13641 		tp->t_acktime = ticks;
13642 		rack_log_progress_event(rack, tp, tp->t_acktime,
13643 					PROGRESS_UPDATE, __LINE__);
13644 		/* Clear out shifts and such */
13645 		tp->t_rxtshift = 0;
13646 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13647 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13648 		rack->rc_tlp_in_progress = 0;
13649 		rack->r_ctl.rc_tlp_cnt_out = 0;
13650 		/* Send recover and snd_nxt must be dragged along */
13651 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
13652 			tp->snd_recover = tp->snd_una;
13653 		if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13654 			tp->snd_nxt = tp->snd_una;
13655 		/*
13656 		 * If the RXT timer is running we want to
13657 		 * stop it, so we can restart a TLP (or new RXT).
13658 		 */
13659 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13660 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13661 #ifdef NETFLIX_HTTP_LOGGING
13662 		tcp_http_check_for_comp(rack->rc_tp, high_seq);
13663 #endif
13664 		tp->snd_wl2 = high_seq;
13665 		tp->t_dupacks = 0;
13666 		if (under_pacing &&
13667 		    (rack->use_fixed_rate == 0) &&
13668 		    (rack->in_probe_rtt == 0) &&
13669 		    rack->rc_gp_dyn_mul &&
13670 		    rack->rc_always_pace) {
13671 			/* Check if we are dragging bottom */
13672 			rack_check_bottom_drag(tp, rack, so, acked);
13673 		}
13674 		if (tp->snd_una == tp->snd_max) {
13675 			tp->t_flags &= ~TF_PREVVALID;
13676 			rack->r_ctl.retran_during_recovery = 0;
13677 			rack->r_ctl.dsack_byte_cnt = 0;
13678 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13679 			if (rack->r_ctl.rc_went_idle_time == 0)
13680 				rack->r_ctl.rc_went_idle_time = 1;
13681 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13682 			if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13683 				tp->t_acktime = 0;
13684 			/* Set so we might enter persists... */
13685 			rack->r_wanted_output = 1;
13686 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13687 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13688 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13689 			    (sbavail(&so->so_snd) == 0) &&
13690 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13691 				/*
13692 				 * The socket was gone and the
13693 				 * peer sent data (not now in the past), time to
13694 				 * reset him.
13695 				 */
13696 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13697 				/* tcp_close will kill the inp pre-log the Reset */
13698 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13699 #ifdef TCP_ACCOUNTING
13700 				rdstc = get_cyclecount();
13701 				if (rdstc > ts_val) {
13702 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13703 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13704 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13705 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13706 					}
13707 				}
13708 #endif
13709 				m_freem(m);
13710 				tp = tcp_close(tp);
13711 				if (tp == NULL) {
13712 #ifdef TCP_ACCOUNTING
13713 					sched_unpin();
13714 #endif
13715 					return (1);
13716 				}
13717 				/*
13718 				 * We would normally do drop-with-reset which would
13719 				 * send back a reset. We can't since we don't have
13720 				 * all the needed bits. Instead lets arrange for
13721 				 * a call to tcp_output(). That way since we
13722 				 * are in the closed state we will generate a reset.
13723 				 *
13724 				 * Note if tcp_accounting is on we don't unpin since
13725 				 * we do that after the goto label.
13726 				 */
13727 				goto send_out_a_rst;
13728 			}
13729 			if ((sbused(&so->so_snd) == 0) &&
13730 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13731 			    (tp->t_flags & TF_SENTFIN)) {
13732 				/*
13733 				 * If we can't receive any more data, then closing user can
13734 				 * proceed. Starting the timer is contrary to the
13735 				 * specification, but if we don't get a FIN we'll hang
13736 				 * forever.
13737 				 *
13738 				 */
13739 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13740 					soisdisconnected(so);
13741 					tcp_timer_activate(tp, TT_2MSL,
13742 							   (tcp_fast_finwait2_recycle ?
13743 							    tcp_finwait2_timeout :
13744 							    TP_MAXIDLE(tp)));
13745 				}
13746 				if (ourfinisacked == 0) {
13747 					/*
13748 					 * We don't change to fin-wait-2 if we have our fin acked
13749 					 * which means we are probably in TCPS_CLOSING.
13750 					 */
13751 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
13752 				}
13753 			}
13754 		}
13755 		/* Wake up the socket if we have room to write more */
13756 		if (sbavail(&so->so_snd)) {
13757 			rack->r_wanted_output = 1;
13758 			if (ctf_progress_timeout_check(tp, true)) {
13759 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13760 							tp, tick, PROGRESS_DROP, __LINE__);
13761 				/*
13762 				 * We cheat here and don't send a RST, we should send one
13763 				 * when the pacer drops the connection.
13764 				 */
13765 #ifdef TCP_ACCOUNTING
13766 				rdstc = get_cyclecount();
13767 				if (rdstc > ts_val) {
13768 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13769 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13770 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13771 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13772 					}
13773 				}
13774 				sched_unpin();
13775 #endif
13776 				(void)tcp_drop(tp, ETIMEDOUT);
13777 				m_freem(m);
13778 				return (1);
13779 			}
13780 		}
13781 		if (ourfinisacked) {
13782 			switch(tp->t_state) {
13783 			case TCPS_CLOSING:
13784 #ifdef TCP_ACCOUNTING
13785 				rdstc = get_cyclecount();
13786 				if (rdstc > ts_val) {
13787 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13788 							(rdstc - ts_val));
13789 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13790 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13791 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13792 					}
13793 				}
13794 				sched_unpin();
13795 #endif
13796 				tcp_twstart(tp);
13797 				m_freem(m);
13798 				return (1);
13799 				break;
13800 			case TCPS_LAST_ACK:
13801 #ifdef TCP_ACCOUNTING
13802 				rdstc = get_cyclecount();
13803 				if (rdstc > ts_val) {
13804 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13805 							(rdstc - ts_val));
13806 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13807 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13808 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13809 					}
13810 				}
13811 				sched_unpin();
13812 #endif
13813 				tp = tcp_close(tp);
13814 				ctf_do_drop(m, tp);
13815 				return (1);
13816 				break;
13817 			case TCPS_FIN_WAIT_1:
13818 #ifdef TCP_ACCOUNTING
13819 				rdstc = get_cyclecount();
13820 				if (rdstc > ts_val) {
13821 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13822 							(rdstc - ts_val));
13823 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13824 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13825 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13826 					}
13827 				}
13828 #endif
13829 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13830 					soisdisconnected(so);
13831 					tcp_timer_activate(tp, TT_2MSL,
13832 							   (tcp_fast_finwait2_recycle ?
13833 							    tcp_finwait2_timeout :
13834 							    TP_MAXIDLE(tp)));
13835 				}
13836 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13837 				break;
13838 			default:
13839 				break;
13840 			}
13841 		}
13842 		if (rack->r_fast_output) {
13843 			/*
13844 			 * We re doing fast output.. can we expand that?
13845 			 */
13846 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13847 		}
13848 #ifdef TCP_ACCOUNTING
13849 		rdstc = get_cyclecount();
13850 		if (rdstc > ts_val) {
13851 			counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13852 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13853 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13854 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13855 			}
13856 		}
13857 
13858 	} else if (win_up_req) {
13859 		rdstc = get_cyclecount();
13860 		if (rdstc > ts_val) {
13861 			counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13862 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13863 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13864 			}
13865 		}
13866 #endif
13867 	}
13868 	/* Now is there a next packet, if so we are done */
13869 	m_freem(m);
13870 	did_out = 0;
13871 	if (nxt_pkt) {
13872 #ifdef TCP_ACCOUNTING
13873 		sched_unpin();
13874 #endif
13875 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13876 		return (0);
13877 	}
13878 	rack_handle_might_revert(tp, rack);
13879 	ctf_calc_rwin(so, tp);
13880 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13881 	send_out_a_rst:
13882 		if (tcp_output(tp) < 0) {
13883 #ifdef TCP_ACCOUNTING
13884 			sched_unpin();
13885 #endif
13886 			return (1);
13887 		}
13888 		did_out = 1;
13889 	}
13890 	rack_free_trim(rack);
13891 #ifdef TCP_ACCOUNTING
13892 	sched_unpin();
13893 #endif
13894 	rack_timer_audit(tp, rack, &so->so_snd);
13895 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13896 	return (0);
13897 }
13898 
13899 
13900 static int
13901 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13902     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13903     int32_t nxt_pkt, struct timeval *tv)
13904 {
13905 #ifdef TCP_ACCOUNTING
13906 	uint64_t ts_val;
13907 #endif
13908 	int32_t thflags, retval, did_out = 0;
13909 	int32_t way_out = 0;
13910 	/*
13911 	 * cts - is the current time from tv (caller gets ts) in microseconds.
13912 	 * ms_cts - is the current time from tv in milliseconds.
13913 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
13914 	 */
13915 	uint32_t cts, us_cts, ms_cts;
13916 	uint32_t tiwin, high_seq;
13917 	struct timespec ts;
13918 	struct tcpopt to;
13919 	struct tcp_rack *rack;
13920 	struct rack_sendmap *rsm;
13921 	int32_t prev_state = 0;
13922 #ifdef TCP_ACCOUNTING
13923 	int ack_val_set = 0xf;
13924 #endif
13925 	int nsegs;
13926 	/*
13927 	 * tv passed from common code is from either M_TSTMP_LRO or
13928 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13929 	 */
13930 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13931 	if (m->m_flags & M_ACKCMP) {
13932 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13933 	}
13934 	if (m->m_flags & M_ACKCMP) {
13935 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13936 	}
13937 	cts = tcp_tv_to_usectick(tv);
13938 	ms_cts =  tcp_tv_to_mssectick(tv);
13939 	nsegs = m->m_pkthdr.lro_nsegs;
13940 	counter_u64_add(rack_proc_non_comp_ack, 1);
13941 	thflags = tcp_get_flags(th);
13942 #ifdef TCP_ACCOUNTING
13943 	sched_pin();
13944 	if (thflags & TH_ACK)
13945 		ts_val = get_cyclecount();
13946 #endif
13947 	if ((m->m_flags & M_TSTMP) ||
13948 	    (m->m_flags & M_TSTMP_LRO)) {
13949 		mbuf_tstmp2timespec(m, &ts);
13950 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13951 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13952 	} else
13953 		rack->r_ctl.act_rcv_time = *tv;
13954 	kern_prefetch(rack, &prev_state);
13955 	prev_state = 0;
13956 	/*
13957 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
13958 	 * the scale is zero.
13959 	 */
13960 	tiwin = th->th_win << tp->snd_scale;
13961 #ifdef TCP_ACCOUNTING
13962 	if (thflags & TH_ACK) {
13963 		/*
13964 		 * We have a tradeoff here. We can either do what we are
13965 		 * doing i.e. pinning to this CPU and then doing the accounting
13966 		 * <or> we could do a critical enter, setup the rdtsc and cpu
13967 		 * as in below, and then validate we are on the same CPU on
13968 		 * exit. I have choosen to not do the critical enter since
13969 		 * that often will gain you a context switch, and instead lock
13970 		 * us (line above this if) to the same CPU with sched_pin(). This
13971 		 * means we may be context switched out for a higher priority
13972 		 * interupt but we won't be moved to another CPU.
13973 		 *
13974 		 * If this occurs (which it won't very often since we most likely
13975 		 * are running this code in interupt context and only a higher
13976 		 * priority will bump us ... clock?) we will falsely add in
13977 		 * to the time the interupt processing time plus the ack processing
13978 		 * time. This is ok since its a rare event.
13979 		 */
13980 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
13981 						    ctf_fixed_maxseg(tp));
13982 	}
13983 #endif
13984 	/*
13985 	 * Parse options on any incoming segment.
13986 	 */
13987 	memset(&to, 0, sizeof(to));
13988 	tcp_dooptions(&to, (u_char *)(th + 1),
13989 	    (th->th_off << 2) - sizeof(struct tcphdr),
13990 	    (thflags & TH_SYN) ? TO_SYN : 0);
13991 	NET_EPOCH_ASSERT();
13992 	INP_WLOCK_ASSERT(tp->t_inpcb);
13993 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
13994 	    __func__));
13995 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
13996 	    __func__));
13997 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13998 	    (tp->t_flags & TF_GPUTINPROG)) {
13999 		/*
14000 		 * We have a goodput in progress
14001 		 * and we have entered a late state.
14002 		 * Do we have enough data in the sb
14003 		 * to handle the GPUT request?
14004 		 */
14005 		uint32_t bytes;
14006 
14007 		bytes = tp->gput_ack - tp->gput_seq;
14008 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
14009 			bytes += tp->gput_seq - tp->snd_una;
14010 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
14011 			/*
14012 			 * There are not enough bytes in the socket
14013 			 * buffer that have been sent to cover this
14014 			 * measurement. Cancel it.
14015 			 */
14016 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14017 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
14018 						   tp->gput_seq,
14019 						   0, 0, 18, __LINE__, NULL, 0);
14020 			tp->t_flags &= ~TF_GPUTINPROG;
14021 		}
14022 	}
14023 	high_seq = th->th_ack;
14024 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14025 		union tcp_log_stackspecific log;
14026 		struct timeval ltv;
14027 #ifdef NETFLIX_HTTP_LOGGING
14028 		struct http_sendfile_track *http_req;
14029 
14030 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
14031 			http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14032 		} else {
14033 			http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14034 		}
14035 #endif
14036 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14037 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14038 		if (rack->rack_no_prr == 0)
14039 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14040 		else
14041 			log.u_bbr.flex1 = 0;
14042 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14043 		log.u_bbr.use_lt_bw <<= 1;
14044 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
14045 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14046 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14047 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14048 		log.u_bbr.flex3 = m->m_flags;
14049 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14050 		log.u_bbr.lost = thflags;
14051 		log.u_bbr.pacing_gain = 0x1;
14052 #ifdef TCP_ACCOUNTING
14053 		log.u_bbr.cwnd_gain = ack_val_set;
14054 #endif
14055 		log.u_bbr.flex7 = 2;
14056 		if (m->m_flags & M_TSTMP) {
14057 			/* Record the hardware timestamp if present */
14058 			mbuf_tstmp2timespec(m, &ts);
14059 			ltv.tv_sec = ts.tv_sec;
14060 			ltv.tv_usec = ts.tv_nsec / 1000;
14061 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14062 		} else if (m->m_flags & M_TSTMP_LRO) {
14063 			/* Record the LRO the arrival timestamp */
14064 			mbuf_tstmp2timespec(m, &ts);
14065 			ltv.tv_sec = ts.tv_sec;
14066 			ltv.tv_usec = ts.tv_nsec / 1000;
14067 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14068 		}
14069 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14070 		/* Log the rcv time */
14071 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14072 #ifdef NETFLIX_HTTP_LOGGING
14073 		log.u_bbr.applimited = tp->t_http_closed;
14074 		log.u_bbr.applimited <<= 8;
14075 		log.u_bbr.applimited |= tp->t_http_open;
14076 		log.u_bbr.applimited <<= 8;
14077 		log.u_bbr.applimited |= tp->t_http_req;
14078 		if (http_req) {
14079 			/* Copy out any client req info */
14080 			/* seconds */
14081 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14082 			/* useconds */
14083 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14084 			log.u_bbr.rttProp = http_req->timestamp;
14085 			log.u_bbr.cur_del_rate = http_req->start;
14086 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14087 				log.u_bbr.flex8 |= 1;
14088 			} else {
14089 				log.u_bbr.flex8 |= 2;
14090 				log.u_bbr.bw_inuse = http_req->end;
14091 			}
14092 			log.u_bbr.flex6 = http_req->start_seq;
14093 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14094 				log.u_bbr.flex8 |= 4;
14095 				log.u_bbr.epoch = http_req->end_seq;
14096 			}
14097 		}
14098 #endif
14099 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14100 		    tlen, &log, true, &ltv);
14101 	}
14102 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14103 		way_out = 4;
14104 		retval = 0;
14105 		m_freem(m);
14106 		goto done_with_input;
14107 	}
14108 	/*
14109 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
14110 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14111 	 */
14112 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14113 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14114 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14115 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14116 #ifdef TCP_ACCOUNTING
14117 		sched_unpin();
14118 #endif
14119 		return (1);
14120 	}
14121 	/*
14122 	 * If timestamps were negotiated during SYN/ACK and a
14123 	 * segment without a timestamp is received, silently drop
14124 	 * the segment, unless it is a RST segment or missing timestamps are
14125 	 * tolerated.
14126 	 * See section 3.2 of RFC 7323.
14127 	 */
14128 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14129 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14130 		way_out = 5;
14131 		retval = 0;
14132 		m_freem(m);
14133 		goto done_with_input;
14134 	}
14135 
14136 	/*
14137 	 * Segment received on connection. Reset idle time and keep-alive
14138 	 * timer. XXX: This should be done after segment validation to
14139 	 * ignore broken/spoofed segs.
14140 	 */
14141 	if  (tp->t_idle_reduce &&
14142 	     (tp->snd_max == tp->snd_una) &&
14143 	     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14144 		counter_u64_add(rack_input_idle_reduces, 1);
14145 		rack_cc_after_idle(rack, tp);
14146 	}
14147 	tp->t_rcvtime = ticks;
14148 #ifdef STATS
14149 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14150 #endif
14151 	if (tiwin > rack->r_ctl.rc_high_rwnd)
14152 		rack->r_ctl.rc_high_rwnd = tiwin;
14153 	/*
14154 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14155 	 * this to occur after we've validated the segment.
14156 	 */
14157 	if (tcp_ecn_input_segment(tp, thflags, iptos))
14158 		rack_cong_signal(tp, CC_ECN, th->th_ack);
14159 
14160 	/*
14161 	 * If echoed timestamp is later than the current time, fall back to
14162 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14163 	 * were used when this connection was established.
14164 	 */
14165 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14166 		to.to_tsecr -= tp->ts_offset;
14167 		if (TSTMP_GT(to.to_tsecr, ms_cts))
14168 			to.to_tsecr = 0;
14169 	}
14170 
14171 	/*
14172 	 * If its the first time in we need to take care of options and
14173 	 * verify we can do SACK for rack!
14174 	 */
14175 	if (rack->r_state == 0) {
14176 		/* Should be init'd by rack_init() */
14177 		KASSERT(rack->rc_inp != NULL,
14178 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
14179 		if (rack->rc_inp == NULL) {
14180 			rack->rc_inp = tp->t_inpcb;
14181 		}
14182 
14183 		/*
14184 		 * Process options only when we get SYN/ACK back. The SYN
14185 		 * case for incoming connections is handled in tcp_syncache.
14186 		 * According to RFC1323 the window field in a SYN (i.e., a
14187 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14188 		 * this is traditional behavior, may need to be cleaned up.
14189 		 */
14190 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14191 			/* Handle parallel SYN for ECN */
14192 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
14193 			if ((to.to_flags & TOF_SCALE) &&
14194 			    (tp->t_flags & TF_REQ_SCALE)) {
14195 				tp->t_flags |= TF_RCVD_SCALE;
14196 				tp->snd_scale = to.to_wscale;
14197 			} else
14198 				tp->t_flags &= ~TF_REQ_SCALE;
14199 			/*
14200 			 * Initial send window.  It will be updated with the
14201 			 * next incoming segment to the scaled value.
14202 			 */
14203 			tp->snd_wnd = th->th_win;
14204 			rack_validate_fo_sendwin_up(tp, rack);
14205 			if ((to.to_flags & TOF_TS) &&
14206 			    (tp->t_flags & TF_REQ_TSTMP)) {
14207 				tp->t_flags |= TF_RCVD_TSTMP;
14208 				tp->ts_recent = to.to_tsval;
14209 				tp->ts_recent_age = cts;
14210 			} else
14211 				tp->t_flags &= ~TF_REQ_TSTMP;
14212 			if (to.to_flags & TOF_MSS) {
14213 				tcp_mss(tp, to.to_mss);
14214 			}
14215 			if ((tp->t_flags & TF_SACK_PERMIT) &&
14216 			    (to.to_flags & TOF_SACKPERM) == 0)
14217 				tp->t_flags &= ~TF_SACK_PERMIT;
14218 			if (IS_FASTOPEN(tp->t_flags)) {
14219 				if (to.to_flags & TOF_FASTOPEN) {
14220 					uint16_t mss;
14221 
14222 					if (to.to_flags & TOF_MSS)
14223 						mss = to.to_mss;
14224 					else
14225 						if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
14226 							mss = TCP6_MSS;
14227 						else
14228 							mss = TCP_MSS;
14229 					tcp_fastopen_update_cache(tp, mss,
14230 					    to.to_tfo_len, to.to_tfo_cookie);
14231 				} else
14232 					tcp_fastopen_disable_path(tp);
14233 			}
14234 		}
14235 		/*
14236 		 * At this point we are at the initial call. Here we decide
14237 		 * if we are doing RACK or not. We do this by seeing if
14238 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
14239 		 * The code now does do dup-ack counting so if you don't
14240 		 * switch back you won't get rack & TLP, but you will still
14241 		 * get this stack.
14242 		 */
14243 
14244 		if ((rack_sack_not_required == 0) &&
14245 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14246 			tcp_switch_back_to_default(tp);
14247 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14248 			    tlen, iptos);
14249 #ifdef TCP_ACCOUNTING
14250 			sched_unpin();
14251 #endif
14252 			return (1);
14253 		}
14254 		tcp_set_hpts(tp->t_inpcb);
14255 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14256 	}
14257 	if (thflags & TH_FIN)
14258 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14259 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14260 	if ((rack->rc_gp_dyn_mul) &&
14261 	    (rack->use_fixed_rate == 0) &&
14262 	    (rack->rc_always_pace)) {
14263 		/* Check in on probertt */
14264 		rack_check_probe_rtt(rack, us_cts);
14265 	}
14266 	rack_clear_rate_sample(rack);
14267 	if ((rack->forced_ack) &&
14268 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
14269 		rack_handle_probe_response(rack, tiwin, us_cts);
14270 	}
14271 	/*
14272 	 * This is the one exception case where we set the rack state
14273 	 * always. All other times (timers etc) we must have a rack-state
14274 	 * set (so we assure we have done the checks above for SACK).
14275 	 */
14276 	rack->r_ctl.rc_rcvtime = cts;
14277 	if (rack->r_state != tp->t_state)
14278 		rack_set_state(tp, rack);
14279 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
14280 	    (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14281 		kern_prefetch(rsm, &prev_state);
14282 	prev_state = rack->r_state;
14283 	retval = (*rack->r_substate) (m, th, so,
14284 	    tp, &to, drop_hdrlen,
14285 	    tlen, tiwin, thflags, nxt_pkt, iptos);
14286 #ifdef INVARIANTS
14287 	if ((retval == 0) &&
14288 	    (tp->t_inpcb == NULL)) {
14289 		panic("retval:%d tp:%p t_inpcb:NULL state:%d",
14290 		    retval, tp, prev_state);
14291 	}
14292 #endif
14293 	if (retval == 0) {
14294 		/*
14295 		 * If retval is 1 the tcb is unlocked and most likely the tp
14296 		 * is gone.
14297 		 */
14298 		INP_WLOCK_ASSERT(tp->t_inpcb);
14299 		if ((rack->rc_gp_dyn_mul) &&
14300 		    (rack->rc_always_pace) &&
14301 		    (rack->use_fixed_rate == 0) &&
14302 		    rack->in_probe_rtt &&
14303 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
14304 			/*
14305 			 * If we are going for target, lets recheck before
14306 			 * we output.
14307 			 */
14308 			rack_check_probe_rtt(rack, us_cts);
14309 		}
14310 		if (rack->set_pacing_done_a_iw == 0) {
14311 			/* How much has been acked? */
14312 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14313 				/* We have enough to set in the pacing segment size */
14314 				rack->set_pacing_done_a_iw = 1;
14315 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
14316 			}
14317 		}
14318 		tcp_rack_xmit_timer_commit(rack, tp);
14319 #ifdef TCP_ACCOUNTING
14320 		/*
14321 		 * If we set the ack_val_se to what ack processing we are doing
14322 		 * we also want to track how many cycles we burned. Note
14323 		 * the bits after tcp_output we let be "free". This is because
14324 		 * we are also tracking the tcp_output times as well. Note the
14325 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
14326 		 * 0xf cannot be returned and is what we initialize it too to
14327 		 * indicate we are not doing the tabulations.
14328 		 */
14329 		if (ack_val_set != 0xf) {
14330 			uint64_t crtsc;
14331 
14332 			crtsc = get_cyclecount();
14333 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14334 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14335 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14336 			}
14337 		}
14338 #endif
14339 		if (nxt_pkt == 0) {
14340 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14341 do_output_now:
14342 				if (tcp_output(tp) < 0)
14343 					return (1);
14344 				did_out = 1;
14345 			}
14346 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14347 			rack_free_trim(rack);
14348 		}
14349 		/* Update any rounds needed */
14350 		if (rack_verbose_logging &&  (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
14351 			union tcp_log_stackspecific log;
14352 			struct timeval tv;
14353 
14354 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14355 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14356 			log.u_bbr.flex1 = high_seq;
14357 			log.u_bbr.flex2 = rack->r_ctl.roundends;
14358 			log.u_bbr.flex3 = rack->r_ctl.current_round;
14359 			log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
14360 			log.u_bbr.flex8 = 9;
14361 			tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
14362 				       0, &log, false, NULL, NULL, 0, &tv);
14363 		}
14364 		/*
14365 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
14366 		 * causes issues when we are just going app limited. Lets
14367 		 * instead use SEQ_GT <or> where its equal but more data
14368 		 * is outstanding.
14369 		 */
14370 		if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) ||
14371 		    ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) {
14372 			rack->r_ctl.current_round++;
14373 			rack->r_ctl.roundends = tp->snd_max;
14374 			if (CC_ALGO(tp)->newround != NULL) {
14375 				CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
14376 			}
14377 		}
14378 		if ((nxt_pkt == 0) &&
14379 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14380 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
14381 		     (tp->t_flags & TF_DELACK) ||
14382 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14383 		      (tp->t_state <= TCPS_CLOSING)))) {
14384 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
14385 			if ((tp->snd_max == tp->snd_una) &&
14386 			    ((tp->t_flags & TF_DELACK) == 0) &&
14387 			    (tcp_in_hpts(rack->rc_inp)) &&
14388 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14389 				/* keep alive not needed if we are hptsi output yet */
14390 				;
14391 			} else {
14392 				int late = 0;
14393 				if (tcp_in_hpts(rack->rc_inp)) {
14394 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14395 						us_cts = tcp_get_usecs(NULL);
14396 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14397 							rack->r_early = 1;
14398 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14399 						} else
14400 							late = 1;
14401 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14402 					}
14403 					tcp_hpts_remove(tp->t_inpcb);
14404 				}
14405 				if (late && (did_out == 0)) {
14406 					/*
14407 					 * We are late in the sending
14408 					 * and we did not call the output
14409 					 * (this probably should not happen).
14410 					 */
14411 					goto do_output_now;
14412 				}
14413 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14414 			}
14415 			way_out = 1;
14416 		} else if (nxt_pkt == 0) {
14417 			/* Do we have the correct timer running? */
14418 			rack_timer_audit(tp, rack, &so->so_snd);
14419 			way_out = 2;
14420 		}
14421 	done_with_input:
14422 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14423 		if (did_out)
14424 			rack->r_wanted_output = 0;
14425 #ifdef INVARIANTS
14426 		if (tp->t_inpcb == NULL) {
14427 			panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
14428 			      did_out,
14429 			      retval, tp, prev_state);
14430 		}
14431 #endif
14432 #ifdef TCP_ACCOUNTING
14433 	} else {
14434 		/*
14435 		 * Track the time (see above).
14436 		 */
14437 		if (ack_val_set != 0xf) {
14438 			uint64_t crtsc;
14439 
14440 			crtsc = get_cyclecount();
14441 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14442 			/*
14443 			 * Note we *DO NOT* increment the per-tcb counters since
14444 			 * in the else the TP may be gone!!
14445 			 */
14446 		}
14447 #endif
14448 	}
14449 #ifdef TCP_ACCOUNTING
14450 	sched_unpin();
14451 #endif
14452 	return (retval);
14453 }
14454 
14455 void
14456 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14457     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14458 {
14459 	struct timeval tv;
14460 
14461 	/* First lets see if we have old packets */
14462 	if (tp->t_in_pkt) {
14463 		if (ctf_do_queued_segments(so, tp, 1)) {
14464 			m_freem(m);
14465 			return;
14466 		}
14467 	}
14468 	if (m->m_flags & M_TSTMP_LRO) {
14469 		tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
14470 		tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
14471 	} else {
14472 		/* Should not be should we kassert instead? */
14473 		tcp_get_usecs(&tv);
14474 	}
14475 	if (rack_do_segment_nounlock(m, th, so, tp,
14476 				     drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14477 		INP_WUNLOCK(tp->t_inpcb);
14478 	}
14479 }
14480 
14481 struct rack_sendmap *
14482 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14483 {
14484 	struct rack_sendmap *rsm = NULL;
14485 	int32_t idx;
14486 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
14487 
14488 	/* Return the next guy to be re-transmitted */
14489 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14490 		return (NULL);
14491 	}
14492 	if (tp->t_flags & TF_SENTFIN) {
14493 		/* retran the end FIN? */
14494 		return (NULL);
14495 	}
14496 	/* ok lets look at this one */
14497 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14498 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14499 		goto check_it;
14500 	}
14501 	rsm = rack_find_lowest_rsm(rack);
14502 	if (rsm == NULL) {
14503 		return (NULL);
14504 	}
14505 check_it:
14506 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14507 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14508 		/*
14509 		 * No sack so we automatically do the 3 strikes and
14510 		 * retransmit (no rack timer would be started).
14511 		 */
14512 
14513 		return (rsm);
14514 	}
14515 	if (rsm->r_flags & RACK_ACKED) {
14516 		return (NULL);
14517 	}
14518 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14519 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14520 		/* Its not yet ready */
14521 		return (NULL);
14522 	}
14523 	srtt = rack_grab_rtt(tp, rack);
14524 	idx = rsm->r_rtr_cnt - 1;
14525 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14526 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14527 	if ((tsused == ts_low) ||
14528 	    (TSTMP_LT(tsused, ts_low))) {
14529 		/* No time since sending */
14530 		return (NULL);
14531 	}
14532 	if ((tsused - ts_low) < thresh) {
14533 		/* It has not been long enough yet */
14534 		return (NULL);
14535 	}
14536 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14537 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
14538 	     (rack->sack_attack_disable == 0))) {
14539 		/*
14540 		 * We have passed the dup-ack threshold <or>
14541 		 * a SACK has indicated this is missing.
14542 		 * Note that if you are a declared attacker
14543 		 * it is only the dup-ack threshold that
14544 		 * will cause retransmits.
14545 		 */
14546 		/* log retransmit reason */
14547 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14548 		rack->r_fast_output = 0;
14549 		return (rsm);
14550 	}
14551 	return (NULL);
14552 }
14553 
14554 static void
14555 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14556 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14557 			   int line, struct rack_sendmap *rsm, uint8_t quality)
14558 {
14559 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14560 		union tcp_log_stackspecific log;
14561 		struct timeval tv;
14562 
14563 		memset(&log, 0, sizeof(log));
14564 		log.u_bbr.flex1 = slot;
14565 		log.u_bbr.flex2 = len;
14566 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14567 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14568 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14569 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14570 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14571 		log.u_bbr.use_lt_bw <<= 1;
14572 		log.u_bbr.use_lt_bw |= rack->r_late;
14573 		log.u_bbr.use_lt_bw <<= 1;
14574 		log.u_bbr.use_lt_bw |= rack->r_early;
14575 		log.u_bbr.use_lt_bw <<= 1;
14576 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14577 		log.u_bbr.use_lt_bw <<= 1;
14578 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14579 		log.u_bbr.use_lt_bw <<= 1;
14580 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14581 		log.u_bbr.use_lt_bw <<= 1;
14582 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14583 		log.u_bbr.use_lt_bw <<= 1;
14584 		log.u_bbr.use_lt_bw |= rack->gp_ready;
14585 		log.u_bbr.pkt_epoch = line;
14586 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14587 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14588 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14589 		log.u_bbr.bw_inuse = bw_est;
14590 		log.u_bbr.delRate = bw;
14591 		if (rack->r_ctl.gp_bw == 0)
14592 			log.u_bbr.cur_del_rate = 0;
14593 		else
14594 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
14595 		log.u_bbr.rttProp = len_time;
14596 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14597 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14598 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14599 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14600 			/* We are in slow start */
14601 			log.u_bbr.flex7 = 1;
14602 		} else {
14603 			/* we are on congestion avoidance */
14604 			log.u_bbr.flex7 = 0;
14605 		}
14606 		log.u_bbr.flex8 = method;
14607 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14608 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14609 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14610 		log.u_bbr.cwnd_gain <<= 1;
14611 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14612 		log.u_bbr.cwnd_gain <<= 1;
14613 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14614 		log.u_bbr.bbr_substate = quality;
14615 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
14616 		    &rack->rc_inp->inp_socket->so_rcv,
14617 		    &rack->rc_inp->inp_socket->so_snd,
14618 		    BBR_LOG_HPTSI_CALC, 0,
14619 		    0, &log, false, &tv);
14620 	}
14621 }
14622 
14623 static uint32_t
14624 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14625 {
14626 	uint32_t new_tso, user_max;
14627 
14628 	user_max = rack->rc_user_set_max_segs * mss;
14629 	if (rack->rc_force_max_seg) {
14630 		return (user_max);
14631 	}
14632 	if (rack->use_fixed_rate &&
14633 	    ((rack->r_ctl.crte == NULL) ||
14634 	     (bw != rack->r_ctl.crte->rate))) {
14635 		/* Use the user mss since we are not exactly matched */
14636 		return (user_max);
14637 	}
14638 	new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14639 	if (new_tso > user_max)
14640 		new_tso = user_max;
14641 	return (new_tso);
14642 }
14643 
14644 static int32_t
14645 pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced)
14646 {
14647 	uint64_t lentim, fill_bw;
14648 
14649 	/* Lets first see if we are full, if so continue with normal rate */
14650 	rack->r_via_fill_cw = 0;
14651 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14652 		return (slot);
14653 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14654 		return (slot);
14655 	if (rack->r_ctl.rc_last_us_rtt == 0)
14656 		return (slot);
14657 	if (rack->rc_pace_fill_if_rttin_range &&
14658 	    (rack->r_ctl.rc_last_us_rtt >=
14659 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14660 		/* The rtt is huge, N * smallest, lets not fill */
14661 		return (slot);
14662 	}
14663 	/*
14664 	 * first lets calculate the b/w based on the last us-rtt
14665 	 * and the sndwnd.
14666 	 */
14667 	fill_bw = rack->r_ctl.cwnd_to_use;
14668 	/* Take the rwnd if its smaller */
14669 	if (fill_bw > rack->rc_tp->snd_wnd)
14670 		fill_bw = rack->rc_tp->snd_wnd;
14671 	if (rack->r_fill_less_agg) {
14672 		/*
14673 		 * Now take away the inflight (this will reduce our
14674 		 * aggressiveness and yeah, if we get that much out in 1RTT
14675 		 * we will have had acks come back and still be behind).
14676 		 */
14677 		fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14678 	}
14679 	/* Now lets make it into a b/w */
14680 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14681 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14682 	/* We are below the min b/w */
14683 	if (non_paced)
14684 		*rate_wanted = fill_bw;
14685 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14686 		return (slot);
14687 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14688 		fill_bw = rack->r_ctl.bw_rate_cap;
14689 	rack->r_via_fill_cw = 1;
14690 	if (rack->r_rack_hw_rate_caps &&
14691 	    (rack->r_ctl.crte != NULL)) {
14692 		uint64_t high_rate;
14693 
14694 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14695 		if (fill_bw > high_rate) {
14696 			/* We are capping bw at the highest rate table entry */
14697 			if (*rate_wanted > high_rate) {
14698 				/* The original rate was also capped */
14699 				rack->r_via_fill_cw = 0;
14700 			}
14701 			rack_log_hdwr_pacing(rack,
14702 					     fill_bw, high_rate, __LINE__,
14703 					     0, 3);
14704 			fill_bw = high_rate;
14705 			if (capped)
14706 				*capped = 1;
14707 		}
14708 	} else if ((rack->r_ctl.crte == NULL) &&
14709 		   (rack->rack_hdrw_pacing == 0) &&
14710 		   (rack->rack_hdw_pace_ena) &&
14711 		   rack->r_rack_hw_rate_caps &&
14712 		   (rack->rack_attempt_hdwr_pace == 0) &&
14713 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
14714 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14715 		/*
14716 		 * Ok we may have a first attempt that is greater than our top rate
14717 		 * lets check.
14718 		 */
14719 		uint64_t high_rate;
14720 
14721 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14722 		if (high_rate) {
14723 			if (fill_bw > high_rate) {
14724 				fill_bw = high_rate;
14725 				if (capped)
14726 					*capped = 1;
14727 			}
14728 		}
14729 	}
14730 	/*
14731 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
14732 	 * in a rtt, what does that time wise equate too?
14733 	 */
14734 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14735 	lentim /= fill_bw;
14736 	*rate_wanted = fill_bw;
14737 	if (non_paced || (lentim < slot)) {
14738 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14739 					   0, lentim, 12, __LINE__, NULL, 0);
14740 		return ((int32_t)lentim);
14741 	} else
14742 		return (slot);
14743 }
14744 
14745 static int32_t
14746 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14747 {
14748 	uint64_t srtt;
14749 	int32_t slot = 0;
14750 	int can_start_hw_pacing = 1;
14751 	int err;
14752 
14753 	if (rack->rc_always_pace == 0) {
14754 		/*
14755 		 * We use the most optimistic possible cwnd/srtt for
14756 		 * sending calculations. This will make our
14757 		 * calculation anticipate getting more through
14758 		 * quicker then possible. But thats ok we don't want
14759 		 * the peer to have a gap in data sending.
14760 		 */
14761 		uint64_t cwnd, tr_perms = 0;
14762 		int32_t reduce = 0;
14763 
14764 	old_method:
14765 		/*
14766 		 * We keep no precise pacing with the old method
14767 		 * instead we use the pacer to mitigate bursts.
14768 		 */
14769 		if (rack->r_ctl.rc_rack_min_rtt)
14770 			srtt = rack->r_ctl.rc_rack_min_rtt;
14771 		else
14772 			srtt = max(tp->t_srtt, 1);
14773 		if (rack->r_ctl.rc_rack_largest_cwnd)
14774 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14775 		else
14776 			cwnd = rack->r_ctl.cwnd_to_use;
14777 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
14778 		tr_perms = (cwnd * 1000) / srtt;
14779 		if (tr_perms == 0) {
14780 			tr_perms = ctf_fixed_maxseg(tp);
14781 		}
14782 		/*
14783 		 * Calculate how long this will take to drain, if
14784 		 * the calculation comes out to zero, thats ok we
14785 		 * will use send_a_lot to possibly spin around for
14786 		 * more increasing tot_len_this_send to the point
14787 		 * that its going to require a pace, or we hit the
14788 		 * cwnd. Which in that case we are just waiting for
14789 		 * a ACK.
14790 		 */
14791 		slot = len / tr_perms;
14792 		/* Now do we reduce the time so we don't run dry? */
14793 		if (slot && rack_slot_reduction) {
14794 			reduce = (slot / rack_slot_reduction);
14795 			if (reduce < slot) {
14796 				slot -= reduce;
14797 			} else
14798 				slot = 0;
14799 		}
14800 		slot *= HPTS_USEC_IN_MSEC;
14801 		if (rack->rc_pace_to_cwnd) {
14802 			uint64_t rate_wanted = 0;
14803 
14804 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14805 			rack->rc_ack_can_sendout_data = 1;
14806 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14807 		} else
14808 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14809 	} else {
14810 		uint64_t bw_est, res, lentim, rate_wanted;
14811 		uint32_t orig_val, segs, oh;
14812 		int capped = 0;
14813 		int prev_fill;
14814 
14815 		if ((rack->r_rr_config == 1) && rsm) {
14816 			return (rack->r_ctl.rc_min_to);
14817 		}
14818 		if (rack->use_fixed_rate) {
14819 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14820 		} else if ((rack->r_ctl.init_rate == 0) &&
14821 #ifdef NETFLIX_PEAKRATE
14822 			   (rack->rc_tp->t_maxpeakrate == 0) &&
14823 #endif
14824 			   (rack->r_ctl.gp_bw == 0)) {
14825 			/* no way to yet do an estimate */
14826 			bw_est = rate_wanted = 0;
14827 		} else {
14828 			bw_est = rack_get_bw(rack);
14829 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14830 		}
14831 		if ((bw_est == 0) || (rate_wanted == 0) ||
14832 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14833 			/*
14834 			 * No way yet to make a b/w estimate or
14835 			 * our raise is set incorrectly.
14836 			 */
14837 			goto old_method;
14838 		}
14839 		/* We need to account for all the overheads */
14840 		segs = (len + segsiz - 1) / segsiz;
14841 		/*
14842 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
14843 		 * and how much data we put in each packet. Yes this
14844 		 * means we may be off if we are larger than 1500 bytes
14845 		 * or smaller. But this just makes us more conservative.
14846 		 */
14847 		if (rack_hw_rate_min &&
14848 		    (bw_est < rack_hw_rate_min))
14849 			can_start_hw_pacing = 0;
14850 		if (ETHERNET_SEGMENT_SIZE > segsiz)
14851 			oh = ETHERNET_SEGMENT_SIZE - segsiz;
14852 		else
14853 			oh = 0;
14854 		segs *= oh;
14855 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14856 		res = lentim / rate_wanted;
14857 		slot = (uint32_t)res;
14858 		orig_val = rack->r_ctl.rc_pace_max_segs;
14859 		if (rack->r_ctl.crte == NULL) {
14860 			/*
14861 			 * Only do this if we are not hardware pacing
14862 			 * since if we are doing hw-pacing below we will
14863 			 * set make a call after setting up or changing
14864 			 * the rate.
14865 			 */
14866 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14867 		} else if (rack->rc_inp->inp_snd_tag == NULL) {
14868 			/*
14869 			 * We lost our rate somehow, this can happen
14870 			 * if the interface changed underneath us.
14871 			 */
14872 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14873 			rack->r_ctl.crte = NULL;
14874 			/* Lets re-allow attempting to setup pacing */
14875 			rack->rack_hdrw_pacing = 0;
14876 			rack->rack_attempt_hdwr_pace = 0;
14877 			rack_log_hdwr_pacing(rack,
14878 					     rate_wanted, bw_est, __LINE__,
14879 					     0, 6);
14880 		}
14881 		/* Did we change the TSO size, if so log it */
14882 		if (rack->r_ctl.rc_pace_max_segs != orig_val)
14883 			rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
14884 		prev_fill = rack->r_via_fill_cw;
14885 		if ((rack->rc_pace_to_cwnd) &&
14886 		    (capped == 0) &&
14887 		    (rack->use_fixed_rate == 0) &&
14888 		    (rack->in_probe_rtt == 0) &&
14889 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14890 			/*
14891 			 * We want to pace at our rate *or* faster to
14892 			 * fill the cwnd to the max if its not full.
14893 			 */
14894 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14895 		}
14896 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14897 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14898 			if ((rack->rack_hdw_pace_ena) &&
14899 			    (can_start_hw_pacing > 0) &&
14900 			    (rack->rack_hdrw_pacing == 0) &&
14901 			    (rack->rack_attempt_hdwr_pace == 0)) {
14902 				/*
14903 				 * Lets attempt to turn on hardware pacing
14904 				 * if we can.
14905 				 */
14906 				rack->rack_attempt_hdwr_pace = 1;
14907 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14908 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
14909 								       rate_wanted,
14910 								       RS_PACING_GEQ,
14911 								       &err, &rack->r_ctl.crte_prev_rate);
14912 				if (rack->r_ctl.crte) {
14913 					rack->rack_hdrw_pacing = 1;
14914 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14915 												 0, rack->r_ctl.crte,
14916 												 NULL);
14917 					rack_log_hdwr_pacing(rack,
14918 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14919 							     err, 0);
14920 					rack->r_ctl.last_hw_bw_req = rate_wanted;
14921 				} else {
14922 					counter_u64_add(rack_hw_pace_init_fail, 1);
14923 				}
14924 			} else if (rack->rack_hdrw_pacing &&
14925 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14926 				/* Do we need to adjust our rate? */
14927 				const struct tcp_hwrate_limit_table *nrte;
14928 
14929 				if (rack->r_up_only &&
14930 				    (rate_wanted < rack->r_ctl.crte->rate)) {
14931 					/**
14932 					 * We have four possible states here
14933 					 * having to do with the previous time
14934 					 * and this time.
14935 					 *   previous  |  this-time
14936 					 * A)     0      |     0   -- fill_cw not in the picture
14937 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
14938 					 * C)     1      |     1   -- all rates from fill_cw
14939 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
14940 					 *
14941 					 * For case A, C and D we don't allow a drop. But for
14942 					 * case B where we now our on our steady rate we do
14943 					 * allow a drop.
14944 					 *
14945 					 */
14946 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14947 						goto done_w_hdwr;
14948 				}
14949 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
14950 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14951 					if (rack_hw_rate_to_low &&
14952 					    (bw_est < rack_hw_rate_to_low)) {
14953 						/*
14954 						 * The pacing rate is too low for hardware, but
14955 						 * do allow hardware pacing to be restarted.
14956 						 */
14957 						rack_log_hdwr_pacing(rack,
14958 							     bw_est, rack->r_ctl.crte->rate, __LINE__,
14959 							     0, 5);
14960 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14961 						rack->r_ctl.crte = NULL;
14962 						rack->rack_attempt_hdwr_pace = 0;
14963 						rack->rack_hdrw_pacing = 0;
14964 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14965 						goto done_w_hdwr;
14966 					}
14967 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
14968 								   rack->rc_tp,
14969 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
14970 								   rate_wanted,
14971 								   RS_PACING_GEQ,
14972 								   &err, &rack->r_ctl.crte_prev_rate);
14973 					if (nrte == NULL) {
14974 						/* Lost the rate */
14975 						rack->rack_hdrw_pacing = 0;
14976 						rack->r_ctl.crte = NULL;
14977 						rack_log_hdwr_pacing(rack,
14978 								     rate_wanted, 0, __LINE__,
14979 								     err, 1);
14980 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14981 						counter_u64_add(rack_hw_pace_lost, 1);
14982 					} else if (nrte != rack->r_ctl.crte) {
14983 						rack->r_ctl.crte = nrte;
14984 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
14985 													 segsiz, 0,
14986 													 rack->r_ctl.crte,
14987 													 NULL);
14988 						rack_log_hdwr_pacing(rack,
14989 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14990 								     err, 2);
14991 						rack->r_ctl.last_hw_bw_req = rate_wanted;
14992 					}
14993 				} else {
14994 					/* We just need to adjust the segment size */
14995 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14996 					rack_log_hdwr_pacing(rack,
14997 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14998 							     0, 4);
14999 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15000 				}
15001 			}
15002 		}
15003 		if ((rack->r_ctl.crte != NULL) &&
15004 		    (rack->r_ctl.crte->rate == rate_wanted)) {
15005 			/*
15006 			 * We need to add a extra if the rates
15007 			 * are exactly matched. The idea is
15008 			 * we want the software to make sure the
15009 			 * queue is empty before adding more, this
15010 			 * gives us N MSS extra pace times where
15011 			 * N is our sysctl
15012 			 */
15013 			slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15014 		}
15015 done_w_hdwr:
15016 		if (rack_limit_time_with_srtt &&
15017 		    (rack->use_fixed_rate == 0) &&
15018 #ifdef NETFLIX_PEAKRATE
15019 		    (rack->rc_tp->t_maxpeakrate == 0) &&
15020 #endif
15021 		    (rack->rack_hdrw_pacing == 0)) {
15022 			/*
15023 			 * Sanity check, we do not allow the pacing delay
15024 			 * to be longer than the SRTT of the path. If it is
15025 			 * a slow path, then adding a packet should increase
15026 			 * the RTT and compensate for this i.e. the srtt will
15027 			 * be greater so the allowed pacing time will be greater.
15028 			 *
15029 			 * Note this restriction is not for where a peak rate
15030 			 * is set, we are doing fixed pacing or hardware pacing.
15031 			 */
15032 			if (rack->rc_tp->t_srtt)
15033 				srtt = rack->rc_tp->t_srtt;
15034 			else
15035 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
15036 			if (srtt < (uint64_t)slot) {
15037 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15038 				slot = srtt;
15039 			}
15040 		}
15041 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15042 	}
15043 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15044 		/*
15045 		 * If this rate is seeing enobufs when it
15046 		 * goes to send then either the nic is out
15047 		 * of gas or we are mis-estimating the time
15048 		 * somehow and not letting the queue empty
15049 		 * completely. Lets add to the pacing time.
15050 		 */
15051 		int hw_boost_delay;
15052 
15053 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15054 		if (hw_boost_delay > rack_enobuf_hw_max)
15055 			hw_boost_delay = rack_enobuf_hw_max;
15056 		else if (hw_boost_delay < rack_enobuf_hw_min)
15057 			hw_boost_delay = rack_enobuf_hw_min;
15058 		slot += hw_boost_delay;
15059 	}
15060 	return (slot);
15061 }
15062 
15063 static void
15064 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15065     tcp_seq startseq, uint32_t sb_offset)
15066 {
15067 	struct rack_sendmap *my_rsm = NULL;
15068 	struct rack_sendmap fe;
15069 
15070 	if (tp->t_state < TCPS_ESTABLISHED) {
15071 		/*
15072 		 * We don't start any measurements if we are
15073 		 * not at least established.
15074 		 */
15075 		return;
15076 	}
15077 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
15078 		/*
15079 		 * We will get no more data into the SB
15080 		 * this means we need to have the data available
15081 		 * before we start a measurement.
15082 		 */
15083 
15084 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) <
15085 		    max(rc_init_window(rack),
15086 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15087 			/* Nope not enough data */
15088 			return;
15089 		}
15090 	}
15091 	tp->t_flags |= TF_GPUTINPROG;
15092 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15093 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15094 	tp->gput_seq = startseq;
15095 	rack->app_limited_needs_set = 0;
15096 	if (rack->in_probe_rtt)
15097 		rack->measure_saw_probe_rtt = 1;
15098 	else if ((rack->measure_saw_probe_rtt) &&
15099 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15100 		rack->measure_saw_probe_rtt = 0;
15101 	if (rack->rc_gp_filled)
15102 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15103 	else {
15104 		/* Special case initial measurement */
15105 		struct timeval tv;
15106 
15107 		tp->gput_ts = tcp_get_usecs(&tv);
15108 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15109 	}
15110 	/*
15111 	 * We take a guess out into the future,
15112 	 * if we have no measurement and no
15113 	 * initial rate, we measure the first
15114 	 * initial-windows worth of data to
15115 	 * speed up getting some GP measurement and
15116 	 * thus start pacing.
15117 	 */
15118 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15119 		rack->app_limited_needs_set = 1;
15120 		tp->gput_ack = startseq + max(rc_init_window(rack),
15121 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15122 		rack_log_pacing_delay_calc(rack,
15123 					   tp->gput_seq,
15124 					   tp->gput_ack,
15125 					   0,
15126 					   tp->gput_ts,
15127 					   rack->r_ctl.rc_app_limited_cnt,
15128 					   9,
15129 					   __LINE__, NULL, 0);
15130 		return;
15131 	}
15132 	if (sb_offset) {
15133 		/*
15134 		 * We are out somewhere in the sb
15135 		 * can we use the already outstanding data?
15136 		 */
15137 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
15138 			/*
15139 			 * Yes first one is good and in this case
15140 			 * the tp->gput_ts is correctly set based on
15141 			 * the last ack that arrived (no need to
15142 			 * set things up when an ack comes in).
15143 			 */
15144 			my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15145 			if ((my_rsm == NULL) ||
15146 			    (my_rsm->r_rtr_cnt != 1)) {
15147 				/* retransmission? */
15148 				goto use_latest;
15149 			}
15150 		} else {
15151 			if (rack->r_ctl.rc_first_appl == NULL) {
15152 				/*
15153 				 * If rc_first_appl is NULL
15154 				 * then the cnt should be 0.
15155 				 * This is probably an error, maybe
15156 				 * a KASSERT would be approprate.
15157 				 */
15158 				goto use_latest;
15159 			}
15160 			/*
15161 			 * If we have a marker pointer to the last one that is
15162 			 * app limited we can use that, but we need to set
15163 			 * things up so that when it gets ack'ed we record
15164 			 * the ack time (if its not already acked).
15165 			 */
15166 			rack->app_limited_needs_set = 1;
15167 			/*
15168 			 * We want to get to the rsm that is either
15169 			 * next with space i.e. over 1 MSS or the one
15170 			 * after that (after the app-limited).
15171 			 */
15172 			my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15173 					 rack->r_ctl.rc_first_appl);
15174 			if (my_rsm) {
15175 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15176 					/* Have to use the next one */
15177 					my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15178 							 my_rsm);
15179 				else {
15180 					/* Use after the first MSS of it is acked */
15181 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15182 					goto start_set;
15183 				}
15184 			}
15185 			if ((my_rsm == NULL) ||
15186 			    (my_rsm->r_rtr_cnt != 1)) {
15187 				/*
15188 				 * Either its a retransmit or
15189 				 * the last is the app-limited one.
15190 				 */
15191 				goto use_latest;
15192 			}
15193 		}
15194 		tp->gput_seq = my_rsm->r_start;
15195 start_set:
15196 		if (my_rsm->r_flags & RACK_ACKED) {
15197 			/*
15198 			 * This one has been acked use the arrival ack time
15199 			 */
15200 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15201 			rack->app_limited_needs_set = 0;
15202 		}
15203 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15204 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15205 		rack_log_pacing_delay_calc(rack,
15206 					   tp->gput_seq,
15207 					   tp->gput_ack,
15208 					   (uint64_t)my_rsm,
15209 					   tp->gput_ts,
15210 					   rack->r_ctl.rc_app_limited_cnt,
15211 					   9,
15212 					   __LINE__, NULL, 0);
15213 		return;
15214 	}
15215 
15216 use_latest:
15217 	/*
15218 	 * We don't know how long we may have been
15219 	 * idle or if this is the first-send. Lets
15220 	 * setup the flag so we will trim off
15221 	 * the first ack'd data so we get a true
15222 	 * measurement.
15223 	 */
15224 	rack->app_limited_needs_set = 1;
15225 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15226 	/* Find this guy so we can pull the send time */
15227 	fe.r_start = startseq;
15228 	my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15229 	if (my_rsm) {
15230 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15231 		if (my_rsm->r_flags & RACK_ACKED) {
15232 			/*
15233 			 * Unlikely since its probably what was
15234 			 * just transmitted (but I am paranoid).
15235 			 */
15236 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15237 			rack->app_limited_needs_set = 0;
15238 		}
15239 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15240 			/* This also is unlikely */
15241 			tp->gput_seq = my_rsm->r_start;
15242 		}
15243 	} else {
15244 		/*
15245 		 * TSNH unless we have some send-map limit,
15246 		 * and even at that it should not be hitting
15247 		 * that limit (we should have stopped sending).
15248 		 */
15249 		struct timeval tv;
15250 
15251 		microuptime(&tv);
15252 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15253 	}
15254 	rack_log_pacing_delay_calc(rack,
15255 				   tp->gput_seq,
15256 				   tp->gput_ack,
15257 				   (uint64_t)my_rsm,
15258 				   tp->gput_ts,
15259 				   rack->r_ctl.rc_app_limited_cnt,
15260 				   9, __LINE__, NULL, 0);
15261 }
15262 
15263 static inline uint32_t
15264 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15265     uint32_t avail, int32_t sb_offset)
15266 {
15267 	uint32_t len;
15268 	uint32_t sendwin;
15269 
15270 	if (tp->snd_wnd > cwnd_to_use)
15271 		sendwin = cwnd_to_use;
15272 	else
15273 		sendwin = tp->snd_wnd;
15274 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
15275 		/* We never want to go over our peers rcv-window */
15276 		len = 0;
15277 	} else {
15278 		uint32_t flight;
15279 
15280 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15281 		if (flight >= sendwin) {
15282 			/*
15283 			 * We have in flight what we are allowed by cwnd (if
15284 			 * it was rwnd blocking it would have hit above out
15285 			 * >= tp->snd_wnd).
15286 			 */
15287 			return (0);
15288 		}
15289 		len = sendwin - flight;
15290 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15291 			/* We would send too much (beyond the rwnd) */
15292 			len = tp->snd_wnd - ctf_outstanding(tp);
15293 		}
15294 		if ((len + sb_offset) > avail) {
15295 			/*
15296 			 * We don't have that much in the SB, how much is
15297 			 * there?
15298 			 */
15299 			len = avail - sb_offset;
15300 		}
15301 	}
15302 	return (len);
15303 }
15304 
15305 static void
15306 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15307 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15308 	     int rsm_is_null, int optlen, int line, uint16_t mode)
15309 {
15310 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15311 		union tcp_log_stackspecific log;
15312 		struct timeval tv;
15313 
15314 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15315 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15316 		log.u_bbr.flex1 = error;
15317 		log.u_bbr.flex2 = flags;
15318 		log.u_bbr.flex3 = rsm_is_null;
15319 		log.u_bbr.flex4 = ipoptlen;
15320 		log.u_bbr.flex5 = tp->rcv_numsacks;
15321 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15322 		log.u_bbr.flex7 = optlen;
15323 		log.u_bbr.flex8 = rack->r_fsb_inited;
15324 		log.u_bbr.applimited = rack->r_fast_output;
15325 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15326 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15327 		log.u_bbr.cwnd_gain = mode;
15328 		log.u_bbr.pkts_out = orig_len;
15329 		log.u_bbr.lt_epoch = len;
15330 		log.u_bbr.delivered = line;
15331 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15332 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15333 		tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15334 			       len, &log, false, NULL, NULL, 0, &tv);
15335 	}
15336 }
15337 
15338 
15339 static struct mbuf *
15340 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15341 		   struct rack_fast_send_blk *fsb,
15342 		   int32_t seglimit, int32_t segsize, int hw_tls)
15343 {
15344 #ifdef KERN_TLS
15345 	struct ktls_session *tls, *ntls;
15346 #ifdef INVARIANTS
15347 	struct mbuf *start;
15348 #endif
15349 #endif
15350 	struct mbuf *m, *n, **np, *smb;
15351 	struct mbuf *top;
15352 	int32_t off, soff;
15353 	int32_t len = *plen;
15354 	int32_t fragsize;
15355 	int32_t len_cp = 0;
15356 	uint32_t mlen, frags;
15357 
15358 	soff = off = the_off;
15359 	smb = m = the_m;
15360 	np = &top;
15361 	top = NULL;
15362 #ifdef KERN_TLS
15363 	if (hw_tls && (m->m_flags & M_EXTPG))
15364 		tls = m->m_epg_tls;
15365 	else
15366 		tls = NULL;
15367 #ifdef INVARIANTS
15368 	start = m;
15369 #endif
15370 #endif
15371 	while (len > 0) {
15372 		if (m == NULL) {
15373 			*plen = len_cp;
15374 			break;
15375 		}
15376 #ifdef KERN_TLS
15377 		if (hw_tls) {
15378 			if (m->m_flags & M_EXTPG)
15379 				ntls = m->m_epg_tls;
15380 			else
15381 				ntls = NULL;
15382 
15383 			/*
15384 			 * Avoid mixing TLS records with handshake
15385 			 * data or TLS records from different
15386 			 * sessions.
15387 			 */
15388 			if (tls != ntls) {
15389 				MPASS(m != start);
15390 				*plen = len_cp;
15391 				break;
15392 			}
15393 		}
15394 #endif
15395 		mlen = min(len, m->m_len - off);
15396 		if (seglimit) {
15397 			/*
15398 			 * For M_EXTPG mbufs, add 3 segments
15399 			 * + 1 in case we are crossing page boundaries
15400 			 * + 2 in case the TLS hdr/trailer are used
15401 			 * It is cheaper to just add the segments
15402 			 * than it is to take the cache miss to look
15403 			 * at the mbuf ext_pgs state in detail.
15404 			 */
15405 			if (m->m_flags & M_EXTPG) {
15406 				fragsize = min(segsize, PAGE_SIZE);
15407 				frags = 3;
15408 			} else {
15409 				fragsize = segsize;
15410 				frags = 0;
15411 			}
15412 
15413 			/* Break if we really can't fit anymore. */
15414 			if ((frags + 1) >= seglimit) {
15415 				*plen =	len_cp;
15416 				break;
15417 			}
15418 
15419 			/*
15420 			 * Reduce size if you can't copy the whole
15421 			 * mbuf. If we can't copy the whole mbuf, also
15422 			 * adjust len so the loop will end after this
15423 			 * mbuf.
15424 			 */
15425 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15426 				mlen = (seglimit - frags - 1) * fragsize;
15427 				len = mlen;
15428 				*plen = len_cp + len;
15429 			}
15430 			frags += howmany(mlen, fragsize);
15431 			if (frags == 0)
15432 				frags++;
15433 			seglimit -= frags;
15434 			KASSERT(seglimit > 0,
15435 			    ("%s: seglimit went too low", __func__));
15436 		}
15437 		n = m_get(M_NOWAIT, m->m_type);
15438 		*np = n;
15439 		if (n == NULL)
15440 			goto nospace;
15441 		n->m_len = mlen;
15442 		soff += mlen;
15443 		len_cp += n->m_len;
15444 		if (m->m_flags & (M_EXT|M_EXTPG)) {
15445 			n->m_data = m->m_data + off;
15446 			mb_dupcl(n, m);
15447 		} else {
15448 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15449 			    (u_int)n->m_len);
15450 		}
15451 		len -= n->m_len;
15452 		off = 0;
15453 		m = m->m_next;
15454 		np = &n->m_next;
15455 		if (len || (soff == smb->m_len)) {
15456 			/*
15457 			 * We have more so we move forward  or
15458 			 * we have consumed the entire mbuf and
15459 			 * len has fell to 0.
15460 			 */
15461 			soff = 0;
15462 			smb = m;
15463 		}
15464 
15465 	}
15466 	if (fsb != NULL) {
15467 		fsb->m = smb;
15468 		fsb->off = soff;
15469 		if (smb) {
15470 			/*
15471 			 * Save off the size of the mbuf. We do
15472 			 * this so that we can recognize when it
15473 			 * has been trimmed by sbcut() as acks
15474 			 * come in.
15475 			 */
15476 			fsb->o_m_len = smb->m_len;
15477 		} else {
15478 			/*
15479 			 * This is the case where the next mbuf went to NULL. This
15480 			 * means with this copy we have sent everything in the sb.
15481 			 * In theory we could clear the fast_output flag, but lets
15482 			 * not since its possible that we could get more added
15483 			 * and acks that call the extend function which would let
15484 			 * us send more.
15485 			 */
15486 			fsb->o_m_len = 0;
15487 		}
15488 	}
15489 	return (top);
15490 nospace:
15491 	if (top)
15492 		m_freem(top);
15493 	return (NULL);
15494 
15495 }
15496 
15497 /*
15498  * This is a copy of m_copym(), taking the TSO segment size/limit
15499  * constraints into account, and advancing the sndptr as it goes.
15500  */
15501 static struct mbuf *
15502 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15503 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15504 {
15505 	struct mbuf *m, *n;
15506 	int32_t soff;
15507 
15508 	soff = rack->r_ctl.fsb.off;
15509 	m = rack->r_ctl.fsb.m;
15510 	if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15511 		/*
15512 		 * The mbuf had the front of it chopped off by an ack
15513 		 * we need to adjust the soff/off by that difference.
15514 		 */
15515 		uint32_t delta;
15516 
15517 		delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15518 		soff -= delta;
15519 	} else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15520 		/*
15521 		 * The mbuf was expanded probably by
15522 		 * a m_compress. Just update o_m_len.
15523 		 */
15524 		rack->r_ctl.fsb.o_m_len = m->m_len;
15525 	}
15526 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15527 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15528 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15529 				 __FUNCTION__,
15530 				 rack, *plen, m, m->m_len));
15531 	/* Save off the right location before we copy and advance */
15532 	*s_soff = soff;
15533 	*s_mb = rack->r_ctl.fsb.m;
15534 	n = rack_fo_base_copym(m, soff, plen,
15535 			       &rack->r_ctl.fsb,
15536 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15537 	return (n);
15538 }
15539 
15540 static int
15541 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15542 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15543 {
15544 	/*
15545 	 * Enter the fast retransmit path. We are given that a sched_pin is
15546 	 * in place (if accounting is compliled in) and the cycle count taken
15547 	 * at the entry is in the ts_val. The concept her is that the rsm
15548 	 * now holds the mbuf offsets and such so we can directly transmit
15549 	 * without a lot of overhead, the len field is already set for
15550 	 * us to prohibit us from sending too much (usually its 1MSS).
15551 	 */
15552 	struct ip *ip = NULL;
15553 	struct udphdr *udp = NULL;
15554 	struct tcphdr *th = NULL;
15555 	struct mbuf *m = NULL;
15556 	struct inpcb *inp;
15557 	uint8_t *cpto;
15558 	struct tcp_log_buffer *lgb;
15559 #ifdef TCP_ACCOUNTING
15560 	uint64_t crtsc;
15561 	int cnt_thru = 1;
15562 #endif
15563 	struct tcpopt to;
15564 	u_char opt[TCP_MAXOLEN];
15565 	uint32_t hdrlen, optlen;
15566 	int32_t slot, segsiz, max_val, tso = 0, error, ulen = 0;
15567 	uint16_t flags;
15568 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
15569 	uint32_t if_hw_tsomaxsegsize;
15570 
15571 #ifdef INET6
15572 	struct ip6_hdr *ip6 = NULL;
15573 
15574 	if (rack->r_is_v6) {
15575 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15576 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15577 	} else
15578 #endif				/* INET6 */
15579 	{
15580 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15581 		hdrlen = sizeof(struct tcpiphdr);
15582 	}
15583 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15584 		goto failed;
15585 	}
15586 	if (doing_tlp) {
15587 		/* Its a TLP add the flag, it may already be there but be sure */
15588 		rsm->r_flags |= RACK_TLP;
15589 	} else {
15590 		/* If it was a TLP it is not not on this retransmit */
15591 		rsm->r_flags &= ~RACK_TLP;
15592 	}
15593 	startseq = rsm->r_start;
15594 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15595 	inp = rack->rc_inp;
15596 	to.to_flags = 0;
15597 	flags = tcp_outflags[tp->t_state];
15598 	if (flags & (TH_SYN|TH_RST)) {
15599 		goto failed;
15600 	}
15601 	if (rsm->r_flags & RACK_HAS_FIN) {
15602 		/* We can't send a FIN here */
15603 		goto failed;
15604 	}
15605 	if (flags & TH_FIN) {
15606 		/* We never send a FIN */
15607 		flags &= ~TH_FIN;
15608 	}
15609 	if (tp->t_flags & TF_RCVD_TSTMP) {
15610 		to.to_tsval = ms_cts + tp->ts_offset;
15611 		to.to_tsecr = tp->ts_recent;
15612 		to.to_flags = TOF_TS;
15613 	}
15614 	optlen = tcp_addoptions(&to, opt);
15615 	hdrlen += optlen;
15616 	udp = rack->r_ctl.fsb.udp;
15617 	if (udp)
15618 		hdrlen += sizeof(struct udphdr);
15619 	if (rack->r_ctl.rc_pace_max_segs)
15620 		max_val = rack->r_ctl.rc_pace_max_segs;
15621 	else if (rack->rc_user_set_max_segs)
15622 		max_val = rack->rc_user_set_max_segs * segsiz;
15623 	else
15624 		max_val = len;
15625 	if ((tp->t_flags & TF_TSO) &&
15626 	    V_tcp_do_tso &&
15627 	    (len > segsiz) &&
15628 	    (tp->t_port == 0))
15629 		tso = 1;
15630 #ifdef INET6
15631 	if (MHLEN < hdrlen + max_linkhdr)
15632 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15633 	else
15634 #endif
15635 		m = m_gethdr(M_NOWAIT, MT_DATA);
15636 	if (m == NULL)
15637 		goto failed;
15638 	m->m_data += max_linkhdr;
15639 	m->m_len = hdrlen;
15640 	th = rack->r_ctl.fsb.th;
15641 	/* Establish the len to send */
15642 	if (len > max_val)
15643 		len = max_val;
15644 	if ((tso) && (len + optlen > tp->t_maxseg)) {
15645 		uint32_t if_hw_tsomax;
15646 		int32_t max_len;
15647 
15648 		/* extract TSO information */
15649 		if_hw_tsomax = tp->t_tsomax;
15650 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15651 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15652 		/*
15653 		 * Check if we should limit by maximum payload
15654 		 * length:
15655 		 */
15656 		if (if_hw_tsomax != 0) {
15657 			/* compute maximum TSO length */
15658 			max_len = (if_hw_tsomax - hdrlen -
15659 				   max_linkhdr);
15660 			if (max_len <= 0) {
15661 				goto failed;
15662 			} else if (len > max_len) {
15663 				len = max_len;
15664 			}
15665 		}
15666 		if (len <= segsiz) {
15667 			/*
15668 			 * In case there are too many small fragments don't
15669 			 * use TSO:
15670 			 */
15671 			tso = 0;
15672 		}
15673 	} else {
15674 		tso = 0;
15675 	}
15676 	if ((tso == 0) && (len > segsiz))
15677 		len = segsiz;
15678 	if ((len == 0) ||
15679 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
15680 		goto failed;
15681 	}
15682 	th->th_seq = htonl(rsm->r_start);
15683 	th->th_ack = htonl(tp->rcv_nxt);
15684 	/*
15685 	 * The PUSH bit should only be applied
15686 	 * if the full retransmission is made. If
15687 	 * we are sending less than this is the
15688 	 * left hand edge and should not have
15689 	 * the PUSH bit.
15690 	 */
15691 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
15692 	    (len == (rsm->r_end - rsm->r_start)))
15693 		flags |= TH_PUSH;
15694 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15695 	if (th->th_win == 0) {
15696 		tp->t_sndzerowin++;
15697 		tp->t_flags |= TF_RXWIN0SENT;
15698 	} else
15699 		tp->t_flags &= ~TF_RXWIN0SENT;
15700 	if (rsm->r_flags & RACK_TLP) {
15701 		/*
15702 		 * TLP should not count in retran count, but
15703 		 * in its own bin
15704 		 */
15705 		counter_u64_add(rack_tlp_retran, 1);
15706 		counter_u64_add(rack_tlp_retran_bytes, len);
15707 	} else {
15708 		tp->t_sndrexmitpack++;
15709 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15710 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15711 	}
15712 #ifdef STATS
15713 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15714 				 len);
15715 #endif
15716 	if (rsm->m == NULL)
15717 		goto failed;
15718 	if (rsm->orig_m_len != rsm->m->m_len) {
15719 		/* Fix up the orig_m_len and possibly the mbuf offset */
15720 		rack_adjust_orig_mlen(rsm);
15721 	}
15722 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15723 	if (len <= segsiz) {
15724 		/*
15725 		 * Must have ran out of mbufs for the copy
15726 		 * shorten it to no longer need tso. Lets
15727 		 * not put on sendalot since we are low on
15728 		 * mbufs.
15729 		 */
15730 		tso = 0;
15731 	}
15732 	if ((m->m_next == NULL) || (len <= 0)){
15733 		goto failed;
15734 	}
15735 	if (udp) {
15736 		if (rack->r_is_v6)
15737 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
15738 		else
15739 			ulen = hdrlen + len - sizeof(struct ip);
15740 		udp->uh_ulen = htons(ulen);
15741 	}
15742 	m->m_pkthdr.rcvif = (struct ifnet *)0;
15743 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
15744 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
15745 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
15746 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
15747 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
15748 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
15749 #ifdef INET6
15750 		if (rack->r_is_v6) {
15751 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
15752 		    ip6->ip6_flow |= htonl(ect << 20);
15753 		}
15754 		else
15755 #endif
15756 		{
15757 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
15758 		    ip->ip_tos |= ect;
15759 		}
15760 	}
15761 	tcp_set_flags(th, flags);
15762 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
15763 #ifdef INET6
15764 	if (rack->r_is_v6) {
15765 		if (tp->t_port) {
15766 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15767 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15768 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15769 			th->th_sum = htons(0);
15770 			UDPSTAT_INC(udps_opackets);
15771 		} else {
15772 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15773 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15774 			th->th_sum = in6_cksum_pseudo(ip6,
15775 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15776 						      0);
15777 		}
15778 	}
15779 #endif
15780 #if defined(INET6) && defined(INET)
15781 	else
15782 #endif
15783 #ifdef INET
15784 	{
15785 		if (tp->t_port) {
15786 			m->m_pkthdr.csum_flags = CSUM_UDP;
15787 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15788 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15789 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15790 			th->th_sum = htons(0);
15791 			UDPSTAT_INC(udps_opackets);
15792 		} else {
15793 			m->m_pkthdr.csum_flags = CSUM_TCP;
15794 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15795 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
15796 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15797 									IPPROTO_TCP + len + optlen));
15798 		}
15799 		/* IP version must be set here for ipv4/ipv6 checking later */
15800 		KASSERT(ip->ip_v == IPVERSION,
15801 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
15802 	}
15803 #endif
15804 	if (tso) {
15805 		KASSERT(len > tp->t_maxseg - optlen,
15806 			("%s: len <= tso_segsz tp:%p", __func__, tp));
15807 		m->m_pkthdr.csum_flags |= CSUM_TSO;
15808 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15809 	}
15810 #ifdef INET6
15811 	if (rack->r_is_v6) {
15812 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15813 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15814 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15815 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15816 		else
15817 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15818 	}
15819 #endif
15820 #if defined(INET) && defined(INET6)
15821 	else
15822 #endif
15823 #ifdef INET
15824 	{
15825 		ip->ip_len = htons(m->m_pkthdr.len);
15826 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15827 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15828 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15829 			if (tp->t_port == 0 || len < V_tcp_minmss) {
15830 				ip->ip_off |= htons(IP_DF);
15831 			}
15832 		} else {
15833 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15834 		}
15835 	}
15836 #endif
15837 	/* Time to copy in our header */
15838 	cpto = mtod(m, uint8_t *);
15839 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15840 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15841 	if (optlen) {
15842 		bcopy(opt, th + 1, optlen);
15843 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15844 	} else {
15845 		th->th_off = sizeof(struct tcphdr) >> 2;
15846 	}
15847 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15848 		union tcp_log_stackspecific log;
15849 
15850 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15851 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15852 		if (rack->rack_no_prr)
15853 			log.u_bbr.flex1 = 0;
15854 		else
15855 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15856 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15857 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15858 		log.u_bbr.flex4 = max_val;
15859 		log.u_bbr.flex5 = 0;
15860 		/* Save off the early/late values */
15861 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15862 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15863 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15864 		if (doing_tlp == 0)
15865 			log.u_bbr.flex8 = 1;
15866 		else
15867 			log.u_bbr.flex8 = 2;
15868 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15869 		log.u_bbr.flex7 = 55;
15870 		log.u_bbr.pkts_out = tp->t_maxseg;
15871 		log.u_bbr.timeStamp = cts;
15872 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15873 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15874 		log.u_bbr.delivered = 0;
15875 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15876 				     len, &log, false, NULL, NULL, 0, tv);
15877 	} else
15878 		lgb = NULL;
15879 #ifdef INET6
15880 	if (rack->r_is_v6) {
15881 		error = ip6_output(m, NULL,
15882 				   &inp->inp_route6,
15883 				   0, NULL, NULL, inp);
15884 	}
15885 #endif
15886 #if defined(INET) && defined(INET6)
15887 	else
15888 #endif
15889 #ifdef INET
15890 	{
15891 		error = ip_output(m, NULL,
15892 				  &inp->inp_route,
15893 				  0, 0, inp);
15894 	}
15895 #endif
15896 	m = NULL;
15897 	if (lgb) {
15898 		lgb->tlb_errno = error;
15899 		lgb = NULL;
15900 	}
15901 	if (error) {
15902 		goto failed;
15903 	}
15904 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15905 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
15906 	if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15907 		rack->rc_tlp_in_progress = 1;
15908 		rack->r_ctl.rc_tlp_cnt_out++;
15909 	}
15910 	if (error == 0) {
15911 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
15912 		if (doing_tlp) {
15913 			rack->rc_last_sent_tlp_past_cumack = 0;
15914 			rack->rc_last_sent_tlp_seq_valid = 1;
15915 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
15916 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
15917 		}
15918 	}
15919 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15920 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
15921 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15922 		rack->r_ctl.retran_during_recovery += len;
15923 	{
15924 		int idx;
15925 
15926 		idx = (len / segsiz) + 3;
15927 		if (idx >= TCP_MSS_ACCT_ATIMER)
15928 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15929 		else
15930 			counter_u64_add(rack_out_size[idx], 1);
15931 	}
15932 	if (tp->t_rtttime == 0) {
15933 		tp->t_rtttime = ticks;
15934 		tp->t_rtseq = startseq;
15935 		KMOD_TCPSTAT_INC(tcps_segstimed);
15936 	}
15937 	counter_u64_add(rack_fto_rsm_send, 1);
15938 	if (error && (error == ENOBUFS)) {
15939 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15940 		if (rack->rc_enobuf < 0x7f)
15941 			rack->rc_enobuf++;
15942 		if (slot < (10 * HPTS_USEC_IN_MSEC))
15943 			slot = 10 * HPTS_USEC_IN_MSEC;
15944 	} else
15945 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15946 	if ((slot == 0) ||
15947 	    (rack->rc_always_pace == 0) ||
15948 	    (rack->r_rr_config == 1)) {
15949 		/*
15950 		 * We have no pacing set or we
15951 		 * are using old-style rack or
15952 		 * we are overriden to use the old 1ms pacing.
15953 		 */
15954 		slot = rack->r_ctl.rc_min_to;
15955 	}
15956 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
15957 	if (rack->r_must_retran) {
15958 		rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
15959 		if ((SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) ||
15960 		    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
15961 			/*
15962 			 * We have retransmitted all we need. If
15963 			 * RACK_MUST_RXT is not set then we need to
15964 			 * not retransmit this guy.
15965 			 */
15966 			rack->r_must_retran = 0;
15967 			rack->r_ctl.rc_out_at_rto = 0;
15968 			if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
15969 				/* Not one we should rxt */
15970 				goto failed;
15971 			} else {
15972 				/* Clear the flag */
15973 				rsm->r_flags &= ~RACK_MUST_RXT;
15974 			}
15975 		} else {
15976 			/* Remove  the flag */
15977 			rsm->r_flags &= ~RACK_MUST_RXT;
15978 		}
15979 	}
15980 #ifdef TCP_ACCOUNTING
15981 	crtsc = get_cyclecount();
15982 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15983 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15984 	}
15985 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15986 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15987 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15988 	}
15989 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15990 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15991 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
15992 	}
15993 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
15994 	sched_unpin();
15995 #endif
15996 	return (0);
15997 failed:
15998 	if (m)
15999 		m_free(m);
16000 	return (-1);
16001 }
16002 
16003 static void
16004 rack_sndbuf_autoscale(struct tcp_rack *rack)
16005 {
16006 	/*
16007 	 * Automatic sizing of send socket buffer.  Often the send buffer
16008 	 * size is not optimally adjusted to the actual network conditions
16009 	 * at hand (delay bandwidth product).  Setting the buffer size too
16010 	 * small limits throughput on links with high bandwidth and high
16011 	 * delay (eg. trans-continental/oceanic links).  Setting the
16012 	 * buffer size too big consumes too much real kernel memory,
16013 	 * especially with many connections on busy servers.
16014 	 *
16015 	 * The criteria to step up the send buffer one notch are:
16016 	 *  1. receive window of remote host is larger than send buffer
16017 	 *     (with a fudge factor of 5/4th);
16018 	 *  2. send buffer is filled to 7/8th with data (so we actually
16019 	 *     have data to make use of it);
16020 	 *  3. send buffer fill has not hit maximal automatic size;
16021 	 *  4. our send window (slow start and cogestion controlled) is
16022 	 *     larger than sent but unacknowledged data in send buffer.
16023 	 *
16024 	 * Note that the rack version moves things much faster since
16025 	 * we want to avoid hitting cache lines in the rack_fast_output()
16026 	 * path so this is called much less often and thus moves
16027 	 * the SB forward by a percentage.
16028 	 */
16029 	struct socket *so;
16030 	struct tcpcb *tp;
16031 	uint32_t sendwin, scaleup;
16032 
16033 	tp = rack->rc_tp;
16034 	so = rack->rc_inp->inp_socket;
16035 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16036 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16037 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16038 		    sbused(&so->so_snd) >=
16039 		    (so->so_snd.sb_hiwat / 8 * 7) &&
16040 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16041 		    sendwin >= (sbused(&so->so_snd) -
16042 		    (tp->snd_nxt - tp->snd_una))) {
16043 			if (rack_autosndbuf_inc)
16044 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16045 			else
16046 				scaleup = V_tcp_autosndbuf_inc;
16047 			if (scaleup < V_tcp_autosndbuf_inc)
16048 				scaleup = V_tcp_autosndbuf_inc;
16049 			scaleup += so->so_snd.sb_hiwat;
16050 			if (scaleup > V_tcp_autosndbuf_max)
16051 				scaleup = V_tcp_autosndbuf_max;
16052 			if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
16053 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16054 		}
16055 	}
16056 }
16057 
16058 static int
16059 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16060 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16061 {
16062 	/*
16063 	 * Enter to do fast output. We are given that the sched_pin is
16064 	 * in place (if accounting is compiled in) and the cycle count taken
16065 	 * at entry is in place in ts_val. The idea here is that
16066 	 * we know how many more bytes needs to be sent (presumably either
16067 	 * during pacing or to fill the cwnd and that was greater than
16068 	 * the max-burst). We have how much to send and all the info we
16069 	 * need to just send.
16070 	 */
16071 	struct ip *ip = NULL;
16072 	struct udphdr *udp = NULL;
16073 	struct tcphdr *th = NULL;
16074 	struct mbuf *m, *s_mb;
16075 	struct inpcb *inp;
16076 	uint8_t *cpto;
16077 	struct tcp_log_buffer *lgb;
16078 #ifdef TCP_ACCOUNTING
16079 	uint64_t crtsc;
16080 #endif
16081 	struct tcpopt to;
16082 	u_char opt[TCP_MAXOLEN];
16083 	uint32_t hdrlen, optlen;
16084 	int cnt_thru = 1;
16085 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
16086 	uint16_t flags;
16087 	uint32_t s_soff;
16088 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
16089 	uint32_t if_hw_tsomaxsegsize;
16090 	uint16_t add_flag = RACK_SENT_FP;
16091 #ifdef INET6
16092 	struct ip6_hdr *ip6 = NULL;
16093 
16094 	if (rack->r_is_v6) {
16095 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16096 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16097 	} else
16098 #endif				/* INET6 */
16099 	{
16100 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16101 		hdrlen = sizeof(struct tcpiphdr);
16102 	}
16103 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16104 		m = NULL;
16105 		goto failed;
16106 	}
16107 	startseq = tp->snd_max;
16108 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16109 	inp = rack->rc_inp;
16110 	len = rack->r_ctl.fsb.left_to_send;
16111 	to.to_flags = 0;
16112 	flags = rack->r_ctl.fsb.tcp_flags;
16113 	if (tp->t_flags & TF_RCVD_TSTMP) {
16114 		to.to_tsval = ms_cts + tp->ts_offset;
16115 		to.to_tsecr = tp->ts_recent;
16116 		to.to_flags = TOF_TS;
16117 	}
16118 	optlen = tcp_addoptions(&to, opt);
16119 	hdrlen += optlen;
16120 	udp = rack->r_ctl.fsb.udp;
16121 	if (udp)
16122 		hdrlen += sizeof(struct udphdr);
16123 	if (rack->r_ctl.rc_pace_max_segs)
16124 		max_val = rack->r_ctl.rc_pace_max_segs;
16125 	else if (rack->rc_user_set_max_segs)
16126 		max_val = rack->rc_user_set_max_segs * segsiz;
16127 	else
16128 		max_val = len;
16129 	if ((tp->t_flags & TF_TSO) &&
16130 	    V_tcp_do_tso &&
16131 	    (len > segsiz) &&
16132 	    (tp->t_port == 0))
16133 		tso = 1;
16134 again:
16135 #ifdef INET6
16136 	if (MHLEN < hdrlen + max_linkhdr)
16137 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16138 	else
16139 #endif
16140 		m = m_gethdr(M_NOWAIT, MT_DATA);
16141 	if (m == NULL)
16142 		goto failed;
16143 	m->m_data += max_linkhdr;
16144 	m->m_len = hdrlen;
16145 	th = rack->r_ctl.fsb.th;
16146 	/* Establish the len to send */
16147 	if (len > max_val)
16148 		len = max_val;
16149 	if ((tso) && (len + optlen > tp->t_maxseg)) {
16150 		uint32_t if_hw_tsomax;
16151 		int32_t max_len;
16152 
16153 		/* extract TSO information */
16154 		if_hw_tsomax = tp->t_tsomax;
16155 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16156 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16157 		/*
16158 		 * Check if we should limit by maximum payload
16159 		 * length:
16160 		 */
16161 		if (if_hw_tsomax != 0) {
16162 			/* compute maximum TSO length */
16163 			max_len = (if_hw_tsomax - hdrlen -
16164 				   max_linkhdr);
16165 			if (max_len <= 0) {
16166 				goto failed;
16167 			} else if (len > max_len) {
16168 				len = max_len;
16169 			}
16170 		}
16171 		if (len <= segsiz) {
16172 			/*
16173 			 * In case there are too many small fragments don't
16174 			 * use TSO:
16175 			 */
16176 			tso = 0;
16177 		}
16178 	} else {
16179 		tso = 0;
16180 	}
16181 	if ((tso == 0) && (len > segsiz))
16182 		len = segsiz;
16183 	if ((len == 0) ||
16184 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
16185 		goto failed;
16186 	}
16187 	sb_offset = tp->snd_max - tp->snd_una;
16188 	th->th_seq = htonl(tp->snd_max);
16189 	th->th_ack = htonl(tp->rcv_nxt);
16190 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16191 	if (th->th_win == 0) {
16192 		tp->t_sndzerowin++;
16193 		tp->t_flags |= TF_RXWIN0SENT;
16194 	} else
16195 		tp->t_flags &= ~TF_RXWIN0SENT;
16196 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
16197 	KMOD_TCPSTAT_INC(tcps_sndpack);
16198 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16199 #ifdef STATS
16200 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16201 				 len);
16202 #endif
16203 	if (rack->r_ctl.fsb.m == NULL)
16204 		goto failed;
16205 
16206 	/* s_mb and s_soff are saved for rack_log_output */
16207 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16208 				    &s_mb, &s_soff);
16209 	if (len <= segsiz) {
16210 		/*
16211 		 * Must have ran out of mbufs for the copy
16212 		 * shorten it to no longer need tso. Lets
16213 		 * not put on sendalot since we are low on
16214 		 * mbufs.
16215 		 */
16216 		tso = 0;
16217 	}
16218 	if (rack->r_ctl.fsb.rfo_apply_push &&
16219 	    (len == rack->r_ctl.fsb.left_to_send)) {
16220 		flags |= TH_PUSH;
16221 		add_flag |= RACK_HAD_PUSH;
16222 	}
16223 	if ((m->m_next == NULL) || (len <= 0)){
16224 		goto failed;
16225 	}
16226 	if (udp) {
16227 		if (rack->r_is_v6)
16228 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
16229 		else
16230 			ulen = hdrlen + len - sizeof(struct ip);
16231 		udp->uh_ulen = htons(ulen);
16232 	}
16233 	m->m_pkthdr.rcvif = (struct ifnet *)0;
16234 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
16235 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
16236 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
16237 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
16238 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
16239 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
16240 #ifdef INET6
16241 		if (rack->r_is_v6) {
16242 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
16243 			ip6->ip6_flow |= htonl(ect << 20);
16244 		}
16245 		else
16246 #endif
16247 		{
16248 			ip->ip_tos &= ~IPTOS_ECN_MASK;
16249 			ip->ip_tos |= ect;
16250 		}
16251 	}
16252 	tcp_set_flags(th, flags);
16253 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
16254 #ifdef INET6
16255 	if (rack->r_is_v6) {
16256 		if (tp->t_port) {
16257 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16258 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16259 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16260 			th->th_sum = htons(0);
16261 			UDPSTAT_INC(udps_opackets);
16262 		} else {
16263 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16264 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16265 			th->th_sum = in6_cksum_pseudo(ip6,
16266 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16267 						      0);
16268 		}
16269 	}
16270 #endif
16271 #if defined(INET6) && defined(INET)
16272 	else
16273 #endif
16274 #ifdef INET
16275 	{
16276 		if (tp->t_port) {
16277 			m->m_pkthdr.csum_flags = CSUM_UDP;
16278 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16279 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16280 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16281 			th->th_sum = htons(0);
16282 			UDPSTAT_INC(udps_opackets);
16283 		} else {
16284 			m->m_pkthdr.csum_flags = CSUM_TCP;
16285 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16286 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
16287 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16288 									IPPROTO_TCP + len + optlen));
16289 		}
16290 		/* IP version must be set here for ipv4/ipv6 checking later */
16291 		KASSERT(ip->ip_v == IPVERSION,
16292 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
16293 	}
16294 #endif
16295 	if (tso) {
16296 		KASSERT(len > tp->t_maxseg - optlen,
16297 			("%s: len <= tso_segsz tp:%p", __func__, tp));
16298 		m->m_pkthdr.csum_flags |= CSUM_TSO;
16299 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16300 	}
16301 #ifdef INET6
16302 	if (rack->r_is_v6) {
16303 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16304 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16305 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16306 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16307 		else
16308 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16309 	}
16310 #endif
16311 #if defined(INET) && defined(INET6)
16312 	else
16313 #endif
16314 #ifdef INET
16315 	{
16316 		ip->ip_len = htons(m->m_pkthdr.len);
16317 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16318 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16319 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16320 			if (tp->t_port == 0 || len < V_tcp_minmss) {
16321 				ip->ip_off |= htons(IP_DF);
16322 			}
16323 		} else {
16324 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16325 		}
16326 	}
16327 #endif
16328 	/* Time to copy in our header */
16329 	cpto = mtod(m, uint8_t *);
16330 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16331 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16332 	if (optlen) {
16333 		bcopy(opt, th + 1, optlen);
16334 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16335 	} else {
16336 		th->th_off = sizeof(struct tcphdr) >> 2;
16337 	}
16338 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16339 		union tcp_log_stackspecific log;
16340 
16341 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16342 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16343 		if (rack->rack_no_prr)
16344 			log.u_bbr.flex1 = 0;
16345 		else
16346 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16347 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16348 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16349 		log.u_bbr.flex4 = max_val;
16350 		log.u_bbr.flex5 = 0;
16351 		/* Save off the early/late values */
16352 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16353 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16354 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16355 		log.u_bbr.flex8 = 0;
16356 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16357 		log.u_bbr.flex7 = 44;
16358 		log.u_bbr.pkts_out = tp->t_maxseg;
16359 		log.u_bbr.timeStamp = cts;
16360 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16361 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16362 		log.u_bbr.delivered = 0;
16363 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16364 				     len, &log, false, NULL, NULL, 0, tv);
16365 	} else
16366 		lgb = NULL;
16367 #ifdef INET6
16368 	if (rack->r_is_v6) {
16369 		error = ip6_output(m, NULL,
16370 				   &inp->inp_route6,
16371 				   0, NULL, NULL, inp);
16372 	}
16373 #endif
16374 #if defined(INET) && defined(INET6)
16375 	else
16376 #endif
16377 #ifdef INET
16378 	{
16379 		error = ip_output(m, NULL,
16380 				  &inp->inp_route,
16381 				  0, 0, inp);
16382 	}
16383 #endif
16384 	if (lgb) {
16385 		lgb->tlb_errno = error;
16386 		lgb = NULL;
16387 	}
16388 	if (error) {
16389 		*send_err = error;
16390 		m = NULL;
16391 		goto failed;
16392 	}
16393 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16394 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16395 	m = NULL;
16396 	if (tp->snd_una == tp->snd_max) {
16397 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
16398 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16399 		tp->t_acktime = ticks;
16400 	}
16401 	if (error == 0)
16402 		tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16403 
16404 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16405 	tot_len += len;
16406 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
16407 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16408 	tp->snd_max += len;
16409 	tp->snd_nxt = tp->snd_max;
16410 	{
16411 		int idx;
16412 
16413 		idx = (len / segsiz) + 3;
16414 		if (idx >= TCP_MSS_ACCT_ATIMER)
16415 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16416 		else
16417 			counter_u64_add(rack_out_size[idx], 1);
16418 	}
16419 	if (len <= rack->r_ctl.fsb.left_to_send)
16420 		rack->r_ctl.fsb.left_to_send -= len;
16421 	else
16422 		rack->r_ctl.fsb.left_to_send = 0;
16423 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
16424 		rack->r_fast_output = 0;
16425 		rack->r_ctl.fsb.left_to_send = 0;
16426 		/* At the end of fast_output scale up the sb */
16427 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16428 		rack_sndbuf_autoscale(rack);
16429 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16430 	}
16431 	if (tp->t_rtttime == 0) {
16432 		tp->t_rtttime = ticks;
16433 		tp->t_rtseq = startseq;
16434 		KMOD_TCPSTAT_INC(tcps_segstimed);
16435 	}
16436 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16437 	    (max_val > len) &&
16438 	    (tso == 0)) {
16439 		max_val -= len;
16440 		len = segsiz;
16441 		th = rack->r_ctl.fsb.th;
16442 		cnt_thru++;
16443 		goto again;
16444 	}
16445 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16446 	counter_u64_add(rack_fto_send, 1);
16447 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16448 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16449 #ifdef TCP_ACCOUNTING
16450 	crtsc = get_cyclecount();
16451 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16452 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16453 	}
16454 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16455 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16456 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16457 	}
16458 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16459 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16460 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16461 	}
16462 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16463 	sched_unpin();
16464 #endif
16465 	return (0);
16466 failed:
16467 	if (m)
16468 		m_free(m);
16469 	rack->r_fast_output = 0;
16470 	return (-1);
16471 }
16472 
16473 static int
16474 rack_output(struct tcpcb *tp)
16475 {
16476 	struct socket *so;
16477 	uint32_t recwin;
16478 	uint32_t sb_offset, s_moff = 0;
16479 	int32_t len, error = 0;
16480 	uint16_t flags;
16481 	struct mbuf *m, *s_mb = NULL;
16482 	struct mbuf *mb;
16483 	uint32_t if_hw_tsomaxsegcount = 0;
16484 	uint32_t if_hw_tsomaxsegsize;
16485 	int32_t segsiz, minseg;
16486 	long tot_len_this_send = 0;
16487 #ifdef INET
16488 	struct ip *ip = NULL;
16489 #endif
16490 #ifdef TCPDEBUG
16491 	struct ipovly *ipov = NULL;
16492 #endif
16493 	struct udphdr *udp = NULL;
16494 	struct tcp_rack *rack;
16495 	struct tcphdr *th;
16496 	uint8_t pass = 0;
16497 	uint8_t mark = 0;
16498 	uint8_t wanted_cookie = 0;
16499 	u_char opt[TCP_MAXOLEN];
16500 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
16501 	uint32_t rack_seq;
16502 
16503 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16504 	unsigned ipsec_optlen = 0;
16505 
16506 #endif
16507 	int32_t idle, sendalot;
16508 	int32_t sub_from_prr = 0;
16509 	volatile int32_t sack_rxmit;
16510 	struct rack_sendmap *rsm = NULL;
16511 	int32_t tso, mtu;
16512 	struct tcpopt to;
16513 	int32_t slot = 0;
16514 	int32_t sup_rack = 0;
16515 	uint32_t cts, ms_cts, delayed, early;
16516 	uint16_t add_flag = RACK_SENT_SP;
16517 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16518 	uint8_t hpts_calling,  doing_tlp = 0;
16519 	uint32_t cwnd_to_use, pace_max_seg;
16520 	int32_t do_a_prefetch = 0;
16521 	int32_t prefetch_rsm = 0;
16522 	int32_t orig_len = 0;
16523 	struct timeval tv;
16524 	int32_t prefetch_so_done = 0;
16525 	struct tcp_log_buffer *lgb;
16526 	struct inpcb *inp;
16527 	struct sockbuf *sb;
16528 	uint64_t ts_val = 0;
16529 #ifdef TCP_ACCOUNTING
16530 	uint64_t crtsc;
16531 #endif
16532 #ifdef INET6
16533 	struct ip6_hdr *ip6 = NULL;
16534 	int32_t isipv6;
16535 #endif
16536 	uint8_t filled_all = 0;
16537 	bool hw_tls = false;
16538 
16539 	/* setup and take the cache hits here */
16540 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16541 #ifdef TCP_ACCOUNTING
16542 	sched_pin();
16543 	ts_val = get_cyclecount();
16544 #endif
16545 	hpts_calling = rack->rc_inp->inp_hpts_calls;
16546 	NET_EPOCH_ASSERT();
16547 	INP_WLOCK_ASSERT(rack->rc_inp);
16548 #ifdef TCP_OFFLOAD
16549 	if (tp->t_flags & TF_TOE) {
16550 #ifdef TCP_ACCOUNTING
16551 		sched_unpin();
16552 #endif
16553 		return (tcp_offload_output(tp));
16554 	}
16555 #endif
16556 	/*
16557 	 * For TFO connections in SYN_RECEIVED, only allow the initial
16558 	 * SYN|ACK and those sent by the retransmit timer.
16559 	 */
16560 	if (IS_FASTOPEN(tp->t_flags) &&
16561 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
16562 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16563 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16564 #ifdef TCP_ACCOUNTING
16565 		sched_unpin();
16566 #endif
16567 		return (0);
16568 	}
16569 #ifdef INET6
16570 	if (rack->r_state) {
16571 		/* Use the cache line loaded if possible */
16572 		isipv6 = rack->r_is_v6;
16573 	} else {
16574 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16575 	}
16576 #endif
16577 	early = 0;
16578 	cts = tcp_get_usecs(&tv);
16579 	ms_cts = tcp_tv_to_mssectick(&tv);
16580 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16581 	    tcp_in_hpts(rack->rc_inp)) {
16582 		/*
16583 		 * We are on the hpts for some timer but not hptsi output.
16584 		 * Remove from the hpts unconditionally.
16585 		 */
16586 		rack_timer_cancel(tp, rack, cts, __LINE__);
16587 	}
16588 	/* Are we pacing and late? */
16589 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16590 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16591 		/* We are delayed */
16592 		delayed = cts - rack->r_ctl.rc_last_output_to;
16593 	} else {
16594 		delayed = 0;
16595 	}
16596 	/* Do the timers, which may override the pacer */
16597 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16598 		int retval;
16599 
16600 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
16601 		    &doing_tlp);
16602 		if (retval != 0) {
16603 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16604 #ifdef TCP_ACCOUNTING
16605 			sched_unpin();
16606 #endif
16607 			/*
16608 			 * If timers want tcp_drop(), then pass error out,
16609 			 * otherwise suppress it.
16610 			 */
16611 			return (retval < 0 ? retval : 0);
16612 		}
16613 	}
16614 	if (rack->rc_in_persist) {
16615 		if (tcp_in_hpts(rack->rc_inp) == 0) {
16616 			/* Timer is not running */
16617 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16618 		}
16619 #ifdef TCP_ACCOUNTING
16620 		sched_unpin();
16621 #endif
16622 		return (0);
16623 	}
16624 	if ((rack->r_timer_override) ||
16625 	    (rack->rc_ack_can_sendout_data) ||
16626 	    (delayed) ||
16627 	    (tp->t_state < TCPS_ESTABLISHED)) {
16628 		rack->rc_ack_can_sendout_data = 0;
16629 		if (tcp_in_hpts(rack->rc_inp))
16630 			tcp_hpts_remove(rack->rc_inp);
16631 	} else if (tcp_in_hpts(rack->rc_inp)) {
16632 		/*
16633 		 * On the hpts you can't pass even if ACKNOW is on, we will
16634 		 * when the hpts fires.
16635 		 */
16636 #ifdef TCP_ACCOUNTING
16637 		crtsc = get_cyclecount();
16638 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16639 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16640 		}
16641 		counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16642 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16643 			tp->tcp_cnt_counters[SND_BLOCKED]++;
16644 		}
16645 		counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16646 		sched_unpin();
16647 #endif
16648 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16649 		return (0);
16650 	}
16651 	rack->rc_inp->inp_hpts_calls = 0;
16652 	/* Finish out both pacing early and late accounting */
16653 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16654 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16655 		early = rack->r_ctl.rc_last_output_to - cts;
16656 	} else
16657 		early = 0;
16658 	if (delayed) {
16659 		rack->r_ctl.rc_agg_delayed += delayed;
16660 		rack->r_late = 1;
16661 	} else if (early) {
16662 		rack->r_ctl.rc_agg_early += early;
16663 		rack->r_early = 1;
16664 	}
16665 	/* Now that early/late accounting is done turn off the flag */
16666 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16667 	rack->r_wanted_output = 0;
16668 	rack->r_timer_override = 0;
16669 	if ((tp->t_state != rack->r_state) &&
16670 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
16671 		rack_set_state(tp, rack);
16672 	}
16673 	if ((rack->r_fast_output) &&
16674 	    (doing_tlp == 0) &&
16675 	    (tp->rcv_numsacks == 0)) {
16676 		int ret;
16677 
16678 		error = 0;
16679 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16680 		if (ret >= 0)
16681 			return(ret);
16682 		else if (error) {
16683 			inp = rack->rc_inp;
16684 			so = inp->inp_socket;
16685 			sb = &so->so_snd;
16686 			goto nomore;
16687 		}
16688 	}
16689 	inp = rack->rc_inp;
16690 	/*
16691 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
16692 	 * only allow the initial SYN or SYN|ACK and those sent
16693 	 * by the retransmit timer.
16694 	 */
16695 	if (IS_FASTOPEN(tp->t_flags) &&
16696 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
16697 	     (tp->t_state == TCPS_SYN_SENT)) &&
16698 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16699 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
16700 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16701 		so = inp->inp_socket;
16702 		sb = &so->so_snd;
16703 		goto just_return_nolock;
16704 	}
16705 	/*
16706 	 * Determine length of data that should be transmitted, and flags
16707 	 * that will be used. If there is some data or critical controls
16708 	 * (SYN, RST) to send, then transmit; otherwise, investigate
16709 	 * further.
16710 	 */
16711 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16712 	if (tp->t_idle_reduce) {
16713 		if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16714 			rack_cc_after_idle(rack, tp);
16715 	}
16716 	tp->t_flags &= ~TF_LASTIDLE;
16717 	if (idle) {
16718 		if (tp->t_flags & TF_MORETOCOME) {
16719 			tp->t_flags |= TF_LASTIDLE;
16720 			idle = 0;
16721 		}
16722 	}
16723 	if ((tp->snd_una == tp->snd_max) &&
16724 	    rack->r_ctl.rc_went_idle_time &&
16725 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16726 		idle = cts - rack->r_ctl.rc_went_idle_time;
16727 		if (idle > rack_min_probertt_hold) {
16728 			/* Count as a probe rtt */
16729 			if (rack->in_probe_rtt == 0) {
16730 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
16731 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16732 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16733 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16734 			} else {
16735 				rack_exit_probertt(rack, cts);
16736 			}
16737 		}
16738 		idle = 0;
16739 	}
16740 	if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16741 		rack_init_fsb_block(tp, rack);
16742 again:
16743 	/*
16744 	 * If we've recently taken a timeout, snd_max will be greater than
16745 	 * snd_nxt.  There may be SACK information that allows us to avoid
16746 	 * resending already delivered data.  Adjust snd_nxt accordingly.
16747 	 */
16748 	sendalot = 0;
16749 	cts = tcp_get_usecs(&tv);
16750 	ms_cts = tcp_tv_to_mssectick(&tv);
16751 	tso = 0;
16752 	mtu = 0;
16753 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16754 	minseg = segsiz;
16755 	if (rack->r_ctl.rc_pace_max_segs == 0)
16756 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16757 	else
16758 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16759 	sb_offset = tp->snd_max - tp->snd_una;
16760 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16761 	flags = tcp_outflags[tp->t_state];
16762 	while (rack->rc_free_cnt < rack_free_cache) {
16763 		rsm = rack_alloc(rack);
16764 		if (rsm == NULL) {
16765 			if (inp->inp_hpts_calls)
16766 				/* Retry in a ms */
16767 				slot = (1 * HPTS_USEC_IN_MSEC);
16768 			so = inp->inp_socket;
16769 			sb = &so->so_snd;
16770 			goto just_return_nolock;
16771 		}
16772 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16773 		rack->rc_free_cnt++;
16774 		rsm = NULL;
16775 	}
16776 	if (inp->inp_hpts_calls)
16777 		inp->inp_hpts_calls = 0;
16778 	sack_rxmit = 0;
16779 	len = 0;
16780 	rsm = NULL;
16781 	if (flags & TH_RST) {
16782 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16783 		so = inp->inp_socket;
16784 		sb = &so->so_snd;
16785 		goto send;
16786 	}
16787 	if (rack->r_ctl.rc_resend) {
16788 		/* Retransmit timer */
16789 		rsm = rack->r_ctl.rc_resend;
16790 		rack->r_ctl.rc_resend = NULL;
16791 		len = rsm->r_end - rsm->r_start;
16792 		sack_rxmit = 1;
16793 		sendalot = 0;
16794 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16795 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16796 			 __func__, __LINE__,
16797 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16798 		sb_offset = rsm->r_start - tp->snd_una;
16799 		if (len >= segsiz)
16800 			len = segsiz;
16801 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16802 		/* We have a retransmit that takes precedence */
16803 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16804 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16805 			/* Enter recovery if not induced by a time-out */
16806 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
16807 		}
16808 #ifdef INVARIANTS
16809 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16810 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16811 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
16812 		}
16813 #endif
16814 		len = rsm->r_end - rsm->r_start;
16815 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16816 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16817 			 __func__, __LINE__,
16818 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16819 		sb_offset = rsm->r_start - tp->snd_una;
16820 		sendalot = 0;
16821 		if (len >= segsiz)
16822 			len = segsiz;
16823 		if (len > 0) {
16824 			sack_rxmit = 1;
16825 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16826 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16827 			    min(len, segsiz));
16828 		}
16829 	} else if (rack->r_ctl.rc_tlpsend) {
16830 		/* Tail loss probe */
16831 		long cwin;
16832 		long tlen;
16833 
16834 		/*
16835 		 * Check if we can do a TLP with a RACK'd packet
16836 		 * this can happen if we are not doing the rack
16837 		 * cheat and we skipped to a TLP and it
16838 		 * went off.
16839 		 */
16840 		rsm = rack->r_ctl.rc_tlpsend;
16841 		/* We are doing a TLP make sure the flag is preent */
16842 		rsm->r_flags |= RACK_TLP;
16843 		rack->r_ctl.rc_tlpsend = NULL;
16844 		sack_rxmit = 1;
16845 		tlen = rsm->r_end - rsm->r_start;
16846 		if (tlen > segsiz)
16847 			tlen = segsiz;
16848 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16849 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16850 			 __func__, __LINE__,
16851 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16852 		sb_offset = rsm->r_start - tp->snd_una;
16853 		cwin = min(tp->snd_wnd, tlen);
16854 		len = cwin;
16855 	}
16856 	if (rack->r_must_retran &&
16857 	    (doing_tlp == 0) &&
16858 	    (rsm == NULL)) {
16859 		/*
16860 		 * Non-Sack and we had a RTO or Sack/non-Sack and a
16861 		 * MTU change, we need to retransmit until we reach
16862 		 * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
16863 		 */
16864 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
16865 			int sendwin, flight;
16866 
16867 			sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16868 			flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16869 			if (flight >= sendwin) {
16870 				so = inp->inp_socket;
16871 				sb = &so->so_snd;
16872 				goto just_return_nolock;
16873 			}
16874 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16875 			if (rsm == NULL) {
16876 				/* TSNH */
16877 				rack->r_must_retran = 0;
16878 				rack->r_ctl.rc_out_at_rto = 0;
16879 				rack->r_must_retran = 0;
16880 				so = inp->inp_socket;
16881 				sb = &so->so_snd;
16882 				goto just_return_nolock;
16883 			}
16884 			if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
16885 				/* It does not have the flag, we are done */
16886 				rack->r_must_retran = 0;
16887 				rack->r_ctl.rc_out_at_rto = 0;
16888 			} else {
16889 				sack_rxmit = 1;
16890 				len = rsm->r_end - rsm->r_start;
16891 				sendalot = 0;
16892 				sb_offset = rsm->r_start - tp->snd_una;
16893 				if (len >= segsiz)
16894 					len = segsiz;
16895 				/*
16896 				 * Delay removing the flag RACK_MUST_RXT so
16897 				 * that the fastpath for retransmit will
16898 				 * work with this rsm.
16899 				 */
16900 
16901 			}
16902 		} else {
16903 			/* We must be done if there is nothing outstanding */
16904 			rack->r_must_retran = 0;
16905 			rack->r_ctl.rc_out_at_rto = 0;
16906 		}
16907 	}
16908 	/*
16909 	 * Enforce a connection sendmap count limit if set
16910 	 * as long as we are not retransmiting.
16911 	 */
16912 	if ((rsm == NULL) &&
16913 	    (rack->do_detection == 0) &&
16914 	    (V_tcp_map_entries_limit > 0) &&
16915 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16916 		counter_u64_add(rack_to_alloc_limited, 1);
16917 		if (!rack->alloc_limit_reported) {
16918 			rack->alloc_limit_reported = 1;
16919 			counter_u64_add(rack_alloc_limited_conns, 1);
16920 		}
16921 		so = inp->inp_socket;
16922 		sb = &so->so_snd;
16923 		goto just_return_nolock;
16924 	}
16925 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16926 		/* we are retransmitting the fin */
16927 		len--;
16928 		if (len) {
16929 			/*
16930 			 * When retransmitting data do *not* include the
16931 			 * FIN. This could happen from a TLP probe.
16932 			 */
16933 			flags &= ~TH_FIN;
16934 		}
16935 	}
16936 	if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16937 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16938 		int ret;
16939 
16940 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
16941 		if (ret == 0)
16942 			return (0);
16943 	}
16944 	if (rsm && (rsm->r_flags & RACK_MUST_RXT)) {
16945 		/*
16946 		 * Clear the flag in prep for the send
16947 		 * note that if we can't get an mbuf
16948 		 * and fail, we won't retransmit this
16949 		 * rsm but that should be ok (its rare).
16950 		 */
16951 		rsm->r_flags &= ~RACK_MUST_RXT;
16952 	}
16953 	so = inp->inp_socket;
16954 	sb = &so->so_snd;
16955 	if (do_a_prefetch == 0) {
16956 		kern_prefetch(sb, &do_a_prefetch);
16957 		do_a_prefetch = 1;
16958 	}
16959 #ifdef NETFLIX_SHARED_CWND
16960 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
16961 	    rack->rack_enable_scwnd) {
16962 		/* We are doing cwnd sharing */
16963 		if (rack->gp_ready &&
16964 		    (rack->rack_attempted_scwnd == 0) &&
16965 		    (rack->r_ctl.rc_scw == NULL) &&
16966 		    tp->t_lib) {
16967 			/* The pcbid is in, lets make an attempt */
16968 			counter_u64_add(rack_try_scwnd, 1);
16969 			rack->rack_attempted_scwnd = 1;
16970 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
16971 								   &rack->r_ctl.rc_scw_index,
16972 								   segsiz);
16973 		}
16974 		if (rack->r_ctl.rc_scw &&
16975 		    (rack->rack_scwnd_is_idle == 1) &&
16976 		    sbavail(&so->so_snd)) {
16977 			/* we are no longer out of data */
16978 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16979 			rack->rack_scwnd_is_idle = 0;
16980 		}
16981 		if (rack->r_ctl.rc_scw) {
16982 			/* First lets update and get the cwnd */
16983 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
16984 								    rack->r_ctl.rc_scw_index,
16985 								    tp->snd_cwnd, tp->snd_wnd, segsiz);
16986 		}
16987 	}
16988 #endif
16989 	/*
16990 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
16991 	 * state flags.
16992 	 */
16993 	if (tp->t_flags & TF_NEEDFIN)
16994 		flags |= TH_FIN;
16995 	if (tp->t_flags & TF_NEEDSYN)
16996 		flags |= TH_SYN;
16997 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
16998 		void *end_rsm;
16999 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17000 		if (end_rsm)
17001 			kern_prefetch(end_rsm, &prefetch_rsm);
17002 		prefetch_rsm = 1;
17003 	}
17004 	SOCKBUF_LOCK(sb);
17005 	/*
17006 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
17007 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17008 	 * negative length.  This can also occur when TCP opens up its
17009 	 * congestion window while receiving additional duplicate acks after
17010 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
17011 	 * the fast-retransmit.
17012 	 *
17013 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
17014 	 * set to snd_una, the sb_offset will be 0, and the length may wind
17015 	 * up 0.
17016 	 *
17017 	 * If sack_rxmit is true we are retransmitting from the scoreboard
17018 	 * in which case len is already set.
17019 	 */
17020 	if ((sack_rxmit == 0) &&
17021 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17022 		uint32_t avail;
17023 
17024 		avail = sbavail(sb);
17025 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17026 			sb_offset = tp->snd_nxt - tp->snd_una;
17027 		else
17028 			sb_offset = 0;
17029 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17030 			if (rack->r_ctl.rc_tlp_new_data) {
17031 				/* TLP is forcing out new data */
17032 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17033 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17034 				}
17035 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17036 					if (tp->snd_wnd > sb_offset)
17037 						len = tp->snd_wnd - sb_offset;
17038 					else
17039 						len = 0;
17040 				} else {
17041 					len = rack->r_ctl.rc_tlp_new_data;
17042 				}
17043 				rack->r_ctl.rc_tlp_new_data = 0;
17044 			}  else {
17045 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17046 			}
17047 			if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17048 				/*
17049 				 * For prr=off, we need to send only 1 MSS
17050 				 * at a time. We do this because another sack could
17051 				 * be arriving that causes us to send retransmits and
17052 				 * we don't want to be on a long pace due to a larger send
17053 				 * that keeps us from sending out the retransmit.
17054 				 */
17055 				len = segsiz;
17056 			}
17057 		} else {
17058 			uint32_t outstanding;
17059 			/*
17060 			 * We are inside of a Fast recovery episode, this
17061 			 * is caused by a SACK or 3 dup acks. At this point
17062 			 * we have sent all the retransmissions and we rely
17063 			 * on PRR to dictate what we will send in the form of
17064 			 * new data.
17065 			 */
17066 
17067 			outstanding = tp->snd_max - tp->snd_una;
17068 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17069 				if (tp->snd_wnd > outstanding) {
17070 					len = tp->snd_wnd - outstanding;
17071 					/* Check to see if we have the data */
17072 					if ((sb_offset + len) > avail) {
17073 						/* It does not all fit */
17074 						if (avail > sb_offset)
17075 							len = avail - sb_offset;
17076 						else
17077 							len = 0;
17078 					}
17079 				} else {
17080 					len = 0;
17081 				}
17082 			} else if (avail > sb_offset) {
17083 				len = avail - sb_offset;
17084 			} else {
17085 				len = 0;
17086 			}
17087 			if (len > 0) {
17088 				if (len > rack->r_ctl.rc_prr_sndcnt) {
17089 					len = rack->r_ctl.rc_prr_sndcnt;
17090 				}
17091 				if (len > 0) {
17092 					sub_from_prr = 1;
17093 				}
17094 			}
17095 			if (len > segsiz) {
17096 				/*
17097 				 * We should never send more than a MSS when
17098 				 * retransmitting or sending new data in prr
17099 				 * mode unless the override flag is on. Most
17100 				 * likely the PRR algorithm is not going to
17101 				 * let us send a lot as well :-)
17102 				 */
17103 				if (rack->r_ctl.rc_prr_sendalot == 0) {
17104 					len = segsiz;
17105 				}
17106 			} else if (len < segsiz) {
17107 				/*
17108 				 * Do we send any? The idea here is if the
17109 				 * send empty's the socket buffer we want to
17110 				 * do it. However if not then lets just wait
17111 				 * for our prr_sndcnt to get bigger.
17112 				 */
17113 				long leftinsb;
17114 
17115 				leftinsb = sbavail(sb) - sb_offset;
17116 				if (leftinsb > len) {
17117 					/* This send does not empty the sb */
17118 					len = 0;
17119 				}
17120 			}
17121 		}
17122 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17123 		/*
17124 		 * If you have not established
17125 		 * and are not doing FAST OPEN
17126 		 * no data please.
17127 		 */
17128 		if ((sack_rxmit == 0) &&
17129 		    (!IS_FASTOPEN(tp->t_flags))){
17130 			len = 0;
17131 			sb_offset = 0;
17132 		}
17133 	}
17134 	if (prefetch_so_done == 0) {
17135 		kern_prefetch(so, &prefetch_so_done);
17136 		prefetch_so_done = 1;
17137 	}
17138 	/*
17139 	 * Lop off SYN bit if it has already been sent.  However, if this is
17140 	 * SYN-SENT state and if segment contains data and if we don't know
17141 	 * that foreign host supports TAO, suppress sending segment.
17142 	 */
17143 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17144 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17145 		/*
17146 		 * When sending additional segments following a TFO SYN|ACK,
17147 		 * do not include the SYN bit.
17148 		 */
17149 		if (IS_FASTOPEN(tp->t_flags) &&
17150 		    (tp->t_state == TCPS_SYN_RECEIVED))
17151 			flags &= ~TH_SYN;
17152 	}
17153 	/*
17154 	 * Be careful not to send data and/or FIN on SYN segments. This
17155 	 * measure is needed to prevent interoperability problems with not
17156 	 * fully conformant TCP implementations.
17157 	 */
17158 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17159 		len = 0;
17160 		flags &= ~TH_FIN;
17161 	}
17162 	/*
17163 	 * On TFO sockets, ensure no data is sent in the following cases:
17164 	 *
17165 	 *  - When retransmitting SYN|ACK on a passively-created socket
17166 	 *
17167 	 *  - When retransmitting SYN on an actively created socket
17168 	 *
17169 	 *  - When sending a zero-length cookie (cookie request) on an
17170 	 *    actively created socket
17171 	 *
17172 	 *  - When the socket is in the CLOSED state (RST is being sent)
17173 	 */
17174 	if (IS_FASTOPEN(tp->t_flags) &&
17175 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17176 	     ((tp->t_state == TCPS_SYN_SENT) &&
17177 	      (tp->t_tfo_client_cookie_len == 0)) ||
17178 	     (flags & TH_RST))) {
17179 		sack_rxmit = 0;
17180 		len = 0;
17181 	}
17182 	/* Without fast-open there should never be data sent on a SYN */
17183 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17184 		tp->snd_nxt = tp->iss;
17185 		len = 0;
17186 	}
17187 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17188 		/* We only send 1 MSS if we have a DSACK block */
17189 		add_flag |= RACK_SENT_W_DSACK;
17190 		len = segsiz;
17191 	}
17192 	orig_len = len;
17193 	if (len <= 0) {
17194 		/*
17195 		 * If FIN has been sent but not acked, but we haven't been
17196 		 * called to retransmit, len will be < 0.  Otherwise, window
17197 		 * shrank after we sent into it.  If window shrank to 0,
17198 		 * cancel pending retransmit, pull snd_nxt back to (closed)
17199 		 * window, and set the persist timer if it isn't already
17200 		 * going.  If the window didn't close completely, just wait
17201 		 * for an ACK.
17202 		 *
17203 		 * We also do a general check here to ensure that we will
17204 		 * set the persist timer when we have data to send, but a
17205 		 * 0-byte window. This makes sure the persist timer is set
17206 		 * even if the packet hits one of the "goto send" lines
17207 		 * below.
17208 		 */
17209 		len = 0;
17210 		if ((tp->snd_wnd == 0) &&
17211 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17212 		    (tp->snd_una == tp->snd_max) &&
17213 		    (sb_offset < (int)sbavail(sb))) {
17214 			rack_enter_persist(tp, rack, cts);
17215 		}
17216 	} else if ((rsm == NULL) &&
17217 		   (doing_tlp == 0) &&
17218 		   (len < pace_max_seg)) {
17219 		/*
17220 		 * We are not sending a maximum sized segment for
17221 		 * some reason. Should we not send anything (think
17222 		 * sws or persists)?
17223 		 */
17224 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17225 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17226 		    (len < minseg) &&
17227 		    (len < (int)(sbavail(sb) - sb_offset))) {
17228 			/*
17229 			 * Here the rwnd is less than
17230 			 * the minimum pacing size, this is not a retransmit,
17231 			 * we are established and
17232 			 * the send is not the last in the socket buffer
17233 			 * we send nothing, and we may enter persists
17234 			 * if nothing is outstanding.
17235 			 */
17236 			len = 0;
17237 			if (tp->snd_max == tp->snd_una) {
17238 				/*
17239 				 * Nothing out we can
17240 				 * go into persists.
17241 				 */
17242 				rack_enter_persist(tp, rack, cts);
17243 			}
17244 		     } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17245 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17246 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17247 			   (len < minseg)) {
17248 			/*
17249 			 * Here we are not retransmitting, and
17250 			 * the cwnd is not so small that we could
17251 			 * not send at least a min size (rxt timer
17252 			 * not having gone off), We have 2 segments or
17253 			 * more already in flight, its not the tail end
17254 			 * of the socket buffer  and the cwnd is blocking
17255 			 * us from sending out a minimum pacing segment size.
17256 			 * Lets not send anything.
17257 			 */
17258 			len = 0;
17259 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17260 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17261 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17262 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17263 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
17264 			/*
17265 			 * Here we have a send window but we have
17266 			 * filled it up and we can't send another pacing segment.
17267 			 * We also have in flight more than 2 segments
17268 			 * and we are not completing the sb i.e. we allow
17269 			 * the last bytes of the sb to go out even if
17270 			 * its not a full pacing segment.
17271 			 */
17272 			len = 0;
17273 		} else if ((rack->r_ctl.crte != NULL) &&
17274 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17275 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17276 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17277 			   (len < (int)(sbavail(sb) - sb_offset))) {
17278 			/*
17279 			 * Here we are doing hardware pacing, this is not a TLP,
17280 			 * we are not sending a pace max segment size, there is rwnd
17281 			 * room to send at least N pace_max_seg, the cwnd is greater
17282 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
17283 			 * more segments in flight and its not the tail of the socket buffer.
17284 			 *
17285 			 * We don't want to send instead we need to get more ack's in to
17286 			 * allow us to send a full pacing segment. Normally, if we are pacing
17287 			 * about the right speed, we should have finished our pacing
17288 			 * send as most of the acks have come back if we are at the
17289 			 * right rate. This is a bit fuzzy since return path delay
17290 			 * can delay the acks, which is why we want to make sure we
17291 			 * have cwnd space to have a bit more than a max pace segments in flight.
17292 			 *
17293 			 * If we have not gotten our acks back we are pacing at too high a
17294 			 * rate delaying will not hurt and will bring our GP estimate down by
17295 			 * injecting the delay. If we don't do this we will send
17296 			 * 2 MSS out in response to the acks being clocked in which
17297 			 * defeats the point of hw-pacing (i.e. to help us get
17298 			 * larger TSO's out).
17299 			 */
17300 			len = 0;
17301 
17302 		}
17303 
17304 	}
17305 	/* len will be >= 0 after this point. */
17306 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17307 	rack_sndbuf_autoscale(rack);
17308 	/*
17309 	 * Decide if we can use TCP Segmentation Offloading (if supported by
17310 	 * hardware).
17311 	 *
17312 	 * TSO may only be used if we are in a pure bulk sending state.  The
17313 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17314 	 * options prevent using TSO.  With TSO the TCP header is the same
17315 	 * (except for the sequence number) for all generated packets.  This
17316 	 * makes it impossible to transmit any options which vary per
17317 	 * generated segment or packet.
17318 	 *
17319 	 * IPv4 handling has a clear separation of ip options and ip header
17320 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17321 	 * the right thing below to provide length of just ip options and thus
17322 	 * checking for ipoptlen is enough to decide if ip options are present.
17323 	 */
17324 	ipoptlen = 0;
17325 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17326 	/*
17327 	 * Pre-calculate here as we save another lookup into the darknesses
17328 	 * of IPsec that way and can actually decide if TSO is ok.
17329 	 */
17330 #ifdef INET6
17331 	if (isipv6 && IPSEC_ENABLED(ipv6))
17332 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
17333 #ifdef INET
17334 	else
17335 #endif
17336 #endif				/* INET6 */
17337 #ifdef INET
17338 		if (IPSEC_ENABLED(ipv4))
17339 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
17340 #endif				/* INET */
17341 #endif
17342 
17343 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17344 	ipoptlen += ipsec_optlen;
17345 #endif
17346 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17347 	    (tp->t_port == 0) &&
17348 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
17349 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17350 	    ipoptlen == 0)
17351 		tso = 1;
17352 	{
17353 		uint32_t outstanding;
17354 
17355 		outstanding = tp->snd_max - tp->snd_una;
17356 		if (tp->t_flags & TF_SENTFIN) {
17357 			/*
17358 			 * If we sent a fin, snd_max is 1 higher than
17359 			 * snd_una
17360 			 */
17361 			outstanding--;
17362 		}
17363 		if (sack_rxmit) {
17364 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17365 				flags &= ~TH_FIN;
17366 		} else {
17367 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17368 				   sbused(sb)))
17369 				flags &= ~TH_FIN;
17370 		}
17371 	}
17372 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17373 	    (long)TCP_MAXWIN << tp->rcv_scale);
17374 
17375 	/*
17376 	 * Sender silly window avoidance.   We transmit under the following
17377 	 * conditions when len is non-zero:
17378 	 *
17379 	 * - We have a full segment (or more with TSO) - This is the last
17380 	 * buffer in a write()/send() and we are either idle or running
17381 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
17382 	 * then 1/2 the maximum send window's worth of data (receiver may be
17383 	 * limited the window size) - we need to retransmit
17384 	 */
17385 	if (len) {
17386 		if (len >= segsiz) {
17387 			goto send;
17388 		}
17389 		/*
17390 		 * NOTE! on localhost connections an 'ack' from the remote
17391 		 * end may occur synchronously with the output and cause us
17392 		 * to flush a buffer queued with moretocome.  XXX
17393 		 *
17394 		 */
17395 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
17396 		    (idle || (tp->t_flags & TF_NODELAY)) &&
17397 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17398 		    (tp->t_flags & TF_NOPUSH) == 0) {
17399 			pass = 2;
17400 			goto send;
17401 		}
17402 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
17403 			pass = 22;
17404 			goto send;
17405 		}
17406 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17407 			pass = 4;
17408 			goto send;
17409 		}
17410 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
17411 			pass = 5;
17412 			goto send;
17413 		}
17414 		if (sack_rxmit) {
17415 			pass = 6;
17416 			goto send;
17417 		}
17418 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17419 		    (ctf_outstanding(tp) < (segsiz * 2))) {
17420 			/*
17421 			 * We have less than two MSS outstanding (delayed ack)
17422 			 * and our rwnd will not let us send a full sized
17423 			 * MSS. Lets go ahead and let this small segment
17424 			 * out because we want to try to have at least two
17425 			 * packets inflight to not be caught by delayed ack.
17426 			 */
17427 			pass = 12;
17428 			goto send;
17429 		}
17430 	}
17431 	/*
17432 	 * Sending of standalone window updates.
17433 	 *
17434 	 * Window updates are important when we close our window due to a
17435 	 * full socket buffer and are opening it again after the application
17436 	 * reads data from it.  Once the window has opened again and the
17437 	 * remote end starts to send again the ACK clock takes over and
17438 	 * provides the most current window information.
17439 	 *
17440 	 * We must avoid the silly window syndrome whereas every read from
17441 	 * the receive buffer, no matter how small, causes a window update
17442 	 * to be sent.  We also should avoid sending a flurry of window
17443 	 * updates when the socket buffer had queued a lot of data and the
17444 	 * application is doing small reads.
17445 	 *
17446 	 * Prevent a flurry of pointless window updates by only sending an
17447 	 * update when we can increase the advertized window by more than
17448 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
17449 	 * full or is very small be more aggressive and send an update
17450 	 * whenever we can increase by two mss sized segments. In all other
17451 	 * situations the ACK's to new incoming data will carry further
17452 	 * window increases.
17453 	 *
17454 	 * Don't send an independent window update if a delayed ACK is
17455 	 * pending (it will get piggy-backed on it) or the remote side
17456 	 * already has done a half-close and won't send more data.  Skip
17457 	 * this if the connection is in T/TCP half-open state.
17458 	 */
17459 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17460 	    !(tp->t_flags & TF_DELACK) &&
17461 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
17462 		/*
17463 		 * "adv" is the amount we could increase the window, taking
17464 		 * into account that we are limited by TCP_MAXWIN <<
17465 		 * tp->rcv_scale.
17466 		 */
17467 		int32_t adv;
17468 		int oldwin;
17469 
17470 		adv = recwin;
17471 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17472 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
17473 			if (adv > oldwin)
17474 			    adv -= oldwin;
17475 			else {
17476 				/* We can't increase the window */
17477 				adv = 0;
17478 			}
17479 		} else
17480 			oldwin = 0;
17481 
17482 		/*
17483 		 * If the new window size ends up being the same as or less
17484 		 * than the old size when it is scaled, then don't force
17485 		 * a window update.
17486 		 */
17487 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17488 			goto dontupdate;
17489 
17490 		if (adv >= (int32_t)(2 * segsiz) &&
17491 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17492 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17493 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17494 			pass = 7;
17495 			goto send;
17496 		}
17497 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17498 			pass = 23;
17499 			goto send;
17500 		}
17501 	}
17502 dontupdate:
17503 
17504 	/*
17505 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17506 	 * is also a catch-all for the retransmit timer timeout case.
17507 	 */
17508 	if (tp->t_flags & TF_ACKNOW) {
17509 		pass = 8;
17510 		goto send;
17511 	}
17512 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17513 		pass = 9;
17514 		goto send;
17515 	}
17516 	/*
17517 	 * If our state indicates that FIN should be sent and we have not
17518 	 * yet done so, then we need to send.
17519 	 */
17520 	if ((flags & TH_FIN) &&
17521 	    (tp->snd_nxt == tp->snd_una)) {
17522 		pass = 11;
17523 		goto send;
17524 	}
17525 	/*
17526 	 * No reason to send a segment, just return.
17527 	 */
17528 just_return:
17529 	SOCKBUF_UNLOCK(sb);
17530 just_return_nolock:
17531 	{
17532 		int app_limited = CTF_JR_SENT_DATA;
17533 
17534 		if (tot_len_this_send > 0) {
17535 			/* Make sure snd_nxt is up to max */
17536 			rack->r_ctl.fsb.recwin = recwin;
17537 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17538 			if ((error == 0) &&
17539 			    rack_use_rfo &&
17540 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
17541 			    (ipoptlen == 0) &&
17542 			    (tp->snd_nxt == tp->snd_max) &&
17543 			    (tp->rcv_numsacks == 0) &&
17544 			    rack->r_fsb_inited &&
17545 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
17546 			    (rack->r_must_retran == 0) &&
17547 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
17548 			    (len > 0) && (orig_len > 0) &&
17549 			    (orig_len > len) &&
17550 			    ((orig_len - len) >= segsiz) &&
17551 			    ((optlen == 0) ||
17552 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17553 				/* We can send at least one more MSS using our fsb */
17554 
17555 				rack->r_fast_output = 1;
17556 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17557 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17558 				rack->r_ctl.fsb.tcp_flags = flags;
17559 				rack->r_ctl.fsb.left_to_send = orig_len - len;
17560 				if (hw_tls)
17561 					rack->r_ctl.fsb.hw_tls = 1;
17562 				else
17563 					rack->r_ctl.fsb.hw_tls = 0;
17564 				KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17565 					("rack:%p left_to_send:%u sbavail:%u out:%u",
17566 					rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17567 					 (tp->snd_max - tp->snd_una)));
17568 				if (rack->r_ctl.fsb.left_to_send < segsiz)
17569 					rack->r_fast_output = 0;
17570 				else {
17571 					if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17572 						rack->r_ctl.fsb.rfo_apply_push = 1;
17573 					else
17574 						rack->r_ctl.fsb.rfo_apply_push = 0;
17575 				}
17576 			} else
17577 				rack->r_fast_output = 0;
17578 
17579 
17580 			rack_log_fsb(rack, tp, so, flags,
17581 				     ipoptlen, orig_len, len, 0,
17582 				     1, optlen, __LINE__, 1);
17583 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17584 				tp->snd_nxt = tp->snd_max;
17585 		} else {
17586 			int end_window = 0;
17587 			uint32_t seq = tp->gput_ack;
17588 
17589 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17590 			if (rsm) {
17591 				/*
17592 				 * Mark the last sent that we just-returned (hinting
17593 				 * that delayed ack may play a role in any rtt measurement).
17594 				 */
17595 				rsm->r_just_ret = 1;
17596 			}
17597 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17598 			rack->r_ctl.rc_agg_delayed = 0;
17599 			rack->r_early = 0;
17600 			rack->r_late = 0;
17601 			rack->r_ctl.rc_agg_early = 0;
17602 			if ((ctf_outstanding(tp) +
17603 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17604 				 minseg)) >= tp->snd_wnd) {
17605 				/* We are limited by the rwnd */
17606 				app_limited = CTF_JR_RWND_LIMITED;
17607 				if (IN_FASTRECOVERY(tp->t_flags))
17608 				    rack->r_ctl.rc_prr_sndcnt = 0;
17609 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
17610 				/* We are limited by whats available -- app limited */
17611 				app_limited = CTF_JR_APP_LIMITED;
17612 				if (IN_FASTRECOVERY(tp->t_flags))
17613 				    rack->r_ctl.rc_prr_sndcnt = 0;
17614 			} else if ((idle == 0) &&
17615 				   ((tp->t_flags & TF_NODELAY) == 0) &&
17616 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17617 				   (len < segsiz)) {
17618 				/*
17619 				 * No delay is not on and the
17620 				 * user is sending less than 1MSS. This
17621 				 * brings out SWS avoidance so we
17622 				 * don't send. Another app-limited case.
17623 				 */
17624 				app_limited = CTF_JR_APP_LIMITED;
17625 			} else if (tp->t_flags & TF_NOPUSH) {
17626 				/*
17627 				 * The user has requested no push of
17628 				 * the last segment and we are
17629 				 * at the last segment. Another app
17630 				 * limited case.
17631 				 */
17632 				app_limited = CTF_JR_APP_LIMITED;
17633 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17634 				/* Its the cwnd */
17635 				app_limited = CTF_JR_CWND_LIMITED;
17636 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
17637 				   (rack->rack_no_prr == 0) &&
17638 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17639 				app_limited = CTF_JR_PRR;
17640 			} else {
17641 				/* Now why here are we not sending? */
17642 #ifdef NOW
17643 #ifdef INVARIANTS
17644 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17645 #endif
17646 #endif
17647 				app_limited = CTF_JR_ASSESSING;
17648 			}
17649 			/*
17650 			 * App limited in some fashion, for our pacing GP
17651 			 * measurements we don't want any gap (even cwnd).
17652 			 * Close  down the measurement window.
17653 			 */
17654 			if (rack_cwnd_block_ends_measure &&
17655 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
17656 			     (app_limited == CTF_JR_PRR))) {
17657 				/*
17658 				 * The reason we are not sending is
17659 				 * the cwnd (or prr). We have been configured
17660 				 * to end the measurement window in
17661 				 * this case.
17662 				 */
17663 				end_window = 1;
17664 			} else if (rack_rwnd_block_ends_measure &&
17665 				   (app_limited == CTF_JR_RWND_LIMITED)) {
17666 				/*
17667 				 * We are rwnd limited and have been
17668 				 * configured to end the measurement
17669 				 * window in this case.
17670 				 */
17671 				end_window = 1;
17672 			} else if (app_limited == CTF_JR_APP_LIMITED) {
17673 				/*
17674 				 * A true application limited period, we have
17675 				 * ran out of data.
17676 				 */
17677 				end_window = 1;
17678 			} else if (app_limited == CTF_JR_ASSESSING) {
17679 				/*
17680 				 * In the assessing case we hit the end of
17681 				 * the if/else and had no known reason
17682 				 * This will panic us under invariants..
17683 				 *
17684 				 * If we get this out in logs we need to
17685 				 * investagate which reason we missed.
17686 				 */
17687 				end_window = 1;
17688 			}
17689 			if (end_window) {
17690 				uint8_t log = 0;
17691 
17692 				/* Adjust the Gput measurement */
17693 				if ((tp->t_flags & TF_GPUTINPROG) &&
17694 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
17695 					tp->gput_ack = tp->snd_max;
17696 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17697 						/*
17698 						 * There is not enough to measure.
17699 						 */
17700 						tp->t_flags &= ~TF_GPUTINPROG;
17701 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17702 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
17703 									   tp->gput_seq,
17704 									   0, 0, 18, __LINE__, NULL, 0);
17705 					} else
17706 						log = 1;
17707 				}
17708 				/* Mark the last packet has app limited */
17709 				rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17710 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17711 					if (rack->r_ctl.rc_app_limited_cnt == 0)
17712 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17713 					else {
17714 						/*
17715 						 * Go out to the end app limited and mark
17716 						 * this new one as next and move the end_appl up
17717 						 * to this guy.
17718 						 */
17719 						if (rack->r_ctl.rc_end_appl)
17720 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17721 						rack->r_ctl.rc_end_appl = rsm;
17722 					}
17723 					rsm->r_flags |= RACK_APP_LIMITED;
17724 					rack->r_ctl.rc_app_limited_cnt++;
17725 				}
17726 				if (log)
17727 					rack_log_pacing_delay_calc(rack,
17728 								   rack->r_ctl.rc_app_limited_cnt, seq,
17729 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17730 			}
17731 		}
17732 		/* Check if we need to go into persists or not */
17733 		if ((tp->snd_max == tp->snd_una) &&
17734 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
17735 		    sbavail(sb) &&
17736 		    (sbavail(sb) > tp->snd_wnd) &&
17737 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17738 			/* Yes lets make sure to move to persist before timer-start */
17739 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17740 		}
17741 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17742 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17743 	}
17744 #ifdef NETFLIX_SHARED_CWND
17745 	if ((sbavail(sb) == 0) &&
17746 	    rack->r_ctl.rc_scw) {
17747 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17748 		rack->rack_scwnd_is_idle = 1;
17749 	}
17750 #endif
17751 #ifdef TCP_ACCOUNTING
17752 	if (tot_len_this_send > 0) {
17753 		crtsc = get_cyclecount();
17754 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17755 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
17756 		}
17757 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17758 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17759 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17760 		}
17761 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17762 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17763 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17764 		}
17765 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17766 	} else {
17767 		crtsc = get_cyclecount();
17768 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17769 			tp->tcp_cnt_counters[SND_LIMITED]++;
17770 		}
17771 		counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17772 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17773 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17774 		}
17775 		counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17776 	}
17777 	sched_unpin();
17778 #endif
17779 	return (0);
17780 
17781 send:
17782 	if (rsm || sack_rxmit)
17783 		counter_u64_add(rack_nfto_resend, 1);
17784 	else
17785 		counter_u64_add(rack_non_fto_send, 1);
17786 	if ((flags & TH_FIN) &&
17787 	    sbavail(sb)) {
17788 		/*
17789 		 * We do not transmit a FIN
17790 		 * with data outstanding. We
17791 		 * need to make it so all data
17792 		 * is acked first.
17793 		 */
17794 		flags &= ~TH_FIN;
17795 	}
17796 	/* Enforce stack imposed max seg size if we have one */
17797 	if (rack->r_ctl.rc_pace_max_segs &&
17798 	    (len > rack->r_ctl.rc_pace_max_segs)) {
17799 		mark = 1;
17800 		len = rack->r_ctl.rc_pace_max_segs;
17801 	}
17802 	SOCKBUF_LOCK_ASSERT(sb);
17803 	if (len > 0) {
17804 		if (len >= segsiz)
17805 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17806 		else
17807 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17808 	}
17809 	/*
17810 	 * Before ESTABLISHED, force sending of initial options unless TCP
17811 	 * set not to do any options. NOTE: we assume that the IP/TCP header
17812 	 * plus TCP options always fit in a single mbuf, leaving room for a
17813 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17814 	 * + optlen <= MCLBYTES
17815 	 */
17816 	optlen = 0;
17817 #ifdef INET6
17818 	if (isipv6)
17819 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17820 	else
17821 #endif
17822 		hdrlen = sizeof(struct tcpiphdr);
17823 
17824 	/*
17825 	 * Compute options for segment. We only have to care about SYN and
17826 	 * established connection segments.  Options for SYN-ACK segments
17827 	 * are handled in TCP syncache.
17828 	 */
17829 	to.to_flags = 0;
17830 	if ((tp->t_flags & TF_NOOPT) == 0) {
17831 		/* Maximum segment size. */
17832 		if (flags & TH_SYN) {
17833 			tp->snd_nxt = tp->iss;
17834 			to.to_mss = tcp_mssopt(&inp->inp_inc);
17835 			if (tp->t_port)
17836 				to.to_mss -= V_tcp_udp_tunneling_overhead;
17837 			to.to_flags |= TOF_MSS;
17838 
17839 			/*
17840 			 * On SYN or SYN|ACK transmits on TFO connections,
17841 			 * only include the TFO option if it is not a
17842 			 * retransmit, as the presence of the TFO option may
17843 			 * have caused the original SYN or SYN|ACK to have
17844 			 * been dropped by a middlebox.
17845 			 */
17846 			if (IS_FASTOPEN(tp->t_flags) &&
17847 			    (tp->t_rxtshift == 0)) {
17848 				if (tp->t_state == TCPS_SYN_RECEIVED) {
17849 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17850 					to.to_tfo_cookie =
17851 						(u_int8_t *)&tp->t_tfo_cookie.server;
17852 					to.to_flags |= TOF_FASTOPEN;
17853 					wanted_cookie = 1;
17854 				} else if (tp->t_state == TCPS_SYN_SENT) {
17855 					to.to_tfo_len =
17856 						tp->t_tfo_client_cookie_len;
17857 					to.to_tfo_cookie =
17858 						tp->t_tfo_cookie.client;
17859 					to.to_flags |= TOF_FASTOPEN;
17860 					wanted_cookie = 1;
17861 					/*
17862 					 * If we wind up having more data to
17863 					 * send with the SYN than can fit in
17864 					 * one segment, don't send any more
17865 					 * until the SYN|ACK comes back from
17866 					 * the other end.
17867 					 */
17868 					sendalot = 0;
17869 				}
17870 			}
17871 		}
17872 		/* Window scaling. */
17873 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17874 			to.to_wscale = tp->request_r_scale;
17875 			to.to_flags |= TOF_SCALE;
17876 		}
17877 		/* Timestamps. */
17878 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
17879 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17880 			to.to_tsval = ms_cts + tp->ts_offset;
17881 			to.to_tsecr = tp->ts_recent;
17882 			to.to_flags |= TOF_TS;
17883 		}
17884 		/* Set receive buffer autosizing timestamp. */
17885 		if (tp->rfbuf_ts == 0 &&
17886 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
17887 			tp->rfbuf_ts = tcp_ts_getticks();
17888 		/* Selective ACK's. */
17889 		if (tp->t_flags & TF_SACK_PERMIT) {
17890 			if (flags & TH_SYN)
17891 				to.to_flags |= TOF_SACKPERM;
17892 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17893 				 tp->rcv_numsacks > 0) {
17894 				to.to_flags |= TOF_SACK;
17895 				to.to_nsacks = tp->rcv_numsacks;
17896 				to.to_sacks = (u_char *)tp->sackblks;
17897 			}
17898 		}
17899 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17900 		/* TCP-MD5 (RFC2385). */
17901 		if (tp->t_flags & TF_SIGNATURE)
17902 			to.to_flags |= TOF_SIGNATURE;
17903 #endif				/* TCP_SIGNATURE */
17904 
17905 		/* Processing the options. */
17906 		hdrlen += optlen = tcp_addoptions(&to, opt);
17907 		/*
17908 		 * If we wanted a TFO option to be added, but it was unable
17909 		 * to fit, ensure no data is sent.
17910 		 */
17911 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17912 		    !(to.to_flags & TOF_FASTOPEN))
17913 			len = 0;
17914 	}
17915 	if (tp->t_port) {
17916 		if (V_tcp_udp_tunneling_port == 0) {
17917 			/* The port was removed?? */
17918 			SOCKBUF_UNLOCK(&so->so_snd);
17919 #ifdef TCP_ACCOUNTING
17920 			crtsc = get_cyclecount();
17921 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17922 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17923 			}
17924 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
17925 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17926 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17927 			}
17928 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
17929 			sched_unpin();
17930 #endif
17931 			return (EHOSTUNREACH);
17932 		}
17933 		hdrlen += sizeof(struct udphdr);
17934 	}
17935 #ifdef INET6
17936 	if (isipv6)
17937 		ipoptlen = ip6_optlen(tp->t_inpcb);
17938 	else
17939 #endif
17940 		if (tp->t_inpcb->inp_options)
17941 			ipoptlen = tp->t_inpcb->inp_options->m_len -
17942 				offsetof(struct ipoption, ipopt_list);
17943 		else
17944 			ipoptlen = 0;
17945 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17946 	ipoptlen += ipsec_optlen;
17947 #endif
17948 
17949 	/*
17950 	 * Adjust data length if insertion of options will bump the packet
17951 	 * length beyond the t_maxseg length. Clear the FIN bit because we
17952 	 * cut off the tail of the segment.
17953 	 */
17954 	if (len + optlen + ipoptlen > tp->t_maxseg) {
17955 		if (tso) {
17956 			uint32_t if_hw_tsomax;
17957 			uint32_t moff;
17958 			int32_t max_len;
17959 
17960 			/* extract TSO information */
17961 			if_hw_tsomax = tp->t_tsomax;
17962 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
17963 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
17964 			KASSERT(ipoptlen == 0,
17965 				("%s: TSO can't do IP options", __func__));
17966 
17967 			/*
17968 			 * Check if we should limit by maximum payload
17969 			 * length:
17970 			 */
17971 			if (if_hw_tsomax != 0) {
17972 				/* compute maximum TSO length */
17973 				max_len = (if_hw_tsomax - hdrlen -
17974 					   max_linkhdr);
17975 				if (max_len <= 0) {
17976 					len = 0;
17977 				} else if (len > max_len) {
17978 					sendalot = 1;
17979 					len = max_len;
17980 					mark = 2;
17981 				}
17982 			}
17983 			/*
17984 			 * Prevent the last segment from being fractional
17985 			 * unless the send sockbuf can be emptied:
17986 			 */
17987 			max_len = (tp->t_maxseg - optlen);
17988 			if ((sb_offset + len) < sbavail(sb)) {
17989 				moff = len % (u_int)max_len;
17990 				if (moff != 0) {
17991 					mark = 3;
17992 					len -= moff;
17993 				}
17994 			}
17995 			/*
17996 			 * In case there are too many small fragments don't
17997 			 * use TSO:
17998 			 */
17999 			if (len <= segsiz) {
18000 				mark = 4;
18001 				tso = 0;
18002 			}
18003 			/*
18004 			 * Send the FIN in a separate segment after the bulk
18005 			 * sending is done. We don't trust the TSO
18006 			 * implementations to clear the FIN flag on all but
18007 			 * the last segment.
18008 			 */
18009 			if (tp->t_flags & TF_NEEDFIN) {
18010 				sendalot = 4;
18011 			}
18012 		} else {
18013 			mark = 5;
18014 			if (optlen + ipoptlen >= tp->t_maxseg) {
18015 				/*
18016 				 * Since we don't have enough space to put
18017 				 * the IP header chain and the TCP header in
18018 				 * one packet as required by RFC 7112, don't
18019 				 * send it. Also ensure that at least one
18020 				 * byte of the payload can be put into the
18021 				 * TCP segment.
18022 				 */
18023 				SOCKBUF_UNLOCK(&so->so_snd);
18024 				error = EMSGSIZE;
18025 				sack_rxmit = 0;
18026 				goto out;
18027 			}
18028 			len = tp->t_maxseg - optlen - ipoptlen;
18029 			sendalot = 5;
18030 		}
18031 	} else {
18032 		tso = 0;
18033 		mark = 6;
18034 	}
18035 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18036 		("%s: len > IP_MAXPACKET", __func__));
18037 #ifdef DIAGNOSTIC
18038 #ifdef INET6
18039 	if (max_linkhdr + hdrlen > MCLBYTES)
18040 #else
18041 		if (max_linkhdr + hdrlen > MHLEN)
18042 #endif
18043 			panic("tcphdr too big");
18044 #endif
18045 
18046 	/*
18047 	 * This KASSERT is here to catch edge cases at a well defined place.
18048 	 * Before, those had triggered (random) panic conditions further
18049 	 * down.
18050 	 */
18051 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18052 	if ((len == 0) &&
18053 	    (flags & TH_FIN) &&
18054 	    (sbused(sb))) {
18055 		/*
18056 		 * We have outstanding data, don't send a fin by itself!.
18057 		 */
18058 		goto just_return;
18059 	}
18060 	/*
18061 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
18062 	 * and initialize the header from the template for sends on this
18063 	 * connection.
18064 	 */
18065 	hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
18066 	if (len) {
18067 		uint32_t max_val;
18068 		uint32_t moff;
18069 
18070 		if (rack->r_ctl.rc_pace_max_segs)
18071 			max_val = rack->r_ctl.rc_pace_max_segs;
18072 		else if (rack->rc_user_set_max_segs)
18073 			max_val = rack->rc_user_set_max_segs * segsiz;
18074 		else
18075 			max_val = len;
18076 		/*
18077 		 * We allow a limit on sending with hptsi.
18078 		 */
18079 		if (len > max_val) {
18080 			mark = 7;
18081 			len = max_val;
18082 		}
18083 #ifdef INET6
18084 		if (MHLEN < hdrlen + max_linkhdr)
18085 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18086 		else
18087 #endif
18088 			m = m_gethdr(M_NOWAIT, MT_DATA);
18089 
18090 		if (m == NULL) {
18091 			SOCKBUF_UNLOCK(sb);
18092 			error = ENOBUFS;
18093 			sack_rxmit = 0;
18094 			goto out;
18095 		}
18096 		m->m_data += max_linkhdr;
18097 		m->m_len = hdrlen;
18098 
18099 		/*
18100 		 * Start the m_copy functions from the closest mbuf to the
18101 		 * sb_offset in the socket buffer chain.
18102 		 */
18103 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
18104 		s_mb = mb;
18105 		s_moff = moff;
18106 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18107 			m_copydata(mb, moff, (int)len,
18108 				   mtod(m, caddr_t)+hdrlen);
18109 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18110 				sbsndptr_adv(sb, mb, len);
18111 			m->m_len += len;
18112 		} else {
18113 			struct sockbuf *msb;
18114 
18115 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18116 				msb = NULL;
18117 			else
18118 				msb = sb;
18119 			m->m_next = tcp_m_copym(
18120 				mb, moff, &len,
18121 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18122 				((rsm == NULL) ? hw_tls : 0)
18123 #ifdef NETFLIX_COPY_ARGS
18124 				, &filled_all
18125 #endif
18126 				);
18127 			if (len <= (tp->t_maxseg - optlen)) {
18128 				/*
18129 				 * Must have ran out of mbufs for the copy
18130 				 * shorten it to no longer need tso. Lets
18131 				 * not put on sendalot since we are low on
18132 				 * mbufs.
18133 				 */
18134 				tso = 0;
18135 			}
18136 			if (m->m_next == NULL) {
18137 				SOCKBUF_UNLOCK(sb);
18138 				(void)m_free(m);
18139 				error = ENOBUFS;
18140 				sack_rxmit = 0;
18141 				goto out;
18142 			}
18143 		}
18144 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18145 			if (rsm && (rsm->r_flags & RACK_TLP)) {
18146 				/*
18147 				 * TLP should not count in retran count, but
18148 				 * in its own bin
18149 				 */
18150 				counter_u64_add(rack_tlp_retran, 1);
18151 				counter_u64_add(rack_tlp_retran_bytes, len);
18152 			} else {
18153 				tp->t_sndrexmitpack++;
18154 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18155 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18156 			}
18157 #ifdef STATS
18158 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18159 						 len);
18160 #endif
18161 		} else {
18162 			KMOD_TCPSTAT_INC(tcps_sndpack);
18163 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18164 #ifdef STATS
18165 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18166 						 len);
18167 #endif
18168 		}
18169 		/*
18170 		 * If we're sending everything we've got, set PUSH. (This
18171 		 * will keep happy those implementations which only give
18172 		 * data to the user when a buffer fills or a PUSH comes in.)
18173 		 */
18174 		if (sb_offset + len == sbused(sb) &&
18175 		    sbused(sb) &&
18176 		    !(flags & TH_SYN)) {
18177 			flags |= TH_PUSH;
18178 			add_flag |= RACK_HAD_PUSH;
18179 		}
18180 
18181 		SOCKBUF_UNLOCK(sb);
18182 	} else {
18183 		SOCKBUF_UNLOCK(sb);
18184 		if (tp->t_flags & TF_ACKNOW)
18185 			KMOD_TCPSTAT_INC(tcps_sndacks);
18186 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
18187 			KMOD_TCPSTAT_INC(tcps_sndctrl);
18188 		else
18189 			KMOD_TCPSTAT_INC(tcps_sndwinup);
18190 
18191 		m = m_gethdr(M_NOWAIT, MT_DATA);
18192 		if (m == NULL) {
18193 			error = ENOBUFS;
18194 			sack_rxmit = 0;
18195 			goto out;
18196 		}
18197 #ifdef INET6
18198 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18199 		    MHLEN >= hdrlen) {
18200 			M_ALIGN(m, hdrlen);
18201 		} else
18202 #endif
18203 			m->m_data += max_linkhdr;
18204 		m->m_len = hdrlen;
18205 	}
18206 	SOCKBUF_UNLOCK_ASSERT(sb);
18207 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18208 #ifdef MAC
18209 	mac_inpcb_create_mbuf(inp, m);
18210 #endif
18211 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18212 #ifdef INET6
18213 		if (isipv6)
18214 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18215 		else
18216 #endif				/* INET6 */
18217 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18218 		th = rack->r_ctl.fsb.th;
18219 		udp = rack->r_ctl.fsb.udp;
18220 		if (udp) {
18221 #ifdef INET6
18222 			if (isipv6)
18223 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18224 			else
18225 #endif				/* INET6 */
18226 				ulen = hdrlen + len - sizeof(struct ip);
18227 			udp->uh_ulen = htons(ulen);
18228 		}
18229 	} else {
18230 #ifdef INET6
18231 		if (isipv6) {
18232 			ip6 = mtod(m, struct ip6_hdr *);
18233 			if (tp->t_port) {
18234 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18235 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18236 				udp->uh_dport = tp->t_port;
18237 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18238 				udp->uh_ulen = htons(ulen);
18239 				th = (struct tcphdr *)(udp + 1);
18240 			} else
18241 				th = (struct tcphdr *)(ip6 + 1);
18242 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
18243 		} else
18244 #endif				/* INET6 */
18245 		{
18246 			ip = mtod(m, struct ip *);
18247 #ifdef TCPDEBUG
18248 			ipov = (struct ipovly *)ip;
18249 #endif
18250 			if (tp->t_port) {
18251 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18252 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18253 				udp->uh_dport = tp->t_port;
18254 				ulen = hdrlen + len - sizeof(struct ip);
18255 				udp->uh_ulen = htons(ulen);
18256 				th = (struct tcphdr *)(udp + 1);
18257 			} else
18258 				th = (struct tcphdr *)(ip + 1);
18259 			tcpip_fillheaders(inp, tp->t_port, ip, th);
18260 		}
18261 	}
18262 	/*
18263 	 * Fill in fields, remembering maximum advertised window for use in
18264 	 * delaying messages about window sizes. If resending a FIN, be sure
18265 	 * not to use a new sequence number.
18266 	 */
18267 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18268 	    tp->snd_nxt == tp->snd_max)
18269 		tp->snd_nxt--;
18270 	/*
18271 	 * If we are starting a connection, send ECN setup SYN packet. If we
18272 	 * are on a retransmit, we may resend those bits a number of times
18273 	 * as per RFC 3168.
18274 	 */
18275 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
18276 		flags |= tcp_ecn_output_syn_sent(tp);
18277 	}
18278 	/* Also handle parallel SYN for ECN */
18279 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18280 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
18281 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
18282 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18283 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18284 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18285 #ifdef INET6
18286 		if (isipv6) {
18287 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18288 			ip6->ip6_flow |= htonl(ect << 20);
18289 		}
18290 		else
18291 #endif
18292 		{
18293 			ip->ip_tos &= ~IPTOS_ECN_MASK;
18294 			ip->ip_tos |= ect;
18295 		}
18296 	}
18297 	/*
18298 	 * If we are doing retransmissions, then snd_nxt will not reflect
18299 	 * the first unsent octet.  For ACK only packets, we do not want the
18300 	 * sequence number of the retransmitted packet, we want the sequence
18301 	 * number of the next unsent octet.  So, if there is no data (and no
18302 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
18303 	 * ti_seq.  But if we are in persist state, snd_max might reflect
18304 	 * one byte beyond the right edge of the window, so use snd_nxt in
18305 	 * that case, since we know we aren't doing a retransmission.
18306 	 * (retransmit and persist are mutually exclusive...)
18307 	 */
18308 	if (sack_rxmit == 0) {
18309 		if (len || (flags & (TH_SYN | TH_FIN))) {
18310 			th->th_seq = htonl(tp->snd_nxt);
18311 			rack_seq = tp->snd_nxt;
18312 		} else {
18313 			th->th_seq = htonl(tp->snd_max);
18314 			rack_seq = tp->snd_max;
18315 		}
18316 	} else {
18317 		th->th_seq = htonl(rsm->r_start);
18318 		rack_seq = rsm->r_start;
18319 	}
18320 	th->th_ack = htonl(tp->rcv_nxt);
18321 	tcp_set_flags(th, flags);
18322 	/*
18323 	 * Calculate receive window.  Don't shrink window, but avoid silly
18324 	 * window syndrome.
18325 	 * If a RST segment is sent, advertise a window of zero.
18326 	 */
18327 	if (flags & TH_RST) {
18328 		recwin = 0;
18329 	} else {
18330 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18331 		    recwin < (long)segsiz) {
18332 			recwin = 0;
18333 		}
18334 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18335 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18336 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18337 	}
18338 
18339 	/*
18340 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18341 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18342 	 * handled in syncache.
18343 	 */
18344 	if (flags & TH_SYN)
18345 		th->th_win = htons((u_short)
18346 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18347 	else {
18348 		/* Avoid shrinking window with window scaling. */
18349 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
18350 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18351 	}
18352 	/*
18353 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18354 	 * window.  This may cause the remote transmitter to stall.  This
18355 	 * flag tells soreceive() to disable delayed acknowledgements when
18356 	 * draining the buffer.  This can occur if the receiver is
18357 	 * attempting to read more data than can be buffered prior to
18358 	 * transmitting on the connection.
18359 	 */
18360 	if (th->th_win == 0) {
18361 		tp->t_sndzerowin++;
18362 		tp->t_flags |= TF_RXWIN0SENT;
18363 	} else
18364 		tp->t_flags &= ~TF_RXWIN0SENT;
18365 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
18366 	/* Now are we using fsb?, if so copy the template data to the mbuf */
18367 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18368 		uint8_t *cpto;
18369 
18370 		cpto = mtod(m, uint8_t *);
18371 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18372 		/*
18373 		 * We have just copied in:
18374 		 * IP/IP6
18375 		 * <optional udphdr>
18376 		 * tcphdr (no options)
18377 		 *
18378 		 * We need to grab the correct pointers into the mbuf
18379 		 * for both the tcp header, and possibly the udp header (if tunneling).
18380 		 * We do this by using the offset in the copy buffer and adding it
18381 		 * to the mbuf base pointer (cpto).
18382 		 */
18383 #ifdef INET6
18384 		if (isipv6)
18385 			ip6 = mtod(m, struct ip6_hdr *);
18386 		else
18387 #endif				/* INET6 */
18388 			ip = mtod(m, struct ip *);
18389 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18390 		/* If we have a udp header lets set it into the mbuf as well */
18391 		if (udp)
18392 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18393 	}
18394 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18395 	if (to.to_flags & TOF_SIGNATURE) {
18396 		/*
18397 		 * Calculate MD5 signature and put it into the place
18398 		 * determined before.
18399 		 * NOTE: since TCP options buffer doesn't point into
18400 		 * mbuf's data, calculate offset and use it.
18401 		 */
18402 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18403 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18404 			/*
18405 			 * Do not send segment if the calculation of MD5
18406 			 * digest has failed.
18407 			 */
18408 			goto out;
18409 		}
18410 	}
18411 #endif
18412 	if (optlen) {
18413 		bcopy(opt, th + 1, optlen);
18414 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18415 	}
18416 	/*
18417 	 * Put TCP length in extended header, and then checksum extended
18418 	 * header and data.
18419 	 */
18420 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18421 #ifdef INET6
18422 	if (isipv6) {
18423 		/*
18424 		 * ip6_plen is not need to be filled now, and will be filled
18425 		 * in ip6_output.
18426 		 */
18427 		if (tp->t_port) {
18428 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18429 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18430 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18431 			th->th_sum = htons(0);
18432 			UDPSTAT_INC(udps_opackets);
18433 		} else {
18434 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18435 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18436 			th->th_sum = in6_cksum_pseudo(ip6,
18437 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18438 						      0);
18439 		}
18440 	}
18441 #endif
18442 #if defined(INET6) && defined(INET)
18443 	else
18444 #endif
18445 #ifdef INET
18446 	{
18447 		if (tp->t_port) {
18448 			m->m_pkthdr.csum_flags = CSUM_UDP;
18449 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18450 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18451 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18452 			th->th_sum = htons(0);
18453 			UDPSTAT_INC(udps_opackets);
18454 		} else {
18455 			m->m_pkthdr.csum_flags = CSUM_TCP;
18456 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18457 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18458 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18459 									IPPROTO_TCP + len + optlen));
18460 		}
18461 		/* IP version must be set here for ipv4/ipv6 checking later */
18462 		KASSERT(ip->ip_v == IPVERSION,
18463 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18464 	}
18465 #endif
18466 	/*
18467 	 * Enable TSO and specify the size of the segments. The TCP pseudo
18468 	 * header checksum is always provided. XXX: Fixme: This is currently
18469 	 * not the case for IPv6.
18470 	 */
18471 	if (tso) {
18472 		KASSERT(len > tp->t_maxseg - optlen,
18473 			("%s: len <= tso_segsz", __func__));
18474 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18475 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18476 	}
18477 	KASSERT(len + hdrlen == m_length(m, NULL),
18478 		("%s: mbuf chain different than expected: %d + %u != %u",
18479 		 __func__, len, hdrlen, m_length(m, NULL)));
18480 
18481 #ifdef TCP_HHOOK
18482 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18483 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
18484 #endif
18485 	/* We're getting ready to send; log now. */
18486 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18487 		union tcp_log_stackspecific log;
18488 
18489 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18490 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18491 		if (rack->rack_no_prr)
18492 			log.u_bbr.flex1 = 0;
18493 		else
18494 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18495 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18496 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18497 		log.u_bbr.flex4 = orig_len;
18498 		if (filled_all)
18499 			log.u_bbr.flex5 = 0x80000000;
18500 		else
18501 			log.u_bbr.flex5 = 0;
18502 		/* Save off the early/late values */
18503 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18504 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18505 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18506 		if (rsm || sack_rxmit) {
18507 			if (doing_tlp)
18508 				log.u_bbr.flex8 = 2;
18509 			else
18510 				log.u_bbr.flex8 = 1;
18511 		} else {
18512 			if (doing_tlp)
18513 				log.u_bbr.flex8 = 3;
18514 			else
18515 				log.u_bbr.flex8 = 0;
18516 		}
18517 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18518 		log.u_bbr.flex7 = mark;
18519 		log.u_bbr.flex7 <<= 8;
18520 		log.u_bbr.flex7 |= pass;
18521 		log.u_bbr.pkts_out = tp->t_maxseg;
18522 		log.u_bbr.timeStamp = cts;
18523 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18524 		log.u_bbr.lt_epoch = cwnd_to_use;
18525 		log.u_bbr.delivered = sendalot;
18526 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18527 				     len, &log, false, NULL, NULL, 0, &tv);
18528 	} else
18529 		lgb = NULL;
18530 
18531 	/*
18532 	 * Fill in IP length and desired time to live and send to IP level.
18533 	 * There should be a better way to handle ttl and tos; we could keep
18534 	 * them in the template, but need a way to checksum without them.
18535 	 */
18536 	/*
18537 	 * m->m_pkthdr.len should have been set before cksum calcuration,
18538 	 * because in6_cksum() need it.
18539 	 */
18540 #ifdef INET6
18541 	if (isipv6) {
18542 		/*
18543 		 * we separately set hoplimit for every segment, since the
18544 		 * user might want to change the value via setsockopt. Also,
18545 		 * desired default hop limit might be changed via Neighbor
18546 		 * Discovery.
18547 		 */
18548 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18549 
18550 		/*
18551 		 * Set the packet size here for the benefit of DTrace
18552 		 * probes. ip6_output() will set it properly; it's supposed
18553 		 * to include the option header lengths as well.
18554 		 */
18555 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18556 
18557 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18558 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18559 		else
18560 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18561 
18562 		if (tp->t_state == TCPS_SYN_SENT)
18563 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18564 
18565 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18566 		/* TODO: IPv6 IP6TOS_ECT bit on */
18567 		error = ip6_output(m,
18568 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18569 				   inp->in6p_outputopts,
18570 #else
18571 				   NULL,
18572 #endif
18573 				   &inp->inp_route6,
18574 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18575 				   NULL, NULL, inp);
18576 
18577 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18578 			mtu = inp->inp_route6.ro_nh->nh_mtu;
18579 	}
18580 #endif				/* INET6 */
18581 #if defined(INET) && defined(INET6)
18582 	else
18583 #endif
18584 #ifdef INET
18585 	{
18586 		ip->ip_len = htons(m->m_pkthdr.len);
18587 #ifdef INET6
18588 		if (inp->inp_vflag & INP_IPV6PROTO)
18589 			ip->ip_ttl = in6_selecthlim(inp, NULL);
18590 #endif				/* INET6 */
18591 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18592 		/*
18593 		 * If we do path MTU discovery, then we set DF on every
18594 		 * packet. This might not be the best thing to do according
18595 		 * to RFC3390 Section 2. However the tcp hostcache migitates
18596 		 * the problem so it affects only the first tcp connection
18597 		 * with a host.
18598 		 *
18599 		 * NB: Don't set DF on small MTU/MSS to have a safe
18600 		 * fallback.
18601 		 */
18602 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18603 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18604 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18605 				ip->ip_off |= htons(IP_DF);
18606 			}
18607 		} else {
18608 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18609 		}
18610 
18611 		if (tp->t_state == TCPS_SYN_SENT)
18612 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18613 
18614 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
18615 
18616 		error = ip_output(m,
18617 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18618 				  inp->inp_options,
18619 #else
18620 				  NULL,
18621 #endif
18622 				  &inp->inp_route,
18623 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18624 				  inp);
18625 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18626 			mtu = inp->inp_route.ro_nh->nh_mtu;
18627 	}
18628 #endif				/* INET */
18629 
18630 out:
18631 	if (lgb) {
18632 		lgb->tlb_errno = error;
18633 		lgb = NULL;
18634 	}
18635 	/*
18636 	 * In transmit state, time the transmission and arrange for the
18637 	 * retransmit.  In persist state, just set snd_max.
18638 	 */
18639 	if (error == 0) {
18640 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18641 		if (rsm && doing_tlp) {
18642 			rack->rc_last_sent_tlp_past_cumack = 0;
18643 			rack->rc_last_sent_tlp_seq_valid = 1;
18644 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18645 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18646 		}
18647 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
18648 		if (rsm && (doing_tlp == 0)) {
18649 			/* Set we retransmitted */
18650 			rack->rc_gp_saw_rec = 1;
18651 		} else {
18652 			if (cwnd_to_use > tp->snd_ssthresh) {
18653 				/* Set we sent in CA */
18654 				rack->rc_gp_saw_ca = 1;
18655 			} else {
18656 				/* Set we sent in SS */
18657 				rack->rc_gp_saw_ss = 1;
18658 			}
18659 		}
18660 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18661 		    (tp->t_flags & TF_SACK_PERMIT) &&
18662 		    tp->rcv_numsacks > 0)
18663 			tcp_clean_dsack_blocks(tp);
18664 		tot_len_this_send += len;
18665 		if (len == 0)
18666 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18667 		else if (len == 1) {
18668 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18669 		} else if (len > 1) {
18670 			int idx;
18671 
18672 			idx = (len / segsiz) + 3;
18673 			if (idx >= TCP_MSS_ACCT_ATIMER)
18674 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18675 			else
18676 				counter_u64_add(rack_out_size[idx], 1);
18677 		}
18678 	}
18679 	if ((rack->rack_no_prr == 0) &&
18680 	    sub_from_prr &&
18681 	    (error == 0)) {
18682 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18683 			rack->r_ctl.rc_prr_sndcnt -= len;
18684 		else
18685 			rack->r_ctl.rc_prr_sndcnt = 0;
18686 	}
18687 	sub_from_prr = 0;
18688 	if (doing_tlp) {
18689 		/* Make sure the TLP is added */
18690 		add_flag |= RACK_TLP;
18691 	} else if (rsm) {
18692 		/* If its a resend without TLP then it must not have the flag */
18693 		rsm->r_flags &= ~RACK_TLP;
18694 	}
18695 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18696 			rack_to_usec_ts(&tv),
18697 			rsm, add_flag, s_mb, s_moff, hw_tls);
18698 
18699 
18700 	if ((error == 0) &&
18701 	    (len > 0) &&
18702 	    (tp->snd_una == tp->snd_max))
18703 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
18704 	{
18705 		tcp_seq startseq = tp->snd_nxt;
18706 
18707 		/* Track our lost count */
18708 		if (rsm && (doing_tlp == 0))
18709 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18710 		/*
18711 		 * Advance snd_nxt over sequence space of this segment.
18712 		 */
18713 		if (error)
18714 			/* We don't log or do anything with errors */
18715 			goto nomore;
18716 		if (doing_tlp == 0) {
18717 			if (rsm == NULL) {
18718 				/*
18719 				 * Not a retransmission of some
18720 				 * sort, new data is going out so
18721 				 * clear our TLP count and flag.
18722 				 */
18723 				rack->rc_tlp_in_progress = 0;
18724 				rack->r_ctl.rc_tlp_cnt_out = 0;
18725 			}
18726 		} else {
18727 			/*
18728 			 * We have just sent a TLP, mark that it is true
18729 			 * and make sure our in progress is set so we
18730 			 * continue to check the count.
18731 			 */
18732 			rack->rc_tlp_in_progress = 1;
18733 			rack->r_ctl.rc_tlp_cnt_out++;
18734 		}
18735 		if (flags & (TH_SYN | TH_FIN)) {
18736 			if (flags & TH_SYN)
18737 				tp->snd_nxt++;
18738 			if (flags & TH_FIN) {
18739 				tp->snd_nxt++;
18740 				tp->t_flags |= TF_SENTFIN;
18741 			}
18742 		}
18743 		/* In the ENOBUFS case we do *not* update snd_max */
18744 		if (sack_rxmit)
18745 			goto nomore;
18746 
18747 		tp->snd_nxt += len;
18748 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18749 			if (tp->snd_una == tp->snd_max) {
18750 				/*
18751 				 * Update the time we just added data since
18752 				 * none was outstanding.
18753 				 */
18754 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18755 				tp->t_acktime = ticks;
18756 			}
18757 			tp->snd_max = tp->snd_nxt;
18758 			/*
18759 			 * Time this transmission if not a retransmission and
18760 			 * not currently timing anything.
18761 			 * This is only relevant in case of switching back to
18762 			 * the base stack.
18763 			 */
18764 			if (tp->t_rtttime == 0) {
18765 				tp->t_rtttime = ticks;
18766 				tp->t_rtseq = startseq;
18767 				KMOD_TCPSTAT_INC(tcps_segstimed);
18768 			}
18769 			if (len &&
18770 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
18771 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18772 		}
18773 		/*
18774 		 * If we are doing FO we need to update the mbuf position and subtract
18775 		 * this happens when the peer sends us duplicate information and
18776 		 * we thus want to send a DSACK.
18777 		 *
18778 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18779 		 * turned off? If not then we are going to echo multiple DSACK blocks
18780 		 * out (with the TSO), which we should not be doing.
18781 		 */
18782 		if (rack->r_fast_output && len) {
18783 			if (rack->r_ctl.fsb.left_to_send > len)
18784 				rack->r_ctl.fsb.left_to_send -= len;
18785 			else
18786 				rack->r_ctl.fsb.left_to_send = 0;
18787 			if (rack->r_ctl.fsb.left_to_send < segsiz)
18788 				rack->r_fast_output = 0;
18789 			if (rack->r_fast_output) {
18790 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18791 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18792 			}
18793 		}
18794 	}
18795 nomore:
18796 	if (error) {
18797 		rack->r_ctl.rc_agg_delayed = 0;
18798 		rack->r_early = 0;
18799 		rack->r_late = 0;
18800 		rack->r_ctl.rc_agg_early = 0;
18801 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
18802 		/*
18803 		 * Failures do not advance the seq counter above. For the
18804 		 * case of ENOBUFS we will fall out and retry in 1ms with
18805 		 * the hpts. Everything else will just have to retransmit
18806 		 * with the timer.
18807 		 *
18808 		 * In any case, we do not want to loop around for another
18809 		 * send without a good reason.
18810 		 */
18811 		sendalot = 0;
18812 		switch (error) {
18813 		case EPERM:
18814 			tp->t_softerror = error;
18815 #ifdef TCP_ACCOUNTING
18816 			crtsc = get_cyclecount();
18817 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18818 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18819 			}
18820 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18821 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18822 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18823 			}
18824 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18825 			sched_unpin();
18826 #endif
18827 			return (error);
18828 		case ENOBUFS:
18829 			/*
18830 			 * Pace us right away to retry in a some
18831 			 * time
18832 			 */
18833 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18834 			if (rack->rc_enobuf < 0x7f)
18835 				rack->rc_enobuf++;
18836 			if (slot < (10 * HPTS_USEC_IN_MSEC))
18837 				slot = 10 * HPTS_USEC_IN_MSEC;
18838 			if (rack->r_ctl.crte != NULL) {
18839 				counter_u64_add(rack_saw_enobuf_hw, 1);
18840 				tcp_rl_log_enobuf(rack->r_ctl.crte);
18841 			}
18842 			counter_u64_add(rack_saw_enobuf, 1);
18843 			goto enobufs;
18844 		case EMSGSIZE:
18845 			/*
18846 			 * For some reason the interface we used initially
18847 			 * to send segments changed to another or lowered
18848 			 * its MTU. If TSO was active we either got an
18849 			 * interface without TSO capabilits or TSO was
18850 			 * turned off. If we obtained mtu from ip_output()
18851 			 * then update it and try again.
18852 			 */
18853 			if (tso)
18854 				tp->t_flags &= ~TF_TSO;
18855 			if (mtu != 0) {
18856 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
18857 				goto again;
18858 			}
18859 			slot = 10 * HPTS_USEC_IN_MSEC;
18860 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18861 #ifdef TCP_ACCOUNTING
18862 			crtsc = get_cyclecount();
18863 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18864 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18865 			}
18866 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18867 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18868 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18869 			}
18870 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18871 			sched_unpin();
18872 #endif
18873 			return (error);
18874 		case ENETUNREACH:
18875 			counter_u64_add(rack_saw_enetunreach, 1);
18876 		case EHOSTDOWN:
18877 		case EHOSTUNREACH:
18878 		case ENETDOWN:
18879 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
18880 				tp->t_softerror = error;
18881 			}
18882 			/* FALLTHROUGH */
18883 		default:
18884 			slot = 10 * HPTS_USEC_IN_MSEC;
18885 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18886 #ifdef TCP_ACCOUNTING
18887 			crtsc = get_cyclecount();
18888 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18889 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18890 			}
18891 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18892 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18893 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18894 			}
18895 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18896 			sched_unpin();
18897 #endif
18898 			return (error);
18899 		}
18900 	} else {
18901 		rack->rc_enobuf = 0;
18902 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18903 			rack->r_ctl.retran_during_recovery += len;
18904 	}
18905 	KMOD_TCPSTAT_INC(tcps_sndtotal);
18906 
18907 	/*
18908 	 * Data sent (as far as we can tell). If this advertises a larger
18909 	 * window than any other segment, then remember the size of the
18910 	 * advertised window. Any pending ACK has now been sent.
18911 	 */
18912 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18913 		tp->rcv_adv = tp->rcv_nxt + recwin;
18914 
18915 	tp->last_ack_sent = tp->rcv_nxt;
18916 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18917 enobufs:
18918 	if (sendalot) {
18919 		/* Do we need to turn off sendalot? */
18920 		if (rack->r_ctl.rc_pace_max_segs &&
18921 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18922 			/* We hit our max. */
18923 			sendalot = 0;
18924 		} else if ((rack->rc_user_set_max_segs) &&
18925 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18926 			/* We hit the user defined max */
18927 			sendalot = 0;
18928 		}
18929 	}
18930 	if ((error == 0) && (flags & TH_FIN))
18931 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18932 	if (flags & TH_RST) {
18933 		/*
18934 		 * We don't send again after sending a RST.
18935 		 */
18936 		slot = 0;
18937 		sendalot = 0;
18938 		if (error == 0)
18939 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18940 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18941 		/*
18942 		 * Get our pacing rate, if an error
18943 		 * occurred in sending (ENOBUF) we would
18944 		 * hit the else if with slot preset. Other
18945 		 * errors return.
18946 		 */
18947 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18948 	}
18949 	if (rsm &&
18950 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18951 	    rack->use_rack_rr) {
18952 		/* Its a retransmit and we use the rack cheat? */
18953 		if ((slot == 0) ||
18954 		    (rack->rc_always_pace == 0) ||
18955 		    (rack->r_rr_config == 1)) {
18956 			/*
18957 			 * We have no pacing set or we
18958 			 * are using old-style rack or
18959 			 * we are overriden to use the old 1ms pacing.
18960 			 */
18961 			slot = rack->r_ctl.rc_min_to;
18962 		}
18963 	}
18964 	/* We have sent clear the flag */
18965 	rack->r_ent_rec_ns = 0;
18966 	if (rack->r_must_retran) {
18967 		if (rsm) {
18968 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
18969 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
18970 				/*
18971 				 * We have retransmitted all.
18972 				 */
18973 				rack->r_must_retran = 0;
18974 				rack->r_ctl.rc_out_at_rto = 0;
18975 			}
18976 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18977 			/*
18978 			 * Sending new data will also kill
18979 			 * the loop.
18980 			 */
18981 			rack->r_must_retran = 0;
18982 			rack->r_ctl.rc_out_at_rto = 0;
18983 		}
18984 	}
18985 	rack->r_ctl.fsb.recwin = recwin;
18986 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
18987 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18988 		/*
18989 		 * We hit an RTO and now have past snd_max at the RTO
18990 		 * clear all the WAS flags.
18991 		 */
18992 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
18993 	}
18994 	if (slot) {
18995 		/* set the rack tcb into the slot N */
18996 		if ((error == 0) &&
18997 		    rack_use_rfo &&
18998 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
18999 		    (rsm == NULL) &&
19000 		    (tp->snd_nxt == tp->snd_max) &&
19001 		    (ipoptlen == 0) &&
19002 		    (tp->rcv_numsacks == 0) &&
19003 		    rack->r_fsb_inited &&
19004 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19005 		    (rack->r_must_retran == 0) &&
19006 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19007 		    (len > 0) && (orig_len > 0) &&
19008 		    (orig_len > len) &&
19009 		    ((orig_len - len) >= segsiz) &&
19010 		    ((optlen == 0) ||
19011 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19012 			/* We can send at least one more MSS using our fsb */
19013 
19014 			rack->r_fast_output = 1;
19015 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19016 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19017 			rack->r_ctl.fsb.tcp_flags = flags;
19018 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19019 			if (hw_tls)
19020 				rack->r_ctl.fsb.hw_tls = 1;
19021 			else
19022 				rack->r_ctl.fsb.hw_tls = 0;
19023 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19024 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19025 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19026 				 (tp->snd_max - tp->snd_una)));
19027 			if (rack->r_ctl.fsb.left_to_send < segsiz)
19028 				rack->r_fast_output = 0;
19029 			else {
19030 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19031 					rack->r_ctl.fsb.rfo_apply_push = 1;
19032 				else
19033 					rack->r_ctl.fsb.rfo_apply_push = 0;
19034 			}
19035 		} else
19036 			rack->r_fast_output = 0;
19037 		rack_log_fsb(rack, tp, so, flags,
19038 			     ipoptlen, orig_len, len, error,
19039 			     (rsm == NULL), optlen, __LINE__, 2);
19040 	} else if (sendalot) {
19041 		int ret;
19042 
19043 		sack_rxmit = 0;
19044 		if ((error == 0) &&
19045 		    rack_use_rfo &&
19046 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19047 		    (rsm == NULL) &&
19048 		    (ipoptlen == 0) &&
19049 		    (tp->rcv_numsacks == 0) &&
19050 		    (tp->snd_nxt == tp->snd_max) &&
19051 		    (rack->r_must_retran == 0) &&
19052 		    rack->r_fsb_inited &&
19053 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19054 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19055 		    (len > 0) && (orig_len > 0) &&
19056 		    (orig_len > len) &&
19057 		    ((orig_len - len) >= segsiz) &&
19058 		    ((optlen == 0) ||
19059 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19060 			/* we can use fast_output for more */
19061 
19062 			rack->r_fast_output = 1;
19063 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19064 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19065 			rack->r_ctl.fsb.tcp_flags = flags;
19066 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19067 			if (hw_tls)
19068 				rack->r_ctl.fsb.hw_tls = 1;
19069 			else
19070 				rack->r_ctl.fsb.hw_tls = 0;
19071 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19072 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19073 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19074 				 (tp->snd_max - tp->snd_una)));
19075 			if (rack->r_ctl.fsb.left_to_send < segsiz) {
19076 				rack->r_fast_output = 0;
19077 			}
19078 			if (rack->r_fast_output) {
19079 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19080 					rack->r_ctl.fsb.rfo_apply_push = 1;
19081 				else
19082 					rack->r_ctl.fsb.rfo_apply_push = 0;
19083 				rack_log_fsb(rack, tp, so, flags,
19084 					     ipoptlen, orig_len, len, error,
19085 					     (rsm == NULL), optlen, __LINE__, 3);
19086 				error = 0;
19087 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19088 				if (ret >= 0)
19089 					return (ret);
19090 			        else if (error)
19091 					goto nomore;
19092 
19093 			}
19094 		}
19095 		goto again;
19096 	}
19097 	/* Assure when we leave that snd_nxt will point to top */
19098 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19099 		tp->snd_nxt = tp->snd_max;
19100 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19101 #ifdef TCP_ACCOUNTING
19102 	crtsc = get_cyclecount() - ts_val;
19103 	if (tot_len_this_send) {
19104 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19105 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
19106 		}
19107 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19108 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19109 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19110 		}
19111 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19112 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19113 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19114 		}
19115 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19116 	} else {
19117 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19118 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
19119 		}
19120 		counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19121 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19122 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19123 		}
19124 		counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19125 	}
19126 	sched_unpin();
19127 #endif
19128 	if (error == ENOBUFS)
19129 		error = 0;
19130 	return (error);
19131 }
19132 
19133 static void
19134 rack_update_seg(struct tcp_rack *rack)
19135 {
19136 	uint32_t orig_val;
19137 
19138 	orig_val = rack->r_ctl.rc_pace_max_segs;
19139 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19140 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
19141 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19142 }
19143 
19144 static void
19145 rack_mtu_change(struct tcpcb *tp)
19146 {
19147 	/*
19148 	 * The MSS may have changed
19149 	 */
19150 	struct tcp_rack *rack;
19151 	struct rack_sendmap *rsm;
19152 
19153 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19154 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19155 		/*
19156 		 * The MTU has changed we need to resend everything
19157 		 * since all we have sent is lost. We first fix
19158 		 * up the mtu though.
19159 		 */
19160 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19161 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
19162 		rack_remxt_tmr(tp);
19163 		rack->r_fast_output = 0;
19164 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19165 						rack->r_ctl.rc_sacked);
19166 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19167 		rack->r_must_retran = 1;
19168 		/* Mark all inflight to needing to be rxt'd */
19169 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19170 			rsm->r_flags |= RACK_MUST_RXT;
19171 		}
19172 	}
19173 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19174 	/* We don't use snd_nxt to retransmit */
19175 	tp->snd_nxt = tp->snd_max;
19176 }
19177 
19178 static int
19179 rack_set_profile(struct tcp_rack *rack, int prof)
19180 {
19181 	int err = EINVAL;
19182 	if (prof == 1) {
19183 		/* pace_always=1 */
19184 		if (rack->rc_always_pace == 0) {
19185 			if (tcp_can_enable_pacing() == 0)
19186 				return (EBUSY);
19187 		}
19188 		rack->rc_always_pace = 1;
19189 		if (rack->use_fixed_rate || rack->gp_ready)
19190 			rack_set_cc_pacing(rack);
19191 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19192 		rack->rack_attempt_hdwr_pace = 0;
19193 		/* cmpack=1 */
19194 		if (rack_use_cmp_acks)
19195 			rack->r_use_cmp_ack = 1;
19196 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19197 		    rack->r_use_cmp_ack)
19198 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19199 		/* scwnd=1 */
19200 		rack->rack_enable_scwnd = 1;
19201 		/* dynamic=100 */
19202 		rack->rc_gp_dyn_mul = 1;
19203 		/* gp_inc_ca */
19204 		rack->r_ctl.rack_per_of_gp_ca = 100;
19205 		/* rrr_conf=3 */
19206 		rack->r_rr_config = 3;
19207 		/* npush=2 */
19208 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19209 		/* fillcw=1 */
19210 		rack->rc_pace_to_cwnd = 1;
19211 		rack->rc_pace_fill_if_rttin_range = 0;
19212 		rack->rtt_limit_mul = 0;
19213 		/* noprr=1 */
19214 		rack->rack_no_prr = 1;
19215 		/* lscwnd=1 */
19216 		rack->r_limit_scw = 1;
19217 		/* gp_inc_rec */
19218 		rack->r_ctl.rack_per_of_gp_rec = 90;
19219 		err = 0;
19220 
19221 	} else if (prof == 3) {
19222 		/* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19223 		/* pace_always=1 */
19224 		if (rack->rc_always_pace == 0) {
19225 			if (tcp_can_enable_pacing() == 0)
19226 				return (EBUSY);
19227 		}
19228 		rack->rc_always_pace = 1;
19229 		if (rack->use_fixed_rate || rack->gp_ready)
19230 			rack_set_cc_pacing(rack);
19231 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19232 		rack->rack_attempt_hdwr_pace = 0;
19233 		/* cmpack=1 */
19234 		if (rack_use_cmp_acks)
19235 			rack->r_use_cmp_ack = 1;
19236 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19237 		    rack->r_use_cmp_ack)
19238 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19239 		/* scwnd=1 */
19240 		rack->rack_enable_scwnd = 1;
19241 		/* dynamic=100 */
19242 		rack->rc_gp_dyn_mul = 1;
19243 		/* gp_inc_ca */
19244 		rack->r_ctl.rack_per_of_gp_ca = 100;
19245 		/* rrr_conf=3 */
19246 		rack->r_rr_config = 3;
19247 		/* npush=2 */
19248 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19249 		/* fillcw=2 */
19250 		rack->rc_pace_to_cwnd = 1;
19251 		rack->r_fill_less_agg = 1;
19252 		rack->rc_pace_fill_if_rttin_range = 0;
19253 		rack->rtt_limit_mul = 0;
19254 		/* noprr=1 */
19255 		rack->rack_no_prr = 1;
19256 		/* lscwnd=1 */
19257 		rack->r_limit_scw = 1;
19258 		/* gp_inc_rec */
19259 		rack->r_ctl.rack_per_of_gp_rec = 90;
19260 		err = 0;
19261 
19262 
19263 	} else if (prof == 2) {
19264 		/* cmpack=1 */
19265 		if (rack->rc_always_pace == 0) {
19266 			if (tcp_can_enable_pacing() == 0)
19267 				return (EBUSY);
19268 		}
19269 		rack->rc_always_pace = 1;
19270 		if (rack->use_fixed_rate || rack->gp_ready)
19271 			rack_set_cc_pacing(rack);
19272 		rack->r_use_cmp_ack = 1;
19273 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19274 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19275 		/* pace_always=1 */
19276 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19277 		/* scwnd=1 */
19278 		rack->rack_enable_scwnd = 1;
19279 		/* dynamic=100 */
19280 		rack->rc_gp_dyn_mul = 1;
19281 		rack->r_ctl.rack_per_of_gp_ca = 100;
19282 		/* rrr_conf=3 */
19283 		rack->r_rr_config = 3;
19284 		/* npush=2 */
19285 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19286 		/* fillcw=1 */
19287 		rack->rc_pace_to_cwnd = 1;
19288 		rack->rc_pace_fill_if_rttin_range = 0;
19289 		rack->rtt_limit_mul = 0;
19290 		/* noprr=1 */
19291 		rack->rack_no_prr = 1;
19292 		/* lscwnd=0 */
19293 		rack->r_limit_scw = 0;
19294 		err = 0;
19295 	} else if (prof == 0) {
19296 		/* This changes things back to the default settings */
19297 		err = 0;
19298 		if (rack->rc_always_pace) {
19299 			tcp_decrement_paced_conn();
19300 			rack_undo_cc_pacing(rack);
19301 			rack->rc_always_pace = 0;
19302 		}
19303 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19304 			rack->rc_always_pace = 1;
19305 			if (rack->use_fixed_rate || rack->gp_ready)
19306 				rack_set_cc_pacing(rack);
19307 		} else
19308 			rack->rc_always_pace = 0;
19309 		if (rack_dsack_std_based & 0x1) {
19310 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19311 			rack->rc_rack_tmr_std_based = 1;
19312 		}
19313 		if (rack_dsack_std_based & 0x2) {
19314 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19315 			rack->rc_rack_use_dsack = 1;
19316 		}
19317 		if (rack_use_cmp_acks)
19318 			rack->r_use_cmp_ack = 1;
19319 		else
19320 			rack->r_use_cmp_ack = 0;
19321 		if (rack_disable_prr)
19322 			rack->rack_no_prr = 1;
19323 		else
19324 			rack->rack_no_prr = 0;
19325 		if (rack_gp_no_rec_chg)
19326 			rack->rc_gp_no_rec_chg = 1;
19327 		else
19328 			rack->rc_gp_no_rec_chg = 0;
19329 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19330 			rack->r_mbuf_queue = 1;
19331 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19332 				rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19333 			rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19334 		} else {
19335 			rack->r_mbuf_queue = 0;
19336 			rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19337 		}
19338 		if (rack_enable_shared_cwnd)
19339 			rack->rack_enable_scwnd = 1;
19340 		else
19341 			rack->rack_enable_scwnd = 0;
19342 		if (rack_do_dyn_mul) {
19343 			/* When dynamic adjustment is on CA needs to start at 100% */
19344 			rack->rc_gp_dyn_mul = 1;
19345 			if (rack_do_dyn_mul >= 100)
19346 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19347 		} else {
19348 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19349 			rack->rc_gp_dyn_mul = 0;
19350 		}
19351 		rack->r_rr_config = 0;
19352 		rack->r_ctl.rc_no_push_at_mrtt = 0;
19353 		rack->rc_pace_to_cwnd = 0;
19354 		rack->rc_pace_fill_if_rttin_range = 0;
19355 		rack->rtt_limit_mul = 0;
19356 
19357 		if (rack_enable_hw_pacing)
19358 			rack->rack_hdw_pace_ena = 1;
19359 		else
19360 			rack->rack_hdw_pace_ena = 0;
19361 		if (rack_disable_prr)
19362 			rack->rack_no_prr = 1;
19363 		else
19364 			rack->rack_no_prr = 0;
19365 		if (rack_limits_scwnd)
19366 			rack->r_limit_scw  = 1;
19367 		else
19368 			rack->r_limit_scw  = 0;
19369 		err = 0;
19370 	}
19371 	return (err);
19372 }
19373 
19374 static int
19375 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19376 {
19377 	struct deferred_opt_list *dol;
19378 
19379 	dol = malloc(sizeof(struct deferred_opt_list),
19380 		     M_TCPFSB, M_NOWAIT|M_ZERO);
19381 	if (dol == NULL) {
19382 		/*
19383 		 * No space yikes -- fail out..
19384 		 */
19385 		return (0);
19386 	}
19387 	dol->optname = sopt_name;
19388 	dol->optval = loptval;
19389 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19390 	return (1);
19391 }
19392 
19393 static int
19394 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19395 		    uint32_t optval, uint64_t loptval)
19396 {
19397 	struct epoch_tracker et;
19398 	struct sockopt sopt;
19399 	struct cc_newreno_opts opt;
19400 	uint64_t val;
19401 	int error = 0;
19402 	uint16_t ca, ss;
19403 
19404 	switch (sopt_name) {
19405 
19406 	case TCP_RACK_DSACK_OPT:
19407 		RACK_OPTS_INC(tcp_rack_dsack_opt);
19408 		if (optval & 0x1) {
19409 			rack->rc_rack_tmr_std_based = 1;
19410 		} else {
19411 			rack->rc_rack_tmr_std_based = 0;
19412 		}
19413 		if (optval & 0x2) {
19414 			rack->rc_rack_use_dsack = 1;
19415 		} else {
19416 			rack->rc_rack_use_dsack = 0;
19417 		}
19418 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19419 		break;
19420 	case TCP_RACK_PACING_BETA:
19421 		RACK_OPTS_INC(tcp_rack_beta);
19422 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19423 			/* This only works for newreno. */
19424 			error = EINVAL;
19425 			break;
19426 		}
19427 		if (rack->rc_pacing_cc_set) {
19428 			/*
19429 			 * Set them into the real CC module
19430 			 * whats in the rack pcb is the old values
19431 			 * to be used on restoral/
19432 			 */
19433 			sopt.sopt_dir = SOPT_SET;
19434 			opt.name = CC_NEWRENO_BETA;
19435 			opt.val = optval;
19436 			if (CC_ALGO(tp)->ctl_output != NULL)
19437 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19438 			else {
19439 				error = ENOENT;
19440 				break;
19441 			}
19442 		} else {
19443 			/*
19444 			 * Not pacing yet so set it into our local
19445 			 * rack pcb storage.
19446 			 */
19447 			rack->r_ctl.rc_saved_beta.beta = optval;
19448 		}
19449 		break;
19450 	case TCP_RACK_TIMER_SLOP:
19451 		RACK_OPTS_INC(tcp_rack_timer_slop);
19452 		rack->r_ctl.timer_slop = optval;
19453 		if (rack->rc_tp->t_srtt) {
19454 			/*
19455 			 * If we have an SRTT lets update t_rxtcur
19456 			 * to have the new slop.
19457 			 */
19458 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19459 					   rack_rto_min, rack_rto_max,
19460 					   rack->r_ctl.timer_slop);
19461 		}
19462 		break;
19463 	case TCP_RACK_PACING_BETA_ECN:
19464 		RACK_OPTS_INC(tcp_rack_beta_ecn);
19465 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19466 			/* This only works for newreno. */
19467 			error = EINVAL;
19468 			break;
19469 		}
19470 		if (rack->rc_pacing_cc_set) {
19471 			/*
19472 			 * Set them into the real CC module
19473 			 * whats in the rack pcb is the old values
19474 			 * to be used on restoral/
19475 			 */
19476 			sopt.sopt_dir = SOPT_SET;
19477 			opt.name = CC_NEWRENO_BETA_ECN;
19478 			opt.val = optval;
19479 			if (CC_ALGO(tp)->ctl_output != NULL)
19480 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19481 			else
19482 				error = ENOENT;
19483 		} else {
19484 			/*
19485 			 * Not pacing yet so set it into our local
19486 			 * rack pcb storage.
19487 			 */
19488 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19489 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19490 		}
19491 		break;
19492 	case TCP_DEFER_OPTIONS:
19493 		RACK_OPTS_INC(tcp_defer_opt);
19494 		if (optval) {
19495 			if (rack->gp_ready) {
19496 				/* Too late */
19497 				error = EINVAL;
19498 				break;
19499 			}
19500 			rack->defer_options = 1;
19501 		} else
19502 			rack->defer_options = 0;
19503 		break;
19504 	case TCP_RACK_MEASURE_CNT:
19505 		RACK_OPTS_INC(tcp_rack_measure_cnt);
19506 		if (optval && (optval <= 0xff)) {
19507 			rack->r_ctl.req_measurements = optval;
19508 		} else
19509 			error = EINVAL;
19510 		break;
19511 	case TCP_REC_ABC_VAL:
19512 		RACK_OPTS_INC(tcp_rec_abc_val);
19513 		if (optval > 0)
19514 			rack->r_use_labc_for_rec = 1;
19515 		else
19516 			rack->r_use_labc_for_rec = 0;
19517 		break;
19518 	case TCP_RACK_ABC_VAL:
19519 		RACK_OPTS_INC(tcp_rack_abc_val);
19520 		if ((optval > 0) && (optval < 255))
19521 			rack->rc_labc = optval;
19522 		else
19523 			error = EINVAL;
19524 		break;
19525 	case TCP_HDWR_UP_ONLY:
19526 		RACK_OPTS_INC(tcp_pacing_up_only);
19527 		if (optval)
19528 			rack->r_up_only = 1;
19529 		else
19530 			rack->r_up_only = 0;
19531 		break;
19532 	case TCP_PACING_RATE_CAP:
19533 		RACK_OPTS_INC(tcp_pacing_rate_cap);
19534 		rack->r_ctl.bw_rate_cap = loptval;
19535 		break;
19536 	case TCP_RACK_PROFILE:
19537 		RACK_OPTS_INC(tcp_profile);
19538 		error = rack_set_profile(rack, optval);
19539 		break;
19540 	case TCP_USE_CMP_ACKS:
19541 		RACK_OPTS_INC(tcp_use_cmp_acks);
19542 		if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19543 			/* You can't turn it off once its on! */
19544 			error = EINVAL;
19545 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19546 			rack->r_use_cmp_ack = 1;
19547 			rack->r_mbuf_queue = 1;
19548 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19549 		}
19550 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19551 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19552 		break;
19553 	case TCP_SHARED_CWND_TIME_LIMIT:
19554 		RACK_OPTS_INC(tcp_lscwnd);
19555 		if (optval)
19556 			rack->r_limit_scw = 1;
19557 		else
19558 			rack->r_limit_scw = 0;
19559 		break;
19560  	case TCP_RACK_PACE_TO_FILL:
19561 		RACK_OPTS_INC(tcp_fillcw);
19562 		if (optval == 0)
19563 			rack->rc_pace_to_cwnd = 0;
19564 		else {
19565 			rack->rc_pace_to_cwnd = 1;
19566 			if (optval > 1)
19567 				rack->r_fill_less_agg = 1;
19568 		}
19569 		if ((optval >= rack_gp_rtt_maxmul) &&
19570 		    rack_gp_rtt_maxmul &&
19571 		    (optval < 0xf)) {
19572 			rack->rc_pace_fill_if_rttin_range = 1;
19573 			rack->rtt_limit_mul = optval;
19574 		} else {
19575 			rack->rc_pace_fill_if_rttin_range = 0;
19576 			rack->rtt_limit_mul = 0;
19577 		}
19578 		break;
19579 	case TCP_RACK_NO_PUSH_AT_MAX:
19580 		RACK_OPTS_INC(tcp_npush);
19581 		if (optval == 0)
19582 			rack->r_ctl.rc_no_push_at_mrtt = 0;
19583 		else if (optval < 0xff)
19584 			rack->r_ctl.rc_no_push_at_mrtt = optval;
19585 		else
19586 			error = EINVAL;
19587 		break;
19588 	case TCP_SHARED_CWND_ENABLE:
19589 		RACK_OPTS_INC(tcp_rack_scwnd);
19590 		if (optval == 0)
19591 			rack->rack_enable_scwnd = 0;
19592 		else
19593 			rack->rack_enable_scwnd = 1;
19594 		break;
19595 	case TCP_RACK_MBUF_QUEUE:
19596 		/* Now do we use the LRO mbuf-queue feature */
19597 		RACK_OPTS_INC(tcp_rack_mbufq);
19598 		if (optval || rack->r_use_cmp_ack)
19599 			rack->r_mbuf_queue = 1;
19600 		else
19601 			rack->r_mbuf_queue = 0;
19602 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19603 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19604 		else
19605 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19606 		break;
19607 	case TCP_RACK_NONRXT_CFG_RATE:
19608 		RACK_OPTS_INC(tcp_rack_cfg_rate);
19609 		if (optval == 0)
19610 			rack->rack_rec_nonrxt_use_cr = 0;
19611 		else
19612 			rack->rack_rec_nonrxt_use_cr = 1;
19613 		break;
19614 	case TCP_NO_PRR:
19615 		RACK_OPTS_INC(tcp_rack_noprr);
19616 		if (optval == 0)
19617 			rack->rack_no_prr = 0;
19618 		else if (optval == 1)
19619 			rack->rack_no_prr = 1;
19620 		else if (optval == 2)
19621 			rack->no_prr_addback = 1;
19622 		else
19623 			error = EINVAL;
19624 		break;
19625 	case TCP_TIMELY_DYN_ADJ:
19626 		RACK_OPTS_INC(tcp_timely_dyn);
19627 		if (optval == 0)
19628 			rack->rc_gp_dyn_mul = 0;
19629 		else {
19630 			rack->rc_gp_dyn_mul = 1;
19631 			if (optval >= 100) {
19632 				/*
19633 				 * If the user sets something 100 or more
19634 				 * its the gp_ca value.
19635 				 */
19636 				rack->r_ctl.rack_per_of_gp_ca  = optval;
19637 			}
19638 		}
19639 		break;
19640 	case TCP_RACK_DO_DETECTION:
19641 		RACK_OPTS_INC(tcp_rack_do_detection);
19642 		if (optval == 0)
19643 			rack->do_detection = 0;
19644 		else
19645 			rack->do_detection = 1;
19646 		break;
19647 	case TCP_RACK_TLP_USE:
19648 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19649 			error = EINVAL;
19650 			break;
19651 		}
19652 		RACK_OPTS_INC(tcp_tlp_use);
19653 		rack->rack_tlp_threshold_use = optval;
19654 		break;
19655 	case TCP_RACK_TLP_REDUCE:
19656 		/* RACK TLP cwnd reduction (bool) */
19657 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
19658 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19659 		break;
19660 	/*  Pacing related ones */
19661 	case TCP_RACK_PACE_ALWAYS:
19662 		/*
19663 		 * zero is old rack method, 1 is new
19664 		 * method using a pacing rate.
19665 		 */
19666 		RACK_OPTS_INC(tcp_rack_pace_always);
19667 		if (optval > 0) {
19668 			if (rack->rc_always_pace) {
19669 				error = EALREADY;
19670 				break;
19671 			} else if (tcp_can_enable_pacing()) {
19672 				rack->rc_always_pace = 1;
19673 				if (rack->use_fixed_rate || rack->gp_ready)
19674 					rack_set_cc_pacing(rack);
19675 			}
19676 			else {
19677 				error = ENOSPC;
19678 				break;
19679 			}
19680 		} else {
19681 			if (rack->rc_always_pace) {
19682 				tcp_decrement_paced_conn();
19683 				rack->rc_always_pace = 0;
19684 				rack_undo_cc_pacing(rack);
19685 			}
19686 		}
19687 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19688 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19689 		else
19690 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19691 		/* A rate may be set irate or other, if so set seg size */
19692 		rack_update_seg(rack);
19693 		break;
19694 	case TCP_BBR_RACK_INIT_RATE:
19695 		RACK_OPTS_INC(tcp_initial_rate);
19696 		val = optval;
19697 		/* Change from kbits per second to bytes per second */
19698 		val *= 1000;
19699 		val /= 8;
19700 		rack->r_ctl.init_rate = val;
19701 		if (rack->rc_init_win != rack_default_init_window) {
19702 			uint32_t win, snt;
19703 
19704 			/*
19705 			 * Options don't always get applied
19706 			 * in the order you think. So in order
19707 			 * to assure we update a cwnd we need
19708 			 * to check and see if we are still
19709 			 * where we should raise the cwnd.
19710 			 */
19711 			win = rc_init_window(rack);
19712 			if (SEQ_GT(tp->snd_max, tp->iss))
19713 				snt = tp->snd_max - tp->iss;
19714 			else
19715 				snt = 0;
19716 			if ((snt < win) &&
19717 			    (tp->snd_cwnd < win))
19718 				tp->snd_cwnd = win;
19719 		}
19720 		if (rack->rc_always_pace)
19721 			rack_update_seg(rack);
19722 		break;
19723 	case TCP_BBR_IWINTSO:
19724 		RACK_OPTS_INC(tcp_initial_win);
19725 		if (optval && (optval <= 0xff)) {
19726 			uint32_t win, snt;
19727 
19728 			rack->rc_init_win = optval;
19729 			win = rc_init_window(rack);
19730 			if (SEQ_GT(tp->snd_max, tp->iss))
19731 				snt = tp->snd_max - tp->iss;
19732 			else
19733 				snt = 0;
19734 			if ((snt < win) &&
19735 			    (tp->t_srtt |
19736 #ifdef NETFLIX_PEAKRATE
19737 			     tp->t_maxpeakrate |
19738 #endif
19739 			     rack->r_ctl.init_rate)) {
19740 				/*
19741 				 * We are not past the initial window
19742 				 * and we have some bases for pacing,
19743 				 * so we need to possibly adjust up
19744 				 * the cwnd. Note even if we don't set
19745 				 * the cwnd, its still ok to raise the rc_init_win
19746 				 * which can be used coming out of idle when we
19747 				 * would have a rate.
19748 				 */
19749 				if (tp->snd_cwnd < win)
19750 					tp->snd_cwnd = win;
19751 			}
19752 			if (rack->rc_always_pace)
19753 				rack_update_seg(rack);
19754 		} else
19755 			error = EINVAL;
19756 		break;
19757 	case TCP_RACK_FORCE_MSEG:
19758 		RACK_OPTS_INC(tcp_rack_force_max_seg);
19759 		if (optval)
19760 			rack->rc_force_max_seg = 1;
19761 		else
19762 			rack->rc_force_max_seg = 0;
19763 		break;
19764 	case TCP_RACK_PACE_MAX_SEG:
19765 		/* Max segments size in a pace in bytes */
19766 		RACK_OPTS_INC(tcp_rack_max_seg);
19767 		rack->rc_user_set_max_segs = optval;
19768 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19769 		break;
19770 	case TCP_RACK_PACE_RATE_REC:
19771 		/* Set the fixed pacing rate in Bytes per second ca */
19772 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19773 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19774 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19775 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19776 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19777 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19778 		rack->use_fixed_rate = 1;
19779 		if (rack->rc_always_pace)
19780 			rack_set_cc_pacing(rack);
19781 		rack_log_pacing_delay_calc(rack,
19782 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19783 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19784 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19785 					   __LINE__, NULL,0);
19786 		break;
19787 
19788 	case TCP_RACK_PACE_RATE_SS:
19789 		/* Set the fixed pacing rate in Bytes per second ca */
19790 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19791 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19792 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19793 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19794 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19795 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19796 		rack->use_fixed_rate = 1;
19797 		if (rack->rc_always_pace)
19798 			rack_set_cc_pacing(rack);
19799 		rack_log_pacing_delay_calc(rack,
19800 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19801 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19802 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19803 					   __LINE__, NULL, 0);
19804 		break;
19805 
19806 	case TCP_RACK_PACE_RATE_CA:
19807 		/* Set the fixed pacing rate in Bytes per second ca */
19808 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19809 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19810 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19811 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19812 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19813 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19814 		rack->use_fixed_rate = 1;
19815 		if (rack->rc_always_pace)
19816 			rack_set_cc_pacing(rack);
19817 		rack_log_pacing_delay_calc(rack,
19818 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19819 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19820 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19821 					   __LINE__, NULL, 0);
19822 		break;
19823 	case TCP_RACK_GP_INCREASE_REC:
19824 		RACK_OPTS_INC(tcp_gp_inc_rec);
19825 		rack->r_ctl.rack_per_of_gp_rec = optval;
19826 		rack_log_pacing_delay_calc(rack,
19827 					   rack->r_ctl.rack_per_of_gp_ss,
19828 					   rack->r_ctl.rack_per_of_gp_ca,
19829 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19830 					   __LINE__, NULL, 0);
19831 		break;
19832 	case TCP_RACK_GP_INCREASE_CA:
19833 		RACK_OPTS_INC(tcp_gp_inc_ca);
19834 		ca = optval;
19835 		if (ca < 100) {
19836 			/*
19837 			 * We don't allow any reduction
19838 			 * over the GP b/w.
19839 			 */
19840 			error = EINVAL;
19841 			break;
19842 		}
19843 		rack->r_ctl.rack_per_of_gp_ca = ca;
19844 		rack_log_pacing_delay_calc(rack,
19845 					   rack->r_ctl.rack_per_of_gp_ss,
19846 					   rack->r_ctl.rack_per_of_gp_ca,
19847 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19848 					   __LINE__, NULL, 0);
19849 		break;
19850 	case TCP_RACK_GP_INCREASE_SS:
19851 		RACK_OPTS_INC(tcp_gp_inc_ss);
19852 		ss = optval;
19853 		if (ss < 100) {
19854 			/*
19855 			 * We don't allow any reduction
19856 			 * over the GP b/w.
19857 			 */
19858 			error = EINVAL;
19859 			break;
19860 		}
19861 		rack->r_ctl.rack_per_of_gp_ss = ss;
19862 		rack_log_pacing_delay_calc(rack,
19863 					   rack->r_ctl.rack_per_of_gp_ss,
19864 					   rack->r_ctl.rack_per_of_gp_ca,
19865 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19866 					   __LINE__, NULL, 0);
19867 		break;
19868 	case TCP_RACK_RR_CONF:
19869 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19870 		if (optval && optval <= 3)
19871 			rack->r_rr_config = optval;
19872 		else
19873 			rack->r_rr_config = 0;
19874 		break;
19875 	case TCP_HDWR_RATE_CAP:
19876 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
19877 		if (optval) {
19878 			if (rack->r_rack_hw_rate_caps == 0)
19879 				rack->r_rack_hw_rate_caps = 1;
19880 			else
19881 				error = EALREADY;
19882 		} else {
19883 			rack->r_rack_hw_rate_caps = 0;
19884 		}
19885 		break;
19886 	case TCP_BBR_HDWR_PACE:
19887 		RACK_OPTS_INC(tcp_hdwr_pacing);
19888 		if (optval){
19889 			if (rack->rack_hdrw_pacing == 0) {
19890 				rack->rack_hdw_pace_ena = 1;
19891 				rack->rack_attempt_hdwr_pace = 0;
19892 			} else
19893 				error = EALREADY;
19894 		} else {
19895 			rack->rack_hdw_pace_ena = 0;
19896 #ifdef RATELIMIT
19897 			if (rack->r_ctl.crte != NULL) {
19898 				rack->rack_hdrw_pacing = 0;
19899 				rack->rack_attempt_hdwr_pace = 0;
19900 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19901 				rack->r_ctl.crte = NULL;
19902 			}
19903 #endif
19904 		}
19905 		break;
19906 	/*  End Pacing related ones */
19907 	case TCP_RACK_PRR_SENDALOT:
19908 		/* Allow PRR to send more than one seg */
19909 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
19910 		rack->r_ctl.rc_prr_sendalot = optval;
19911 		break;
19912 	case TCP_RACK_MIN_TO:
19913 		/* Minimum time between rack t-o's in ms */
19914 		RACK_OPTS_INC(tcp_rack_min_to);
19915 		rack->r_ctl.rc_min_to = optval;
19916 		break;
19917 	case TCP_RACK_EARLY_SEG:
19918 		/* If early recovery max segments */
19919 		RACK_OPTS_INC(tcp_rack_early_seg);
19920 		rack->r_ctl.rc_early_recovery_segs = optval;
19921 		break;
19922 	case TCP_RACK_ENABLE_HYSTART:
19923 	{
19924 		if (optval) {
19925 			tp->ccv->flags |= CCF_HYSTART_ALLOWED;
19926 			if (rack_do_hystart > RACK_HYSTART_ON)
19927 				tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND;
19928 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
19929 				tp->ccv->flags |= CCF_HYSTART_CONS_SSTH;
19930 		} else {
19931 			tp->ccv->flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
19932 		}
19933 	}
19934 	break;
19935 	case TCP_RACK_REORD_THRESH:
19936 		/* RACK reorder threshold (shift amount) */
19937 		RACK_OPTS_INC(tcp_rack_reord_thresh);
19938 		if ((optval > 0) && (optval < 31))
19939 			rack->r_ctl.rc_reorder_shift = optval;
19940 		else
19941 			error = EINVAL;
19942 		break;
19943 	case TCP_RACK_REORD_FADE:
19944 		/* Does reordering fade after ms time */
19945 		RACK_OPTS_INC(tcp_rack_reord_fade);
19946 		rack->r_ctl.rc_reorder_fade = optval;
19947 		break;
19948 	case TCP_RACK_TLP_THRESH:
19949 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
19950 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
19951 		if (optval)
19952 			rack->r_ctl.rc_tlp_threshold = optval;
19953 		else
19954 			error = EINVAL;
19955 		break;
19956 	case TCP_BBR_USE_RACK_RR:
19957 		RACK_OPTS_INC(tcp_rack_rr);
19958 		if (optval)
19959 			rack->use_rack_rr = 1;
19960 		else
19961 			rack->use_rack_rr = 0;
19962 		break;
19963 	case TCP_FAST_RSM_HACK:
19964 		RACK_OPTS_INC(tcp_rack_fastrsm_hack);
19965 		if (optval)
19966 			rack->fast_rsm_hack = 1;
19967 		else
19968 			rack->fast_rsm_hack = 0;
19969 		break;
19970 	case TCP_RACK_PKT_DELAY:
19971 		/* RACK added ms i.e. rack-rtt + reord + N */
19972 		RACK_OPTS_INC(tcp_rack_pkt_delay);
19973 		rack->r_ctl.rc_pkt_delay = optval;
19974 		break;
19975 	case TCP_DELACK:
19976 		RACK_OPTS_INC(tcp_rack_delayed_ack);
19977 		if (optval == 0)
19978 			tp->t_delayed_ack = 0;
19979 		else
19980 			tp->t_delayed_ack = 1;
19981 		if (tp->t_flags & TF_DELACK) {
19982 			tp->t_flags &= ~TF_DELACK;
19983 			tp->t_flags |= TF_ACKNOW;
19984 			NET_EPOCH_ENTER(et);
19985 			rack_output(tp);
19986 			NET_EPOCH_EXIT(et);
19987 		}
19988 		break;
19989 
19990 	case TCP_BBR_RACK_RTT_USE:
19991 		RACK_OPTS_INC(tcp_rack_rtt_use);
19992 		if ((optval != USE_RTT_HIGH) &&
19993 		    (optval != USE_RTT_LOW) &&
19994 		    (optval != USE_RTT_AVG))
19995 			error = EINVAL;
19996 		else
19997 			rack->r_ctl.rc_rate_sample_method = optval;
19998 		break;
19999 	case TCP_DATA_AFTER_CLOSE:
20000 		RACK_OPTS_INC(tcp_data_after_close);
20001 		if (optval)
20002 			rack->rc_allow_data_af_clo = 1;
20003 		else
20004 			rack->rc_allow_data_af_clo = 0;
20005 		break;
20006 	default:
20007 		break;
20008 	}
20009 #ifdef NETFLIX_STATS
20010 	tcp_log_socket_option(tp, sopt_name, optval, error);
20011 #endif
20012 	return (error);
20013 }
20014 
20015 
20016 static void
20017 rack_apply_deferred_options(struct tcp_rack *rack)
20018 {
20019 	struct deferred_opt_list *dol, *sdol;
20020 	uint32_t s_optval;
20021 
20022 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20023 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20024 		/* Disadvantage of deferal is you loose the error return */
20025 		s_optval = (uint32_t)dol->optval;
20026 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20027 		free(dol, M_TCPDO);
20028 	}
20029 }
20030 
20031 static void
20032 rack_hw_tls_change(struct tcpcb *tp, int chg)
20033 {
20034 	/*
20035 	 * HW tls state has changed.. fix all
20036 	 * rsm's in flight.
20037 	 */
20038 	struct tcp_rack *rack;
20039 	struct rack_sendmap *rsm;
20040 
20041 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20042 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
20043 		if (chg)
20044 			rsm->r_hw_tls = 1;
20045 		else
20046 			rsm->r_hw_tls = 0;
20047 	}
20048 	if (chg)
20049 		rack->r_ctl.fsb.hw_tls = 1;
20050 	else
20051 		rack->r_ctl.fsb.hw_tls = 0;
20052 }
20053 
20054 static int
20055 rack_pru_options(struct tcpcb *tp, int flags)
20056 {
20057 	if (flags & PRUS_OOB)
20058 		return (EOPNOTSUPP);
20059 	return (0);
20060 }
20061 
20062 static struct tcp_function_block __tcp_rack = {
20063 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
20064 	.tfb_tcp_output = rack_output,
20065 	.tfb_do_queued_segments = ctf_do_queued_segments,
20066 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
20067 	.tfb_tcp_do_segment = rack_do_segment,
20068 	.tfb_tcp_ctloutput = rack_ctloutput,
20069 	.tfb_tcp_fb_init = rack_init,
20070 	.tfb_tcp_fb_fini = rack_fini,
20071 	.tfb_tcp_timer_stop_all = rack_stopall,
20072 	.tfb_tcp_timer_activate = rack_timer_activate,
20073 	.tfb_tcp_timer_active = rack_timer_active,
20074 	.tfb_tcp_timer_stop = rack_timer_stop,
20075 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20076 	.tfb_tcp_handoff_ok = rack_handoff_ok,
20077 	.tfb_tcp_mtu_chg = rack_mtu_change,
20078 	.tfb_pru_options = rack_pru_options,
20079 	.tfb_hwtls_change = rack_hw_tls_change,
20080 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20081 };
20082 
20083 /*
20084  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20085  * socket option arguments.  When it re-acquires the lock after the copy, it
20086  * has to revalidate that the connection is still valid for the socket
20087  * option.
20088  */
20089 static int
20090 rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
20091 {
20092 #ifdef INET6
20093 	struct ip6_hdr *ip6;
20094 #endif
20095 #ifdef INET
20096 	struct ip *ip;
20097 #endif
20098 	struct tcpcb *tp;
20099 	struct tcp_rack *rack;
20100 	uint64_t loptval;
20101 	int32_t error = 0, optval;
20102 
20103 	tp = intotcpcb(inp);
20104 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20105 	if (rack == NULL) {
20106 		INP_WUNLOCK(inp);
20107 		return (EINVAL);
20108 	}
20109 #ifdef INET6
20110 	ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20111 #endif
20112 #ifdef INET
20113 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20114 #endif
20115 
20116 	switch (sopt->sopt_level) {
20117 #ifdef INET6
20118 	case IPPROTO_IPV6:
20119 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
20120 		switch (sopt->sopt_name) {
20121 		case IPV6_USE_MIN_MTU:
20122 			tcp6_use_min_mtu(tp);
20123 			break;
20124 		case IPV6_TCLASS:
20125 			/*
20126 			 * The DSCP codepoint has changed, update the fsb.
20127 			 */
20128 			ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20129 			    (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20130 			break;
20131 		}
20132 		INP_WUNLOCK(inp);
20133 		return (0);
20134 #endif
20135 #ifdef INET
20136 	case IPPROTO_IP:
20137 		switch (sopt->sopt_name) {
20138 		case IP_TOS:
20139 			/*
20140 			 * The DSCP codepoint has changed, update the fsb.
20141 			 */
20142 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
20143 			break;
20144 		case IP_TTL:
20145 			/*
20146 			 * The TTL has changed, update the fsb.
20147 			 */
20148 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20149 			break;
20150 		}
20151 		INP_WUNLOCK(inp);
20152 		return (0);
20153 #endif
20154 	}
20155 
20156 	switch (sopt->sopt_name) {
20157 	case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
20158 	/*  Pacing related ones */
20159 	case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
20160 	case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
20161 	case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
20162 	case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
20163 	case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
20164 	case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
20165 	case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
20166 	case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
20167 	case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
20168 	case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
20169 	case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
20170 	case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
20171 	case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
20172 	case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
20173 	case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
20174 	case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
20175        /* End pacing related */
20176 	case TCP_FAST_RSM_HACK:			/*  URL:frsm_hack */
20177 	case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20178 	case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
20179 	case TCP_RACK_MIN_TO:			/*  URL:min_to */
20180 	case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
20181 	case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
20182 	case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
20183 	case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
20184 	case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
20185 	case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
20186 	case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
20187 	case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
20188 	case TCP_RACK_DO_DETECTION:		/*  URL:detect */
20189 	case TCP_NO_PRR:			/*  URL:noprr */
20190 	case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
20191 	case TCP_DATA_AFTER_CLOSE:		/*  no URL */
20192 	case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
20193 	case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
20194 	case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
20195 	case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
20196 	case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
20197 	case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
20198 	case TCP_RACK_PROFILE:			/*  URL:profile */
20199 	case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
20200 	case TCP_RACK_ABC_VAL:			/*  URL:labc */
20201 	case TCP_REC_ABC_VAL:			/*  URL:reclabc */
20202 	case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
20203 	case TCP_DEFER_OPTIONS:			/*  URL:defer */
20204 	case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
20205 	case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
20206 	case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
20207 	case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
20208 	case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
20209 		break;
20210 	default:
20211 		/* Filter off all unknown options to the base stack */
20212 		return (tcp_default_ctloutput(inp, sopt));
20213 		break;
20214 	}
20215 	INP_WUNLOCK(inp);
20216 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20217 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20218 		/*
20219 		 * We truncate it down to 32 bits for the socket-option trace this
20220 		 * means rates > 34Gbps won't show right, but thats probably ok.
20221 		 */
20222 		optval = (uint32_t)loptval;
20223 	} else {
20224 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20225 		/* Save it in 64 bit form too */
20226 		loptval = optval;
20227 	}
20228 	if (error)
20229 		return (error);
20230 	INP_WLOCK(inp);
20231 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
20232 		INP_WUNLOCK(inp);
20233 		return (ECONNRESET);
20234 	}
20235 	if (tp->t_fb != &__tcp_rack) {
20236 		INP_WUNLOCK(inp);
20237 		return (ENOPROTOOPT);
20238 	}
20239 	if (rack->defer_options && (rack->gp_ready == 0) &&
20240 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20241 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20242 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20243 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20244 		/* Options are beind deferred */
20245 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20246 			INP_WUNLOCK(inp);
20247 			return (0);
20248 		} else {
20249 			/* No memory to defer, fail */
20250 			INP_WUNLOCK(inp);
20251 			return (ENOMEM);
20252 		}
20253 	}
20254 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20255 	INP_WUNLOCK(inp);
20256 	return (error);
20257 }
20258 
20259 static void
20260 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20261 {
20262 
20263 	INP_WLOCK_ASSERT(tp->t_inpcb);
20264 	bzero(ti, sizeof(*ti));
20265 
20266 	ti->tcpi_state = tp->t_state;
20267 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20268 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20269 	if (tp->t_flags & TF_SACK_PERMIT)
20270 		ti->tcpi_options |= TCPI_OPT_SACK;
20271 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20272 		ti->tcpi_options |= TCPI_OPT_WSCALE;
20273 		ti->tcpi_snd_wscale = tp->snd_scale;
20274 		ti->tcpi_rcv_wscale = tp->rcv_scale;
20275 	}
20276 	if (tp->t_flags2 & TF2_ECN_PERMIT)
20277 		ti->tcpi_options |= TCPI_OPT_ECN;
20278 	if (tp->t_flags & TF_FASTOPEN)
20279 		ti->tcpi_options |= TCPI_OPT_TFO;
20280 	/* still kept in ticks is t_rcvtime */
20281 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20282 	/* Since we hold everything in precise useconds this is easy */
20283 	ti->tcpi_rtt = tp->t_srtt;
20284 	ti->tcpi_rttvar = tp->t_rttvar;
20285 	ti->tcpi_rto = tp->t_rxtcur;
20286 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20287 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
20288 	/*
20289 	 * FreeBSD-specific extension fields for tcp_info.
20290 	 */
20291 	ti->tcpi_rcv_space = tp->rcv_wnd;
20292 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
20293 	ti->tcpi_snd_wnd = tp->snd_wnd;
20294 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
20295 	ti->tcpi_snd_nxt = tp->snd_nxt;
20296 	ti->tcpi_snd_mss = tp->t_maxseg;
20297 	ti->tcpi_rcv_mss = tp->t_maxseg;
20298 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20299 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20300 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20301 #ifdef NETFLIX_STATS
20302 	ti->tcpi_total_tlp = tp->t_sndtlppack;
20303 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20304 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20305 #endif
20306 #ifdef TCP_OFFLOAD
20307 	if (tp->t_flags & TF_TOE) {
20308 		ti->tcpi_options |= TCPI_OPT_TOE;
20309 		tcp_offload_tcp_info(tp, ti);
20310 	}
20311 #endif
20312 }
20313 
20314 static int
20315 rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
20316 {
20317 	struct tcpcb *tp;
20318 	struct tcp_rack *rack;
20319 	int32_t error, optval;
20320 	uint64_t val, loptval;
20321 	struct	tcp_info ti;
20322 	/*
20323 	 * Because all our options are either boolean or an int, we can just
20324 	 * pull everything into optval and then unlock and copy. If we ever
20325 	 * add a option that is not a int, then this will have quite an
20326 	 * impact to this routine.
20327 	 */
20328 	error = 0;
20329 	tp = intotcpcb(inp);
20330 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20331 	if (rack == NULL) {
20332 		INP_WUNLOCK(inp);
20333 		return (EINVAL);
20334 	}
20335 	switch (sopt->sopt_name) {
20336 	case TCP_INFO:
20337 		/* First get the info filled */
20338 		rack_fill_info(tp, &ti);
20339 		/* Fix up the rtt related fields if needed */
20340 		INP_WUNLOCK(inp);
20341 		error = sooptcopyout(sopt, &ti, sizeof ti);
20342 		return (error);
20343 	/*
20344 	 * Beta is the congestion control value for NewReno that influences how
20345 	 * much of a backoff happens when loss is detected. It is normally set
20346 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20347 	 * when you exit recovery.
20348 	 */
20349 	case TCP_RACK_PACING_BETA:
20350 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20351 			error = EINVAL;
20352 		else if (rack->rc_pacing_cc_set == 0)
20353 			optval = rack->r_ctl.rc_saved_beta.beta;
20354 		else {
20355 			/*
20356 			 * Reach out into the CC data and report back what
20357 			 * I have previously set. Yeah it looks hackish but
20358 			 * we don't want to report the saved values.
20359 			 */
20360 			if (tp->ccv->cc_data)
20361 				optval = ((struct newreno *)tp->ccv->cc_data)->beta;
20362 			else
20363 				error = EINVAL;
20364 		}
20365 		break;
20366 		/*
20367 		 * Beta_ecn is the congestion control value for NewReno that influences how
20368 		 * much of a backoff happens when a ECN mark is detected. It is normally set
20369 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20370 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
20371 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
20372 		 */
20373 
20374 	case TCP_RACK_PACING_BETA_ECN:
20375 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20376 			error = EINVAL;
20377 		else if (rack->rc_pacing_cc_set == 0)
20378 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20379 		else {
20380 			/*
20381 			 * Reach out into the CC data and report back what
20382 			 * I have previously set. Yeah it looks hackish but
20383 			 * we don't want to report the saved values.
20384 			 */
20385 			if (tp->ccv->cc_data)
20386 				optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
20387 			else
20388 				error = EINVAL;
20389 		}
20390 		break;
20391 	case TCP_RACK_DSACK_OPT:
20392 		optval = 0;
20393 		if (rack->rc_rack_tmr_std_based) {
20394 			optval |= 1;
20395 		}
20396 		if (rack->rc_rack_use_dsack) {
20397 			optval |= 2;
20398 		}
20399 		break;
20400  	case TCP_RACK_ENABLE_HYSTART:
20401 	{
20402 		if (tp->ccv->flags & CCF_HYSTART_ALLOWED) {
20403 			optval = RACK_HYSTART_ON;
20404 			if (tp->ccv->flags & CCF_HYSTART_CAN_SH_CWND)
20405 				optval = RACK_HYSTART_ON_W_SC;
20406 			if (tp->ccv->flags & CCF_HYSTART_CONS_SSTH)
20407 				optval = RACK_HYSTART_ON_W_SC_C;
20408 		} else {
20409 			optval = RACK_HYSTART_OFF;
20410 		}
20411 	}
20412 	break;
20413 	case TCP_FAST_RSM_HACK:
20414 		optval = rack->fast_rsm_hack;
20415 		break;
20416 	case TCP_DEFER_OPTIONS:
20417 		optval = rack->defer_options;
20418 		break;
20419 	case TCP_RACK_MEASURE_CNT:
20420 		optval = rack->r_ctl.req_measurements;
20421 		break;
20422 	case TCP_REC_ABC_VAL:
20423 		optval = rack->r_use_labc_for_rec;
20424 		break;
20425 	case TCP_RACK_ABC_VAL:
20426 		optval = rack->rc_labc;
20427 		break;
20428 	case TCP_HDWR_UP_ONLY:
20429 		optval= rack->r_up_only;
20430 		break;
20431 	case TCP_PACING_RATE_CAP:
20432 		loptval = rack->r_ctl.bw_rate_cap;
20433 		break;
20434 	case TCP_RACK_PROFILE:
20435 		/* You cannot retrieve a profile, its write only */
20436 		error = EINVAL;
20437 		break;
20438 	case TCP_USE_CMP_ACKS:
20439 		optval = rack->r_use_cmp_ack;
20440 		break;
20441 	case TCP_RACK_PACE_TO_FILL:
20442 		optval = rack->rc_pace_to_cwnd;
20443 		if (optval && rack->r_fill_less_agg)
20444 			optval++;
20445 		break;
20446 	case TCP_RACK_NO_PUSH_AT_MAX:
20447 		optval = rack->r_ctl.rc_no_push_at_mrtt;
20448 		break;
20449 	case TCP_SHARED_CWND_ENABLE:
20450 		optval = rack->rack_enable_scwnd;
20451 		break;
20452 	case TCP_RACK_NONRXT_CFG_RATE:
20453 		optval = rack->rack_rec_nonrxt_use_cr;
20454 		break;
20455 	case TCP_NO_PRR:
20456 		if (rack->rack_no_prr  == 1)
20457 			optval = 1;
20458 		else if (rack->no_prr_addback == 1)
20459 			optval = 2;
20460 		else
20461 			optval = 0;
20462 		break;
20463 	case TCP_RACK_DO_DETECTION:
20464 		optval = rack->do_detection;
20465 		break;
20466 	case TCP_RACK_MBUF_QUEUE:
20467 		/* Now do we use the LRO mbuf-queue feature */
20468 		optval = rack->r_mbuf_queue;
20469 		break;
20470 	case TCP_TIMELY_DYN_ADJ:
20471 		optval = rack->rc_gp_dyn_mul;
20472 		break;
20473 	case TCP_BBR_IWINTSO:
20474 		optval = rack->rc_init_win;
20475 		break;
20476 	case TCP_RACK_TLP_REDUCE:
20477 		/* RACK TLP cwnd reduction (bool) */
20478 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20479 		break;
20480 	case TCP_BBR_RACK_INIT_RATE:
20481 		val = rack->r_ctl.init_rate;
20482 		/* convert to kbits per sec */
20483 		val *= 8;
20484 		val /= 1000;
20485 		optval = (uint32_t)val;
20486 		break;
20487 	case TCP_RACK_FORCE_MSEG:
20488 		optval = rack->rc_force_max_seg;
20489 		break;
20490 	case TCP_RACK_PACE_MAX_SEG:
20491 		/* Max segments in a pace */
20492 		optval = rack->rc_user_set_max_segs;
20493 		break;
20494 	case TCP_RACK_PACE_ALWAYS:
20495 		/* Use the always pace method */
20496 		optval = rack->rc_always_pace;
20497 		break;
20498 	case TCP_RACK_PRR_SENDALOT:
20499 		/* Allow PRR to send more than one seg */
20500 		optval = rack->r_ctl.rc_prr_sendalot;
20501 		break;
20502 	case TCP_RACK_MIN_TO:
20503 		/* Minimum time between rack t-o's in ms */
20504 		optval = rack->r_ctl.rc_min_to;
20505 		break;
20506 	case TCP_RACK_EARLY_SEG:
20507 		/* If early recovery max segments */
20508 		optval = rack->r_ctl.rc_early_recovery_segs;
20509 		break;
20510 	case TCP_RACK_REORD_THRESH:
20511 		/* RACK reorder threshold (shift amount) */
20512 		optval = rack->r_ctl.rc_reorder_shift;
20513 		break;
20514 	case TCP_RACK_REORD_FADE:
20515 		/* Does reordering fade after ms time */
20516 		optval = rack->r_ctl.rc_reorder_fade;
20517 		break;
20518 	case TCP_BBR_USE_RACK_RR:
20519 		/* Do we use the rack cheat for rxt */
20520 		optval = rack->use_rack_rr;
20521 		break;
20522 	case TCP_RACK_RR_CONF:
20523 		optval = rack->r_rr_config;
20524 		break;
20525 	case TCP_HDWR_RATE_CAP:
20526 		optval = rack->r_rack_hw_rate_caps;
20527 		break;
20528 	case TCP_BBR_HDWR_PACE:
20529 		optval = rack->rack_hdw_pace_ena;
20530 		break;
20531 	case TCP_RACK_TLP_THRESH:
20532 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20533 		optval = rack->r_ctl.rc_tlp_threshold;
20534 		break;
20535 	case TCP_RACK_PKT_DELAY:
20536 		/* RACK added ms i.e. rack-rtt + reord + N */
20537 		optval = rack->r_ctl.rc_pkt_delay;
20538 		break;
20539 	case TCP_RACK_TLP_USE:
20540 		optval = rack->rack_tlp_threshold_use;
20541 		break;
20542 	case TCP_RACK_PACE_RATE_CA:
20543 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20544 		break;
20545 	case TCP_RACK_PACE_RATE_SS:
20546 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20547 		break;
20548 	case TCP_RACK_PACE_RATE_REC:
20549 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20550 		break;
20551 	case TCP_RACK_GP_INCREASE_SS:
20552 		optval = rack->r_ctl.rack_per_of_gp_ca;
20553 		break;
20554 	case TCP_RACK_GP_INCREASE_CA:
20555 		optval = rack->r_ctl.rack_per_of_gp_ss;
20556 		break;
20557 	case TCP_BBR_RACK_RTT_USE:
20558 		optval = rack->r_ctl.rc_rate_sample_method;
20559 		break;
20560 	case TCP_DELACK:
20561 		optval = tp->t_delayed_ack;
20562 		break;
20563 	case TCP_DATA_AFTER_CLOSE:
20564 		optval = rack->rc_allow_data_af_clo;
20565 		break;
20566 	case TCP_SHARED_CWND_TIME_LIMIT:
20567 		optval = rack->r_limit_scw;
20568 		break;
20569 	case TCP_RACK_TIMER_SLOP:
20570 		optval = rack->r_ctl.timer_slop;
20571 		break;
20572 	default:
20573 		return (tcp_default_ctloutput(inp, sopt));
20574 		break;
20575 	}
20576 	INP_WUNLOCK(inp);
20577 	if (error == 0) {
20578 		if (TCP_PACING_RATE_CAP)
20579 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
20580 		else
20581 			error = sooptcopyout(sopt, &optval, sizeof optval);
20582 	}
20583 	return (error);
20584 }
20585 
20586 static int
20587 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt)
20588 {
20589 	if (sopt->sopt_dir == SOPT_SET) {
20590 		return (rack_set_sockopt(inp, sopt));
20591 	} else if (sopt->sopt_dir == SOPT_GET) {
20592 		return (rack_get_sockopt(inp, sopt));
20593 	} else {
20594 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
20595 	}
20596 }
20597 
20598 static const char *rack_stack_names[] = {
20599 	__XSTRING(STACKNAME),
20600 #ifdef STACKALIAS
20601 	__XSTRING(STACKALIAS),
20602 #endif
20603 };
20604 
20605 static int
20606 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20607 {
20608 	memset(mem, 0, size);
20609 	return (0);
20610 }
20611 
20612 static void
20613 rack_dtor(void *mem, int32_t size, void *arg)
20614 {
20615 
20616 }
20617 
20618 static bool rack_mod_inited = false;
20619 
20620 static int
20621 tcp_addrack(module_t mod, int32_t type, void *data)
20622 {
20623 	int32_t err = 0;
20624 	int num_stacks;
20625 
20626 	switch (type) {
20627 	case MOD_LOAD:
20628 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20629 		    sizeof(struct rack_sendmap),
20630 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20631 
20632 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20633 		    sizeof(struct tcp_rack),
20634 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20635 
20636 		sysctl_ctx_init(&rack_sysctl_ctx);
20637 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20638 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20639 		    OID_AUTO,
20640 #ifdef STACKALIAS
20641 		    __XSTRING(STACKALIAS),
20642 #else
20643 		    __XSTRING(STACKNAME),
20644 #endif
20645 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20646 		    "");
20647 		if (rack_sysctl_root == NULL) {
20648 			printf("Failed to add sysctl node\n");
20649 			err = EFAULT;
20650 			goto free_uma;
20651 		}
20652 		rack_init_sysctls();
20653 		num_stacks = nitems(rack_stack_names);
20654 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20655 		    rack_stack_names, &num_stacks);
20656 		if (err) {
20657 			printf("Failed to register %s stack name for "
20658 			    "%s module\n", rack_stack_names[num_stacks],
20659 			    __XSTRING(MODNAME));
20660 			sysctl_ctx_free(&rack_sysctl_ctx);
20661 free_uma:
20662 			uma_zdestroy(rack_zone);
20663 			uma_zdestroy(rack_pcb_zone);
20664 			rack_counter_destroy();
20665 			printf("Failed to register rack module -- err:%d\n", err);
20666 			return (err);
20667 		}
20668 		tcp_lro_reg_mbufq();
20669 		rack_mod_inited = true;
20670 		break;
20671 	case MOD_QUIESCE:
20672 		err = deregister_tcp_functions(&__tcp_rack, true, false);
20673 		break;
20674 	case MOD_UNLOAD:
20675 		err = deregister_tcp_functions(&__tcp_rack, false, true);
20676 		if (err == EBUSY)
20677 			break;
20678 		if (rack_mod_inited) {
20679 			uma_zdestroy(rack_zone);
20680 			uma_zdestroy(rack_pcb_zone);
20681 			sysctl_ctx_free(&rack_sysctl_ctx);
20682 			rack_counter_destroy();
20683 			rack_mod_inited = false;
20684 		}
20685 		tcp_lro_dereg_mbufq();
20686 		err = 0;
20687 		break;
20688 	default:
20689 		return (EOPNOTSUPP);
20690 	}
20691 	return (err);
20692 }
20693 
20694 static moduledata_t tcp_rack = {
20695 	.name = __XSTRING(MODNAME),
20696 	.evhand = tcp_addrack,
20697 	.priv = 0
20698 };
20699 
20700 MODULE_VERSION(MODNAME, 1);
20701 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20702 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
20703