xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 56cc89585d392d8c8f120b2e4b8faa7df236b4ea) 
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, int );
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 		return;
566 	}
567 	rack->rc_pacing_cc_set = 1;
568 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
569 		/* Not new-reno we can't play games with beta! */
570 		goto out;
571 	}
572 	ptr = ((struct newreno *)tp->ccv->cc_data);
573 	if (CC_ALGO(tp)->ctl_output == NULL)  {
574 		/* Huh, why does new_reno no longer have a set function? */
575 		goto out;
576 	}
577 	if (ptr == NULL) {
578 		/* Just the default values */
579 		old.beta = V_newreno_beta_ecn;
580 		old.beta_ecn = V_newreno_beta_ecn;
581 		old.newreno_flags = 0;
582 	} else {
583 		old.beta = ptr->beta;
584 		old.beta_ecn = ptr->beta_ecn;
585 		old.newreno_flags = ptr->newreno_flags;
586 	}
587 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
588 	sopt.sopt_dir = SOPT_SET;
589 	opt.name = CC_NEWRENO_BETA;
590 	opt.val = rack->r_ctl.rc_saved_beta.beta;
591 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
592 	if (error)  {
593 		goto out;
594 	}
595 	/*
596 	 * Hack alert we need to set in our newreno_flags
597 	 * so that Abe behavior is also applied.
598 	 */
599 	((struct newreno *)tp->ccv->cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
600 	opt.name = CC_NEWRENO_BETA_ECN;
601 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
602 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
603 	if (error) {
604 		goto out;
605 	}
606 	/* Save off the original values for restoral */
607 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
608 out:
609 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
610 		union tcp_log_stackspecific log;
611 		struct timeval tv;
612 
613 		ptr = ((struct newreno *)tp->ccv->cc_data);
614 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
615 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
616 		if (ptr) {
617 			log.u_bbr.flex1 = ptr->beta;
618 			log.u_bbr.flex2 = ptr->beta_ecn;
619 			log.u_bbr.flex3 = ptr->newreno_flags;
620 		}
621 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
622 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
623 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
624 		log.u_bbr.flex7 = rack->gp_ready;
625 		log.u_bbr.flex7 <<= 1;
626 		log.u_bbr.flex7 |= rack->use_fixed_rate;
627 		log.u_bbr.flex7 <<= 1;
628 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
629 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
630 		log.u_bbr.flex8 = 3;
631 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
632 			       0, &log, false, NULL, NULL, 0, &tv);
633 	}
634 }
635 
636 static void
637 rack_undo_cc_pacing(struct tcp_rack *rack)
638 {
639 	struct newreno old, *ptr;
640 	struct tcpcb *tp;
641 
642 	if (rack->rc_pacing_cc_set == 0)
643 		return;
644 	tp = rack->rc_tp;
645 	rack->rc_pacing_cc_set = 0;
646 	if (tp->cc_algo == NULL)
647 		/* Tcb is leaving */
648 		return;
649 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
650 		/* Not new-reno nothing to do! */
651 		return;
652 	}
653 	ptr = ((struct newreno *)tp->ccv->cc_data);
654 	if (ptr == NULL) {
655 		/*
656 		 * This happens at rack_fini() if the
657 		 * cc module gets freed on us. In that
658 		 * case we loose our "new" settings but
659 		 * thats ok, since the tcb is going away anyway.
660 		 */
661 		return;
662 	}
663 	/* Grab out our set values */
664 	memcpy(&old, ptr, sizeof(struct newreno));
665 	/* Copy back in the original values */
666 	memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
667 	/* Now save back the values we had set in (for when pacing is restored) */
668 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
669 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
670 		union tcp_log_stackspecific log;
671 		struct timeval tv;
672 
673 		ptr = ((struct newreno *)tp->ccv->cc_data);
674 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
675 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
676 		log.u_bbr.flex1 = ptr->beta;
677 		log.u_bbr.flex2 = ptr->beta_ecn;
678 		log.u_bbr.flex3 = ptr->newreno_flags;
679 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
680 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
681 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
682 		log.u_bbr.flex7 = rack->gp_ready;
683 		log.u_bbr.flex7 <<= 1;
684 		log.u_bbr.flex7 |= rack->use_fixed_rate;
685 		log.u_bbr.flex7 <<= 1;
686 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
687 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
688 		log.u_bbr.flex8 = 4;
689 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
690 			       0, &log, false, NULL, NULL, 0, &tv);
691 	}
692 }
693 
694 #ifdef NETFLIX_PEAKRATE
695 static inline void
696 rack_update_peakrate_thr(struct tcpcb *tp)
697 {
698 	/* Keep in mind that t_maxpeakrate is in B/s. */
699 	uint64_t peak;
700 	peak = uqmax((tp->t_maxseg * 2),
701 		     (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
702 	tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
703 }
704 #endif
705 
706 static int
707 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
708 {
709 	uint32_t stat;
710 	int32_t error;
711 
712 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
713 	if (error || req->newptr == NULL)
714 		return error;
715 
716 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
717 	if (error)
718 		return (error);
719 	if (stat == 1) {
720 #ifdef INVARIANTS
721 		printf("Clearing RACK counters\n");
722 #endif
723 		counter_u64_zero(rack_tlp_tot);
724 		counter_u64_zero(rack_tlp_newdata);
725 		counter_u64_zero(rack_tlp_retran);
726 		counter_u64_zero(rack_tlp_retran_bytes);
727 		counter_u64_zero(rack_to_tot);
728 		counter_u64_zero(rack_saw_enobuf);
729 		counter_u64_zero(rack_saw_enobuf_hw);
730 		counter_u64_zero(rack_saw_enetunreach);
731 		counter_u64_zero(rack_persists_sends);
732 		counter_u64_zero(rack_persists_acks);
733 		counter_u64_zero(rack_persists_loss);
734 		counter_u64_zero(rack_persists_lost_ends);
735 #ifdef INVARIANTS
736 		counter_u64_zero(rack_adjust_map_bw);
737 #endif
738 		counter_u64_zero(rack_to_alloc_hard);
739 		counter_u64_zero(rack_to_alloc_emerg);
740 		counter_u64_zero(rack_sack_proc_all);
741 		counter_u64_zero(rack_fto_send);
742 		counter_u64_zero(rack_fto_rsm_send);
743 		counter_u64_zero(rack_extended_rfo);
744 		counter_u64_zero(rack_hw_pace_init_fail);
745 		counter_u64_zero(rack_hw_pace_lost);
746 		counter_u64_zero(rack_non_fto_send);
747 		counter_u64_zero(rack_nfto_resend);
748 		counter_u64_zero(rack_sack_proc_short);
749 		counter_u64_zero(rack_sack_proc_restart);
750 		counter_u64_zero(rack_to_alloc);
751 		counter_u64_zero(rack_to_alloc_limited);
752 		counter_u64_zero(rack_alloc_limited_conns);
753 		counter_u64_zero(rack_split_limited);
754 		counter_u64_zero(rack_multi_single_eq);
755 		counter_u64_zero(rack_proc_non_comp_ack);
756 		counter_u64_zero(rack_sack_attacks_detected);
757 		counter_u64_zero(rack_sack_attacks_reversed);
758 		counter_u64_zero(rack_sack_used_next_merge);
759 		counter_u64_zero(rack_sack_used_prev_merge);
760 		counter_u64_zero(rack_sack_splits);
761 		counter_u64_zero(rack_sack_skipped_acked);
762 		counter_u64_zero(rack_ack_total);
763 		counter_u64_zero(rack_express_sack);
764 		counter_u64_zero(rack_sack_total);
765 		counter_u64_zero(rack_move_none);
766 		counter_u64_zero(rack_move_some);
767 		counter_u64_zero(rack_try_scwnd);
768 		counter_u64_zero(rack_collapsed_win);
769 	}
770 	rack_clear_counter = 0;
771 	return (0);
772 }
773 
774 static void
775 rack_init_sysctls(void)
776 {
777 	struct sysctl_oid *rack_counters;
778 	struct sysctl_oid *rack_attack;
779 	struct sysctl_oid *rack_pacing;
780 	struct sysctl_oid *rack_timely;
781 	struct sysctl_oid *rack_timers;
782 	struct sysctl_oid *rack_tlp;
783 	struct sysctl_oid *rack_misc;
784 	struct sysctl_oid *rack_features;
785 	struct sysctl_oid *rack_measure;
786 	struct sysctl_oid *rack_probertt;
787 	struct sysctl_oid *rack_hw_pacing;
788 
789 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
790 	    SYSCTL_CHILDREN(rack_sysctl_root),
791 	    OID_AUTO,
792 	    "sack_attack",
793 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
794 	    "Rack Sack Attack Counters and Controls");
795 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
796 	    SYSCTL_CHILDREN(rack_sysctl_root),
797 	    OID_AUTO,
798 	    "stats",
799 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
800 	    "Rack Counters");
801 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
802 	    SYSCTL_CHILDREN(rack_sysctl_root),
803 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
804 	    &rack_rate_sample_method , USE_RTT_LOW,
805 	    "What method should we use for rate sampling 0=high, 1=low ");
806 	/* Probe rtt related controls */
807 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
808 	    SYSCTL_CHILDREN(rack_sysctl_root),
809 	    OID_AUTO,
810 	    "probertt",
811 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
812 	    "ProbeRTT related Controls");
813 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
814 	    SYSCTL_CHILDREN(rack_probertt),
815 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
816 	    &rack_atexit_prtt_hbp, 130,
817 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
818 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
819 	    SYSCTL_CHILDREN(rack_probertt),
820 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
821 	    &rack_atexit_prtt, 130,
822 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
823 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
824 	    SYSCTL_CHILDREN(rack_probertt),
825 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
826 	    &rack_per_of_gp_probertt, 60,
827 	    "What percentage of goodput do we pace at in probertt");
828 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
829 	    SYSCTL_CHILDREN(rack_probertt),
830 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
831 	    &rack_per_of_gp_probertt_reduce, 10,
832 	    "What percentage of goodput do we reduce every gp_srtt");
833 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
834 	    SYSCTL_CHILDREN(rack_probertt),
835 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
836 	    &rack_per_of_gp_lowthresh, 40,
837 	    "What percentage of goodput do we allow the multiplier to fall to");
838 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
839 	    SYSCTL_CHILDREN(rack_probertt),
840 	    OID_AUTO, "time_between", CTLFLAG_RW,
841 	    & rack_time_between_probertt, 96000000,
842 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
843 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
844 	    SYSCTL_CHILDREN(rack_probertt),
845 	    OID_AUTO, "safety", CTLFLAG_RW,
846 	    &rack_probe_rtt_safety_val, 2000000,
847 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
848 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
849 	    SYSCTL_CHILDREN(rack_probertt),
850 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
851 	    &rack_probe_rtt_sets_cwnd, 0,
852 	    "Do we set the cwnd too (if always_lower is on)");
853 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
854 	    SYSCTL_CHILDREN(rack_probertt),
855 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
856 	    &rack_max_drain_wait, 2,
857 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
858 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
859 	    SYSCTL_CHILDREN(rack_probertt),
860 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
861 	    &rack_must_drain, 1,
862 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
863 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
864 	    SYSCTL_CHILDREN(rack_probertt),
865 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
866 	    &rack_probertt_use_min_rtt_entry, 1,
867 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
868 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
869 	    SYSCTL_CHILDREN(rack_probertt),
870 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
871 	    &rack_probertt_use_min_rtt_exit, 0,
872 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
873 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
874 	    SYSCTL_CHILDREN(rack_probertt),
875 	    OID_AUTO, "length_div", CTLFLAG_RW,
876 	    &rack_probertt_gpsrtt_cnt_div, 0,
877 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
878 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
879 	    SYSCTL_CHILDREN(rack_probertt),
880 	    OID_AUTO, "length_mul", CTLFLAG_RW,
881 	    &rack_probertt_gpsrtt_cnt_mul, 0,
882 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
883 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
884 	    SYSCTL_CHILDREN(rack_probertt),
885 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
886 	    &rack_min_probertt_hold, 200000,
887 	    "What is the minimum time we hold probertt at target");
888 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
889 	    SYSCTL_CHILDREN(rack_probertt),
890 	    OID_AUTO, "filter_life", CTLFLAG_RW,
891 	    &rack_probertt_filter_life, 10000000,
892 	    "What is the time for the filters life in useconds");
893 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
894 	    SYSCTL_CHILDREN(rack_probertt),
895 	    OID_AUTO, "lower_within", CTLFLAG_RW,
896 	    &rack_probertt_lower_within, 10,
897 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
898 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
899 	    SYSCTL_CHILDREN(rack_probertt),
900 	    OID_AUTO, "must_move", CTLFLAG_RW,
901 	    &rack_min_rtt_movement, 250,
902 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
903 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
904 	    SYSCTL_CHILDREN(rack_probertt),
905 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
906 	    &rack_probertt_clear_is, 1,
907 	    "Do we clear I/S counts on exiting probe-rtt");
908 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
909 	    SYSCTL_CHILDREN(rack_probertt),
910 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
911 	    &rack_max_drain_hbp, 1,
912 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
913 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
914 	    SYSCTL_CHILDREN(rack_probertt),
915 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
916 	    &rack_hbp_thresh, 3,
917 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
918 	/* Pacing related sysctls */
919 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
920 	    SYSCTL_CHILDREN(rack_sysctl_root),
921 	    OID_AUTO,
922 	    "pacing",
923 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
924 	    "Pacing related Controls");
925 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
926 	    SYSCTL_CHILDREN(rack_pacing),
927 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
928 	    &rack_max_per_above, 30,
929 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
930 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
931 	    SYSCTL_CHILDREN(rack_pacing),
932 	    OID_AUTO, "pace_to_one", CTLFLAG_RW,
933 	    &rack_pace_one_seg, 0,
934 	    "Do we allow low b/w pacing of 1MSS instead of two");
935 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
936 	    SYSCTL_CHILDREN(rack_pacing),
937 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
938 	    &rack_limit_time_with_srtt, 0,
939 	    "Do we limit pacing time based on srtt");
940 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
941 	    SYSCTL_CHILDREN(rack_pacing),
942 	    OID_AUTO, "init_win", CTLFLAG_RW,
943 	    &rack_default_init_window, 0,
944 	    "Do we have a rack initial window 0 = system default");
945 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
946 	    SYSCTL_CHILDREN(rack_pacing),
947 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
948 	    &rack_per_of_gp_ss, 250,
949 	    "If non zero, what percentage of goodput to pace at in slow start");
950 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
951 	    SYSCTL_CHILDREN(rack_pacing),
952 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
953 	    &rack_per_of_gp_ca, 150,
954 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
955 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
956 	    SYSCTL_CHILDREN(rack_pacing),
957 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
958 	    &rack_per_of_gp_rec, 200,
959 	    "If non zero, what percentage of goodput to pace at in recovery");
960 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
961 	    SYSCTL_CHILDREN(rack_pacing),
962 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
963 	    &rack_hptsi_segments, 40,
964 	    "What size is the max for TSO segments in pacing and burst mitigation");
965 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
966 	    SYSCTL_CHILDREN(rack_pacing),
967 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
968 	    &rack_slot_reduction, 4,
969 	    "When doing only burst mitigation what is the reduce divisor");
970 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
971 	    SYSCTL_CHILDREN(rack_sysctl_root),
972 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
973 	    &rack_pace_every_seg, 0,
974 	    "If set we use pacing, if clear we use only the original burst mitigation");
975 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
976 	    SYSCTL_CHILDREN(rack_pacing),
977 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
978 	    &rack_bw_rate_cap, 0,
979 	    "If set we apply this value to the absolute rate cap used by pacing");
980 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
981 	    SYSCTL_CHILDREN(rack_sysctl_root),
982 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
983 	    &rack_req_measurements, 1,
984 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
985 	/* Hardware pacing */
986 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
987 	    SYSCTL_CHILDREN(rack_sysctl_root),
988 	    OID_AUTO,
989 	    "hdwr_pacing",
990 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
991 	    "Pacing related Controls");
992 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
993 	    SYSCTL_CHILDREN(rack_hw_pacing),
994 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
995 	    &rack_hw_rwnd_factor, 2,
996 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
997 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
998 	    SYSCTL_CHILDREN(rack_hw_pacing),
999 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1000 	    &rack_enobuf_hw_boost_mult, 2,
1001 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1002 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1003 	    SYSCTL_CHILDREN(rack_hw_pacing),
1004 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1005 	    &rack_enobuf_hw_max, 2,
1006 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1007 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1008 	    SYSCTL_CHILDREN(rack_hw_pacing),
1009 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1010 	    &rack_enobuf_hw_min, 2,
1011 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1012 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1013 	    SYSCTL_CHILDREN(rack_hw_pacing),
1014 	    OID_AUTO, "enable", CTLFLAG_RW,
1015 	    &rack_enable_hw_pacing, 0,
1016 	    "Should RACK attempt to use hw pacing?");
1017 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1018 	    SYSCTL_CHILDREN(rack_hw_pacing),
1019 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1020 	    &rack_hw_rate_caps, 1,
1021 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1022 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1023 	    SYSCTL_CHILDREN(rack_hw_pacing),
1024 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1025 	    &rack_hw_rate_min, 0,
1026 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1027 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1028 	    SYSCTL_CHILDREN(rack_hw_pacing),
1029 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1030 	    &rack_hw_rate_to_low, 0,
1031 	    "If we fall below this rate, dis-engage hw pacing?");
1032 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1033 	    SYSCTL_CHILDREN(rack_hw_pacing),
1034 	    OID_AUTO, "up_only", CTLFLAG_RW,
1035 	    &rack_hw_up_only, 1,
1036 	    "Do we allow hw pacing to lower the rate selected?");
1037 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1038 	    SYSCTL_CHILDREN(rack_hw_pacing),
1039 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1040 	    &rack_hw_pace_extra_slots, 2,
1041 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1042 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1043 	    SYSCTL_CHILDREN(rack_sysctl_root),
1044 	    OID_AUTO,
1045 	    "timely",
1046 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1047 	    "Rack Timely RTT Controls");
1048 	/* Timely based GP dynmics */
1049 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1050 	    SYSCTL_CHILDREN(rack_timely),
1051 	    OID_AUTO, "upper", CTLFLAG_RW,
1052 	    &rack_gp_per_bw_mul_up, 2,
1053 	    "Rack timely upper range for equal b/w (in percentage)");
1054 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1055 	    SYSCTL_CHILDREN(rack_timely),
1056 	    OID_AUTO, "lower", CTLFLAG_RW,
1057 	    &rack_gp_per_bw_mul_down, 4,
1058 	    "Rack timely lower range for equal b/w (in percentage)");
1059 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1060 	    SYSCTL_CHILDREN(rack_timely),
1061 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1062 	    &rack_gp_rtt_maxmul, 3,
1063 	    "Rack timely multipler of lowest rtt for rtt_max");
1064 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1065 	    SYSCTL_CHILDREN(rack_timely),
1066 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1067 	    &rack_gp_rtt_mindiv, 4,
1068 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1069 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1070 	    SYSCTL_CHILDREN(rack_timely),
1071 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1072 	    &rack_gp_rtt_minmul, 1,
1073 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1074 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1075 	    SYSCTL_CHILDREN(rack_timely),
1076 	    OID_AUTO, "decrease", CTLFLAG_RW,
1077 	    &rack_gp_decrease_per, 20,
1078 	    "Rack timely decrease percentage of our GP multiplication factor");
1079 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1080 	    SYSCTL_CHILDREN(rack_timely),
1081 	    OID_AUTO, "increase", CTLFLAG_RW,
1082 	    &rack_gp_increase_per, 2,
1083 	    "Rack timely increase perentage of our GP multiplication factor");
1084 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1085 	    SYSCTL_CHILDREN(rack_timely),
1086 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1087 	    &rack_per_lower_bound, 50,
1088 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1089 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1090 	    SYSCTL_CHILDREN(rack_timely),
1091 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1092 	    &rack_per_upper_bound_ss, 0,
1093 	    "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1094 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1095 	    SYSCTL_CHILDREN(rack_timely),
1096 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1097 	    &rack_per_upper_bound_ca, 0,
1098 	    "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1099 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1100 	    SYSCTL_CHILDREN(rack_timely),
1101 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1102 	    &rack_do_dyn_mul, 0,
1103 	    "Rack timely do we enable dynmaic timely goodput by default");
1104 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1105 	    SYSCTL_CHILDREN(rack_timely),
1106 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1107 	    &rack_gp_no_rec_chg, 1,
1108 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1109 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1110 	    SYSCTL_CHILDREN(rack_timely),
1111 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1112 	    &rack_timely_dec_clear, 6,
1113 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1114 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1115 	    SYSCTL_CHILDREN(rack_timely),
1116 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1117 	    &rack_timely_max_push_rise, 3,
1118 	    "Rack timely how many times do we push up with b/w increase");
1119 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1120 	    SYSCTL_CHILDREN(rack_timely),
1121 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1122 	    &rack_timely_max_push_drop, 3,
1123 	    "Rack timely how many times do we push back on b/w decent");
1124 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1125 	    SYSCTL_CHILDREN(rack_timely),
1126 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1127 	    &rack_timely_min_segs, 4,
1128 	    "Rack timely when setting the cwnd what is the min num segments");
1129 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1130 	    SYSCTL_CHILDREN(rack_timely),
1131 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1132 	    &rack_use_max_for_nobackoff, 0,
1133 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1134 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1135 	    SYSCTL_CHILDREN(rack_timely),
1136 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1137 	    &rack_timely_int_timely_only, 0,
1138 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1139 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1140 	    SYSCTL_CHILDREN(rack_timely),
1141 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1142 	    &rack_timely_no_stopping, 0,
1143 	    "Rack timely don't stop increase");
1144 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1145 	    SYSCTL_CHILDREN(rack_timely),
1146 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1147 	    &rack_down_raise_thresh, 100,
1148 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1149 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1150 	    SYSCTL_CHILDREN(rack_timely),
1151 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1152 	    &rack_req_segs, 1,
1153 	    "Bottom dragging if not these many segments outstanding and room");
1154 
1155 	/* TLP and Rack related parameters */
1156 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1157 	    SYSCTL_CHILDREN(rack_sysctl_root),
1158 	    OID_AUTO,
1159 	    "tlp",
1160 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1161 	    "TLP and Rack related Controls");
1162 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1163 	    SYSCTL_CHILDREN(rack_tlp),
1164 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1165 	    &use_rack_rr, 1,
1166 	    "Do we use Rack Rapid Recovery");
1167 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1168 	    SYSCTL_CHILDREN(rack_tlp),
1169 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1170 	    &rack_max_abc_post_recovery, 2,
1171 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1172 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1173 	    SYSCTL_CHILDREN(rack_tlp),
1174 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1175 	    &rack_non_rxt_use_cr, 0,
1176 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1177 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1178 	    SYSCTL_CHILDREN(rack_tlp),
1179 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1180 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1181 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1182 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1183 	    SYSCTL_CHILDREN(rack_tlp),
1184 	    OID_AUTO, "limit", CTLFLAG_RW,
1185 	    &rack_tlp_limit, 2,
1186 	    "How many TLP's can be sent without sending new data");
1187 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1188 	    SYSCTL_CHILDREN(rack_tlp),
1189 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1190 	    &rack_tlp_use_greater, 1,
1191 	    "Should we use the rack_rtt time if its greater than srtt");
1192 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1193 	    SYSCTL_CHILDREN(rack_tlp),
1194 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1195 	    &rack_tlp_min, 10000,
1196 	    "TLP minimum timeout per the specification (in microseconds)");
1197 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1198 	    SYSCTL_CHILDREN(rack_tlp),
1199 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1200 	    &rack_always_send_oldest, 0,
1201 	    "Should we always send the oldest TLP and RACK-TLP");
1202 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1203 	    SYSCTL_CHILDREN(rack_tlp),
1204 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1205 	    &rack_limited_retran, 0,
1206 	    "How many times can a rack timeout drive out sends");
1207 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1208 	    SYSCTL_CHILDREN(rack_tlp),
1209 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1210 	    &rack_lower_cwnd_at_tlp, 0,
1211 	    "When a TLP completes a retran should we enter recovery");
1212 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1213 	    SYSCTL_CHILDREN(rack_tlp),
1214 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1215 	    &rack_reorder_thresh, 2,
1216 	    "What factor for rack will be added when seeing reordering (shift right)");
1217 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1218 	    SYSCTL_CHILDREN(rack_tlp),
1219 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1220 	    &rack_tlp_thresh, 1,
1221 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1222 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1223 	    SYSCTL_CHILDREN(rack_tlp),
1224 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1225 	    &rack_reorder_fade, 60000000,
1226 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1227 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1228 	    SYSCTL_CHILDREN(rack_tlp),
1229 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1230 	    &rack_pkt_delay, 1000,
1231 	    "Extra RACK time (in microseconds) besides reordering thresh");
1232 
1233 	/* Timer related controls */
1234 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1235 	    SYSCTL_CHILDREN(rack_sysctl_root),
1236 	    OID_AUTO,
1237 	    "timers",
1238 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1239 	    "Timer related controls");
1240 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1241 	    SYSCTL_CHILDREN(rack_timers),
1242 	    OID_AUTO, "persmin", CTLFLAG_RW,
1243 	    &rack_persist_min, 250000,
1244 	    "What is the minimum time in microseconds between persists");
1245 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1246 	    SYSCTL_CHILDREN(rack_timers),
1247 	    OID_AUTO, "persmax", CTLFLAG_RW,
1248 	    &rack_persist_max, 2000000,
1249 	    "What is the largest delay in microseconds between persists");
1250 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1251 	    SYSCTL_CHILDREN(rack_timers),
1252 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1253 	    &rack_delayed_ack_time, 40000,
1254 	    "Delayed ack time (40ms in microseconds)");
1255 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1256 	    SYSCTL_CHILDREN(rack_timers),
1257 	    OID_AUTO, "minrto", CTLFLAG_RW,
1258 	    &rack_rto_min, 30000,
1259 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1260 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1261 	    SYSCTL_CHILDREN(rack_timers),
1262 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1263 	    &rack_rto_max, 4000000,
1264 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1265 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1266 	    SYSCTL_CHILDREN(rack_timers),
1267 	    OID_AUTO, "minto", CTLFLAG_RW,
1268 	    &rack_min_to, 1000,
1269 	    "Minimum rack timeout in microseconds");
1270 	/* Measure controls */
1271 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1272 	    SYSCTL_CHILDREN(rack_sysctl_root),
1273 	    OID_AUTO,
1274 	    "measure",
1275 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1276 	    "Measure related controls");
1277 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1278 	    SYSCTL_CHILDREN(rack_measure),
1279 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1280 	    &rack_wma_divisor, 8,
1281 	    "When doing b/w calculation what is the  divisor for the WMA");
1282 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1283 	    SYSCTL_CHILDREN(rack_measure),
1284 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1285 	    &rack_cwnd_block_ends_measure, 0,
1286 	    "Does a cwnd just-return end the measurement window (app limited)");
1287 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1288 	    SYSCTL_CHILDREN(rack_measure),
1289 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1290 	    &rack_rwnd_block_ends_measure, 0,
1291 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1292 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1293 	    SYSCTL_CHILDREN(rack_measure),
1294 	    OID_AUTO, "min_target", CTLFLAG_RW,
1295 	    &rack_def_data_window, 20,
1296 	    "What is the minimum target window (in mss) for a GP measurements");
1297 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1298 	    SYSCTL_CHILDREN(rack_measure),
1299 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1300 	    &rack_goal_bdp, 2,
1301 	    "What is the goal BDP to measure");
1302 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1303 	    SYSCTL_CHILDREN(rack_measure),
1304 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1305 	    &rack_min_srtts, 1,
1306 	    "What is the goal BDP to measure");
1307 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1308 	    SYSCTL_CHILDREN(rack_measure),
1309 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1310 	    &rack_min_measure_usec, 0,
1311 	    "What is the Minimum time time for a measurement if 0, this is off");
1312 	/* Features */
1313 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1314 	    SYSCTL_CHILDREN(rack_sysctl_root),
1315 	    OID_AUTO,
1316 	    "features",
1317 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1318 	    "Feature controls");
1319 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1320 	    SYSCTL_CHILDREN(rack_features),
1321 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1322 	    &rack_use_cmp_acks, 1,
1323 	    "Should RACK have LRO send compressed acks");
1324 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1325 	    SYSCTL_CHILDREN(rack_features),
1326 	    OID_AUTO, "fsb", CTLFLAG_RW,
1327 	    &rack_use_fsb, 1,
1328 	    "Should RACK use the fast send block?");
1329 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1330 	    SYSCTL_CHILDREN(rack_features),
1331 	    OID_AUTO, "rfo", CTLFLAG_RW,
1332 	    &rack_use_rfo, 1,
1333 	    "Should RACK use rack_fast_output()?");
1334 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1335 	    SYSCTL_CHILDREN(rack_features),
1336 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1337 	    &rack_use_rsm_rfo, 1,
1338 	    "Should RACK use rack_fast_rsm_output()?");
1339 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1340 	    SYSCTL_CHILDREN(rack_features),
1341 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1342 	    &rack_enable_mqueue_for_nonpaced, 0,
1343 	    "Should RACK use mbuf queuing for non-paced connections");
1344 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1345 	    SYSCTL_CHILDREN(rack_features),
1346 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1347 	    &rack_do_hystart, 0,
1348 	    "Should RACK enable HyStart++ on connections?");
1349 	/* Misc rack controls */
1350 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1351 	    SYSCTL_CHILDREN(rack_sysctl_root),
1352 	    OID_AUTO,
1353 	    "misc",
1354 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1355 	    "Misc related controls");
1356 #ifdef TCP_ACCOUNTING
1357 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1358 	    SYSCTL_CHILDREN(rack_misc),
1359 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1360 	    &rack_tcp_accounting, 0,
1361 	    "Should we turn on TCP accounting for all rack sessions?");
1362 #endif
1363 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1364 	    SYSCTL_CHILDREN(rack_misc),
1365 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1366 	    &rack_apply_rtt_with_reduced_conf, 0,
1367 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1368 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1369 	    SYSCTL_CHILDREN(rack_misc),
1370 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1371 	    &rack_dsack_std_based, 3,
1372 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1373 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1374 	    SYSCTL_CHILDREN(rack_misc),
1375 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1376 	    &rack_prr_addbackmax, 2,
1377 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1378 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1379 	    SYSCTL_CHILDREN(rack_misc),
1380 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1381 	    &rack_stats_gets_ms_rtt, 1,
1382 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1383 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1384 	    SYSCTL_CHILDREN(rack_misc),
1385 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1386 	    &rack_client_low_buf, 0,
1387 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1388 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1389 	    SYSCTL_CHILDREN(rack_misc),
1390 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1391 	    &rack_def_profile, 0,
1392 	    "Should RACK use a default profile (0=no, num == profile num)?");
1393 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1394 	    SYSCTL_CHILDREN(rack_misc),
1395 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1396 	    &rack_enable_shared_cwnd, 1,
1397 	    "Should RACK try to use the shared cwnd on connections where allowed");
1398 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1399 	    SYSCTL_CHILDREN(rack_misc),
1400 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1401 	    &rack_limits_scwnd, 1,
1402 	    "Should RACK place low end time limits on the shared cwnd feature");
1403 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1404 	    SYSCTL_CHILDREN(rack_misc),
1405 	    OID_AUTO, "iMac_dack", CTLFLAG_RW,
1406 	    &rack_use_imac_dack, 0,
1407 	    "Should RACK try to emulate iMac delayed ack");
1408 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1409 	    SYSCTL_CHILDREN(rack_misc),
1410 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1411 	    &rack_disable_prr, 0,
1412 	    "Should RACK not use prr and only pace (must have pacing on)");
1413 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1414 	    SYSCTL_CHILDREN(rack_misc),
1415 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1416 	    &rack_verbose_logging, 0,
1417 	    "Should RACK black box logging be verbose");
1418 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1419 	    SYSCTL_CHILDREN(rack_misc),
1420 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1421 	    &rack_ignore_data_after_close, 1,
1422 	    "Do we hold off sending a RST until all pending data is ack'd");
1423 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1424 	    SYSCTL_CHILDREN(rack_misc),
1425 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1426 	    &rack_sack_not_required, 1,
1427 	    "Do we allow rack to run on connections not supporting SACK");
1428 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1429 	    SYSCTL_CHILDREN(rack_misc),
1430 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1431 	    &rack_send_a_lot_in_prr, 1,
1432 	    "Send a lot in prr");
1433 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1434 	    SYSCTL_CHILDREN(rack_misc),
1435 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1436 	    &rack_autosndbuf_inc, 20,
1437 	    "What percentage should rack scale up its snd buffer by?");
1438 	/* Sack Attacker detection stuff */
1439 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1440 	    SYSCTL_CHILDREN(rack_attack),
1441 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1442 	    &rack_highest_sack_thresh_seen, 0,
1443 	    "Highest sack to ack ratio seen");
1444 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1445 	    SYSCTL_CHILDREN(rack_attack),
1446 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1447 	    &rack_highest_move_thresh_seen, 0,
1448 	    "Highest move to non-move ratio seen");
1449 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1450 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1451 	    SYSCTL_CHILDREN(rack_attack),
1452 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1453 	    &rack_ack_total,
1454 	    "Total number of Ack's");
1455 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1456 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1457 	    SYSCTL_CHILDREN(rack_attack),
1458 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1459 	    &rack_express_sack,
1460 	    "Total expresss number of Sack's");
1461 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1462 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1463 	    SYSCTL_CHILDREN(rack_attack),
1464 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1465 	    &rack_sack_total,
1466 	    "Total number of SACKs");
1467 	rack_move_none = counter_u64_alloc(M_WAITOK);
1468 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1469 	    SYSCTL_CHILDREN(rack_attack),
1470 	    OID_AUTO, "move_none", CTLFLAG_RD,
1471 	    &rack_move_none,
1472 	    "Total number of SACK index reuse of postions under threshold");
1473 	rack_move_some = counter_u64_alloc(M_WAITOK);
1474 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1475 	    SYSCTL_CHILDREN(rack_attack),
1476 	    OID_AUTO, "move_some", CTLFLAG_RD,
1477 	    &rack_move_some,
1478 	    "Total number of SACK index reuse of postions over threshold");
1479 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1480 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1481 	    SYSCTL_CHILDREN(rack_attack),
1482 	    OID_AUTO, "attacks", CTLFLAG_RD,
1483 	    &rack_sack_attacks_detected,
1484 	    "Total number of SACK attackers that had sack disabled");
1485 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1486 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1487 	    SYSCTL_CHILDREN(rack_attack),
1488 	    OID_AUTO, "reversed", CTLFLAG_RD,
1489 	    &rack_sack_attacks_reversed,
1490 	    "Total number of SACK attackers that were later determined false positive");
1491 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1492 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1493 	    SYSCTL_CHILDREN(rack_attack),
1494 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1495 	    &rack_sack_used_next_merge,
1496 	    "Total number of times we used the next merge");
1497 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1498 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1499 	    SYSCTL_CHILDREN(rack_attack),
1500 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1501 	    &rack_sack_used_prev_merge,
1502 	    "Total number of times we used the prev merge");
1503 	/* Counters */
1504 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1505 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1506 	    SYSCTL_CHILDREN(rack_counters),
1507 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1508 	    &rack_fto_send, "Total number of rack_fast_output sends");
1509 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1510 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1511 	    SYSCTL_CHILDREN(rack_counters),
1512 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1513 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1514 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1515 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1516 	    SYSCTL_CHILDREN(rack_counters),
1517 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1518 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1519 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1520 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1521 	    SYSCTL_CHILDREN(rack_counters),
1522 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1523 	    &rack_non_fto_send, "Total number of rack_output first sends");
1524 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1525 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1526 	    SYSCTL_CHILDREN(rack_counters),
1527 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1528 	    &rack_extended_rfo, "Total number of times we extended rfo");
1529 
1530 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1531 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1532 	    SYSCTL_CHILDREN(rack_counters),
1533 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1534 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1535 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1536 
1537 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1538 	    SYSCTL_CHILDREN(rack_counters),
1539 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1540 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1541 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1542 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1543 	    SYSCTL_CHILDREN(rack_counters),
1544 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1545 	    &rack_tlp_tot,
1546 	    "Total number of tail loss probe expirations");
1547 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1548 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1549 	    SYSCTL_CHILDREN(rack_counters),
1550 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1551 	    &rack_tlp_newdata,
1552 	    "Total number of tail loss probe sending new data");
1553 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1554 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1555 	    SYSCTL_CHILDREN(rack_counters),
1556 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1557 	    &rack_tlp_retran,
1558 	    "Total number of tail loss probe sending retransmitted data");
1559 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1560 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1561 	    SYSCTL_CHILDREN(rack_counters),
1562 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1563 	    &rack_tlp_retran_bytes,
1564 	    "Total bytes of tail loss probe sending retransmitted data");
1565 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1566 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1567 	    SYSCTL_CHILDREN(rack_counters),
1568 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1569 	    &rack_to_tot,
1570 	    "Total number of times the rack to expired");
1571 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1572 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1573 	    SYSCTL_CHILDREN(rack_counters),
1574 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1575 	    &rack_saw_enobuf,
1576 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1577 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1578 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1579 	    SYSCTL_CHILDREN(rack_counters),
1580 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1581 	    &rack_saw_enobuf_hw,
1582 	    "Total number of times a send returned enobuf for hdwr paced connections");
1583 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1584 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1585 	    SYSCTL_CHILDREN(rack_counters),
1586 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1587 	    &rack_saw_enetunreach,
1588 	    "Total number of times a send received a enetunreachable");
1589 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1590 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1591 	    SYSCTL_CHILDREN(rack_counters),
1592 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1593 	    &rack_hot_alloc,
1594 	    "Total allocations from the top of our list");
1595 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1596 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1597 	    SYSCTL_CHILDREN(rack_counters),
1598 	    OID_AUTO, "allocs", CTLFLAG_RD,
1599 	    &rack_to_alloc,
1600 	    "Total allocations of tracking structures");
1601 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1602 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1603 	    SYSCTL_CHILDREN(rack_counters),
1604 	    OID_AUTO, "allochard", CTLFLAG_RD,
1605 	    &rack_to_alloc_hard,
1606 	    "Total allocations done with sleeping the hard way");
1607 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1608 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1609 	    SYSCTL_CHILDREN(rack_counters),
1610 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1611 	    &rack_to_alloc_emerg,
1612 	    "Total allocations done from emergency cache");
1613 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1614 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1615 	    SYSCTL_CHILDREN(rack_counters),
1616 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1617 	    &rack_to_alloc_limited,
1618 	    "Total allocations dropped due to limit");
1619 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1620 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1621 	    SYSCTL_CHILDREN(rack_counters),
1622 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1623 	    &rack_alloc_limited_conns,
1624 	    "Connections with allocations dropped due to limit");
1625 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1626 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1627 	    SYSCTL_CHILDREN(rack_counters),
1628 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1629 	    &rack_split_limited,
1630 	    "Split allocations dropped due to limit");
1631 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1632 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1633 	    SYSCTL_CHILDREN(rack_counters),
1634 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1635 	    &rack_persists_sends,
1636 	    "Number of times we sent a persist probe");
1637 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1638 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1639 	    SYSCTL_CHILDREN(rack_counters),
1640 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1641 	    &rack_persists_acks,
1642 	    "Number of times a persist probe was acked");
1643 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1644 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1645 	    SYSCTL_CHILDREN(rack_counters),
1646 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1647 	    &rack_persists_loss,
1648 	    "Number of times we detected a lost persist probe (no ack)");
1649 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1650 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1651 	    SYSCTL_CHILDREN(rack_counters),
1652 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1653 	    &rack_persists_lost_ends,
1654 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1655 #ifdef INVARIANTS
1656 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1657 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1658 	    SYSCTL_CHILDREN(rack_counters),
1659 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1660 	    &rack_adjust_map_bw,
1661 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1662 #endif
1663 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1664 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1665 	    SYSCTL_CHILDREN(rack_counters),
1666 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1667 	    &rack_multi_single_eq,
1668 	    "Number of compressed acks total represented");
1669 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1670 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1671 	    SYSCTL_CHILDREN(rack_counters),
1672 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1673 	    &rack_proc_non_comp_ack,
1674 	    "Number of non compresseds acks that we processed");
1675 
1676 
1677 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1678 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1679 	    SYSCTL_CHILDREN(rack_counters),
1680 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1681 	    &rack_sack_proc_all,
1682 	    "Total times we had to walk whole list for sack processing");
1683 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1684 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1685 	    SYSCTL_CHILDREN(rack_counters),
1686 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1687 	    &rack_sack_proc_restart,
1688 	    "Total times we had to walk whole list due to a restart");
1689 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1690 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1691 	    SYSCTL_CHILDREN(rack_counters),
1692 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1693 	    &rack_sack_proc_short,
1694 	    "Total times we took shortcut for sack processing");
1695 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1696 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1697 	    SYSCTL_CHILDREN(rack_attack),
1698 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1699 	    &rack_sack_skipped_acked,
1700 	    "Total number of times we skipped previously sacked");
1701 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1702 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1703 	    SYSCTL_CHILDREN(rack_attack),
1704 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1705 	    &rack_sack_splits,
1706 	    "Total number of times we did the old fashion tree split");
1707 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1708 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1709 	    SYSCTL_CHILDREN(rack_counters),
1710 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1711 	    &rack_input_idle_reduces,
1712 	    "Total number of idle reductions on input");
1713 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1714 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1715 	    SYSCTL_CHILDREN(rack_counters),
1716 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1717 	    &rack_collapsed_win,
1718 	    "Total number of collapsed windows");
1719 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1720 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1721 	    SYSCTL_CHILDREN(rack_counters),
1722 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1723 	    &rack_try_scwnd,
1724 	    "Total number of scwnd attempts");
1725 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1726 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1727 	    OID_AUTO, "outsize", CTLFLAG_RD,
1728 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1729 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1730 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1731 	    OID_AUTO, "opts", CTLFLAG_RD,
1732 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1733 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1734 	    SYSCTL_CHILDREN(rack_sysctl_root),
1735 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1736 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1737 }
1738 
1739 static __inline int
1740 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1741 {
1742 	if (SEQ_GEQ(b->r_start, a->r_start) &&
1743 	    SEQ_LT(b->r_start, a->r_end)) {
1744 		/*
1745 		 * The entry b is within the
1746 		 * block a. i.e.:
1747 		 * a --   |-------------|
1748 		 * b --   |----|
1749 		 * <or>
1750 		 * b --       |------|
1751 		 * <or>
1752 		 * b --       |-----------|
1753 		 */
1754 		return (0);
1755 	} else if (SEQ_GEQ(b->r_start, a->r_end)) {
1756 		/*
1757 		 * b falls as either the next
1758 		 * sequence block after a so a
1759 		 * is said to be smaller than b.
1760 		 * i.e:
1761 		 * a --   |------|
1762 		 * b --          |--------|
1763 		 * or
1764 		 * b --              |-----|
1765 		 */
1766 		return (1);
1767 	}
1768 	/*
1769 	 * Whats left is where a is
1770 	 * larger than b. i.e:
1771 	 * a --         |-------|
1772 	 * b --  |---|
1773 	 * or even possibly
1774 	 * b --   |--------------|
1775 	 */
1776 	return (-1);
1777 }
1778 
1779 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1780 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1781 
1782 static uint32_t
1783 rc_init_window(struct tcp_rack *rack)
1784 {
1785 	uint32_t win;
1786 
1787 	if (rack->rc_init_win == 0) {
1788 		/*
1789 		 * Nothing set by the user, use the system stack
1790 		 * default.
1791 		 */
1792 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1793 	}
1794 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1795 	return (win);
1796 }
1797 
1798 static uint64_t
1799 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1800 {
1801 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1802 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1803 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1804 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1805 	else
1806 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1807 }
1808 
1809 static uint64_t
1810 rack_get_bw(struct tcp_rack *rack)
1811 {
1812 	if (rack->use_fixed_rate) {
1813 		/* Return the fixed pacing rate */
1814 		return (rack_get_fixed_pacing_bw(rack));
1815 	}
1816 	if (rack->r_ctl.gp_bw == 0) {
1817 		/*
1818 		 * We have yet no b/w measurement,
1819 		 * if we have a user set initial bw
1820 		 * return it. If we don't have that and
1821 		 * we have an srtt, use the tcp IW (10) to
1822 		 * calculate a fictional b/w over the SRTT
1823 		 * which is more or less a guess. Note
1824 		 * we don't use our IW from rack on purpose
1825 		 * so if we have like IW=30, we are not
1826 		 * calculating a "huge" b/w.
1827 		 */
1828 		uint64_t bw, srtt;
1829 		if (rack->r_ctl.init_rate)
1830 			return (rack->r_ctl.init_rate);
1831 
1832 		/* Has the user set a max peak rate? */
1833 #ifdef NETFLIX_PEAKRATE
1834 		if (rack->rc_tp->t_maxpeakrate)
1835 			return (rack->rc_tp->t_maxpeakrate);
1836 #endif
1837 		/* Ok lets come up with the IW guess, if we have a srtt */
1838 		if (rack->rc_tp->t_srtt == 0) {
1839 			/*
1840 			 * Go with old pacing method
1841 			 * i.e. burst mitigation only.
1842 			 */
1843 			return (0);
1844 		}
1845 		/* Ok lets get the initial TCP win (not racks) */
1846 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1847 		srtt = (uint64_t)rack->rc_tp->t_srtt;
1848 		bw *= (uint64_t)USECS_IN_SECOND;
1849 		bw /= srtt;
1850 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1851 			bw = rack->r_ctl.bw_rate_cap;
1852 		return (bw);
1853 	} else {
1854 		uint64_t bw;
1855 
1856 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
1857 			/* Averaging is done, we can return the value */
1858 			bw = rack->r_ctl.gp_bw;
1859 		} else {
1860 			/* Still doing initial average must calculate */
1861 			bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
1862 		}
1863 #ifdef NETFLIX_PEAKRATE
1864 		if ((rack->rc_tp->t_maxpeakrate) &&
1865 		    (bw > rack->rc_tp->t_maxpeakrate)) {
1866 			/* The user has set a peak rate to pace at
1867 			 * don't allow us to pace faster than that.
1868 			 */
1869 			return (rack->rc_tp->t_maxpeakrate);
1870 		}
1871 #endif
1872 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1873 			bw = rack->r_ctl.bw_rate_cap;
1874 		return (bw);
1875 	}
1876 }
1877 
1878 static uint16_t
1879 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1880 {
1881 	if (rack->use_fixed_rate) {
1882 		return (100);
1883 	} else if (rack->in_probe_rtt && (rsm == NULL))
1884 		return (rack->r_ctl.rack_per_of_gp_probertt);
1885 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
1886 		  rack->r_ctl.rack_per_of_gp_rec)) {
1887 		if (rsm) {
1888 			/* a retransmission always use the recovery rate */
1889 			return (rack->r_ctl.rack_per_of_gp_rec);
1890 		} else if (rack->rack_rec_nonrxt_use_cr) {
1891 			/* Directed to use the configured rate */
1892 			goto configured_rate;
1893 		} else if (rack->rack_no_prr &&
1894 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1895 			/* No PRR, lets just use the b/w estimate only */
1896 			return (100);
1897 		} else {
1898 			/*
1899 			 * Here we may have a non-retransmit but we
1900 			 * have no overrides, so just use the recovery
1901 			 * rate (prr is in effect).
1902 			 */
1903 			return (rack->r_ctl.rack_per_of_gp_rec);
1904 		}
1905 	}
1906 configured_rate:
1907 	/* For the configured rate we look at our cwnd vs the ssthresh */
1908 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1909 		return (rack->r_ctl.rack_per_of_gp_ss);
1910 	else
1911 		return (rack->r_ctl.rack_per_of_gp_ca);
1912 }
1913 
1914 static void
1915 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
1916 {
1917 	/*
1918 	 * Types of logs (mod value)
1919 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
1920 	 * 2 = a dsack round begins, persist is reset to 16.
1921 	 * 3 = a dsack round ends
1922 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
1923 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
1924 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
1925 	 */
1926 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1927 		union tcp_log_stackspecific log;
1928 		struct timeval tv;
1929 
1930 		memset(&log, 0, sizeof(log));
1931 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
1932 		log.u_bbr.flex1 <<= 1;
1933 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
1934 		log.u_bbr.flex1 <<= 1;
1935 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
1936 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
1937 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
1938 		log.u_bbr.flex4 = flex4;
1939 		log.u_bbr.flex5 = flex5;
1940 		log.u_bbr.flex6 = flex6;
1941 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
1942 		log.u_bbr.flex8 = mod;
1943 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1944 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1945 		    &rack->rc_inp->inp_socket->so_rcv,
1946 		    &rack->rc_inp->inp_socket->so_snd,
1947 		    RACK_DSACK_HANDLING, 0,
1948 		    0, &log, false, &tv);
1949 	}
1950 }
1951 
1952 static void
1953 rack_log_hdwr_pacing(struct tcp_rack *rack,
1954 		     uint64_t rate, uint64_t hw_rate, int line,
1955 		     int error, uint16_t mod)
1956 {
1957 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1958 		union tcp_log_stackspecific log;
1959 		struct timeval tv;
1960 		const struct ifnet *ifp;
1961 
1962 		memset(&log, 0, sizeof(log));
1963 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
1964 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
1965 		if (rack->r_ctl.crte) {
1966 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
1967 		} else if (rack->rc_inp->inp_route.ro_nh &&
1968 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
1969 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
1970 		} else
1971 			ifp = NULL;
1972 		if (ifp) {
1973 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
1974 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
1975 		}
1976 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1977 		log.u_bbr.bw_inuse = rate;
1978 		log.u_bbr.flex5 = line;
1979 		log.u_bbr.flex6 = error;
1980 		log.u_bbr.flex7 = mod;
1981 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
1982 		log.u_bbr.flex8 = rack->use_fixed_rate;
1983 		log.u_bbr.flex8 <<= 1;
1984 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
1985 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
1986 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
1987 		if (rack->r_ctl.crte)
1988 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
1989 		else
1990 			log.u_bbr.cur_del_rate = 0;
1991 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
1992 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1993 		    &rack->rc_inp->inp_socket->so_rcv,
1994 		    &rack->rc_inp->inp_socket->so_snd,
1995 		    BBR_LOG_HDWR_PACE, 0,
1996 		    0, &log, false, &tv);
1997 	}
1998 }
1999 
2000 static uint64_t
2001 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2002 {
2003 	/*
2004 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2005 	 */
2006 	uint64_t bw_est, high_rate;
2007 	uint64_t gain;
2008 
2009 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2010 	bw_est = bw * gain;
2011 	bw_est /= (uint64_t)100;
2012 	/* Never fall below the minimum (def 64kbps) */
2013 	if (bw_est < RACK_MIN_BW)
2014 		bw_est = RACK_MIN_BW;
2015 	if (rack->r_rack_hw_rate_caps) {
2016 		/* Rate caps are in place */
2017 		if (rack->r_ctl.crte != NULL) {
2018 			/* We have a hdwr rate already */
2019 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2020 			if (bw_est >= high_rate) {
2021 				/* We are capping bw at the highest rate table entry */
2022 				rack_log_hdwr_pacing(rack,
2023 						     bw_est, high_rate, __LINE__,
2024 						     0, 3);
2025 				bw_est = high_rate;
2026 				if (capped)
2027 					*capped = 1;
2028 			}
2029 		} else if ((rack->rack_hdrw_pacing == 0) &&
2030 			   (rack->rack_hdw_pace_ena) &&
2031 			   (rack->rack_attempt_hdwr_pace == 0) &&
2032 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2033 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2034 			/*
2035 			 * Special case, we have not yet attempted hardware
2036 			 * pacing, and yet we may, when we do, find out if we are
2037 			 * above the highest rate. We need to know the maxbw for the interface
2038 			 * in question (if it supports ratelimiting). We get back
2039 			 * a 0, if the interface is not found in the RL lists.
2040 			 */
2041 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2042 			if (high_rate) {
2043 				/* Yep, we have a rate is it above this rate? */
2044 				if (bw_est > high_rate) {
2045 					bw_est = high_rate;
2046 					if (capped)
2047 						*capped = 1;
2048 				}
2049 			}
2050 		}
2051 	}
2052 	return (bw_est);
2053 }
2054 
2055 static void
2056 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2057 {
2058 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2059 		union tcp_log_stackspecific log;
2060 		struct timeval tv;
2061 
2062 		if ((mod != 1) && (rack_verbose_logging == 0)) {
2063 			/*
2064 			 * We get 3 values currently for mod
2065 			 * 1 - We are retransmitting and this tells the reason.
2066 			 * 2 - We are clearing a dup-ack count.
2067 			 * 3 - We are incrementing a dup-ack count.
2068 			 *
2069 			 * The clear/increment are only logged
2070 			 * if you have BBverbose on.
2071 			 */
2072 			return;
2073 		}
2074 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2075 		log.u_bbr.flex1 = tsused;
2076 		log.u_bbr.flex2 = thresh;
2077 		log.u_bbr.flex3 = rsm->r_flags;
2078 		log.u_bbr.flex4 = rsm->r_dupack;
2079 		log.u_bbr.flex5 = rsm->r_start;
2080 		log.u_bbr.flex6 = rsm->r_end;
2081 		log.u_bbr.flex8 = mod;
2082 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2083 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2084 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2085 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2086 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2087 		log.u_bbr.pacing_gain = rack->r_must_retran;
2088 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2089 		    &rack->rc_inp->inp_socket->so_rcv,
2090 		    &rack->rc_inp->inp_socket->so_snd,
2091 		    BBR_LOG_SETTINGS_CHG, 0,
2092 		    0, &log, false, &tv);
2093 	}
2094 }
2095 
2096 static void
2097 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2098 {
2099 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2100 		union tcp_log_stackspecific log;
2101 		struct timeval tv;
2102 
2103 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2104 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2105 		log.u_bbr.flex2 = to;
2106 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2107 		log.u_bbr.flex4 = slot;
2108 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2109 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2110 		log.u_bbr.flex7 = rack->rc_in_persist;
2111 		log.u_bbr.flex8 = which;
2112 		if (rack->rack_no_prr)
2113 			log.u_bbr.pkts_out = 0;
2114 		else
2115 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2116 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2117 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2118 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2119 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2120 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2121 		log.u_bbr.pacing_gain = rack->r_must_retran;
2122 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
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 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2514 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2515 		    &rack->rc_inp->inp_socket->so_rcv,
2516 		    &rack->rc_inp->inp_socket->so_snd,
2517 		    BBR_LOG_JUSTRET, 0,
2518 		    tlen, &log, false, &tv);
2519 	}
2520 }
2521 
2522 static void
2523 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2524 		   struct timeval *tv, uint32_t flags_on_entry)
2525 {
2526 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2527 		union tcp_log_stackspecific log;
2528 
2529 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2530 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2531 		log.u_bbr.flex1 = line;
2532 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2533 		log.u_bbr.flex3 = flags_on_entry;
2534 		log.u_bbr.flex4 = us_cts;
2535 		if (rack->rack_no_prr)
2536 			log.u_bbr.flex5 = 0;
2537 		else
2538 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2539 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2540 		log.u_bbr.flex7 = hpts_removed;
2541 		log.u_bbr.flex8 = 1;
2542 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2543 		log.u_bbr.timeStamp = us_cts;
2544 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2545 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2546 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2547 		log.u_bbr.pacing_gain = rack->r_must_retran;
2548 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2549 		    &rack->rc_inp->inp_socket->so_rcv,
2550 		    &rack->rc_inp->inp_socket->so_snd,
2551 		    BBR_LOG_TIMERCANC, 0,
2552 		    0, &log, false, tv);
2553 	}
2554 }
2555 
2556 static void
2557 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2558 			  uint32_t flex1, uint32_t flex2,
2559 			  uint32_t flex3, uint32_t flex4,
2560 			  uint32_t flex5, uint32_t flex6,
2561 			  uint16_t flex7, uint8_t mod)
2562 {
2563 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2564 		union tcp_log_stackspecific log;
2565 		struct timeval tv;
2566 
2567 		if (mod == 1) {
2568 			/* No you can't use 1, its for the real to cancel */
2569 			return;
2570 		}
2571 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2572 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2573 		log.u_bbr.flex1 = flex1;
2574 		log.u_bbr.flex2 = flex2;
2575 		log.u_bbr.flex3 = flex3;
2576 		log.u_bbr.flex4 = flex4;
2577 		log.u_bbr.flex5 = flex5;
2578 		log.u_bbr.flex6 = flex6;
2579 		log.u_bbr.flex7 = flex7;
2580 		log.u_bbr.flex8 = mod;
2581 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2582 		    &rack->rc_inp->inp_socket->so_rcv,
2583 		    &rack->rc_inp->inp_socket->so_snd,
2584 		    BBR_LOG_TIMERCANC, 0,
2585 		    0, &log, false, &tv);
2586 	}
2587 }
2588 
2589 static void
2590 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2591 {
2592 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2593 		union tcp_log_stackspecific log;
2594 		struct timeval tv;
2595 
2596 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2597 		log.u_bbr.flex1 = timers;
2598 		log.u_bbr.flex2 = ret;
2599 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2600 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2601 		log.u_bbr.flex5 = cts;
2602 		if (rack->rack_no_prr)
2603 			log.u_bbr.flex6 = 0;
2604 		else
2605 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2606 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2607 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2608 		log.u_bbr.pacing_gain = rack->r_must_retran;
2609 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2610 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2611 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2612 		    &rack->rc_inp->inp_socket->so_rcv,
2613 		    &rack->rc_inp->inp_socket->so_snd,
2614 		    BBR_LOG_TO_PROCESS, 0,
2615 		    0, &log, false, &tv);
2616 	}
2617 }
2618 
2619 static void
2620 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
2621 {
2622 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2623 		union tcp_log_stackspecific log;
2624 		struct timeval tv;
2625 
2626 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2627 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2628 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2629 		if (rack->rack_no_prr)
2630 			log.u_bbr.flex3 = 0;
2631 		else
2632 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2633 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2634 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2635 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2636 		log.u_bbr.flex7 = line;
2637 		log.u_bbr.flex8 = frm;
2638 		log.u_bbr.pkts_out = orig_cwnd;
2639 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2640 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2641 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2642 		log.u_bbr.use_lt_bw <<= 1;
2643 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2644 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2645 		    &rack->rc_inp->inp_socket->so_rcv,
2646 		    &rack->rc_inp->inp_socket->so_snd,
2647 		    BBR_LOG_BBRUPD, 0,
2648 		    0, &log, false, &tv);
2649 	}
2650 }
2651 
2652 #ifdef NETFLIX_EXP_DETECTION
2653 static void
2654 rack_log_sad(struct tcp_rack *rack, int event)
2655 {
2656 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2657 		union tcp_log_stackspecific log;
2658 		struct timeval tv;
2659 
2660 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2661 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
2662 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2663 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2664 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2665 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2666 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2667 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2668 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2669 		log.u_bbr.lt_epoch |= rack->do_detection;
2670 		log.u_bbr.applimited = tcp_map_minimum;
2671 		log.u_bbr.flex7 = rack->sack_attack_disable;
2672 		log.u_bbr.flex8 = event;
2673 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2674 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2675 		log.u_bbr.delivered = tcp_sad_decay_val;
2676 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2677 		    &rack->rc_inp->inp_socket->so_rcv,
2678 		    &rack->rc_inp->inp_socket->so_snd,
2679 		    TCP_SAD_DETECTION, 0,
2680 		    0, &log, false, &tv);
2681 	}
2682 }
2683 #endif
2684 
2685 static void
2686 rack_counter_destroy(void)
2687 {
2688 	counter_u64_free(rack_fto_send);
2689 	counter_u64_free(rack_fto_rsm_send);
2690 	counter_u64_free(rack_nfto_resend);
2691 	counter_u64_free(rack_hw_pace_init_fail);
2692 	counter_u64_free(rack_hw_pace_lost);
2693 	counter_u64_free(rack_non_fto_send);
2694 	counter_u64_free(rack_extended_rfo);
2695 	counter_u64_free(rack_ack_total);
2696 	counter_u64_free(rack_express_sack);
2697 	counter_u64_free(rack_sack_total);
2698 	counter_u64_free(rack_move_none);
2699 	counter_u64_free(rack_move_some);
2700 	counter_u64_free(rack_sack_attacks_detected);
2701 	counter_u64_free(rack_sack_attacks_reversed);
2702 	counter_u64_free(rack_sack_used_next_merge);
2703 	counter_u64_free(rack_sack_used_prev_merge);
2704 	counter_u64_free(rack_tlp_tot);
2705 	counter_u64_free(rack_tlp_newdata);
2706 	counter_u64_free(rack_tlp_retran);
2707 	counter_u64_free(rack_tlp_retran_bytes);
2708 	counter_u64_free(rack_to_tot);
2709 	counter_u64_free(rack_saw_enobuf);
2710 	counter_u64_free(rack_saw_enobuf_hw);
2711 	counter_u64_free(rack_saw_enetunreach);
2712 	counter_u64_free(rack_hot_alloc);
2713 	counter_u64_free(rack_to_alloc);
2714 	counter_u64_free(rack_to_alloc_hard);
2715 	counter_u64_free(rack_to_alloc_emerg);
2716 	counter_u64_free(rack_to_alloc_limited);
2717 	counter_u64_free(rack_alloc_limited_conns);
2718 	counter_u64_free(rack_split_limited);
2719 	counter_u64_free(rack_multi_single_eq);
2720 	counter_u64_free(rack_proc_non_comp_ack);
2721 	counter_u64_free(rack_sack_proc_all);
2722 	counter_u64_free(rack_sack_proc_restart);
2723 	counter_u64_free(rack_sack_proc_short);
2724 	counter_u64_free(rack_sack_skipped_acked);
2725 	counter_u64_free(rack_sack_splits);
2726 	counter_u64_free(rack_input_idle_reduces);
2727 	counter_u64_free(rack_collapsed_win);
2728 	counter_u64_free(rack_try_scwnd);
2729 	counter_u64_free(rack_persists_sends);
2730 	counter_u64_free(rack_persists_acks);
2731 	counter_u64_free(rack_persists_loss);
2732 	counter_u64_free(rack_persists_lost_ends);
2733 #ifdef INVARIANTS
2734 	counter_u64_free(rack_adjust_map_bw);
2735 #endif
2736 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2737 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2738 }
2739 
2740 static struct rack_sendmap *
2741 rack_alloc(struct tcp_rack *rack)
2742 {
2743 	struct rack_sendmap *rsm;
2744 
2745 	/*
2746 	 * First get the top of the list it in
2747 	 * theory is the "hottest" rsm we have,
2748 	 * possibly just freed by ack processing.
2749 	 */
2750 	if (rack->rc_free_cnt > rack_free_cache) {
2751 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2752 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2753 		counter_u64_add(rack_hot_alloc, 1);
2754 		rack->rc_free_cnt--;
2755 		return (rsm);
2756 	}
2757 	/*
2758 	 * Once we get under our free cache we probably
2759 	 * no longer have a "hot" one available. Lets
2760 	 * get one from UMA.
2761 	 */
2762 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
2763 	if (rsm) {
2764 		rack->r_ctl.rc_num_maps_alloced++;
2765 		counter_u64_add(rack_to_alloc, 1);
2766 		return (rsm);
2767 	}
2768 	/*
2769 	 * Dig in to our aux rsm's (the last two) since
2770 	 * UMA failed to get us one.
2771 	 */
2772 	if (rack->rc_free_cnt) {
2773 		counter_u64_add(rack_to_alloc_emerg, 1);
2774 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2775 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2776 		rack->rc_free_cnt--;
2777 		return (rsm);
2778 	}
2779 	return (NULL);
2780 }
2781 
2782 static struct rack_sendmap *
2783 rack_alloc_full_limit(struct tcp_rack *rack)
2784 {
2785 	if ((V_tcp_map_entries_limit > 0) &&
2786 	    (rack->do_detection == 0) &&
2787 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2788 		counter_u64_add(rack_to_alloc_limited, 1);
2789 		if (!rack->alloc_limit_reported) {
2790 			rack->alloc_limit_reported = 1;
2791 			counter_u64_add(rack_alloc_limited_conns, 1);
2792 		}
2793 		return (NULL);
2794 	}
2795 	return (rack_alloc(rack));
2796 }
2797 
2798 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2799 static struct rack_sendmap *
2800 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2801 {
2802 	struct rack_sendmap *rsm;
2803 
2804 	if (limit_type) {
2805 		/* currently there is only one limit type */
2806 		if (V_tcp_map_split_limit > 0 &&
2807 		    (rack->do_detection == 0) &&
2808 		    rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2809 			counter_u64_add(rack_split_limited, 1);
2810 			if (!rack->alloc_limit_reported) {
2811 				rack->alloc_limit_reported = 1;
2812 				counter_u64_add(rack_alloc_limited_conns, 1);
2813 			}
2814 			return (NULL);
2815 		}
2816 	}
2817 
2818 	/* allocate and mark in the limit type, if set */
2819 	rsm = rack_alloc(rack);
2820 	if (rsm != NULL && limit_type) {
2821 		rsm->r_limit_type = limit_type;
2822 		rack->r_ctl.rc_num_split_allocs++;
2823 	}
2824 	return (rsm);
2825 }
2826 
2827 static void
2828 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2829 {
2830 	if (rsm->r_flags & RACK_APP_LIMITED) {
2831 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
2832 			rack->r_ctl.rc_app_limited_cnt--;
2833 		}
2834 	}
2835 	if (rsm->r_limit_type) {
2836 		/* currently there is only one limit type */
2837 		rack->r_ctl.rc_num_split_allocs--;
2838 	}
2839 	if (rsm == rack->r_ctl.rc_first_appl) {
2840 		if (rack->r_ctl.rc_app_limited_cnt == 0)
2841 			rack->r_ctl.rc_first_appl = NULL;
2842 		else {
2843 			/* Follow the next one out */
2844 			struct rack_sendmap fe;
2845 
2846 			fe.r_start = rsm->r_nseq_appl;
2847 			rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2848 		}
2849 	}
2850 	if (rsm == rack->r_ctl.rc_resend)
2851 		rack->r_ctl.rc_resend = NULL;
2852 	if (rsm == rack->r_ctl.rc_end_appl)
2853 		rack->r_ctl.rc_end_appl = NULL;
2854 	if (rack->r_ctl.rc_tlpsend == rsm)
2855 		rack->r_ctl.rc_tlpsend = NULL;
2856 	if (rack->r_ctl.rc_sacklast == rsm)
2857 		rack->r_ctl.rc_sacklast = NULL;
2858 	memset(rsm, 0, sizeof(struct rack_sendmap));
2859 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
2860 	rack->rc_free_cnt++;
2861 }
2862 
2863 static void
2864 rack_free_trim(struct tcp_rack *rack)
2865 {
2866 	struct rack_sendmap *rsm;
2867 
2868 	/*
2869 	 * Free up all the tail entries until
2870 	 * we get our list down to the limit.
2871 	 */
2872 	while (rack->rc_free_cnt > rack_free_cache) {
2873 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
2874 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2875 		rack->rc_free_cnt--;
2876 		uma_zfree(rack_zone, rsm);
2877 	}
2878 }
2879 
2880 
2881 static uint32_t
2882 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2883 {
2884 	uint64_t srtt, bw, len, tim;
2885 	uint32_t segsiz, def_len, minl;
2886 
2887 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2888 	def_len = rack_def_data_window * segsiz;
2889 	if (rack->rc_gp_filled == 0) {
2890 		/*
2891 		 * We have no measurement (IW is in flight?) so
2892 		 * we can only guess using our data_window sysctl
2893 		 * value (usually 20MSS).
2894 		 */
2895 		return (def_len);
2896 	}
2897 	/*
2898 	 * Now we have a number of factors to consider.
2899 	 *
2900 	 * 1) We have a desired BDP which is usually
2901 	 *    at least 2.
2902 	 * 2) We have a minimum number of rtt's usually 1 SRTT
2903 	 *    but we allow it too to be more.
2904 	 * 3) We want to make sure a measurement last N useconds (if
2905 	 *    we have set rack_min_measure_usec.
2906 	 *
2907 	 * We handle the first concern here by trying to create a data
2908 	 * window of max(rack_def_data_window, DesiredBDP). The
2909 	 * second concern we handle in not letting the measurement
2910 	 * window end normally until at least the required SRTT's
2911 	 * have gone by which is done further below in
2912 	 * rack_enough_for_measurement(). Finally the third concern
2913 	 * we also handle here by calculating how long that time
2914 	 * would take at the current BW and then return the
2915 	 * max of our first calculation and that length. Note
2916 	 * that if rack_min_measure_usec is 0, we don't deal
2917 	 * with concern 3. Also for both Concern 1 and 3 an
2918 	 * application limited period could end the measurement
2919 	 * earlier.
2920 	 *
2921 	 * So lets calculate the BDP with the "known" b/w using
2922 	 * the SRTT has our rtt and then multiply it by the
2923 	 * goal.
2924 	 */
2925 	bw = rack_get_bw(rack);
2926 	srtt = (uint64_t)tp->t_srtt;
2927 	len = bw * srtt;
2928 	len /= (uint64_t)HPTS_USEC_IN_SEC;
2929 	len *= max(1, rack_goal_bdp);
2930 	/* Now we need to round up to the nearest MSS */
2931 	len = roundup(len, segsiz);
2932 	if (rack_min_measure_usec) {
2933 		/* Now calculate our min length for this b/w */
2934 		tim = rack_min_measure_usec;
2935 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
2936 		if (minl == 0)
2937 			minl = 1;
2938 		minl = roundup(minl, segsiz);
2939 		if (len < minl)
2940 			len = minl;
2941 	}
2942 	/*
2943 	 * Now if we have a very small window we want
2944 	 * to attempt to get the window that is
2945 	 * as small as possible. This happens on
2946 	 * low b/w connections and we don't want to
2947 	 * span huge numbers of rtt's between measurements.
2948 	 *
2949 	 * We basically include 2 over our "MIN window" so
2950 	 * that the measurement can be shortened (possibly) by
2951 	 * an ack'ed packet.
2952 	 */
2953 	if (len < def_len)
2954 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
2955 	else
2956 		return (max((uint32_t)len, def_len));
2957 
2958 }
2959 
2960 static int
2961 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
2962 {
2963 	uint32_t tim, srtts, segsiz;
2964 
2965 	/*
2966 	 * Has enough time passed for the GP measurement to be valid?
2967 	 */
2968 	if ((tp->snd_max == tp->snd_una) ||
2969 	    (th_ack == tp->snd_max)){
2970 		/* All is acked */
2971 		*quality = RACK_QUALITY_ALLACKED;
2972 		return (1);
2973 	}
2974 	if (SEQ_LT(th_ack, tp->gput_seq)) {
2975 		/* Not enough bytes yet */
2976 		return (0);
2977 	}
2978 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2979 	if (SEQ_LT(th_ack, tp->gput_ack) &&
2980 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
2981 		/* Not enough bytes yet */
2982 		return (0);
2983 	}
2984 	if (rack->r_ctl.rc_first_appl &&
2985 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
2986 		/*
2987 		 * We are up to the app limited send point
2988 		 * we have to measure irrespective of the time..
2989 		 */
2990 		*quality = RACK_QUALITY_APPLIMITED;
2991 		return (1);
2992 	}
2993 	/* Now what about time? */
2994 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
2995 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
2996 	if (tim >= srtts) {
2997 		*quality = RACK_QUALITY_HIGH;
2998 		return (1);
2999 	}
3000 	/* Nope not even a full SRTT has passed */
3001 	return (0);
3002 }
3003 
3004 static void
3005 rack_log_timely(struct tcp_rack *rack,
3006 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3007 		uint64_t up_bnd, int line, uint8_t method)
3008 {
3009 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3010 		union tcp_log_stackspecific log;
3011 		struct timeval tv;
3012 
3013 		memset(&log, 0, sizeof(log));
3014 		log.u_bbr.flex1 = logged;
3015 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3016 		log.u_bbr.flex2 <<= 4;
3017 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3018 		log.u_bbr.flex2 <<= 4;
3019 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3020 		log.u_bbr.flex2 <<= 4;
3021 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3022 		log.u_bbr.flex3 = rack->rc_gp_incr;
3023 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3024 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3025 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3026 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3027 		log.u_bbr.flex8 = method;
3028 		log.u_bbr.cur_del_rate = cur_bw;
3029 		log.u_bbr.delRate = low_bnd;
3030 		log.u_bbr.bw_inuse = up_bnd;
3031 		log.u_bbr.rttProp = rack_get_bw(rack);
3032 		log.u_bbr.pkt_epoch = line;
3033 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3034 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3035 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3036 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3037 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3038 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3039 		log.u_bbr.cwnd_gain <<= 1;
3040 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3041 		log.u_bbr.cwnd_gain <<= 1;
3042 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3043 		log.u_bbr.cwnd_gain <<= 1;
3044 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3045 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3046 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3047 		    &rack->rc_inp->inp_socket->so_rcv,
3048 		    &rack->rc_inp->inp_socket->so_snd,
3049 		    TCP_TIMELY_WORK, 0,
3050 		    0, &log, false, &tv);
3051 	}
3052 }
3053 
3054 static int
3055 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3056 {
3057 	/*
3058 	 * Before we increase we need to know if
3059 	 * the estimate just made was less than
3060 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3061 	 *
3062 	 * If we already are pacing at a fast enough
3063 	 * rate to push us faster there is no sense of
3064 	 * increasing.
3065 	 *
3066 	 * We first caculate our actual pacing rate (ss or ca multipler
3067 	 * times our cur_bw).
3068 	 *
3069 	 * Then we take the last measured rate and multipy by our
3070 	 * maximum pacing overage to give us a max allowable rate.
3071 	 *
3072 	 * If our act_rate is smaller than our max_allowable rate
3073 	 * then we should increase. Else we should hold steady.
3074 	 *
3075 	 */
3076 	uint64_t act_rate, max_allow_rate;
3077 
3078 	if (rack_timely_no_stopping)
3079 		return (1);
3080 
3081 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3082 		/*
3083 		 * Initial startup case or
3084 		 * everything is acked case.
3085 		 */
3086 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3087 				__LINE__, 9);
3088 		return (1);
3089 	}
3090 	if (mult <= 100) {
3091 		/*
3092 		 * We can always pace at or slightly above our rate.
3093 		 */
3094 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3095 				__LINE__, 9);
3096 		return (1);
3097 	}
3098 	act_rate = cur_bw * (uint64_t)mult;
3099 	act_rate /= 100;
3100 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3101 	max_allow_rate /= 100;
3102 	if (act_rate < max_allow_rate) {
3103 		/*
3104 		 * Here the rate we are actually pacing at
3105 		 * is smaller than 10% above our last measurement.
3106 		 * This means we are pacing below what we would
3107 		 * like to try to achieve (plus some wiggle room).
3108 		 */
3109 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3110 				__LINE__, 9);
3111 		return (1);
3112 	} else {
3113 		/*
3114 		 * Here we are already pacing at least rack_max_per_above(10%)
3115 		 * what we are getting back. This indicates most likely
3116 		 * that we are being limited (cwnd/rwnd/app) and can't
3117 		 * get any more b/w. There is no sense of trying to
3118 		 * raise up the pacing rate its not speeding us up
3119 		 * and we already are pacing faster than we are getting.
3120 		 */
3121 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3122 				__LINE__, 8);
3123 		return (0);
3124 	}
3125 }
3126 
3127 static void
3128 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3129 {
3130 	/*
3131 	 * When we drag bottom, we want to assure
3132 	 * that no multiplier is below 1.0, if so
3133 	 * we want to restore it to at least that.
3134 	 */
3135 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3136 		/* This is unlikely we usually do not touch recovery */
3137 		rack->r_ctl.rack_per_of_gp_rec = 100;
3138 	}
3139 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3140 		rack->r_ctl.rack_per_of_gp_ca = 100;
3141 	}
3142 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3143 		rack->r_ctl.rack_per_of_gp_ss = 100;
3144 	}
3145 }
3146 
3147 static void
3148 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3149 {
3150 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3151 		rack->r_ctl.rack_per_of_gp_ca = 100;
3152 	}
3153 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3154 		rack->r_ctl.rack_per_of_gp_ss = 100;
3155 	}
3156 }
3157 
3158 static void
3159 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3160 {
3161 	int32_t  calc, logged, plus;
3162 
3163 	logged = 0;
3164 
3165 	if (override) {
3166 		/*
3167 		 * override is passed when we are
3168 		 * loosing b/w and making one last
3169 		 * gasp at trying to not loose out
3170 		 * to a new-reno flow.
3171 		 */
3172 		goto extra_boost;
3173 	}
3174 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3175 	if (rack->rc_gp_incr &&
3176 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3177 		/*
3178 		 * Reset and get 5 strokes more before the boost. Note
3179 		 * that the count is 0 based so we have to add one.
3180 		 */
3181 extra_boost:
3182 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3183 		rack->rc_gp_timely_inc_cnt = 0;
3184 	} else
3185 		plus = (uint32_t)rack_gp_increase_per;
3186 	/* Must be at least 1% increase for true timely increases */
3187 	if ((plus < 1) &&
3188 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3189 		plus = 1;
3190 	if (rack->rc_gp_saw_rec &&
3191 	    (rack->rc_gp_no_rec_chg == 0) &&
3192 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3193 				  rack->r_ctl.rack_per_of_gp_rec)) {
3194 		/* We have been in recovery ding it too */
3195 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3196 		if (calc > 0xffff)
3197 			calc = 0xffff;
3198 		logged |= 1;
3199 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3200 		if (rack_per_upper_bound_ss &&
3201 		    (rack->rc_dragged_bottom == 0) &&
3202 		    (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3203 			rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3204 	}
3205 	if (rack->rc_gp_saw_ca &&
3206 	    (rack->rc_gp_saw_ss == 0) &&
3207 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3208 				  rack->r_ctl.rack_per_of_gp_ca)) {
3209 		/* In CA */
3210 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3211 		if (calc > 0xffff)
3212 			calc = 0xffff;
3213 		logged |= 2;
3214 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3215 		if (rack_per_upper_bound_ca &&
3216 		    (rack->rc_dragged_bottom == 0) &&
3217 		    (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3218 			rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3219 	}
3220 	if (rack->rc_gp_saw_ss &&
3221 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3222 				  rack->r_ctl.rack_per_of_gp_ss)) {
3223 		/* In SS */
3224 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3225 		if (calc > 0xffff)
3226 			calc = 0xffff;
3227 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3228 		if (rack_per_upper_bound_ss &&
3229 		    (rack->rc_dragged_bottom == 0) &&
3230 		    (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3231 			rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3232 		logged |= 4;
3233 	}
3234 	if (logged &&
3235 	    (rack->rc_gp_incr == 0)){
3236 		/* Go into increment mode */
3237 		rack->rc_gp_incr = 1;
3238 		rack->rc_gp_timely_inc_cnt = 0;
3239 	}
3240 	if (rack->rc_gp_incr &&
3241 	    logged &&
3242 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3243 		rack->rc_gp_timely_inc_cnt++;
3244 	}
3245 	rack_log_timely(rack,  logged, plus, 0, 0,
3246 			__LINE__, 1);
3247 }
3248 
3249 static uint32_t
3250 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3251 {
3252 	/*
3253 	 * norm_grad = rtt_diff / minrtt;
3254 	 * new_per = curper * (1 - B * norm_grad)
3255 	 *
3256 	 * B = rack_gp_decrease_per (default 10%)
3257 	 * rtt_dif = input var current rtt-diff
3258 	 * curper = input var current percentage
3259 	 * minrtt = from rack filter
3260 	 *
3261 	 */
3262 	uint64_t perf;
3263 
3264 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3265 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3266 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3267 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3268 		     (uint64_t)1000000)) /
3269 		(uint64_t)1000000);
3270 	if (perf > curper) {
3271 		/* TSNH */
3272 		perf = curper - 1;
3273 	}
3274 	return ((uint32_t)perf);
3275 }
3276 
3277 static uint32_t
3278 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3279 {
3280 	/*
3281 	 *                                   highrttthresh
3282 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3283 	 *                                     gp_srtt
3284 	 *
3285 	 * B = rack_gp_decrease_per (default 10%)
3286 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3287 	 */
3288 	uint64_t perf;
3289 	uint32_t highrttthresh;
3290 
3291 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3292 
3293 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3294 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3295 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3296 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3297 	return (perf);
3298 }
3299 
3300 static void
3301 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3302 {
3303 	uint64_t logvar, logvar2, logvar3;
3304 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3305 
3306 	if (rack->rc_gp_incr) {
3307 		/* Turn off increment counting */
3308 		rack->rc_gp_incr = 0;
3309 		rack->rc_gp_timely_inc_cnt = 0;
3310 	}
3311 	ss_red = ca_red = rec_red = 0;
3312 	logged = 0;
3313 	/* Calculate the reduction value */
3314 	if (rtt_diff < 0) {
3315 		rtt_diff *= -1;
3316 	}
3317 	/* Must be at least 1% reduction */
3318 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3319 		/* We have been in recovery ding it too */
3320 		if (timely_says == 2) {
3321 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3322 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3323 			if (alt < new_per)
3324 				val = alt;
3325 			else
3326 				val = new_per;
3327 		} else
3328 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3329 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
3330 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3331 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3332 		} else {
3333 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3334 			rec_red = 0;
3335 		}
3336 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3337 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3338 		logged |= 1;
3339 	}
3340 	if (rack->rc_gp_saw_ss) {
3341 		/* Sent in SS */
3342 		if (timely_says == 2) {
3343 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3344 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3345 			if (alt < new_per)
3346 				val = alt;
3347 			else
3348 				val = new_per;
3349 		} else
3350 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3351 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3352 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3353 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3354 		} else {
3355 			ss_red = new_per;
3356 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3357 			logvar = new_per;
3358 			logvar <<= 32;
3359 			logvar |= alt;
3360 			logvar2 = (uint32_t)rtt;
3361 			logvar2 <<= 32;
3362 			logvar2 |= (uint32_t)rtt_diff;
3363 			logvar3 = rack_gp_rtt_maxmul;
3364 			logvar3 <<= 32;
3365 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3366 			rack_log_timely(rack, timely_says,
3367 					logvar2, logvar3,
3368 					logvar, __LINE__, 10);
3369 		}
3370 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3371 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3372 		logged |= 4;
3373 	} else if (rack->rc_gp_saw_ca) {
3374 		/* Sent in CA */
3375 		if (timely_says == 2) {
3376 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3377 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3378 			if (alt < new_per)
3379 				val = alt;
3380 			else
3381 				val = new_per;
3382 		} else
3383 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3384 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
3385 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3386 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3387 		} else {
3388 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3389 			ca_red = 0;
3390 			logvar = new_per;
3391 			logvar <<= 32;
3392 			logvar |= alt;
3393 			logvar2 = (uint32_t)rtt;
3394 			logvar2 <<= 32;
3395 			logvar2 |= (uint32_t)rtt_diff;
3396 			logvar3 = rack_gp_rtt_maxmul;
3397 			logvar3 <<= 32;
3398 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3399 			rack_log_timely(rack, timely_says,
3400 					logvar2, logvar3,
3401 					logvar, __LINE__, 10);
3402 		}
3403 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3404 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3405 		logged |= 2;
3406 	}
3407 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
3408 		rack->rc_gp_timely_dec_cnt++;
3409 		if (rack_timely_dec_clear &&
3410 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3411 			rack->rc_gp_timely_dec_cnt = 0;
3412 	}
3413 	logvar = ss_red;
3414 	logvar <<= 32;
3415 	logvar |= ca_red;
3416 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3417 			__LINE__, 2);
3418 }
3419 
3420 static void
3421 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3422 		     uint32_t rtt, uint32_t line, uint8_t reas)
3423 {
3424 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3425 		union tcp_log_stackspecific log;
3426 		struct timeval tv;
3427 
3428 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3429 		log.u_bbr.flex1 = line;
3430 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3431 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3432 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3433 		log.u_bbr.flex5 = rtt;
3434 		log.u_bbr.flex6 = rack->rc_highly_buffered;
3435 		log.u_bbr.flex6 <<= 1;
3436 		log.u_bbr.flex6 |= rack->forced_ack;
3437 		log.u_bbr.flex6 <<= 1;
3438 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3439 		log.u_bbr.flex6 <<= 1;
3440 		log.u_bbr.flex6 |= rack->in_probe_rtt;
3441 		log.u_bbr.flex6 <<= 1;
3442 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3443 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3444 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3445 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3446 		log.u_bbr.flex8 = reas;
3447 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3448 		log.u_bbr.delRate = rack_get_bw(rack);
3449 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3450 		log.u_bbr.cur_del_rate <<= 32;
3451 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3452 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3453 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3454 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3455 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3456 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3457 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3458 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3459 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3460 		log.u_bbr.rttProp = us_cts;
3461 		log.u_bbr.rttProp <<= 32;
3462 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3463 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3464 		    &rack->rc_inp->inp_socket->so_rcv,
3465 		    &rack->rc_inp->inp_socket->so_snd,
3466 		    BBR_LOG_RTT_SHRINKS, 0,
3467 		    0, &log, false, &rack->r_ctl.act_rcv_time);
3468 	}
3469 }
3470 
3471 static void
3472 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3473 {
3474 	uint64_t bwdp;
3475 
3476 	bwdp = rack_get_bw(rack);
3477 	bwdp *= (uint64_t)rtt;
3478 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3479 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3480 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3481 		/*
3482 		 * A window protocol must be able to have 4 packets
3483 		 * outstanding as the floor in order to function
3484 		 * (especially considering delayed ack :D).
3485 		 */
3486 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3487 	}
3488 }
3489 
3490 static void
3491 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3492 {
3493 	/**
3494 	 * ProbeRTT is a bit different in rack_pacing than in
3495 	 * BBR. It is like BBR in that it uses the lowering of
3496 	 * the RTT as a signal that we saw something new and
3497 	 * counts from there for how long between. But it is
3498 	 * different in that its quite simple. It does not
3499 	 * play with the cwnd and wait until we get down
3500 	 * to N segments outstanding and hold that for
3501 	 * 200ms. Instead it just sets the pacing reduction
3502 	 * rate to a set percentage (70 by default) and hold
3503 	 * that for a number of recent GP Srtt's.
3504 	 */
3505 	uint32_t segsiz;
3506 
3507 	if (rack->rc_gp_dyn_mul == 0)
3508 		return;
3509 
3510 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3511 		/* We are idle */
3512 		return;
3513 	}
3514 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3515 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3516 		/*
3517 		 * Stop the goodput now, the idea here is
3518 		 * that future measurements with in_probe_rtt
3519 		 * won't register if they are not greater so
3520 		 * we want to get what info (if any) is available
3521 		 * now.
3522 		 */
3523 		rack_do_goodput_measurement(rack->rc_tp, rack,
3524 					    rack->rc_tp->snd_una, __LINE__,
3525 					    RACK_QUALITY_PROBERTT);
3526 	}
3527 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3528 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3529 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3530 		     rack->r_ctl.rc_pace_min_segs);
3531 	rack->in_probe_rtt = 1;
3532 	rack->measure_saw_probe_rtt = 1;
3533 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3534 	rack->r_ctl.rc_time_probertt_starts = 0;
3535 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3536 	if (rack_probertt_use_min_rtt_entry)
3537 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3538 	else
3539 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3540 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3541 			     __LINE__, RACK_RTTS_ENTERPROBE);
3542 }
3543 
3544 static void
3545 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3546 {
3547 	struct rack_sendmap *rsm;
3548 	uint32_t segsiz;
3549 
3550 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3551 		     rack->r_ctl.rc_pace_min_segs);
3552 	rack->in_probe_rtt = 0;
3553 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3554 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3555 		/*
3556 		 * Stop the goodput now, the idea here is
3557 		 * that future measurements with in_probe_rtt
3558 		 * won't register if they are not greater so
3559 		 * we want to get what info (if any) is available
3560 		 * now.
3561 		 */
3562 		rack_do_goodput_measurement(rack->rc_tp, rack,
3563 					    rack->rc_tp->snd_una, __LINE__,
3564 					    RACK_QUALITY_PROBERTT);
3565 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3566 		/*
3567 		 * We don't have enough data to make a measurement.
3568 		 * So lets just stop and start here after exiting
3569 		 * probe-rtt. We probably are not interested in
3570 		 * the results anyway.
3571 		 */
3572 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3573 	}
3574 	/*
3575 	 * Measurements through the current snd_max are going
3576 	 * to be limited by the slower pacing rate.
3577 	 *
3578 	 * We need to mark these as app-limited so we
3579 	 * don't collapse the b/w.
3580 	 */
3581 	rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3582 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3583 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3584 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3585 		else {
3586 			/*
3587 			 * Go out to the end app limited and mark
3588 			 * this new one as next and move the end_appl up
3589 			 * to this guy.
3590 			 */
3591 			if (rack->r_ctl.rc_end_appl)
3592 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3593 			rack->r_ctl.rc_end_appl = rsm;
3594 		}
3595 		rsm->r_flags |= RACK_APP_LIMITED;
3596 		rack->r_ctl.rc_app_limited_cnt++;
3597 	}
3598 	/*
3599 	 * Now, we need to examine our pacing rate multipliers.
3600 	 * If its under 100%, we need to kick it back up to
3601 	 * 100%. We also don't let it be over our "max" above
3602 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3603 	 * Note setting clamp_atexit_prtt to 0 has the effect
3604 	 * of setting CA/SS to 100% always at exit (which is
3605 	 * the default behavior).
3606 	 */
3607 	if (rack_probertt_clear_is) {
3608 		rack->rc_gp_incr = 0;
3609 		rack->rc_gp_bwred = 0;
3610 		rack->rc_gp_timely_inc_cnt = 0;
3611 		rack->rc_gp_timely_dec_cnt = 0;
3612 	}
3613 	/* Do we do any clamping at exit? */
3614 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3615 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3616 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3617 	}
3618 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3619 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3620 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3621 	}
3622 	/*
3623 	 * Lets set rtt_diff to 0, so that we will get a "boost"
3624 	 * after exiting.
3625 	 */
3626 	rack->r_ctl.rc_rtt_diff = 0;
3627 
3628 	/* Clear all flags so we start fresh */
3629 	rack->rc_tp->t_bytes_acked = 0;
3630 	rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3631 	/*
3632 	 * If configured to, set the cwnd and ssthresh to
3633 	 * our targets.
3634 	 */
3635 	if (rack_probe_rtt_sets_cwnd) {
3636 		uint64_t ebdp;
3637 		uint32_t setto;
3638 
3639 		/* Set ssthresh so we get into CA once we hit our target */
3640 		if (rack_probertt_use_min_rtt_exit == 1) {
3641 			/* Set to min rtt */
3642 			rack_set_prtt_target(rack, segsiz,
3643 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3644 		} else if (rack_probertt_use_min_rtt_exit == 2) {
3645 			/* Set to current gp rtt */
3646 			rack_set_prtt_target(rack, segsiz,
3647 					     rack->r_ctl.rc_gp_srtt);
3648 		} else if (rack_probertt_use_min_rtt_exit == 3) {
3649 			/* Set to entry gp rtt */
3650 			rack_set_prtt_target(rack, segsiz,
3651 					     rack->r_ctl.rc_entry_gp_rtt);
3652 		} else {
3653 			uint64_t sum;
3654 			uint32_t setval;
3655 
3656 			sum = rack->r_ctl.rc_entry_gp_rtt;
3657 			sum *= 10;
3658 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3659 			if (sum >= 20) {
3660 				/*
3661 				 * A highly buffered path needs
3662 				 * cwnd space for timely to work.
3663 				 * Lets set things up as if
3664 				 * we are heading back here again.
3665 				 */
3666 				setval = rack->r_ctl.rc_entry_gp_rtt;
3667 			} else if (sum >= 15) {
3668 				/*
3669 				 * Lets take the smaller of the
3670 				 * two since we are just somewhat
3671 				 * buffered.
3672 				 */
3673 				setval = rack->r_ctl.rc_gp_srtt;
3674 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
3675 					setval = rack->r_ctl.rc_entry_gp_rtt;
3676 			} else {
3677 				/*
3678 				 * Here we are not highly buffered
3679 				 * and should pick the min we can to
3680 				 * keep from causing loss.
3681 				 */
3682 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3683 			}
3684 			rack_set_prtt_target(rack, segsiz,
3685 					     setval);
3686 		}
3687 		if (rack_probe_rtt_sets_cwnd > 1) {
3688 			/* There is a percentage here to boost */
3689 			ebdp = rack->r_ctl.rc_target_probertt_flight;
3690 			ebdp *= rack_probe_rtt_sets_cwnd;
3691 			ebdp /= 100;
3692 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3693 		} else
3694 			setto = rack->r_ctl.rc_target_probertt_flight;
3695 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3696 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3697 			/* Enforce a min */
3698 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3699 		}
3700 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
3701 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3702 	}
3703 	rack_log_rtt_shrinks(rack,  us_cts,
3704 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3705 			     __LINE__, RACK_RTTS_EXITPROBE);
3706 	/* Clear times last so log has all the info */
3707 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3708 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3709 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3710 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
3711 }
3712 
3713 static void
3714 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3715 {
3716 	/* Check in on probe-rtt */
3717 	if (rack->rc_gp_filled == 0) {
3718 		/* We do not do p-rtt unless we have gp measurements */
3719 		return;
3720 	}
3721 	if (rack->in_probe_rtt) {
3722 		uint64_t no_overflow;
3723 		uint32_t endtime, must_stay;
3724 
3725 		if (rack->r_ctl.rc_went_idle_time &&
3726 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3727 			/*
3728 			 * We went idle during prtt, just exit now.
3729 			 */
3730 			rack_exit_probertt(rack, us_cts);
3731 		} else if (rack_probe_rtt_safety_val &&
3732 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3733 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3734 			/*
3735 			 * Probe RTT safety value triggered!
3736 			 */
3737 			rack_log_rtt_shrinks(rack,  us_cts,
3738 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3739 					     __LINE__, RACK_RTTS_SAFETY);
3740 			rack_exit_probertt(rack, us_cts);
3741 		}
3742 		/* Calculate the max we will wait */
3743 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3744 		if (rack->rc_highly_buffered)
3745 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3746 		/* Calculate the min we must wait */
3747 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3748 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3749 		    TSTMP_LT(us_cts, endtime)) {
3750 			uint32_t calc;
3751 			/* Do we lower more? */
3752 no_exit:
3753 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3754 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3755 			else
3756 				calc = 0;
3757 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3758 			if (calc) {
3759 				/* Maybe */
3760 				calc *= rack_per_of_gp_probertt_reduce;
3761 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3762 				/* Limit it too */
3763 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3764 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3765 			}
3766 			/* We must reach target or the time set */
3767 			return;
3768 		}
3769 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
3770 			if ((TSTMP_LT(us_cts, must_stay) &&
3771 			     rack->rc_highly_buffered) ||
3772 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3773 			      rack->r_ctl.rc_target_probertt_flight)) {
3774 				/* We are not past the must_stay time */
3775 				goto no_exit;
3776 			}
3777 			rack_log_rtt_shrinks(rack,  us_cts,
3778 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3779 					     __LINE__, RACK_RTTS_REACHTARGET);
3780 			rack->r_ctl.rc_time_probertt_starts = us_cts;
3781 			if (rack->r_ctl.rc_time_probertt_starts == 0)
3782 				rack->r_ctl.rc_time_probertt_starts = 1;
3783 			/* Restore back to our rate we want to pace at in prtt */
3784 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3785 		}
3786 		/*
3787 		 * Setup our end time, some number of gp_srtts plus 200ms.
3788 		 */
3789 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3790 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3791 		if (rack_probertt_gpsrtt_cnt_div)
3792 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3793 		else
3794 			endtime = 0;
3795 		endtime += rack_min_probertt_hold;
3796 		endtime += rack->r_ctl.rc_time_probertt_starts;
3797 		if (TSTMP_GEQ(us_cts,  endtime)) {
3798 			/* yes, exit probertt */
3799 			rack_exit_probertt(rack, us_cts);
3800 		}
3801 
3802 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3803 		/* Go into probertt, its been too long since we went lower */
3804 		rack_enter_probertt(rack, us_cts);
3805 	}
3806 }
3807 
3808 static void
3809 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3810 		       uint32_t rtt, int32_t rtt_diff)
3811 {
3812 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
3813 	uint32_t losses;
3814 
3815 	if ((rack->rc_gp_dyn_mul == 0) ||
3816 	    (rack->use_fixed_rate) ||
3817 	    (rack->in_probe_rtt) ||
3818 	    (rack->rc_always_pace == 0)) {
3819 		/* No dynamic GP multipler in play */
3820 		return;
3821 	}
3822 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3823 	cur_bw = rack_get_bw(rack);
3824 	/* Calculate our up and down range */
3825 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3826 	up_bnd /= 100;
3827 	up_bnd += rack->r_ctl.last_gp_comp_bw;
3828 
3829 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3830 	subfr /= 100;
3831 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3832 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3833 		/*
3834 		 * This is the case where our RTT is above
3835 		 * the max target and we have been configured
3836 		 * to just do timely no bonus up stuff in that case.
3837 		 *
3838 		 * There are two configurations, set to 1, and we
3839 		 * just do timely if we are over our max. If its
3840 		 * set above 1 then we slam the multipliers down
3841 		 * to 100 and then decrement per timely.
3842 		 */
3843 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3844 				__LINE__, 3);
3845 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3846 			rack_validate_multipliers_at_or_below_100(rack);
3847 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3848 	} else if ((last_bw_est < low_bnd) && !losses) {
3849 		/*
3850 		 * We are decreasing this is a bit complicated this
3851 		 * means we are loosing ground. This could be
3852 		 * because another flow entered and we are competing
3853 		 * for b/w with it. This will push the RTT up which
3854 		 * makes timely unusable unless we want to get shoved
3855 		 * into a corner and just be backed off (the age
3856 		 * old problem with delay based CC).
3857 		 *
3858 		 * On the other hand if it was a route change we
3859 		 * would like to stay somewhat contained and not
3860 		 * blow out the buffers.
3861 		 */
3862 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3863 				__LINE__, 3);
3864 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3865 		if (rack->rc_gp_bwred == 0) {
3866 			/* Go into reduction counting */
3867 			rack->rc_gp_bwred = 1;
3868 			rack->rc_gp_timely_dec_cnt = 0;
3869 		}
3870 		if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3871 		    (timely_says == 0)) {
3872 			/*
3873 			 * Push another time with a faster pacing
3874 			 * to try to gain back (we include override to
3875 			 * get a full raise factor).
3876 			 */
3877 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3878 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3879 			    (timely_says == 0) ||
3880 			    (rack_down_raise_thresh == 0)) {
3881 				/*
3882 				 * Do an override up in b/w if we were
3883 				 * below the threshold or if the threshold
3884 				 * is zero we always do the raise.
3885 				 */
3886 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3887 			} else {
3888 				/* Log it stays the same */
3889 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3890 						__LINE__, 11);
3891 			}
3892 			rack->rc_gp_timely_dec_cnt++;
3893 			/* We are not incrementing really no-count */
3894 			rack->rc_gp_incr = 0;
3895 			rack->rc_gp_timely_inc_cnt = 0;
3896 		} else {
3897 			/*
3898 			 * Lets just use the RTT
3899 			 * information and give up
3900 			 * pushing.
3901 			 */
3902 			goto use_timely;
3903 		}
3904 	} else if ((timely_says != 2) &&
3905 		    !losses &&
3906 		    (last_bw_est > up_bnd)) {
3907 		/*
3908 		 * We are increasing b/w lets keep going, updating
3909 		 * our b/w and ignoring any timely input, unless
3910 		 * of course we are at our max raise (if there is one).
3911 		 */
3912 
3913 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3914 				__LINE__, 3);
3915 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3916 		if (rack->rc_gp_saw_ss &&
3917 		    rack_per_upper_bound_ss &&
3918 		     (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3919 			    /*
3920 			     * In cases where we can't go higher
3921 			     * we should just use timely.
3922 			     */
3923 			    goto use_timely;
3924 		}
3925 		if (rack->rc_gp_saw_ca &&
3926 		    rack_per_upper_bound_ca &&
3927 		    (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
3928 			    /*
3929 			     * In cases where we can't go higher
3930 			     * we should just use timely.
3931 			     */
3932 			    goto use_timely;
3933 		}
3934 		rack->rc_gp_bwred = 0;
3935 		rack->rc_gp_timely_dec_cnt = 0;
3936 		/* You get a set number of pushes if timely is trying to reduce */
3937 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
3938 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3939 		} else {
3940 			/* Log it stays the same */
3941 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
3942 			    __LINE__, 12);
3943 		}
3944 		return;
3945 	} else {
3946 		/*
3947 		 * We are staying between the lower and upper range bounds
3948 		 * so use timely to decide.
3949 		 */
3950 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3951 				__LINE__, 3);
3952 use_timely:
3953 		if (timely_says) {
3954 			rack->rc_gp_incr = 0;
3955 			rack->rc_gp_timely_inc_cnt = 0;
3956 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
3957 			    !losses &&
3958 			    (last_bw_est < low_bnd)) {
3959 				/* We are loosing ground */
3960 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3961 				rack->rc_gp_timely_dec_cnt++;
3962 				/* We are not incrementing really no-count */
3963 				rack->rc_gp_incr = 0;
3964 				rack->rc_gp_timely_inc_cnt = 0;
3965 			} else
3966 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3967 		} else {
3968 			rack->rc_gp_bwred = 0;
3969 			rack->rc_gp_timely_dec_cnt = 0;
3970 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3971 		}
3972 	}
3973 }
3974 
3975 static int32_t
3976 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
3977 {
3978 	int32_t timely_says;
3979 	uint64_t log_mult, log_rtt_a_diff;
3980 
3981 	log_rtt_a_diff = rtt;
3982 	log_rtt_a_diff <<= 32;
3983 	log_rtt_a_diff |= (uint32_t)rtt_diff;
3984 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
3985 		    rack_gp_rtt_maxmul)) {
3986 		/* Reduce the b/w multipler */
3987 		timely_says = 2;
3988 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3989 		log_mult <<= 32;
3990 		log_mult |= prev_rtt;
3991 		rack_log_timely(rack,  timely_says, log_mult,
3992 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3993 				log_rtt_a_diff, __LINE__, 4);
3994 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3995 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
3996 			    max(rack_gp_rtt_mindiv , 1)))) {
3997 		/* Increase the b/w multipler */
3998 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3999 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4000 			 max(rack_gp_rtt_mindiv , 1));
4001 		log_mult <<= 32;
4002 		log_mult |= prev_rtt;
4003 		timely_says = 0;
4004 		rack_log_timely(rack,  timely_says, log_mult ,
4005 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4006 				log_rtt_a_diff, __LINE__, 5);
4007 	} else {
4008 		/*
4009 		 * Use a gradient to find it the timely gradient
4010 		 * is:
4011 		 * grad = rc_rtt_diff / min_rtt;
4012 		 *
4013 		 * anything below or equal to 0 will be
4014 		 * a increase indication. Anything above
4015 		 * zero is a decrease. Note we take care
4016 		 * of the actual gradient calculation
4017 		 * in the reduction (its not needed for
4018 		 * increase).
4019 		 */
4020 		log_mult = prev_rtt;
4021 		if (rtt_diff <= 0) {
4022 			/*
4023 			 * Rttdiff is less than zero, increase the
4024 			 * b/w multipler (its 0 or negative)
4025 			 */
4026 			timely_says = 0;
4027 			rack_log_timely(rack,  timely_says, log_mult,
4028 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4029 		} else {
4030 			/* Reduce the b/w multipler */
4031 			timely_says = 1;
4032 			rack_log_timely(rack,  timely_says, log_mult,
4033 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4034 		}
4035 	}
4036 	return (timely_says);
4037 }
4038 
4039 static void
4040 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4041 			    tcp_seq th_ack, int line, uint8_t quality)
4042 {
4043 	uint64_t tim, bytes_ps, ltim, stim, utim;
4044 	uint32_t segsiz, bytes, reqbytes, us_cts;
4045 	int32_t gput, new_rtt_diff, timely_says;
4046 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4047 	int did_add = 0;
4048 
4049 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4050 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4051 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4052 		tim = us_cts - tp->gput_ts;
4053 	else
4054 		tim = 0;
4055 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4056 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4057 	else
4058 		stim = 0;
4059 	/*
4060 	 * Use the larger of the send time or ack time. This prevents us
4061 	 * from being influenced by ack artifacts to come up with too
4062 	 * high of measurement. Note that since we are spanning over many more
4063 	 * bytes in most of our measurements hopefully that is less likely to
4064 	 * occur.
4065 	 */
4066 	if (tim > stim)
4067 		utim = max(tim, 1);
4068 	else
4069 		utim = max(stim, 1);
4070 	/* Lets get a msec time ltim too for the old stuff */
4071 	ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4072 	gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4073 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4074 	if ((tim == 0) && (stim == 0)) {
4075 		/*
4076 		 * Invalid measurement time, maybe
4077 		 * all on one ack/one send?
4078 		 */
4079 		bytes = 0;
4080 		bytes_ps = 0;
4081 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4082 					   0, 0, 0, 10, __LINE__, NULL, quality);
4083 		goto skip_measurement;
4084 	}
4085 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4086 		/* We never made a us_rtt measurement? */
4087 		bytes = 0;
4088 		bytes_ps = 0;
4089 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4090 					   0, 0, 0, 10, __LINE__, NULL, quality);
4091 		goto skip_measurement;
4092 	}
4093 	/*
4094 	 * Calculate the maximum possible b/w this connection
4095 	 * could have. We base our calculation on the lowest
4096 	 * rtt we have seen during the measurement and the
4097 	 * largest rwnd the client has given us in that time. This
4098 	 * forms a BDP that is the maximum that we could ever
4099 	 * get to the client. Anything larger is not valid.
4100 	 *
4101 	 * I originally had code here that rejected measurements
4102 	 * where the time was less than 1/2 the latest us_rtt.
4103 	 * But after thinking on that I realized its wrong since
4104 	 * say you had a 150Mbps or even 1Gbps link, and you
4105 	 * were a long way away.. example I am in Europe (100ms rtt)
4106 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4107 	 * bytes my time would be 1.2ms, and yet my rtt would say
4108 	 * the measurement was invalid the time was < 50ms. The
4109 	 * same thing is true for 150Mb (8ms of time).
4110 	 *
4111 	 * A better way I realized is to look at what the maximum
4112 	 * the connection could possibly do. This is gated on
4113 	 * the lowest RTT we have seen and the highest rwnd.
4114 	 * We should in theory never exceed that, if we are
4115 	 * then something on the path is storing up packets
4116 	 * and then feeding them all at once to our endpoint
4117 	 * messing up our measurement.
4118 	 */
4119 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4120 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4121 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4122 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4123 		/* No measurement can be made */
4124 		bytes = 0;
4125 		bytes_ps = 0;
4126 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4127 					   0, 0, 0, 10, __LINE__, NULL, quality);
4128 		goto skip_measurement;
4129 	} else
4130 		bytes = (th_ack - tp->gput_seq);
4131 	bytes_ps = (uint64_t)bytes;
4132 	/*
4133 	 * Don't measure a b/w for pacing unless we have gotten at least
4134 	 * an initial windows worth of data in this measurement interval.
4135 	 *
4136 	 * Small numbers of bytes get badly influenced by delayed ack and
4137 	 * other artifacts. Note we take the initial window or our
4138 	 * defined minimum GP (defaulting to 10 which hopefully is the
4139 	 * IW).
4140 	 */
4141 	if (rack->rc_gp_filled == 0) {
4142 		/*
4143 		 * The initial estimate is special. We
4144 		 * have blasted out an IW worth of packets
4145 		 * without a real valid ack ts results. We
4146 		 * then setup the app_limited_needs_set flag,
4147 		 * this should get the first ack in (probably 2
4148 		 * MSS worth) to be recorded as the timestamp.
4149 		 * We thus allow a smaller number of bytes i.e.
4150 		 * IW - 2MSS.
4151 		 */
4152 		reqbytes -= (2 * segsiz);
4153 		/* Also lets fill previous for our first measurement to be neutral */
4154 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4155 	}
4156 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4157 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4158 					   rack->r_ctl.rc_app_limited_cnt,
4159 					   0, 0, 10, __LINE__, NULL, quality);
4160 		goto skip_measurement;
4161 	}
4162 	/*
4163 	 * We now need to calculate the Timely like status so
4164 	 * we can update (possibly) the b/w multipliers.
4165 	 */
4166 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4167 	if (rack->rc_gp_filled == 0) {
4168 		/* No previous reading */
4169 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4170 	} else {
4171 		if (rack->measure_saw_probe_rtt == 0) {
4172 			/*
4173 			 * We don't want a probertt to be counted
4174 			 * since it will be negative incorrectly. We
4175 			 * expect to be reducing the RTT when we
4176 			 * pace at a slower rate.
4177 			 */
4178 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4179 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4180 		}
4181 	}
4182 	timely_says = rack_make_timely_judgement(rack,
4183 		rack->r_ctl.rc_gp_srtt,
4184 		rack->r_ctl.rc_rtt_diff,
4185 	        rack->r_ctl.rc_prev_gp_srtt
4186 		);
4187 	bytes_ps *= HPTS_USEC_IN_SEC;
4188 	bytes_ps /= utim;
4189 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4190 		/*
4191 		 * Something is on path playing
4192 		 * since this b/w is not possible based
4193 		 * on our BDP (highest rwnd and lowest rtt
4194 		 * we saw in the measurement window).
4195 		 *
4196 		 * Another option here would be to
4197 		 * instead skip the measurement.
4198 		 */
4199 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4200 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4201 					   11, __LINE__, NULL, quality);
4202 		bytes_ps = rack->r_ctl.last_max_bw;
4203 	}
4204 	/* We store gp for b/w in bytes per second */
4205 	if (rack->rc_gp_filled == 0) {
4206 		/* Initial measurement */
4207 		if (bytes_ps) {
4208 			rack->r_ctl.gp_bw = bytes_ps;
4209 			rack->rc_gp_filled = 1;
4210 			rack->r_ctl.num_measurements = 1;
4211 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4212 		} else {
4213 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4214 						   rack->r_ctl.rc_app_limited_cnt,
4215 						   0, 0, 10, __LINE__, NULL, quality);
4216 		}
4217 		if (tcp_in_hpts(rack->rc_inp) &&
4218 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4219 			/*
4220 			 * Ok we can't trust the pacer in this case
4221 			 * where we transition from un-paced to paced.
4222 			 * Or for that matter when the burst mitigation
4223 			 * was making a wild guess and got it wrong.
4224 			 * Stop the pacer and clear up all the aggregate
4225 			 * delays etc.
4226 			 */
4227 			tcp_hpts_remove(rack->rc_inp);
4228 			rack->r_ctl.rc_hpts_flags = 0;
4229 			rack->r_ctl.rc_last_output_to = 0;
4230 		}
4231 		did_add = 2;
4232 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4233 		/* Still a small number run an average */
4234 		rack->r_ctl.gp_bw += bytes_ps;
4235 		addpart = rack->r_ctl.num_measurements;
4236 		rack->r_ctl.num_measurements++;
4237 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4238 			/* We have collected enought to move forward */
4239 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4240 		}
4241 		did_add = 3;
4242 	} else {
4243 		/*
4244 		 * We want to take 1/wma of the goodput and add in to 7/8th
4245 		 * of the old value weighted by the srtt. So if your measurement
4246 		 * period is say 2 SRTT's long you would get 1/4 as the
4247 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
4248 		 *
4249 		 * But we must be careful not to take too much i.e. if the
4250 		 * srtt is say 20ms and the measurement is taken over
4251 		 * 400ms our weight would be 400/20 i.e. 20. On the
4252 		 * other hand if we get a measurement over 1ms with a
4253 		 * 10ms rtt we only want to take a much smaller portion.
4254 		 */
4255 		if (rack->r_ctl.num_measurements < 0xff) {
4256 			rack->r_ctl.num_measurements++;
4257 		}
4258 		srtt = (uint64_t)tp->t_srtt;
4259 		if (srtt == 0) {
4260 			/*
4261 			 * Strange why did t_srtt go back to zero?
4262 			 */
4263 			if (rack->r_ctl.rc_rack_min_rtt)
4264 				srtt = rack->r_ctl.rc_rack_min_rtt;
4265 			else
4266 				srtt = HPTS_USEC_IN_MSEC;
4267 		}
4268 		/*
4269 		 * XXXrrs: Note for reviewers, in playing with
4270 		 * dynamic pacing I discovered this GP calculation
4271 		 * as done originally leads to some undesired results.
4272 		 * Basically you can get longer measurements contributing
4273 		 * too much to the WMA. Thus I changed it if you are doing
4274 		 * dynamic adjustments to only do the aportioned adjustment
4275 		 * if we have a very small (time wise) measurement. Longer
4276 		 * measurements just get there weight (defaulting to 1/8)
4277 		 * add to the WMA. We may want to think about changing
4278 		 * this to always do that for both sides i.e. dynamic
4279 		 * and non-dynamic... but considering lots of folks
4280 		 * were playing with this I did not want to change the
4281 		 * calculation per.se. without your thoughts.. Lawerence?
4282 		 * Peter??
4283 		 */
4284 		if (rack->rc_gp_dyn_mul == 0) {
4285 			subpart = rack->r_ctl.gp_bw * utim;
4286 			subpart /= (srtt * 8);
4287 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
4288 				/*
4289 				 * The b/w update takes no more
4290 				 * away then 1/2 our running total
4291 				 * so factor it in.
4292 				 */
4293 				addpart = bytes_ps * utim;
4294 				addpart /= (srtt * 8);
4295 			} else {
4296 				/*
4297 				 * Don't allow a single measurement
4298 				 * to account for more than 1/2 of the
4299 				 * WMA. This could happen on a retransmission
4300 				 * where utim becomes huge compared to
4301 				 * srtt (multiple retransmissions when using
4302 				 * the sending rate which factors in all the
4303 				 * transmissions from the first one).
4304 				 */
4305 				subpart = rack->r_ctl.gp_bw / 2;
4306 				addpart = bytes_ps / 2;
4307 			}
4308 			resid_bw = rack->r_ctl.gp_bw - subpart;
4309 			rack->r_ctl.gp_bw = resid_bw + addpart;
4310 			did_add = 1;
4311 		} else {
4312 			if ((utim / srtt) <= 1) {
4313 				/*
4314 				 * The b/w update was over a small period
4315 				 * of time. The idea here is to prevent a small
4316 				 * measurement time period from counting
4317 				 * too much. So we scale it based on the
4318 				 * time so it attributes less than 1/rack_wma_divisor
4319 				 * of its measurement.
4320 				 */
4321 				subpart = rack->r_ctl.gp_bw * utim;
4322 				subpart /= (srtt * rack_wma_divisor);
4323 				addpart = bytes_ps * utim;
4324 				addpart /= (srtt * rack_wma_divisor);
4325 			} else {
4326 				/*
4327 				 * The scaled measurement was long
4328 				 * enough so lets just add in the
4329 				 * portion of the measurement i.e. 1/rack_wma_divisor
4330 				 */
4331 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4332 				addpart = bytes_ps / rack_wma_divisor;
4333 			}
4334 			if ((rack->measure_saw_probe_rtt == 0) ||
4335 		            (bytes_ps > rack->r_ctl.gp_bw)) {
4336 				/*
4337 				 * For probe-rtt we only add it in
4338 				 * if its larger, all others we just
4339 				 * add in.
4340 				 */
4341 				did_add = 1;
4342 				resid_bw = rack->r_ctl.gp_bw - subpart;
4343 				rack->r_ctl.gp_bw = resid_bw + addpart;
4344 			}
4345 		}
4346 	}
4347 	if ((rack->gp_ready == 0) &&
4348 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4349 		/* We have enough measurements now */
4350 		rack->gp_ready = 1;
4351 		rack_set_cc_pacing(rack);
4352 		if (rack->defer_options)
4353 			rack_apply_deferred_options(rack);
4354 	}
4355 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4356 				   rack_get_bw(rack), 22, did_add, NULL, quality);
4357 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
4358 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4359 		rack_update_multiplier(rack, timely_says, bytes_ps,
4360 				       rack->r_ctl.rc_gp_srtt,
4361 				       rack->r_ctl.rc_rtt_diff);
4362 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4363 				   rack_get_bw(rack), 3, line, NULL, quality);
4364 	/* reset the gp srtt and setup the new prev */
4365 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4366 	/* Record the lost count for the next measurement */
4367 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4368 	/*
4369 	 * We restart our diffs based on the gpsrtt in the
4370 	 * measurement window.
4371 	 */
4372 	rack->rc_gp_rtt_set = 0;
4373 	rack->rc_gp_saw_rec = 0;
4374 	rack->rc_gp_saw_ca = 0;
4375 	rack->rc_gp_saw_ss = 0;
4376 	rack->rc_dragged_bottom = 0;
4377 skip_measurement:
4378 
4379 #ifdef STATS
4380 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4381 				 gput);
4382 	/*
4383 	 * XXXLAS: This is a temporary hack, and should be
4384 	 * chained off VOI_TCP_GPUT when stats(9) grows an
4385 	 * API to deal with chained VOIs.
4386 	 */
4387 	if (tp->t_stats_gput_prev > 0)
4388 		stats_voi_update_abs_s32(tp->t_stats,
4389 					 VOI_TCP_GPUT_ND,
4390 					 ((gput - tp->t_stats_gput_prev) * 100) /
4391 					 tp->t_stats_gput_prev);
4392 #endif
4393 	tp->t_flags &= ~TF_GPUTINPROG;
4394 	tp->t_stats_gput_prev = gput;
4395 	/*
4396 	 * Now are we app limited now and there is space from where we
4397 	 * were to where we want to go?
4398 	 *
4399 	 * We don't do the other case i.e. non-applimited here since
4400 	 * the next send will trigger us picking up the missing data.
4401 	 */
4402 	if (rack->r_ctl.rc_first_appl &&
4403 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
4404 	    rack->r_ctl.rc_app_limited_cnt &&
4405 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4406 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4407 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4408 		/*
4409 		 * Yep there is enough outstanding to make a measurement here.
4410 		 */
4411 		struct rack_sendmap *rsm, fe;
4412 
4413 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4414 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4415 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4416 		rack->app_limited_needs_set = 0;
4417 		tp->gput_seq = th_ack;
4418 		if (rack->in_probe_rtt)
4419 			rack->measure_saw_probe_rtt = 1;
4420 		else if ((rack->measure_saw_probe_rtt) &&
4421 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4422 			rack->measure_saw_probe_rtt = 0;
4423 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4424 			/* There is a full window to gain info from */
4425 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4426 		} else {
4427 			/* We can only measure up to the applimited point */
4428 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4429 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4430 				/*
4431 				 * We don't have enough to make a measurement.
4432 				 */
4433 				tp->t_flags &= ~TF_GPUTINPROG;
4434 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4435 							   0, 0, 0, 6, __LINE__, NULL, quality);
4436 				return;
4437 			}
4438 		}
4439 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
4440 			/*
4441 			 * We will get no more data into the SB
4442 			 * this means we need to have the data available
4443 			 * before we start a measurement.
4444 			 */
4445 			if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4446 				/* Nope not enough data. */
4447 				return;
4448 			}
4449 		}
4450 		tp->t_flags |= TF_GPUTINPROG;
4451 		/*
4452 		 * Now we need to find the timestamp of the send at tp->gput_seq
4453 		 * for the send based measurement.
4454 		 */
4455 		fe.r_start = tp->gput_seq;
4456 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4457 		if (rsm) {
4458 			/* Ok send-based limit is set */
4459 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4460 				/*
4461 				 * Move back to include the earlier part
4462 				 * so our ack time lines up right (this may
4463 				 * make an overlapping measurement but thats
4464 				 * ok).
4465 				 */
4466 				tp->gput_seq = rsm->r_start;
4467 			}
4468 			if (rsm->r_flags & RACK_ACKED)
4469 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4470 			else
4471 				rack->app_limited_needs_set = 1;
4472 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4473 		} else {
4474 			/*
4475 			 * If we don't find the rsm due to some
4476 			 * send-limit set the current time, which
4477 			 * basically disables the send-limit.
4478 			 */
4479 			struct timeval tv;
4480 
4481 			microuptime(&tv);
4482 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4483 		}
4484 		rack_log_pacing_delay_calc(rack,
4485 					   tp->gput_seq,
4486 					   tp->gput_ack,
4487 					   (uint64_t)rsm,
4488 					   tp->gput_ts,
4489 					   rack->r_ctl.rc_app_limited_cnt,
4490 					   9,
4491 					   __LINE__, NULL, quality);
4492 	}
4493 }
4494 
4495 /*
4496  * CC wrapper hook functions
4497  */
4498 static void
4499 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4500     uint16_t type, int32_t recovery)
4501 {
4502 	uint32_t prior_cwnd, acked;
4503 	struct tcp_log_buffer *lgb = NULL;
4504 	uint8_t labc_to_use, quality;
4505 
4506 	INP_WLOCK_ASSERT(tp->t_inpcb);
4507 	tp->ccv->nsegs = nsegs;
4508 	acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4509 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4510 		uint32_t max;
4511 
4512 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4513 		if (tp->ccv->bytes_this_ack > max) {
4514 			tp->ccv->bytes_this_ack = max;
4515 		}
4516 	}
4517 #ifdef STATS
4518 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4519 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4520 #endif
4521 	quality = RACK_QUALITY_NONE;
4522 	if ((tp->t_flags & TF_GPUTINPROG) &&
4523 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4524 		/* Measure the Goodput */
4525 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4526 #ifdef NETFLIX_PEAKRATE
4527 		if ((type == CC_ACK) &&
4528 		    (tp->t_maxpeakrate)) {
4529 			/*
4530 			 * We update t_peakrate_thr. This gives us roughly
4531 			 * one update per round trip time. Note
4532 			 * it will only be used if pace_always is off i.e
4533 			 * we don't do this for paced flows.
4534 			 */
4535 			rack_update_peakrate_thr(tp);
4536 		}
4537 #endif
4538 	}
4539 	/* Which way our we limited, if not cwnd limited no advance in CA */
4540 	if (tp->snd_cwnd <= tp->snd_wnd)
4541 		tp->ccv->flags |= CCF_CWND_LIMITED;
4542 	else
4543 		tp->ccv->flags &= ~CCF_CWND_LIMITED;
4544 	if (tp->snd_cwnd > tp->snd_ssthresh) {
4545 		tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4546 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4547 		/* For the setting of a window past use the actual scwnd we are using */
4548 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4549 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4550 			tp->ccv->flags |= CCF_ABC_SENTAWND;
4551 		}
4552 	} else {
4553 		tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4554 		tp->t_bytes_acked = 0;
4555 	}
4556 	prior_cwnd = tp->snd_cwnd;
4557 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4558 	    (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4559 		labc_to_use = rack->rc_labc;
4560 	else
4561 		labc_to_use = rack_max_abc_post_recovery;
4562 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4563 		union tcp_log_stackspecific log;
4564 		struct timeval tv;
4565 
4566 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4567 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4568 		log.u_bbr.flex1 = th_ack;
4569 		log.u_bbr.flex2 = tp->ccv->flags;
4570 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4571 		log.u_bbr.flex4 = tp->ccv->nsegs;
4572 		log.u_bbr.flex5 = labc_to_use;
4573 		log.u_bbr.flex6 = prior_cwnd;
4574 		log.u_bbr.flex7 = V_tcp_do_newsack;
4575 		log.u_bbr.flex8 = 1;
4576 		lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4577 				     0, &log, false, NULL, NULL, 0, &tv);
4578 	}
4579 	if (CC_ALGO(tp)->ack_received != NULL) {
4580 		/* XXXLAS: Find a way to live without this */
4581 		tp->ccv->curack = th_ack;
4582 		tp->ccv->labc = labc_to_use;
4583 		tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4584 		CC_ALGO(tp)->ack_received(tp->ccv, type);
4585 	}
4586 	if (lgb) {
4587 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4588 	}
4589 	if (rack->r_must_retran) {
4590 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4591 			/*
4592 			 * We now are beyond the rxt point so lets disable
4593 			 * the flag.
4594 			 */
4595 			rack->r_ctl.rc_out_at_rto = 0;
4596 			rack->r_must_retran = 0;
4597 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4598 			/*
4599 			 * Only decrement the rc_out_at_rto if the cwnd advances
4600 			 * at least a whole segment. Otherwise next time the peer
4601 			 * acks, we won't be able to send this generaly happens
4602 			 * when we are in Congestion Avoidance.
4603 			 */
4604 			if (acked <= rack->r_ctl.rc_out_at_rto){
4605 				rack->r_ctl.rc_out_at_rto -= acked;
4606 			} else {
4607 				rack->r_ctl.rc_out_at_rto = 0;
4608 			}
4609 		}
4610 	}
4611 #ifdef STATS
4612 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4613 #endif
4614 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4615 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4616 	}
4617 #ifdef NETFLIX_PEAKRATE
4618 	/* we enforce max peak rate if it is set and we are not pacing */
4619 	if ((rack->rc_always_pace == 0) &&
4620 	    tp->t_peakrate_thr &&
4621 	    (tp->snd_cwnd > tp->t_peakrate_thr)) {
4622 		tp->snd_cwnd = tp->t_peakrate_thr;
4623 	}
4624 #endif
4625 }
4626 
4627 static void
4628 tcp_rack_partialack(struct tcpcb *tp)
4629 {
4630 	struct tcp_rack *rack;
4631 
4632 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4633 	INP_WLOCK_ASSERT(tp->t_inpcb);
4634 	/*
4635 	 * If we are doing PRR and have enough
4636 	 * room to send <or> we are pacing and prr
4637 	 * is disabled we will want to see if we
4638 	 * can send data (by setting r_wanted_output to
4639 	 * true).
4640 	 */
4641 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4642 	    rack->rack_no_prr)
4643 		rack->r_wanted_output = 1;
4644 }
4645 
4646 static void
4647 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4648 {
4649 	struct tcp_rack *rack;
4650 	uint32_t orig_cwnd;
4651 
4652 	orig_cwnd = tp->snd_cwnd;
4653 	INP_WLOCK_ASSERT(tp->t_inpcb);
4654 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4655 	/* only alert CC if we alerted when we entered */
4656 	if (CC_ALGO(tp)->post_recovery != NULL) {
4657 		tp->ccv->curack = th_ack;
4658 		CC_ALGO(tp)->post_recovery(tp->ccv);
4659 		if (tp->snd_cwnd < tp->snd_ssthresh) {
4660 			/*
4661 			 * Rack has burst control and pacing
4662 			 * so lets not set this any lower than
4663 			 * snd_ssthresh per RFC-6582 (option 2).
4664 			 */
4665 			tp->snd_cwnd = tp->snd_ssthresh;
4666 		}
4667 	}
4668 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4669 		union tcp_log_stackspecific log;
4670 		struct timeval tv;
4671 
4672 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4673 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4674 		log.u_bbr.flex1 = th_ack;
4675 		log.u_bbr.flex2 = tp->ccv->flags;
4676 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4677 		log.u_bbr.flex4 = tp->ccv->nsegs;
4678 		log.u_bbr.flex5 = V_tcp_abc_l_var;
4679 		log.u_bbr.flex6 = orig_cwnd;
4680 		log.u_bbr.flex7 = V_tcp_do_newsack;
4681 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4682 		log.u_bbr.flex8 = 2;
4683 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4684 			       0, &log, false, NULL, NULL, 0, &tv);
4685 	}
4686 	if ((rack->rack_no_prr == 0) &&
4687 	    (rack->no_prr_addback == 0) &&
4688 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
4689 		/*
4690 		 * Suck the next prr cnt back into cwnd, but
4691 		 * only do that if we are not application limited.
4692 		 */
4693 		if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4694 			/*
4695 			 * We are allowed to add back to the cwnd the amount we did
4696 			 * not get out if:
4697 			 * a) no_prr_addback is off.
4698 			 * b) we are not app limited
4699 			 * c) we are doing prr
4700 			 * <and>
4701 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4702 			 */
4703 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4704 					    rack->r_ctl.rc_prr_sndcnt);
4705 		}
4706 		rack->r_ctl.rc_prr_sndcnt = 0;
4707 		rack_log_to_prr(rack, 1, 0, __LINE__);
4708 	}
4709 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
4710 	tp->snd_recover = tp->snd_una;
4711 	if (rack->r_ctl.dsack_persist) {
4712 		rack->r_ctl.dsack_persist--;
4713 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4714 			rack->r_ctl.num_dsack = 0;
4715 		}
4716 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4717 	}
4718 	EXIT_RECOVERY(tp->t_flags);
4719 }
4720 
4721 static void
4722 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
4723 {
4724 	struct tcp_rack *rack;
4725 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4726 
4727 	INP_WLOCK_ASSERT(tp->t_inpcb);
4728 #ifdef STATS
4729 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4730 #endif
4731 	if (IN_RECOVERY(tp->t_flags) == 0) {
4732 		in_rec_at_entry = 0;
4733 		ssthresh_enter = tp->snd_ssthresh;
4734 		cwnd_enter = tp->snd_cwnd;
4735 	} else
4736 		in_rec_at_entry = 1;
4737 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4738 	switch (type) {
4739 	case CC_NDUPACK:
4740 		tp->t_flags &= ~TF_WASFRECOVERY;
4741 		tp->t_flags &= ~TF_WASCRECOVERY;
4742 		if (!IN_FASTRECOVERY(tp->t_flags)) {
4743 			rack->r_ctl.rc_prr_delivered = 0;
4744 			rack->r_ctl.rc_prr_out = 0;
4745 			if (rack->rack_no_prr == 0) {
4746 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4747 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
4748 			}
4749 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4750 			tp->snd_recover = tp->snd_max;
4751 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4752 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4753 		}
4754 		break;
4755 	case CC_ECN:
4756 		if (!IN_CONGRECOVERY(tp->t_flags) ||
4757 		    /*
4758 		     * Allow ECN reaction on ACK to CWR, if
4759 		     * that data segment was also CE marked.
4760 		     */
4761 		    SEQ_GEQ(ack, tp->snd_recover)) {
4762 			EXIT_CONGRECOVERY(tp->t_flags);
4763 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4764 			tp->snd_recover = tp->snd_max + 1;
4765 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4766 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4767 		}
4768 		break;
4769 	case CC_RTO:
4770 		tp->t_dupacks = 0;
4771 		tp->t_bytes_acked = 0;
4772 		EXIT_RECOVERY(tp->t_flags);
4773 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4774 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4775 		orig_cwnd = tp->snd_cwnd;
4776 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
4777 		rack_log_to_prr(rack, 16, orig_cwnd, line);
4778 		if (tp->t_flags2 & TF2_ECN_PERMIT)
4779 			tp->t_flags2 |= TF2_ECN_SND_CWR;
4780 		break;
4781 	case CC_RTO_ERR:
4782 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4783 		/* RTO was unnecessary, so reset everything. */
4784 		tp->snd_cwnd = tp->snd_cwnd_prev;
4785 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
4786 		tp->snd_recover = tp->snd_recover_prev;
4787 		if (tp->t_flags & TF_WASFRECOVERY) {
4788 			ENTER_FASTRECOVERY(tp->t_flags);
4789 			tp->t_flags &= ~TF_WASFRECOVERY;
4790 		}
4791 		if (tp->t_flags & TF_WASCRECOVERY) {
4792 			ENTER_CONGRECOVERY(tp->t_flags);
4793 			tp->t_flags &= ~TF_WASCRECOVERY;
4794 		}
4795 		tp->snd_nxt = tp->snd_max;
4796 		tp->t_badrxtwin = 0;
4797 		break;
4798 	}
4799 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4800 	    (type != CC_RTO)){
4801 		tp->ccv->curack = ack;
4802 		CC_ALGO(tp)->cong_signal(tp->ccv, type);
4803 	}
4804 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4805 		rack_log_to_prr(rack, 15, cwnd_enter, line);
4806 		rack->r_ctl.dsack_byte_cnt = 0;
4807 		rack->r_ctl.retran_during_recovery = 0;
4808 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4809 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4810 		rack->r_ent_rec_ns = 1;
4811 	}
4812 }
4813 
4814 static inline void
4815 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4816 {
4817 	uint32_t i_cwnd;
4818 
4819 	INP_WLOCK_ASSERT(tp->t_inpcb);
4820 
4821 #ifdef NETFLIX_STATS
4822 	KMOD_TCPSTAT_INC(tcps_idle_restarts);
4823 	if (tp->t_state == TCPS_ESTABLISHED)
4824 		KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4825 #endif
4826 	if (CC_ALGO(tp)->after_idle != NULL)
4827 		CC_ALGO(tp)->after_idle(tp->ccv);
4828 
4829 	if (tp->snd_cwnd == 1)
4830 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
4831 	else
4832 		i_cwnd = rc_init_window(rack);
4833 
4834 	/*
4835 	 * Being idle is no different than the initial window. If the cc
4836 	 * clamps it down below the initial window raise it to the initial
4837 	 * window.
4838 	 */
4839 	if (tp->snd_cwnd < i_cwnd) {
4840 		tp->snd_cwnd = i_cwnd;
4841 	}
4842 }
4843 
4844 /*
4845  * Indicate whether this ack should be delayed.  We can delay the ack if
4846  * following conditions are met:
4847  *	- There is no delayed ack timer in progress.
4848  *	- Our last ack wasn't a 0-sized window. We never want to delay
4849  *	  the ack that opens up a 0-sized window.
4850  *	- LRO wasn't used for this segment. We make sure by checking that the
4851  *	  segment size is not larger than the MSS.
4852  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
4853  *	  connection.
4854  */
4855 #define DELAY_ACK(tp, tlen)			 \
4856 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4857 	((tp->t_flags & TF_DELACK) == 0) &&	 \
4858 	(tlen <= tp->t_maxseg) &&		 \
4859 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4860 
4861 static struct rack_sendmap *
4862 rack_find_lowest_rsm(struct tcp_rack *rack)
4863 {
4864 	struct rack_sendmap *rsm;
4865 
4866 	/*
4867 	 * Walk the time-order transmitted list looking for an rsm that is
4868 	 * not acked. This will be the one that was sent the longest time
4869 	 * ago that is still outstanding.
4870 	 */
4871 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4872 		if (rsm->r_flags & RACK_ACKED) {
4873 			continue;
4874 		}
4875 		goto finish;
4876 	}
4877 finish:
4878 	return (rsm);
4879 }
4880 
4881 static struct rack_sendmap *
4882 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4883 {
4884 	struct rack_sendmap *prsm;
4885 
4886 	/*
4887 	 * Walk the sequence order list backward until we hit and arrive at
4888 	 * the highest seq not acked. In theory when this is called it
4889 	 * should be the last segment (which it was not).
4890 	 */
4891 	prsm = rsm;
4892 	RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4893 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4894 			continue;
4895 		}
4896 		return (prsm);
4897 	}
4898 	return (NULL);
4899 }
4900 
4901 static uint32_t
4902 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4903 {
4904 	int32_t lro;
4905 	uint32_t thresh;
4906 
4907 	/*
4908 	 * lro is the flag we use to determine if we have seen reordering.
4909 	 * If it gets set we have seen reordering. The reorder logic either
4910 	 * works in one of two ways:
4911 	 *
4912 	 * If reorder-fade is configured, then we track the last time we saw
4913 	 * re-ordering occur. If we reach the point where enough time as
4914 	 * passed we no longer consider reordering has occuring.
4915 	 *
4916 	 * Or if reorder-face is 0, then once we see reordering we consider
4917 	 * the connection to alway be subject to reordering and just set lro
4918 	 * to 1.
4919 	 *
4920 	 * In the end if lro is non-zero we add the extra time for
4921 	 * reordering in.
4922 	 */
4923 	if (srtt == 0)
4924 		srtt = 1;
4925 	if (rack->r_ctl.rc_reorder_ts) {
4926 		if (rack->r_ctl.rc_reorder_fade) {
4927 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
4928 				lro = cts - rack->r_ctl.rc_reorder_ts;
4929 				if (lro == 0) {
4930 					/*
4931 					 * No time as passed since the last
4932 					 * reorder, mark it as reordering.
4933 					 */
4934 					lro = 1;
4935 				}
4936 			} else {
4937 				/* Negative time? */
4938 				lro = 0;
4939 			}
4940 			if (lro > rack->r_ctl.rc_reorder_fade) {
4941 				/* Turn off reordering seen too */
4942 				rack->r_ctl.rc_reorder_ts = 0;
4943 				lro = 0;
4944 			}
4945 		} else {
4946 			/* Reodering does not fade */
4947 			lro = 1;
4948 		}
4949 	} else {
4950 		lro = 0;
4951 	}
4952 	if (rack->rc_rack_tmr_std_based == 0) {
4953 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
4954 	} else {
4955 		/* Standards based pkt-delay is 1/4 srtt */
4956 		thresh = srtt +  (srtt >> 2);
4957 	}
4958 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
4959 		/* It must be set, if not you get 1/4 rtt */
4960 		if (rack->r_ctl.rc_reorder_shift)
4961 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
4962 		else
4963 			thresh += (srtt >> 2);
4964 	}
4965 	if (rack->rc_rack_use_dsack &&
4966 	    lro &&
4967 	    (rack->r_ctl.num_dsack > 0)) {
4968 		/*
4969 		 * We only increase the reordering window if we
4970 		 * have seen reordering <and> we have a DSACK count.
4971 		 */
4972 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
4973 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
4974 	}
4975 	/* SRTT * 2 is the ceiling */
4976 	if (thresh > (srtt * 2)) {
4977 		thresh = srtt * 2;
4978 	}
4979 	/* And we don't want it above the RTO max either */
4980 	if (thresh > rack_rto_max) {
4981 		thresh = rack_rto_max;
4982 	}
4983 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
4984 	return (thresh);
4985 }
4986 
4987 static uint32_t
4988 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
4989 		     struct rack_sendmap *rsm, uint32_t srtt)
4990 {
4991 	struct rack_sendmap *prsm;
4992 	uint32_t thresh, len;
4993 	int segsiz;
4994 
4995 	if (srtt == 0)
4996 		srtt = 1;
4997 	if (rack->r_ctl.rc_tlp_threshold)
4998 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
4999 	else
5000 		thresh = (srtt * 2);
5001 
5002 	/* Get the previous sent packet, if any */
5003 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5004 	len = rsm->r_end - rsm->r_start;
5005 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5006 		/* Exactly like the ID */
5007 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5008 			uint32_t alt_thresh;
5009 			/*
5010 			 * Compensate for delayed-ack with the d-ack time.
5011 			 */
5012 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5013 			if (alt_thresh > thresh)
5014 				thresh = alt_thresh;
5015 		}
5016 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5017 		/* 2.1 behavior */
5018 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5019 		if (prsm && (len <= segsiz)) {
5020 			/*
5021 			 * Two packets outstanding, thresh should be (2*srtt) +
5022 			 * possible inter-packet delay (if any).
5023 			 */
5024 			uint32_t inter_gap = 0;
5025 			int idx, nidx;
5026 
5027 			idx = rsm->r_rtr_cnt - 1;
5028 			nidx = prsm->r_rtr_cnt - 1;
5029 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5030 				/* Yes it was sent later (or at the same time) */
5031 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5032 			}
5033 			thresh += inter_gap;
5034 		} else if (len <= segsiz) {
5035 			/*
5036 			 * Possibly compensate for delayed-ack.
5037 			 */
5038 			uint32_t alt_thresh;
5039 
5040 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5041 			if (alt_thresh > thresh)
5042 				thresh = alt_thresh;
5043 		}
5044 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5045 		/* 2.2 behavior */
5046 		if (len <= segsiz) {
5047 			uint32_t alt_thresh;
5048 			/*
5049 			 * Compensate for delayed-ack with the d-ack time.
5050 			 */
5051 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5052 			if (alt_thresh > thresh)
5053 				thresh = alt_thresh;
5054 		}
5055 	}
5056 	/* Not above an RTO */
5057 	if (thresh > tp->t_rxtcur) {
5058 		thresh = tp->t_rxtcur;
5059 	}
5060 	/* Not above a RTO max */
5061 	if (thresh > rack_rto_max) {
5062 		thresh = rack_rto_max;
5063 	}
5064 	/* Apply user supplied min TLP */
5065 	if (thresh < rack_tlp_min) {
5066 		thresh = rack_tlp_min;
5067 	}
5068 	return (thresh);
5069 }
5070 
5071 static uint32_t
5072 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5073 {
5074 	/*
5075 	 * We want the rack_rtt which is the
5076 	 * last rtt we measured. However if that
5077 	 * does not exist we fallback to the srtt (which
5078 	 * we probably will never do) and then as a last
5079 	 * resort we use RACK_INITIAL_RTO if no srtt is
5080 	 * yet set.
5081 	 */
5082 	if (rack->rc_rack_rtt)
5083 		return (rack->rc_rack_rtt);
5084 	else if (tp->t_srtt == 0)
5085 		return (RACK_INITIAL_RTO);
5086 	return (tp->t_srtt);
5087 }
5088 
5089 static struct rack_sendmap *
5090 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5091 {
5092 	/*
5093 	 * Check to see that we don't need to fall into recovery. We will
5094 	 * need to do so if our oldest transmit is past the time we should
5095 	 * have had an ack.
5096 	 */
5097 	struct tcp_rack *rack;
5098 	struct rack_sendmap *rsm;
5099 	int32_t idx;
5100 	uint32_t srtt, thresh;
5101 
5102 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5103 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5104 		return (NULL);
5105 	}
5106 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5107 	if (rsm == NULL)
5108 		return (NULL);
5109 
5110 
5111 	if (rsm->r_flags & RACK_ACKED) {
5112 		rsm = rack_find_lowest_rsm(rack);
5113 		if (rsm == NULL)
5114 			return (NULL);
5115 	}
5116 	idx = rsm->r_rtr_cnt - 1;
5117 	srtt = rack_grab_rtt(tp, rack);
5118 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5119 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5120 		return (NULL);
5121 	}
5122 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5123 		return (NULL);
5124 	}
5125 	/* Ok if we reach here we are over-due and this guy can be sent */
5126 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
5127 	return (rsm);
5128 }
5129 
5130 static uint32_t
5131 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5132 {
5133 	int32_t t;
5134 	int32_t tt;
5135 	uint32_t ret_val;
5136 
5137 	t = (tp->t_srtt + (tp->t_rttvar << 2));
5138 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5139  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5140 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5141 	ret_val = (uint32_t)tt;
5142 	return (ret_val);
5143 }
5144 
5145 static uint32_t
5146 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5147 {
5148 	/*
5149 	 * Start the FR timer, we do this based on getting the first one in
5150 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
5151 	 * events we need to stop the running timer (if its running) before
5152 	 * starting the new one.
5153 	 */
5154 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5155 	uint32_t srtt_cur;
5156 	int32_t idx;
5157 	int32_t is_tlp_timer = 0;
5158 	struct rack_sendmap *rsm;
5159 
5160 	if (rack->t_timers_stopped) {
5161 		/* All timers have been stopped none are to run */
5162 		return (0);
5163 	}
5164 	if (rack->rc_in_persist) {
5165 		/* We can't start any timer in persists */
5166 		return (rack_get_persists_timer_val(tp, rack));
5167 	}
5168 	rack->rc_on_min_to = 0;
5169 	if ((tp->t_state < TCPS_ESTABLISHED) ||
5170 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5171 		goto activate_rxt;
5172 	}
5173 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5174 	if ((rsm == NULL) || sup_rack) {
5175 		/* Nothing on the send map or no rack */
5176 activate_rxt:
5177 		time_since_sent = 0;
5178 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5179 		if (rsm) {
5180 			/*
5181 			 * Should we discount the RTX timer any?
5182 			 *
5183 			 * We want to discount it the smallest amount.
5184 			 * If a timer (Rack/TLP or RXT) has gone off more
5185 			 * recently thats the discount we want to use (now - timer time).
5186 			 * If the retransmit of the oldest packet was more recent then
5187 			 * we want to use that (now - oldest-packet-last_transmit_time).
5188 			 *
5189 			 */
5190 			idx = rsm->r_rtr_cnt - 1;
5191 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5192 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5193 			else
5194 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5195 			if (TSTMP_GT(cts, tstmp_touse))
5196 			    time_since_sent = cts - tstmp_touse;
5197 		}
5198 		if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5199 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5200 			to = tp->t_rxtcur;
5201 			if (to > time_since_sent)
5202 				to -= time_since_sent;
5203 			else
5204 				to = rack->r_ctl.rc_min_to;
5205 			if (to == 0)
5206 				to = 1;
5207 			/* Special case for KEEPINIT */
5208 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5209 			    (TP_KEEPINIT(tp) != 0) &&
5210 			    rsm) {
5211 				/*
5212 				 * We have to put a ceiling on the rxt timer
5213 				 * of the keep-init timeout.
5214 				 */
5215 				uint32_t max_time, red;
5216 
5217 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5218 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5219 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5220 					if (red < max_time)
5221 						max_time -= red;
5222 					else
5223 						max_time = 1;
5224 				}
5225 				/* Reduce timeout to the keep value if needed */
5226 				if (max_time < to)
5227 					to = max_time;
5228 			}
5229 			return (to);
5230 		}
5231 		return (0);
5232 	}
5233 	if (rsm->r_flags & RACK_ACKED) {
5234 		rsm = rack_find_lowest_rsm(rack);
5235 		if (rsm == NULL) {
5236 			/* No lowest? */
5237 			goto activate_rxt;
5238 		}
5239 	}
5240 	if (rack->sack_attack_disable) {
5241 		/*
5242 		 * We don't want to do
5243 		 * any TLP's if you are an attacker.
5244 		 * Though if you are doing what
5245 		 * is expected you may still have
5246 		 * SACK-PASSED marks.
5247 		 */
5248 		goto activate_rxt;
5249 	}
5250 	/* Convert from ms to usecs */
5251 	if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5252 		if ((tp->t_flags & TF_SENTFIN) &&
5253 		    ((tp->snd_max - tp->snd_una) == 1) &&
5254 		    (rsm->r_flags & RACK_HAS_FIN)) {
5255 			/*
5256 			 * We don't start a rack timer if all we have is a
5257 			 * FIN outstanding.
5258 			 */
5259 			goto activate_rxt;
5260 		}
5261 		if ((rack->use_rack_rr == 0) &&
5262 		    (IN_FASTRECOVERY(tp->t_flags)) &&
5263 		    (rack->rack_no_prr == 0) &&
5264 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5265 			/*
5266 			 * We are not cheating, in recovery  and
5267 			 * not enough ack's to yet get our next
5268 			 * retransmission out.
5269 			 *
5270 			 * Note that classified attackers do not
5271 			 * get to use the rack-cheat.
5272 			 */
5273 			goto activate_tlp;
5274 		}
5275 		srtt = rack_grab_rtt(tp, rack);
5276 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
5277 		idx = rsm->r_rtr_cnt - 1;
5278 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5279 		if (SEQ_GEQ(exp, cts)) {
5280 			to = exp - cts;
5281 			if (to < rack->r_ctl.rc_min_to) {
5282 				to = rack->r_ctl.rc_min_to;
5283 				if (rack->r_rr_config == 3)
5284 					rack->rc_on_min_to = 1;
5285 			}
5286 		} else {
5287 			to = rack->r_ctl.rc_min_to;
5288 			if (rack->r_rr_config == 3)
5289 				rack->rc_on_min_to = 1;
5290 		}
5291 	} else {
5292 		/* Ok we need to do a TLP not RACK */
5293 activate_tlp:
5294 		if ((rack->rc_tlp_in_progress != 0) &&
5295 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5296 			/*
5297 			 * The previous send was a TLP and we have sent
5298 			 * N TLP's without sending new data.
5299 			 */
5300 			goto activate_rxt;
5301 		}
5302 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5303 		if (rsm == NULL) {
5304 			/* We found no rsm to TLP with. */
5305 			goto activate_rxt;
5306 		}
5307 		if (rsm->r_flags & RACK_HAS_FIN) {
5308 			/* If its a FIN we dont do TLP */
5309 			rsm = NULL;
5310 			goto activate_rxt;
5311 		}
5312 		idx = rsm->r_rtr_cnt - 1;
5313 		time_since_sent = 0;
5314 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5315 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5316 		else
5317 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5318 		if (TSTMP_GT(cts, tstmp_touse))
5319 		    time_since_sent = cts - tstmp_touse;
5320 		is_tlp_timer = 1;
5321 		if (tp->t_srtt) {
5322 			if ((rack->rc_srtt_measure_made == 0) &&
5323 			    (tp->t_srtt == 1)) {
5324 				/*
5325 				 * If another stack as run and set srtt to 1,
5326 				 * then the srtt was 0, so lets use the initial.
5327 				 */
5328 				srtt = RACK_INITIAL_RTO;
5329 			} else {
5330 				srtt_cur = tp->t_srtt;
5331 				srtt = srtt_cur;
5332 			}
5333 		} else
5334 			srtt = RACK_INITIAL_RTO;
5335 		/*
5336 		 * If the SRTT is not keeping up and the
5337 		 * rack RTT has spiked we want to use
5338 		 * the last RTT not the smoothed one.
5339 		 */
5340 		if (rack_tlp_use_greater &&
5341 		    tp->t_srtt &&
5342 		    (srtt < rack_grab_rtt(tp, rack))) {
5343 			srtt = rack_grab_rtt(tp, rack);
5344 		}
5345 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5346 		if (thresh > time_since_sent) {
5347 			to = thresh - time_since_sent;
5348 		} else {
5349 			to = rack->r_ctl.rc_min_to;
5350 			rack_log_alt_to_to_cancel(rack,
5351 						  thresh,		/* flex1 */
5352 						  time_since_sent,	/* flex2 */
5353 						  tstmp_touse,		/* flex3 */
5354 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5355 						  (uint32_t)rsm->r_tim_lastsent[idx],
5356 						  srtt,
5357 						  idx, 99);
5358 		}
5359 		if (to < rack_tlp_min) {
5360 			to = rack_tlp_min;
5361 		}
5362 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5363 			/*
5364 			 * If the TLP time works out to larger than the max
5365 			 * RTO lets not do TLP.. just RTO.
5366 			 */
5367 			goto activate_rxt;
5368 		}
5369 	}
5370 	if (is_tlp_timer == 0) {
5371 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5372 	} else {
5373 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5374 	}
5375 	if (to == 0)
5376 		to = 1;
5377 	return (to);
5378 }
5379 
5380 static void
5381 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5382 {
5383 	if (rack->rc_in_persist == 0) {
5384 		if (tp->t_flags & TF_GPUTINPROG) {
5385 			/*
5386 			 * Stop the goodput now, the calling of the
5387 			 * measurement function clears the flag.
5388 			 */
5389 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5390 						    RACK_QUALITY_PERSIST);
5391 		}
5392 #ifdef NETFLIX_SHARED_CWND
5393 		if (rack->r_ctl.rc_scw) {
5394 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5395 			rack->rack_scwnd_is_idle = 1;
5396 		}
5397 #endif
5398 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5399 		if (rack->r_ctl.rc_went_idle_time == 0)
5400 			rack->r_ctl.rc_went_idle_time = 1;
5401 		rack_timer_cancel(tp, rack, cts, __LINE__);
5402 		rack->r_ctl.persist_lost_ends = 0;
5403 		rack->probe_not_answered = 0;
5404 		rack->forced_ack = 0;
5405 		tp->t_rxtshift = 0;
5406 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5407 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5408 		rack->rc_in_persist = 1;
5409 	}
5410 }
5411 
5412 static void
5413 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5414 {
5415 	if (tcp_in_hpts(rack->rc_inp)) {
5416 		tcp_hpts_remove(rack->rc_inp);
5417 		rack->r_ctl.rc_hpts_flags = 0;
5418 	}
5419 #ifdef NETFLIX_SHARED_CWND
5420 	if (rack->r_ctl.rc_scw) {
5421 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5422 		rack->rack_scwnd_is_idle = 0;
5423 	}
5424 #endif
5425 	if (rack->rc_gp_dyn_mul &&
5426 	    (rack->use_fixed_rate == 0) &&
5427 	    (rack->rc_always_pace)) {
5428 		/*
5429 		 * Do we count this as if a probe-rtt just
5430 		 * finished?
5431 		 */
5432 		uint32_t time_idle, idle_min;
5433 
5434 		time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5435 		idle_min = rack_min_probertt_hold;
5436 		if (rack_probertt_gpsrtt_cnt_div) {
5437 			uint64_t extra;
5438 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5439 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
5440 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5441 			idle_min += (uint32_t)extra;
5442 		}
5443 		if (time_idle >= idle_min) {
5444 			/* Yes, we count it as a probe-rtt. */
5445 			uint32_t us_cts;
5446 
5447 			us_cts = tcp_get_usecs(NULL);
5448 			if (rack->in_probe_rtt == 0) {
5449 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5450 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5451 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5452 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5453 			} else {
5454 				rack_exit_probertt(rack, us_cts);
5455 			}
5456 		}
5457 	}
5458 	rack->rc_in_persist = 0;
5459 	rack->r_ctl.rc_went_idle_time = 0;
5460 	tp->t_rxtshift = 0;
5461 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5462 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5463 	rack->r_ctl.rc_agg_delayed = 0;
5464 	rack->r_early = 0;
5465 	rack->r_late = 0;
5466 	rack->r_ctl.rc_agg_early = 0;
5467 }
5468 
5469 static void
5470 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5471 		   struct hpts_diag *diag, struct timeval *tv)
5472 {
5473 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5474 		union tcp_log_stackspecific log;
5475 
5476 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5477 		log.u_bbr.flex1 = diag->p_nxt_slot;
5478 		log.u_bbr.flex2 = diag->p_cur_slot;
5479 		log.u_bbr.flex3 = diag->slot_req;
5480 		log.u_bbr.flex4 = diag->inp_hptsslot;
5481 		log.u_bbr.flex5 = diag->slot_remaining;
5482 		log.u_bbr.flex6 = diag->need_new_to;
5483 		log.u_bbr.flex7 = diag->p_hpts_active;
5484 		log.u_bbr.flex8 = diag->p_on_min_sleep;
5485 		/* Hijack other fields as needed */
5486 		log.u_bbr.epoch = diag->have_slept;
5487 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
5488 		log.u_bbr.pkts_out = diag->co_ret;
5489 		log.u_bbr.applimited = diag->hpts_sleep_time;
5490 		log.u_bbr.delivered = diag->p_prev_slot;
5491 		log.u_bbr.inflight = diag->p_runningslot;
5492 		log.u_bbr.bw_inuse = diag->wheel_slot;
5493 		log.u_bbr.rttProp = diag->wheel_cts;
5494 		log.u_bbr.timeStamp = cts;
5495 		log.u_bbr.delRate = diag->maxslots;
5496 		log.u_bbr.cur_del_rate = diag->p_curtick;
5497 		log.u_bbr.cur_del_rate <<= 32;
5498 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
5499 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5500 		    &rack->rc_inp->inp_socket->so_rcv,
5501 		    &rack->rc_inp->inp_socket->so_snd,
5502 		    BBR_LOG_HPTSDIAG, 0,
5503 		    0, &log, false, tv);
5504 	}
5505 
5506 }
5507 
5508 static void
5509 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5510 {
5511 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5512 		union tcp_log_stackspecific log;
5513 		struct timeval tv;
5514 
5515 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5516 		log.u_bbr.flex1 = sb->sb_flags;
5517 		log.u_bbr.flex2 = len;
5518 		log.u_bbr.flex3 = sb->sb_state;
5519 		log.u_bbr.flex8 = type;
5520 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5521 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5522 		    &rack->rc_inp->inp_socket->so_rcv,
5523 		    &rack->rc_inp->inp_socket->so_snd,
5524 		    TCP_LOG_SB_WAKE, 0,
5525 		    len, &log, false, &tv);
5526 	}
5527 }
5528 
5529 static void
5530 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5531       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5532 {
5533 	struct hpts_diag diag;
5534 	struct inpcb *inp;
5535 	struct timeval tv;
5536 	uint32_t delayed_ack = 0;
5537 	uint32_t hpts_timeout;
5538 	uint32_t entry_slot = slot;
5539 	uint8_t stopped;
5540 	uint32_t left = 0;
5541 	uint32_t us_cts;
5542 
5543 	inp = tp->t_inpcb;
5544 	if ((tp->t_state == TCPS_CLOSED) ||
5545 	    (tp->t_state == TCPS_LISTEN)) {
5546 		return;
5547 	}
5548 	if (tcp_in_hpts(inp)) {
5549 		/* Already on the pacer */
5550 		return;
5551 	}
5552 	stopped = rack->rc_tmr_stopped;
5553 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5554 		left = rack->r_ctl.rc_timer_exp - cts;
5555 	}
5556 	rack->r_ctl.rc_timer_exp = 0;
5557 	rack->r_ctl.rc_hpts_flags = 0;
5558 	us_cts = tcp_get_usecs(&tv);
5559 	/* Now early/late accounting */
5560 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5561 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5562 		/*
5563 		 * We have a early carry over set,
5564 		 * we can always add more time so we
5565 		 * can always make this compensation.
5566 		 *
5567 		 * Note if ack's are allowed to wake us do not
5568 		 * penalize the next timer for being awoke
5569 		 * by an ack aka the rc_agg_early (non-paced mode).
5570 		 */
5571 		slot += rack->r_ctl.rc_agg_early;
5572 		rack->r_early = 0;
5573 		rack->r_ctl.rc_agg_early = 0;
5574 	}
5575 	if (rack->r_late) {
5576 		/*
5577 		 * This is harder, we can
5578 		 * compensate some but it
5579 		 * really depends on what
5580 		 * the current pacing time is.
5581 		 */
5582 		if (rack->r_ctl.rc_agg_delayed >= slot) {
5583 			/*
5584 			 * We can't compensate for it all.
5585 			 * And we have to have some time
5586 			 * on the clock. We always have a min
5587 			 * 10 slots (10 x 10 i.e. 100 usecs).
5588 			 */
5589 			if (slot <= HPTS_TICKS_PER_SLOT) {
5590 				/* We gain delay */
5591 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5592 				slot = HPTS_TICKS_PER_SLOT;
5593 			} else {
5594 				/* We take off some */
5595 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5596 				slot = HPTS_TICKS_PER_SLOT;
5597 			}
5598 		} else {
5599 			slot -= rack->r_ctl.rc_agg_delayed;
5600 			rack->r_ctl.rc_agg_delayed = 0;
5601 			/* Make sure we have 100 useconds at minimum */
5602 			if (slot < HPTS_TICKS_PER_SLOT) {
5603 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5604 				slot = HPTS_TICKS_PER_SLOT;
5605 			}
5606 			if (rack->r_ctl.rc_agg_delayed == 0)
5607 				rack->r_late = 0;
5608 		}
5609 	}
5610 	if (slot) {
5611 		/* We are pacing too */
5612 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5613 	}
5614 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5615 #ifdef NETFLIX_EXP_DETECTION
5616 	if (rack->sack_attack_disable &&
5617 	    (slot < tcp_sad_pacing_interval)) {
5618 		/*
5619 		 * We have a potential attacker on
5620 		 * the line. We have possibly some
5621 		 * (or now) pacing time set. We want to
5622 		 * slow down the processing of sacks by some
5623 		 * amount (if it is an attacker). Set the default
5624 		 * slot for attackers in place (unless the orginal
5625 		 * interval is longer). Its stored in
5626 		 * micro-seconds, so lets convert to msecs.
5627 		 */
5628 		slot = tcp_sad_pacing_interval;
5629 	}
5630 #endif
5631 	if (tp->t_flags & TF_DELACK) {
5632 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
5633 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5634 	}
5635 	if (delayed_ack && ((hpts_timeout == 0) ||
5636 			    (delayed_ack < hpts_timeout)))
5637 		hpts_timeout = delayed_ack;
5638 	else
5639 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5640 	/*
5641 	 * If no timers are going to run and we will fall off the hptsi
5642 	 * wheel, we resort to a keep-alive timer if its configured.
5643 	 */
5644 	if ((hpts_timeout == 0) &&
5645 	    (slot == 0)) {
5646 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5647 		    (tp->t_state <= TCPS_CLOSING)) {
5648 			/*
5649 			 * Ok we have no timer (persists, rack, tlp, rxt  or
5650 			 * del-ack), we don't have segments being paced. So
5651 			 * all that is left is the keepalive timer.
5652 			 */
5653 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5654 				/* Get the established keep-alive time */
5655 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5656 			} else {
5657 				/*
5658 				 * Get the initial setup keep-alive time,
5659 				 * note that this is probably not going to
5660 				 * happen, since rack will be running a rxt timer
5661 				 * if a SYN of some sort is outstanding. It is
5662 				 * actually handled in rack_timeout_rxt().
5663 				 */
5664 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5665 			}
5666 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5667 			if (rack->in_probe_rtt) {
5668 				/*
5669 				 * We want to instead not wake up a long time from
5670 				 * now but to wake up about the time we would
5671 				 * exit probe-rtt and initiate a keep-alive ack.
5672 				 * This will get us out of probe-rtt and update
5673 				 * our min-rtt.
5674 				 */
5675 				hpts_timeout = rack_min_probertt_hold;
5676 			}
5677 		}
5678 	}
5679 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5680 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5681 		/*
5682 		 * RACK, TLP, persists and RXT timers all are restartable
5683 		 * based on actions input .. i.e we received a packet (ack
5684 		 * or sack) and that changes things (rw, or snd_una etc).
5685 		 * Thus we can restart them with a new value. For
5686 		 * keep-alive, delayed_ack we keep track of what was left
5687 		 * and restart the timer with a smaller value.
5688 		 */
5689 		if (left < hpts_timeout)
5690 			hpts_timeout = left;
5691 	}
5692 	if (hpts_timeout) {
5693 		/*
5694 		 * Hack alert for now we can't time-out over 2,147,483
5695 		 * seconds (a bit more than 596 hours), which is probably ok
5696 		 * :).
5697 		 */
5698 		if (hpts_timeout > 0x7ffffffe)
5699 			hpts_timeout = 0x7ffffffe;
5700 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5701 	}
5702 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5703 	if ((rack->gp_ready == 0) &&
5704 	    (rack->use_fixed_rate == 0) &&
5705 	    (hpts_timeout < slot) &&
5706 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5707 		/*
5708 		 * We have no good estimate yet for the
5709 		 * old clunky burst mitigation or the
5710 		 * real pacing. And the tlp or rxt is smaller
5711 		 * than the pacing calculation. Lets not
5712 		 * pace that long since we know the calculation
5713 		 * so far is not accurate.
5714 		 */
5715 		slot = hpts_timeout;
5716 	}
5717 	/**
5718 	 * Turn off all the flags for queuing by default. The
5719 	 * flags have important meanings to what happens when
5720 	 * LRO interacts with the transport. Most likely (by default now)
5721 	 * mbuf_queueing and ack compression are on. So the transport
5722 	 * has a couple of flags that control what happens (if those
5723 	 * are not on then these flags won't have any effect since it
5724 	 * won't go through the queuing LRO path).
5725 	 *
5726 	 * INP_MBUF_QUEUE_READY - This flags says that I am busy
5727 	 *                        pacing output, so don't disturb. But
5728 	 *                        it also means LRO can wake me if there
5729 	 *                        is a SACK arrival.
5730 	 *
5731 	 * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5732 	 *                       with the above flag (QUEUE_READY) and
5733 	 *                       when present it says don't even wake me
5734 	 *                       if a SACK arrives.
5735 	 *
5736 	 * The idea behind these flags is that if we are pacing we
5737 	 * set the MBUF_QUEUE_READY and only get woken up if
5738 	 * a SACK arrives (which could change things) or if
5739 	 * our pacing timer expires. If, however, we have a rack
5740 	 * timer running, then we don't even want a sack to wake
5741 	 * us since the rack timer has to expire before we can send.
5742 	 *
5743 	 * Other cases should usually have none of the flags set
5744 	 * so LRO can call into us.
5745 	 */
5746 	inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5747 	if (slot) {
5748 		rack->r_ctl.rc_last_output_to = us_cts + slot;
5749 		/*
5750 		 * A pacing timer (slot) is being set, in
5751 		 * such a case we cannot send (we are blocked by
5752 		 * the timer). So lets tell LRO that it should not
5753 		 * wake us unless there is a SACK. Note this only
5754 		 * will be effective if mbuf queueing is on or
5755 		 * compressed acks are being processed.
5756 		 */
5757 		inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5758 		/*
5759 		 * But wait if we have a Rack timer running
5760 		 * even a SACK should not disturb us (with
5761 		 * the exception of r_rr_config 3).
5762 		 */
5763 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5764 		    (rack->r_rr_config != 3))
5765 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5766 		if (rack->rc_ack_can_sendout_data) {
5767 			/*
5768 			 * Ahh but wait, this is that special case
5769 			 * where the pacing timer can be disturbed
5770 			 * backout the changes (used for non-paced
5771 			 * burst limiting).
5772 			 */
5773 			inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5774 		}
5775 		if ((rack->use_rack_rr) &&
5776 		    (rack->r_rr_config < 2) &&
5777 		    ((hpts_timeout) && (hpts_timeout < slot))) {
5778 			/*
5779 			 * Arrange for the hpts to kick back in after the
5780 			 * t-o if the t-o does not cause a send.
5781 			 */
5782 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5783 						   __LINE__, &diag);
5784 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5785 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5786 		} else {
5787 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
5788 						   __LINE__, &diag);
5789 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5790 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5791 		}
5792 	} else if (hpts_timeout) {
5793 		/*
5794 		 * With respect to inp_flags2 here, lets let any new acks wake
5795 		 * us up here. Since we are not pacing (no pacing timer), output
5796 		 * can happen so we should let it. If its a Rack timer, then any inbound
5797 		 * packet probably won't change the sending (we will be blocked)
5798 		 * but it may change the prr stats so letting it in (the set defaults
5799 		 * at the start of this block) are good enough.
5800 		 */
5801 		(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5802 					   __LINE__, &diag);
5803 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5804 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5805 	} else {
5806 		/* No timer starting */
5807 #ifdef INVARIANTS
5808 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5809 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5810 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5811 		}
5812 #endif
5813 	}
5814 	rack->rc_tmr_stopped = 0;
5815 	if (slot)
5816 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5817 }
5818 
5819 /*
5820  * RACK Timer, here we simply do logging and house keeping.
5821  * the normal rack_output() function will call the
5822  * appropriate thing to check if we need to do a RACK retransmit.
5823  * We return 1, saying don't proceed with rack_output only
5824  * when all timers have been stopped (destroyed PCB?).
5825  */
5826 static int
5827 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5828 {
5829 	/*
5830 	 * This timer simply provides an internal trigger to send out data.
5831 	 * The check_recovery_mode call will see if there are needed
5832 	 * retransmissions, if so we will enter fast-recovery. The output
5833 	 * call may or may not do the same thing depending on sysctl
5834 	 * settings.
5835 	 */
5836 	struct rack_sendmap *rsm;
5837 
5838 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5839 		return (1);
5840 	}
5841 	counter_u64_add(rack_to_tot, 1);
5842 	if (rack->r_state && (rack->r_state != tp->t_state))
5843 		rack_set_state(tp, rack);
5844 	rack->rc_on_min_to = 0;
5845 	rsm = rack_check_recovery_mode(tp, cts);
5846 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5847 	if (rsm) {
5848 		rack->r_ctl.rc_resend = rsm;
5849 		rack->r_timer_override = 1;
5850 		if (rack->use_rack_rr) {
5851 			/*
5852 			 * Don't accumulate extra pacing delay
5853 			 * we are allowing the rack timer to
5854 			 * over-ride pacing i.e. rrr takes precedence
5855 			 * if the pacing interval is longer than the rrr
5856 			 * time (in other words we get the min pacing
5857 			 * time versus rrr pacing time).
5858 			 */
5859 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5860 		}
5861 	}
5862 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5863 	if (rsm == NULL) {
5864 		/* restart a timer and return 1 */
5865 		rack_start_hpts_timer(rack, tp, cts,
5866 				      0, 0, 0);
5867 		return (1);
5868 	}
5869 	return (0);
5870 }
5871 
5872 static void
5873 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5874 {
5875 	if (rsm->m->m_len > rsm->orig_m_len) {
5876 		/*
5877 		 * Mbuf grew, caused by sbcompress, our offset does
5878 		 * not change.
5879 		 */
5880 		rsm->orig_m_len = rsm->m->m_len;
5881 	} else if (rsm->m->m_len < rsm->orig_m_len) {
5882 		/*
5883 		 * Mbuf shrank, trimmed off the top by an ack, our
5884 		 * offset changes.
5885 		 */
5886 		rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
5887 		rsm->orig_m_len = rsm->m->m_len;
5888 	}
5889 }
5890 
5891 static void
5892 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
5893 {
5894 	struct mbuf *m;
5895 	uint32_t soff;
5896 
5897 	if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
5898 		/* Fix up the orig_m_len and possibly the mbuf offset */
5899 		rack_adjust_orig_mlen(src_rsm);
5900 	}
5901 	m = src_rsm->m;
5902 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
5903 	while (soff >= m->m_len) {
5904 		/* Move out past this mbuf */
5905 		soff -= m->m_len;
5906 		m = m->m_next;
5907 		KASSERT((m != NULL),
5908 			("rsm:%p nrsm:%p hit at soff:%u null m",
5909 			 src_rsm, rsm, soff));
5910 	}
5911 	rsm->m = m;
5912 	rsm->soff = soff;
5913 	rsm->orig_m_len = m->m_len;
5914 }
5915 
5916 static __inline void
5917 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5918 	       struct rack_sendmap *rsm, uint32_t start)
5919 {
5920 	int idx;
5921 
5922 	nrsm->r_start = start;
5923 	nrsm->r_end = rsm->r_end;
5924 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
5925 	nrsm->r_flags = rsm->r_flags;
5926 	nrsm->r_dupack = rsm->r_dupack;
5927 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
5928 	nrsm->r_rtr_bytes = 0;
5929 	nrsm->r_fas = rsm->r_fas;
5930 	rsm->r_end = nrsm->r_start;
5931 	nrsm->r_just_ret = rsm->r_just_ret;
5932 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
5933 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
5934 	}
5935 	/* Now if we have SYN flag we keep it on the left edge */
5936 	if (nrsm->r_flags & RACK_HAS_SYN)
5937 		nrsm->r_flags &= ~RACK_HAS_SYN;
5938 	/* Now if we have a FIN flag we keep it on the right edge */
5939 	if (rsm->r_flags & RACK_HAS_FIN)
5940 		rsm->r_flags &= ~RACK_HAS_FIN;
5941 	/* Push bit must go to the right edge as well */
5942 	if (rsm->r_flags & RACK_HAD_PUSH)
5943 		rsm->r_flags &= ~RACK_HAD_PUSH;
5944 	/* Clone over the state of the hw_tls flag */
5945 	nrsm->r_hw_tls = rsm->r_hw_tls;
5946 	/*
5947 	 * Now we need to find nrsm's new location in the mbuf chain
5948 	 * we basically calculate a new offset, which is soff +
5949 	 * how much is left in original rsm. Then we walk out the mbuf
5950 	 * chain to find the righ postion, it may be the same mbuf
5951 	 * or maybe not.
5952 	 */
5953 	KASSERT(((rsm->m != NULL) ||
5954 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
5955 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
5956 	if (rsm->m)
5957 		rack_setup_offset_for_rsm(rsm, nrsm);
5958 }
5959 
5960 static struct rack_sendmap *
5961 rack_merge_rsm(struct tcp_rack *rack,
5962 	       struct rack_sendmap *l_rsm,
5963 	       struct rack_sendmap *r_rsm)
5964 {
5965 	/*
5966 	 * We are merging two ack'd RSM's,
5967 	 * the l_rsm is on the left (lower seq
5968 	 * values) and the r_rsm is on the right
5969 	 * (higher seq value). The simplest way
5970 	 * to merge these is to move the right
5971 	 * one into the left. I don't think there
5972 	 * is any reason we need to try to find
5973 	 * the oldest (or last oldest retransmitted).
5974 	 */
5975 #ifdef INVARIANTS
5976 	struct rack_sendmap *rm;
5977 #endif
5978 	rack_log_map_chg(rack->rc_tp, rack, NULL,
5979 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
5980 	l_rsm->r_end = r_rsm->r_end;
5981 	if (l_rsm->r_dupack < r_rsm->r_dupack)
5982 		l_rsm->r_dupack = r_rsm->r_dupack;
5983 	if (r_rsm->r_rtr_bytes)
5984 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
5985 	if (r_rsm->r_in_tmap) {
5986 		/* This really should not happen */
5987 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
5988 		r_rsm->r_in_tmap = 0;
5989 	}
5990 
5991 	/* Now the flags */
5992 	if (r_rsm->r_flags & RACK_HAS_FIN)
5993 		l_rsm->r_flags |= RACK_HAS_FIN;
5994 	if (r_rsm->r_flags & RACK_TLP)
5995 		l_rsm->r_flags |= RACK_TLP;
5996 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
5997 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
5998 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
5999 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6000 		/*
6001 		 * If both are app-limited then let the
6002 		 * free lower the count. If right is app
6003 		 * limited and left is not, transfer.
6004 		 */
6005 		l_rsm->r_flags |= RACK_APP_LIMITED;
6006 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
6007 		if (r_rsm == rack->r_ctl.rc_first_appl)
6008 			rack->r_ctl.rc_first_appl = l_rsm;
6009 	}
6010 #ifndef INVARIANTS
6011 	(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6012 #else
6013 	rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6014 	if (rm != r_rsm) {
6015 		panic("removing head in rack:%p rsm:%p rm:%p",
6016 		      rack, r_rsm, rm);
6017 	}
6018 #endif
6019 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6020 		/* Transfer the split limit to the map we free */
6021 		r_rsm->r_limit_type = l_rsm->r_limit_type;
6022 		l_rsm->r_limit_type = 0;
6023 	}
6024 	rack_free(rack, r_rsm);
6025 	return (l_rsm);
6026 }
6027 
6028 /*
6029  * TLP Timer, here we simply setup what segment we want to
6030  * have the TLP expire on, the normal rack_output() will then
6031  * send it out.
6032  *
6033  * We return 1, saying don't proceed with rack_output only
6034  * when all timers have been stopped (destroyed PCB?).
6035  */
6036 static int
6037 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6038 {
6039 	/*
6040 	 * Tail Loss Probe.
6041 	 */
6042 	struct rack_sendmap *rsm = NULL;
6043 #ifdef INVARIANTS
6044 	struct rack_sendmap *insret;
6045 #endif
6046 	struct socket *so;
6047 	uint32_t amm;
6048 	uint32_t out, avail;
6049 	int collapsed_win = 0;
6050 
6051 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6052 		return (1);
6053 	}
6054 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6055 		/* Its not time yet */
6056 		return (0);
6057 	}
6058 	if (ctf_progress_timeout_check(tp, true)) {
6059 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6060 		return (-ETIMEDOUT);	/* tcp_drop() */
6061 	}
6062 	/*
6063 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
6064 	 * need to figure out how to force a full MSS segment out.
6065 	 */
6066 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6067 	rack->r_ctl.retran_during_recovery = 0;
6068 	rack->r_ctl.dsack_byte_cnt = 0;
6069 	counter_u64_add(rack_tlp_tot, 1);
6070 	if (rack->r_state && (rack->r_state != tp->t_state))
6071 		rack_set_state(tp, rack);
6072 	so = tp->t_inpcb->inp_socket;
6073 	avail = sbavail(&so->so_snd);
6074 	out = tp->snd_max - tp->snd_una;
6075 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
6076 		/* special case, we need a retransmission */
6077 		collapsed_win = 1;
6078 		goto need_retran;
6079 	}
6080 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6081 		rack->r_ctl.dsack_persist--;
6082 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6083 			rack->r_ctl.num_dsack = 0;
6084 		}
6085 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6086 	}
6087 	if ((tp->t_flags & TF_GPUTINPROG) &&
6088 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6089 		/*
6090 		 * If this is the second in a row
6091 		 * TLP and we are doing a measurement
6092 		 * its time to abandon the measurement.
6093 		 * Something is likely broken on
6094 		 * the clients network and measuring a
6095 		 * broken network does us no good.
6096 		 */
6097 		tp->t_flags &= ~TF_GPUTINPROG;
6098 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6099 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6100 					   tp->gput_seq,
6101 					   0, 0, 18, __LINE__, NULL, 0);
6102 	}
6103 	/*
6104 	 * Check our send oldest always settings, and if
6105 	 * there is an oldest to send jump to the need_retran.
6106 	 */
6107 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6108 		goto need_retran;
6109 
6110 	if (avail > out) {
6111 		/* New data is available */
6112 		amm = avail - out;
6113 		if (amm > ctf_fixed_maxseg(tp)) {
6114 			amm = ctf_fixed_maxseg(tp);
6115 			if ((amm + out) > tp->snd_wnd) {
6116 				/* We are rwnd limited */
6117 				goto need_retran;
6118 			}
6119 		} else if (amm < ctf_fixed_maxseg(tp)) {
6120 			/* not enough to fill a MTU */
6121 			goto need_retran;
6122 		}
6123 		if (IN_FASTRECOVERY(tp->t_flags)) {
6124 			/* Unlikely */
6125 			if (rack->rack_no_prr == 0) {
6126 				if (out + amm <= tp->snd_wnd) {
6127 					rack->r_ctl.rc_prr_sndcnt = amm;
6128 					rack->r_ctl.rc_tlp_new_data = amm;
6129 					rack_log_to_prr(rack, 4, 0, __LINE__);
6130 				}
6131 			} else
6132 				goto need_retran;
6133 		} else {
6134 			/* Set the send-new override */
6135 			if (out + amm <= tp->snd_wnd)
6136 				rack->r_ctl.rc_tlp_new_data = amm;
6137 			else
6138 				goto need_retran;
6139 		}
6140 		rack->r_ctl.rc_tlpsend = NULL;
6141 		counter_u64_add(rack_tlp_newdata, 1);
6142 		goto send;
6143 	}
6144 need_retran:
6145 	/*
6146 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
6147 	 * optionally the first un-acked segment.
6148 	 */
6149 	if (collapsed_win == 0) {
6150 		if (rack_always_send_oldest)
6151 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6152 		else {
6153 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6154 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6155 				rsm = rack_find_high_nonack(rack, rsm);
6156 			}
6157 		}
6158 		if (rsm == NULL) {
6159 #ifdef TCP_BLACKBOX
6160 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6161 #endif
6162 			goto out;
6163 		}
6164 	} else {
6165 		/*
6166 		 * We must find the last segment
6167 		 * that was acceptable by the client.
6168 		 */
6169 		RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6170 			if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6171 				/* Found one */
6172 				break;
6173 			}
6174 		}
6175 		if (rsm == NULL) {
6176 			/* None? if so send the first */
6177 			rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6178 			if (rsm == NULL) {
6179 #ifdef TCP_BLACKBOX
6180 				tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6181 #endif
6182 				goto out;
6183 			}
6184 		}
6185 	}
6186 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6187 		/*
6188 		 * We need to split this the last segment in two.
6189 		 */
6190 		struct rack_sendmap *nrsm;
6191 
6192 		nrsm = rack_alloc_full_limit(rack);
6193 		if (nrsm == NULL) {
6194 			/*
6195 			 * No memory to split, we will just exit and punt
6196 			 * off to the RXT timer.
6197 			 */
6198 			goto out;
6199 		}
6200 		rack_clone_rsm(rack, nrsm, rsm,
6201 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
6202 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6203 #ifndef INVARIANTS
6204 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6205 #else
6206 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6207 		if (insret != NULL) {
6208 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6209 			      nrsm, insret, rack, rsm);
6210 		}
6211 #endif
6212 		if (rsm->r_in_tmap) {
6213 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6214 			nrsm->r_in_tmap = 1;
6215 		}
6216 		rsm = nrsm;
6217 	}
6218 	rack->r_ctl.rc_tlpsend = rsm;
6219 send:
6220 	/* Make sure output path knows we are doing a TLP */
6221 	*doing_tlp = 1;
6222 	rack->r_timer_override = 1;
6223 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6224 	return (0);
6225 out:
6226 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6227 	return (0);
6228 }
6229 
6230 /*
6231  * Delayed ack Timer, here we simply need to setup the
6232  * ACK_NOW flag and remove the DELACK flag. From there
6233  * the output routine will send the ack out.
6234  *
6235  * We only return 1, saying don't proceed, if all timers
6236  * are stopped (destroyed PCB?).
6237  */
6238 static int
6239 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6240 {
6241 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6242 		return (1);
6243 	}
6244 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6245 	tp->t_flags &= ~TF_DELACK;
6246 	tp->t_flags |= TF_ACKNOW;
6247 	KMOD_TCPSTAT_INC(tcps_delack);
6248 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6249 	return (0);
6250 }
6251 
6252 /*
6253  * Persists timer, here we simply send the
6254  * same thing as a keepalive will.
6255  * the one byte send.
6256  *
6257  * We only return 1, saying don't proceed, if all timers
6258  * are stopped (destroyed PCB?).
6259  */
6260 static int
6261 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6262 {
6263 	struct tcptemp *t_template;
6264 #ifdef INVARIANTS
6265 	struct inpcb *inp = tp->t_inpcb;
6266 #endif
6267 	int32_t retval = 1;
6268 
6269 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6270 		return (1);
6271 	}
6272 	if (rack->rc_in_persist == 0)
6273 		return (0);
6274 	if (ctf_progress_timeout_check(tp, false)) {
6275 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6276 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6277 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6278 		return (-ETIMEDOUT);	/* tcp_drop() */
6279 	}
6280 	KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6281 	/*
6282 	 * Persistence timer into zero window. Force a byte to be output, if
6283 	 * possible.
6284 	 */
6285 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
6286 	/*
6287 	 * Hack: if the peer is dead/unreachable, we do not time out if the
6288 	 * window is closed.  After a full backoff, drop the connection if
6289 	 * the idle time (no responses to probes) reaches the maximum
6290 	 * backoff that we would use if retransmitting.
6291 	 */
6292 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6293 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6294 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6295 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6296 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6297 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6298 		retval = -ETIMEDOUT;	/* tcp_drop() */
6299 		goto out;
6300 	}
6301 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6302 	    tp->snd_una == tp->snd_max)
6303 		rack_exit_persist(tp, rack, cts);
6304 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6305 	/*
6306 	 * If the user has closed the socket then drop a persisting
6307 	 * connection after a much reduced timeout.
6308 	 */
6309 	if (tp->t_state > TCPS_CLOSE_WAIT &&
6310 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6311 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6312 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6313 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6314 		retval = -ETIMEDOUT;	/* tcp_drop() */
6315 		goto out;
6316 	}
6317 	t_template = tcpip_maketemplate(rack->rc_inp);
6318 	if (t_template) {
6319 		/* only set it if we were answered */
6320 		if (rack->forced_ack == 0) {
6321 			rack->forced_ack = 1;
6322 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6323 		} else {
6324 			rack->probe_not_answered = 1;
6325 			counter_u64_add(rack_persists_loss, 1);
6326 			rack->r_ctl.persist_lost_ends++;
6327 		}
6328 		counter_u64_add(rack_persists_sends, 1);
6329 		tcp_respond(tp, t_template->tt_ipgen,
6330 			    &t_template->tt_t, (struct mbuf *)NULL,
6331 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6332 		/* This sends an ack */
6333 		if (tp->t_flags & TF_DELACK)
6334 			tp->t_flags &= ~TF_DELACK;
6335 		free(t_template, M_TEMP);
6336 	}
6337 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6338 		tp->t_rxtshift++;
6339 out:
6340 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6341 	rack_start_hpts_timer(rack, tp, cts,
6342 			      0, 0, 0);
6343 	return (retval);
6344 }
6345 
6346 /*
6347  * If a keepalive goes off, we had no other timers
6348  * happening. We always return 1 here since this
6349  * routine either drops the connection or sends
6350  * out a segment with respond.
6351  */
6352 static int
6353 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6354 {
6355 	struct tcptemp *t_template;
6356 	struct inpcb *inp;
6357 
6358 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6359 		return (1);
6360 	}
6361 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6362 	inp = tp->t_inpcb;
6363 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6364 	/*
6365 	 * Keep-alive timer went off; send something or drop connection if
6366 	 * idle for too long.
6367 	 */
6368 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
6369 	if (tp->t_state < TCPS_ESTABLISHED)
6370 		goto dropit;
6371 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6372 	    tp->t_state <= TCPS_CLOSING) {
6373 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6374 			goto dropit;
6375 		/*
6376 		 * Send a packet designed to force a response if the peer is
6377 		 * up and reachable: either an ACK if the connection is
6378 		 * still alive, or an RST if the peer has closed the
6379 		 * connection due to timeout or reboot. Using sequence
6380 		 * number tp->snd_una-1 causes the transmitted zero-length
6381 		 * segment to lie outside the receive window; by the
6382 		 * protocol spec, this requires the correspondent TCP to
6383 		 * respond.
6384 		 */
6385 		KMOD_TCPSTAT_INC(tcps_keepprobe);
6386 		t_template = tcpip_maketemplate(inp);
6387 		if (t_template) {
6388 			if (rack->forced_ack == 0) {
6389 				rack->forced_ack = 1;
6390 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6391 			} else {
6392 				rack->probe_not_answered = 1;
6393 			}
6394 			tcp_respond(tp, t_template->tt_ipgen,
6395 			    &t_template->tt_t, (struct mbuf *)NULL,
6396 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6397 			free(t_template, M_TEMP);
6398 		}
6399 	}
6400 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6401 	return (1);
6402 dropit:
6403 	KMOD_TCPSTAT_INC(tcps_keepdrops);
6404 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6405 	return (-ETIMEDOUT);	/* tcp_drop() */
6406 }
6407 
6408 /*
6409  * Retransmit helper function, clear up all the ack
6410  * flags and take care of important book keeping.
6411  */
6412 static void
6413 rack_remxt_tmr(struct tcpcb *tp)
6414 {
6415 	/*
6416 	 * The retransmit timer went off, all sack'd blocks must be
6417 	 * un-acked.
6418 	 */
6419 	struct rack_sendmap *rsm, *trsm = NULL;
6420 	struct tcp_rack *rack;
6421 
6422 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6423 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6424 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6425 	if (rack->r_state && (rack->r_state != tp->t_state))
6426 		rack_set_state(tp, rack);
6427 	/*
6428 	 * Ideally we would like to be able to
6429 	 * mark SACK-PASS on anything not acked here.
6430 	 *
6431 	 * However, if we do that we would burst out
6432 	 * all that data 1ms apart. This would be unwise,
6433 	 * so for now we will just let the normal rxt timer
6434 	 * and tlp timer take care of it.
6435 	 *
6436 	 * Also we really need to stick them back in sequence
6437 	 * order. This way we send in the proper order and any
6438 	 * sacks that come floating in will "re-ack" the data.
6439 	 * To do this we zap the tmap with an INIT and then
6440 	 * walk through and place every rsm in the RB tree
6441 	 * back in its seq ordered place.
6442 	 */
6443 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
6444 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6445 		rsm->r_dupack = 0;
6446 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6447 		/* We must re-add it back to the tlist */
6448 		if (trsm == NULL) {
6449 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6450 		} else {
6451 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6452 		}
6453 		rsm->r_in_tmap = 1;
6454 		trsm = rsm;
6455 		if (rsm->r_flags & RACK_ACKED)
6456 			rsm->r_flags |= RACK_WAS_ACKED;
6457 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6458 		rsm->r_flags |= RACK_MUST_RXT;
6459 	}
6460 	/* Clear the count (we just un-acked them) */
6461 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6462 	rack->r_ctl.rc_sacked = 0;
6463 	rack->r_ctl.rc_sacklast = NULL;
6464 	rack->r_ctl.rc_agg_delayed = 0;
6465 	rack->r_early = 0;
6466 	rack->r_ctl.rc_agg_early = 0;
6467 	rack->r_late = 0;
6468 	/* Clear the tlp rtx mark */
6469 	rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6470 	if (rack->r_ctl.rc_resend != NULL)
6471 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6472 	rack->r_ctl.rc_prr_sndcnt = 0;
6473 	rack_log_to_prr(rack, 6, 0, __LINE__);
6474 	rack->r_timer_override = 1;
6475 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6476 #ifdef NETFLIX_EXP_DETECTION
6477 	    || (rack->sack_attack_disable != 0)
6478 #endif
6479 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6480 		/*
6481 		 * For non-sack customers new data
6482 		 * needs to go out as retransmits until
6483 		 * we retransmit up to snd_max.
6484 		 */
6485 		rack->r_must_retran = 1;
6486 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6487 						rack->r_ctl.rc_sacked);
6488 	}
6489 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6490 }
6491 
6492 static void
6493 rack_convert_rtts(struct tcpcb *tp)
6494 {
6495 	if (tp->t_srtt > 1) {
6496 		uint32_t val, frac;
6497 
6498 		val = tp->t_srtt >> TCP_RTT_SHIFT;
6499 		frac = tp->t_srtt & 0x1f;
6500 		tp->t_srtt = TICKS_2_USEC(val);
6501 		/*
6502 		 * frac is the fractional part of the srtt (if any)
6503 		 * but its in ticks and every bit represents
6504 		 * 1/32nd of a hz.
6505 		 */
6506 		if (frac) {
6507 			if (hz == 1000) {
6508 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6509 			} else {
6510 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6511 			}
6512 			tp->t_srtt += frac;
6513 		}
6514 	}
6515 	if (tp->t_rttvar) {
6516 		uint32_t val, frac;
6517 
6518 		val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6519 		frac = tp->t_rttvar & 0x1f;
6520 		tp->t_rttvar = TICKS_2_USEC(val);
6521 		/*
6522 		 * frac is the fractional part of the srtt (if any)
6523 		 * but its in ticks and every bit represents
6524 		 * 1/32nd of a hz.
6525 		 */
6526 		if (frac) {
6527 			if (hz == 1000) {
6528 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6529 			} else {
6530 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6531 			}
6532 			tp->t_rttvar += frac;
6533 		}
6534 	}
6535 	tp->t_rxtcur = RACK_REXMTVAL(tp);
6536 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6537 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6538 	}
6539 	if (tp->t_rxtcur > rack_rto_max) {
6540 		tp->t_rxtcur = rack_rto_max;
6541 	}
6542 }
6543 
6544 static void
6545 rack_cc_conn_init(struct tcpcb *tp)
6546 {
6547 	struct tcp_rack *rack;
6548 	uint32_t srtt;
6549 
6550 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6551 	srtt = tp->t_srtt;
6552 	cc_conn_init(tp);
6553 	/*
6554 	 * Now convert to rack's internal format,
6555 	 * if required.
6556 	 */
6557 	if ((srtt == 0) && (tp->t_srtt != 0))
6558 		rack_convert_rtts(tp);
6559 	/*
6560 	 * We want a chance to stay in slowstart as
6561 	 * we create a connection. TCP spec says that
6562 	 * initially ssthresh is infinite. For our
6563 	 * purposes that is the snd_wnd.
6564 	 */
6565 	if (tp->snd_ssthresh < tp->snd_wnd) {
6566 		tp->snd_ssthresh = tp->snd_wnd;
6567 	}
6568 	/*
6569 	 * We also want to assure a IW worth of
6570 	 * data can get inflight.
6571 	 */
6572 	if (rc_init_window(rack) < tp->snd_cwnd)
6573 		tp->snd_cwnd = rc_init_window(rack);
6574 }
6575 
6576 /*
6577  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6578  * we will setup to retransmit the lowest seq number outstanding.
6579  */
6580 static int
6581 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6582 {
6583 	int32_t rexmt;
6584 	int32_t retval = 0;
6585 	bool isipv6;
6586 
6587 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6588 		return (1);
6589 	}
6590 	if ((tp->t_flags & TF_GPUTINPROG) &&
6591 	    (tp->t_rxtshift)) {
6592 		/*
6593 		 * We have had a second timeout
6594 		 * measurements on successive rxt's are not profitable.
6595 		 * It is unlikely to be of any use (the network is
6596 		 * broken or the client went away).
6597 		 */
6598 		tp->t_flags &= ~TF_GPUTINPROG;
6599 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6600 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6601 					   tp->gput_seq,
6602 					   0, 0, 18, __LINE__, NULL, 0);
6603 	}
6604 	if (ctf_progress_timeout_check(tp, false)) {
6605 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6606 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6607 		return (-ETIMEDOUT);	/* tcp_drop() */
6608 	}
6609 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6610 	rack->r_ctl.retran_during_recovery = 0;
6611 	rack->r_ctl.dsack_byte_cnt = 0;
6612 	if (IN_FASTRECOVERY(tp->t_flags))
6613 		tp->t_flags |= TF_WASFRECOVERY;
6614 	else
6615 		tp->t_flags &= ~TF_WASFRECOVERY;
6616 	if (IN_CONGRECOVERY(tp->t_flags))
6617 		tp->t_flags |= TF_WASCRECOVERY;
6618 	else
6619 		tp->t_flags &= ~TF_WASCRECOVERY;
6620 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6621 	    (tp->snd_una == tp->snd_max)) {
6622 		/* Nothing outstanding .. nothing to do */
6623 		return (0);
6624 	}
6625 	if (rack->r_ctl.dsack_persist) {
6626 		rack->r_ctl.dsack_persist--;
6627 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6628 			rack->r_ctl.num_dsack = 0;
6629 		}
6630 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6631 	}
6632 	/*
6633 	 * Rack can only run one timer  at a time, so we cannot
6634 	 * run a KEEPINIT (gating SYN sending) and a retransmit
6635 	 * timer for the SYN. So if we are in a front state and
6636 	 * have a KEEPINIT timer we need to check the first transmit
6637 	 * against now to see if we have exceeded the KEEPINIT time
6638 	 * (if one is set).
6639 	 */
6640 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6641 	    (TP_KEEPINIT(tp) != 0)) {
6642 		struct rack_sendmap *rsm;
6643 
6644 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6645 		if (rsm) {
6646 			/* Ok we have something outstanding to test keepinit with */
6647 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6648 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6649 				/* We have exceeded the KEEPINIT time */
6650 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6651 				goto drop_it;
6652 			}
6653 		}
6654 	}
6655 	/*
6656 	 * Retransmission timer went off.  Message has not been acked within
6657 	 * retransmit interval.  Back off to a longer retransmit interval
6658 	 * and retransmit one segment.
6659 	 */
6660 	rack_remxt_tmr(tp);
6661 	if ((rack->r_ctl.rc_resend == NULL) ||
6662 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6663 		/*
6664 		 * If the rwnd collapsed on
6665 		 * the one we are retransmitting
6666 		 * it does not count against the
6667 		 * rxt count.
6668 		 */
6669 		tp->t_rxtshift++;
6670 	}
6671 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6672 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6673 drop_it:
6674 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
6675 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6676 		/* XXXGL: previously t_softerror was casted to uint16_t */
6677 		MPASS(tp->t_softerror >= 0);
6678 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6679 		goto out;	/* tcp_drop() */
6680 	}
6681 	if (tp->t_state == TCPS_SYN_SENT) {
6682 		/*
6683 		 * If the SYN was retransmitted, indicate CWND to be limited
6684 		 * to 1 segment in cc_conn_init().
6685 		 */
6686 		tp->snd_cwnd = 1;
6687 	} else if (tp->t_rxtshift == 1) {
6688 		/*
6689 		 * first retransmit; record ssthresh and cwnd so they can be
6690 		 * recovered if this turns out to be a "bad" retransmit. A
6691 		 * retransmit is considered "bad" if an ACK for this segment
6692 		 * is received within RTT/2 interval; the assumption here is
6693 		 * that the ACK was already in flight.  See "On Estimating
6694 		 * End-to-End Network Path Properties" by Allman and Paxson
6695 		 * for more details.
6696 		 */
6697 		tp->snd_cwnd_prev = tp->snd_cwnd;
6698 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
6699 		tp->snd_recover_prev = tp->snd_recover;
6700 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6701 		tp->t_flags |= TF_PREVVALID;
6702 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6703 		tp->t_flags &= ~TF_PREVVALID;
6704 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6705 	if ((tp->t_state == TCPS_SYN_SENT) ||
6706 	    (tp->t_state == TCPS_SYN_RECEIVED))
6707 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6708 	else
6709 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6710 
6711 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6712 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6713 	/*
6714 	 * We enter the path for PLMTUD if connection is established or, if
6715 	 * connection is FIN_WAIT_1 status, reason for the last is that if
6716 	 * amount of data we send is very small, we could send it in couple
6717 	 * of packets and process straight to FIN. In that case we won't
6718 	 * catch ESTABLISHED state.
6719 	 */
6720 #ifdef INET6
6721 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6722 #else
6723 	isipv6 = false;
6724 #endif
6725 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
6726 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6727 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6728 	    ((tp->t_state == TCPS_ESTABLISHED) ||
6729 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
6730 		/*
6731 		 * Idea here is that at each stage of mtu probe (usually,
6732 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
6733 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6734 		 * should take care of that.
6735 		 */
6736 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6737 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6738 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6739 		    tp->t_rxtshift % 2 == 0)) {
6740 			/*
6741 			 * Enter Path MTU Black-hole Detection mechanism: -
6742 			 * Disable Path MTU Discovery (IP "DF" bit). -
6743 			 * Reduce MTU to lower value than what we negotiated
6744 			 * with peer.
6745 			 */
6746 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6747 				/* Record that we may have found a black hole. */
6748 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6749 				/* Keep track of previous MSS. */
6750 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6751 			}
6752 
6753 			/*
6754 			 * Reduce the MSS to blackhole value or to the
6755 			 * default in an attempt to retransmit.
6756 			 */
6757 #ifdef INET6
6758 			if (isipv6 &&
6759 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6760 				/* Use the sysctl tuneable blackhole MSS. */
6761 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6762 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6763 			} else if (isipv6) {
6764 				/* Use the default MSS. */
6765 				tp->t_maxseg = V_tcp_v6mssdflt;
6766 				/*
6767 				 * Disable Path MTU Discovery when we switch
6768 				 * to minmss.
6769 				 */
6770 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6771 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6772 			}
6773 #endif
6774 #if defined(INET6) && defined(INET)
6775 			else
6776 #endif
6777 #ifdef INET
6778 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6779 				/* Use the sysctl tuneable blackhole MSS. */
6780 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6781 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6782 			} else {
6783 				/* Use the default MSS. */
6784 				tp->t_maxseg = V_tcp_mssdflt;
6785 				/*
6786 				 * Disable Path MTU Discovery when we switch
6787 				 * to minmss.
6788 				 */
6789 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6790 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6791 			}
6792 #endif
6793 		} else {
6794 			/*
6795 			 * If further retransmissions are still unsuccessful
6796 			 * with a lowered MTU, maybe this isn't a blackhole
6797 			 * and we restore the previous MSS and blackhole
6798 			 * detection flags. The limit '6' is determined by
6799 			 * giving each probe stage (1448, 1188, 524) 2
6800 			 * chances to recover.
6801 			 */
6802 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6803 			    (tp->t_rxtshift >= 6)) {
6804 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6805 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6806 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6807 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6808 			}
6809 		}
6810 	}
6811 	/*
6812 	 * Disable RFC1323 and SACK if we haven't got any response to
6813 	 * our third SYN to work-around some broken terminal servers
6814 	 * (most of which have hopefully been retired) that have bad VJ
6815 	 * header compression code which trashes TCP segments containing
6816 	 * unknown-to-them TCP options.
6817 	 */
6818 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6819 	    (tp->t_rxtshift == 3))
6820 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6821 	/*
6822 	 * If we backed off this far, our srtt estimate is probably bogus.
6823 	 * Clobber it so we'll take the next rtt measurement as our srtt;
6824 	 * move the current srtt into rttvar to keep the current retransmit
6825 	 * times until then.
6826 	 */
6827 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6828 #ifdef INET6
6829 		if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6830 			in6_losing(tp->t_inpcb);
6831 		else
6832 #endif
6833 			in_losing(tp->t_inpcb);
6834 		tp->t_rttvar += tp->t_srtt;
6835 		tp->t_srtt = 0;
6836 	}
6837 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6838 	tp->snd_recover = tp->snd_max;
6839 	tp->t_flags |= TF_ACKNOW;
6840 	tp->t_rtttime = 0;
6841 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
6842 out:
6843 	return (retval);
6844 }
6845 
6846 static int
6847 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6848 {
6849 	int32_t ret = 0;
6850 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6851 
6852 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6853 	    (tp->t_flags & TF_GPUTINPROG)) {
6854 		/*
6855 		 * We have a goodput in progress
6856 		 * and we have entered a late state.
6857 		 * Do we have enough data in the sb
6858 		 * to handle the GPUT request?
6859 		 */
6860 		uint32_t bytes;
6861 
6862 		bytes = tp->gput_ack - tp->gput_seq;
6863 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
6864 			bytes += tp->gput_seq - tp->snd_una;
6865 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
6866 			/*
6867 			 * There are not enough bytes in the socket
6868 			 * buffer that have been sent to cover this
6869 			 * measurement. Cancel it.
6870 			 */
6871 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6872 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
6873 						   tp->gput_seq,
6874 						   0, 0, 18, __LINE__, NULL, 0);
6875 			tp->t_flags &= ~TF_GPUTINPROG;
6876 		}
6877 	}
6878 	if (timers == 0) {
6879 		return (0);
6880 	}
6881 	if (tp->t_state == TCPS_LISTEN) {
6882 		/* no timers on listen sockets */
6883 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6884 			return (0);
6885 		return (1);
6886 	}
6887 	if ((timers & PACE_TMR_RACK) &&
6888 	    rack->rc_on_min_to) {
6889 		/*
6890 		 * For the rack timer when we
6891 		 * are on a min-timeout (which means rrr_conf = 3)
6892 		 * we don't want to check the timer. It may
6893 		 * be going off for a pace and thats ok we
6894 		 * want to send the retransmit (if its ready).
6895 		 *
6896 		 * If its on a normal rack timer (non-min) then
6897 		 * we will check if its expired.
6898 		 */
6899 		goto skip_time_check;
6900 	}
6901 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6902 		uint32_t left;
6903 
6904 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6905 			ret = -1;
6906 			rack_log_to_processing(rack, cts, ret, 0);
6907 			return (0);
6908 		}
6909 		if (hpts_calling == 0) {
6910 			/*
6911 			 * A user send or queued mbuf (sack) has called us? We
6912 			 * return 0 and let the pacing guards
6913 			 * deal with it if they should or
6914 			 * should not cause a send.
6915 			 */
6916 			ret = -2;
6917 			rack_log_to_processing(rack, cts, ret, 0);
6918 			return (0);
6919 		}
6920 		/*
6921 		 * Ok our timer went off early and we are not paced false
6922 		 * alarm, go back to sleep.
6923 		 */
6924 		ret = -3;
6925 		left = rack->r_ctl.rc_timer_exp - cts;
6926 		tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
6927 		rack_log_to_processing(rack, cts, ret, left);
6928 		return (1);
6929 	}
6930 skip_time_check:
6931 	rack->rc_tmr_stopped = 0;
6932 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6933 	if (timers & PACE_TMR_DELACK) {
6934 		ret = rack_timeout_delack(tp, rack, cts);
6935 	} else if (timers & PACE_TMR_RACK) {
6936 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6937 		rack->r_fast_output = 0;
6938 		ret = rack_timeout_rack(tp, rack, cts);
6939 	} else if (timers & PACE_TMR_TLP) {
6940 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6941 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
6942 	} else if (timers & PACE_TMR_RXT) {
6943 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6944 		rack->r_fast_output = 0;
6945 		ret = rack_timeout_rxt(tp, rack, cts);
6946 	} else if (timers & PACE_TMR_PERSIT) {
6947 		ret = rack_timeout_persist(tp, rack, cts);
6948 	} else if (timers & PACE_TMR_KEEP) {
6949 		ret = rack_timeout_keepalive(tp, rack, cts);
6950 	}
6951 	rack_log_to_processing(rack, cts, ret, timers);
6952 	return (ret);
6953 }
6954 
6955 static void
6956 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6957 {
6958 	struct timeval tv;
6959 	uint32_t us_cts, flags_on_entry;
6960 	uint8_t hpts_removed = 0;
6961 
6962 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
6963 	us_cts = tcp_get_usecs(&tv);
6964 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
6965 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
6966 	     ((tp->snd_max - tp->snd_una) == 0))) {
6967 		tcp_hpts_remove(rack->rc_inp);
6968 		hpts_removed = 1;
6969 		/* If we were not delayed cancel out the flag. */
6970 		if ((tp->snd_max - tp->snd_una) == 0)
6971 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6972 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6973 	}
6974 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
6975 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
6976 		if (tcp_in_hpts(rack->rc_inp) &&
6977 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
6978 			/*
6979 			 * Canceling timer's when we have no output being
6980 			 * paced. We also must remove ourselves from the
6981 			 * hpts.
6982 			 */
6983 			tcp_hpts_remove(rack->rc_inp);
6984 			hpts_removed = 1;
6985 		}
6986 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
6987 	}
6988 	if (hpts_removed == 0)
6989 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6990 }
6991 
6992 static void
6993 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
6994 {
6995 	return;
6996 }
6997 
6998 static int
6999 rack_stopall(struct tcpcb *tp)
7000 {
7001 	struct tcp_rack *rack;
7002 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7003 	rack->t_timers_stopped = 1;
7004 	return (0);
7005 }
7006 
7007 static void
7008 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7009 {
7010 	return;
7011 }
7012 
7013 static int
7014 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7015 {
7016 	return (0);
7017 }
7018 
7019 static void
7020 rack_stop_all_timers(struct tcpcb *tp)
7021 {
7022 	struct tcp_rack *rack;
7023 
7024 	/*
7025 	 * Assure no timers are running.
7026 	 */
7027 	if (tcp_timer_active(tp, TT_PERSIST)) {
7028 		/* We enter in persists, set the flag appropriately */
7029 		rack = (struct tcp_rack *)tp->t_fb_ptr;
7030 		rack->rc_in_persist = 1;
7031 	}
7032 	tcp_timer_suspend(tp, TT_PERSIST);
7033 	tcp_timer_suspend(tp, TT_REXMT);
7034 	tcp_timer_suspend(tp, TT_KEEP);
7035 	tcp_timer_suspend(tp, TT_DELACK);
7036 }
7037 
7038 static void
7039 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7040     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7041 {
7042 	int32_t idx;
7043 
7044 	rsm->r_rtr_cnt++;
7045 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7046 	rsm->r_dupack = 0;
7047 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7048 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7049 		rsm->r_flags |= RACK_OVERMAX;
7050 	}
7051 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7052 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7053 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7054 	}
7055 	idx = rsm->r_rtr_cnt - 1;
7056 	rsm->r_tim_lastsent[idx] = ts;
7057 	/*
7058 	 * Here we don't add in the len of send, since its already
7059 	 * in snduna <->snd_max.
7060 	 */
7061 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
7062 				     rack->r_ctl.rc_sacked);
7063 	if (rsm->r_flags & RACK_ACKED) {
7064 		/* Problably MTU discovery messing with us */
7065 		rsm->r_flags &= ~RACK_ACKED;
7066 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7067 	}
7068 	if (rsm->r_in_tmap) {
7069 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7070 		rsm->r_in_tmap = 0;
7071 	}
7072 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7073 	rsm->r_in_tmap = 1;
7074 	/* Take off the must retransmit flag, if its on */
7075 	if (rsm->r_flags & RACK_MUST_RXT) {
7076 		if (rack->r_must_retran)
7077 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
7078 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
7079 			/*
7080 			 * We have retransmitted all we need. Clear
7081 			 * any must retransmit flags.
7082 			 */
7083 			rack->r_must_retran = 0;
7084 			rack->r_ctl.rc_out_at_rto = 0;
7085 		}
7086 		rsm->r_flags &= ~RACK_MUST_RXT;
7087 	}
7088 	if (rsm->r_flags & RACK_SACK_PASSED) {
7089 		/* We have retransmitted due to the SACK pass */
7090 		rsm->r_flags &= ~RACK_SACK_PASSED;
7091 		rsm->r_flags |= RACK_WAS_SACKPASS;
7092 	}
7093 }
7094 
7095 static uint32_t
7096 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7097     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7098 {
7099 	/*
7100 	 * We (re-)transmitted starting at rsm->r_start for some length
7101 	 * (possibly less than r_end.
7102 	 */
7103 	struct rack_sendmap *nrsm;
7104 #ifdef INVARIANTS
7105 	struct rack_sendmap *insret;
7106 #endif
7107 	uint32_t c_end;
7108 	int32_t len;
7109 
7110 	len = *lenp;
7111 	c_end = rsm->r_start + len;
7112 	if (SEQ_GEQ(c_end, rsm->r_end)) {
7113 		/*
7114 		 * We retransmitted the whole piece or more than the whole
7115 		 * slopping into the next rsm.
7116 		 */
7117 		rack_update_rsm(tp, rack, rsm, ts, add_flag);
7118 		if (c_end == rsm->r_end) {
7119 			*lenp = 0;
7120 			return (0);
7121 		} else {
7122 			int32_t act_len;
7123 
7124 			/* Hangs over the end return whats left */
7125 			act_len = rsm->r_end - rsm->r_start;
7126 			*lenp = (len - act_len);
7127 			return (rsm->r_end);
7128 		}
7129 		/* We don't get out of this block. */
7130 	}
7131 	/*
7132 	 * Here we retransmitted less than the whole thing which means we
7133 	 * have to split this into what was transmitted and what was not.
7134 	 */
7135 	nrsm = rack_alloc_full_limit(rack);
7136 	if (nrsm == NULL) {
7137 		/*
7138 		 * We can't get memory, so lets not proceed.
7139 		 */
7140 		*lenp = 0;
7141 		return (0);
7142 	}
7143 	/*
7144 	 * So here we are going to take the original rsm and make it what we
7145 	 * retransmitted. nrsm will be the tail portion we did not
7146 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7147 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7148 	 * 1, 6 and the new piece will be 6, 11.
7149 	 */
7150 	rack_clone_rsm(rack, nrsm, rsm, c_end);
7151 	nrsm->r_dupack = 0;
7152 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7153 #ifndef INVARIANTS
7154 	(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7155 #else
7156 	insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7157 	if (insret != NULL) {
7158 		panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7159 		      nrsm, insret, rack, rsm);
7160 	}
7161 #endif
7162 	if (rsm->r_in_tmap) {
7163 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7164 		nrsm->r_in_tmap = 1;
7165 	}
7166 	rsm->r_flags &= (~RACK_HAS_FIN);
7167 	rack_update_rsm(tp, rack, rsm, ts, add_flag);
7168 	/* Log a split of rsm into rsm and nrsm */
7169 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7170 	*lenp = 0;
7171 	return (0);
7172 }
7173 
7174 static void
7175 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7176 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
7177 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7178 {
7179 	struct tcp_rack *rack;
7180 	struct rack_sendmap *rsm, *nrsm, fe;
7181 #ifdef INVARIANTS
7182 	struct rack_sendmap *insret;
7183 #endif
7184 	register uint32_t snd_max, snd_una;
7185 
7186 	/*
7187 	 * Add to the RACK log of packets in flight or retransmitted. If
7188 	 * there is a TS option we will use the TS echoed, if not we will
7189 	 * grab a TS.
7190 	 *
7191 	 * Retransmissions will increment the count and move the ts to its
7192 	 * proper place. Note that if options do not include TS's then we
7193 	 * won't be able to effectively use the ACK for an RTT on a retran.
7194 	 *
7195 	 * Notes about r_start and r_end. Lets consider a send starting at
7196 	 * sequence 1 for 10 bytes. In such an example the r_start would be
7197 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7198 	 * This means that r_end is actually the first sequence for the next
7199 	 * slot (11).
7200 	 *
7201 	 */
7202 	/*
7203 	 * If err is set what do we do XXXrrs? should we not add the thing?
7204 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
7205 	 * i.e. proceed with add ** do this for now.
7206 	 */
7207 	INP_WLOCK_ASSERT(tp->t_inpcb);
7208 	if (err)
7209 		/*
7210 		 * We don't log errors -- we could but snd_max does not
7211 		 * advance in this case either.
7212 		 */
7213 		return;
7214 
7215 	if (th_flags & TH_RST) {
7216 		/*
7217 		 * We don't log resets and we return immediately from
7218 		 * sending
7219 		 */
7220 		return;
7221 	}
7222 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7223 	snd_una = tp->snd_una;
7224 	snd_max = tp->snd_max;
7225 	if (th_flags & (TH_SYN | TH_FIN)) {
7226 		/*
7227 		 * The call to rack_log_output is made before bumping
7228 		 * snd_max. This means we can record one extra byte on a SYN
7229 		 * or FIN if seq_out is adding more on and a FIN is present
7230 		 * (and we are not resending).
7231 		 */
7232 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7233 			len++;
7234 		if (th_flags & TH_FIN)
7235 			len++;
7236 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
7237 			/*
7238 			 * The add/update as not been done for the FIN/SYN
7239 			 * yet.
7240 			 */
7241 			snd_max = tp->snd_nxt;
7242 		}
7243 	}
7244 	if (SEQ_LEQ((seq_out + len), snd_una)) {
7245 		/* Are sending an old segment to induce an ack (keep-alive)? */
7246 		return;
7247 	}
7248 	if (SEQ_LT(seq_out, snd_una)) {
7249 		/* huh? should we panic? */
7250 		uint32_t end;
7251 
7252 		end = seq_out + len;
7253 		seq_out = snd_una;
7254 		if (SEQ_GEQ(end, seq_out))
7255 			len = end - seq_out;
7256 		else
7257 			len = 0;
7258 	}
7259 	if (len == 0) {
7260 		/* We don't log zero window probes */
7261 		return;
7262 	}
7263 	if (IN_FASTRECOVERY(tp->t_flags)) {
7264 		rack->r_ctl.rc_prr_out += len;
7265 	}
7266 	/* First question is it a retransmission or new? */
7267 	if (seq_out == snd_max) {
7268 		/* Its new */
7269 again:
7270 		rsm = rack_alloc(rack);
7271 		if (rsm == NULL) {
7272 			/*
7273 			 * Hmm out of memory and the tcb got destroyed while
7274 			 * we tried to wait.
7275 			 */
7276 			return;
7277 		}
7278 		if (th_flags & TH_FIN) {
7279 			rsm->r_flags = RACK_HAS_FIN|add_flag;
7280 		} else {
7281 			rsm->r_flags = add_flag;
7282 		}
7283 		if (hw_tls)
7284 			rsm->r_hw_tls = 1;
7285 		rsm->r_tim_lastsent[0] = cts;
7286 		rsm->r_rtr_cnt = 1;
7287 		rsm->r_rtr_bytes = 0;
7288 		if (th_flags & TH_SYN) {
7289 			/* The data space is one beyond snd_una */
7290 			rsm->r_flags |= RACK_HAS_SYN;
7291 		}
7292 		rsm->r_start = seq_out;
7293 		rsm->r_end = rsm->r_start + len;
7294 		rsm->r_dupack = 0;
7295 		/*
7296 		 * save off the mbuf location that
7297 		 * sndmbuf_noadv returned (which is
7298 		 * where we started copying from)..
7299 		 */
7300 		rsm->m = s_mb;
7301 		rsm->soff = s_moff;
7302 		/*
7303 		 * Here we do add in the len of send, since its not yet
7304 		 * reflected in in snduna <->snd_max
7305 		 */
7306 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7307 					      rack->r_ctl.rc_sacked) +
7308 			      (rsm->r_end - rsm->r_start));
7309 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7310 		if (rsm->m) {
7311 			if (rsm->m->m_len <= rsm->soff) {
7312 				/*
7313 				 * XXXrrs Question, will this happen?
7314 				 *
7315 				 * If sbsndptr is set at the correct place
7316 				 * then s_moff should always be somewhere
7317 				 * within rsm->m. But if the sbsndptr was
7318 				 * off then that won't be true. If it occurs
7319 				 * we need to walkout to the correct location.
7320 				 */
7321 				struct mbuf *lm;
7322 
7323 				lm = rsm->m;
7324 				while (lm->m_len <= rsm->soff) {
7325 					rsm->soff -= lm->m_len;
7326 					lm = lm->m_next;
7327 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7328 							     __func__, rack, s_moff, s_mb, rsm->soff));
7329 				}
7330 				rsm->m = lm;
7331 			}
7332 			rsm->orig_m_len = rsm->m->m_len;
7333 		} else
7334 			rsm->orig_m_len = 0;
7335 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7336 		/* Log a new rsm */
7337 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7338 #ifndef INVARIANTS
7339 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7340 #else
7341 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7342 		if (insret != NULL) {
7343 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7344 			      nrsm, insret, rack, rsm);
7345 		}
7346 #endif
7347 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7348 		rsm->r_in_tmap = 1;
7349 		/*
7350 		 * Special case detection, is there just a single
7351 		 * packet outstanding when we are not in recovery?
7352 		 *
7353 		 * If this is true mark it so.
7354 		 */
7355 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7356 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7357 			struct rack_sendmap *prsm;
7358 
7359 			prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7360 			if (prsm)
7361 				prsm->r_one_out_nr = 1;
7362 		}
7363 		return;
7364 	}
7365 	/*
7366 	 * If we reach here its a retransmission and we need to find it.
7367 	 */
7368 	memset(&fe, 0, sizeof(fe));
7369 more:
7370 	if (hintrsm && (hintrsm->r_start == seq_out)) {
7371 		rsm = hintrsm;
7372 		hintrsm = NULL;
7373 	} else {
7374 		/* No hints sorry */
7375 		rsm = NULL;
7376 	}
7377 	if ((rsm) && (rsm->r_start == seq_out)) {
7378 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7379 		if (len == 0) {
7380 			return;
7381 		} else {
7382 			goto more;
7383 		}
7384 	}
7385 	/* Ok it was not the last pointer go through it the hard way. */
7386 refind:
7387 	fe.r_start = seq_out;
7388 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7389 	if (rsm) {
7390 		if (rsm->r_start == seq_out) {
7391 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7392 			if (len == 0) {
7393 				return;
7394 			} else {
7395 				goto refind;
7396 			}
7397 		}
7398 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7399 			/* Transmitted within this piece */
7400 			/*
7401 			 * Ok we must split off the front and then let the
7402 			 * update do the rest
7403 			 */
7404 			nrsm = rack_alloc_full_limit(rack);
7405 			if (nrsm == NULL) {
7406 				rack_update_rsm(tp, rack, rsm, cts, add_flag);
7407 				return;
7408 			}
7409 			/*
7410 			 * copy rsm to nrsm and then trim the front of rsm
7411 			 * to not include this part.
7412 			 */
7413 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
7414 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7415 #ifndef INVARIANTS
7416 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7417 #else
7418 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7419 			if (insret != NULL) {
7420 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7421 				      nrsm, insret, rack, rsm);
7422 			}
7423 #endif
7424 			if (rsm->r_in_tmap) {
7425 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7426 				nrsm->r_in_tmap = 1;
7427 			}
7428 			rsm->r_flags &= (~RACK_HAS_FIN);
7429 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7430 			if (len == 0) {
7431 				return;
7432 			} else if (len > 0)
7433 				goto refind;
7434 		}
7435 	}
7436 	/*
7437 	 * Hmm not found in map did they retransmit both old and on into the
7438 	 * new?
7439 	 */
7440 	if (seq_out == tp->snd_max) {
7441 		goto again;
7442 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
7443 #ifdef INVARIANTS
7444 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7445 		       seq_out, len, tp->snd_una, tp->snd_max);
7446 		printf("Starting Dump of all rack entries\n");
7447 		RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7448 			printf("rsm:%p start:%u end:%u\n",
7449 			       rsm, rsm->r_start, rsm->r_end);
7450 		}
7451 		printf("Dump complete\n");
7452 		panic("seq_out not found rack:%p tp:%p",
7453 		      rack, tp);
7454 #endif
7455 	} else {
7456 #ifdef INVARIANTS
7457 		/*
7458 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
7459 		 * flag)
7460 		 */
7461 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7462 		      seq_out, len, tp->snd_max, tp);
7463 #endif
7464 	}
7465 }
7466 
7467 /*
7468  * Record one of the RTT updates from an ack into
7469  * our sample structure.
7470  */
7471 
7472 static void
7473 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7474 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7475 {
7476 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7477 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7478 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7479 	}
7480 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7481 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7482 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7483 	}
7484 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7485 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7486 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
7487 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7488 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7489 	}
7490 	if ((confidence == 1) &&
7491 	    ((rsm == NULL) ||
7492 	     (rsm->r_just_ret) ||
7493 	     (rsm->r_one_out_nr &&
7494 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7495 		/*
7496 		 * If the rsm had a just return
7497 		 * hit it then we can't trust the
7498 		 * rtt measurement for buffer deterimination
7499 		 * Note that a confidence of 2, indicates
7500 		 * SACK'd which overrides the r_just_ret or
7501 		 * the r_one_out_nr. If it was a CUM-ACK and
7502 		 * we had only two outstanding, but get an
7503 		 * ack for only 1. Then that also lowers our
7504 		 * confidence.
7505 		 */
7506 		confidence = 0;
7507 	}
7508 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7509 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7510 		if (rack->r_ctl.rack_rs.confidence == 0) {
7511 			/*
7512 			 * We take anything with no current confidence
7513 			 * saved.
7514 			 */
7515 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7516 			rack->r_ctl.rack_rs.confidence = confidence;
7517 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7518 		} else if (confidence || rack->r_ctl.rack_rs.confidence) {
7519 			/*
7520 			 * Once we have a confident number,
7521 			 * we can update it with a smaller
7522 			 * value since this confident number
7523 			 * may include the DSACK time until
7524 			 * the next segment (the second one) arrived.
7525 			 */
7526 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7527 			rack->r_ctl.rack_rs.confidence = confidence;
7528 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7529 		}
7530 	}
7531 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7532 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7533 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7534 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
7535 }
7536 
7537 /*
7538  * Collect new round-trip time estimate
7539  * and update averages and current timeout.
7540  */
7541 static void
7542 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7543 {
7544 	int32_t delta;
7545 	int32_t rtt;
7546 
7547 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7548 		/* No valid sample */
7549 		return;
7550 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7551 		/* We are to use the lowest RTT seen in a single ack */
7552 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7553 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7554 		/* We are to use the highest RTT seen in a single ack */
7555 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7556 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7557 		/* We are to use the average RTT seen in a single ack */
7558 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7559 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7560 	} else {
7561 #ifdef INVARIANTS
7562 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7563 #endif
7564 		return;
7565 	}
7566 	if (rtt == 0)
7567 		rtt = 1;
7568 	if (rack->rc_gp_rtt_set == 0) {
7569 		/*
7570 		 * With no RTT we have to accept
7571 		 * even one we are not confident of.
7572 		 */
7573 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7574 		rack->rc_gp_rtt_set = 1;
7575 	} else if (rack->r_ctl.rack_rs.confidence) {
7576 		/* update the running gp srtt */
7577 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7578 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7579 	}
7580 	if (rack->r_ctl.rack_rs.confidence) {
7581 		/*
7582 		 * record the low and high for highly buffered path computation,
7583 		 * we only do this if we are confident (not a retransmission).
7584 		 */
7585 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7586 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7587 		}
7588 		if (rack->rc_highly_buffered == 0) {
7589 			/*
7590 			 * Currently once we declare a path has
7591 			 * highly buffered there is no going
7592 			 * back, which may be a problem...
7593 			 */
7594 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7595 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7596 						     rack->r_ctl.rc_highest_us_rtt,
7597 						     rack->r_ctl.rc_lowest_us_rtt,
7598 						     RACK_RTTS_SEEHBP);
7599 				rack->rc_highly_buffered = 1;
7600 			}
7601 		}
7602 	}
7603 	if ((rack->r_ctl.rack_rs.confidence) ||
7604 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7605 		/*
7606 		 * If we are highly confident of it <or> it was
7607 		 * never retransmitted we accept it as the last us_rtt.
7608 		 */
7609 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7610 		/* The lowest rtt can be set if its was not retransmited */
7611 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7612 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7613 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
7614 				rack->r_ctl.rc_lowest_us_rtt = 1;
7615 		}
7616 	}
7617 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7618 	if (tp->t_srtt != 0) {
7619 		/*
7620 		 * We keep a simple srtt in microseconds, like our rtt
7621 		 * measurement. We don't need to do any tricks with shifting
7622 		 * etc. Instead we just add in 1/8th of the new measurement
7623 		 * and subtract out 1/8 of the old srtt. We do the same with
7624 		 * the variance after finding the absolute value of the
7625 		 * difference between this sample and the current srtt.
7626 		 */
7627 		delta = tp->t_srtt - rtt;
7628 		/* Take off 1/8th of the current sRTT */
7629 		tp->t_srtt -= (tp->t_srtt >> 3);
7630 		/* Add in 1/8th of the new RTT just measured */
7631 		tp->t_srtt += (rtt >> 3);
7632 		if (tp->t_srtt <= 0)
7633 			tp->t_srtt = 1;
7634 		/* Now lets make the absolute value of the variance */
7635 		if (delta < 0)
7636 			delta = -delta;
7637 		/* Subtract out 1/8th */
7638 		tp->t_rttvar -= (tp->t_rttvar >> 3);
7639 		/* Add in 1/8th of the new variance we just saw */
7640 		tp->t_rttvar += (delta >> 3);
7641 		if (tp->t_rttvar <= 0)
7642 			tp->t_rttvar = 1;
7643 		if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7644 			tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7645 	} else {
7646 		/*
7647 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
7648 		 * variance to half the rtt (so our first retransmit happens
7649 		 * at 3*rtt).
7650 		 */
7651 		tp->t_srtt = rtt;
7652 		tp->t_rttvar = rtt >> 1;
7653 		tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7654 	}
7655 	rack->rc_srtt_measure_made = 1;
7656 	KMOD_TCPSTAT_INC(tcps_rttupdated);
7657 	tp->t_rttupdated++;
7658 #ifdef STATS
7659 	if (rack_stats_gets_ms_rtt == 0) {
7660 		/* Send in the microsecond rtt used for rxt timeout purposes */
7661 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7662 	} else if (rack_stats_gets_ms_rtt == 1) {
7663 		/* Send in the millisecond rtt used for rxt timeout purposes */
7664 		int32_t ms_rtt;
7665 
7666 		/* Round up */
7667 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7668 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7669 	} else if (rack_stats_gets_ms_rtt == 2) {
7670 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
7671 		int32_t ms_rtt;
7672 
7673 		/* Round up */
7674 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7675 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7676 	}  else {
7677 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
7678 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7679 	}
7680 
7681 #endif
7682 	/*
7683 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
7684 	 * way we do the smoothing, srtt and rttvar will each average +1/2
7685 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
7686 	 * tick of rounding and 1 extra tick because of +-1/2 tick
7687 	 * uncertainty in the firing of the timer.  The bias will give us
7688 	 * exactly the 1.5 tick we need.  But, because the bias is
7689 	 * statistical, we have to test that we don't drop below the minimum
7690 	 * feasible timer (which is 2 ticks).
7691 	 */
7692 	tp->t_rxtshift = 0;
7693 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7694 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7695 	rack_log_rtt_sample(rack, rtt);
7696 	tp->t_softerror = 0;
7697 }
7698 
7699 
7700 static void
7701 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7702 {
7703 	/*
7704 	 * Apply to filter the inbound us-rtt at us_cts.
7705 	 */
7706 	uint32_t old_rtt;
7707 
7708 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7709 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7710 			       us_rtt, us_cts);
7711 	if (old_rtt > us_rtt) {
7712 		/* We just hit a new lower rtt time */
7713 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7714 				     __LINE__, RACK_RTTS_NEWRTT);
7715 		/*
7716 		 * Only count it if its lower than what we saw within our
7717 		 * calculated range.
7718 		 */
7719 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7720 			if (rack_probertt_lower_within &&
7721 			    rack->rc_gp_dyn_mul &&
7722 			    (rack->use_fixed_rate == 0) &&
7723 			    (rack->rc_always_pace)) {
7724 				/*
7725 				 * We are seeing a new lower rtt very close
7726 				 * to the time that we would have entered probe-rtt.
7727 				 * This is probably due to the fact that a peer flow
7728 				 * has entered probe-rtt. Lets go in now too.
7729 				 */
7730 				uint32_t val;
7731 
7732 				val = rack_probertt_lower_within * rack_time_between_probertt;
7733 				val /= 100;
7734 				if ((rack->in_probe_rtt == 0)  &&
7735 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
7736 					rack_enter_probertt(rack, us_cts);
7737 				}
7738 			}
7739 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7740 		}
7741 	}
7742 }
7743 
7744 static int
7745 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7746     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7747 {
7748 	uint32_t us_rtt;
7749 	int32_t i, all;
7750 	uint32_t t, len_acked;
7751 
7752 	if ((rsm->r_flags & RACK_ACKED) ||
7753 	    (rsm->r_flags & RACK_WAS_ACKED))
7754 		/* Already done */
7755 		return (0);
7756 	if (rsm->r_no_rtt_allowed) {
7757 		/* Not allowed */
7758 		return (0);
7759 	}
7760 	if (ack_type == CUM_ACKED) {
7761 		if (SEQ_GT(th_ack, rsm->r_end)) {
7762 			len_acked = rsm->r_end - rsm->r_start;
7763 			all = 1;
7764 		} else {
7765 			len_acked = th_ack - rsm->r_start;
7766 			all = 0;
7767 		}
7768 	} else {
7769 		len_acked = rsm->r_end - rsm->r_start;
7770 		all = 0;
7771 	}
7772 	if (rsm->r_rtr_cnt == 1) {
7773 
7774 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7775 		if ((int)t <= 0)
7776 			t = 1;
7777 		if (!tp->t_rttlow || tp->t_rttlow > t)
7778 			tp->t_rttlow = t;
7779 		if (!rack->r_ctl.rc_rack_min_rtt ||
7780 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7781 			rack->r_ctl.rc_rack_min_rtt = t;
7782 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
7783 				rack->r_ctl.rc_rack_min_rtt = 1;
7784 			}
7785 		}
7786 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7787 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7788 		else
7789 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7790 		if (us_rtt == 0)
7791 			us_rtt = 1;
7792 		if (CC_ALGO(tp)->rttsample != NULL) {
7793 			/* Kick the RTT to the CC */
7794 			CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
7795 		}
7796 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7797 		if (ack_type == SACKED) {
7798 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7799 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7800 		} else {
7801 			/*
7802 			 * We need to setup what our confidence
7803 			 * is in this ack.
7804 			 *
7805 			 * If the rsm was app limited and it is
7806 			 * less than a mss in length (the end
7807 			 * of the send) then we have a gap. If we
7808 			 * were app limited but say we were sending
7809 			 * multiple MSS's then we are more confident
7810 			 * int it.
7811 			 *
7812 			 * When we are not app-limited then we see if
7813 			 * the rsm is being included in the current
7814 			 * measurement, we tell this by the app_limited_needs_set
7815 			 * flag.
7816 			 *
7817 			 * Note that being cwnd blocked is not applimited
7818 			 * as well as the pacing delay between packets which
7819 			 * are sending only 1 or 2 MSS's also will show up
7820 			 * in the RTT. We probably need to examine this algorithm
7821 			 * a bit more and enhance it to account for the delay
7822 			 * between rsm's. We could do that by saving off the
7823 			 * pacing delay of each rsm (in an rsm) and then
7824 			 * factoring that in somehow though for now I am
7825 			 * not sure how :)
7826 			 */
7827 			int calc_conf = 0;
7828 
7829 			if (rsm->r_flags & RACK_APP_LIMITED) {
7830 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7831 					calc_conf = 0;
7832 				else
7833 					calc_conf = 1;
7834 			} else if (rack->app_limited_needs_set == 0) {
7835 				calc_conf = 1;
7836 			} else {
7837 				calc_conf = 0;
7838 			}
7839 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7840 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7841 					    calc_conf, rsm, rsm->r_rtr_cnt);
7842 		}
7843 		if ((rsm->r_flags & RACK_TLP) &&
7844 		    (!IN_FASTRECOVERY(tp->t_flags))) {
7845 			/* Segment was a TLP and our retrans matched */
7846 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7847 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
7848 			}
7849 		}
7850 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7851 			/* New more recent rack_tmit_time */
7852 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7853 			rack->rc_rack_rtt = t;
7854 		}
7855 		return (1);
7856 	}
7857 	/*
7858 	 * We clear the soft/rxtshift since we got an ack.
7859 	 * There is no assurance we will call the commit() function
7860 	 * so we need to clear these to avoid incorrect handling.
7861 	 */
7862 	tp->t_rxtshift = 0;
7863 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7864 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7865 	tp->t_softerror = 0;
7866 	if (to && (to->to_flags & TOF_TS) &&
7867 	    (ack_type == CUM_ACKED) &&
7868 	    (to->to_tsecr) &&
7869 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7870 		/*
7871 		 * Now which timestamp does it match? In this block the ACK
7872 		 * must be coming from a previous transmission.
7873 		 */
7874 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
7875 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7876 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7877 				if ((int)t <= 0)
7878 					t = 1;
7879 				if (CC_ALGO(tp)->rttsample != NULL) {
7880 					/*
7881 					 * Kick the RTT to the CC, here
7882 					 * we lie a bit in that we know the
7883 					 * retransmission is correct even though
7884 					 * we retransmitted. This is because
7885 					 * we match the timestamps.
7886 					 */
7887 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
7888 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
7889 					else
7890 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
7891 					CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
7892 				}
7893 				if ((i + 1) < rsm->r_rtr_cnt) {
7894 					/*
7895 					 * The peer ack'd from our previous
7896 					 * transmission. We have a spurious
7897 					 * retransmission and thus we dont
7898 					 * want to update our rack_rtt.
7899 					 *
7900 					 * Hmm should there be a CC revert here?
7901 					 *
7902 					 */
7903 					return (0);
7904 				}
7905 				if (!tp->t_rttlow || tp->t_rttlow > t)
7906 					tp->t_rttlow = t;
7907 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7908 					rack->r_ctl.rc_rack_min_rtt = t;
7909 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
7910 						rack->r_ctl.rc_rack_min_rtt = 1;
7911 					}
7912 				}
7913 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7914 					   (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7915 					/* New more recent rack_tmit_time */
7916 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7917 					rack->rc_rack_rtt = t;
7918 				}
7919 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7920 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7921 						    rsm->r_rtr_cnt);
7922 				return (1);
7923 			}
7924 		}
7925 		goto ts_not_found;
7926 	} else {
7927 		/*
7928 		 * Ok its a SACK block that we retransmitted. or a windows
7929 		 * machine without timestamps. We can tell nothing from the
7930 		 * time-stamp since its not there or the time the peer last
7931 		 * recieved a segment that moved forward its cum-ack point.
7932 		 */
7933 ts_not_found:
7934 		i = rsm->r_rtr_cnt - 1;
7935 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7936 		if ((int)t <= 0)
7937 			t = 1;
7938 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7939 			/*
7940 			 * We retransmitted and the ack came back in less
7941 			 * than the smallest rtt we have observed. We most
7942 			 * likely did an improper retransmit as outlined in
7943 			 * 6.2 Step 2 point 2 in the rack-draft so we
7944 			 * don't want to update our rack_rtt. We in
7945 			 * theory (in future) might want to think about reverting our
7946 			 * cwnd state but we won't for now.
7947 			 */
7948 			return (0);
7949 		} else if (rack->r_ctl.rc_rack_min_rtt) {
7950 			/*
7951 			 * We retransmitted it and the retransmit did the
7952 			 * job.
7953 			 */
7954 			if (!rack->r_ctl.rc_rack_min_rtt ||
7955 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7956 				rack->r_ctl.rc_rack_min_rtt = t;
7957 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
7958 					rack->r_ctl.rc_rack_min_rtt = 1;
7959 				}
7960 			}
7961 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7962 				/* New more recent rack_tmit_time */
7963 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7964 				rack->rc_rack_rtt = t;
7965 			}
7966 			return (1);
7967 		}
7968 	}
7969 	return (0);
7970 }
7971 
7972 /*
7973  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7974  */
7975 static void
7976 rack_log_sack_passed(struct tcpcb *tp,
7977     struct tcp_rack *rack, struct rack_sendmap *rsm)
7978 {
7979 	struct rack_sendmap *nrsm;
7980 
7981 	nrsm = rsm;
7982 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7983 	    rack_head, r_tnext) {
7984 		if (nrsm == rsm) {
7985 			/* Skip orginal segment he is acked */
7986 			continue;
7987 		}
7988 		if (nrsm->r_flags & RACK_ACKED) {
7989 			/*
7990 			 * Skip ack'd segments, though we
7991 			 * should not see these, since tmap
7992 			 * should not have ack'd segments.
7993 			 */
7994 			continue;
7995 		}
7996 		if (nrsm->r_flags & RACK_SACK_PASSED) {
7997 			/*
7998 			 * We found one that is already marked
7999 			 * passed, we have been here before and
8000 			 * so all others below this are marked.
8001 			 */
8002 			break;
8003 		}
8004 		nrsm->r_flags |= RACK_SACK_PASSED;
8005 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8006 	}
8007 }
8008 
8009 static void
8010 rack_need_set_test(struct tcpcb *tp,
8011 		   struct tcp_rack *rack,
8012 		   struct rack_sendmap *rsm,
8013 		   tcp_seq th_ack,
8014 		   int line,
8015 		   int use_which)
8016 {
8017 
8018 	if ((tp->t_flags & TF_GPUTINPROG) &&
8019 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8020 		/*
8021 		 * We were app limited, and this ack
8022 		 * butts up or goes beyond the point where we want
8023 		 * to start our next measurement. We need
8024 		 * to record the new gput_ts as here and
8025 		 * possibly update the start sequence.
8026 		 */
8027 		uint32_t seq, ts;
8028 
8029 		if (rsm->r_rtr_cnt > 1) {
8030 			/*
8031 			 * This is a retransmit, can we
8032 			 * really make any assessment at this
8033 			 * point?  We are not really sure of
8034 			 * the timestamp, is it this or the
8035 			 * previous transmission?
8036 			 *
8037 			 * Lets wait for something better that
8038 			 * is not retransmitted.
8039 			 */
8040 			return;
8041 		}
8042 		seq = tp->gput_seq;
8043 		ts = tp->gput_ts;
8044 		rack->app_limited_needs_set = 0;
8045 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8046 		/* Do we start at a new end? */
8047 		if ((use_which == RACK_USE_BEG) &&
8048 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8049 			/*
8050 			 * When we get an ACK that just eats
8051 			 * up some of the rsm, we set RACK_USE_BEG
8052 			 * since whats at r_start (i.e. th_ack)
8053 			 * is left unacked and thats where the
8054 			 * measurement not starts.
8055 			 */
8056 			tp->gput_seq = rsm->r_start;
8057 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8058 		}
8059 		if ((use_which == RACK_USE_END) &&
8060 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8061 			    /*
8062 			     * We use the end when the cumack
8063 			     * is moving forward and completely
8064 			     * deleting the rsm passed so basically
8065 			     * r_end holds th_ack.
8066 			     *
8067 			     * For SACK's we also want to use the end
8068 			     * since this piece just got sacked and
8069 			     * we want to target anything after that
8070 			     * in our measurement.
8071 			     */
8072 			    tp->gput_seq = rsm->r_end;
8073 			    rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8074 		}
8075 		if (use_which == RACK_USE_END_OR_THACK) {
8076 			/*
8077 			 * special case for ack moving forward,
8078 			 * not a sack, we need to move all the
8079 			 * way up to where this ack cum-ack moves
8080 			 * to.
8081 			 */
8082 			if (SEQ_GT(th_ack, rsm->r_end))
8083 				tp->gput_seq = th_ack;
8084 			else
8085 				tp->gput_seq = rsm->r_end;
8086 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8087 		}
8088 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8089 			/*
8090 			 * We moved beyond this guy's range, re-calculate
8091 			 * the new end point.
8092 			 */
8093 			if (rack->rc_gp_filled == 0) {
8094 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8095 			} else {
8096 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8097 			}
8098 		}
8099 		/*
8100 		 * We are moving the goal post, we may be able to clear the
8101 		 * measure_saw_probe_rtt flag.
8102 		 */
8103 		if ((rack->in_probe_rtt == 0) &&
8104 		    (rack->measure_saw_probe_rtt) &&
8105 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8106 			rack->measure_saw_probe_rtt = 0;
8107 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8108 					   seq, tp->gput_seq, 0, 5, line, NULL, 0);
8109 		if (rack->rc_gp_filled &&
8110 		    ((tp->gput_ack - tp->gput_seq) <
8111 		     max(rc_init_window(rack), (MIN_GP_WIN *
8112 						ctf_fixed_maxseg(tp))))) {
8113 			uint32_t ideal_amount;
8114 
8115 			ideal_amount = rack_get_measure_window(tp, rack);
8116 			if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8117 				/*
8118 				 * There is no sense of continuing this measurement
8119 				 * because its too small to gain us anything we
8120 				 * trust. Skip it and that way we can start a new
8121 				 * measurement quicker.
8122 				 */
8123 				tp->t_flags &= ~TF_GPUTINPROG;
8124 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8125 							   0, 0, 0, 6, __LINE__, NULL, 0);
8126 			} else {
8127 				/*
8128 				 * Reset the window further out.
8129 				 */
8130 				tp->gput_ack = tp->gput_seq + ideal_amount;
8131 			}
8132 		}
8133 	}
8134 }
8135 
8136 static inline int
8137 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8138 {
8139 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8140 		/* Behind our TLP definition or right at */
8141 		return (0);
8142 	}
8143 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8144 		/* The start is beyond or right at our end of TLP definition */
8145 		return (0);
8146 	}
8147 	/* It has to be a sub-part of the original TLP recorded */
8148 	return (1);
8149 }
8150 
8151 
8152 static uint32_t
8153 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8154 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8155 {
8156 	uint32_t start, end, changed = 0;
8157 	struct rack_sendmap stack_map;
8158 	struct rack_sendmap *rsm, *nrsm, fe, *prev, *next;
8159 #ifdef INVARIANTS
8160 	struct rack_sendmap *insret;
8161 #endif
8162 	int32_t used_ref = 1;
8163 	int moved = 0;
8164 
8165 	start = sack->start;
8166 	end = sack->end;
8167 	rsm = *prsm;
8168 	memset(&fe, 0, sizeof(fe));
8169 do_rest_ofb:
8170 	if ((rsm == NULL) ||
8171 	    (SEQ_LT(end, rsm->r_start)) ||
8172 	    (SEQ_GEQ(start, rsm->r_end)) ||
8173 	    (SEQ_LT(start, rsm->r_start))) {
8174 		/*
8175 		 * We are not in the right spot,
8176 		 * find the correct spot in the tree.
8177 		 */
8178 		used_ref = 0;
8179 		fe.r_start = start;
8180 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8181 		moved++;
8182 	}
8183 	if (rsm == NULL) {
8184 		/* TSNH */
8185 		goto out;
8186 	}
8187 	/* Ok we have an ACK for some piece of this rsm */
8188 	if (rsm->r_start != start) {
8189 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8190 			/*
8191 			 * Before any splitting or hookery is
8192 			 * done is it a TLP of interest i.e. rxt?
8193 			 */
8194 			if ((rsm->r_flags & RACK_TLP) &&
8195 			    (rsm->r_rtr_cnt > 1)) {
8196 				/*
8197 				 * We are splitting a rxt TLP, check
8198 				 * if we need to save off the start/end
8199 				 */
8200 				if (rack->rc_last_tlp_acked_set &&
8201 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8202 					/*
8203 					 * We already turned this on since we are inside
8204 					 * the previous one was a partially sack now we
8205 					 * are getting another one (maybe all of it).
8206 					 *
8207 					 */
8208 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8209 					/*
8210 					 * Lets make sure we have all of it though.
8211 					 */
8212 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8213 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8214 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8215 								     rack->r_ctl.last_tlp_acked_end);
8216 					}
8217 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8218 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8219 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8220 								     rack->r_ctl.last_tlp_acked_end);
8221 					}
8222 				} else {
8223 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8224 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8225 					rack->rc_last_tlp_past_cumack = 0;
8226 					rack->rc_last_tlp_acked_set = 1;
8227 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8228 				}
8229 			}
8230 			/**
8231 			 * Need to split this in two pieces the before and after,
8232 			 * the before remains in the map, the after must be
8233 			 * added. In other words we have:
8234 			 * rsm        |--------------|
8235 			 * sackblk        |------->
8236 			 * rsm will become
8237 			 *     rsm    |---|
8238 			 * and nrsm will be  the sacked piece
8239 			 *     nrsm       |----------|
8240 			 *
8241 			 * But before we start down that path lets
8242 			 * see if the sack spans over on top of
8243 			 * the next guy and it is already sacked.
8244 			 *
8245 			 */
8246 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8247 			if (next && (next->r_flags & RACK_ACKED) &&
8248 			    SEQ_GEQ(end, next->r_start)) {
8249 				/**
8250 				 * So the next one is already acked, and
8251 				 * we can thus by hookery use our stack_map
8252 				 * to reflect the piece being sacked and
8253 				 * then adjust the two tree entries moving
8254 				 * the start and ends around. So we start like:
8255 				 *  rsm     |------------|             (not-acked)
8256 				 *  next                 |-----------| (acked)
8257 				 *  sackblk        |-------->
8258 				 *  We want to end like so:
8259 				 *  rsm     |------|                   (not-acked)
8260 				 *  next           |-----------------| (acked)
8261 				 *  nrsm           |-----|
8262 				 * Where nrsm is a temporary stack piece we
8263 				 * use to update all the gizmos.
8264 				 */
8265 				/* Copy up our fudge block */
8266 				nrsm = &stack_map;
8267 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8268 				/* Now adjust our tree blocks */
8269 				rsm->r_end = start;
8270 				next->r_start = start;
8271 				/* Now we must adjust back where next->m is */
8272 				rack_setup_offset_for_rsm(rsm, next);
8273 
8274 				/* We don't need to adjust rsm, it did not change */
8275 				/* Clear out the dup ack count of the remainder */
8276 				rsm->r_dupack = 0;
8277 				rsm->r_just_ret = 0;
8278 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8279 				/* Now lets make sure our fudge block is right */
8280 				nrsm->r_start = start;
8281 				/* Now lets update all the stats and such */
8282 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8283 				if (rack->app_limited_needs_set)
8284 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8285 				changed += (nrsm->r_end - nrsm->r_start);
8286 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8287 				if (nrsm->r_flags & RACK_SACK_PASSED) {
8288 					rack->r_ctl.rc_reorder_ts = cts;
8289 				}
8290 				/*
8291 				 * Now we want to go up from rsm (the
8292 				 * one left un-acked) to the next one
8293 				 * in the tmap. We do this so when
8294 				 * we walk backwards we include marking
8295 				 * sack-passed on rsm (The one passed in
8296 				 * is skipped since it is generally called
8297 				 * on something sacked before removing it
8298 				 * from the tmap).
8299 				 */
8300 				if (rsm->r_in_tmap) {
8301 					nrsm = TAILQ_NEXT(rsm, r_tnext);
8302 					/*
8303 					 * Now that we have the next
8304 					 * one walk backwards from there.
8305 					 */
8306 					if (nrsm && nrsm->r_in_tmap)
8307 						rack_log_sack_passed(tp, rack, nrsm);
8308 				}
8309 				/* Now are we done? */
8310 				if (SEQ_LT(end, next->r_end) ||
8311 				    (end == next->r_end)) {
8312 					/* Done with block */
8313 					goto out;
8314 				}
8315 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8316 				counter_u64_add(rack_sack_used_next_merge, 1);
8317 				/* Postion for the next block */
8318 				start = next->r_end;
8319 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8320 				if (rsm == NULL)
8321 					goto out;
8322 			} else {
8323 				/**
8324 				 * We can't use any hookery here, so we
8325 				 * need to split the map. We enter like
8326 				 * so:
8327 				 *  rsm      |--------|
8328 				 *  sackblk       |----->
8329 				 * We will add the new block nrsm and
8330 				 * that will be the new portion, and then
8331 				 * fall through after reseting rsm. So we
8332 				 * split and look like this:
8333 				 *  rsm      |----|
8334 				 *  sackblk       |----->
8335 				 *  nrsm          |---|
8336 				 * We then fall through reseting
8337 				 * rsm to nrsm, so the next block
8338 				 * picks it up.
8339 				 */
8340 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8341 				if (nrsm == NULL) {
8342 					/*
8343 					 * failed XXXrrs what can we do but loose the sack
8344 					 * info?
8345 					 */
8346 					goto out;
8347 				}
8348 				counter_u64_add(rack_sack_splits, 1);
8349 				rack_clone_rsm(rack, nrsm, rsm, start);
8350 				rsm->r_just_ret = 0;
8351 #ifndef INVARIANTS
8352 				(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8353 #else
8354 				insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8355 				if (insret != NULL) {
8356 					panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8357 					      nrsm, insret, rack, rsm);
8358 				}
8359 #endif
8360 				if (rsm->r_in_tmap) {
8361 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8362 					nrsm->r_in_tmap = 1;
8363 				}
8364 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8365 				rsm->r_flags &= (~RACK_HAS_FIN);
8366 				/* Position us to point to the new nrsm that starts the sack blk */
8367 				rsm = nrsm;
8368 			}
8369 		} else {
8370 			/* Already sacked this piece */
8371 			counter_u64_add(rack_sack_skipped_acked, 1);
8372 			moved++;
8373 			if (end == rsm->r_end) {
8374 				/* Done with block */
8375 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8376 				goto out;
8377 			} else if (SEQ_LT(end, rsm->r_end)) {
8378 				/* A partial sack to a already sacked block */
8379 				moved++;
8380 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8381 				goto out;
8382 			} else {
8383 				/*
8384 				 * The end goes beyond this guy
8385 				 * repostion the start to the
8386 				 * next block.
8387 				 */
8388 				start = rsm->r_end;
8389 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8390 				if (rsm == NULL)
8391 					goto out;
8392 			}
8393 		}
8394 	}
8395 	if (SEQ_GEQ(end, rsm->r_end)) {
8396 		/**
8397 		 * The end of this block is either beyond this guy or right
8398 		 * at this guy. I.e.:
8399 		 *  rsm ---                 |-----|
8400 		 *  end                     |-----|
8401 		 *  <or>
8402 		 *  end                     |---------|
8403 		 */
8404 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8405 			/*
8406 			 * Is it a TLP of interest?
8407 			 */
8408 			if ((rsm->r_flags & RACK_TLP) &&
8409 			    (rsm->r_rtr_cnt > 1)) {
8410 				/*
8411 				 * We are splitting a rxt TLP, check
8412 				 * if we need to save off the start/end
8413 				 */
8414 				if (rack->rc_last_tlp_acked_set &&
8415 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8416 					/*
8417 					 * We already turned this on since we are inside
8418 					 * the previous one was a partially sack now we
8419 					 * are getting another one (maybe all of it).
8420 					 */
8421 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8422 					/*
8423 					 * Lets make sure we have all of it though.
8424 					 */
8425 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8426 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8427 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8428 								     rack->r_ctl.last_tlp_acked_end);
8429 					}
8430 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8431 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8432 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8433 								     rack->r_ctl.last_tlp_acked_end);
8434 					}
8435 				} else {
8436 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8437 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8438 					rack->rc_last_tlp_past_cumack = 0;
8439 					rack->rc_last_tlp_acked_set = 1;
8440 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8441 				}
8442 			}
8443 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8444 			changed += (rsm->r_end - rsm->r_start);
8445 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8446 			if (rsm->r_in_tmap) /* should be true */
8447 				rack_log_sack_passed(tp, rack, rsm);
8448 			/* Is Reordering occuring? */
8449 			if (rsm->r_flags & RACK_SACK_PASSED) {
8450 				rsm->r_flags &= ~RACK_SACK_PASSED;
8451 				rack->r_ctl.rc_reorder_ts = cts;
8452 			}
8453 			if (rack->app_limited_needs_set)
8454 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8455 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8456 			rsm->r_flags |= RACK_ACKED;
8457 			if (rsm->r_in_tmap) {
8458 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8459 				rsm->r_in_tmap = 0;
8460 			}
8461 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8462 		} else {
8463 			counter_u64_add(rack_sack_skipped_acked, 1);
8464 			moved++;
8465 		}
8466 		if (end == rsm->r_end) {
8467 			/* This block only - done, setup for next */
8468 			goto out;
8469 		}
8470 		/*
8471 		 * There is more not coverend by this rsm move on
8472 		 * to the next block in the RB tree.
8473 		 */
8474 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8475 		start = rsm->r_end;
8476 		rsm = nrsm;
8477 		if (rsm == NULL)
8478 			goto out;
8479 		goto do_rest_ofb;
8480 	}
8481 	/**
8482 	 * The end of this sack block is smaller than
8483 	 * our rsm i.e.:
8484 	 *  rsm ---                 |-----|
8485 	 *  end                     |--|
8486 	 */
8487 	if ((rsm->r_flags & RACK_ACKED) == 0) {
8488 		/*
8489 		 * Is it a TLP of interest?
8490 		 */
8491 		if ((rsm->r_flags & RACK_TLP) &&
8492 		    (rsm->r_rtr_cnt > 1)) {
8493 			/*
8494 			 * We are splitting a rxt TLP, check
8495 			 * if we need to save off the start/end
8496 			 */
8497 			if (rack->rc_last_tlp_acked_set &&
8498 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8499 				/*
8500 				 * We already turned this on since we are inside
8501 				 * the previous one was a partially sack now we
8502 				 * are getting another one (maybe all of it).
8503 				 */
8504 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8505 				/*
8506 				 * Lets make sure we have all of it though.
8507 				 */
8508 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8509 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8510 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8511 							     rack->r_ctl.last_tlp_acked_end);
8512 				}
8513 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8514 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8515 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8516 							     rack->r_ctl.last_tlp_acked_end);
8517 				}
8518 			} else {
8519 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8520 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8521 				rack->rc_last_tlp_past_cumack = 0;
8522 				rack->rc_last_tlp_acked_set = 1;
8523 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8524 			}
8525 		}
8526 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8527 		if (prev &&
8528 		    (prev->r_flags & RACK_ACKED)) {
8529 			/**
8530 			 * Goal, we want the right remainder of rsm to shrink
8531 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
8532 			 * We want to expand prev to go all the way
8533 			 * to prev->r_end <- end.
8534 			 * so in the tree we have before:
8535 			 *   prev     |--------|         (acked)
8536 			 *   rsm               |-------| (non-acked)
8537 			 *   sackblk           |-|
8538 			 * We churn it so we end up with
8539 			 *   prev     |----------|       (acked)
8540 			 *   rsm                 |-----| (non-acked)
8541 			 *   nrsm              |-| (temporary)
8542 			 *
8543 			 * Note if either prev/rsm is a TLP we don't
8544 			 * do this.
8545 			 */
8546 			nrsm = &stack_map;
8547 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8548 			prev->r_end = end;
8549 			rsm->r_start = end;
8550 			/* Now adjust nrsm (stack copy) to be
8551 			 * the one that is the small
8552 			 * piece that was "sacked".
8553 			 */
8554 			nrsm->r_end = end;
8555 			rsm->r_dupack = 0;
8556 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8557 			/*
8558 			 * Now that the rsm has had its start moved forward
8559 			 * lets go ahead and get its new place in the world.
8560 			 */
8561 			rack_setup_offset_for_rsm(prev, rsm);
8562 			/*
8563 			 * Now nrsm is our new little piece
8564 			 * that is acked (which was merged
8565 			 * to prev). Update the rtt and changed
8566 			 * based on that. Also check for reordering.
8567 			 */
8568 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8569 			if (rack->app_limited_needs_set)
8570 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8571 			changed += (nrsm->r_end - nrsm->r_start);
8572 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8573 			if (nrsm->r_flags & RACK_SACK_PASSED) {
8574 				rack->r_ctl.rc_reorder_ts = cts;
8575 			}
8576 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8577 			rsm = prev;
8578 			counter_u64_add(rack_sack_used_prev_merge, 1);
8579 		} else {
8580 			/**
8581 			 * This is the case where our previous
8582 			 * block is not acked either, so we must
8583 			 * split the block in two.
8584 			 */
8585 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8586 			if (nrsm == NULL) {
8587 				/* failed rrs what can we do but loose the sack info? */
8588 				goto out;
8589 			}
8590 			if ((rsm->r_flags & RACK_TLP) &&
8591 			    (rsm->r_rtr_cnt > 1)) {
8592 				/*
8593 				 * We are splitting a rxt TLP, check
8594 				 * if we need to save off the start/end
8595 				 */
8596 				if (rack->rc_last_tlp_acked_set &&
8597 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8598 					    /*
8599 					     * We already turned this on since this block is inside
8600 					     * the previous one was a partially sack now we
8601 					     * are getting another one (maybe all of it).
8602 					     */
8603 					    rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8604 					    /*
8605 					     * Lets make sure we have all of it though.
8606 					     */
8607 					    if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8608 						    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8609 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8610 									 rack->r_ctl.last_tlp_acked_end);
8611 					    }
8612 					    if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8613 						    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8614 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8615 									 rack->r_ctl.last_tlp_acked_end);
8616 					    }
8617 				    } else {
8618 					    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8619 					    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8620 					    rack->rc_last_tlp_acked_set = 1;
8621 					    rack->rc_last_tlp_past_cumack = 0;
8622 					    rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8623 				    }
8624 			}
8625 			/**
8626 			 * In this case nrsm becomes
8627 			 * nrsm->r_start = end;
8628 			 * nrsm->r_end = rsm->r_end;
8629 			 * which is un-acked.
8630 			 * <and>
8631 			 * rsm->r_end = nrsm->r_start;
8632 			 * i.e. the remaining un-acked
8633 			 * piece is left on the left
8634 			 * hand side.
8635 			 *
8636 			 * So we start like this
8637 			 * rsm      |----------| (not acked)
8638 			 * sackblk  |---|
8639 			 * build it so we have
8640 			 * rsm      |---|         (acked)
8641 			 * nrsm         |------|  (not acked)
8642 			 */
8643 			counter_u64_add(rack_sack_splits, 1);
8644 			rack_clone_rsm(rack, nrsm, rsm, end);
8645 			rsm->r_flags &= (~RACK_HAS_FIN);
8646 			rsm->r_just_ret = 0;
8647 #ifndef INVARIANTS
8648 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8649 #else
8650 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8651 			if (insret != NULL) {
8652 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8653 				      nrsm, insret, rack, rsm);
8654 			}
8655 #endif
8656 			if (rsm->r_in_tmap) {
8657 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8658 				nrsm->r_in_tmap = 1;
8659 			}
8660 			nrsm->r_dupack = 0;
8661 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8662 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8663 			changed += (rsm->r_end - rsm->r_start);
8664 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8665 			if (rsm->r_in_tmap) /* should be true */
8666 				rack_log_sack_passed(tp, rack, rsm);
8667 			/* Is Reordering occuring? */
8668 			if (rsm->r_flags & RACK_SACK_PASSED) {
8669 				rsm->r_flags &= ~RACK_SACK_PASSED;
8670 				rack->r_ctl.rc_reorder_ts = cts;
8671 			}
8672 			if (rack->app_limited_needs_set)
8673 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8674 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8675 			rsm->r_flags |= RACK_ACKED;
8676 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8677 			if (rsm->r_in_tmap) {
8678 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8679 				rsm->r_in_tmap = 0;
8680 			}
8681 		}
8682 	} else if (start != end){
8683 		/*
8684 		 * The block was already acked.
8685 		 */
8686 		counter_u64_add(rack_sack_skipped_acked, 1);
8687 		moved++;
8688 	}
8689 out:
8690 	if (rsm &&
8691 	    ((rsm->r_flags & RACK_TLP) == 0) &&
8692 	    (rsm->r_flags & RACK_ACKED)) {
8693 		/*
8694 		 * Now can we merge where we worked
8695 		 * with either the previous or
8696 		 * next block?
8697 		 */
8698 		next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8699 		while (next) {
8700 			if (next->r_flags & RACK_TLP)
8701 				break;
8702 			if (next->r_flags & RACK_ACKED) {
8703 			/* yep this and next can be merged */
8704 				rsm = rack_merge_rsm(rack, rsm, next);
8705 				next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8706 			} else
8707 				break;
8708 		}
8709 		/* Now what about the previous? */
8710 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8711 		while (prev) {
8712 			if (prev->r_flags & RACK_TLP)
8713 				break;
8714 			if (prev->r_flags & RACK_ACKED) {
8715 				/* yep the previous and this can be merged */
8716 				rsm = rack_merge_rsm(rack, prev, rsm);
8717 				prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8718 			} else
8719 				break;
8720 		}
8721 	}
8722 	if (used_ref == 0) {
8723 		counter_u64_add(rack_sack_proc_all, 1);
8724 	} else {
8725 		counter_u64_add(rack_sack_proc_short, 1);
8726 	}
8727 	/* Save off the next one for quick reference. */
8728 	if (rsm)
8729 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8730 	else
8731 		nrsm = NULL;
8732 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
8733 	/* Pass back the moved. */
8734 	*moved_two = moved;
8735 	return (changed);
8736 }
8737 
8738 static void inline
8739 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8740 {
8741 	struct rack_sendmap *tmap;
8742 
8743 	tmap = NULL;
8744 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
8745 		/* Its no longer sacked, mark it so */
8746 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8747 #ifdef INVARIANTS
8748 		if (rsm->r_in_tmap) {
8749 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
8750 			      rack, rsm, rsm->r_flags);
8751 		}
8752 #endif
8753 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8754 		/* Rebuild it into our tmap */
8755 		if (tmap == NULL) {
8756 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8757 			tmap = rsm;
8758 		} else {
8759 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8760 			tmap = rsm;
8761 		}
8762 		tmap->r_in_tmap = 1;
8763 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8764 	}
8765 	/*
8766 	 * Now lets possibly clear the sack filter so we start
8767 	 * recognizing sacks that cover this area.
8768 	 */
8769 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8770 
8771 }
8772 
8773 static void
8774 rack_do_decay(struct tcp_rack *rack)
8775 {
8776 	struct timeval res;
8777 
8778 #define	timersub(tvp, uvp, vvp)						\
8779 	do {								\
8780 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
8781 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
8782 		if ((vvp)->tv_usec < 0) {				\
8783 			(vvp)->tv_sec--;				\
8784 			(vvp)->tv_usec += 1000000;			\
8785 		}							\
8786 	} while (0)
8787 
8788 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8789 #undef timersub
8790 
8791 	rack->r_ctl.input_pkt++;
8792 	if ((rack->rc_in_persist) ||
8793 	    (res.tv_sec >= 1) ||
8794 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8795 		/*
8796 		 * Check for decay of non-SAD,
8797 		 * we want all SAD detection metrics to
8798 		 * decay 1/4 per second (or more) passed.
8799 		 */
8800 #ifdef NETFLIX_EXP_DETECTION
8801 		uint32_t pkt_delta;
8802 
8803 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8804 #endif
8805 		/* Update our saved tracking values */
8806 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8807 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8808 		/* Now do we escape without decay? */
8809 #ifdef NETFLIX_EXP_DETECTION
8810 		if (rack->rc_in_persist ||
8811 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8812 		    (pkt_delta < tcp_sad_low_pps)){
8813 			/*
8814 			 * We don't decay idle connections
8815 			 * or ones that have a low input pps.
8816 			 */
8817 			return;
8818 		}
8819 		/* Decay the counters */
8820 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8821 							tcp_sad_decay_val);
8822 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8823 							 tcp_sad_decay_val);
8824 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8825 							       tcp_sad_decay_val);
8826 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8827 								tcp_sad_decay_val);
8828 #endif
8829 	}
8830 }
8831 
8832 static void
8833 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8834 {
8835 	struct rack_sendmap *rsm;
8836 #ifdef INVARIANTS
8837 	struct rack_sendmap *rm;
8838 #endif
8839 
8840 	/*
8841 	 * The ACK point is advancing to th_ack, we must drop off
8842 	 * the packets in the rack log and calculate any eligble
8843 	 * RTT's.
8844 	 */
8845 	rack->r_wanted_output = 1;
8846 
8847 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
8848 	if ((rack->rc_last_tlp_acked_set == 1)&&
8849 	    (rack->rc_last_tlp_past_cumack == 1) &&
8850 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
8851 		/*
8852 		 * We have reached the point where our last rack
8853 		 * tlp retransmit sequence is ahead of the cum-ack.
8854 		 * This can only happen when the cum-ack moves all
8855 		 * the way around (its been a full 2^^31+1 bytes
8856 		 * or more since we sent a retransmitted TLP). Lets
8857 		 * turn off the valid flag since its not really valid.
8858 		 *
8859 		 * Note since sack's also turn on this event we have
8860 		 * a complication, we have to wait to age it out until
8861 		 * the cum-ack is by the TLP before checking which is
8862 		 * what the next else clause does.
8863 		 */
8864 		rack_log_dsack_event(rack, 9, __LINE__,
8865 				     rack->r_ctl.last_tlp_acked_start,
8866 				     rack->r_ctl.last_tlp_acked_end);
8867 		rack->rc_last_tlp_acked_set = 0;
8868 		rack->rc_last_tlp_past_cumack = 0;
8869 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
8870 		   (rack->rc_last_tlp_past_cumack == 0) &&
8871 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
8872 		/*
8873 		 * It is safe to start aging TLP's out.
8874 		 */
8875 		rack->rc_last_tlp_past_cumack = 1;
8876 	}
8877 	/* We do the same for the tlp send seq as well */
8878 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8879 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
8880 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
8881 		rack_log_dsack_event(rack, 9, __LINE__,
8882 				     rack->r_ctl.last_sent_tlp_seq,
8883 				     (rack->r_ctl.last_sent_tlp_seq +
8884 				      rack->r_ctl.last_sent_tlp_len));
8885 		rack->rc_last_sent_tlp_seq_valid = 0;
8886 		rack->rc_last_sent_tlp_past_cumack = 0;
8887 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8888 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
8889 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
8890 		/*
8891 		 * It is safe to start aging TLP's send.
8892 		 */
8893 		rack->rc_last_sent_tlp_past_cumack = 1;
8894 	}
8895 more:
8896 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8897 	if (rsm == NULL) {
8898 		if ((th_ack - 1) == tp->iss) {
8899 			/*
8900 			 * For the SYN incoming case we will not
8901 			 * have called tcp_output for the sending of
8902 			 * the SYN, so there will be no map. All
8903 			 * other cases should probably be a panic.
8904 			 */
8905 			return;
8906 		}
8907 		if (tp->t_flags & TF_SENTFIN) {
8908 			/* if we sent a FIN we often will not have map */
8909 			return;
8910 		}
8911 #ifdef INVARIANTS
8912 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8913 		      tp,
8914 		      tp->t_state, th_ack, rack,
8915 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
8916 #endif
8917 		return;
8918 	}
8919 	if (SEQ_LT(th_ack, rsm->r_start)) {
8920 		/* Huh map is missing this */
8921 #ifdef INVARIANTS
8922 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8923 		       rsm->r_start,
8924 		       th_ack, tp->t_state, rack->r_state);
8925 #endif
8926 		return;
8927 	}
8928 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8929 
8930 	/* Now was it a retransmitted TLP? */
8931 	if ((rsm->r_flags & RACK_TLP) &&
8932 	    (rsm->r_rtr_cnt > 1)) {
8933 		/*
8934 		 * Yes, this rsm was a TLP and retransmitted, remember that
8935 		 * since if a DSACK comes back on this we don't want
8936 		 * to think of it as a reordered segment. This may
8937 		 * get updated again with possibly even other TLPs
8938 		 * in flight, but thats ok. Only when we don't send
8939 		 * a retransmitted TLP for 1/2 the sequences space
8940 		 * will it get turned off (above).
8941 		 */
8942 		if (rack->rc_last_tlp_acked_set &&
8943 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8944 			/*
8945 			 * We already turned this on since the end matches,
8946 			 * the previous one was a partially ack now we
8947 			 * are getting another one (maybe all of it).
8948 			 */
8949 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8950 			/*
8951 			 * Lets make sure we have all of it though.
8952 			 */
8953 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8954 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8955 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8956 						     rack->r_ctl.last_tlp_acked_end);
8957 			}
8958 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8959 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8960 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8961 						     rack->r_ctl.last_tlp_acked_end);
8962 			}
8963 		} else {
8964 			rack->rc_last_tlp_past_cumack = 1;
8965 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8966 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8967 			rack->rc_last_tlp_acked_set = 1;
8968 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8969 		}
8970 	}
8971 	/* Now do we consume the whole thing? */
8972 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
8973 		/* Its all consumed. */
8974 		uint32_t left;
8975 		uint8_t newly_acked;
8976 
8977 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8978 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8979 		rsm->r_rtr_bytes = 0;
8980 		/* Record the time of highest cumack sent */
8981 		rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8982 #ifndef INVARIANTS
8983 		(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8984 #else
8985 		rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8986 		if (rm != rsm) {
8987 			panic("removing head in rack:%p rsm:%p rm:%p",
8988 			      rack, rsm, rm);
8989 		}
8990 #endif
8991 		if (rsm->r_in_tmap) {
8992 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8993 			rsm->r_in_tmap = 0;
8994 		}
8995 		newly_acked = 1;
8996 		if (rsm->r_flags & RACK_ACKED) {
8997 			/*
8998 			 * It was acked on the scoreboard -- remove
8999 			 * it from total
9000 			 */
9001 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
9002 			newly_acked = 0;
9003 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
9004 			/*
9005 			 * There are segments ACKED on the
9006 			 * scoreboard further up. We are seeing
9007 			 * reordering.
9008 			 */
9009 			rsm->r_flags &= ~RACK_SACK_PASSED;
9010 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9011 			rsm->r_flags |= RACK_ACKED;
9012 			rack->r_ctl.rc_reorder_ts = cts;
9013 			if (rack->r_ent_rec_ns) {
9014 				/*
9015 				 * We have sent no more, and we saw an sack
9016 				 * then ack arrive.
9017 				 */
9018 				rack->r_might_revert = 1;
9019 			}
9020 		}
9021 		if ((rsm->r_flags & RACK_TO_REXT) &&
9022 		    (tp->t_flags & TF_RCVD_TSTMP) &&
9023 		    (to->to_flags & TOF_TS) &&
9024 		    (to->to_tsecr != 0) &&
9025 		    (tp->t_flags & TF_PREVVALID)) {
9026 			/*
9027 			 * We can use the timestamp to see
9028 			 * if this retransmission was from the
9029 			 * first transmit. If so we made a mistake.
9030 			 */
9031 			tp->t_flags &= ~TF_PREVVALID;
9032 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9033 				/* The first transmit is what this ack is for */
9034 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
9035 			}
9036 		}
9037 		left = th_ack - rsm->r_end;
9038 		if (rack->app_limited_needs_set && newly_acked)
9039 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9040 		/* Free back to zone */
9041 		rack_free(rack, rsm);
9042 		if (left) {
9043 			goto more;
9044 		}
9045 		/* Check for reneging */
9046 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9047 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9048 			/*
9049 			 * The peer has moved snd_una up to
9050 			 * the edge of this send, i.e. one
9051 			 * that it had previously acked. The only
9052 			 * way that can be true if the peer threw
9053 			 * away data (space issues) that it had
9054 			 * previously sacked (else it would have
9055 			 * given us snd_una up to (rsm->r_end).
9056 			 * We need to undo the acked markings here.
9057 			 *
9058 			 * Note we have to look to make sure th_ack is
9059 			 * our rsm->r_start in case we get an old ack
9060 			 * where th_ack is behind snd_una.
9061 			 */
9062 			rack_peer_reneges(rack, rsm, th_ack);
9063 		}
9064 		return;
9065 	}
9066 	if (rsm->r_flags & RACK_ACKED) {
9067 		/*
9068 		 * It was acked on the scoreboard -- remove it from
9069 		 * total for the part being cum-acked.
9070 		 */
9071 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9072 	}
9073 	/*
9074 	 * Clear the dup ack count for
9075 	 * the piece that remains.
9076 	 */
9077 	rsm->r_dupack = 0;
9078 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9079 	if (rsm->r_rtr_bytes) {
9080 		/*
9081 		 * It was retransmitted adjust the
9082 		 * sack holes for what was acked.
9083 		 */
9084 		int ack_am;
9085 
9086 		ack_am = (th_ack - rsm->r_start);
9087 		if (ack_am >= rsm->r_rtr_bytes) {
9088 			rack->r_ctl.rc_holes_rxt -= ack_am;
9089 			rsm->r_rtr_bytes -= ack_am;
9090 		}
9091 	}
9092 	/*
9093 	 * Update where the piece starts and record
9094 	 * the time of send of highest cumack sent.
9095 	 */
9096 	rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9097 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9098 	/* Now we need to move our offset forward too */
9099 	if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9100 		/* Fix up the orig_m_len and possibly the mbuf offset */
9101 		rack_adjust_orig_mlen(rsm);
9102 	}
9103 	rsm->soff += (th_ack - rsm->r_start);
9104 	rsm->r_start = th_ack;
9105 	/* Now do we need to move the mbuf fwd too? */
9106 	if (rsm->m) {
9107 		while (rsm->soff >= rsm->m->m_len) {
9108 			rsm->soff -= rsm->m->m_len;
9109 			rsm->m = rsm->m->m_next;
9110 			KASSERT((rsm->m != NULL),
9111 				(" nrsm:%p hit at soff:%u null m",
9112 				 rsm, rsm->soff));
9113 		}
9114 		rsm->orig_m_len = rsm->m->m_len;
9115 	}
9116 	if (rack->app_limited_needs_set)
9117 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9118 }
9119 
9120 static void
9121 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9122 {
9123 	struct rack_sendmap *rsm;
9124 	int sack_pass_fnd = 0;
9125 
9126 	if (rack->r_might_revert) {
9127 		/*
9128 		 * Ok we have reordering, have not sent anything, we
9129 		 * might want to revert the congestion state if nothing
9130 		 * further has SACK_PASSED on it. Lets check.
9131 		 *
9132 		 * We also get here when we have DSACKs come in for
9133 		 * all the data that we FR'd. Note that a rxt or tlp
9134 		 * timer clears this from happening.
9135 		 */
9136 
9137 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9138 			if (rsm->r_flags & RACK_SACK_PASSED) {
9139 				sack_pass_fnd = 1;
9140 				break;
9141 			}
9142 		}
9143 		if (sack_pass_fnd == 0) {
9144 			/*
9145 			 * We went into recovery
9146 			 * incorrectly due to reordering!
9147 			 */
9148 			int orig_cwnd;
9149 
9150 			rack->r_ent_rec_ns = 0;
9151 			orig_cwnd = tp->snd_cwnd;
9152 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9153 			tp->snd_recover = tp->snd_una;
9154 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
9155 			EXIT_RECOVERY(tp->t_flags);
9156 		}
9157 		rack->r_might_revert = 0;
9158 	}
9159 }
9160 
9161 #ifdef NETFLIX_EXP_DETECTION
9162 static void
9163 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9164 {
9165 	if ((rack->do_detection || tcp_force_detection) &&
9166 	    tcp_sack_to_ack_thresh &&
9167 	    tcp_sack_to_move_thresh &&
9168 	    ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9169 		/*
9170 		 * We have thresholds set to find
9171 		 * possible attackers and disable sack.
9172 		 * Check them.
9173 		 */
9174 		uint64_t ackratio, moveratio, movetotal;
9175 
9176 		/* Log detecting */
9177 		rack_log_sad(rack, 1);
9178 		ackratio = (uint64_t)(rack->r_ctl.sack_count);
9179 		ackratio *= (uint64_t)(1000);
9180 		if (rack->r_ctl.ack_count)
9181 			ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9182 		else {
9183 			/* We really should not hit here */
9184 			ackratio = 1000;
9185 		}
9186 		if ((rack->sack_attack_disable == 0) &&
9187 		    (ackratio > rack_highest_sack_thresh_seen))
9188 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9189 		movetotal = rack->r_ctl.sack_moved_extra;
9190 		movetotal += rack->r_ctl.sack_noextra_move;
9191 		moveratio = rack->r_ctl.sack_moved_extra;
9192 		moveratio *= (uint64_t)1000;
9193 		if (movetotal)
9194 			moveratio /= movetotal;
9195 		else {
9196 			/* No moves, thats pretty good */
9197 			moveratio = 0;
9198 		}
9199 		if ((rack->sack_attack_disable == 0) &&
9200 		    (moveratio > rack_highest_move_thresh_seen))
9201 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
9202 		if (rack->sack_attack_disable == 0) {
9203 			if ((ackratio > tcp_sack_to_ack_thresh) &&
9204 			    (moveratio > tcp_sack_to_move_thresh)) {
9205 				/* Disable sack processing */
9206 				rack->sack_attack_disable = 1;
9207 				if (rack->r_rep_attack == 0) {
9208 					rack->r_rep_attack = 1;
9209 					counter_u64_add(rack_sack_attacks_detected, 1);
9210 				}
9211 				if (tcp_attack_on_turns_on_logging) {
9212 					/*
9213 					 * Turn on logging, used for debugging
9214 					 * false positives.
9215 					 */
9216 					rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9217 				}
9218 				/* Clamp the cwnd at flight size */
9219 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9220 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9221 				rack_log_sad(rack, 2);
9222 			}
9223 		} else {
9224 			/* We are sack-disabled check for false positives */
9225 			if ((ackratio <= tcp_restoral_thresh) ||
9226 			    (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9227 				rack->sack_attack_disable = 0;
9228 				rack_log_sad(rack, 3);
9229 				/* Restart counting */
9230 				rack->r_ctl.sack_count = 0;
9231 				rack->r_ctl.sack_moved_extra = 0;
9232 				rack->r_ctl.sack_noextra_move = 1;
9233 				rack->r_ctl.ack_count = max(1,
9234 				      (bytes_this_ack / segsiz));
9235 
9236 				if (rack->r_rep_reverse == 0) {
9237 					rack->r_rep_reverse = 1;
9238 					counter_u64_add(rack_sack_attacks_reversed, 1);
9239 				}
9240 				/* Restore the cwnd */
9241 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9242 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9243 			}
9244 		}
9245 	}
9246 }
9247 #endif
9248 
9249 static int
9250 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9251 {
9252 
9253 	uint32_t am, l_end;
9254 	int was_tlp = 0;
9255 
9256 	if (SEQ_GT(end, start))
9257 		am = end - start;
9258 	else
9259 		am = 0;
9260 	if ((rack->rc_last_tlp_acked_set ) &&
9261 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9262 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9263 		/*
9264 		 * The DSACK is because of a TLP which we don't
9265 		 * do anything with the reordering window over since
9266 		 * it was not reordering that caused the DSACK but
9267 		 * our previous retransmit TLP.
9268 		 */
9269 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
9270 		was_tlp = 1;
9271 		goto skip_dsack_round;
9272 	}
9273 	if (rack->rc_last_sent_tlp_seq_valid) {
9274 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9275 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9276 		    (SEQ_LEQ(end, l_end))) {
9277 			/*
9278 			 * This dsack is from the last sent TLP, ignore it
9279 			 * for reordering purposes.
9280 			 */
9281 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
9282 			was_tlp = 1;
9283 			goto skip_dsack_round;
9284 		}
9285 	}
9286 	if (rack->rc_dsack_round_seen == 0) {
9287 		rack->rc_dsack_round_seen = 1;
9288 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9289 		rack->r_ctl.num_dsack++;
9290 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
9291 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9292 	}
9293 skip_dsack_round:
9294 	/*
9295 	 * We keep track of how many DSACK blocks we get
9296 	 * after a recovery incident.
9297 	 */
9298 	rack->r_ctl.dsack_byte_cnt += am;
9299 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9300 	    rack->r_ctl.retran_during_recovery &&
9301 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9302 		/*
9303 		 * False recovery most likely culprit is reordering. If
9304 		 * nothing else is missing we need to revert.
9305 		 */
9306 		rack->r_might_revert = 1;
9307 		rack_handle_might_revert(rack->rc_tp, rack);
9308 		rack->r_might_revert = 0;
9309 		rack->r_ctl.retran_during_recovery = 0;
9310 		rack->r_ctl.dsack_byte_cnt = 0;
9311 	}
9312 	return (was_tlp);
9313 }
9314 
9315 static void
9316 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9317 {
9318 	/* Deal with changed and PRR here (in recovery only) */
9319 	uint32_t pipe, snd_una;
9320 
9321 	rack->r_ctl.rc_prr_delivered += changed;
9322 
9323 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9324 		/*
9325 		 * It is all outstanding, we are application limited
9326 		 * and thus we don't need more room to send anything.
9327 		 * Note we use tp->snd_una here and not th_ack because
9328 		 * the data as yet not been cut from the sb.
9329 		 */
9330 		rack->r_ctl.rc_prr_sndcnt = 0;
9331 		return;
9332 	}
9333 	/* Compute prr_sndcnt */
9334 	if (SEQ_GT(tp->snd_una, th_ack)) {
9335 		snd_una = tp->snd_una;
9336 	} else {
9337 		snd_una = th_ack;
9338 	}
9339 	pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9340 	if (pipe > tp->snd_ssthresh) {
9341 		long sndcnt;
9342 
9343 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9344 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
9345 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9346 		else {
9347 			rack->r_ctl.rc_prr_sndcnt = 0;
9348 			rack_log_to_prr(rack, 9, 0, __LINE__);
9349 			sndcnt = 0;
9350 		}
9351 		sndcnt++;
9352 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9353 			sndcnt -= rack->r_ctl.rc_prr_out;
9354 		else
9355 			sndcnt = 0;
9356 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
9357 		rack_log_to_prr(rack, 10, 0, __LINE__);
9358 	} else {
9359 		uint32_t limit;
9360 
9361 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9362 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9363 		else
9364 			limit = 0;
9365 		if (changed > limit)
9366 			limit = changed;
9367 		limit += ctf_fixed_maxseg(tp);
9368 		if (tp->snd_ssthresh > pipe) {
9369 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9370 			rack_log_to_prr(rack, 11, 0, __LINE__);
9371 		} else {
9372 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9373 			rack_log_to_prr(rack, 12, 0, __LINE__);
9374 		}
9375 	}
9376 }
9377 
9378 static void
9379 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9380 {
9381 	uint32_t changed;
9382 	struct tcp_rack *rack;
9383 	struct rack_sendmap *rsm;
9384 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9385 	register uint32_t th_ack;
9386 	int32_t i, j, k, num_sack_blks = 0;
9387 	uint32_t cts, acked, ack_point;
9388 	int loop_start = 0, moved_two = 0;
9389 	uint32_t tsused;
9390 
9391 
9392 	INP_WLOCK_ASSERT(tp->t_inpcb);
9393 	if (tcp_get_flags(th) & TH_RST) {
9394 		/* We don't log resets */
9395 		return;
9396 	}
9397 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9398 	cts = tcp_get_usecs(NULL);
9399 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9400 	changed = 0;
9401 	th_ack = th->th_ack;
9402 	if (rack->sack_attack_disable == 0)
9403 		rack_do_decay(rack);
9404 	if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9405 		/*
9406 		 * You only get credit for
9407 		 * MSS and greater (and you get extra
9408 		 * credit for larger cum-ack moves).
9409 		 */
9410 		int ac;
9411 
9412 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9413 		rack->r_ctl.ack_count += ac;
9414 		counter_u64_add(rack_ack_total, ac);
9415 	}
9416 	if (rack->r_ctl.ack_count > 0xfff00000) {
9417 		/*
9418 		 * reduce the number to keep us under
9419 		 * a uint32_t.
9420 		 */
9421 		rack->r_ctl.ack_count /= 2;
9422 		rack->r_ctl.sack_count /= 2;
9423 	}
9424 	if (SEQ_GT(th_ack, tp->snd_una)) {
9425 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9426 		tp->t_acktime = ticks;
9427 	}
9428 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
9429 		changed = th_ack - rsm->r_start;
9430 	if (changed) {
9431 		rack_process_to_cumack(tp, rack, th_ack, cts, to);
9432 	}
9433 	if ((to->to_flags & TOF_SACK) == 0) {
9434 		/* We are done nothing left and no sack. */
9435 		rack_handle_might_revert(tp, rack);
9436 		/*
9437 		 * For cases where we struck a dup-ack
9438 		 * with no SACK, add to the changes so
9439 		 * PRR will work right.
9440 		 */
9441 		if (dup_ack_struck && (changed == 0)) {
9442 			changed += ctf_fixed_maxseg(rack->rc_tp);
9443 		}
9444 		goto out;
9445 	}
9446 	/* Sack block processing */
9447 	if (SEQ_GT(th_ack, tp->snd_una))
9448 		ack_point = th_ack;
9449 	else
9450 		ack_point = tp->snd_una;
9451 	for (i = 0; i < to->to_nsacks; i++) {
9452 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
9453 		      &sack, sizeof(sack));
9454 		sack.start = ntohl(sack.start);
9455 		sack.end = ntohl(sack.end);
9456 		if (SEQ_GT(sack.end, sack.start) &&
9457 		    SEQ_GT(sack.start, ack_point) &&
9458 		    SEQ_LT(sack.start, tp->snd_max) &&
9459 		    SEQ_GT(sack.end, ack_point) &&
9460 		    SEQ_LEQ(sack.end, tp->snd_max)) {
9461 			sack_blocks[num_sack_blks] = sack;
9462 			num_sack_blks++;
9463 		} else if (SEQ_LEQ(sack.start, th_ack) &&
9464 			   SEQ_LEQ(sack.end, th_ack)) {
9465 			int was_tlp;
9466 
9467 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9468 			/*
9469 			 * Its a D-SACK block.
9470 			 */
9471 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9472 		}
9473 	}
9474 	if (rack->rc_dsack_round_seen) {
9475 		/* Is the dsack roound over? */
9476 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9477 			/* Yes it is */
9478 			rack->rc_dsack_round_seen = 0;
9479 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9480 		}
9481 	}
9482 	/*
9483 	 * Sort the SACK blocks so we can update the rack scoreboard with
9484 	 * just one pass.
9485 	 */
9486 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9487 					 num_sack_blks, th->th_ack);
9488 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9489 	if (num_sack_blks == 0) {
9490 		/* Nothing to sack (DSACKs?) */
9491 		goto out_with_totals;
9492 	}
9493 	if (num_sack_blks < 2) {
9494 		/* Only one, we don't need to sort */
9495 		goto do_sack_work;
9496 	}
9497 	/* Sort the sacks */
9498 	for (i = 0; i < num_sack_blks; i++) {
9499 		for (j = i + 1; j < num_sack_blks; j++) {
9500 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9501 				sack = sack_blocks[i];
9502 				sack_blocks[i] = sack_blocks[j];
9503 				sack_blocks[j] = sack;
9504 			}
9505 		}
9506 	}
9507 	/*
9508 	 * Now are any of the sack block ends the same (yes some
9509 	 * implementations send these)?
9510 	 */
9511 again:
9512 	if (num_sack_blks == 0)
9513 		goto out_with_totals;
9514 	if (num_sack_blks > 1) {
9515 		for (i = 0; i < num_sack_blks; i++) {
9516 			for (j = i + 1; j < num_sack_blks; j++) {
9517 				if (sack_blocks[i].end == sack_blocks[j].end) {
9518 					/*
9519 					 * Ok these two have the same end we
9520 					 * want the smallest end and then
9521 					 * throw away the larger and start
9522 					 * again.
9523 					 */
9524 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9525 						/*
9526 						 * The second block covers
9527 						 * more area use that
9528 						 */
9529 						sack_blocks[i].start = sack_blocks[j].start;
9530 					}
9531 					/*
9532 					 * Now collapse out the dup-sack and
9533 					 * lower the count
9534 					 */
9535 					for (k = (j + 1); k < num_sack_blks; k++) {
9536 						sack_blocks[j].start = sack_blocks[k].start;
9537 						sack_blocks[j].end = sack_blocks[k].end;
9538 						j++;
9539 					}
9540 					num_sack_blks--;
9541 					goto again;
9542 				}
9543 			}
9544 		}
9545 	}
9546 do_sack_work:
9547 	/*
9548 	 * First lets look to see if
9549 	 * we have retransmitted and
9550 	 * can use the transmit next?
9551 	 */
9552 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9553 	if (rsm &&
9554 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9555 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9556 		/*
9557 		 * We probably did the FR and the next
9558 		 * SACK in continues as we would expect.
9559 		 */
9560 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9561 		if (acked) {
9562 			rack->r_wanted_output = 1;
9563 			changed += acked;
9564 		}
9565 		if (num_sack_blks == 1) {
9566 			/*
9567 			 * This is what we would expect from
9568 			 * a normal implementation to happen
9569 			 * after we have retransmitted the FR,
9570 			 * i.e the sack-filter pushes down
9571 			 * to 1 block and the next to be retransmitted
9572 			 * is the sequence in the sack block (has more
9573 			 * are acked). Count this as ACK'd data to boost
9574 			 * up the chances of recovering any false positives.
9575 			 */
9576 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9577 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9578 			counter_u64_add(rack_express_sack, 1);
9579 			if (rack->r_ctl.ack_count > 0xfff00000) {
9580 				/*
9581 				 * reduce the number to keep us under
9582 				 * a uint32_t.
9583 				 */
9584 				rack->r_ctl.ack_count /= 2;
9585 				rack->r_ctl.sack_count /= 2;
9586 			}
9587 			goto out_with_totals;
9588 		} else {
9589 			/*
9590 			 * Start the loop through the
9591 			 * rest of blocks, past the first block.
9592 			 */
9593 			moved_two = 0;
9594 			loop_start = 1;
9595 		}
9596 	}
9597 	/* Its a sack of some sort */
9598 	rack->r_ctl.sack_count++;
9599 	if (rack->r_ctl.sack_count > 0xfff00000) {
9600 		/*
9601 		 * reduce the number to keep us under
9602 		 * a uint32_t.
9603 		 */
9604 		rack->r_ctl.ack_count /= 2;
9605 		rack->r_ctl.sack_count /= 2;
9606 	}
9607 	counter_u64_add(rack_sack_total, 1);
9608 	if (rack->sack_attack_disable) {
9609 		/* An attacker disablement is in place */
9610 		if (num_sack_blks > 1) {
9611 			rack->r_ctl.sack_count += (num_sack_blks - 1);
9612 			rack->r_ctl.sack_moved_extra++;
9613 			counter_u64_add(rack_move_some, 1);
9614 			if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9615 				rack->r_ctl.sack_moved_extra /= 2;
9616 				rack->r_ctl.sack_noextra_move /= 2;
9617 			}
9618 		}
9619 		goto out;
9620 	}
9621 	rsm = rack->r_ctl.rc_sacklast;
9622 	for (i = loop_start; i < num_sack_blks; i++) {
9623 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9624 		if (acked) {
9625 			rack->r_wanted_output = 1;
9626 			changed += acked;
9627 		}
9628 		if (moved_two) {
9629 			/*
9630 			 * If we did not get a SACK for at least a MSS and
9631 			 * had to move at all, or if we moved more than our
9632 			 * threshold, it counts against the "extra" move.
9633 			 */
9634 			rack->r_ctl.sack_moved_extra += moved_two;
9635 			counter_u64_add(rack_move_some, 1);
9636 		} else {
9637 			/*
9638 			 * else we did not have to move
9639 			 * any more than we would expect.
9640 			 */
9641 			rack->r_ctl.sack_noextra_move++;
9642 			counter_u64_add(rack_move_none, 1);
9643 		}
9644 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9645 			/*
9646 			 * If the SACK was not a full MSS then
9647 			 * we add to sack_count the number of
9648 			 * MSS's (or possibly more than
9649 			 * a MSS if its a TSO send) we had to skip by.
9650 			 */
9651 			rack->r_ctl.sack_count += moved_two;
9652 			counter_u64_add(rack_sack_total, moved_two);
9653 		}
9654 		/*
9655 		 * Now we need to setup for the next
9656 		 * round. First we make sure we won't
9657 		 * exceed the size of our uint32_t on
9658 		 * the various counts, and then clear out
9659 		 * moved_two.
9660 		 */
9661 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9662 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9663 			rack->r_ctl.sack_moved_extra /= 2;
9664 			rack->r_ctl.sack_noextra_move /= 2;
9665 		}
9666 		if (rack->r_ctl.sack_count > 0xfff00000) {
9667 			rack->r_ctl.ack_count /= 2;
9668 			rack->r_ctl.sack_count /= 2;
9669 		}
9670 		moved_two = 0;
9671 	}
9672 out_with_totals:
9673 	if (num_sack_blks > 1) {
9674 		/*
9675 		 * You get an extra stroke if
9676 		 * you have more than one sack-blk, this
9677 		 * could be where we are skipping forward
9678 		 * and the sack-filter is still working, or
9679 		 * it could be an attacker constantly
9680 		 * moving us.
9681 		 */
9682 		rack->r_ctl.sack_moved_extra++;
9683 		counter_u64_add(rack_move_some, 1);
9684 	}
9685 out:
9686 #ifdef NETFLIX_EXP_DETECTION
9687 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9688 #endif
9689 	if (changed) {
9690 		/* Something changed cancel the rack timer */
9691 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9692 	}
9693 	tsused = tcp_get_usecs(NULL);
9694 	rsm = tcp_rack_output(tp, rack, tsused);
9695 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9696 	    rsm &&
9697 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
9698 		/* Enter recovery */
9699 		entered_recovery = 1;
9700 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9701 		/*
9702 		 * When we enter recovery we need to assure we send
9703 		 * one packet.
9704 		 */
9705 		if (rack->rack_no_prr == 0) {
9706 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9707 			rack_log_to_prr(rack, 8, 0, __LINE__);
9708 		}
9709 		rack->r_timer_override = 1;
9710 		rack->r_early = 0;
9711 		rack->r_ctl.rc_agg_early = 0;
9712 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
9713 		   rsm &&
9714 		   (rack->r_rr_config == 3)) {
9715 		/*
9716 		 * Assure we can output and we get no
9717 		 * remembered pace time except the retransmit.
9718 		 */
9719 		rack->r_timer_override = 1;
9720 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9721 		rack->r_ctl.rc_resend = rsm;
9722 	}
9723 	if (IN_FASTRECOVERY(tp->t_flags) &&
9724 	    (rack->rack_no_prr == 0) &&
9725 	    (entered_recovery == 0)) {
9726 		rack_update_prr(tp, rack, changed, th_ack);
9727 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9728 		     ((tcp_in_hpts(rack->rc_inp) == 0) &&
9729 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9730 			/*
9731 			 * If you are pacing output you don't want
9732 			 * to override.
9733 			 */
9734 			rack->r_early = 0;
9735 			rack->r_ctl.rc_agg_early = 0;
9736 			rack->r_timer_override = 1;
9737 		}
9738 	}
9739 }
9740 
9741 static void
9742 rack_strike_dupack(struct tcp_rack *rack)
9743 {
9744 	struct rack_sendmap *rsm;
9745 
9746 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9747 	while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9748 		rsm = TAILQ_NEXT(rsm, r_tnext);
9749 		if (rsm->r_flags & RACK_MUST_RXT) {
9750 			/* Sendmap entries that are marked to
9751 			 * be retransmitted do not need dupack's
9752 			 * struck. We get these marks for a number
9753 			 * of reasons (rxt timeout with no sack,
9754 			 * mtu change, or rwnd collapses). When
9755 			 * these events occur, we know we must retransmit
9756 			 * them and mark the sendmap entries. Dupack counting
9757 			 * is not needed since we are already set to retransmit
9758 			 * it as soon as we can.
9759 			 */
9760 			continue;
9761 		}
9762 	}
9763 	if (rsm && (rsm->r_dupack < 0xff)) {
9764 		rsm->r_dupack++;
9765 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9766 			struct timeval tv;
9767 			uint32_t cts;
9768 			/*
9769 			 * Here we see if we need to retransmit. For
9770 			 * a SACK type connection if enough time has passed
9771 			 * we will get a return of the rsm. For a non-sack
9772 			 * connection we will get the rsm returned if the
9773 			 * dupack value is 3 or more.
9774 			 */
9775 			cts = tcp_get_usecs(&tv);
9776 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9777 			if (rack->r_ctl.rc_resend != NULL) {
9778 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9779 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9780 							 rack->rc_tp->snd_una, __LINE__);
9781 				}
9782 				rack->r_wanted_output = 1;
9783 				rack->r_timer_override = 1;
9784 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9785 			}
9786 		} else {
9787 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9788 		}
9789 	}
9790 }
9791 
9792 static void
9793 rack_check_bottom_drag(struct tcpcb *tp,
9794 		       struct tcp_rack *rack,
9795 		       struct socket *so, int32_t acked)
9796 {
9797 	uint32_t segsiz, minseg;
9798 
9799 	segsiz = ctf_fixed_maxseg(tp);
9800 	minseg = segsiz;
9801 
9802 	if (tp->snd_max == tp->snd_una) {
9803 		/*
9804 		 * We are doing dynamic pacing and we are way
9805 		 * under. Basically everything got acked while
9806 		 * we were still waiting on the pacer to expire.
9807 		 *
9808 		 * This means we need to boost the b/w in
9809 		 * addition to any earlier boosting of
9810 		 * the multipler.
9811 		 */
9812 		rack->rc_dragged_bottom = 1;
9813 		rack_validate_multipliers_at_or_above100(rack);
9814 		/*
9815 		 * Lets use the segment bytes acked plus
9816 		 * the lowest RTT seen as the basis to
9817 		 * form a b/w estimate. This will be off
9818 		 * due to the fact that the true estimate
9819 		 * should be around 1/2 the time of the RTT
9820 		 * but we can settle for that.
9821 		 */
9822 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9823 		    acked) {
9824 			uint64_t bw, calc_bw, rtt;
9825 
9826 			rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9827 			if (rtt == 0) {
9828 				/* no us sample is there a ms one? */
9829 				if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9830 					rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9831 				} else {
9832 					goto no_measurement;
9833 				}
9834 			}
9835 			bw = acked;
9836 			calc_bw = bw * 1000000;
9837 			calc_bw /= rtt;
9838 			if (rack->r_ctl.last_max_bw &&
9839 			    (rack->r_ctl.last_max_bw < calc_bw)) {
9840 				/*
9841 				 * If we have a last calculated max bw
9842 				 * enforce it.
9843 				 */
9844 				calc_bw = rack->r_ctl.last_max_bw;
9845 			}
9846 			/* now plop it in */
9847 			if (rack->rc_gp_filled == 0) {
9848 				if (calc_bw > ONE_POINT_TWO_MEG) {
9849 					/*
9850 					 * If we have no measurement
9851 					 * don't let us set in more than
9852 					 * 1.2Mbps. If we are still too
9853 					 * low after pacing with this we
9854 					 * will hopefully have a max b/w
9855 					 * available to sanity check things.
9856 					 */
9857 					calc_bw = ONE_POINT_TWO_MEG;
9858 				}
9859 				rack->r_ctl.rc_rtt_diff = 0;
9860 				rack->r_ctl.gp_bw = calc_bw;
9861 				rack->rc_gp_filled = 1;
9862 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9863 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9864 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9865 			} else if (calc_bw > rack->r_ctl.gp_bw) {
9866 				rack->r_ctl.rc_rtt_diff = 0;
9867 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9868 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9869 				rack->r_ctl.gp_bw = calc_bw;
9870 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9871 			} else
9872 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9873 			if ((rack->gp_ready == 0) &&
9874 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9875 				/* We have enough measurements now */
9876 				rack->gp_ready = 1;
9877 				rack_set_cc_pacing(rack);
9878 				if (rack->defer_options)
9879 					rack_apply_deferred_options(rack);
9880 			}
9881 			/*
9882 			 * For acks over 1mss we do a extra boost to simulate
9883 			 * where we would get 2 acks (we want 110 for the mul).
9884 			 */
9885 			if (acked > segsiz)
9886 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9887 		} else {
9888 			/*
9889 			 * zero rtt possibly?, settle for just an old increase.
9890 			 */
9891 no_measurement:
9892 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
9893 		}
9894 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9895 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9896 					       minseg)) &&
9897 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9898 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9899 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9900 		    (segsiz * rack_req_segs))) {
9901 		/*
9902 		 * We are doing dynamic GP pacing and
9903 		 * we have everything except 1MSS or less
9904 		 * bytes left out. We are still pacing away.
9905 		 * And there is data that could be sent, This
9906 		 * means we are inserting delayed ack time in
9907 		 * our measurements because we are pacing too slow.
9908 		 */
9909 		rack_validate_multipliers_at_or_above100(rack);
9910 		rack->rc_dragged_bottom = 1;
9911 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
9912 	}
9913 }
9914 
9915 
9916 
9917 static void
9918 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9919 {
9920 	/*
9921 	 * The fast output path is enabled and we
9922 	 * have moved the cumack forward. Lets see if
9923 	 * we can expand forward the fast path length by
9924 	 * that amount. What we would ideally like to
9925 	 * do is increase the number of bytes in the
9926 	 * fast path block (left_to_send) by the
9927 	 * acked amount. However we have to gate that
9928 	 * by two factors:
9929 	 * 1) The amount outstanding and the rwnd of the peer
9930 	 *    (i.e. we don't want to exceed the rwnd of the peer).
9931 	 *    <and>
9932 	 * 2) The amount of data left in the socket buffer (i.e.
9933 	 *    we can't send beyond what is in the buffer).
9934 	 *
9935 	 * Note that this does not take into account any increase
9936 	 * in the cwnd. We will only extend the fast path by
9937 	 * what was acked.
9938 	 */
9939 	uint32_t new_total, gating_val;
9940 
9941 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9942 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9943 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9944 	if (new_total <= gating_val) {
9945 		/* We can increase left_to_send by the acked amount */
9946 		counter_u64_add(rack_extended_rfo, 1);
9947 		rack->r_ctl.fsb.left_to_send = new_total;
9948 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9949 			("rack:%p left_to_send:%u sbavail:%u out:%u",
9950 			 rack, rack->r_ctl.fsb.left_to_send,
9951 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
9952 			 (tp->snd_max - tp->snd_una)));
9953 
9954 	}
9955 }
9956 
9957 static void
9958 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9959 {
9960 	/*
9961 	 * Here any sendmap entry that points to the
9962 	 * beginning mbuf must be adjusted to the correct
9963 	 * offset. This must be called with:
9964 	 * 1) The socket buffer locked
9965 	 * 2) snd_una adjusted to its new postion.
9966 	 *
9967 	 * Note that (2) implies rack_ack_received has also
9968 	 * been called.
9969 	 *
9970 	 * We grab the first mbuf in the socket buffer and
9971 	 * then go through the front of the sendmap, recalculating
9972 	 * the stored offset for any sendmap entry that has
9973 	 * that mbuf. We must use the sb functions to do this
9974 	 * since its possible an add was done has well as
9975 	 * the subtraction we may have just completed. This should
9976 	 * not be a penalty though, since we just referenced the sb
9977 	 * to go in and trim off the mbufs that we freed (of course
9978 	 * there will be a penalty for the sendmap references though).
9979 	 */
9980 	struct mbuf *m;
9981 	struct rack_sendmap *rsm;
9982 
9983 	SOCKBUF_LOCK_ASSERT(sb);
9984 	m = sb->sb_mb;
9985 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9986 	if ((rsm == NULL) || (m == NULL)) {
9987 		/* Nothing outstanding */
9988 		return;
9989 	}
9990 	while (rsm->m && (rsm->m == m)) {
9991 		/* one to adjust */
9992 #ifdef INVARIANTS
9993 		struct mbuf *tm;
9994 		uint32_t soff;
9995 
9996 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9997 		if (rsm->orig_m_len != m->m_len) {
9998 			rack_adjust_orig_mlen(rsm);
9999 		}
10000 		if (rsm->soff != soff) {
10001 			/*
10002 			 * This is not a fatal error, we anticipate it
10003 			 * might happen (the else code), so we count it here
10004 			 * so that under invariant we can see that it really
10005 			 * does happen.
10006 			 */
10007 			counter_u64_add(rack_adjust_map_bw, 1);
10008 		}
10009 		rsm->m = tm;
10010 		rsm->soff = soff;
10011 		if (tm)
10012 			rsm->orig_m_len = rsm->m->m_len;
10013 		else
10014 			rsm->orig_m_len = 0;
10015 #else
10016 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
10017 		if (rsm->m)
10018 			rsm->orig_m_len = rsm->m->m_len;
10019 		else
10020 			rsm->orig_m_len = 0;
10021 #endif
10022 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
10023 			      rsm);
10024 		if (rsm == NULL)
10025 			break;
10026 	}
10027 }
10028 
10029 /*
10030  * Return value of 1, we do not need to call rack_process_data().
10031  * return value of 0, rack_process_data can be called.
10032  * For ret_val if its 0 the TCP is locked, if its non-zero
10033  * its unlocked and probably unsafe to touch the TCB.
10034  */
10035 static int
10036 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10037     struct tcpcb *tp, struct tcpopt *to,
10038     uint32_t tiwin, int32_t tlen,
10039     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10040 {
10041 	int32_t ourfinisacked = 0;
10042 	int32_t nsegs, acked_amount;
10043 	int32_t acked;
10044 	struct mbuf *mfree;
10045 	struct tcp_rack *rack;
10046 	int32_t under_pacing = 0;
10047 	int32_t recovery = 0;
10048 
10049 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10050 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
10051 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10052 				      &rack->r_ctl.challenge_ack_ts,
10053 				      &rack->r_ctl.challenge_ack_cnt);
10054 		rack->r_wanted_output = 1;
10055 		return (1);
10056 	}
10057 	if (rack->gp_ready &&
10058 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10059 		under_pacing = 1;
10060 	}
10061 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10062 		int in_rec, dup_ack_struck = 0;
10063 
10064 		in_rec = IN_FASTRECOVERY(tp->t_flags);
10065 		if (rack->rc_in_persist) {
10066 			tp->t_rxtshift = 0;
10067 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10068 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10069 		}
10070 		if ((th->th_ack == tp->snd_una) &&
10071 		    (tiwin == tp->snd_wnd) &&
10072 		    ((to->to_flags & TOF_SACK) == 0)) {
10073 			rack_strike_dupack(rack);
10074 			dup_ack_struck = 1;
10075 		}
10076 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10077 	}
10078 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10079 		/*
10080 		 * Old ack, behind (or duplicate to) the last one rcv'd
10081 		 * Note: We mark reordering is occuring if its
10082 		 * less than and we have not closed our window.
10083 		 */
10084 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10085 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10086 		}
10087 		return (0);
10088 	}
10089 	/*
10090 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10091 	 * something we sent.
10092 	 */
10093 	if (tp->t_flags & TF_NEEDSYN) {
10094 		/*
10095 		 * T/TCP: Connection was half-synchronized, and our SYN has
10096 		 * been ACK'd (so connection is now fully synchronized).  Go
10097 		 * to non-starred state, increment snd_una for ACK of SYN,
10098 		 * and check if we can do window scaling.
10099 		 */
10100 		tp->t_flags &= ~TF_NEEDSYN;
10101 		tp->snd_una++;
10102 		/* Do window scaling? */
10103 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10104 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10105 			tp->rcv_scale = tp->request_r_scale;
10106 			/* Send window already scaled. */
10107 		}
10108 	}
10109 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10110 	INP_WLOCK_ASSERT(tp->t_inpcb);
10111 
10112 	acked = BYTES_THIS_ACK(tp, th);
10113 	if (acked) {
10114 		/*
10115 		 * Any time we move the cum-ack forward clear
10116 		 * keep-alive tied probe-not-answered. The
10117 		 * persists clears its own on entry.
10118 		 */
10119 		rack->probe_not_answered = 0;
10120 	}
10121 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10122 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10123 	/*
10124 	 * If we just performed our first retransmit, and the ACK arrives
10125 	 * within our recovery window, then it was a mistake to do the
10126 	 * retransmit in the first place.  Recover our original cwnd and
10127 	 * ssthresh, and proceed to transmit where we left off.
10128 	 */
10129 	if ((tp->t_flags & TF_PREVVALID) &&
10130 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10131 		tp->t_flags &= ~TF_PREVVALID;
10132 		if (tp->t_rxtshift == 1 &&
10133 		    (int)(ticks - tp->t_badrxtwin) < 0)
10134 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10135 	}
10136 	if (acked) {
10137 		/* assure we are not backed off */
10138 		tp->t_rxtshift = 0;
10139 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10140 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10141 		rack->rc_tlp_in_progress = 0;
10142 		rack->r_ctl.rc_tlp_cnt_out = 0;
10143 		/*
10144 		 * If it is the RXT timer we want to
10145 		 * stop it, so we can restart a TLP.
10146 		 */
10147 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10148 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10149 #ifdef NETFLIX_HTTP_LOGGING
10150 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10151 #endif
10152 	}
10153 	/*
10154 	 * If we have a timestamp reply, update smoothed round trip time. If
10155 	 * no timestamp is present but transmit timer is running and timed
10156 	 * sequence number was acked, update smoothed round trip time. Since
10157 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
10158 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
10159 	 * timer.
10160 	 *
10161 	 * Some boxes send broken timestamp replies during the SYN+ACK
10162 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10163 	 * and blow up the retransmit timer.
10164 	 */
10165 	/*
10166 	 * If all outstanding data is acked, stop retransmit timer and
10167 	 * remember to restart (more output or persist). If there is more
10168 	 * data to be acked, restart retransmit timer, using current
10169 	 * (possibly backed-off) value.
10170 	 */
10171 	if (acked == 0) {
10172 		if (ofia)
10173 			*ofia = ourfinisacked;
10174 		return (0);
10175 	}
10176 	if (IN_RECOVERY(tp->t_flags)) {
10177 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10178 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
10179 			tcp_rack_partialack(tp);
10180 		} else {
10181 			rack_post_recovery(tp, th->th_ack);
10182 			recovery = 1;
10183 		}
10184 	}
10185 	/*
10186 	 * Let the congestion control algorithm update congestion control
10187 	 * related information. This typically means increasing the
10188 	 * congestion window.
10189 	 */
10190 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10191 	SOCKBUF_LOCK(&so->so_snd);
10192 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
10193 	tp->snd_wnd -= acked_amount;
10194 	mfree = sbcut_locked(&so->so_snd, acked_amount);
10195 	if ((sbused(&so->so_snd) == 0) &&
10196 	    (acked > acked_amount) &&
10197 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
10198 	    (tp->t_flags & TF_SENTFIN)) {
10199 		/*
10200 		 * We must be sure our fin
10201 		 * was sent and acked (we can be
10202 		 * in FIN_WAIT_1 without having
10203 		 * sent the fin).
10204 		 */
10205 		ourfinisacked = 1;
10206 	}
10207 	tp->snd_una = th->th_ack;
10208 	if (acked_amount && sbavail(&so->so_snd))
10209 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10210 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10211 	/* NB: sowwakeup_locked() does an implicit unlock. */
10212 	sowwakeup_locked(so);
10213 	m_freem(mfree);
10214 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
10215 		tp->snd_recover = tp->snd_una;
10216 
10217 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10218 		tp->snd_nxt = tp->snd_una;
10219 	}
10220 	if (under_pacing &&
10221 	    (rack->use_fixed_rate == 0) &&
10222 	    (rack->in_probe_rtt == 0) &&
10223 	    rack->rc_gp_dyn_mul &&
10224 	    rack->rc_always_pace) {
10225 		/* Check if we are dragging bottom */
10226 		rack_check_bottom_drag(tp, rack, so, acked);
10227 	}
10228 	if (tp->snd_una == tp->snd_max) {
10229 		/* Nothing left outstanding */
10230 		tp->t_flags &= ~TF_PREVVALID;
10231 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10232 		rack->r_ctl.retran_during_recovery = 0;
10233 		rack->r_ctl.dsack_byte_cnt = 0;
10234 		if (rack->r_ctl.rc_went_idle_time == 0)
10235 			rack->r_ctl.rc_went_idle_time = 1;
10236 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10237 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10238 			tp->t_acktime = 0;
10239 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10240 		/* Set need output so persist might get set */
10241 		rack->r_wanted_output = 1;
10242 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10243 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10244 		    (sbavail(&so->so_snd) == 0) &&
10245 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10246 			/*
10247 			 * The socket was gone and the
10248 			 * peer sent data (now or in the past), time to
10249 			 * reset him.
10250 			 */
10251 			*ret_val = 1;
10252 			/* tcp_close will kill the inp pre-log the Reset */
10253 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10254 			tp = tcp_close(tp);
10255 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10256 			return (1);
10257 		}
10258 	}
10259 	if (ofia)
10260 		*ofia = ourfinisacked;
10261 	return (0);
10262 }
10263 
10264 static void
10265 rack_collapsed_window(struct tcp_rack *rack)
10266 {
10267 	/*
10268 	 * Now we must walk the
10269 	 * send map and divide the
10270 	 * ones left stranded. These
10271 	 * guys can't cause us to abort
10272 	 * the connection and are really
10273 	 * "unsent". However if a buggy
10274 	 * client actually did keep some
10275 	 * of the data i.e. collapsed the win
10276 	 * and refused to ack and then opened
10277 	 * the win and acked that data. We would
10278 	 * get into an ack war, the simplier
10279 	 * method then of just pretending we
10280 	 * did not send those segments something
10281 	 * won't work.
10282 	 */
10283 	struct rack_sendmap *rsm, *nrsm, fe;
10284 #ifdef INVARIANTS
10285 	struct rack_sendmap *insret;
10286 #endif
10287 	tcp_seq max_seq;
10288 
10289 	max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10290 	memset(&fe, 0, sizeof(fe));
10291 	fe.r_start = max_seq;
10292 	/* Find the first seq past or at maxseq */
10293 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10294 	if (rsm == NULL) {
10295 		/* Nothing to do strange */
10296 		rack->rc_has_collapsed = 0;
10297 		return;
10298 	}
10299 	/*
10300 	 * Now do we need to split at
10301 	 * the collapse point?
10302 	 */
10303 	if (SEQ_GT(max_seq, rsm->r_start)) {
10304 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10305 		if (nrsm == NULL) {
10306 			/* We can't get a rsm, mark all? */
10307 			nrsm = rsm;
10308 			goto no_split;
10309 		}
10310 		/* Clone it */
10311 		rack_clone_rsm(rack, nrsm, rsm, max_seq);
10312 #ifndef INVARIANTS
10313 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10314 #else
10315 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10316 		if (insret != NULL) {
10317 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10318 			      nrsm, insret, rack, rsm);
10319 		}
10320 #endif
10321 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
10322 		if (rsm->r_in_tmap) {
10323 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10324 			nrsm->r_in_tmap = 1;
10325 		}
10326 		/*
10327 		 * Set in the new RSM as the
10328 		 * collapsed starting point
10329 		 */
10330 		rsm = nrsm;
10331 	}
10332 no_split:
10333 	counter_u64_add(rack_collapsed_win, 1);
10334 	RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10335 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
10336 	}
10337 	rack->rc_has_collapsed = 1;
10338 }
10339 
10340 static void
10341 rack_un_collapse_window(struct tcp_rack *rack)
10342 {
10343 	struct rack_sendmap *rsm;
10344 	int cnt = 0;;
10345 
10346 	rack->r_ctl.rc_out_at_rto = 0;
10347 	rack->r_ctl.rc_snd_max_at_rto = rack->rc_tp->snd_una;
10348 	RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
10349 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
10350 			rsm->r_flags &= ~RACK_RWND_COLLAPSED;
10351 			rsm->r_flags |= RACK_MUST_RXT;
10352 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
10353 				rack->r_ctl.rc_snd_max_at_rto = rsm->r_end;
10354 				rack->r_ctl.rc_out_at_rto += (rsm->r_end - rsm->r_start);
10355 			}
10356 			cnt++;
10357 		}
10358 		else
10359 			break;
10360 	}
10361 	rack->rc_has_collapsed = 0;
10362 	if (cnt) {
10363 		rack->r_must_retran = 1;
10364 	}
10365 }
10366 
10367 static void
10368 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10369 			int32_t tlen, int32_t tfo_syn)
10370 {
10371 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
10372 		if (rack->rc_dack_mode &&
10373 		    (tlen > 500) &&
10374 		    (rack->rc_dack_toggle == 1)) {
10375 			goto no_delayed_ack;
10376 		}
10377 		rack_timer_cancel(tp, rack,
10378 				  rack->r_ctl.rc_rcvtime, __LINE__);
10379 		tp->t_flags |= TF_DELACK;
10380 	} else {
10381 no_delayed_ack:
10382 		rack->r_wanted_output = 1;
10383 		tp->t_flags |= TF_ACKNOW;
10384 		if (rack->rc_dack_mode) {
10385 			if (tp->t_flags & TF_DELACK)
10386 				rack->rc_dack_toggle = 1;
10387 			else
10388 				rack->rc_dack_toggle = 0;
10389 		}
10390 	}
10391 }
10392 
10393 static void
10394 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10395 {
10396 	/*
10397 	 * If fast output is in progress, lets validate that
10398 	 * the new window did not shrink on us and make it
10399 	 * so fast output should end.
10400 	 */
10401 	if (rack->r_fast_output) {
10402 		uint32_t out;
10403 
10404 		/*
10405 		 * Calculate what we will send if left as is
10406 		 * and compare that to our send window.
10407 		 */
10408 		out = ctf_outstanding(tp);
10409 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10410 			/* ok we have an issue */
10411 			if (out >= tp->snd_wnd) {
10412 				/* Turn off fast output the window is met or collapsed */
10413 				rack->r_fast_output = 0;
10414 			} else {
10415 				/* we have some room left */
10416 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10417 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10418 					/* If not at least 1 full segment never mind */
10419 					rack->r_fast_output = 0;
10420 				}
10421 			}
10422 		}
10423 	}
10424 }
10425 
10426 
10427 /*
10428  * Return value of 1, the TCB is unlocked and most
10429  * likely gone, return value of 0, the TCP is still
10430  * locked.
10431  */
10432 static int
10433 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10434     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10435     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10436 {
10437 	/*
10438 	 * Update window information. Don't look at window if no ACK: TAC's
10439 	 * send garbage on first SYN.
10440 	 */
10441 	int32_t nsegs;
10442 	int32_t tfo_syn;
10443 	struct tcp_rack *rack;
10444 
10445 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10446 	INP_WLOCK_ASSERT(tp->t_inpcb);
10447 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10448 	if ((thflags & TH_ACK) &&
10449 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10450 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10451 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10452 		/* keep track of pure window updates */
10453 		if (tlen == 0 &&
10454 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10455 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10456 		tp->snd_wnd = tiwin;
10457 		rack_validate_fo_sendwin_up(tp, rack);
10458 		tp->snd_wl1 = th->th_seq;
10459 		tp->snd_wl2 = th->th_ack;
10460 		if (tp->snd_wnd > tp->max_sndwnd)
10461 			tp->max_sndwnd = tp->snd_wnd;
10462 		rack->r_wanted_output = 1;
10463 	} else if (thflags & TH_ACK) {
10464 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10465 			tp->snd_wnd = tiwin;
10466 			rack_validate_fo_sendwin_up(tp, rack);
10467 			tp->snd_wl1 = th->th_seq;
10468 			tp->snd_wl2 = th->th_ack;
10469 		}
10470 	}
10471 	if (tp->snd_wnd < ctf_outstanding(tp))
10472 		/* The peer collapsed the window */
10473 		rack_collapsed_window(rack);
10474 	else if (rack->rc_has_collapsed)
10475 		rack_un_collapse_window(rack);
10476 	/* Was persist timer active and now we have window space? */
10477 	if ((rack->rc_in_persist != 0) &&
10478 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10479 				rack->r_ctl.rc_pace_min_segs))) {
10480 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10481 		tp->snd_nxt = tp->snd_max;
10482 		/* Make sure we output to start the timer */
10483 		rack->r_wanted_output = 1;
10484 	}
10485 	/* Do we enter persists? */
10486 	if ((rack->rc_in_persist == 0) &&
10487 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10488 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10489 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10490 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10491 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10492 		/*
10493 		 * Here the rwnd is less than
10494 		 * the pacing size, we are established,
10495 		 * nothing is outstanding, and there is
10496 		 * data to send. Enter persists.
10497 		 */
10498 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10499 	}
10500 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10501 		m_freem(m);
10502 		return (0);
10503 	}
10504 	/*
10505 	 * don't process the URG bit, ignore them drag
10506 	 * along the up.
10507 	 */
10508 	tp->rcv_up = tp->rcv_nxt;
10509 	INP_WLOCK_ASSERT(tp->t_inpcb);
10510 
10511 	/*
10512 	 * Process the segment text, merging it into the TCP sequencing
10513 	 * queue, and arranging for acknowledgment of receipt if necessary.
10514 	 * This process logically involves adjusting tp->rcv_wnd as data is
10515 	 * presented to the user (this happens in tcp_usrreq.c, case
10516 	 * PRU_RCVD).  If a FIN has already been received on this connection
10517 	 * then we just ignore the text.
10518 	 */
10519 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10520 		   IS_FASTOPEN(tp->t_flags));
10521 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10522 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10523 		tcp_seq save_start = th->th_seq;
10524 		tcp_seq save_rnxt  = tp->rcv_nxt;
10525 		int     save_tlen  = tlen;
10526 
10527 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10528 		/*
10529 		 * Insert segment which includes th into TCP reassembly
10530 		 * queue with control block tp.  Set thflags to whether
10531 		 * reassembly now includes a segment with FIN.  This handles
10532 		 * the common case inline (segment is the next to be
10533 		 * received on an established connection, and the queue is
10534 		 * empty), avoiding linkage into and removal from the queue
10535 		 * and repetition of various conversions. Set DELACK for
10536 		 * segments received in order, but ack immediately when
10537 		 * segments are out of order (so fast retransmit can work).
10538 		 */
10539 		if (th->th_seq == tp->rcv_nxt &&
10540 		    SEGQ_EMPTY(tp) &&
10541 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
10542 		    tfo_syn)) {
10543 #ifdef NETFLIX_SB_LIMITS
10544 			u_int mcnt, appended;
10545 
10546 			if (so->so_rcv.sb_shlim) {
10547 				mcnt = m_memcnt(m);
10548 				appended = 0;
10549 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10550 				    CFO_NOSLEEP, NULL) == false) {
10551 					counter_u64_add(tcp_sb_shlim_fails, 1);
10552 					m_freem(m);
10553 					return (0);
10554 				}
10555 			}
10556 #endif
10557 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10558 			tp->rcv_nxt += tlen;
10559 			if (tlen &&
10560 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10561 			    (tp->t_fbyte_in == 0)) {
10562 				tp->t_fbyte_in = ticks;
10563 				if (tp->t_fbyte_in == 0)
10564 					tp->t_fbyte_in = 1;
10565 				if (tp->t_fbyte_out && tp->t_fbyte_in)
10566 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10567 			}
10568 			thflags = tcp_get_flags(th) & TH_FIN;
10569 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10570 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10571 			SOCKBUF_LOCK(&so->so_rcv);
10572 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10573 				m_freem(m);
10574 			} else
10575 #ifdef NETFLIX_SB_LIMITS
10576 				appended =
10577 #endif
10578 					sbappendstream_locked(&so->so_rcv, m, 0);
10579 
10580 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10581 			/* NB: sorwakeup_locked() does an implicit unlock. */
10582 			sorwakeup_locked(so);
10583 #ifdef NETFLIX_SB_LIMITS
10584 			if (so->so_rcv.sb_shlim && appended != mcnt)
10585 				counter_fo_release(so->so_rcv.sb_shlim,
10586 				    mcnt - appended);
10587 #endif
10588 		} else {
10589 			/*
10590 			 * XXX: Due to the header drop above "th" is
10591 			 * theoretically invalid by now.  Fortunately
10592 			 * m_adj() doesn't actually frees any mbufs when
10593 			 * trimming from the head.
10594 			 */
10595 			tcp_seq temp = save_start;
10596 
10597 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
10598 			tp->t_flags |= TF_ACKNOW;
10599 			if (tp->t_flags & TF_WAKESOR) {
10600 				tp->t_flags &= ~TF_WAKESOR;
10601 				/* NB: sorwakeup_locked() does an implicit unlock. */
10602 				sorwakeup_locked(so);
10603 			}
10604 		}
10605 		if ((tp->t_flags & TF_SACK_PERMIT) &&
10606 		    (save_tlen > 0) &&
10607 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
10608 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10609 				/*
10610 				 * DSACK actually handled in the fastpath
10611 				 * above.
10612 				 */
10613 				RACK_OPTS_INC(tcp_sack_path_1);
10614 				tcp_update_sack_list(tp, save_start,
10615 				    save_start + save_tlen);
10616 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10617 				if ((tp->rcv_numsacks >= 1) &&
10618 				    (tp->sackblks[0].end == save_start)) {
10619 					/*
10620 					 * Partial overlap, recorded at todrop
10621 					 * above.
10622 					 */
10623 					RACK_OPTS_INC(tcp_sack_path_2a);
10624 					tcp_update_sack_list(tp,
10625 					    tp->sackblks[0].start,
10626 					    tp->sackblks[0].end);
10627 				} else {
10628 					RACK_OPTS_INC(tcp_sack_path_2b);
10629 					tcp_update_dsack_list(tp, save_start,
10630 					    save_start + save_tlen);
10631 				}
10632 			} else if (tlen >= save_tlen) {
10633 				/* Update of sackblks. */
10634 				RACK_OPTS_INC(tcp_sack_path_3);
10635 				tcp_update_dsack_list(tp, save_start,
10636 				    save_start + save_tlen);
10637 			} else if (tlen > 0) {
10638 				RACK_OPTS_INC(tcp_sack_path_4);
10639 				tcp_update_dsack_list(tp, save_start,
10640 				    save_start + tlen);
10641 			}
10642 		}
10643 	} else {
10644 		m_freem(m);
10645 		thflags &= ~TH_FIN;
10646 	}
10647 
10648 	/*
10649 	 * If FIN is received ACK the FIN and let the user know that the
10650 	 * connection is closing.
10651 	 */
10652 	if (thflags & TH_FIN) {
10653 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10654 			/* The socket upcall is handled by socantrcvmore. */
10655 			socantrcvmore(so);
10656 			/*
10657 			 * If connection is half-synchronized (ie NEEDSYN
10658 			 * flag on) then delay ACK, so it may be piggybacked
10659 			 * when SYN is sent. Otherwise, since we received a
10660 			 * FIN then no more input can be expected, send ACK
10661 			 * now.
10662 			 */
10663 			if (tp->t_flags & TF_NEEDSYN) {
10664 				rack_timer_cancel(tp, rack,
10665 				    rack->r_ctl.rc_rcvtime, __LINE__);
10666 				tp->t_flags |= TF_DELACK;
10667 			} else {
10668 				tp->t_flags |= TF_ACKNOW;
10669 			}
10670 			tp->rcv_nxt++;
10671 		}
10672 		switch (tp->t_state) {
10673 			/*
10674 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
10675 			 * CLOSE_WAIT state.
10676 			 */
10677 		case TCPS_SYN_RECEIVED:
10678 			tp->t_starttime = ticks;
10679 			/* FALLTHROUGH */
10680 		case TCPS_ESTABLISHED:
10681 			rack_timer_cancel(tp, rack,
10682 			    rack->r_ctl.rc_rcvtime, __LINE__);
10683 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
10684 			break;
10685 
10686 			/*
10687 			 * If still in FIN_WAIT_1 STATE FIN has not been
10688 			 * acked so enter the CLOSING state.
10689 			 */
10690 		case TCPS_FIN_WAIT_1:
10691 			rack_timer_cancel(tp, rack,
10692 			    rack->r_ctl.rc_rcvtime, __LINE__);
10693 			tcp_state_change(tp, TCPS_CLOSING);
10694 			break;
10695 
10696 			/*
10697 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
10698 			 * starting the time-wait timer, turning off the
10699 			 * other standard timers.
10700 			 */
10701 		case TCPS_FIN_WAIT_2:
10702 			rack_timer_cancel(tp, rack,
10703 			    rack->r_ctl.rc_rcvtime, __LINE__);
10704 			tcp_twstart(tp);
10705 			return (1);
10706 		}
10707 	}
10708 	/*
10709 	 * Return any desired output.
10710 	 */
10711 	if ((tp->t_flags & TF_ACKNOW) ||
10712 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10713 		rack->r_wanted_output = 1;
10714 	}
10715 	INP_WLOCK_ASSERT(tp->t_inpcb);
10716 	return (0);
10717 }
10718 
10719 /*
10720  * Here nothing is really faster, its just that we
10721  * have broken out the fast-data path also just like
10722  * the fast-ack.
10723  */
10724 static int
10725 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10726     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10727     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10728 {
10729 	int32_t nsegs;
10730 	int32_t newsize = 0;	/* automatic sockbuf scaling */
10731 	struct tcp_rack *rack;
10732 #ifdef NETFLIX_SB_LIMITS
10733 	u_int mcnt, appended;
10734 #endif
10735 #ifdef TCPDEBUG
10736 	/*
10737 	 * The size of tcp_saveipgen must be the size of the max ip header,
10738 	 * now IPv6.
10739 	 */
10740 	u_char tcp_saveipgen[IP6_HDR_LEN];
10741 	struct tcphdr tcp_savetcp;
10742 	short ostate = 0;
10743 
10744 #endif
10745 	/*
10746 	 * If last ACK falls within this segment's sequence numbers, record
10747 	 * the timestamp. NOTE that the test is modified according to the
10748 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10749 	 */
10750 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10751 		return (0);
10752 	}
10753 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10754 		return (0);
10755 	}
10756 	if (tiwin && tiwin != tp->snd_wnd) {
10757 		return (0);
10758 	}
10759 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10760 		return (0);
10761 	}
10762 	if (__predict_false((to->to_flags & TOF_TS) &&
10763 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10764 		return (0);
10765 	}
10766 	if (__predict_false((th->th_ack != tp->snd_una))) {
10767 		return (0);
10768 	}
10769 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10770 		return (0);
10771 	}
10772 	if ((to->to_flags & TOF_TS) != 0 &&
10773 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10774 		tp->ts_recent_age = tcp_ts_getticks();
10775 		tp->ts_recent = to->to_tsval;
10776 	}
10777 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10778 	/*
10779 	 * This is a pure, in-sequence data packet with nothing on the
10780 	 * reassembly queue and we have enough buffer space to take it.
10781 	 */
10782 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10783 
10784 #ifdef NETFLIX_SB_LIMITS
10785 	if (so->so_rcv.sb_shlim) {
10786 		mcnt = m_memcnt(m);
10787 		appended = 0;
10788 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10789 		    CFO_NOSLEEP, NULL) == false) {
10790 			counter_u64_add(tcp_sb_shlim_fails, 1);
10791 			m_freem(m);
10792 			return (1);
10793 		}
10794 	}
10795 #endif
10796 	/* Clean receiver SACK report if present */
10797 	if (tp->rcv_numsacks)
10798 		tcp_clean_sackreport(tp);
10799 	KMOD_TCPSTAT_INC(tcps_preddat);
10800 	tp->rcv_nxt += tlen;
10801 	if (tlen &&
10802 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10803 	    (tp->t_fbyte_in == 0)) {
10804 		tp->t_fbyte_in = ticks;
10805 		if (tp->t_fbyte_in == 0)
10806 			tp->t_fbyte_in = 1;
10807 		if (tp->t_fbyte_out && tp->t_fbyte_in)
10808 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10809 	}
10810 	/*
10811 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10812 	 */
10813 	tp->snd_wl1 = th->th_seq;
10814 	/*
10815 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10816 	 */
10817 	tp->rcv_up = tp->rcv_nxt;
10818 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10819 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10820 #ifdef TCPDEBUG
10821 	if (so->so_options & SO_DEBUG)
10822 		tcp_trace(TA_INPUT, ostate, tp,
10823 		    (void *)tcp_saveipgen, &tcp_savetcp, 0);
10824 #endif
10825 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10826 
10827 	/* Add data to socket buffer. */
10828 	SOCKBUF_LOCK(&so->so_rcv);
10829 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10830 		m_freem(m);
10831 	} else {
10832 		/*
10833 		 * Set new socket buffer size. Give up when limit is
10834 		 * reached.
10835 		 */
10836 		if (newsize)
10837 			if (!sbreserve_locked(&so->so_rcv,
10838 			    newsize, so, NULL))
10839 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10840 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10841 #ifdef NETFLIX_SB_LIMITS
10842 		appended =
10843 #endif
10844 			sbappendstream_locked(&so->so_rcv, m, 0);
10845 		ctf_calc_rwin(so, tp);
10846 	}
10847 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10848 	/* NB: sorwakeup_locked() does an implicit unlock. */
10849 	sorwakeup_locked(so);
10850 #ifdef NETFLIX_SB_LIMITS
10851 	if (so->so_rcv.sb_shlim && mcnt != appended)
10852 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10853 #endif
10854 	rack_handle_delayed_ack(tp, rack, tlen, 0);
10855 	if (tp->snd_una == tp->snd_max)
10856 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10857 	return (1);
10858 }
10859 
10860 /*
10861  * This subfunction is used to try to highly optimize the
10862  * fast path. We again allow window updates that are
10863  * in sequence to remain in the fast-path. We also add
10864  * in the __predict's to attempt to help the compiler.
10865  * Note that if we return a 0, then we can *not* process
10866  * it and the caller should push the packet into the
10867  * slow-path.
10868  */
10869 static int
10870 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10871     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10872     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10873 {
10874 	int32_t acked;
10875 	int32_t nsegs;
10876 #ifdef TCPDEBUG
10877 	/*
10878 	 * The size of tcp_saveipgen must be the size of the max ip header,
10879 	 * now IPv6.
10880 	 */
10881 	u_char tcp_saveipgen[IP6_HDR_LEN];
10882 	struct tcphdr tcp_savetcp;
10883 	short ostate = 0;
10884 #endif
10885 	int32_t under_pacing = 0;
10886 	struct tcp_rack *rack;
10887 
10888 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10889 		/* Old ack, behind (or duplicate to) the last one rcv'd */
10890 		return (0);
10891 	}
10892 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10893 		/* Above what we have sent? */
10894 		return (0);
10895 	}
10896 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10897 		/* We are retransmitting */
10898 		return (0);
10899 	}
10900 	if (__predict_false(tiwin == 0)) {
10901 		/* zero window */
10902 		return (0);
10903 	}
10904 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10905 		/* We need a SYN or a FIN, unlikely.. */
10906 		return (0);
10907 	}
10908 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10909 		/* Timestamp is behind .. old ack with seq wrap? */
10910 		return (0);
10911 	}
10912 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10913 		/* Still recovering */
10914 		return (0);
10915 	}
10916 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10917 	if (rack->r_ctl.rc_sacked) {
10918 		/* We have sack holes on our scoreboard */
10919 		return (0);
10920 	}
10921 	/* Ok if we reach here, we can process a fast-ack */
10922 	if (rack->gp_ready &&
10923 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10924 		under_pacing = 1;
10925 	}
10926 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10927 	rack_log_ack(tp, to, th, 0, 0);
10928 	/* Did the window get updated? */
10929 	if (tiwin != tp->snd_wnd) {
10930 		tp->snd_wnd = tiwin;
10931 		rack_validate_fo_sendwin_up(tp, rack);
10932 		tp->snd_wl1 = th->th_seq;
10933 		if (tp->snd_wnd > tp->max_sndwnd)
10934 			tp->max_sndwnd = tp->snd_wnd;
10935 	}
10936 	/* Do we exit persists? */
10937 	if ((rack->rc_in_persist != 0) &&
10938 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10939 			       rack->r_ctl.rc_pace_min_segs))) {
10940 		rack_exit_persist(tp, rack, cts);
10941 	}
10942 	/* Do we enter persists? */
10943 	if ((rack->rc_in_persist == 0) &&
10944 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10945 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10946 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10947 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10948 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10949 		/*
10950 		 * Here the rwnd is less than
10951 		 * the pacing size, we are established,
10952 		 * nothing is outstanding, and there is
10953 		 * data to send. Enter persists.
10954 		 */
10955 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10956 	}
10957 	/*
10958 	 * If last ACK falls within this segment's sequence numbers, record
10959 	 * the timestamp. NOTE that the test is modified according to the
10960 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10961 	 */
10962 	if ((to->to_flags & TOF_TS) != 0 &&
10963 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10964 		tp->ts_recent_age = tcp_ts_getticks();
10965 		tp->ts_recent = to->to_tsval;
10966 	}
10967 	/*
10968 	 * This is a pure ack for outstanding data.
10969 	 */
10970 	KMOD_TCPSTAT_INC(tcps_predack);
10971 
10972 	/*
10973 	 * "bad retransmit" recovery.
10974 	 */
10975 	if ((tp->t_flags & TF_PREVVALID) &&
10976 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10977 		tp->t_flags &= ~TF_PREVVALID;
10978 		if (tp->t_rxtshift == 1 &&
10979 		    (int)(ticks - tp->t_badrxtwin) < 0)
10980 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10981 	}
10982 	/*
10983 	 * Recalculate the transmit timer / rtt.
10984 	 *
10985 	 * Some boxes send broken timestamp replies during the SYN+ACK
10986 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10987 	 * and blow up the retransmit timer.
10988 	 */
10989 	acked = BYTES_THIS_ACK(tp, th);
10990 
10991 #ifdef TCP_HHOOK
10992 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10993 	hhook_run_tcp_est_in(tp, th, to);
10994 #endif
10995 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10996 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10997 	if (acked) {
10998 		struct mbuf *mfree;
10999 
11000 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
11001 		SOCKBUF_LOCK(&so->so_snd);
11002 		mfree = sbcut_locked(&so->so_snd, acked);
11003 		tp->snd_una = th->th_ack;
11004 		/* Note we want to hold the sb lock through the sendmap adjust */
11005 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
11006 		/* Wake up the socket if we have room to write more */
11007 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11008 		sowwakeup_locked(so);
11009 		m_freem(mfree);
11010 		tp->t_rxtshift = 0;
11011 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11012 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11013 		rack->rc_tlp_in_progress = 0;
11014 		rack->r_ctl.rc_tlp_cnt_out = 0;
11015 		/*
11016 		 * If it is the RXT timer we want to
11017 		 * stop it, so we can restart a TLP.
11018 		 */
11019 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11020 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11021 #ifdef NETFLIX_HTTP_LOGGING
11022 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
11023 #endif
11024 	}
11025 	/*
11026 	 * Let the congestion control algorithm update congestion control
11027 	 * related information. This typically means increasing the
11028 	 * congestion window.
11029 	 */
11030 	if (tp->snd_wnd < ctf_outstanding(tp)) {
11031 		/* The peer collapsed the window */
11032 		rack_collapsed_window(rack);
11033 	} else if (rack->rc_has_collapsed)
11034 		rack_un_collapse_window(rack);
11035 
11036 	/*
11037 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
11038 	 */
11039 	tp->snd_wl2 = th->th_ack;
11040 	tp->t_dupacks = 0;
11041 	m_freem(m);
11042 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
11043 
11044 	/*
11045 	 * If all outstanding data are acked, stop retransmit timer,
11046 	 * otherwise restart timer using current (possibly backed-off)
11047 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
11048 	 * If data are ready to send, let tcp_output decide between more
11049 	 * output or persist.
11050 	 */
11051 #ifdef TCPDEBUG
11052 	if (so->so_options & SO_DEBUG)
11053 		tcp_trace(TA_INPUT, ostate, tp,
11054 		    (void *)tcp_saveipgen,
11055 		    &tcp_savetcp, 0);
11056 #endif
11057 	if (under_pacing &&
11058 	    (rack->use_fixed_rate == 0) &&
11059 	    (rack->in_probe_rtt == 0) &&
11060 	    rack->rc_gp_dyn_mul &&
11061 	    rack->rc_always_pace) {
11062 		/* Check if we are dragging bottom */
11063 		rack_check_bottom_drag(tp, rack, so, acked);
11064 	}
11065 	if (tp->snd_una == tp->snd_max) {
11066 		tp->t_flags &= ~TF_PREVVALID;
11067 		rack->r_ctl.retran_during_recovery = 0;
11068 		rack->r_ctl.dsack_byte_cnt = 0;
11069 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11070 		if (rack->r_ctl.rc_went_idle_time == 0)
11071 			rack->r_ctl.rc_went_idle_time = 1;
11072 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11073 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
11074 			tp->t_acktime = 0;
11075 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11076 	}
11077 	if (acked && rack->r_fast_output)
11078 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11079 	if (sbavail(&so->so_snd)) {
11080 		rack->r_wanted_output = 1;
11081 	}
11082 	return (1);
11083 }
11084 
11085 /*
11086  * Return value of 1, the TCB is unlocked and most
11087  * likely gone, return value of 0, the TCP is still
11088  * locked.
11089  */
11090 static int
11091 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11092     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11093     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11094 {
11095 	int32_t ret_val = 0;
11096 	int32_t todrop;
11097 	int32_t ourfinisacked = 0;
11098 	struct tcp_rack *rack;
11099 
11100 	ctf_calc_rwin(so, tp);
11101 	/*
11102 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
11103 	 * SYN, drop the input. if seg contains a RST, then drop the
11104 	 * connection. if seg does not contain SYN, then drop it. Otherwise
11105 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
11106 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
11107 	 * contains an ECE and ECN support is enabled, the stream is ECN
11108 	 * capable. if SYN has been acked change to ESTABLISHED else
11109 	 * SYN_RCVD state arrange for segment to be acked (eventually)
11110 	 * continue processing rest of data/controls.
11111 	 */
11112 	if ((thflags & TH_ACK) &&
11113 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
11114 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11115 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11116 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11117 		return (1);
11118 	}
11119 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11120 		TCP_PROBE5(connect__refused, NULL, tp,
11121 		    mtod(m, const char *), tp, th);
11122 		tp = tcp_drop(tp, ECONNREFUSED);
11123 		ctf_do_drop(m, tp);
11124 		return (1);
11125 	}
11126 	if (thflags & TH_RST) {
11127 		ctf_do_drop(m, tp);
11128 		return (1);
11129 	}
11130 	if (!(thflags & TH_SYN)) {
11131 		ctf_do_drop(m, tp);
11132 		return (1);
11133 	}
11134 	tp->irs = th->th_seq;
11135 	tcp_rcvseqinit(tp);
11136 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11137 	if (thflags & TH_ACK) {
11138 		int tfo_partial = 0;
11139 
11140 		KMOD_TCPSTAT_INC(tcps_connects);
11141 		soisconnected(so);
11142 #ifdef MAC
11143 		mac_socketpeer_set_from_mbuf(m, so);
11144 #endif
11145 		/* Do window scaling on this connection? */
11146 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11147 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11148 			tp->rcv_scale = tp->request_r_scale;
11149 		}
11150 		tp->rcv_adv += min(tp->rcv_wnd,
11151 		    TCP_MAXWIN << tp->rcv_scale);
11152 		/*
11153 		 * If not all the data that was sent in the TFO SYN
11154 		 * has been acked, resend the remainder right away.
11155 		 */
11156 		if (IS_FASTOPEN(tp->t_flags) &&
11157 		    (tp->snd_una != tp->snd_max)) {
11158 			tp->snd_nxt = th->th_ack;
11159 			tfo_partial = 1;
11160 		}
11161 		/*
11162 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
11163 		 * will be turned on later.
11164 		 */
11165 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11166 			rack_timer_cancel(tp, rack,
11167 					  rack->r_ctl.rc_rcvtime, __LINE__);
11168 			tp->t_flags |= TF_DELACK;
11169 		} else {
11170 			rack->r_wanted_output = 1;
11171 			tp->t_flags |= TF_ACKNOW;
11172 			rack->rc_dack_toggle = 0;
11173 		}
11174 
11175 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
11176 
11177 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
11178 			/*
11179 			 * We advance snd_una for the
11180 			 * fast open case. If th_ack is
11181 			 * acknowledging data beyond
11182 			 * snd_una we can't just call
11183 			 * ack-processing since the
11184 			 * data stream in our send-map
11185 			 * will start at snd_una + 1 (one
11186 			 * beyond the SYN). If its just
11187 			 * equal we don't need to do that
11188 			 * and there is no send_map.
11189 			 */
11190 			tp->snd_una++;
11191 		}
11192 		/*
11193 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11194 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11195 		 */
11196 		tp->t_starttime = ticks;
11197 		if (tp->t_flags & TF_NEEDFIN) {
11198 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
11199 			tp->t_flags &= ~TF_NEEDFIN;
11200 			thflags &= ~TH_SYN;
11201 		} else {
11202 			tcp_state_change(tp, TCPS_ESTABLISHED);
11203 			TCP_PROBE5(connect__established, NULL, tp,
11204 			    mtod(m, const char *), tp, th);
11205 			rack_cc_conn_init(tp);
11206 		}
11207 	} else {
11208 		/*
11209 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
11210 		 * open.  If segment contains CC option and there is a
11211 		 * cached CC, apply TAO test. If it succeeds, connection is *
11212 		 * half-synchronized. Otherwise, do 3-way handshake:
11213 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11214 		 * there was no CC option, clear cached CC value.
11215 		 */
11216 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
11217 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
11218 	}
11219 	INP_WLOCK_ASSERT(tp->t_inpcb);
11220 	/*
11221 	 * Advance th->th_seq to correspond to first data byte. If data,
11222 	 * trim to stay within window, dropping FIN if necessary.
11223 	 */
11224 	th->th_seq++;
11225 	if (tlen > tp->rcv_wnd) {
11226 		todrop = tlen - tp->rcv_wnd;
11227 		m_adj(m, -todrop);
11228 		tlen = tp->rcv_wnd;
11229 		thflags &= ~TH_FIN;
11230 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11231 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11232 	}
11233 	tp->snd_wl1 = th->th_seq - 1;
11234 	tp->rcv_up = th->th_seq;
11235 	/*
11236 	 * Client side of transaction: already sent SYN and data. If the
11237 	 * remote host used T/TCP to validate the SYN, our data will be
11238 	 * ACK'd; if so, enter normal data segment processing in the middle
11239 	 * of step 5, ack processing. Otherwise, goto step 6.
11240 	 */
11241 	if (thflags & TH_ACK) {
11242 		/* For syn-sent we need to possibly update the rtt */
11243 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11244 			uint32_t t, mcts;
11245 
11246 			mcts = tcp_ts_getticks();
11247 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11248 			if (!tp->t_rttlow || tp->t_rttlow > t)
11249 				tp->t_rttlow = t;
11250 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11251 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11252 			tcp_rack_xmit_timer_commit(rack, tp);
11253 		}
11254 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11255 			return (ret_val);
11256 		/* We may have changed to FIN_WAIT_1 above */
11257 		if (tp->t_state == TCPS_FIN_WAIT_1) {
11258 			/*
11259 			 * In FIN_WAIT_1 STATE in addition to the processing
11260 			 * for the ESTABLISHED state if our FIN is now
11261 			 * acknowledged then enter FIN_WAIT_2.
11262 			 */
11263 			if (ourfinisacked) {
11264 				/*
11265 				 * If we can't receive any more data, then
11266 				 * closing user can proceed. Starting the
11267 				 * timer is contrary to the specification,
11268 				 * but if we don't get a FIN we'll hang
11269 				 * forever.
11270 				 *
11271 				 * XXXjl: we should release the tp also, and
11272 				 * use a compressed state.
11273 				 */
11274 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11275 					soisdisconnected(so);
11276 					tcp_timer_activate(tp, TT_2MSL,
11277 					    (tcp_fast_finwait2_recycle ?
11278 					    tcp_finwait2_timeout :
11279 					    TP_MAXIDLE(tp)));
11280 				}
11281 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
11282 			}
11283 		}
11284 	}
11285 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11286 	   tiwin, thflags, nxt_pkt));
11287 }
11288 
11289 /*
11290  * Return value of 1, the TCB is unlocked and most
11291  * likely gone, return value of 0, the TCP is still
11292  * locked.
11293  */
11294 static int
11295 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11296     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11297     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11298 {
11299 	struct tcp_rack *rack;
11300 	int32_t ret_val = 0;
11301 	int32_t ourfinisacked = 0;
11302 
11303 	ctf_calc_rwin(so, tp);
11304 	if ((thflags & TH_ACK) &&
11305 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11306 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11307 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11308 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11309 		return (1);
11310 	}
11311 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11312 	if (IS_FASTOPEN(tp->t_flags)) {
11313 		/*
11314 		 * When a TFO connection is in SYN_RECEIVED, the
11315 		 * only valid packets are the initial SYN, a
11316 		 * retransmit/copy of the initial SYN (possibly with
11317 		 * a subset of the original data), a valid ACK, a
11318 		 * FIN, or a RST.
11319 		 */
11320 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11321 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11322 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11323 			return (1);
11324 		} else if (thflags & TH_SYN) {
11325 			/* non-initial SYN is ignored */
11326 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11327 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11328 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11329 				ctf_do_drop(m, NULL);
11330 				return (0);
11331 			}
11332 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11333 			ctf_do_drop(m, NULL);
11334 			return (0);
11335 		}
11336 	}
11337 
11338 	if ((thflags & TH_RST) ||
11339 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11340 		return (__ctf_process_rst(m, th, so, tp,
11341 					  &rack->r_ctl.challenge_ack_ts,
11342 					  &rack->r_ctl.challenge_ack_cnt));
11343 	/*
11344 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11345 	 * it's less than ts_recent, drop it.
11346 	 */
11347 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11348 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11349 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11350 			return (ret_val);
11351 	}
11352 	/*
11353 	 * In the SYN-RECEIVED state, validate that the packet belongs to
11354 	 * this connection before trimming the data to fit the receive
11355 	 * window.  Check the sequence number versus IRS since we know the
11356 	 * sequence numbers haven't wrapped.  This is a partial fix for the
11357 	 * "LAND" DoS attack.
11358 	 */
11359 	if (SEQ_LT(th->th_seq, tp->irs)) {
11360 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11361 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11362 		return (1);
11363 	}
11364 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11365 			      &rack->r_ctl.challenge_ack_ts,
11366 			      &rack->r_ctl.challenge_ack_cnt)) {
11367 		return (ret_val);
11368 	}
11369 	/*
11370 	 * If last ACK falls within this segment's sequence numbers, record
11371 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11372 	 * from the latest proposal of the tcplw@cray.com list (Braden
11373 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11374 	 * with our earlier PAWS tests, so this check should be solely
11375 	 * predicated on the sequence space of this segment. 3) That we
11376 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11377 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11378 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11379 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11380 	 * p.869. In such cases, we can still calculate the RTT correctly
11381 	 * when RCV.NXT == Last.ACK.Sent.
11382 	 */
11383 	if ((to->to_flags & TOF_TS) != 0 &&
11384 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11385 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11386 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11387 		tp->ts_recent_age = tcp_ts_getticks();
11388 		tp->ts_recent = to->to_tsval;
11389 	}
11390 	tp->snd_wnd = tiwin;
11391 	rack_validate_fo_sendwin_up(tp, rack);
11392 	/*
11393 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11394 	 * is on (half-synchronized state), then queue data for later
11395 	 * processing; else drop segment and return.
11396 	 */
11397 	if ((thflags & TH_ACK) == 0) {
11398 		if (IS_FASTOPEN(tp->t_flags)) {
11399 			rack_cc_conn_init(tp);
11400 		}
11401 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11402 		    tiwin, thflags, nxt_pkt));
11403 	}
11404 	KMOD_TCPSTAT_INC(tcps_connects);
11405 	soisconnected(so);
11406 	/* Do window scaling? */
11407 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11408 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11409 		tp->rcv_scale = tp->request_r_scale;
11410 	}
11411 	/*
11412 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11413 	 * FIN-WAIT-1
11414 	 */
11415 	tp->t_starttime = ticks;
11416 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11417 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11418 		tp->t_tfo_pending = NULL;
11419 	}
11420 	if (tp->t_flags & TF_NEEDFIN) {
11421 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
11422 		tp->t_flags &= ~TF_NEEDFIN;
11423 	} else {
11424 		tcp_state_change(tp, TCPS_ESTABLISHED);
11425 		TCP_PROBE5(accept__established, NULL, tp,
11426 		    mtod(m, const char *), tp, th);
11427 		/*
11428 		 * TFO connections call cc_conn_init() during SYN
11429 		 * processing.  Calling it again here for such connections
11430 		 * is not harmless as it would undo the snd_cwnd reduction
11431 		 * that occurs when a TFO SYN|ACK is retransmitted.
11432 		 */
11433 		if (!IS_FASTOPEN(tp->t_flags))
11434 			rack_cc_conn_init(tp);
11435 	}
11436 	/*
11437 	 * Account for the ACK of our SYN prior to
11438 	 * regular ACK processing below, except for
11439 	 * simultaneous SYN, which is handled later.
11440 	 */
11441 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11442 		tp->snd_una++;
11443 	/*
11444 	 * If segment contains data or ACK, will call tcp_reass() later; if
11445 	 * not, do so now to pass queued data to user.
11446 	 */
11447 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
11448 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11449 		    (struct mbuf *)0);
11450 		if (tp->t_flags & TF_WAKESOR) {
11451 			tp->t_flags &= ~TF_WAKESOR;
11452 			/* NB: sorwakeup_locked() does an implicit unlock. */
11453 			sorwakeup_locked(so);
11454 		}
11455 	}
11456 	tp->snd_wl1 = th->th_seq - 1;
11457 	/* For syn-recv we need to possibly update the rtt */
11458 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11459 		uint32_t t, mcts;
11460 
11461 		mcts = tcp_ts_getticks();
11462 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11463 		if (!tp->t_rttlow || tp->t_rttlow > t)
11464 			tp->t_rttlow = t;
11465 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11466 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11467 		tcp_rack_xmit_timer_commit(rack, tp);
11468 	}
11469 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11470 		return (ret_val);
11471 	}
11472 	if (tp->t_state == TCPS_FIN_WAIT_1) {
11473 		/* We could have went to FIN_WAIT_1 (or EST) above */
11474 		/*
11475 		 * In FIN_WAIT_1 STATE in addition to the processing for the
11476 		 * ESTABLISHED state if our FIN is now acknowledged then
11477 		 * enter FIN_WAIT_2.
11478 		 */
11479 		if (ourfinisacked) {
11480 			/*
11481 			 * If we can't receive any more data, then closing
11482 			 * user can proceed. Starting the timer is contrary
11483 			 * to the specification, but if we don't get a FIN
11484 			 * we'll hang forever.
11485 			 *
11486 			 * XXXjl: we should release the tp also, and use a
11487 			 * compressed state.
11488 			 */
11489 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11490 				soisdisconnected(so);
11491 				tcp_timer_activate(tp, TT_2MSL,
11492 				    (tcp_fast_finwait2_recycle ?
11493 				    tcp_finwait2_timeout :
11494 				    TP_MAXIDLE(tp)));
11495 			}
11496 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
11497 		}
11498 	}
11499 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11500 	    tiwin, thflags, nxt_pkt));
11501 }
11502 
11503 /*
11504  * Return value of 1, the TCB is unlocked and most
11505  * likely gone, return value of 0, the TCP is still
11506  * locked.
11507  */
11508 static int
11509 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11510     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11511     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11512 {
11513 	int32_t ret_val = 0;
11514 	struct tcp_rack *rack;
11515 
11516 	/*
11517 	 * Header prediction: check for the two common cases of a
11518 	 * uni-directional data xfer.  If the packet has no control flags,
11519 	 * is in-sequence, the window didn't change and we're not
11520 	 * retransmitting, it's a candidate.  If the length is zero and the
11521 	 * ack moved forward, we're the sender side of the xfer.  Just free
11522 	 * the data acked & wake any higher level process that was blocked
11523 	 * waiting for space.  If the length is non-zero and the ack didn't
11524 	 * move, we're the receiver side.  If we're getting packets in-order
11525 	 * (the reassembly queue is empty), add the data toc The socket
11526 	 * buffer and note that we need a delayed ack. Make sure that the
11527 	 * hidden state-flags are also off. Since we check for
11528 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11529 	 */
11530 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11531 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11532 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11533 	    __predict_true(SEGQ_EMPTY(tp)) &&
11534 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
11535 		if (tlen == 0) {
11536 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11537 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11538 				return (0);
11539 			}
11540 		} else {
11541 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11542 			    tiwin, nxt_pkt, iptos)) {
11543 				return (0);
11544 			}
11545 		}
11546 	}
11547 	ctf_calc_rwin(so, tp);
11548 
11549 	if ((thflags & TH_RST) ||
11550 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11551 		return (__ctf_process_rst(m, th, so, tp,
11552 					  &rack->r_ctl.challenge_ack_ts,
11553 					  &rack->r_ctl.challenge_ack_cnt));
11554 
11555 	/*
11556 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11557 	 * synchronized state.
11558 	 */
11559 	if (thflags & TH_SYN) {
11560 		ctf_challenge_ack(m, th, tp, &ret_val);
11561 		return (ret_val);
11562 	}
11563 	/*
11564 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11565 	 * it's less than ts_recent, drop it.
11566 	 */
11567 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11568 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11569 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11570 			return (ret_val);
11571 	}
11572 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11573 			      &rack->r_ctl.challenge_ack_ts,
11574 			      &rack->r_ctl.challenge_ack_cnt)) {
11575 		return (ret_val);
11576 	}
11577 	/*
11578 	 * If last ACK falls within this segment's sequence numbers, record
11579 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11580 	 * from the latest proposal of the tcplw@cray.com list (Braden
11581 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11582 	 * with our earlier PAWS tests, so this check should be solely
11583 	 * predicated on the sequence space of this segment. 3) That we
11584 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11585 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11586 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11587 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11588 	 * p.869. In such cases, we can still calculate the RTT correctly
11589 	 * when RCV.NXT == Last.ACK.Sent.
11590 	 */
11591 	if ((to->to_flags & TOF_TS) != 0 &&
11592 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11593 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11594 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11595 		tp->ts_recent_age = tcp_ts_getticks();
11596 		tp->ts_recent = to->to_tsval;
11597 	}
11598 	/*
11599 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11600 	 * is on (half-synchronized state), then queue data for later
11601 	 * processing; else drop segment and return.
11602 	 */
11603 	if ((thflags & TH_ACK) == 0) {
11604 		if (tp->t_flags & TF_NEEDSYN) {
11605 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11606 			    tiwin, thflags, nxt_pkt));
11607 
11608 		} else if (tp->t_flags & TF_ACKNOW) {
11609 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11610 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11611 			return (ret_val);
11612 		} else {
11613 			ctf_do_drop(m, NULL);
11614 			return (0);
11615 		}
11616 	}
11617 	/*
11618 	 * Ack processing.
11619 	 */
11620 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11621 		return (ret_val);
11622 	}
11623 	if (sbavail(&so->so_snd)) {
11624 		if (ctf_progress_timeout_check(tp, true)) {
11625 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11626 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11627 			return (1);
11628 		}
11629 	}
11630 	/* State changes only happen in rack_process_data() */
11631 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11632 	    tiwin, thflags, nxt_pkt));
11633 }
11634 
11635 /*
11636  * Return value of 1, the TCB is unlocked and most
11637  * likely gone, return value of 0, the TCP is still
11638  * locked.
11639  */
11640 static int
11641 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11642     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11643     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11644 {
11645 	int32_t ret_val = 0;
11646 	struct tcp_rack *rack;
11647 
11648 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11649 	ctf_calc_rwin(so, tp);
11650 	if ((thflags & TH_RST) ||
11651 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11652 		return (__ctf_process_rst(m, th, so, tp,
11653 					  &rack->r_ctl.challenge_ack_ts,
11654 					  &rack->r_ctl.challenge_ack_cnt));
11655 	/*
11656 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11657 	 * synchronized state.
11658 	 */
11659 	if (thflags & TH_SYN) {
11660 		ctf_challenge_ack(m, th, tp, &ret_val);
11661 		return (ret_val);
11662 	}
11663 	/*
11664 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11665 	 * it's less than ts_recent, drop it.
11666 	 */
11667 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11668 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11669 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11670 			return (ret_val);
11671 	}
11672 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11673 			      &rack->r_ctl.challenge_ack_ts,
11674 			      &rack->r_ctl.challenge_ack_cnt)) {
11675 		return (ret_val);
11676 	}
11677 	/*
11678 	 * If last ACK falls within this segment's sequence numbers, record
11679 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11680 	 * from the latest proposal of the tcplw@cray.com list (Braden
11681 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11682 	 * with our earlier PAWS tests, so this check should be solely
11683 	 * predicated on the sequence space of this segment. 3) That we
11684 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11685 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11686 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11687 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11688 	 * p.869. In such cases, we can still calculate the RTT correctly
11689 	 * when RCV.NXT == Last.ACK.Sent.
11690 	 */
11691 	if ((to->to_flags & TOF_TS) != 0 &&
11692 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11693 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11694 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11695 		tp->ts_recent_age = tcp_ts_getticks();
11696 		tp->ts_recent = to->to_tsval;
11697 	}
11698 	/*
11699 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11700 	 * is on (half-synchronized state), then queue data for later
11701 	 * processing; else drop segment and return.
11702 	 */
11703 	if ((thflags & TH_ACK) == 0) {
11704 		if (tp->t_flags & TF_NEEDSYN) {
11705 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11706 			    tiwin, thflags, nxt_pkt));
11707 
11708 		} else if (tp->t_flags & TF_ACKNOW) {
11709 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11710 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11711 			return (ret_val);
11712 		} else {
11713 			ctf_do_drop(m, NULL);
11714 			return (0);
11715 		}
11716 	}
11717 	/*
11718 	 * Ack processing.
11719 	 */
11720 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11721 		return (ret_val);
11722 	}
11723 	if (sbavail(&so->so_snd)) {
11724 		if (ctf_progress_timeout_check(tp, true)) {
11725 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11726 						tp, tick, PROGRESS_DROP, __LINE__);
11727 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11728 			return (1);
11729 		}
11730 	}
11731 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11732 	    tiwin, thflags, nxt_pkt));
11733 }
11734 
11735 static int
11736 rack_check_data_after_close(struct mbuf *m,
11737     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11738 {
11739 	struct tcp_rack *rack;
11740 
11741 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11742 	if (rack->rc_allow_data_af_clo == 0) {
11743 	close_now:
11744 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11745 		/* tcp_close will kill the inp pre-log the Reset */
11746 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11747 		tp = tcp_close(tp);
11748 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11749 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11750 		return (1);
11751 	}
11752 	if (sbavail(&so->so_snd) == 0)
11753 		goto close_now;
11754 	/* Ok we allow data that is ignored and a followup reset */
11755 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11756 	tp->rcv_nxt = th->th_seq + *tlen;
11757 	tp->t_flags2 |= TF2_DROP_AF_DATA;
11758 	rack->r_wanted_output = 1;
11759 	*tlen = 0;
11760 	return (0);
11761 }
11762 
11763 /*
11764  * Return value of 1, the TCB is unlocked and most
11765  * likely gone, return value of 0, the TCP is still
11766  * locked.
11767  */
11768 static int
11769 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11770     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11771     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11772 {
11773 	int32_t ret_val = 0;
11774 	int32_t ourfinisacked = 0;
11775 	struct tcp_rack *rack;
11776 
11777 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11778 	ctf_calc_rwin(so, tp);
11779 
11780 	if ((thflags & TH_RST) ||
11781 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11782 		return (__ctf_process_rst(m, th, so, tp,
11783 					  &rack->r_ctl.challenge_ack_ts,
11784 					  &rack->r_ctl.challenge_ack_cnt));
11785 	/*
11786 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11787 	 * synchronized state.
11788 	 */
11789 	if (thflags & TH_SYN) {
11790 		ctf_challenge_ack(m, th, tp, &ret_val);
11791 		return (ret_val);
11792 	}
11793 	/*
11794 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11795 	 * it's less than ts_recent, drop it.
11796 	 */
11797 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11798 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11799 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11800 			return (ret_val);
11801 	}
11802 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11803 			      &rack->r_ctl.challenge_ack_ts,
11804 			      &rack->r_ctl.challenge_ack_cnt)) {
11805 		return (ret_val);
11806 	}
11807 	/*
11808 	 * If new data are received on a connection after the user processes
11809 	 * are gone, then RST the other end.
11810 	 */
11811 	if ((so->so_state & SS_NOFDREF) && tlen) {
11812 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
11813 			return (1);
11814 	}
11815 	/*
11816 	 * If last ACK falls within this segment's sequence numbers, record
11817 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11818 	 * from the latest proposal of the tcplw@cray.com list (Braden
11819 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11820 	 * with our earlier PAWS tests, so this check should be solely
11821 	 * predicated on the sequence space of this segment. 3) That we
11822 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11823 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11824 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11825 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11826 	 * p.869. In such cases, we can still calculate the RTT correctly
11827 	 * when RCV.NXT == Last.ACK.Sent.
11828 	 */
11829 	if ((to->to_flags & TOF_TS) != 0 &&
11830 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11831 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11832 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11833 		tp->ts_recent_age = tcp_ts_getticks();
11834 		tp->ts_recent = to->to_tsval;
11835 	}
11836 	/*
11837 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11838 	 * is on (half-synchronized state), then queue data for later
11839 	 * processing; else drop segment and return.
11840 	 */
11841 	if ((thflags & TH_ACK) == 0) {
11842 		if (tp->t_flags & TF_NEEDSYN) {
11843 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11844 			    tiwin, thflags, nxt_pkt));
11845 		} else if (tp->t_flags & TF_ACKNOW) {
11846 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11847 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11848 			return (ret_val);
11849 		} else {
11850 			ctf_do_drop(m, NULL);
11851 			return (0);
11852 		}
11853 	}
11854 	/*
11855 	 * Ack processing.
11856 	 */
11857 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11858 		return (ret_val);
11859 	}
11860 	if (ourfinisacked) {
11861 		/*
11862 		 * If we can't receive any more data, then closing user can
11863 		 * proceed. Starting the timer is contrary to the
11864 		 * specification, but if we don't get a FIN we'll hang
11865 		 * forever.
11866 		 *
11867 		 * XXXjl: we should release the tp also, and use a
11868 		 * compressed state.
11869 		 */
11870 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11871 			soisdisconnected(so);
11872 			tcp_timer_activate(tp, TT_2MSL,
11873 			    (tcp_fast_finwait2_recycle ?
11874 			    tcp_finwait2_timeout :
11875 			    TP_MAXIDLE(tp)));
11876 		}
11877 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
11878 	}
11879 	if (sbavail(&so->so_snd)) {
11880 		if (ctf_progress_timeout_check(tp, true)) {
11881 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11882 						tp, tick, PROGRESS_DROP, __LINE__);
11883 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11884 			return (1);
11885 		}
11886 	}
11887 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11888 	    tiwin, thflags, nxt_pkt));
11889 }
11890 
11891 /*
11892  * Return value of 1, the TCB is unlocked and most
11893  * likely gone, return value of 0, the TCP is still
11894  * locked.
11895  */
11896 static int
11897 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11898     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11899     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11900 {
11901 	int32_t ret_val = 0;
11902 	int32_t ourfinisacked = 0;
11903 	struct tcp_rack *rack;
11904 
11905 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11906 	ctf_calc_rwin(so, tp);
11907 
11908 	if ((thflags & TH_RST) ||
11909 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11910 		return (__ctf_process_rst(m, th, so, tp,
11911 					  &rack->r_ctl.challenge_ack_ts,
11912 					  &rack->r_ctl.challenge_ack_cnt));
11913 	/*
11914 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11915 	 * synchronized state.
11916 	 */
11917 	if (thflags & TH_SYN) {
11918 		ctf_challenge_ack(m, th, tp, &ret_val);
11919 		return (ret_val);
11920 	}
11921 	/*
11922 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11923 	 * it's less than ts_recent, drop it.
11924 	 */
11925 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11926 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11927 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11928 			return (ret_val);
11929 	}
11930 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11931 			      &rack->r_ctl.challenge_ack_ts,
11932 			      &rack->r_ctl.challenge_ack_cnt)) {
11933 		return (ret_val);
11934 	}
11935 	/*
11936 	 * If new data are received on a connection after the user processes
11937 	 * are gone, then RST the other end.
11938 	 */
11939 	if ((so->so_state & SS_NOFDREF) && tlen) {
11940 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
11941 			return (1);
11942 	}
11943 	/*
11944 	 * If last ACK falls within this segment's sequence numbers, record
11945 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11946 	 * from the latest proposal of the tcplw@cray.com list (Braden
11947 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11948 	 * with our earlier PAWS tests, so this check should be solely
11949 	 * predicated on the sequence space of this segment. 3) That we
11950 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11951 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11952 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11953 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11954 	 * p.869. In such cases, we can still calculate the RTT correctly
11955 	 * when RCV.NXT == Last.ACK.Sent.
11956 	 */
11957 	if ((to->to_flags & TOF_TS) != 0 &&
11958 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11959 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11960 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11961 		tp->ts_recent_age = tcp_ts_getticks();
11962 		tp->ts_recent = to->to_tsval;
11963 	}
11964 	/*
11965 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11966 	 * is on (half-synchronized state), then queue data for later
11967 	 * processing; else drop segment and return.
11968 	 */
11969 	if ((thflags & TH_ACK) == 0) {
11970 		if (tp->t_flags & TF_NEEDSYN) {
11971 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11972 			    tiwin, thflags, nxt_pkt));
11973 		} else if (tp->t_flags & TF_ACKNOW) {
11974 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11975 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11976 			return (ret_val);
11977 		} else {
11978 			ctf_do_drop(m, NULL);
11979 			return (0);
11980 		}
11981 	}
11982 	/*
11983 	 * Ack processing.
11984 	 */
11985 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11986 		return (ret_val);
11987 	}
11988 	if (ourfinisacked) {
11989 		tcp_twstart(tp);
11990 		m_freem(m);
11991 		return (1);
11992 	}
11993 	if (sbavail(&so->so_snd)) {
11994 		if (ctf_progress_timeout_check(tp, true)) {
11995 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11996 						tp, tick, PROGRESS_DROP, __LINE__);
11997 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11998 			return (1);
11999 		}
12000 	}
12001 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12002 	    tiwin, thflags, nxt_pkt));
12003 }
12004 
12005 /*
12006  * Return value of 1, the TCB is unlocked and most
12007  * likely gone, return value of 0, the TCP is still
12008  * locked.
12009  */
12010 static int
12011 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12012     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12013     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12014 {
12015 	int32_t ret_val = 0;
12016 	int32_t ourfinisacked = 0;
12017 	struct tcp_rack *rack;
12018 
12019 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12020 	ctf_calc_rwin(so, tp);
12021 
12022 	if ((thflags & TH_RST) ||
12023 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12024 		return (__ctf_process_rst(m, th, so, tp,
12025 					  &rack->r_ctl.challenge_ack_ts,
12026 					  &rack->r_ctl.challenge_ack_cnt));
12027 	/*
12028 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12029 	 * synchronized state.
12030 	 */
12031 	if (thflags & TH_SYN) {
12032 		ctf_challenge_ack(m, th, tp, &ret_val);
12033 		return (ret_val);
12034 	}
12035 	/*
12036 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12037 	 * it's less than ts_recent, drop it.
12038 	 */
12039 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12040 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12041 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12042 			return (ret_val);
12043 	}
12044 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12045 			      &rack->r_ctl.challenge_ack_ts,
12046 			      &rack->r_ctl.challenge_ack_cnt)) {
12047 		return (ret_val);
12048 	}
12049 	/*
12050 	 * If new data are received on a connection after the user processes
12051 	 * are gone, then RST the other end.
12052 	 */
12053 	if ((so->so_state & SS_NOFDREF) && tlen) {
12054 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12055 			return (1);
12056 	}
12057 	/*
12058 	 * If last ACK falls within this segment's sequence numbers, record
12059 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12060 	 * from the latest proposal of the tcplw@cray.com list (Braden
12061 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12062 	 * with our earlier PAWS tests, so this check should be solely
12063 	 * predicated on the sequence space of this segment. 3) That we
12064 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12065 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12066 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12067 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12068 	 * p.869. In such cases, we can still calculate the RTT correctly
12069 	 * when RCV.NXT == Last.ACK.Sent.
12070 	 */
12071 	if ((to->to_flags & TOF_TS) != 0 &&
12072 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12073 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12074 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12075 		tp->ts_recent_age = tcp_ts_getticks();
12076 		tp->ts_recent = to->to_tsval;
12077 	}
12078 	/*
12079 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12080 	 * is on (half-synchronized state), then queue data for later
12081 	 * processing; else drop segment and return.
12082 	 */
12083 	if ((thflags & TH_ACK) == 0) {
12084 		if (tp->t_flags & TF_NEEDSYN) {
12085 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12086 			    tiwin, thflags, nxt_pkt));
12087 		} else if (tp->t_flags & TF_ACKNOW) {
12088 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12089 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12090 			return (ret_val);
12091 		} else {
12092 			ctf_do_drop(m, NULL);
12093 			return (0);
12094 		}
12095 	}
12096 	/*
12097 	 * case TCPS_LAST_ACK: Ack processing.
12098 	 */
12099 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12100 		return (ret_val);
12101 	}
12102 	if (ourfinisacked) {
12103 		tp = tcp_close(tp);
12104 		ctf_do_drop(m, tp);
12105 		return (1);
12106 	}
12107 	if (sbavail(&so->so_snd)) {
12108 		if (ctf_progress_timeout_check(tp, true)) {
12109 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12110 						tp, tick, PROGRESS_DROP, __LINE__);
12111 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12112 			return (1);
12113 		}
12114 	}
12115 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12116 	    tiwin, thflags, nxt_pkt));
12117 }
12118 
12119 /*
12120  * Return value of 1, the TCB is unlocked and most
12121  * likely gone, return value of 0, the TCP is still
12122  * locked.
12123  */
12124 static int
12125 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12126     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12127     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12128 {
12129 	int32_t ret_val = 0;
12130 	int32_t ourfinisacked = 0;
12131 	struct tcp_rack *rack;
12132 
12133 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12134 	ctf_calc_rwin(so, tp);
12135 
12136 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
12137 	if ((thflags & TH_RST) ||
12138 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12139 		return (__ctf_process_rst(m, th, so, tp,
12140 					  &rack->r_ctl.challenge_ack_ts,
12141 					  &rack->r_ctl.challenge_ack_cnt));
12142 	/*
12143 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12144 	 * synchronized state.
12145 	 */
12146 	if (thflags & TH_SYN) {
12147 		ctf_challenge_ack(m, th, tp, &ret_val);
12148 		return (ret_val);
12149 	}
12150 	/*
12151 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12152 	 * it's less than ts_recent, drop it.
12153 	 */
12154 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12155 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12156 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12157 			return (ret_val);
12158 	}
12159 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12160 			      &rack->r_ctl.challenge_ack_ts,
12161 			      &rack->r_ctl.challenge_ack_cnt)) {
12162 		return (ret_val);
12163 	}
12164 	/*
12165 	 * If new data are received on a connection after the user processes
12166 	 * are gone, then RST the other end.
12167 	 */
12168 	if ((so->so_state & SS_NOFDREF) &&
12169 	    tlen) {
12170 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12171 			return (1);
12172 	}
12173 	/*
12174 	 * If last ACK falls within this segment's sequence numbers, record
12175 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12176 	 * from the latest proposal of the tcplw@cray.com list (Braden
12177 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12178 	 * with our earlier PAWS tests, so this check should be solely
12179 	 * predicated on the sequence space of this segment. 3) That we
12180 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12181 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12182 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12183 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12184 	 * p.869. In such cases, we can still calculate the RTT correctly
12185 	 * when RCV.NXT == Last.ACK.Sent.
12186 	 */
12187 	if ((to->to_flags & TOF_TS) != 0 &&
12188 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12189 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12190 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12191 		tp->ts_recent_age = tcp_ts_getticks();
12192 		tp->ts_recent = to->to_tsval;
12193 	}
12194 	/*
12195 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12196 	 * is on (half-synchronized state), then queue data for later
12197 	 * processing; else drop segment and return.
12198 	 */
12199 	if ((thflags & TH_ACK) == 0) {
12200 		if (tp->t_flags & TF_NEEDSYN) {
12201 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12202 			    tiwin, thflags, nxt_pkt));
12203 		} else if (tp->t_flags & TF_ACKNOW) {
12204 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12205 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12206 			return (ret_val);
12207 		} else {
12208 			ctf_do_drop(m, NULL);
12209 			return (0);
12210 		}
12211 	}
12212 	/*
12213 	 * Ack processing.
12214 	 */
12215 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12216 		return (ret_val);
12217 	}
12218 	if (sbavail(&so->so_snd)) {
12219 		if (ctf_progress_timeout_check(tp, true)) {
12220 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12221 						tp, tick, PROGRESS_DROP, __LINE__);
12222 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12223 			return (1);
12224 		}
12225 	}
12226 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12227 	    tiwin, thflags, nxt_pkt));
12228 }
12229 
12230 static void inline
12231 rack_clear_rate_sample(struct tcp_rack *rack)
12232 {
12233 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12234 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12235 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12236 }
12237 
12238 static void
12239 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12240 {
12241 	uint64_t bw_est, rate_wanted;
12242 	int chged = 0;
12243 	uint32_t user_max, orig_min, orig_max;
12244 
12245 	orig_min = rack->r_ctl.rc_pace_min_segs;
12246 	orig_max = rack->r_ctl.rc_pace_max_segs;
12247 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12248 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12249 		chged = 1;
12250 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12251 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12252 		if (user_max != rack->r_ctl.rc_pace_max_segs)
12253 			chged = 1;
12254 	}
12255 	if (rack->rc_force_max_seg) {
12256 		rack->r_ctl.rc_pace_max_segs = user_max;
12257 	} else if (rack->use_fixed_rate) {
12258 		bw_est = rack_get_bw(rack);
12259 		if ((rack->r_ctl.crte == NULL) ||
12260 		    (bw_est != rack->r_ctl.crte->rate)) {
12261 			rack->r_ctl.rc_pace_max_segs = user_max;
12262 		} else {
12263 			/* We are pacing right at the hardware rate */
12264 			uint32_t segsiz;
12265 
12266 			segsiz = min(ctf_fixed_maxseg(tp),
12267 				     rack->r_ctl.rc_pace_min_segs);
12268 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12269 				                           tp, bw_est, segsiz, 0,
12270 							   rack->r_ctl.crte, NULL);
12271 		}
12272 	} else if (rack->rc_always_pace) {
12273 		if (rack->r_ctl.gp_bw ||
12274 #ifdef NETFLIX_PEAKRATE
12275 		    rack->rc_tp->t_maxpeakrate ||
12276 #endif
12277 		    rack->r_ctl.init_rate) {
12278 			/* We have a rate of some sort set */
12279 			uint32_t  orig;
12280 
12281 			bw_est = rack_get_bw(rack);
12282 			orig = rack->r_ctl.rc_pace_max_segs;
12283 			if (fill_override)
12284 				rate_wanted = *fill_override;
12285 			else
12286 				rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12287 			if (rate_wanted) {
12288 				/* We have something */
12289 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12290 										   rate_wanted,
12291 										   ctf_fixed_maxseg(rack->rc_tp));
12292 			} else
12293 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12294 			if (orig != rack->r_ctl.rc_pace_max_segs)
12295 				chged = 1;
12296 		} else if ((rack->r_ctl.gp_bw == 0) &&
12297 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
12298 			/*
12299 			 * If we have nothing limit us to bursting
12300 			 * out IW sized pieces.
12301 			 */
12302 			chged = 1;
12303 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12304 		}
12305 	}
12306 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12307 		chged = 1;
12308 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12309 	}
12310 	if (chged)
12311 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12312 }
12313 
12314 
12315 static void
12316 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12317 {
12318 #ifdef INET6
12319 	struct ip6_hdr *ip6 = NULL;
12320 #endif
12321 #ifdef INET
12322 	struct ip *ip = NULL;
12323 #endif
12324 	struct udphdr *udp = NULL;
12325 
12326 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
12327 #ifdef INET6
12328 	if (rack->r_is_v6) {
12329 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12330 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12331 		if (tp->t_port) {
12332 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12333 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12334 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12335 			udp->uh_dport = tp->t_port;
12336 			rack->r_ctl.fsb.udp = udp;
12337 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12338 		} else
12339 		{
12340 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12341 			rack->r_ctl.fsb.udp = NULL;
12342 		}
12343 		tcpip_fillheaders(rack->rc_inp,
12344 				  tp->t_port,
12345 				  ip6, rack->r_ctl.fsb.th);
12346 	} else
12347 #endif				/* INET6 */
12348 	{
12349 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12350 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12351 		if (tp->t_port) {
12352 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12353 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12354 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12355 			udp->uh_dport = tp->t_port;
12356 			rack->r_ctl.fsb.udp = udp;
12357 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12358 		} else
12359 		{
12360 			rack->r_ctl.fsb.udp = NULL;
12361 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12362 		}
12363 		tcpip_fillheaders(rack->rc_inp,
12364 				  tp->t_port,
12365 				  ip, rack->r_ctl.fsb.th);
12366 	}
12367 	rack->r_fsb_inited = 1;
12368 }
12369 
12370 static int
12371 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12372 {
12373 	/*
12374 	 * Allocate the larger of spaces V6 if available else just
12375 	 * V4 and include udphdr (overbook)
12376 	 */
12377 #ifdef INET6
12378 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12379 #else
12380 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12381 #endif
12382 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12383 					    M_TCPFSB, M_NOWAIT|M_ZERO);
12384 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12385 		return (ENOMEM);
12386 	}
12387 	rack->r_fsb_inited = 0;
12388 	return (0);
12389 }
12390 
12391 static int
12392 rack_init(struct tcpcb *tp)
12393 {
12394 	struct tcp_rack *rack = NULL;
12395 #ifdef INVARIANTS
12396 	struct rack_sendmap *insret;
12397 #endif
12398 	uint32_t iwin, snt, us_cts;
12399 	int err;
12400 
12401 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12402 	if (tp->t_fb_ptr == NULL) {
12403 		/*
12404 		 * We need to allocate memory but cant. The INP and INP_INFO
12405 		 * locks and they are recursive (happens during setup. So a
12406 		 * scheme to drop the locks fails :(
12407 		 *
12408 		 */
12409 		return (ENOMEM);
12410 	}
12411 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12412 
12413 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12414 	RB_INIT(&rack->r_ctl.rc_mtree);
12415 	TAILQ_INIT(&rack->r_ctl.rc_free);
12416 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
12417 	rack->rc_tp = tp;
12418 	rack->rc_inp = tp->t_inpcb;
12419 	/* Set the flag */
12420 	rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12421 	/* Probably not needed but lets be sure */
12422 	rack_clear_rate_sample(rack);
12423 	/*
12424 	 * Save off the default values, socket options will poke
12425 	 * at these if pacing is not on or we have not yet
12426 	 * reached where pacing is on (gp_ready/fixed enabled).
12427 	 * When they get set into the CC module (when gp_ready
12428 	 * is enabled or we enable fixed) then we will set these
12429 	 * values into the CC and place in here the old values
12430 	 * so we have a restoral. Then we will set the flag
12431 	 * rc_pacing_cc_set. That way whenever we turn off pacing
12432 	 * or switch off this stack, we will know to go restore
12433 	 * the saved values.
12434 	 */
12435 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12436 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12437 	/* We want abe like behavior as well */
12438 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12439 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12440 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12441 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12442 	rack->r_ctl.roundends = tp->snd_max;
12443 	if (use_rack_rr)
12444 		rack->use_rack_rr = 1;
12445 	if (V_tcp_delack_enabled)
12446 		tp->t_delayed_ack = 1;
12447 	else
12448 		tp->t_delayed_ack = 0;
12449 #ifdef TCP_ACCOUNTING
12450 	if (rack_tcp_accounting) {
12451 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12452 	}
12453 #endif
12454 	if (rack_enable_shared_cwnd)
12455 		rack->rack_enable_scwnd = 1;
12456 	rack->rc_user_set_max_segs = rack_hptsi_segments;
12457 	rack->rc_force_max_seg = 0;
12458 	if (rack_use_imac_dack)
12459 		rack->rc_dack_mode = 1;
12460 	TAILQ_INIT(&rack->r_ctl.opt_list);
12461 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12462 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12463 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12464 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12465 	rack->r_ctl.rc_highest_us_rtt = 0;
12466 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12467 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12468 	if (rack_use_cmp_acks)
12469 		rack->r_use_cmp_ack = 1;
12470 	if (rack_disable_prr)
12471 		rack->rack_no_prr = 1;
12472 	if (rack_gp_no_rec_chg)
12473 		rack->rc_gp_no_rec_chg = 1;
12474 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12475 		rack->rc_always_pace = 1;
12476 		if (rack->use_fixed_rate || rack->gp_ready)
12477 			rack_set_cc_pacing(rack);
12478 	} else
12479 		rack->rc_always_pace = 0;
12480 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12481 		rack->r_mbuf_queue = 1;
12482 	else
12483 		rack->r_mbuf_queue = 0;
12484 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12485 		tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12486 	else
12487 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12488 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
12489 	if (rack_limits_scwnd)
12490 		rack->r_limit_scw = 1;
12491 	else
12492 		rack->r_limit_scw = 0;
12493 	rack->rc_labc = V_tcp_abc_l_var;
12494 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12495 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12496 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12497 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12498 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12499 	rack->r_ctl.rc_min_to = rack_min_to;
12500 	microuptime(&rack->r_ctl.act_rcv_time);
12501 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12502 	rack->rc_init_win = rack_default_init_window;
12503 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12504 	if (rack_hw_up_only)
12505 		rack->r_up_only = 1;
12506 	if (rack_do_dyn_mul) {
12507 		/* When dynamic adjustment is on CA needs to start at 100% */
12508 		rack->rc_gp_dyn_mul = 1;
12509 		if (rack_do_dyn_mul >= 100)
12510 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12511 	} else
12512 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12513 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12514 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12515 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12516 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12517 				rack_probertt_filter_life);
12518 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12519 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12520 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
12521 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12522 	rack->r_ctl.rc_time_probertt_starts = 0;
12523 	if (rack_dsack_std_based & 0x1) {
12524 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12525 		rack->rc_rack_tmr_std_based = 1;
12526 	}
12527 	if (rack_dsack_std_based & 0x2) {
12528 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12529 		rack->rc_rack_use_dsack = 1;
12530 	}
12531 	/* We require at least one measurement, even if the sysctl is 0 */
12532 	if (rack_req_measurements)
12533 		rack->r_ctl.req_measurements = rack_req_measurements;
12534 	else
12535 		rack->r_ctl.req_measurements = 1;
12536 	if (rack_enable_hw_pacing)
12537 		rack->rack_hdw_pace_ena = 1;
12538 	if (rack_hw_rate_caps)
12539 		rack->r_rack_hw_rate_caps = 1;
12540 	/* Do we force on detection? */
12541 #ifdef NETFLIX_EXP_DETECTION
12542 	if (tcp_force_detection)
12543 		rack->do_detection = 1;
12544 	else
12545 #endif
12546 		rack->do_detection = 0;
12547 	if (rack_non_rxt_use_cr)
12548 		rack->rack_rec_nonrxt_use_cr = 1;
12549 	err = rack_init_fsb(tp, rack);
12550 	if (err) {
12551 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12552 		tp->t_fb_ptr = NULL;
12553 		return (err);
12554 	}
12555 	if (tp->snd_una != tp->snd_max) {
12556 		/* Create a send map for the current outstanding data */
12557 		struct rack_sendmap *rsm;
12558 
12559 		rsm = rack_alloc(rack);
12560 		if (rsm == NULL) {
12561 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12562 			tp->t_fb_ptr = NULL;
12563 			return (ENOMEM);
12564 		}
12565 		rsm->r_no_rtt_allowed = 1;
12566 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12567 		rsm->r_rtr_cnt = 1;
12568 		rsm->r_rtr_bytes = 0;
12569 		if (tp->t_flags & TF_SENTFIN) {
12570 			rsm->r_end = tp->snd_max - 1;
12571 			rsm->r_flags |= RACK_HAS_FIN;
12572 		} else {
12573 			rsm->r_end = tp->snd_max;
12574 		}
12575 		if (tp->snd_una == tp->iss) {
12576 			/* The data space is one beyond snd_una */
12577 			rsm->r_flags |= RACK_HAS_SYN;
12578 			rsm->r_start = tp->iss;
12579 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12580 		} else
12581 			rsm->r_start = tp->snd_una;
12582 		rsm->r_dupack = 0;
12583 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12584 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12585 			if (rsm->m)
12586 				rsm->orig_m_len = rsm->m->m_len;
12587 			else
12588 				rsm->orig_m_len = 0;
12589 		} else {
12590 			/*
12591 			 * This can happen if we have a stand-alone FIN or
12592 			 *  SYN.
12593 			 */
12594 			rsm->m = NULL;
12595 			rsm->orig_m_len = 0;
12596 			rsm->soff = 0;
12597 		}
12598 #ifndef INVARIANTS
12599 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12600 #else
12601 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12602 		if (insret != NULL) {
12603 			panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12604 			      insret, rack, rsm);
12605 		}
12606 #endif
12607 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12608 		rsm->r_in_tmap = 1;
12609 	}
12610 	/*
12611 	 * Timers in Rack are kept in microseconds so lets
12612 	 * convert any initial incoming variables
12613 	 * from ticks into usecs. Note that we
12614 	 * also change the values of t_srtt and t_rttvar, if
12615 	 * they are non-zero. They are kept with a 5
12616 	 * bit decimal so we have to carefully convert
12617 	 * these to get the full precision.
12618 	 */
12619 	rack_convert_rtts(tp);
12620 	tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12621 	if (rack_do_hystart) {
12622 		tp->ccv->flags |= CCF_HYSTART_ALLOWED;
12623 		if (rack_do_hystart > 1)
12624 			tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND;
12625 		if (rack_do_hystart > 2)
12626 			tp->ccv->flags |= CCF_HYSTART_CONS_SSTH;
12627 	}
12628 	if (rack_def_profile)
12629 		rack_set_profile(rack, rack_def_profile);
12630 	/* Cancel the GP measurement in progress */
12631 	tp->t_flags &= ~TF_GPUTINPROG;
12632 	if (SEQ_GT(tp->snd_max, tp->iss))
12633 		snt = tp->snd_max - tp->iss;
12634 	else
12635 		snt = 0;
12636 	iwin = rc_init_window(rack);
12637 	if (snt < iwin) {
12638 		/* We are not past the initial window
12639 		 * so we need to make sure cwnd is
12640 		 * correct.
12641 		 */
12642 		if (tp->snd_cwnd < iwin)
12643 			tp->snd_cwnd = iwin;
12644 		/*
12645 		 * If we are within the initial window
12646 		 * we want ssthresh to be unlimited. Setting
12647 		 * it to the rwnd (which the default stack does
12648 		 * and older racks) is not really a good idea
12649 		 * since we want to be in SS and grow both the
12650 		 * cwnd and the rwnd (via dynamic rwnd growth). If
12651 		 * we set it to the rwnd then as the peer grows its
12652 		 * rwnd we will be stuck in CA and never hit SS.
12653 		 *
12654 		 * Its far better to raise it up high (this takes the
12655 		 * risk that there as been a loss already, probably
12656 		 * we should have an indicator in all stacks of loss
12657 		 * but we don't), but considering the normal use this
12658 		 * is a risk worth taking. The consequences of not
12659 		 * hitting SS are far worse than going one more time
12660 		 * into it early on (before we have sent even a IW).
12661 		 * It is highly unlikely that we will have had a loss
12662 		 * before getting the IW out.
12663 		 */
12664 		tp->snd_ssthresh = 0xffffffff;
12665 	}
12666 	rack_stop_all_timers(tp);
12667 	/* Lets setup the fsb block */
12668 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12669 	rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12670 			     __LINE__, RACK_RTTS_INIT);
12671 	return (0);
12672 }
12673 
12674 static int
12675 rack_handoff_ok(struct tcpcb *tp)
12676 {
12677 	if ((tp->t_state == TCPS_CLOSED) ||
12678 	    (tp->t_state == TCPS_LISTEN)) {
12679 		/* Sure no problem though it may not stick */
12680 		return (0);
12681 	}
12682 	if ((tp->t_state == TCPS_SYN_SENT) ||
12683 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
12684 		/*
12685 		 * We really don't know if you support sack,
12686 		 * you have to get to ESTAB or beyond to tell.
12687 		 */
12688 		return (EAGAIN);
12689 	}
12690 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12691 		/*
12692 		 * Rack will only send a FIN after all data is acknowledged.
12693 		 * So in this case we have more data outstanding. We can't
12694 		 * switch stacks until either all data and only the FIN
12695 		 * is left (in which case rack_init() now knows how
12696 		 * to deal with that) <or> all is acknowledged and we
12697 		 * are only left with incoming data, though why you
12698 		 * would want to switch to rack after all data is acknowledged
12699 		 * I have no idea (rrs)!
12700 		 */
12701 		return (EAGAIN);
12702 	}
12703 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12704 		return (0);
12705 	}
12706 	/*
12707 	 * If we reach here we don't do SACK on this connection so we can
12708 	 * never do rack.
12709 	 */
12710 	return (EINVAL);
12711 }
12712 
12713 
12714 static void
12715 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12716 {
12717 	if (tp->t_fb_ptr) {
12718 		struct tcp_rack *rack;
12719 		struct rack_sendmap *rsm, *nrsm;
12720 #ifdef INVARIANTS
12721 		struct rack_sendmap *rm;
12722 #endif
12723 
12724 		rack = (struct tcp_rack *)tp->t_fb_ptr;
12725 		if (tp->t_in_pkt) {
12726 			/*
12727 			 * It is unsafe to process the packets since a
12728 			 * reset may be lurking in them (its rare but it
12729 			 * can occur). If we were to find a RST, then we
12730 			 * would end up dropping the connection and the
12731 			 * INP lock, so when we return the caller (tcp_usrreq)
12732 			 * will blow up when it trys to unlock the inp.
12733 			 */
12734 			struct mbuf *save, *m;
12735 
12736 			m = tp->t_in_pkt;
12737 			tp->t_in_pkt = NULL;
12738 			tp->t_tail_pkt = NULL;
12739 			while (m) {
12740 				save = m->m_nextpkt;
12741 				m->m_nextpkt = NULL;
12742 				m_freem(m);
12743 				m = save;
12744 			}
12745 		}
12746 		tp->t_flags &= ~TF_FORCEDATA;
12747 #ifdef NETFLIX_SHARED_CWND
12748 		if (rack->r_ctl.rc_scw) {
12749 			uint32_t limit;
12750 
12751 			if (rack->r_limit_scw)
12752 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12753 			else
12754 				limit = 0;
12755 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12756 						  rack->r_ctl.rc_scw_index,
12757 						  limit);
12758 			rack->r_ctl.rc_scw = NULL;
12759 		}
12760 #endif
12761 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
12762 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12763 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12764 			rack->r_ctl.fsb.th = NULL;
12765 		}
12766 		/* Convert back to ticks, with  */
12767 		if (tp->t_srtt > 1) {
12768 			uint32_t val, frac;
12769 
12770 			val = USEC_2_TICKS(tp->t_srtt);
12771 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12772 			tp->t_srtt = val << TCP_RTT_SHIFT;
12773 			/*
12774 			 * frac is the fractional part here is left
12775 			 * over from converting to hz and shifting.
12776 			 * We need to convert this to the 5 bit
12777 			 * remainder.
12778 			 */
12779 			if (frac) {
12780 				if (hz == 1000) {
12781 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12782 				} else {
12783 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12784 				}
12785 				tp->t_srtt += frac;
12786 			}
12787 		}
12788 		if (tp->t_rttvar) {
12789 			uint32_t val, frac;
12790 
12791 			val = USEC_2_TICKS(tp->t_rttvar);
12792 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12793 			tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12794 			/*
12795 			 * frac is the fractional part here is left
12796 			 * over from converting to hz and shifting.
12797 			 * We need to convert this to the 5 bit
12798 			 * remainder.
12799 			 */
12800 			if (frac) {
12801 				if (hz == 1000) {
12802 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12803 				} else {
12804 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12805 				}
12806 				tp->t_rttvar += frac;
12807 			}
12808 		}
12809 		tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12810 		tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12811 		if (rack->rc_always_pace) {
12812 			tcp_decrement_paced_conn();
12813 			rack_undo_cc_pacing(rack);
12814 			rack->rc_always_pace = 0;
12815 		}
12816 		/* Clean up any options if they were not applied */
12817 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12818 			struct deferred_opt_list *dol;
12819 
12820 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12821 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12822 			free(dol, M_TCPDO);
12823 		}
12824 		/* rack does not use force data but other stacks may clear it */
12825 		if (rack->r_ctl.crte != NULL) {
12826 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12827 			rack->rack_hdrw_pacing = 0;
12828 			rack->r_ctl.crte = NULL;
12829 		}
12830 #ifdef TCP_BLACKBOX
12831 		tcp_log_flowend(tp);
12832 #endif
12833 		RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12834 #ifndef INVARIANTS
12835 			(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12836 #else
12837 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12838 			if (rm != rsm) {
12839 				panic("At fini, rack:%p rsm:%p rm:%p",
12840 				      rack, rsm, rm);
12841 			}
12842 #endif
12843 			uma_zfree(rack_zone, rsm);
12844 		}
12845 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12846 		while (rsm) {
12847 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12848 			uma_zfree(rack_zone, rsm);
12849 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12850 		}
12851 		rack->rc_free_cnt = 0;
12852 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12853 		tp->t_fb_ptr = NULL;
12854 	}
12855 	if (tp->t_inpcb) {
12856 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12857 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12858 		tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12859 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
12860 		/* Cancel the GP measurement in progress */
12861 		tp->t_flags &= ~TF_GPUTINPROG;
12862 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
12863 	}
12864 	/* Make sure snd_nxt is correctly set */
12865 	tp->snd_nxt = tp->snd_max;
12866 }
12867 
12868 static void
12869 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12870 {
12871 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12872 		rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12873 	}
12874 	switch (tp->t_state) {
12875 	case TCPS_SYN_SENT:
12876 		rack->r_state = TCPS_SYN_SENT;
12877 		rack->r_substate = rack_do_syn_sent;
12878 		break;
12879 	case TCPS_SYN_RECEIVED:
12880 		rack->r_state = TCPS_SYN_RECEIVED;
12881 		rack->r_substate = rack_do_syn_recv;
12882 		break;
12883 	case TCPS_ESTABLISHED:
12884 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12885 		rack->r_state = TCPS_ESTABLISHED;
12886 		rack->r_substate = rack_do_established;
12887 		break;
12888 	case TCPS_CLOSE_WAIT:
12889 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12890 		rack->r_state = TCPS_CLOSE_WAIT;
12891 		rack->r_substate = rack_do_close_wait;
12892 		break;
12893 	case TCPS_FIN_WAIT_1:
12894 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12895 		rack->r_state = TCPS_FIN_WAIT_1;
12896 		rack->r_substate = rack_do_fin_wait_1;
12897 		break;
12898 	case TCPS_CLOSING:
12899 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12900 		rack->r_state = TCPS_CLOSING;
12901 		rack->r_substate = rack_do_closing;
12902 		break;
12903 	case TCPS_LAST_ACK:
12904 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12905 		rack->r_state = TCPS_LAST_ACK;
12906 		rack->r_substate = rack_do_lastack;
12907 		break;
12908 	case TCPS_FIN_WAIT_2:
12909 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12910 		rack->r_state = TCPS_FIN_WAIT_2;
12911 		rack->r_substate = rack_do_fin_wait_2;
12912 		break;
12913 	case TCPS_LISTEN:
12914 	case TCPS_CLOSED:
12915 	case TCPS_TIME_WAIT:
12916 	default:
12917 		break;
12918 	};
12919 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12920 		rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12921 
12922 }
12923 
12924 static void
12925 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12926 {
12927 	/*
12928 	 * We received an ack, and then did not
12929 	 * call send or were bounced out due to the
12930 	 * hpts was running. Now a timer is up as well, is
12931 	 * it the right timer?
12932 	 */
12933 	struct rack_sendmap *rsm;
12934 	int tmr_up;
12935 
12936 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12937 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12938 		return;
12939 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12940 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12941 	    (tmr_up == PACE_TMR_RXT)) {
12942 		/* Should be an RXT */
12943 		return;
12944 	}
12945 	if (rsm == NULL) {
12946 		/* Nothing outstanding? */
12947 		if (tp->t_flags & TF_DELACK) {
12948 			if (tmr_up == PACE_TMR_DELACK)
12949 				/* We are supposed to have delayed ack up and we do */
12950 				return;
12951 		} else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12952 			/*
12953 			 * if we hit enobufs then we would expect the possiblity
12954 			 * of nothing outstanding and the RXT up (and the hptsi timer).
12955 			 */
12956 			return;
12957 		} else if (((V_tcp_always_keepalive ||
12958 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12959 			    (tp->t_state <= TCPS_CLOSING)) &&
12960 			   (tmr_up == PACE_TMR_KEEP) &&
12961 			   (tp->snd_max == tp->snd_una)) {
12962 			/* We should have keep alive up and we do */
12963 			return;
12964 		}
12965 	}
12966 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12967 		   ((tmr_up == PACE_TMR_TLP) ||
12968 		    (tmr_up == PACE_TMR_RACK) ||
12969 		    (tmr_up == PACE_TMR_RXT))) {
12970 		/*
12971 		 * Either a Rack, TLP or RXT is fine if  we
12972 		 * have outstanding data.
12973 		 */
12974 		return;
12975 	} else if (tmr_up == PACE_TMR_DELACK) {
12976 		/*
12977 		 * If the delayed ack was going to go off
12978 		 * before the rtx/tlp/rack timer were going to
12979 		 * expire, then that would be the timer in control.
12980 		 * Note we don't check the time here trusting the
12981 		 * code is correct.
12982 		 */
12983 		return;
12984 	}
12985 	/*
12986 	 * Ok the timer originally started is not what we want now.
12987 	 * We will force the hpts to be stopped if any, and restart
12988 	 * with the slot set to what was in the saved slot.
12989 	 */
12990 	if (tcp_in_hpts(rack->rc_inp)) {
12991 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12992 			uint32_t us_cts;
12993 
12994 			us_cts = tcp_get_usecs(NULL);
12995 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12996 				rack->r_early = 1;
12997 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12998 			}
12999 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13000 		}
13001 		tcp_hpts_remove(tp->t_inpcb);
13002 	}
13003 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13004 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13005 }
13006 
13007 
13008 static void
13009 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)
13010 {
13011 	if ((SEQ_LT(tp->snd_wl1, seq) ||
13012 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
13013 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
13014 		/* keep track of pure window updates */
13015 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
13016 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13017 		tp->snd_wnd = tiwin;
13018 		rack_validate_fo_sendwin_up(tp, rack);
13019 		tp->snd_wl1 = seq;
13020 		tp->snd_wl2 = ack;
13021 		if (tp->snd_wnd > tp->max_sndwnd)
13022 			tp->max_sndwnd = tp->snd_wnd;
13023 	    rack->r_wanted_output = 1;
13024 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
13025 		tp->snd_wnd = tiwin;
13026 		rack_validate_fo_sendwin_up(tp, rack);
13027 		tp->snd_wl1 = seq;
13028 		tp->snd_wl2 = ack;
13029 	} else {
13030 		/* Not a valid win update */
13031 		return;
13032 	}
13033 	if (tp->snd_wnd > tp->max_sndwnd)
13034 		tp->max_sndwnd = tp->snd_wnd;
13035 	if (tp->snd_wnd < (tp->snd_max - high_seq)) {
13036 		/* The peer collapsed the window */
13037 		rack_collapsed_window(rack);
13038 	} else if (rack->rc_has_collapsed)
13039 		rack_un_collapse_window(rack);
13040 	/* Do we exit persists? */
13041 	if ((rack->rc_in_persist != 0) &&
13042 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13043 				rack->r_ctl.rc_pace_min_segs))) {
13044 		rack_exit_persist(tp, rack, cts);
13045 	}
13046 	/* Do we enter persists? */
13047 	if ((rack->rc_in_persist == 0) &&
13048 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13049 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13050 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13051 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
13052 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
13053 		/*
13054 		 * Here the rwnd is less than
13055 		 * the pacing size, we are established,
13056 		 * nothing is outstanding, and there is
13057 		 * data to send. Enter persists.
13058 		 */
13059 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13060 	}
13061 }
13062 
13063 static void
13064 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13065 {
13066 
13067 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13068 		union tcp_log_stackspecific log;
13069 		struct timeval ltv;
13070 		char tcp_hdr_buf[60];
13071 		struct tcphdr *th;
13072 		struct timespec ts;
13073 		uint32_t orig_snd_una;
13074 		uint8_t xx = 0;
13075 
13076 #ifdef NETFLIX_HTTP_LOGGING
13077 		struct http_sendfile_track *http_req;
13078 
13079 		if (SEQ_GT(ae->ack, tp->snd_una)) {
13080 			http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13081 		} else {
13082 			http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13083 		}
13084 #endif
13085 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13086 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13087 		if (rack->rack_no_prr == 0)
13088 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13089 		else
13090 			log.u_bbr.flex1 = 0;
13091 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13092 		log.u_bbr.use_lt_bw <<= 1;
13093 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
13094 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13095 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13096 		log.u_bbr.pkts_out = tp->t_maxseg;
13097 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13098 		log.u_bbr.flex7 = 1;
13099 		log.u_bbr.lost = ae->flags;
13100 		log.u_bbr.cwnd_gain = ackval;
13101 		log.u_bbr.pacing_gain = 0x2;
13102 		if (ae->flags & TSTMP_HDWR) {
13103 			/* Record the hardware timestamp if present */
13104 			log.u_bbr.flex3 = M_TSTMP;
13105 			ts.tv_sec = ae->timestamp / 1000000000;
13106 			ts.tv_nsec = ae->timestamp % 1000000000;
13107 			ltv.tv_sec = ts.tv_sec;
13108 			ltv.tv_usec = ts.tv_nsec / 1000;
13109 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13110 		} else if (ae->flags & TSTMP_LRO) {
13111 			/* Record the LRO the arrival timestamp */
13112 			log.u_bbr.flex3 = M_TSTMP_LRO;
13113 			ts.tv_sec = ae->timestamp / 1000000000;
13114 			ts.tv_nsec = ae->timestamp % 1000000000;
13115 			ltv.tv_sec = ts.tv_sec;
13116 			ltv.tv_usec = ts.tv_nsec / 1000;
13117 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13118 		}
13119 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13120 		/* Log the rcv time */
13121 		log.u_bbr.delRate = ae->timestamp;
13122 #ifdef NETFLIX_HTTP_LOGGING
13123 		log.u_bbr.applimited = tp->t_http_closed;
13124 		log.u_bbr.applimited <<= 8;
13125 		log.u_bbr.applimited |= tp->t_http_open;
13126 		log.u_bbr.applimited <<= 8;
13127 		log.u_bbr.applimited |= tp->t_http_req;
13128 		if (http_req) {
13129 			/* Copy out any client req info */
13130 			/* seconds */
13131 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13132 			/* useconds */
13133 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13134 			log.u_bbr.rttProp = http_req->timestamp;
13135 			log.u_bbr.cur_del_rate = http_req->start;
13136 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13137 				log.u_bbr.flex8 |= 1;
13138 			} else {
13139 				log.u_bbr.flex8 |= 2;
13140 				log.u_bbr.bw_inuse = http_req->end;
13141 			}
13142 			log.u_bbr.flex6 = http_req->start_seq;
13143 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13144 				log.u_bbr.flex8 |= 4;
13145 				log.u_bbr.epoch = http_req->end_seq;
13146 			}
13147 		}
13148 #endif
13149 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13150 		th = (struct tcphdr *)tcp_hdr_buf;
13151 		th->th_seq = ae->seq;
13152 		th->th_ack = ae->ack;
13153 		th->th_win = ae->win;
13154 		/* Now fill in the ports */
13155 		th->th_sport = tp->t_inpcb->inp_fport;
13156 		th->th_dport = tp->t_inpcb->inp_lport;
13157 		tcp_set_flags(th, ae->flags);
13158 		/* Now do we have a timestamp option? */
13159 		if (ae->flags & HAS_TSTMP) {
13160 			u_char *cp;
13161 			uint32_t val;
13162 
13163 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13164 			cp = (u_char *)(th + 1);
13165 			*cp = TCPOPT_NOP;
13166 			cp++;
13167 			*cp = TCPOPT_NOP;
13168 			cp++;
13169 			*cp = TCPOPT_TIMESTAMP;
13170 			cp++;
13171 			*cp = TCPOLEN_TIMESTAMP;
13172 			cp++;
13173 			val = htonl(ae->ts_value);
13174 			bcopy((char *)&val,
13175 			      (char *)cp, sizeof(uint32_t));
13176 			val = htonl(ae->ts_echo);
13177 			bcopy((char *)&val,
13178 			      (char *)(cp + 4), sizeof(uint32_t));
13179 		} else
13180 			th->th_off = (sizeof(struct tcphdr) >> 2);
13181 
13182 		/*
13183 		 * For sane logging we need to play a little trick.
13184 		 * If the ack were fully processed we would have moved
13185 		 * snd_una to high_seq, but since compressed acks are
13186 		 * processed in two phases, at this point (logging) snd_una
13187 		 * won't be advanced. So we would see multiple acks showing
13188 		 * the advancement. We can prevent that by "pretending" that
13189 		 * snd_una was advanced and then un-advancing it so that the
13190 		 * logging code has the right value for tlb_snd_una.
13191 		 */
13192 		if (tp->snd_una != high_seq) {
13193 			orig_snd_una = tp->snd_una;
13194 			tp->snd_una = high_seq;
13195 			xx = 1;
13196 		} else
13197 			xx = 0;
13198 		TCP_LOG_EVENTP(tp, th,
13199 			       &tp->t_inpcb->inp_socket->so_rcv,
13200 			       &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
13201 			       0, &log, true, &ltv);
13202 		if (xx) {
13203 			tp->snd_una = orig_snd_una;
13204 		}
13205 	}
13206 
13207 }
13208 
13209 static void
13210 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13211 {
13212 	uint32_t us_rtt;
13213 	/*
13214 	 * A persist or keep-alive was forced out, update our
13215 	 * min rtt time. Note now worry about lost responses.
13216 	 * When a subsequent keep-alive or persist times out
13217 	 * and forced_ack is still on, then the last probe
13218 	 * was not responded to. In such cases we have a
13219 	 * sysctl that controls the behavior. Either we apply
13220 	 * the rtt but with reduced confidence (0). Or we just
13221 	 * plain don't apply the rtt estimate. Having data flow
13222 	 * will clear the probe_not_answered flag i.e. cum-ack
13223 	 * move forward <or> exiting and reentering persists.
13224 	 */
13225 
13226 	rack->forced_ack = 0;
13227 	rack->rc_tp->t_rxtshift = 0;
13228 	if ((rack->rc_in_persist &&
13229 	     (tiwin == rack->rc_tp->snd_wnd)) ||
13230 	    (rack->rc_in_persist == 0)) {
13231 		/*
13232 		 * In persists only apply the RTT update if this is
13233 		 * a response to our window probe. And that
13234 		 * means the rwnd sent must match the current
13235 		 * snd_wnd. If it does not, then we got a
13236 		 * window update ack instead. For keepalive
13237 		 * we allow the answer no matter what the window.
13238 		 *
13239 		 * Note that if the probe_not_answered is set then
13240 		 * the forced_ack_ts is the oldest one i.e. the first
13241 		 * probe sent that might have been lost. This assures
13242 		 * us that if we do calculate an RTT it is longer not
13243 		 * some short thing.
13244 		 */
13245 		if (rack->rc_in_persist)
13246 			counter_u64_add(rack_persists_acks, 1);
13247 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13248 		if (us_rtt == 0)
13249 			us_rtt = 1;
13250 		if (rack->probe_not_answered == 0) {
13251 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13252 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13253 		} else {
13254 			/* We have a retransmitted probe here too */
13255 			if (rack_apply_rtt_with_reduced_conf) {
13256 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13257 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13258 			}
13259 		}
13260 	}
13261 }
13262 
13263 static int
13264 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13265 {
13266 	/*
13267 	 * Handle a "special" compressed ack mbuf. Each incoming
13268 	 * ack has only four possible dispositions:
13269 	 *
13270 	 * A) It moves the cum-ack forward
13271 	 * B) It is behind the cum-ack.
13272 	 * C) It is a window-update ack.
13273 	 * D) It is a dup-ack.
13274 	 *
13275 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13276 	 * in the incoming mbuf. We also need to still pay attention
13277 	 * to nxt_pkt since there may be another packet after this
13278 	 * one.
13279 	 */
13280 #ifdef TCP_ACCOUNTING
13281 	uint64_t ts_val;
13282 	uint64_t rdstc;
13283 #endif
13284 	int segsiz;
13285 	struct timespec ts;
13286 	struct tcp_rack *rack;
13287 	struct tcp_ackent *ae;
13288 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13289 	int cnt, i, did_out, ourfinisacked = 0;
13290 	struct tcpopt to_holder, *to = NULL;
13291 #ifdef TCP_ACCOUNTING
13292 	int win_up_req = 0;
13293 #endif
13294 	int nsegs = 0;
13295 	int under_pacing = 1;
13296 	int recovery = 0;
13297 #ifdef TCP_ACCOUNTING
13298 	sched_pin();
13299 #endif
13300 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13301 	if (rack->gp_ready &&
13302 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13303 		under_pacing = 0;
13304 	else
13305 		under_pacing = 1;
13306 
13307 	if (rack->r_state != tp->t_state)
13308 		rack_set_state(tp, rack);
13309 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13310 	    (tp->t_flags & TF_GPUTINPROG)) {
13311 		/*
13312 		 * We have a goodput in progress
13313 		 * and we have entered a late state.
13314 		 * Do we have enough data in the sb
13315 		 * to handle the GPUT request?
13316 		 */
13317 		uint32_t bytes;
13318 
13319 		bytes = tp->gput_ack - tp->gput_seq;
13320 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
13321 			bytes += tp->gput_seq - tp->snd_una;
13322 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
13323 			/*
13324 			 * There are not enough bytes in the socket
13325 			 * buffer that have been sent to cover this
13326 			 * measurement. Cancel it.
13327 			 */
13328 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13329 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
13330 						   tp->gput_seq,
13331 						   0, 0, 18, __LINE__, NULL, 0);
13332 			tp->t_flags &= ~TF_GPUTINPROG;
13333 		}
13334 	}
13335 	to = &to_holder;
13336 	to->to_flags = 0;
13337 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13338 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13339 	cnt = m->m_len / sizeof(struct tcp_ackent);
13340 	counter_u64_add(rack_multi_single_eq, cnt);
13341 	high_seq = tp->snd_una;
13342 	the_win = tp->snd_wnd;
13343 	win_seq = tp->snd_wl1;
13344 	win_upd_ack = tp->snd_wl2;
13345 	cts = tcp_tv_to_usectick(tv);
13346 	ms_cts = tcp_tv_to_mssectick(tv);
13347 	rack->r_ctl.rc_rcvtime = cts;
13348 	segsiz = ctf_fixed_maxseg(tp);
13349 	if ((rack->rc_gp_dyn_mul) &&
13350 	    (rack->use_fixed_rate == 0) &&
13351 	    (rack->rc_always_pace)) {
13352 		/* Check in on probertt */
13353 		rack_check_probe_rtt(rack, cts);
13354 	}
13355 	for (i = 0; i < cnt; i++) {
13356 #ifdef TCP_ACCOUNTING
13357 		ts_val = get_cyclecount();
13358 #endif
13359 		rack_clear_rate_sample(rack);
13360 		ae = ((mtod(m, struct tcp_ackent *)) + i);
13361 		/* Setup the window */
13362 		tiwin = ae->win << tp->snd_scale;
13363 		if (tiwin > rack->r_ctl.rc_high_rwnd)
13364 			rack->r_ctl.rc_high_rwnd = tiwin;
13365 		/* figure out the type of ack */
13366 		if (SEQ_LT(ae->ack, high_seq)) {
13367 			/* Case B*/
13368 			ae->ack_val_set = ACK_BEHIND;
13369 		} else if (SEQ_GT(ae->ack, high_seq)) {
13370 			/* Case A */
13371 			ae->ack_val_set = ACK_CUMACK;
13372 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13373 			/* Case D */
13374 			ae->ack_val_set = ACK_DUPACK;
13375 		} else {
13376 			/* Case C */
13377 			ae->ack_val_set = ACK_RWND;
13378 		}
13379 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13380 		/* Validate timestamp */
13381 		if (ae->flags & HAS_TSTMP) {
13382 			/* Setup for a timestamp */
13383 			to->to_flags = TOF_TS;
13384 			ae->ts_echo -= tp->ts_offset;
13385 			to->to_tsecr = ae->ts_echo;
13386 			to->to_tsval = ae->ts_value;
13387 			/*
13388 			 * If echoed timestamp is later than the current time, fall back to
13389 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13390 			 * were used when this connection was established.
13391 			 */
13392 			if (TSTMP_GT(ae->ts_echo, ms_cts))
13393 				to->to_tsecr = 0;
13394 			if (tp->ts_recent &&
13395 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13396 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13397 #ifdef TCP_ACCOUNTING
13398 					rdstc = get_cyclecount();
13399 					if (rdstc > ts_val) {
13400 						counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13401 								(rdstc - ts_val));
13402 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13403 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13404 						}
13405 					}
13406 #endif
13407 					continue;
13408 				}
13409 			}
13410 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13411 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13412 				tp->ts_recent_age = tcp_ts_getticks();
13413 				tp->ts_recent = ae->ts_value;
13414 			}
13415 		} else {
13416 			/* Setup for a no options */
13417 			to->to_flags = 0;
13418 		}
13419 		/* Update the rcv time and perform idle reduction possibly */
13420 		if  (tp->t_idle_reduce &&
13421 		     (tp->snd_max == tp->snd_una) &&
13422 		     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13423 			counter_u64_add(rack_input_idle_reduces, 1);
13424 			rack_cc_after_idle(rack, tp);
13425 		}
13426 		tp->t_rcvtime = ticks;
13427 		/* Now what about ECN? */
13428 		if (tcp_ecn_input_segment(tp, ae->flags, ae->codepoint))
13429 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
13430 #ifdef TCP_ACCOUNTING
13431 		/* Count for the specific type of ack in */
13432 		counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13433 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13434 			tp->tcp_cnt_counters[ae->ack_val_set]++;
13435 		}
13436 #endif
13437 		/*
13438 		 * Note how we could move up these in the determination
13439 		 * above, but we don't so that way the timestamp checks (and ECN)
13440 		 * is done first before we do any processing on the ACK.
13441 		 * The non-compressed path through the code has this
13442 		 * weakness (noted by @jtl) that it actually does some
13443 		 * processing before verifying the timestamp information.
13444 		 * We don't take that path here which is why we set
13445 		 * the ack_val_set first, do the timestamp and ecn
13446 		 * processing, and then look at what we have setup.
13447 		 */
13448 		if (ae->ack_val_set == ACK_BEHIND) {
13449 			/*
13450 			 * Case B flag reordering, if window is not closed
13451 			 * or it could be a keep-alive or persists
13452 			 */
13453 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13454 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13455 			}
13456 		} else if (ae->ack_val_set == ACK_DUPACK) {
13457 			/* Case D */
13458 			rack_strike_dupack(rack);
13459 		} else if (ae->ack_val_set == ACK_RWND) {
13460 			/* Case C */
13461 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13462 				ts.tv_sec = ae->timestamp / 1000000000;
13463 				ts.tv_nsec = ae->timestamp % 1000000000;
13464 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13465 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13466 			} else {
13467 				rack->r_ctl.act_rcv_time = *tv;
13468 			}
13469 			if (rack->forced_ack) {
13470 				rack_handle_probe_response(rack, tiwin,
13471 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13472 			}
13473 #ifdef TCP_ACCOUNTING
13474 			win_up_req = 1;
13475 #endif
13476 			win_upd_ack = ae->ack;
13477 			win_seq = ae->seq;
13478 			the_win = tiwin;
13479 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13480 		} else {
13481 			/* Case A */
13482 			if (SEQ_GT(ae->ack, tp->snd_max)) {
13483 				/*
13484 				 * We just send an ack since the incoming
13485 				 * ack is beyond the largest seq we sent.
13486 				 */
13487 				if ((tp->t_flags & TF_ACKNOW) == 0) {
13488 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13489 					if (tp->t_flags && TF_ACKNOW)
13490 						rack->r_wanted_output = 1;
13491 				}
13492 			} else {
13493 				nsegs++;
13494 				/* If the window changed setup to update */
13495 				if (tiwin != tp->snd_wnd) {
13496 					win_upd_ack = ae->ack;
13497 					win_seq = ae->seq;
13498 					the_win = tiwin;
13499 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13500 				}
13501 #ifdef TCP_ACCOUNTING
13502 				/* Account for the acks */
13503 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13504 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13505 				}
13506 				counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13507 						(((ae->ack - high_seq) + segsiz - 1) / segsiz));
13508 #endif
13509 				high_seq = ae->ack;
13510 				if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
13511 					union tcp_log_stackspecific log;
13512 					struct timeval tv;
13513 
13514 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13515 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13516 					log.u_bbr.flex1 = high_seq;
13517 					log.u_bbr.flex2 = rack->r_ctl.roundends;
13518 					log.u_bbr.flex3 = rack->r_ctl.current_round;
13519 					log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
13520 					log.u_bbr.flex8 = 8;
13521 					tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
13522 						       0, &log, false, NULL, NULL, 0, &tv);
13523 				}
13524 				/*
13525 				 * The draft (v3) calls for us to use SEQ_GEQ, but that
13526 				 * causes issues when we are just going app limited. Lets
13527 				 * instead use SEQ_GT <or> where its equal but more data
13528 				 * is outstanding.
13529 				 */
13530 				if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) ||
13531 				    ((high_seq == rack->r_ctl.roundends) &&
13532 				     SEQ_GT(tp->snd_max, tp->snd_una))) {
13533 					rack->r_ctl.current_round++;
13534 					rack->r_ctl.roundends = tp->snd_max;
13535 					if (CC_ALGO(tp)->newround != NULL) {
13536 						CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
13537 					}
13538 				}
13539 				/* Setup our act_rcv_time */
13540 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13541 					ts.tv_sec = ae->timestamp / 1000000000;
13542 					ts.tv_nsec = ae->timestamp % 1000000000;
13543 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13544 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13545 				} else {
13546 					rack->r_ctl.act_rcv_time = *tv;
13547 				}
13548 				rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13549 				if (rack->rc_dsack_round_seen) {
13550 					/* Is the dsack round over? */
13551 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13552 						/* Yes it is */
13553 						rack->rc_dsack_round_seen = 0;
13554 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13555 					}
13556 				}
13557 			}
13558 		}
13559 		/* And lets be sure to commit the rtt measurements for this ack */
13560 		tcp_rack_xmit_timer_commit(rack, tp);
13561 #ifdef TCP_ACCOUNTING
13562 		rdstc = get_cyclecount();
13563 		if (rdstc > ts_val) {
13564 			counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13565 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13566 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13567 				if (ae->ack_val_set == ACK_CUMACK)
13568 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13569 			}
13570 		}
13571 #endif
13572 	}
13573 #ifdef TCP_ACCOUNTING
13574 	ts_val = get_cyclecount();
13575 #endif
13576 	acked_amount = acked = (high_seq - tp->snd_una);
13577 	if (acked) {
13578 		/*
13579 		 * Clear the probe not answered flag
13580 		 * since cum-ack moved forward.
13581 		 */
13582 		rack->probe_not_answered = 0;
13583 		if (rack->sack_attack_disable == 0)
13584 			rack_do_decay(rack);
13585 		if (acked >= segsiz) {
13586 			/*
13587 			 * You only get credit for
13588 			 * MSS and greater (and you get extra
13589 			 * credit for larger cum-ack moves).
13590 			 */
13591 			int ac;
13592 
13593 			ac = acked / segsiz;
13594 			rack->r_ctl.ack_count += ac;
13595 			counter_u64_add(rack_ack_total, ac);
13596 		}
13597 		if (rack->r_ctl.ack_count > 0xfff00000) {
13598 			/*
13599 			 * reduce the number to keep us under
13600 			 * a uint32_t.
13601 			 */
13602 			rack->r_ctl.ack_count /= 2;
13603 			rack->r_ctl.sack_count /= 2;
13604 		}
13605 		if (tp->t_flags & TF_NEEDSYN) {
13606 			/*
13607 			 * T/TCP: Connection was half-synchronized, and our SYN has
13608 			 * been ACK'd (so connection is now fully synchronized).  Go
13609 			 * to non-starred state, increment snd_una for ACK of SYN,
13610 			 * and check if we can do window scaling.
13611 			 */
13612 			tp->t_flags &= ~TF_NEEDSYN;
13613 			tp->snd_una++;
13614 			acked_amount = acked = (high_seq - tp->snd_una);
13615 		}
13616 		if (acked > sbavail(&so->so_snd))
13617 			acked_amount = sbavail(&so->so_snd);
13618 #ifdef NETFLIX_EXP_DETECTION
13619 		/*
13620 		 * We only care on a cum-ack move if we are in a sack-disabled
13621 		 * state. We have already added in to the ack_count, and we never
13622 		 * would disable on a cum-ack move, so we only care to do the
13623 		 * detection if it may "undo" it, i.e. we were in disabled already.
13624 		 */
13625 		if (rack->sack_attack_disable)
13626 			rack_do_detection(tp, rack, acked_amount, segsiz);
13627 #endif
13628 		if (IN_FASTRECOVERY(tp->t_flags) &&
13629 		    (rack->rack_no_prr == 0))
13630 			rack_update_prr(tp, rack, acked_amount, high_seq);
13631 		if (IN_RECOVERY(tp->t_flags)) {
13632 			if (SEQ_LT(high_seq, tp->snd_recover) &&
13633 			    (SEQ_LT(high_seq, tp->snd_max))) {
13634 				tcp_rack_partialack(tp);
13635 			} else {
13636 				rack_post_recovery(tp, high_seq);
13637 				recovery = 1;
13638 			}
13639 		}
13640 		/* Handle the rack-log-ack part (sendmap) */
13641 		if ((sbused(&so->so_snd) == 0) &&
13642 		    (acked > acked_amount) &&
13643 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13644 		    (tp->t_flags & TF_SENTFIN)) {
13645 			/*
13646 			 * We must be sure our fin
13647 			 * was sent and acked (we can be
13648 			 * in FIN_WAIT_1 without having
13649 			 * sent the fin).
13650 			 */
13651 			ourfinisacked = 1;
13652 			/*
13653 			 * Lets make sure snd_una is updated
13654 			 * since most likely acked_amount = 0 (it
13655 			 * should be).
13656 			 */
13657 			tp->snd_una = high_seq;
13658 		}
13659 		/* Did we make a RTO error? */
13660 		if ((tp->t_flags & TF_PREVVALID) &&
13661 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13662 			tp->t_flags &= ~TF_PREVVALID;
13663 			if (tp->t_rxtshift == 1 &&
13664 			    (int)(ticks - tp->t_badrxtwin) < 0)
13665 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
13666 		}
13667 		/* Handle the data in the socket buffer */
13668 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13669 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13670 		if (acked_amount > 0) {
13671 			struct mbuf *mfree;
13672 
13673 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13674 			SOCKBUF_LOCK(&so->so_snd);
13675 			mfree = sbcut_locked(&so->so_snd, acked_amount);
13676 			tp->snd_una = high_seq;
13677 			/* Note we want to hold the sb lock through the sendmap adjust */
13678 			rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13679 			/* Wake up the socket if we have room to write more */
13680 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13681 			sowwakeup_locked(so);
13682 			m_freem(mfree);
13683 		}
13684 		/* update progress */
13685 		tp->t_acktime = ticks;
13686 		rack_log_progress_event(rack, tp, tp->t_acktime,
13687 					PROGRESS_UPDATE, __LINE__);
13688 		/* Clear out shifts and such */
13689 		tp->t_rxtshift = 0;
13690 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13691 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13692 		rack->rc_tlp_in_progress = 0;
13693 		rack->r_ctl.rc_tlp_cnt_out = 0;
13694 		/* Send recover and snd_nxt must be dragged along */
13695 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
13696 			tp->snd_recover = tp->snd_una;
13697 		if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13698 			tp->snd_nxt = tp->snd_una;
13699 		/*
13700 		 * If the RXT timer is running we want to
13701 		 * stop it, so we can restart a TLP (or new RXT).
13702 		 */
13703 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13704 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13705 #ifdef NETFLIX_HTTP_LOGGING
13706 		tcp_http_check_for_comp(rack->rc_tp, high_seq);
13707 #endif
13708 		tp->snd_wl2 = high_seq;
13709 		tp->t_dupacks = 0;
13710 		if (under_pacing &&
13711 		    (rack->use_fixed_rate == 0) &&
13712 		    (rack->in_probe_rtt == 0) &&
13713 		    rack->rc_gp_dyn_mul &&
13714 		    rack->rc_always_pace) {
13715 			/* Check if we are dragging bottom */
13716 			rack_check_bottom_drag(tp, rack, so, acked);
13717 		}
13718 		if (tp->snd_una == tp->snd_max) {
13719 			tp->t_flags &= ~TF_PREVVALID;
13720 			rack->r_ctl.retran_during_recovery = 0;
13721 			rack->r_ctl.dsack_byte_cnt = 0;
13722 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13723 			if (rack->r_ctl.rc_went_idle_time == 0)
13724 				rack->r_ctl.rc_went_idle_time = 1;
13725 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13726 			if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13727 				tp->t_acktime = 0;
13728 			/* Set so we might enter persists... */
13729 			rack->r_wanted_output = 1;
13730 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13731 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13732 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13733 			    (sbavail(&so->so_snd) == 0) &&
13734 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13735 				/*
13736 				 * The socket was gone and the
13737 				 * peer sent data (not now in the past), time to
13738 				 * reset him.
13739 				 */
13740 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13741 				/* tcp_close will kill the inp pre-log the Reset */
13742 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13743 #ifdef TCP_ACCOUNTING
13744 				rdstc = get_cyclecount();
13745 				if (rdstc > ts_val) {
13746 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13747 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13748 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13749 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13750 					}
13751 				}
13752 #endif
13753 				m_freem(m);
13754 				tp = tcp_close(tp);
13755 				if (tp == NULL) {
13756 #ifdef TCP_ACCOUNTING
13757 					sched_unpin();
13758 #endif
13759 					return (1);
13760 				}
13761 				/*
13762 				 * We would normally do drop-with-reset which would
13763 				 * send back a reset. We can't since we don't have
13764 				 * all the needed bits. Instead lets arrange for
13765 				 * a call to tcp_output(). That way since we
13766 				 * are in the closed state we will generate a reset.
13767 				 *
13768 				 * Note if tcp_accounting is on we don't unpin since
13769 				 * we do that after the goto label.
13770 				 */
13771 				goto send_out_a_rst;
13772 			}
13773 			if ((sbused(&so->so_snd) == 0) &&
13774 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13775 			    (tp->t_flags & TF_SENTFIN)) {
13776 				/*
13777 				 * If we can't receive any more data, then closing user can
13778 				 * proceed. Starting the timer is contrary to the
13779 				 * specification, but if we don't get a FIN we'll hang
13780 				 * forever.
13781 				 *
13782 				 */
13783 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13784 					soisdisconnected(so);
13785 					tcp_timer_activate(tp, TT_2MSL,
13786 							   (tcp_fast_finwait2_recycle ?
13787 							    tcp_finwait2_timeout :
13788 							    TP_MAXIDLE(tp)));
13789 				}
13790 				if (ourfinisacked == 0) {
13791 					/*
13792 					 * We don't change to fin-wait-2 if we have our fin acked
13793 					 * which means we are probably in TCPS_CLOSING.
13794 					 */
13795 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
13796 				}
13797 			}
13798 		}
13799 		/* Wake up the socket if we have room to write more */
13800 		if (sbavail(&so->so_snd)) {
13801 			rack->r_wanted_output = 1;
13802 			if (ctf_progress_timeout_check(tp, true)) {
13803 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13804 							tp, tick, PROGRESS_DROP, __LINE__);
13805 				/*
13806 				 * We cheat here and don't send a RST, we should send one
13807 				 * when the pacer drops the connection.
13808 				 */
13809 #ifdef TCP_ACCOUNTING
13810 				rdstc = get_cyclecount();
13811 				if (rdstc > ts_val) {
13812 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13813 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13814 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13815 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13816 					}
13817 				}
13818 				sched_unpin();
13819 #endif
13820 				(void)tcp_drop(tp, ETIMEDOUT);
13821 				m_freem(m);
13822 				return (1);
13823 			}
13824 		}
13825 		if (ourfinisacked) {
13826 			switch(tp->t_state) {
13827 			case TCPS_CLOSING:
13828 #ifdef TCP_ACCOUNTING
13829 				rdstc = get_cyclecount();
13830 				if (rdstc > ts_val) {
13831 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13832 							(rdstc - ts_val));
13833 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13834 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13835 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13836 					}
13837 				}
13838 				sched_unpin();
13839 #endif
13840 				tcp_twstart(tp);
13841 				m_freem(m);
13842 				return (1);
13843 				break;
13844 			case TCPS_LAST_ACK:
13845 #ifdef TCP_ACCOUNTING
13846 				rdstc = get_cyclecount();
13847 				if (rdstc > ts_val) {
13848 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13849 							(rdstc - ts_val));
13850 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13851 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13852 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13853 					}
13854 				}
13855 				sched_unpin();
13856 #endif
13857 				tp = tcp_close(tp);
13858 				ctf_do_drop(m, tp);
13859 				return (1);
13860 				break;
13861 			case TCPS_FIN_WAIT_1:
13862 #ifdef TCP_ACCOUNTING
13863 				rdstc = get_cyclecount();
13864 				if (rdstc > ts_val) {
13865 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13866 							(rdstc - ts_val));
13867 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13868 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13869 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13870 					}
13871 				}
13872 #endif
13873 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13874 					soisdisconnected(so);
13875 					tcp_timer_activate(tp, TT_2MSL,
13876 							   (tcp_fast_finwait2_recycle ?
13877 							    tcp_finwait2_timeout :
13878 							    TP_MAXIDLE(tp)));
13879 				}
13880 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13881 				break;
13882 			default:
13883 				break;
13884 			}
13885 		}
13886 		if (rack->r_fast_output) {
13887 			/*
13888 			 * We re doing fast output.. can we expand that?
13889 			 */
13890 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13891 		}
13892 #ifdef TCP_ACCOUNTING
13893 		rdstc = get_cyclecount();
13894 		if (rdstc > ts_val) {
13895 			counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13896 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13897 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13898 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13899 			}
13900 		}
13901 
13902 	} else if (win_up_req) {
13903 		rdstc = get_cyclecount();
13904 		if (rdstc > ts_val) {
13905 			counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13906 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13907 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13908 			}
13909 		}
13910 #endif
13911 	}
13912 	/* Now is there a next packet, if so we are done */
13913 	m_freem(m);
13914 	did_out = 0;
13915 	if (nxt_pkt) {
13916 #ifdef TCP_ACCOUNTING
13917 		sched_unpin();
13918 #endif
13919 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13920 		return (0);
13921 	}
13922 	rack_handle_might_revert(tp, rack);
13923 	ctf_calc_rwin(so, tp);
13924 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13925 	send_out_a_rst:
13926 		if (tcp_output(tp) < 0) {
13927 #ifdef TCP_ACCOUNTING
13928 			sched_unpin();
13929 #endif
13930 			return (1);
13931 		}
13932 		did_out = 1;
13933 	}
13934 	rack_free_trim(rack);
13935 #ifdef TCP_ACCOUNTING
13936 	sched_unpin();
13937 #endif
13938 	rack_timer_audit(tp, rack, &so->so_snd);
13939 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13940 	return (0);
13941 }
13942 
13943 
13944 static int
13945 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13946     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13947     int32_t nxt_pkt, struct timeval *tv)
13948 {
13949 #ifdef TCP_ACCOUNTING
13950 	uint64_t ts_val;
13951 #endif
13952 	int32_t thflags, retval, did_out = 0;
13953 	int32_t way_out = 0;
13954 	/*
13955 	 * cts - is the current time from tv (caller gets ts) in microseconds.
13956 	 * ms_cts - is the current time from tv in milliseconds.
13957 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
13958 	 */
13959 	uint32_t cts, us_cts, ms_cts;
13960 	uint32_t tiwin, high_seq;
13961 	struct timespec ts;
13962 	struct tcpopt to;
13963 	struct tcp_rack *rack;
13964 	struct rack_sendmap *rsm;
13965 	int32_t prev_state = 0;
13966 #ifdef TCP_ACCOUNTING
13967 	int ack_val_set = 0xf;
13968 #endif
13969 	int nsegs;
13970 	/*
13971 	 * tv passed from common code is from either M_TSTMP_LRO or
13972 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13973 	 */
13974 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13975 	if (m->m_flags & M_ACKCMP) {
13976 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13977 	}
13978 	if (m->m_flags & M_ACKCMP) {
13979 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13980 	}
13981 	cts = tcp_tv_to_usectick(tv);
13982 	ms_cts =  tcp_tv_to_mssectick(tv);
13983 	nsegs = m->m_pkthdr.lro_nsegs;
13984 	counter_u64_add(rack_proc_non_comp_ack, 1);
13985 	thflags = tcp_get_flags(th);
13986 #ifdef TCP_ACCOUNTING
13987 	sched_pin();
13988 	if (thflags & TH_ACK)
13989 		ts_val = get_cyclecount();
13990 #endif
13991 	if ((m->m_flags & M_TSTMP) ||
13992 	    (m->m_flags & M_TSTMP_LRO)) {
13993 		mbuf_tstmp2timespec(m, &ts);
13994 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13995 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13996 	} else
13997 		rack->r_ctl.act_rcv_time = *tv;
13998 	kern_prefetch(rack, &prev_state);
13999 	prev_state = 0;
14000 	/*
14001 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
14002 	 * the scale is zero.
14003 	 */
14004 	tiwin = th->th_win << tp->snd_scale;
14005 #ifdef TCP_ACCOUNTING
14006 	if (thflags & TH_ACK) {
14007 		/*
14008 		 * We have a tradeoff here. We can either do what we are
14009 		 * doing i.e. pinning to this CPU and then doing the accounting
14010 		 * <or> we could do a critical enter, setup the rdtsc and cpu
14011 		 * as in below, and then validate we are on the same CPU on
14012 		 * exit. I have choosen to not do the critical enter since
14013 		 * that often will gain you a context switch, and instead lock
14014 		 * us (line above this if) to the same CPU with sched_pin(). This
14015 		 * means we may be context switched out for a higher priority
14016 		 * interupt but we won't be moved to another CPU.
14017 		 *
14018 		 * If this occurs (which it won't very often since we most likely
14019 		 * are running this code in interupt context and only a higher
14020 		 * priority will bump us ... clock?) we will falsely add in
14021 		 * to the time the interupt processing time plus the ack processing
14022 		 * time. This is ok since its a rare event.
14023 		 */
14024 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
14025 						    ctf_fixed_maxseg(tp));
14026 	}
14027 #endif
14028 	/*
14029 	 * Parse options on any incoming segment.
14030 	 */
14031 	memset(&to, 0, sizeof(to));
14032 	tcp_dooptions(&to, (u_char *)(th + 1),
14033 	    (th->th_off << 2) - sizeof(struct tcphdr),
14034 	    (thflags & TH_SYN) ? TO_SYN : 0);
14035 	NET_EPOCH_ASSERT();
14036 	INP_WLOCK_ASSERT(tp->t_inpcb);
14037 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
14038 	    __func__));
14039 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
14040 	    __func__));
14041 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
14042 	    (tp->t_flags & TF_GPUTINPROG)) {
14043 		/*
14044 		 * We have a goodput in progress
14045 		 * and we have entered a late state.
14046 		 * Do we have enough data in the sb
14047 		 * to handle the GPUT request?
14048 		 */
14049 		uint32_t bytes;
14050 
14051 		bytes = tp->gput_ack - tp->gput_seq;
14052 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
14053 			bytes += tp->gput_seq - tp->snd_una;
14054 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
14055 			/*
14056 			 * There are not enough bytes in the socket
14057 			 * buffer that have been sent to cover this
14058 			 * measurement. Cancel it.
14059 			 */
14060 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14061 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
14062 						   tp->gput_seq,
14063 						   0, 0, 18, __LINE__, NULL, 0);
14064 			tp->t_flags &= ~TF_GPUTINPROG;
14065 		}
14066 	}
14067 	high_seq = th->th_ack;
14068 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14069 		union tcp_log_stackspecific log;
14070 		struct timeval ltv;
14071 #ifdef NETFLIX_HTTP_LOGGING
14072 		struct http_sendfile_track *http_req;
14073 
14074 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
14075 			http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14076 		} else {
14077 			http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14078 		}
14079 #endif
14080 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14081 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14082 		if (rack->rack_no_prr == 0)
14083 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14084 		else
14085 			log.u_bbr.flex1 = 0;
14086 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14087 		log.u_bbr.use_lt_bw <<= 1;
14088 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
14089 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14090 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14091 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14092 		log.u_bbr.flex3 = m->m_flags;
14093 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14094 		log.u_bbr.lost = thflags;
14095 		log.u_bbr.pacing_gain = 0x1;
14096 #ifdef TCP_ACCOUNTING
14097 		log.u_bbr.cwnd_gain = ack_val_set;
14098 #endif
14099 		log.u_bbr.flex7 = 2;
14100 		if (m->m_flags & M_TSTMP) {
14101 			/* Record the hardware timestamp if present */
14102 			mbuf_tstmp2timespec(m, &ts);
14103 			ltv.tv_sec = ts.tv_sec;
14104 			ltv.tv_usec = ts.tv_nsec / 1000;
14105 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14106 		} else if (m->m_flags & M_TSTMP_LRO) {
14107 			/* Record the LRO the arrival timestamp */
14108 			mbuf_tstmp2timespec(m, &ts);
14109 			ltv.tv_sec = ts.tv_sec;
14110 			ltv.tv_usec = ts.tv_nsec / 1000;
14111 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14112 		}
14113 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14114 		/* Log the rcv time */
14115 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14116 #ifdef NETFLIX_HTTP_LOGGING
14117 		log.u_bbr.applimited = tp->t_http_closed;
14118 		log.u_bbr.applimited <<= 8;
14119 		log.u_bbr.applimited |= tp->t_http_open;
14120 		log.u_bbr.applimited <<= 8;
14121 		log.u_bbr.applimited |= tp->t_http_req;
14122 		if (http_req) {
14123 			/* Copy out any client req info */
14124 			/* seconds */
14125 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14126 			/* useconds */
14127 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14128 			log.u_bbr.rttProp = http_req->timestamp;
14129 			log.u_bbr.cur_del_rate = http_req->start;
14130 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14131 				log.u_bbr.flex8 |= 1;
14132 			} else {
14133 				log.u_bbr.flex8 |= 2;
14134 				log.u_bbr.bw_inuse = http_req->end;
14135 			}
14136 			log.u_bbr.flex6 = http_req->start_seq;
14137 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14138 				log.u_bbr.flex8 |= 4;
14139 				log.u_bbr.epoch = http_req->end_seq;
14140 			}
14141 		}
14142 #endif
14143 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14144 		    tlen, &log, true, &ltv);
14145 	}
14146 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14147 		way_out = 4;
14148 		retval = 0;
14149 		m_freem(m);
14150 		goto done_with_input;
14151 	}
14152 	/*
14153 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
14154 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14155 	 */
14156 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14157 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14158 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14159 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14160 #ifdef TCP_ACCOUNTING
14161 		sched_unpin();
14162 #endif
14163 		return (1);
14164 	}
14165 	/*
14166 	 * If timestamps were negotiated during SYN/ACK and a
14167 	 * segment without a timestamp is received, silently drop
14168 	 * the segment, unless it is a RST segment or missing timestamps are
14169 	 * tolerated.
14170 	 * See section 3.2 of RFC 7323.
14171 	 */
14172 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14173 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14174 		way_out = 5;
14175 		retval = 0;
14176 		m_freem(m);
14177 		goto done_with_input;
14178 	}
14179 
14180 	/*
14181 	 * Segment received on connection. Reset idle time and keep-alive
14182 	 * timer. XXX: This should be done after segment validation to
14183 	 * ignore broken/spoofed segs.
14184 	 */
14185 	if  (tp->t_idle_reduce &&
14186 	     (tp->snd_max == tp->snd_una) &&
14187 	     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14188 		counter_u64_add(rack_input_idle_reduces, 1);
14189 		rack_cc_after_idle(rack, tp);
14190 	}
14191 	tp->t_rcvtime = ticks;
14192 #ifdef STATS
14193 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14194 #endif
14195 	if (tiwin > rack->r_ctl.rc_high_rwnd)
14196 		rack->r_ctl.rc_high_rwnd = tiwin;
14197 	/*
14198 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14199 	 * this to occur after we've validated the segment.
14200 	 */
14201 	if (tcp_ecn_input_segment(tp, thflags, iptos))
14202 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
14203 
14204 	/*
14205 	 * If echoed timestamp is later than the current time, fall back to
14206 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14207 	 * were used when this connection was established.
14208 	 */
14209 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14210 		to.to_tsecr -= tp->ts_offset;
14211 		if (TSTMP_GT(to.to_tsecr, ms_cts))
14212 			to.to_tsecr = 0;
14213 	}
14214 
14215 	/*
14216 	 * If its the first time in we need to take care of options and
14217 	 * verify we can do SACK for rack!
14218 	 */
14219 	if (rack->r_state == 0) {
14220 		/* Should be init'd by rack_init() */
14221 		KASSERT(rack->rc_inp != NULL,
14222 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
14223 		if (rack->rc_inp == NULL) {
14224 			rack->rc_inp = tp->t_inpcb;
14225 		}
14226 
14227 		/*
14228 		 * Process options only when we get SYN/ACK back. The SYN
14229 		 * case for incoming connections is handled in tcp_syncache.
14230 		 * According to RFC1323 the window field in a SYN (i.e., a
14231 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14232 		 * this is traditional behavior, may need to be cleaned up.
14233 		 */
14234 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14235 			/* Handle parallel SYN for ECN */
14236 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
14237 			if ((to.to_flags & TOF_SCALE) &&
14238 			    (tp->t_flags & TF_REQ_SCALE)) {
14239 				tp->t_flags |= TF_RCVD_SCALE;
14240 				tp->snd_scale = to.to_wscale;
14241 			} else
14242 				tp->t_flags &= ~TF_REQ_SCALE;
14243 			/*
14244 			 * Initial send window.  It will be updated with the
14245 			 * next incoming segment to the scaled value.
14246 			 */
14247 			tp->snd_wnd = th->th_win;
14248 			rack_validate_fo_sendwin_up(tp, rack);
14249 			if ((to.to_flags & TOF_TS) &&
14250 			    (tp->t_flags & TF_REQ_TSTMP)) {
14251 				tp->t_flags |= TF_RCVD_TSTMP;
14252 				tp->ts_recent = to.to_tsval;
14253 				tp->ts_recent_age = cts;
14254 			} else
14255 				tp->t_flags &= ~TF_REQ_TSTMP;
14256 			if (to.to_flags & TOF_MSS) {
14257 				tcp_mss(tp, to.to_mss);
14258 			}
14259 			if ((tp->t_flags & TF_SACK_PERMIT) &&
14260 			    (to.to_flags & TOF_SACKPERM) == 0)
14261 				tp->t_flags &= ~TF_SACK_PERMIT;
14262 			if (IS_FASTOPEN(tp->t_flags)) {
14263 				if (to.to_flags & TOF_FASTOPEN) {
14264 					uint16_t mss;
14265 
14266 					if (to.to_flags & TOF_MSS)
14267 						mss = to.to_mss;
14268 					else
14269 						if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
14270 							mss = TCP6_MSS;
14271 						else
14272 							mss = TCP_MSS;
14273 					tcp_fastopen_update_cache(tp, mss,
14274 					    to.to_tfo_len, to.to_tfo_cookie);
14275 				} else
14276 					tcp_fastopen_disable_path(tp);
14277 			}
14278 		}
14279 		/*
14280 		 * At this point we are at the initial call. Here we decide
14281 		 * if we are doing RACK or not. We do this by seeing if
14282 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
14283 		 * The code now does do dup-ack counting so if you don't
14284 		 * switch back you won't get rack & TLP, but you will still
14285 		 * get this stack.
14286 		 */
14287 
14288 		if ((rack_sack_not_required == 0) &&
14289 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14290 			tcp_switch_back_to_default(tp);
14291 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14292 			    tlen, iptos);
14293 #ifdef TCP_ACCOUNTING
14294 			sched_unpin();
14295 #endif
14296 			return (1);
14297 		}
14298 		tcp_set_hpts(tp->t_inpcb);
14299 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14300 	}
14301 	if (thflags & TH_FIN)
14302 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14303 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14304 	if ((rack->rc_gp_dyn_mul) &&
14305 	    (rack->use_fixed_rate == 0) &&
14306 	    (rack->rc_always_pace)) {
14307 		/* Check in on probertt */
14308 		rack_check_probe_rtt(rack, us_cts);
14309 	}
14310 	rack_clear_rate_sample(rack);
14311 	if ((rack->forced_ack) &&
14312 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
14313 		rack_handle_probe_response(rack, tiwin, us_cts);
14314 	}
14315 	/*
14316 	 * This is the one exception case where we set the rack state
14317 	 * always. All other times (timers etc) we must have a rack-state
14318 	 * set (so we assure we have done the checks above for SACK).
14319 	 */
14320 	rack->r_ctl.rc_rcvtime = cts;
14321 	if (rack->r_state != tp->t_state)
14322 		rack_set_state(tp, rack);
14323 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
14324 	    (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14325 		kern_prefetch(rsm, &prev_state);
14326 	prev_state = rack->r_state;
14327 	retval = (*rack->r_substate) (m, th, so,
14328 	    tp, &to, drop_hdrlen,
14329 	    tlen, tiwin, thflags, nxt_pkt, iptos);
14330 #ifdef INVARIANTS
14331 	if ((retval == 0) &&
14332 	    (tp->t_inpcb == NULL)) {
14333 		panic("retval:%d tp:%p t_inpcb:NULL state:%d",
14334 		    retval, tp, prev_state);
14335 	}
14336 #endif
14337 	if (retval == 0) {
14338 		/*
14339 		 * If retval is 1 the tcb is unlocked and most likely the tp
14340 		 * is gone.
14341 		 */
14342 		INP_WLOCK_ASSERT(tp->t_inpcb);
14343 		if ((rack->rc_gp_dyn_mul) &&
14344 		    (rack->rc_always_pace) &&
14345 		    (rack->use_fixed_rate == 0) &&
14346 		    rack->in_probe_rtt &&
14347 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
14348 			/*
14349 			 * If we are going for target, lets recheck before
14350 			 * we output.
14351 			 */
14352 			rack_check_probe_rtt(rack, us_cts);
14353 		}
14354 		if (rack->set_pacing_done_a_iw == 0) {
14355 			/* How much has been acked? */
14356 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14357 				/* We have enough to set in the pacing segment size */
14358 				rack->set_pacing_done_a_iw = 1;
14359 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
14360 			}
14361 		}
14362 		tcp_rack_xmit_timer_commit(rack, tp);
14363 #ifdef TCP_ACCOUNTING
14364 		/*
14365 		 * If we set the ack_val_se to what ack processing we are doing
14366 		 * we also want to track how many cycles we burned. Note
14367 		 * the bits after tcp_output we let be "free". This is because
14368 		 * we are also tracking the tcp_output times as well. Note the
14369 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
14370 		 * 0xf cannot be returned and is what we initialize it too to
14371 		 * indicate we are not doing the tabulations.
14372 		 */
14373 		if (ack_val_set != 0xf) {
14374 			uint64_t crtsc;
14375 
14376 			crtsc = get_cyclecount();
14377 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14378 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14379 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14380 			}
14381 		}
14382 #endif
14383 		if (nxt_pkt == 0) {
14384 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14385 do_output_now:
14386 				if (tcp_output(tp) < 0)
14387 					return (1);
14388 				did_out = 1;
14389 			}
14390 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14391 			rack_free_trim(rack);
14392 		}
14393 		/* Update any rounds needed */
14394 		if (rack_verbose_logging &&  (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
14395 			union tcp_log_stackspecific log;
14396 			struct timeval tv;
14397 
14398 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14399 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14400 			log.u_bbr.flex1 = high_seq;
14401 			log.u_bbr.flex2 = rack->r_ctl.roundends;
14402 			log.u_bbr.flex3 = rack->r_ctl.current_round;
14403 			log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
14404 			log.u_bbr.flex8 = 9;
14405 			tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
14406 				       0, &log, false, NULL, NULL, 0, &tv);
14407 		}
14408 		/*
14409 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
14410 		 * causes issues when we are just going app limited. Lets
14411 		 * instead use SEQ_GT <or> where its equal but more data
14412 		 * is outstanding.
14413 		 */
14414 		if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) ||
14415 		    ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) {
14416 			rack->r_ctl.current_round++;
14417 			rack->r_ctl.roundends = tp->snd_max;
14418 			if (CC_ALGO(tp)->newround != NULL) {
14419 				CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
14420 			}
14421 		}
14422 		if ((nxt_pkt == 0) &&
14423 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14424 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
14425 		     (tp->t_flags & TF_DELACK) ||
14426 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14427 		      (tp->t_state <= TCPS_CLOSING)))) {
14428 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
14429 			if ((tp->snd_max == tp->snd_una) &&
14430 			    ((tp->t_flags & TF_DELACK) == 0) &&
14431 			    (tcp_in_hpts(rack->rc_inp)) &&
14432 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14433 				/* keep alive not needed if we are hptsi output yet */
14434 				;
14435 			} else {
14436 				int late = 0;
14437 				if (tcp_in_hpts(rack->rc_inp)) {
14438 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14439 						us_cts = tcp_get_usecs(NULL);
14440 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14441 							rack->r_early = 1;
14442 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14443 						} else
14444 							late = 1;
14445 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14446 					}
14447 					tcp_hpts_remove(tp->t_inpcb);
14448 				}
14449 				if (late && (did_out == 0)) {
14450 					/*
14451 					 * We are late in the sending
14452 					 * and we did not call the output
14453 					 * (this probably should not happen).
14454 					 */
14455 					goto do_output_now;
14456 				}
14457 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14458 			}
14459 			way_out = 1;
14460 		} else if (nxt_pkt == 0) {
14461 			/* Do we have the correct timer running? */
14462 			rack_timer_audit(tp, rack, &so->so_snd);
14463 			way_out = 2;
14464 		}
14465 	done_with_input:
14466 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14467 		if (did_out)
14468 			rack->r_wanted_output = 0;
14469 #ifdef INVARIANTS
14470 		if (tp->t_inpcb == NULL) {
14471 			panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
14472 			      did_out,
14473 			      retval, tp, prev_state);
14474 		}
14475 #endif
14476 #ifdef TCP_ACCOUNTING
14477 	} else {
14478 		/*
14479 		 * Track the time (see above).
14480 		 */
14481 		if (ack_val_set != 0xf) {
14482 			uint64_t crtsc;
14483 
14484 			crtsc = get_cyclecount();
14485 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14486 			/*
14487 			 * Note we *DO NOT* increment the per-tcb counters since
14488 			 * in the else the TP may be gone!!
14489 			 */
14490 		}
14491 #endif
14492 	}
14493 #ifdef TCP_ACCOUNTING
14494 	sched_unpin();
14495 #endif
14496 	return (retval);
14497 }
14498 
14499 void
14500 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14501     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14502 {
14503 	struct timeval tv;
14504 
14505 	/* First lets see if we have old packets */
14506 	if (tp->t_in_pkt) {
14507 		if (ctf_do_queued_segments(so, tp, 1)) {
14508 			m_freem(m);
14509 			return;
14510 		}
14511 	}
14512 	if (m->m_flags & M_TSTMP_LRO) {
14513 		tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
14514 		tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
14515 	} else {
14516 		/* Should not be should we kassert instead? */
14517 		tcp_get_usecs(&tv);
14518 	}
14519 	if (rack_do_segment_nounlock(m, th, so, tp,
14520 				     drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14521 		INP_WUNLOCK(tp->t_inpcb);
14522 	}
14523 }
14524 
14525 struct rack_sendmap *
14526 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14527 {
14528 	struct rack_sendmap *rsm = NULL;
14529 	int32_t idx;
14530 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
14531 
14532 	/* Return the next guy to be re-transmitted */
14533 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14534 		return (NULL);
14535 	}
14536 	if (tp->t_flags & TF_SENTFIN) {
14537 		/* retran the end FIN? */
14538 		return (NULL);
14539 	}
14540 	/* ok lets look at this one */
14541 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14542 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
14543 		return (rsm);
14544 	}
14545 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14546 		goto check_it;
14547 	}
14548 	rsm = rack_find_lowest_rsm(rack);
14549 	if (rsm == NULL) {
14550 		return (NULL);
14551 	}
14552 check_it:
14553 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14554 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14555 		/*
14556 		 * No sack so we automatically do the 3 strikes and
14557 		 * retransmit (no rack timer would be started).
14558 		 */
14559 
14560 		return (rsm);
14561 	}
14562 	if (rsm->r_flags & RACK_ACKED) {
14563 		return (NULL);
14564 	}
14565 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14566 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14567 		/* Its not yet ready */
14568 		return (NULL);
14569 	}
14570 	srtt = rack_grab_rtt(tp, rack);
14571 	idx = rsm->r_rtr_cnt - 1;
14572 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14573 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14574 	if ((tsused == ts_low) ||
14575 	    (TSTMP_LT(tsused, ts_low))) {
14576 		/* No time since sending */
14577 		return (NULL);
14578 	}
14579 	if ((tsused - ts_low) < thresh) {
14580 		/* It has not been long enough yet */
14581 		return (NULL);
14582 	}
14583 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14584 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
14585 	     (rack->sack_attack_disable == 0))) {
14586 		/*
14587 		 * We have passed the dup-ack threshold <or>
14588 		 * a SACK has indicated this is missing.
14589 		 * Note that if you are a declared attacker
14590 		 * it is only the dup-ack threshold that
14591 		 * will cause retransmits.
14592 		 */
14593 		/* log retransmit reason */
14594 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14595 		rack->r_fast_output = 0;
14596 		return (rsm);
14597 	}
14598 	return (NULL);
14599 }
14600 
14601 static void
14602 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14603 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14604 			   int line, struct rack_sendmap *rsm, uint8_t quality)
14605 {
14606 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14607 		union tcp_log_stackspecific log;
14608 		struct timeval tv;
14609 
14610 		memset(&log, 0, sizeof(log));
14611 		log.u_bbr.flex1 = slot;
14612 		log.u_bbr.flex2 = len;
14613 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14614 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14615 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14616 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14617 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14618 		log.u_bbr.use_lt_bw <<= 1;
14619 		log.u_bbr.use_lt_bw |= rack->r_late;
14620 		log.u_bbr.use_lt_bw <<= 1;
14621 		log.u_bbr.use_lt_bw |= rack->r_early;
14622 		log.u_bbr.use_lt_bw <<= 1;
14623 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14624 		log.u_bbr.use_lt_bw <<= 1;
14625 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14626 		log.u_bbr.use_lt_bw <<= 1;
14627 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14628 		log.u_bbr.use_lt_bw <<= 1;
14629 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14630 		log.u_bbr.use_lt_bw <<= 1;
14631 		log.u_bbr.use_lt_bw |= rack->gp_ready;
14632 		log.u_bbr.pkt_epoch = line;
14633 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14634 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14635 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14636 		log.u_bbr.bw_inuse = bw_est;
14637 		log.u_bbr.delRate = bw;
14638 		if (rack->r_ctl.gp_bw == 0)
14639 			log.u_bbr.cur_del_rate = 0;
14640 		else
14641 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
14642 		log.u_bbr.rttProp = len_time;
14643 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14644 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14645 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14646 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14647 			/* We are in slow start */
14648 			log.u_bbr.flex7 = 1;
14649 		} else {
14650 			/* we are on congestion avoidance */
14651 			log.u_bbr.flex7 = 0;
14652 		}
14653 		log.u_bbr.flex8 = method;
14654 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14655 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14656 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14657 		log.u_bbr.cwnd_gain <<= 1;
14658 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14659 		log.u_bbr.cwnd_gain <<= 1;
14660 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14661 		log.u_bbr.bbr_substate = quality;
14662 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
14663 		    &rack->rc_inp->inp_socket->so_rcv,
14664 		    &rack->rc_inp->inp_socket->so_snd,
14665 		    BBR_LOG_HPTSI_CALC, 0,
14666 		    0, &log, false, &tv);
14667 	}
14668 }
14669 
14670 static uint32_t
14671 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14672 {
14673 	uint32_t new_tso, user_max;
14674 
14675 	user_max = rack->rc_user_set_max_segs * mss;
14676 	if (rack->rc_force_max_seg) {
14677 		return (user_max);
14678 	}
14679 	if (rack->use_fixed_rate &&
14680 	    ((rack->r_ctl.crte == NULL) ||
14681 	     (bw != rack->r_ctl.crte->rate))) {
14682 		/* Use the user mss since we are not exactly matched */
14683 		return (user_max);
14684 	}
14685 	new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14686 	if (new_tso > user_max)
14687 		new_tso = user_max;
14688 	return (new_tso);
14689 }
14690 
14691 static int32_t
14692 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)
14693 {
14694 	uint64_t lentim, fill_bw;
14695 
14696 	/* Lets first see if we are full, if so continue with normal rate */
14697 	rack->r_via_fill_cw = 0;
14698 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14699 		return (slot);
14700 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14701 		return (slot);
14702 	if (rack->r_ctl.rc_last_us_rtt == 0)
14703 		return (slot);
14704 	if (rack->rc_pace_fill_if_rttin_range &&
14705 	    (rack->r_ctl.rc_last_us_rtt >=
14706 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14707 		/* The rtt is huge, N * smallest, lets not fill */
14708 		return (slot);
14709 	}
14710 	/*
14711 	 * first lets calculate the b/w based on the last us-rtt
14712 	 * and the sndwnd.
14713 	 */
14714 	fill_bw = rack->r_ctl.cwnd_to_use;
14715 	/* Take the rwnd if its smaller */
14716 	if (fill_bw > rack->rc_tp->snd_wnd)
14717 		fill_bw = rack->rc_tp->snd_wnd;
14718 	if (rack->r_fill_less_agg) {
14719 		/*
14720 		 * Now take away the inflight (this will reduce our
14721 		 * aggressiveness and yeah, if we get that much out in 1RTT
14722 		 * we will have had acks come back and still be behind).
14723 		 */
14724 		fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14725 	}
14726 	/* Now lets make it into a b/w */
14727 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14728 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14729 	/* We are below the min b/w */
14730 	if (non_paced)
14731 		*rate_wanted = fill_bw;
14732 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14733 		return (slot);
14734 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14735 		fill_bw = rack->r_ctl.bw_rate_cap;
14736 	rack->r_via_fill_cw = 1;
14737 	if (rack->r_rack_hw_rate_caps &&
14738 	    (rack->r_ctl.crte != NULL)) {
14739 		uint64_t high_rate;
14740 
14741 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14742 		if (fill_bw > high_rate) {
14743 			/* We are capping bw at the highest rate table entry */
14744 			if (*rate_wanted > high_rate) {
14745 				/* The original rate was also capped */
14746 				rack->r_via_fill_cw = 0;
14747 			}
14748 			rack_log_hdwr_pacing(rack,
14749 					     fill_bw, high_rate, __LINE__,
14750 					     0, 3);
14751 			fill_bw = high_rate;
14752 			if (capped)
14753 				*capped = 1;
14754 		}
14755 	} else if ((rack->r_ctl.crte == NULL) &&
14756 		   (rack->rack_hdrw_pacing == 0) &&
14757 		   (rack->rack_hdw_pace_ena) &&
14758 		   rack->r_rack_hw_rate_caps &&
14759 		   (rack->rack_attempt_hdwr_pace == 0) &&
14760 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
14761 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14762 		/*
14763 		 * Ok we may have a first attempt that is greater than our top rate
14764 		 * lets check.
14765 		 */
14766 		uint64_t high_rate;
14767 
14768 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14769 		if (high_rate) {
14770 			if (fill_bw > high_rate) {
14771 				fill_bw = high_rate;
14772 				if (capped)
14773 					*capped = 1;
14774 			}
14775 		}
14776 	}
14777 	/*
14778 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
14779 	 * in a rtt, what does that time wise equate too?
14780 	 */
14781 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14782 	lentim /= fill_bw;
14783 	*rate_wanted = fill_bw;
14784 	if (non_paced || (lentim < slot)) {
14785 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14786 					   0, lentim, 12, __LINE__, NULL, 0);
14787 		return ((int32_t)lentim);
14788 	} else
14789 		return (slot);
14790 }
14791 
14792 static int32_t
14793 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14794 {
14795 	uint64_t srtt;
14796 	int32_t slot = 0;
14797 	int can_start_hw_pacing = 1;
14798 	int err;
14799 
14800 	if (rack->rc_always_pace == 0) {
14801 		/*
14802 		 * We use the most optimistic possible cwnd/srtt for
14803 		 * sending calculations. This will make our
14804 		 * calculation anticipate getting more through
14805 		 * quicker then possible. But thats ok we don't want
14806 		 * the peer to have a gap in data sending.
14807 		 */
14808 		uint64_t cwnd, tr_perms = 0;
14809 		int32_t reduce = 0;
14810 
14811 	old_method:
14812 		/*
14813 		 * We keep no precise pacing with the old method
14814 		 * instead we use the pacer to mitigate bursts.
14815 		 */
14816 		if (rack->r_ctl.rc_rack_min_rtt)
14817 			srtt = rack->r_ctl.rc_rack_min_rtt;
14818 		else
14819 			srtt = max(tp->t_srtt, 1);
14820 		if (rack->r_ctl.rc_rack_largest_cwnd)
14821 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14822 		else
14823 			cwnd = rack->r_ctl.cwnd_to_use;
14824 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
14825 		tr_perms = (cwnd * 1000) / srtt;
14826 		if (tr_perms == 0) {
14827 			tr_perms = ctf_fixed_maxseg(tp);
14828 		}
14829 		/*
14830 		 * Calculate how long this will take to drain, if
14831 		 * the calculation comes out to zero, thats ok we
14832 		 * will use send_a_lot to possibly spin around for
14833 		 * more increasing tot_len_this_send to the point
14834 		 * that its going to require a pace, or we hit the
14835 		 * cwnd. Which in that case we are just waiting for
14836 		 * a ACK.
14837 		 */
14838 		slot = len / tr_perms;
14839 		/* Now do we reduce the time so we don't run dry? */
14840 		if (slot && rack_slot_reduction) {
14841 			reduce = (slot / rack_slot_reduction);
14842 			if (reduce < slot) {
14843 				slot -= reduce;
14844 			} else
14845 				slot = 0;
14846 		}
14847 		slot *= HPTS_USEC_IN_MSEC;
14848 		if (rack->rc_pace_to_cwnd) {
14849 			uint64_t rate_wanted = 0;
14850 
14851 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14852 			rack->rc_ack_can_sendout_data = 1;
14853 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14854 		} else
14855 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14856 	} else {
14857 		uint64_t bw_est, res, lentim, rate_wanted;
14858 		uint32_t orig_val, segs, oh;
14859 		int capped = 0;
14860 		int prev_fill;
14861 
14862 		if ((rack->r_rr_config == 1) && rsm) {
14863 			return (rack->r_ctl.rc_min_to);
14864 		}
14865 		if (rack->use_fixed_rate) {
14866 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14867 		} else if ((rack->r_ctl.init_rate == 0) &&
14868 #ifdef NETFLIX_PEAKRATE
14869 			   (rack->rc_tp->t_maxpeakrate == 0) &&
14870 #endif
14871 			   (rack->r_ctl.gp_bw == 0)) {
14872 			/* no way to yet do an estimate */
14873 			bw_est = rate_wanted = 0;
14874 		} else {
14875 			bw_est = rack_get_bw(rack);
14876 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14877 		}
14878 		if ((bw_est == 0) || (rate_wanted == 0) ||
14879 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14880 			/*
14881 			 * No way yet to make a b/w estimate or
14882 			 * our raise is set incorrectly.
14883 			 */
14884 			goto old_method;
14885 		}
14886 		/* We need to account for all the overheads */
14887 		segs = (len + segsiz - 1) / segsiz;
14888 		/*
14889 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
14890 		 * and how much data we put in each packet. Yes this
14891 		 * means we may be off if we are larger than 1500 bytes
14892 		 * or smaller. But this just makes us more conservative.
14893 		 */
14894 		if (rack_hw_rate_min &&
14895 		    (bw_est < rack_hw_rate_min))
14896 			can_start_hw_pacing = 0;
14897 		if (ETHERNET_SEGMENT_SIZE > segsiz)
14898 			oh = ETHERNET_SEGMENT_SIZE - segsiz;
14899 		else
14900 			oh = 0;
14901 		segs *= oh;
14902 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14903 		res = lentim / rate_wanted;
14904 		slot = (uint32_t)res;
14905 		orig_val = rack->r_ctl.rc_pace_max_segs;
14906 		if (rack->r_ctl.crte == NULL) {
14907 			/*
14908 			 * Only do this if we are not hardware pacing
14909 			 * since if we are doing hw-pacing below we will
14910 			 * set make a call after setting up or changing
14911 			 * the rate.
14912 			 */
14913 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14914 		} else if (rack->rc_inp->inp_snd_tag == NULL) {
14915 			/*
14916 			 * We lost our rate somehow, this can happen
14917 			 * if the interface changed underneath us.
14918 			 */
14919 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14920 			rack->r_ctl.crte = NULL;
14921 			/* Lets re-allow attempting to setup pacing */
14922 			rack->rack_hdrw_pacing = 0;
14923 			rack->rack_attempt_hdwr_pace = 0;
14924 			rack_log_hdwr_pacing(rack,
14925 					     rate_wanted, bw_est, __LINE__,
14926 					     0, 6);
14927 		}
14928 		/* Did we change the TSO size, if so log it */
14929 		if (rack->r_ctl.rc_pace_max_segs != orig_val)
14930 			rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
14931 		prev_fill = rack->r_via_fill_cw;
14932 		if ((rack->rc_pace_to_cwnd) &&
14933 		    (capped == 0) &&
14934 		    (rack->use_fixed_rate == 0) &&
14935 		    (rack->in_probe_rtt == 0) &&
14936 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14937 			/*
14938 			 * We want to pace at our rate *or* faster to
14939 			 * fill the cwnd to the max if its not full.
14940 			 */
14941 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14942 		}
14943 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14944 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14945 			if ((rack->rack_hdw_pace_ena) &&
14946 			    (can_start_hw_pacing > 0) &&
14947 			    (rack->rack_hdrw_pacing == 0) &&
14948 			    (rack->rack_attempt_hdwr_pace == 0)) {
14949 				/*
14950 				 * Lets attempt to turn on hardware pacing
14951 				 * if we can.
14952 				 */
14953 				rack->rack_attempt_hdwr_pace = 1;
14954 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14955 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
14956 								       rate_wanted,
14957 								       RS_PACING_GEQ,
14958 								       &err, &rack->r_ctl.crte_prev_rate);
14959 				if (rack->r_ctl.crte) {
14960 					rack->rack_hdrw_pacing = 1;
14961 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14962 												 0, rack->r_ctl.crte,
14963 												 NULL);
14964 					rack_log_hdwr_pacing(rack,
14965 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14966 							     err, 0);
14967 					rack->r_ctl.last_hw_bw_req = rate_wanted;
14968 				} else {
14969 					counter_u64_add(rack_hw_pace_init_fail, 1);
14970 				}
14971 			} else if (rack->rack_hdrw_pacing &&
14972 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14973 				/* Do we need to adjust our rate? */
14974 				const struct tcp_hwrate_limit_table *nrte;
14975 
14976 				if (rack->r_up_only &&
14977 				    (rate_wanted < rack->r_ctl.crte->rate)) {
14978 					/**
14979 					 * We have four possible states here
14980 					 * having to do with the previous time
14981 					 * and this time.
14982 					 *   previous  |  this-time
14983 					 * A)     0      |     0   -- fill_cw not in the picture
14984 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
14985 					 * C)     1      |     1   -- all rates from fill_cw
14986 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
14987 					 *
14988 					 * For case A, C and D we don't allow a drop. But for
14989 					 * case B where we now our on our steady rate we do
14990 					 * allow a drop.
14991 					 *
14992 					 */
14993 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14994 						goto done_w_hdwr;
14995 				}
14996 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
14997 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14998 					if (rack_hw_rate_to_low &&
14999 					    (bw_est < rack_hw_rate_to_low)) {
15000 						/*
15001 						 * The pacing rate is too low for hardware, but
15002 						 * do allow hardware pacing to be restarted.
15003 						 */
15004 						rack_log_hdwr_pacing(rack,
15005 							     bw_est, rack->r_ctl.crte->rate, __LINE__,
15006 							     0, 5);
15007 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15008 						rack->r_ctl.crte = NULL;
15009 						rack->rack_attempt_hdwr_pace = 0;
15010 						rack->rack_hdrw_pacing = 0;
15011 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15012 						goto done_w_hdwr;
15013 					}
15014 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
15015 								   rack->rc_tp,
15016 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
15017 								   rate_wanted,
15018 								   RS_PACING_GEQ,
15019 								   &err, &rack->r_ctl.crte_prev_rate);
15020 					if (nrte == NULL) {
15021 						/* Lost the rate */
15022 						rack->rack_hdrw_pacing = 0;
15023 						rack->r_ctl.crte = NULL;
15024 						rack_log_hdwr_pacing(rack,
15025 								     rate_wanted, 0, __LINE__,
15026 								     err, 1);
15027 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15028 						counter_u64_add(rack_hw_pace_lost, 1);
15029 					} else if (nrte != rack->r_ctl.crte) {
15030 						rack->r_ctl.crte = nrte;
15031 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
15032 													 segsiz, 0,
15033 													 rack->r_ctl.crte,
15034 													 NULL);
15035 						rack_log_hdwr_pacing(rack,
15036 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15037 								     err, 2);
15038 						rack->r_ctl.last_hw_bw_req = rate_wanted;
15039 					}
15040 				} else {
15041 					/* We just need to adjust the segment size */
15042 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15043 					rack_log_hdwr_pacing(rack,
15044 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15045 							     0, 4);
15046 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15047 				}
15048 			}
15049 		}
15050 		if ((rack->r_ctl.crte != NULL) &&
15051 		    (rack->r_ctl.crte->rate == rate_wanted)) {
15052 			/*
15053 			 * We need to add a extra if the rates
15054 			 * are exactly matched. The idea is
15055 			 * we want the software to make sure the
15056 			 * queue is empty before adding more, this
15057 			 * gives us N MSS extra pace times where
15058 			 * N is our sysctl
15059 			 */
15060 			slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15061 		}
15062 done_w_hdwr:
15063 		if (rack_limit_time_with_srtt &&
15064 		    (rack->use_fixed_rate == 0) &&
15065 #ifdef NETFLIX_PEAKRATE
15066 		    (rack->rc_tp->t_maxpeakrate == 0) &&
15067 #endif
15068 		    (rack->rack_hdrw_pacing == 0)) {
15069 			/*
15070 			 * Sanity check, we do not allow the pacing delay
15071 			 * to be longer than the SRTT of the path. If it is
15072 			 * a slow path, then adding a packet should increase
15073 			 * the RTT and compensate for this i.e. the srtt will
15074 			 * be greater so the allowed pacing time will be greater.
15075 			 *
15076 			 * Note this restriction is not for where a peak rate
15077 			 * is set, we are doing fixed pacing or hardware pacing.
15078 			 */
15079 			if (rack->rc_tp->t_srtt)
15080 				srtt = rack->rc_tp->t_srtt;
15081 			else
15082 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
15083 			if (srtt < (uint64_t)slot) {
15084 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15085 				slot = srtt;
15086 			}
15087 		}
15088 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15089 	}
15090 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15091 		/*
15092 		 * If this rate is seeing enobufs when it
15093 		 * goes to send then either the nic is out
15094 		 * of gas or we are mis-estimating the time
15095 		 * somehow and not letting the queue empty
15096 		 * completely. Lets add to the pacing time.
15097 		 */
15098 		int hw_boost_delay;
15099 
15100 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15101 		if (hw_boost_delay > rack_enobuf_hw_max)
15102 			hw_boost_delay = rack_enobuf_hw_max;
15103 		else if (hw_boost_delay < rack_enobuf_hw_min)
15104 			hw_boost_delay = rack_enobuf_hw_min;
15105 		slot += hw_boost_delay;
15106 	}
15107 	return (slot);
15108 }
15109 
15110 static void
15111 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15112     tcp_seq startseq, uint32_t sb_offset)
15113 {
15114 	struct rack_sendmap *my_rsm = NULL;
15115 	struct rack_sendmap fe;
15116 
15117 	if (tp->t_state < TCPS_ESTABLISHED) {
15118 		/*
15119 		 * We don't start any measurements if we are
15120 		 * not at least established.
15121 		 */
15122 		return;
15123 	}
15124 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
15125 		/*
15126 		 * We will get no more data into the SB
15127 		 * this means we need to have the data available
15128 		 * before we start a measurement.
15129 		 */
15130 
15131 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) <
15132 		    max(rc_init_window(rack),
15133 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15134 			/* Nope not enough data */
15135 			return;
15136 		}
15137 	}
15138 	tp->t_flags |= TF_GPUTINPROG;
15139 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15140 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15141 	tp->gput_seq = startseq;
15142 	rack->app_limited_needs_set = 0;
15143 	if (rack->in_probe_rtt)
15144 		rack->measure_saw_probe_rtt = 1;
15145 	else if ((rack->measure_saw_probe_rtt) &&
15146 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15147 		rack->measure_saw_probe_rtt = 0;
15148 	if (rack->rc_gp_filled)
15149 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15150 	else {
15151 		/* Special case initial measurement */
15152 		struct timeval tv;
15153 
15154 		tp->gput_ts = tcp_get_usecs(&tv);
15155 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15156 	}
15157 	/*
15158 	 * We take a guess out into the future,
15159 	 * if we have no measurement and no
15160 	 * initial rate, we measure the first
15161 	 * initial-windows worth of data to
15162 	 * speed up getting some GP measurement and
15163 	 * thus start pacing.
15164 	 */
15165 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15166 		rack->app_limited_needs_set = 1;
15167 		tp->gput_ack = startseq + max(rc_init_window(rack),
15168 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15169 		rack_log_pacing_delay_calc(rack,
15170 					   tp->gput_seq,
15171 					   tp->gput_ack,
15172 					   0,
15173 					   tp->gput_ts,
15174 					   rack->r_ctl.rc_app_limited_cnt,
15175 					   9,
15176 					   __LINE__, NULL, 0);
15177 		return;
15178 	}
15179 	if (sb_offset) {
15180 		/*
15181 		 * We are out somewhere in the sb
15182 		 * can we use the already outstanding data?
15183 		 */
15184 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
15185 			/*
15186 			 * Yes first one is good and in this case
15187 			 * the tp->gput_ts is correctly set based on
15188 			 * the last ack that arrived (no need to
15189 			 * set things up when an ack comes in).
15190 			 */
15191 			my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15192 			if ((my_rsm == NULL) ||
15193 			    (my_rsm->r_rtr_cnt != 1)) {
15194 				/* retransmission? */
15195 				goto use_latest;
15196 			}
15197 		} else {
15198 			if (rack->r_ctl.rc_first_appl == NULL) {
15199 				/*
15200 				 * If rc_first_appl is NULL
15201 				 * then the cnt should be 0.
15202 				 * This is probably an error, maybe
15203 				 * a KASSERT would be approprate.
15204 				 */
15205 				goto use_latest;
15206 			}
15207 			/*
15208 			 * If we have a marker pointer to the last one that is
15209 			 * app limited we can use that, but we need to set
15210 			 * things up so that when it gets ack'ed we record
15211 			 * the ack time (if its not already acked).
15212 			 */
15213 			rack->app_limited_needs_set = 1;
15214 			/*
15215 			 * We want to get to the rsm that is either
15216 			 * next with space i.e. over 1 MSS or the one
15217 			 * after that (after the app-limited).
15218 			 */
15219 			my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15220 					 rack->r_ctl.rc_first_appl);
15221 			if (my_rsm) {
15222 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15223 					/* Have to use the next one */
15224 					my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15225 							 my_rsm);
15226 				else {
15227 					/* Use after the first MSS of it is acked */
15228 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15229 					goto start_set;
15230 				}
15231 			}
15232 			if ((my_rsm == NULL) ||
15233 			    (my_rsm->r_rtr_cnt != 1)) {
15234 				/*
15235 				 * Either its a retransmit or
15236 				 * the last is the app-limited one.
15237 				 */
15238 				goto use_latest;
15239 			}
15240 		}
15241 		tp->gput_seq = my_rsm->r_start;
15242 start_set:
15243 		if (my_rsm->r_flags & RACK_ACKED) {
15244 			/*
15245 			 * This one has been acked use the arrival ack time
15246 			 */
15247 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15248 			rack->app_limited_needs_set = 0;
15249 		}
15250 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15251 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15252 		rack_log_pacing_delay_calc(rack,
15253 					   tp->gput_seq,
15254 					   tp->gput_ack,
15255 					   (uint64_t)my_rsm,
15256 					   tp->gput_ts,
15257 					   rack->r_ctl.rc_app_limited_cnt,
15258 					   9,
15259 					   __LINE__, NULL, 0);
15260 		return;
15261 	}
15262 
15263 use_latest:
15264 	/*
15265 	 * We don't know how long we may have been
15266 	 * idle or if this is the first-send. Lets
15267 	 * setup the flag so we will trim off
15268 	 * the first ack'd data so we get a true
15269 	 * measurement.
15270 	 */
15271 	rack->app_limited_needs_set = 1;
15272 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15273 	/* Find this guy so we can pull the send time */
15274 	fe.r_start = startseq;
15275 	my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15276 	if (my_rsm) {
15277 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15278 		if (my_rsm->r_flags & RACK_ACKED) {
15279 			/*
15280 			 * Unlikely since its probably what was
15281 			 * just transmitted (but I am paranoid).
15282 			 */
15283 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15284 			rack->app_limited_needs_set = 0;
15285 		}
15286 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15287 			/* This also is unlikely */
15288 			tp->gput_seq = my_rsm->r_start;
15289 		}
15290 	} else {
15291 		/*
15292 		 * TSNH unless we have some send-map limit,
15293 		 * and even at that it should not be hitting
15294 		 * that limit (we should have stopped sending).
15295 		 */
15296 		struct timeval tv;
15297 
15298 		microuptime(&tv);
15299 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15300 	}
15301 	rack_log_pacing_delay_calc(rack,
15302 				   tp->gput_seq,
15303 				   tp->gput_ack,
15304 				   (uint64_t)my_rsm,
15305 				   tp->gput_ts,
15306 				   rack->r_ctl.rc_app_limited_cnt,
15307 				   9, __LINE__, NULL, 0);
15308 }
15309 
15310 static inline uint32_t
15311 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15312     uint32_t avail, int32_t sb_offset)
15313 {
15314 	uint32_t len;
15315 	uint32_t sendwin;
15316 
15317 	if (tp->snd_wnd > cwnd_to_use)
15318 		sendwin = cwnd_to_use;
15319 	else
15320 		sendwin = tp->snd_wnd;
15321 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
15322 		/* We never want to go over our peers rcv-window */
15323 		len = 0;
15324 	} else {
15325 		uint32_t flight;
15326 
15327 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15328 		if (flight >= sendwin) {
15329 			/*
15330 			 * We have in flight what we are allowed by cwnd (if
15331 			 * it was rwnd blocking it would have hit above out
15332 			 * >= tp->snd_wnd).
15333 			 */
15334 			return (0);
15335 		}
15336 		len = sendwin - flight;
15337 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15338 			/* We would send too much (beyond the rwnd) */
15339 			len = tp->snd_wnd - ctf_outstanding(tp);
15340 		}
15341 		if ((len + sb_offset) > avail) {
15342 			/*
15343 			 * We don't have that much in the SB, how much is
15344 			 * there?
15345 			 */
15346 			len = avail - sb_offset;
15347 		}
15348 	}
15349 	return (len);
15350 }
15351 
15352 static void
15353 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15354 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15355 	     int rsm_is_null, int optlen, int line, uint16_t mode)
15356 {
15357 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15358 		union tcp_log_stackspecific log;
15359 		struct timeval tv;
15360 
15361 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15362 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15363 		log.u_bbr.flex1 = error;
15364 		log.u_bbr.flex2 = flags;
15365 		log.u_bbr.flex3 = rsm_is_null;
15366 		log.u_bbr.flex4 = ipoptlen;
15367 		log.u_bbr.flex5 = tp->rcv_numsacks;
15368 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15369 		log.u_bbr.flex7 = optlen;
15370 		log.u_bbr.flex8 = rack->r_fsb_inited;
15371 		log.u_bbr.applimited = rack->r_fast_output;
15372 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15373 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15374 		log.u_bbr.cwnd_gain = mode;
15375 		log.u_bbr.pkts_out = orig_len;
15376 		log.u_bbr.lt_epoch = len;
15377 		log.u_bbr.delivered = line;
15378 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15379 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15380 		tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15381 			       len, &log, false, NULL, NULL, 0, &tv);
15382 	}
15383 }
15384 
15385 
15386 static struct mbuf *
15387 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15388 		   struct rack_fast_send_blk *fsb,
15389 		   int32_t seglimit, int32_t segsize, int hw_tls)
15390 {
15391 #ifdef KERN_TLS
15392 	struct ktls_session *tls, *ntls;
15393 #ifdef INVARIANTS
15394 	struct mbuf *start;
15395 #endif
15396 #endif
15397 	struct mbuf *m, *n, **np, *smb;
15398 	struct mbuf *top;
15399 	int32_t off, soff;
15400 	int32_t len = *plen;
15401 	int32_t fragsize;
15402 	int32_t len_cp = 0;
15403 	uint32_t mlen, frags;
15404 
15405 	soff = off = the_off;
15406 	smb = m = the_m;
15407 	np = &top;
15408 	top = NULL;
15409 #ifdef KERN_TLS
15410 	if (hw_tls && (m->m_flags & M_EXTPG))
15411 		tls = m->m_epg_tls;
15412 	else
15413 		tls = NULL;
15414 #ifdef INVARIANTS
15415 	start = m;
15416 #endif
15417 #endif
15418 	while (len > 0) {
15419 		if (m == NULL) {
15420 			*plen = len_cp;
15421 			break;
15422 		}
15423 #ifdef KERN_TLS
15424 		if (hw_tls) {
15425 			if (m->m_flags & M_EXTPG)
15426 				ntls = m->m_epg_tls;
15427 			else
15428 				ntls = NULL;
15429 
15430 			/*
15431 			 * Avoid mixing TLS records with handshake
15432 			 * data or TLS records from different
15433 			 * sessions.
15434 			 */
15435 			if (tls != ntls) {
15436 				MPASS(m != start);
15437 				*plen = len_cp;
15438 				break;
15439 			}
15440 		}
15441 #endif
15442 		mlen = min(len, m->m_len - off);
15443 		if (seglimit) {
15444 			/*
15445 			 * For M_EXTPG mbufs, add 3 segments
15446 			 * + 1 in case we are crossing page boundaries
15447 			 * + 2 in case the TLS hdr/trailer are used
15448 			 * It is cheaper to just add the segments
15449 			 * than it is to take the cache miss to look
15450 			 * at the mbuf ext_pgs state in detail.
15451 			 */
15452 			if (m->m_flags & M_EXTPG) {
15453 				fragsize = min(segsize, PAGE_SIZE);
15454 				frags = 3;
15455 			} else {
15456 				fragsize = segsize;
15457 				frags = 0;
15458 			}
15459 
15460 			/* Break if we really can't fit anymore. */
15461 			if ((frags + 1) >= seglimit) {
15462 				*plen =	len_cp;
15463 				break;
15464 			}
15465 
15466 			/*
15467 			 * Reduce size if you can't copy the whole
15468 			 * mbuf. If we can't copy the whole mbuf, also
15469 			 * adjust len so the loop will end after this
15470 			 * mbuf.
15471 			 */
15472 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15473 				mlen = (seglimit - frags - 1) * fragsize;
15474 				len = mlen;
15475 				*plen = len_cp + len;
15476 			}
15477 			frags += howmany(mlen, fragsize);
15478 			if (frags == 0)
15479 				frags++;
15480 			seglimit -= frags;
15481 			KASSERT(seglimit > 0,
15482 			    ("%s: seglimit went too low", __func__));
15483 		}
15484 		n = m_get(M_NOWAIT, m->m_type);
15485 		*np = n;
15486 		if (n == NULL)
15487 			goto nospace;
15488 		n->m_len = mlen;
15489 		soff += mlen;
15490 		len_cp += n->m_len;
15491 		if (m->m_flags & (M_EXT|M_EXTPG)) {
15492 			n->m_data = m->m_data + off;
15493 			mb_dupcl(n, m);
15494 		} else {
15495 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15496 			    (u_int)n->m_len);
15497 		}
15498 		len -= n->m_len;
15499 		off = 0;
15500 		m = m->m_next;
15501 		np = &n->m_next;
15502 		if (len || (soff == smb->m_len)) {
15503 			/*
15504 			 * We have more so we move forward  or
15505 			 * we have consumed the entire mbuf and
15506 			 * len has fell to 0.
15507 			 */
15508 			soff = 0;
15509 			smb = m;
15510 		}
15511 
15512 	}
15513 	if (fsb != NULL) {
15514 		fsb->m = smb;
15515 		fsb->off = soff;
15516 		if (smb) {
15517 			/*
15518 			 * Save off the size of the mbuf. We do
15519 			 * this so that we can recognize when it
15520 			 * has been trimmed by sbcut() as acks
15521 			 * come in.
15522 			 */
15523 			fsb->o_m_len = smb->m_len;
15524 		} else {
15525 			/*
15526 			 * This is the case where the next mbuf went to NULL. This
15527 			 * means with this copy we have sent everything in the sb.
15528 			 * In theory we could clear the fast_output flag, but lets
15529 			 * not since its possible that we could get more added
15530 			 * and acks that call the extend function which would let
15531 			 * us send more.
15532 			 */
15533 			fsb->o_m_len = 0;
15534 		}
15535 	}
15536 	return (top);
15537 nospace:
15538 	if (top)
15539 		m_freem(top);
15540 	return (NULL);
15541 
15542 }
15543 
15544 /*
15545  * This is a copy of m_copym(), taking the TSO segment size/limit
15546  * constraints into account, and advancing the sndptr as it goes.
15547  */
15548 static struct mbuf *
15549 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15550 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15551 {
15552 	struct mbuf *m, *n;
15553 	int32_t soff;
15554 
15555 	soff = rack->r_ctl.fsb.off;
15556 	m = rack->r_ctl.fsb.m;
15557 	if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15558 		/*
15559 		 * The mbuf had the front of it chopped off by an ack
15560 		 * we need to adjust the soff/off by that difference.
15561 		 */
15562 		uint32_t delta;
15563 
15564 		delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15565 		soff -= delta;
15566 	} else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15567 		/*
15568 		 * The mbuf was expanded probably by
15569 		 * a m_compress. Just update o_m_len.
15570 		 */
15571 		rack->r_ctl.fsb.o_m_len = m->m_len;
15572 	}
15573 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15574 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15575 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15576 				 __FUNCTION__,
15577 				 rack, *plen, m, m->m_len));
15578 	/* Save off the right location before we copy and advance */
15579 	*s_soff = soff;
15580 	*s_mb = rack->r_ctl.fsb.m;
15581 	n = rack_fo_base_copym(m, soff, plen,
15582 			       &rack->r_ctl.fsb,
15583 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15584 	return (n);
15585 }
15586 
15587 static int
15588 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15589 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15590 {
15591 	/*
15592 	 * Enter the fast retransmit path. We are given that a sched_pin is
15593 	 * in place (if accounting is compliled in) and the cycle count taken
15594 	 * at the entry is in the ts_val. The concept her is that the rsm
15595 	 * now holds the mbuf offsets and such so we can directly transmit
15596 	 * without a lot of overhead, the len field is already set for
15597 	 * us to prohibit us from sending too much (usually its 1MSS).
15598 	 */
15599 	struct ip *ip = NULL;
15600 	struct udphdr *udp = NULL;
15601 	struct tcphdr *th = NULL;
15602 	struct mbuf *m = NULL;
15603 	struct inpcb *inp;
15604 	uint8_t *cpto;
15605 	struct tcp_log_buffer *lgb;
15606 #ifdef TCP_ACCOUNTING
15607 	uint64_t crtsc;
15608 	int cnt_thru = 1;
15609 #endif
15610 	struct tcpopt to;
15611 	u_char opt[TCP_MAXOLEN];
15612 	uint32_t hdrlen, optlen;
15613 	int32_t slot, segsiz, max_val, tso = 0, error, ulen = 0;
15614 	uint16_t flags;
15615 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
15616 	uint32_t if_hw_tsomaxsegsize;
15617 
15618 #ifdef INET6
15619 	struct ip6_hdr *ip6 = NULL;
15620 
15621 	if (rack->r_is_v6) {
15622 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15623 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15624 	} else
15625 #endif				/* INET6 */
15626 	{
15627 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15628 		hdrlen = sizeof(struct tcpiphdr);
15629 	}
15630 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15631 		goto failed;
15632 	}
15633 	if (doing_tlp) {
15634 		/* Its a TLP add the flag, it may already be there but be sure */
15635 		rsm->r_flags |= RACK_TLP;
15636 	} else {
15637 		/* If it was a TLP it is not not on this retransmit */
15638 		rsm->r_flags &= ~RACK_TLP;
15639 	}
15640 	startseq = rsm->r_start;
15641 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15642 	inp = rack->rc_inp;
15643 	to.to_flags = 0;
15644 	flags = tcp_outflags[tp->t_state];
15645 	if (flags & (TH_SYN|TH_RST)) {
15646 		goto failed;
15647 	}
15648 	if (rsm->r_flags & RACK_HAS_FIN) {
15649 		/* We can't send a FIN here */
15650 		goto failed;
15651 	}
15652 	if (flags & TH_FIN) {
15653 		/* We never send a FIN */
15654 		flags &= ~TH_FIN;
15655 	}
15656 	if (tp->t_flags & TF_RCVD_TSTMP) {
15657 		to.to_tsval = ms_cts + tp->ts_offset;
15658 		to.to_tsecr = tp->ts_recent;
15659 		to.to_flags = TOF_TS;
15660 	}
15661 	optlen = tcp_addoptions(&to, opt);
15662 	hdrlen += optlen;
15663 	udp = rack->r_ctl.fsb.udp;
15664 	if (udp)
15665 		hdrlen += sizeof(struct udphdr);
15666 	if (rack->r_ctl.rc_pace_max_segs)
15667 		max_val = rack->r_ctl.rc_pace_max_segs;
15668 	else if (rack->rc_user_set_max_segs)
15669 		max_val = rack->rc_user_set_max_segs * segsiz;
15670 	else
15671 		max_val = len;
15672 	if ((tp->t_flags & TF_TSO) &&
15673 	    V_tcp_do_tso &&
15674 	    (len > segsiz) &&
15675 	    (tp->t_port == 0))
15676 		tso = 1;
15677 #ifdef INET6
15678 	if (MHLEN < hdrlen + max_linkhdr)
15679 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15680 	else
15681 #endif
15682 		m = m_gethdr(M_NOWAIT, MT_DATA);
15683 	if (m == NULL)
15684 		goto failed;
15685 	m->m_data += max_linkhdr;
15686 	m->m_len = hdrlen;
15687 	th = rack->r_ctl.fsb.th;
15688 	/* Establish the len to send */
15689 	if (len > max_val)
15690 		len = max_val;
15691 	if ((tso) && (len + optlen > tp->t_maxseg)) {
15692 		uint32_t if_hw_tsomax;
15693 		int32_t max_len;
15694 
15695 		/* extract TSO information */
15696 		if_hw_tsomax = tp->t_tsomax;
15697 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15698 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15699 		/*
15700 		 * Check if we should limit by maximum payload
15701 		 * length:
15702 		 */
15703 		if (if_hw_tsomax != 0) {
15704 			/* compute maximum TSO length */
15705 			max_len = (if_hw_tsomax - hdrlen -
15706 				   max_linkhdr);
15707 			if (max_len <= 0) {
15708 				goto failed;
15709 			} else if (len > max_len) {
15710 				len = max_len;
15711 			}
15712 		}
15713 		if (len <= segsiz) {
15714 			/*
15715 			 * In case there are too many small fragments don't
15716 			 * use TSO:
15717 			 */
15718 			tso = 0;
15719 		}
15720 	} else {
15721 		tso = 0;
15722 	}
15723 	if ((tso == 0) && (len > segsiz))
15724 		len = segsiz;
15725 	if ((len == 0) ||
15726 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
15727 		goto failed;
15728 	}
15729 	th->th_seq = htonl(rsm->r_start);
15730 	th->th_ack = htonl(tp->rcv_nxt);
15731 	/*
15732 	 * The PUSH bit should only be applied
15733 	 * if the full retransmission is made. If
15734 	 * we are sending less than this is the
15735 	 * left hand edge and should not have
15736 	 * the PUSH bit.
15737 	 */
15738 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
15739 	    (len == (rsm->r_end - rsm->r_start)))
15740 		flags |= TH_PUSH;
15741 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15742 	if (th->th_win == 0) {
15743 		tp->t_sndzerowin++;
15744 		tp->t_flags |= TF_RXWIN0SENT;
15745 	} else
15746 		tp->t_flags &= ~TF_RXWIN0SENT;
15747 	if (rsm->r_flags & RACK_TLP) {
15748 		/*
15749 		 * TLP should not count in retran count, but
15750 		 * in its own bin
15751 		 */
15752 		counter_u64_add(rack_tlp_retran, 1);
15753 		counter_u64_add(rack_tlp_retran_bytes, len);
15754 	} else {
15755 		tp->t_sndrexmitpack++;
15756 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15757 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15758 	}
15759 #ifdef STATS
15760 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15761 				 len);
15762 #endif
15763 	if (rsm->m == NULL)
15764 		goto failed;
15765 	if (rsm->orig_m_len != rsm->m->m_len) {
15766 		/* Fix up the orig_m_len and possibly the mbuf offset */
15767 		rack_adjust_orig_mlen(rsm);
15768 	}
15769 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15770 	if (len <= segsiz) {
15771 		/*
15772 		 * Must have ran out of mbufs for the copy
15773 		 * shorten it to no longer need tso. Lets
15774 		 * not put on sendalot since we are low on
15775 		 * mbufs.
15776 		 */
15777 		tso = 0;
15778 	}
15779 	if ((m->m_next == NULL) || (len <= 0)){
15780 		goto failed;
15781 	}
15782 	if (udp) {
15783 		if (rack->r_is_v6)
15784 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
15785 		else
15786 			ulen = hdrlen + len - sizeof(struct ip);
15787 		udp->uh_ulen = htons(ulen);
15788 	}
15789 	m->m_pkthdr.rcvif = (struct ifnet *)0;
15790 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
15791 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
15792 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
15793 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
15794 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
15795 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
15796 #ifdef INET6
15797 		if (rack->r_is_v6) {
15798 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
15799 		    ip6->ip6_flow |= htonl(ect << 20);
15800 		}
15801 		else
15802 #endif
15803 		{
15804 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
15805 		    ip->ip_tos |= ect;
15806 		}
15807 	}
15808 	tcp_set_flags(th, flags);
15809 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
15810 #ifdef INET6
15811 	if (rack->r_is_v6) {
15812 		if (tp->t_port) {
15813 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15814 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15815 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15816 			th->th_sum = htons(0);
15817 			UDPSTAT_INC(udps_opackets);
15818 		} else {
15819 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15820 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15821 			th->th_sum = in6_cksum_pseudo(ip6,
15822 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15823 						      0);
15824 		}
15825 	}
15826 #endif
15827 #if defined(INET6) && defined(INET)
15828 	else
15829 #endif
15830 #ifdef INET
15831 	{
15832 		if (tp->t_port) {
15833 			m->m_pkthdr.csum_flags = CSUM_UDP;
15834 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15835 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15836 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15837 			th->th_sum = htons(0);
15838 			UDPSTAT_INC(udps_opackets);
15839 		} else {
15840 			m->m_pkthdr.csum_flags = CSUM_TCP;
15841 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15842 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
15843 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15844 									IPPROTO_TCP + len + optlen));
15845 		}
15846 		/* IP version must be set here for ipv4/ipv6 checking later */
15847 		KASSERT(ip->ip_v == IPVERSION,
15848 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
15849 	}
15850 #endif
15851 	if (tso) {
15852 		KASSERT(len > tp->t_maxseg - optlen,
15853 			("%s: len <= tso_segsz tp:%p", __func__, tp));
15854 		m->m_pkthdr.csum_flags |= CSUM_TSO;
15855 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15856 	}
15857 #ifdef INET6
15858 	if (rack->r_is_v6) {
15859 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15860 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15861 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15862 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15863 		else
15864 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15865 	}
15866 #endif
15867 #if defined(INET) && defined(INET6)
15868 	else
15869 #endif
15870 #ifdef INET
15871 	{
15872 		ip->ip_len = htons(m->m_pkthdr.len);
15873 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15874 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15875 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15876 			if (tp->t_port == 0 || len < V_tcp_minmss) {
15877 				ip->ip_off |= htons(IP_DF);
15878 			}
15879 		} else {
15880 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15881 		}
15882 	}
15883 #endif
15884 	/* Time to copy in our header */
15885 	cpto = mtod(m, uint8_t *);
15886 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15887 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15888 	if (optlen) {
15889 		bcopy(opt, th + 1, optlen);
15890 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15891 	} else {
15892 		th->th_off = sizeof(struct tcphdr) >> 2;
15893 	}
15894 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15895 		union tcp_log_stackspecific log;
15896 
15897 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15898 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15899 		if (rack->rack_no_prr)
15900 			log.u_bbr.flex1 = 0;
15901 		else
15902 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15903 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15904 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15905 		log.u_bbr.flex4 = max_val;
15906 		log.u_bbr.flex5 = 0;
15907 		/* Save off the early/late values */
15908 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15909 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15910 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15911 		if (doing_tlp == 0)
15912 			log.u_bbr.flex8 = 1;
15913 		else
15914 			log.u_bbr.flex8 = 2;
15915 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15916 		log.u_bbr.flex7 = 55;
15917 		log.u_bbr.pkts_out = tp->t_maxseg;
15918 		log.u_bbr.timeStamp = cts;
15919 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15920 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15921 		log.u_bbr.delivered = 0;
15922 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15923 				     len, &log, false, NULL, NULL, 0, tv);
15924 	} else
15925 		lgb = NULL;
15926 #ifdef INET6
15927 	if (rack->r_is_v6) {
15928 		error = ip6_output(m, NULL,
15929 				   &inp->inp_route6,
15930 				   0, NULL, NULL, inp);
15931 	}
15932 #endif
15933 #if defined(INET) && defined(INET6)
15934 	else
15935 #endif
15936 #ifdef INET
15937 	{
15938 		error = ip_output(m, NULL,
15939 				  &inp->inp_route,
15940 				  0, 0, inp);
15941 	}
15942 #endif
15943 	m = NULL;
15944 	if (lgb) {
15945 		lgb->tlb_errno = error;
15946 		lgb = NULL;
15947 	}
15948 	if (error) {
15949 		goto failed;
15950 	}
15951 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15952 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
15953 	if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15954 		rack->rc_tlp_in_progress = 1;
15955 		rack->r_ctl.rc_tlp_cnt_out++;
15956 	}
15957 	if (error == 0) {
15958 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
15959 		if (doing_tlp) {
15960 			rack->rc_last_sent_tlp_past_cumack = 0;
15961 			rack->rc_last_sent_tlp_seq_valid = 1;
15962 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
15963 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
15964 		}
15965 	}
15966 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15967 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
15968 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15969 		rack->r_ctl.retran_during_recovery += len;
15970 	{
15971 		int idx;
15972 
15973 		idx = (len / segsiz) + 3;
15974 		if (idx >= TCP_MSS_ACCT_ATIMER)
15975 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15976 		else
15977 			counter_u64_add(rack_out_size[idx], 1);
15978 	}
15979 	if (tp->t_rtttime == 0) {
15980 		tp->t_rtttime = ticks;
15981 		tp->t_rtseq = startseq;
15982 		KMOD_TCPSTAT_INC(tcps_segstimed);
15983 	}
15984 	counter_u64_add(rack_fto_rsm_send, 1);
15985 	if (error && (error == ENOBUFS)) {
15986 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15987 		if (rack->rc_enobuf < 0x7f)
15988 			rack->rc_enobuf++;
15989 		if (slot < (10 * HPTS_USEC_IN_MSEC))
15990 			slot = 10 * HPTS_USEC_IN_MSEC;
15991 	} else
15992 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15993 	if ((slot == 0) ||
15994 	    (rack->rc_always_pace == 0) ||
15995 	    (rack->r_rr_config == 1)) {
15996 		/*
15997 		 * We have no pacing set or we
15998 		 * are using old-style rack or
15999 		 * we are overriden to use the old 1ms pacing.
16000 		 */
16001 		slot = rack->r_ctl.rc_min_to;
16002 	}
16003 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
16004 #ifdef TCP_ACCOUNTING
16005 	crtsc = get_cyclecount();
16006 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16007 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16008 	}
16009 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16010 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16011 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16012 	}
16013 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16014 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16015 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
16016 	}
16017 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
16018 	sched_unpin();
16019 #endif
16020 	return (0);
16021 failed:
16022 	if (m)
16023 		m_free(m);
16024 	return (-1);
16025 }
16026 
16027 static void
16028 rack_sndbuf_autoscale(struct tcp_rack *rack)
16029 {
16030 	/*
16031 	 * Automatic sizing of send socket buffer.  Often the send buffer
16032 	 * size is not optimally adjusted to the actual network conditions
16033 	 * at hand (delay bandwidth product).  Setting the buffer size too
16034 	 * small limits throughput on links with high bandwidth and high
16035 	 * delay (eg. trans-continental/oceanic links).  Setting the
16036 	 * buffer size too big consumes too much real kernel memory,
16037 	 * especially with many connections on busy servers.
16038 	 *
16039 	 * The criteria to step up the send buffer one notch are:
16040 	 *  1. receive window of remote host is larger than send buffer
16041 	 *     (with a fudge factor of 5/4th);
16042 	 *  2. send buffer is filled to 7/8th with data (so we actually
16043 	 *     have data to make use of it);
16044 	 *  3. send buffer fill has not hit maximal automatic size;
16045 	 *  4. our send window (slow start and cogestion controlled) is
16046 	 *     larger than sent but unacknowledged data in send buffer.
16047 	 *
16048 	 * Note that the rack version moves things much faster since
16049 	 * we want to avoid hitting cache lines in the rack_fast_output()
16050 	 * path so this is called much less often and thus moves
16051 	 * the SB forward by a percentage.
16052 	 */
16053 	struct socket *so;
16054 	struct tcpcb *tp;
16055 	uint32_t sendwin, scaleup;
16056 
16057 	tp = rack->rc_tp;
16058 	so = rack->rc_inp->inp_socket;
16059 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16060 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16061 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16062 		    sbused(&so->so_snd) >=
16063 		    (so->so_snd.sb_hiwat / 8 * 7) &&
16064 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16065 		    sendwin >= (sbused(&so->so_snd) -
16066 		    (tp->snd_nxt - tp->snd_una))) {
16067 			if (rack_autosndbuf_inc)
16068 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16069 			else
16070 				scaleup = V_tcp_autosndbuf_inc;
16071 			if (scaleup < V_tcp_autosndbuf_inc)
16072 				scaleup = V_tcp_autosndbuf_inc;
16073 			scaleup += so->so_snd.sb_hiwat;
16074 			if (scaleup > V_tcp_autosndbuf_max)
16075 				scaleup = V_tcp_autosndbuf_max;
16076 			if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
16077 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16078 		}
16079 	}
16080 }
16081 
16082 static int
16083 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16084 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16085 {
16086 	/*
16087 	 * Enter to do fast output. We are given that the sched_pin is
16088 	 * in place (if accounting is compiled in) and the cycle count taken
16089 	 * at entry is in place in ts_val. The idea here is that
16090 	 * we know how many more bytes needs to be sent (presumably either
16091 	 * during pacing or to fill the cwnd and that was greater than
16092 	 * the max-burst). We have how much to send and all the info we
16093 	 * need to just send.
16094 	 */
16095 	struct ip *ip = NULL;
16096 	struct udphdr *udp = NULL;
16097 	struct tcphdr *th = NULL;
16098 	struct mbuf *m, *s_mb;
16099 	struct inpcb *inp;
16100 	uint8_t *cpto;
16101 	struct tcp_log_buffer *lgb;
16102 #ifdef TCP_ACCOUNTING
16103 	uint64_t crtsc;
16104 #endif
16105 	struct tcpopt to;
16106 	u_char opt[TCP_MAXOLEN];
16107 	uint32_t hdrlen, optlen;
16108 	int cnt_thru = 1;
16109 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
16110 	uint16_t flags;
16111 	uint32_t s_soff;
16112 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
16113 	uint32_t if_hw_tsomaxsegsize;
16114 	uint16_t add_flag = RACK_SENT_FP;
16115 #ifdef INET6
16116 	struct ip6_hdr *ip6 = NULL;
16117 
16118 	if (rack->r_is_v6) {
16119 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16120 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16121 	} else
16122 #endif				/* INET6 */
16123 	{
16124 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16125 		hdrlen = sizeof(struct tcpiphdr);
16126 	}
16127 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16128 		m = NULL;
16129 		goto failed;
16130 	}
16131 	startseq = tp->snd_max;
16132 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16133 	inp = rack->rc_inp;
16134 	len = rack->r_ctl.fsb.left_to_send;
16135 	to.to_flags = 0;
16136 	flags = rack->r_ctl.fsb.tcp_flags;
16137 	if (tp->t_flags & TF_RCVD_TSTMP) {
16138 		to.to_tsval = ms_cts + tp->ts_offset;
16139 		to.to_tsecr = tp->ts_recent;
16140 		to.to_flags = TOF_TS;
16141 	}
16142 	optlen = tcp_addoptions(&to, opt);
16143 	hdrlen += optlen;
16144 	udp = rack->r_ctl.fsb.udp;
16145 	if (udp)
16146 		hdrlen += sizeof(struct udphdr);
16147 	if (rack->r_ctl.rc_pace_max_segs)
16148 		max_val = rack->r_ctl.rc_pace_max_segs;
16149 	else if (rack->rc_user_set_max_segs)
16150 		max_val = rack->rc_user_set_max_segs * segsiz;
16151 	else
16152 		max_val = len;
16153 	if ((tp->t_flags & TF_TSO) &&
16154 	    V_tcp_do_tso &&
16155 	    (len > segsiz) &&
16156 	    (tp->t_port == 0))
16157 		tso = 1;
16158 again:
16159 #ifdef INET6
16160 	if (MHLEN < hdrlen + max_linkhdr)
16161 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16162 	else
16163 #endif
16164 		m = m_gethdr(M_NOWAIT, MT_DATA);
16165 	if (m == NULL)
16166 		goto failed;
16167 	m->m_data += max_linkhdr;
16168 	m->m_len = hdrlen;
16169 	th = rack->r_ctl.fsb.th;
16170 	/* Establish the len to send */
16171 	if (len > max_val)
16172 		len = max_val;
16173 	if ((tso) && (len + optlen > tp->t_maxseg)) {
16174 		uint32_t if_hw_tsomax;
16175 		int32_t max_len;
16176 
16177 		/* extract TSO information */
16178 		if_hw_tsomax = tp->t_tsomax;
16179 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16180 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16181 		/*
16182 		 * Check if we should limit by maximum payload
16183 		 * length:
16184 		 */
16185 		if (if_hw_tsomax != 0) {
16186 			/* compute maximum TSO length */
16187 			max_len = (if_hw_tsomax - hdrlen -
16188 				   max_linkhdr);
16189 			if (max_len <= 0) {
16190 				goto failed;
16191 			} else if (len > max_len) {
16192 				len = max_len;
16193 			}
16194 		}
16195 		if (len <= segsiz) {
16196 			/*
16197 			 * In case there are too many small fragments don't
16198 			 * use TSO:
16199 			 */
16200 			tso = 0;
16201 		}
16202 	} else {
16203 		tso = 0;
16204 	}
16205 	if ((tso == 0) && (len > segsiz))
16206 		len = segsiz;
16207 	if ((len == 0) ||
16208 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
16209 		goto failed;
16210 	}
16211 	sb_offset = tp->snd_max - tp->snd_una;
16212 	th->th_seq = htonl(tp->snd_max);
16213 	th->th_ack = htonl(tp->rcv_nxt);
16214 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16215 	if (th->th_win == 0) {
16216 		tp->t_sndzerowin++;
16217 		tp->t_flags |= TF_RXWIN0SENT;
16218 	} else
16219 		tp->t_flags &= ~TF_RXWIN0SENT;
16220 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
16221 	KMOD_TCPSTAT_INC(tcps_sndpack);
16222 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16223 #ifdef STATS
16224 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16225 				 len);
16226 #endif
16227 	if (rack->r_ctl.fsb.m == NULL)
16228 		goto failed;
16229 
16230 	/* s_mb and s_soff are saved for rack_log_output */
16231 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16232 				    &s_mb, &s_soff);
16233 	if (len <= segsiz) {
16234 		/*
16235 		 * Must have ran out of mbufs for the copy
16236 		 * shorten it to no longer need tso. Lets
16237 		 * not put on sendalot since we are low on
16238 		 * mbufs.
16239 		 */
16240 		tso = 0;
16241 	}
16242 	if (rack->r_ctl.fsb.rfo_apply_push &&
16243 	    (len == rack->r_ctl.fsb.left_to_send)) {
16244 		flags |= TH_PUSH;
16245 		add_flag |= RACK_HAD_PUSH;
16246 	}
16247 	if ((m->m_next == NULL) || (len <= 0)){
16248 		goto failed;
16249 	}
16250 	if (udp) {
16251 		if (rack->r_is_v6)
16252 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
16253 		else
16254 			ulen = hdrlen + len - sizeof(struct ip);
16255 		udp->uh_ulen = htons(ulen);
16256 	}
16257 	m->m_pkthdr.rcvif = (struct ifnet *)0;
16258 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
16259 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
16260 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
16261 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
16262 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
16263 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
16264 #ifdef INET6
16265 		if (rack->r_is_v6) {
16266 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
16267 			ip6->ip6_flow |= htonl(ect << 20);
16268 		}
16269 		else
16270 #endif
16271 		{
16272 			ip->ip_tos &= ~IPTOS_ECN_MASK;
16273 			ip->ip_tos |= ect;
16274 		}
16275 	}
16276 	tcp_set_flags(th, flags);
16277 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
16278 #ifdef INET6
16279 	if (rack->r_is_v6) {
16280 		if (tp->t_port) {
16281 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16282 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16283 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16284 			th->th_sum = htons(0);
16285 			UDPSTAT_INC(udps_opackets);
16286 		} else {
16287 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16288 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16289 			th->th_sum = in6_cksum_pseudo(ip6,
16290 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16291 						      0);
16292 		}
16293 	}
16294 #endif
16295 #if defined(INET6) && defined(INET)
16296 	else
16297 #endif
16298 #ifdef INET
16299 	{
16300 		if (tp->t_port) {
16301 			m->m_pkthdr.csum_flags = CSUM_UDP;
16302 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16303 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16304 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16305 			th->th_sum = htons(0);
16306 			UDPSTAT_INC(udps_opackets);
16307 		} else {
16308 			m->m_pkthdr.csum_flags = CSUM_TCP;
16309 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16310 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
16311 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16312 									IPPROTO_TCP + len + optlen));
16313 		}
16314 		/* IP version must be set here for ipv4/ipv6 checking later */
16315 		KASSERT(ip->ip_v == IPVERSION,
16316 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
16317 	}
16318 #endif
16319 	if (tso) {
16320 		KASSERT(len > tp->t_maxseg - optlen,
16321 			("%s: len <= tso_segsz tp:%p", __func__, tp));
16322 		m->m_pkthdr.csum_flags |= CSUM_TSO;
16323 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16324 	}
16325 #ifdef INET6
16326 	if (rack->r_is_v6) {
16327 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16328 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16329 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16330 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16331 		else
16332 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16333 	}
16334 #endif
16335 #if defined(INET) && defined(INET6)
16336 	else
16337 #endif
16338 #ifdef INET
16339 	{
16340 		ip->ip_len = htons(m->m_pkthdr.len);
16341 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16342 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16343 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16344 			if (tp->t_port == 0 || len < V_tcp_minmss) {
16345 				ip->ip_off |= htons(IP_DF);
16346 			}
16347 		} else {
16348 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16349 		}
16350 	}
16351 #endif
16352 	/* Time to copy in our header */
16353 	cpto = mtod(m, uint8_t *);
16354 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16355 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16356 	if (optlen) {
16357 		bcopy(opt, th + 1, optlen);
16358 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16359 	} else {
16360 		th->th_off = sizeof(struct tcphdr) >> 2;
16361 	}
16362 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16363 		union tcp_log_stackspecific log;
16364 
16365 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16366 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16367 		if (rack->rack_no_prr)
16368 			log.u_bbr.flex1 = 0;
16369 		else
16370 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16371 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16372 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16373 		log.u_bbr.flex4 = max_val;
16374 		log.u_bbr.flex5 = 0;
16375 		/* Save off the early/late values */
16376 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16377 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16378 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16379 		log.u_bbr.flex8 = 0;
16380 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16381 		log.u_bbr.flex7 = 44;
16382 		log.u_bbr.pkts_out = tp->t_maxseg;
16383 		log.u_bbr.timeStamp = cts;
16384 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16385 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16386 		log.u_bbr.delivered = 0;
16387 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16388 				     len, &log, false, NULL, NULL, 0, tv);
16389 	} else
16390 		lgb = NULL;
16391 #ifdef INET6
16392 	if (rack->r_is_v6) {
16393 		error = ip6_output(m, NULL,
16394 				   &inp->inp_route6,
16395 				   0, NULL, NULL, inp);
16396 	}
16397 #endif
16398 #if defined(INET) && defined(INET6)
16399 	else
16400 #endif
16401 #ifdef INET
16402 	{
16403 		error = ip_output(m, NULL,
16404 				  &inp->inp_route,
16405 				  0, 0, inp);
16406 	}
16407 #endif
16408 	if (lgb) {
16409 		lgb->tlb_errno = error;
16410 		lgb = NULL;
16411 	}
16412 	if (error) {
16413 		*send_err = error;
16414 		m = NULL;
16415 		goto failed;
16416 	}
16417 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16418 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16419 	m = NULL;
16420 	if (tp->snd_una == tp->snd_max) {
16421 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
16422 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16423 		tp->t_acktime = ticks;
16424 	}
16425 	if (error == 0)
16426 		tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16427 
16428 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16429 	tot_len += len;
16430 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
16431 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16432 	tp->snd_max += len;
16433 	tp->snd_nxt = tp->snd_max;
16434 	{
16435 		int idx;
16436 
16437 		idx = (len / segsiz) + 3;
16438 		if (idx >= TCP_MSS_ACCT_ATIMER)
16439 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16440 		else
16441 			counter_u64_add(rack_out_size[idx], 1);
16442 	}
16443 	if (len <= rack->r_ctl.fsb.left_to_send)
16444 		rack->r_ctl.fsb.left_to_send -= len;
16445 	else
16446 		rack->r_ctl.fsb.left_to_send = 0;
16447 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
16448 		rack->r_fast_output = 0;
16449 		rack->r_ctl.fsb.left_to_send = 0;
16450 		/* At the end of fast_output scale up the sb */
16451 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16452 		rack_sndbuf_autoscale(rack);
16453 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16454 	}
16455 	if (tp->t_rtttime == 0) {
16456 		tp->t_rtttime = ticks;
16457 		tp->t_rtseq = startseq;
16458 		KMOD_TCPSTAT_INC(tcps_segstimed);
16459 	}
16460 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16461 	    (max_val > len) &&
16462 	    (tso == 0)) {
16463 		max_val -= len;
16464 		len = segsiz;
16465 		th = rack->r_ctl.fsb.th;
16466 		cnt_thru++;
16467 		goto again;
16468 	}
16469 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16470 	counter_u64_add(rack_fto_send, 1);
16471 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16472 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16473 #ifdef TCP_ACCOUNTING
16474 	crtsc = get_cyclecount();
16475 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16476 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16477 	}
16478 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16479 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16480 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16481 	}
16482 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16483 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16484 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16485 	}
16486 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16487 	sched_unpin();
16488 #endif
16489 	return (0);
16490 failed:
16491 	if (m)
16492 		m_free(m);
16493 	rack->r_fast_output = 0;
16494 	return (-1);
16495 }
16496 
16497 static int
16498 rack_output(struct tcpcb *tp)
16499 {
16500 	struct socket *so;
16501 	uint32_t recwin;
16502 	uint32_t sb_offset, s_moff = 0;
16503 	int32_t len, error = 0;
16504 	uint16_t flags;
16505 	struct mbuf *m, *s_mb = NULL;
16506 	struct mbuf *mb;
16507 	uint32_t if_hw_tsomaxsegcount = 0;
16508 	uint32_t if_hw_tsomaxsegsize;
16509 	int32_t segsiz, minseg;
16510 	long tot_len_this_send = 0;
16511 #ifdef INET
16512 	struct ip *ip = NULL;
16513 #endif
16514 #ifdef TCPDEBUG
16515 	struct ipovly *ipov = NULL;
16516 #endif
16517 	struct udphdr *udp = NULL;
16518 	struct tcp_rack *rack;
16519 	struct tcphdr *th;
16520 	uint8_t pass = 0;
16521 	uint8_t mark = 0;
16522 	uint8_t wanted_cookie = 0;
16523 	u_char opt[TCP_MAXOLEN];
16524 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
16525 	uint32_t rack_seq;
16526 
16527 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16528 	unsigned ipsec_optlen = 0;
16529 
16530 #endif
16531 	int32_t idle, sendalot;
16532 	int32_t sub_from_prr = 0;
16533 	volatile int32_t sack_rxmit;
16534 	struct rack_sendmap *rsm = NULL;
16535 	int32_t tso, mtu;
16536 	struct tcpopt to;
16537 	int32_t slot = 0;
16538 	int32_t sup_rack = 0;
16539 	uint32_t cts, ms_cts, delayed, early;
16540 	uint16_t add_flag = RACK_SENT_SP;
16541 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16542 	uint8_t hpts_calling,  doing_tlp = 0;
16543 	uint32_t cwnd_to_use, pace_max_seg;
16544 	int32_t do_a_prefetch = 0;
16545 	int32_t prefetch_rsm = 0;
16546 	int32_t orig_len = 0;
16547 	struct timeval tv;
16548 	int32_t prefetch_so_done = 0;
16549 	struct tcp_log_buffer *lgb;
16550 	struct inpcb *inp;
16551 	struct sockbuf *sb;
16552 	uint64_t ts_val = 0;
16553 #ifdef TCP_ACCOUNTING
16554 	uint64_t crtsc;
16555 #endif
16556 #ifdef INET6
16557 	struct ip6_hdr *ip6 = NULL;
16558 	int32_t isipv6;
16559 #endif
16560 	uint8_t filled_all = 0;
16561 	bool hw_tls = false;
16562 
16563 	/* setup and take the cache hits here */
16564 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16565 #ifdef TCP_ACCOUNTING
16566 	sched_pin();
16567 	ts_val = get_cyclecount();
16568 #endif
16569 	hpts_calling = rack->rc_inp->inp_hpts_calls;
16570 	NET_EPOCH_ASSERT();
16571 	INP_WLOCK_ASSERT(rack->rc_inp);
16572 #ifdef TCP_OFFLOAD
16573 	if (tp->t_flags & TF_TOE) {
16574 #ifdef TCP_ACCOUNTING
16575 		sched_unpin();
16576 #endif
16577 		return (tcp_offload_output(tp));
16578 	}
16579 #endif
16580 	/*
16581 	 * For TFO connections in SYN_RECEIVED, only allow the initial
16582 	 * SYN|ACK and those sent by the retransmit timer.
16583 	 */
16584 	if (IS_FASTOPEN(tp->t_flags) &&
16585 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
16586 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16587 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16588 #ifdef TCP_ACCOUNTING
16589 		sched_unpin();
16590 #endif
16591 		return (0);
16592 	}
16593 #ifdef INET6
16594 	if (rack->r_state) {
16595 		/* Use the cache line loaded if possible */
16596 		isipv6 = rack->r_is_v6;
16597 	} else {
16598 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16599 	}
16600 #endif
16601 	early = 0;
16602 	cts = tcp_get_usecs(&tv);
16603 	ms_cts = tcp_tv_to_mssectick(&tv);
16604 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16605 	    tcp_in_hpts(rack->rc_inp)) {
16606 		/*
16607 		 * We are on the hpts for some timer but not hptsi output.
16608 		 * Remove from the hpts unconditionally.
16609 		 */
16610 		rack_timer_cancel(tp, rack, cts, __LINE__);
16611 	}
16612 	/* Are we pacing and late? */
16613 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16614 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16615 		/* We are delayed */
16616 		delayed = cts - rack->r_ctl.rc_last_output_to;
16617 	} else {
16618 		delayed = 0;
16619 	}
16620 	/* Do the timers, which may override the pacer */
16621 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16622 		int retval;
16623 
16624 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
16625 		    &doing_tlp);
16626 		if (retval != 0) {
16627 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16628 #ifdef TCP_ACCOUNTING
16629 			sched_unpin();
16630 #endif
16631 			/*
16632 			 * If timers want tcp_drop(), then pass error out,
16633 			 * otherwise suppress it.
16634 			 */
16635 			return (retval < 0 ? retval : 0);
16636 		}
16637 	}
16638 	if (rack->rc_in_persist) {
16639 		if (tcp_in_hpts(rack->rc_inp) == 0) {
16640 			/* Timer is not running */
16641 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16642 		}
16643 #ifdef TCP_ACCOUNTING
16644 		sched_unpin();
16645 #endif
16646 		return (0);
16647 	}
16648 	if ((rack->r_timer_override) ||
16649 	    (rack->rc_ack_can_sendout_data) ||
16650 	    (delayed) ||
16651 	    (tp->t_state < TCPS_ESTABLISHED)) {
16652 		rack->rc_ack_can_sendout_data = 0;
16653 		if (tcp_in_hpts(rack->rc_inp))
16654 			tcp_hpts_remove(rack->rc_inp);
16655 	} else if (tcp_in_hpts(rack->rc_inp)) {
16656 		/*
16657 		 * On the hpts you can't pass even if ACKNOW is on, we will
16658 		 * when the hpts fires.
16659 		 */
16660 #ifdef TCP_ACCOUNTING
16661 		crtsc = get_cyclecount();
16662 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16663 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16664 		}
16665 		counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16666 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16667 			tp->tcp_cnt_counters[SND_BLOCKED]++;
16668 		}
16669 		counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16670 		sched_unpin();
16671 #endif
16672 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16673 		return (0);
16674 	}
16675 	rack->rc_inp->inp_hpts_calls = 0;
16676 	/* Finish out both pacing early and late accounting */
16677 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16678 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16679 		early = rack->r_ctl.rc_last_output_to - cts;
16680 	} else
16681 		early = 0;
16682 	if (delayed) {
16683 		rack->r_ctl.rc_agg_delayed += delayed;
16684 		rack->r_late = 1;
16685 	} else if (early) {
16686 		rack->r_ctl.rc_agg_early += early;
16687 		rack->r_early = 1;
16688 	}
16689 	/* Now that early/late accounting is done turn off the flag */
16690 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16691 	rack->r_wanted_output = 0;
16692 	rack->r_timer_override = 0;
16693 	if ((tp->t_state != rack->r_state) &&
16694 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
16695 		rack_set_state(tp, rack);
16696 	}
16697 	if ((rack->r_fast_output) &&
16698 	    (doing_tlp == 0) &&
16699 	    (tp->rcv_numsacks == 0)) {
16700 		int ret;
16701 
16702 		error = 0;
16703 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16704 		if (ret >= 0)
16705 			return(ret);
16706 		else if (error) {
16707 			inp = rack->rc_inp;
16708 			so = inp->inp_socket;
16709 			sb = &so->so_snd;
16710 			goto nomore;
16711 		}
16712 	}
16713 	inp = rack->rc_inp;
16714 	/*
16715 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
16716 	 * only allow the initial SYN or SYN|ACK and those sent
16717 	 * by the retransmit timer.
16718 	 */
16719 	if (IS_FASTOPEN(tp->t_flags) &&
16720 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
16721 	     (tp->t_state == TCPS_SYN_SENT)) &&
16722 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16723 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
16724 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16725 		so = inp->inp_socket;
16726 		sb = &so->so_snd;
16727 		goto just_return_nolock;
16728 	}
16729 	/*
16730 	 * Determine length of data that should be transmitted, and flags
16731 	 * that will be used. If there is some data or critical controls
16732 	 * (SYN, RST) to send, then transmit; otherwise, investigate
16733 	 * further.
16734 	 */
16735 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16736 	if (tp->t_idle_reduce) {
16737 		if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16738 			rack_cc_after_idle(rack, tp);
16739 	}
16740 	tp->t_flags &= ~TF_LASTIDLE;
16741 	if (idle) {
16742 		if (tp->t_flags & TF_MORETOCOME) {
16743 			tp->t_flags |= TF_LASTIDLE;
16744 			idle = 0;
16745 		}
16746 	}
16747 	if ((tp->snd_una == tp->snd_max) &&
16748 	    rack->r_ctl.rc_went_idle_time &&
16749 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16750 		idle = cts - rack->r_ctl.rc_went_idle_time;
16751 		if (idle > rack_min_probertt_hold) {
16752 			/* Count as a probe rtt */
16753 			if (rack->in_probe_rtt == 0) {
16754 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
16755 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16756 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16757 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16758 			} else {
16759 				rack_exit_probertt(rack, cts);
16760 			}
16761 		}
16762 		idle = 0;
16763 	}
16764 	if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16765 		rack_init_fsb_block(tp, rack);
16766 again:
16767 	/*
16768 	 * If we've recently taken a timeout, snd_max will be greater than
16769 	 * snd_nxt.  There may be SACK information that allows us to avoid
16770 	 * resending already delivered data.  Adjust snd_nxt accordingly.
16771 	 */
16772 	sendalot = 0;
16773 	cts = tcp_get_usecs(&tv);
16774 	ms_cts = tcp_tv_to_mssectick(&tv);
16775 	tso = 0;
16776 	mtu = 0;
16777 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16778 	minseg = segsiz;
16779 	if (rack->r_ctl.rc_pace_max_segs == 0)
16780 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16781 	else
16782 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16783 	sb_offset = tp->snd_max - tp->snd_una;
16784 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16785 	flags = tcp_outflags[tp->t_state];
16786 	while (rack->rc_free_cnt < rack_free_cache) {
16787 		rsm = rack_alloc(rack);
16788 		if (rsm == NULL) {
16789 			if (inp->inp_hpts_calls)
16790 				/* Retry in a ms */
16791 				slot = (1 * HPTS_USEC_IN_MSEC);
16792 			so = inp->inp_socket;
16793 			sb = &so->so_snd;
16794 			goto just_return_nolock;
16795 		}
16796 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16797 		rack->rc_free_cnt++;
16798 		rsm = NULL;
16799 	}
16800 	if (inp->inp_hpts_calls)
16801 		inp->inp_hpts_calls = 0;
16802 	sack_rxmit = 0;
16803 	len = 0;
16804 	rsm = NULL;
16805 	if (flags & TH_RST) {
16806 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16807 		so = inp->inp_socket;
16808 		sb = &so->so_snd;
16809 		goto send;
16810 	}
16811 	if (rack->r_ctl.rc_resend) {
16812 		/* Retransmit timer */
16813 		rsm = rack->r_ctl.rc_resend;
16814 		rack->r_ctl.rc_resend = NULL;
16815 		len = rsm->r_end - rsm->r_start;
16816 		sack_rxmit = 1;
16817 		sendalot = 0;
16818 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16819 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16820 			 __func__, __LINE__,
16821 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16822 		sb_offset = rsm->r_start - tp->snd_una;
16823 		if (len >= segsiz)
16824 			len = segsiz;
16825 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16826 		/* We have a retransmit that takes precedence */
16827 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16828 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
16829 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16830 			/* Enter recovery if not induced by a time-out */
16831 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
16832 		}
16833 #ifdef INVARIANTS
16834 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16835 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16836 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
16837 		}
16838 #endif
16839 		len = rsm->r_end - rsm->r_start;
16840 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16841 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16842 			 __func__, __LINE__,
16843 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16844 		sb_offset = rsm->r_start - tp->snd_una;
16845 		sendalot = 0;
16846 		if (len >= segsiz)
16847 			len = segsiz;
16848 		if (len > 0) {
16849 			sack_rxmit = 1;
16850 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16851 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16852 			    min(len, segsiz));
16853 		}
16854 	} else if (rack->r_ctl.rc_tlpsend) {
16855 		/* Tail loss probe */
16856 		long cwin;
16857 		long tlen;
16858 
16859 		/*
16860 		 * Check if we can do a TLP with a RACK'd packet
16861 		 * this can happen if we are not doing the rack
16862 		 * cheat and we skipped to a TLP and it
16863 		 * went off.
16864 		 */
16865 		rsm = rack->r_ctl.rc_tlpsend;
16866 		/* We are doing a TLP make sure the flag is preent */
16867 		rsm->r_flags |= RACK_TLP;
16868 		rack->r_ctl.rc_tlpsend = NULL;
16869 		sack_rxmit = 1;
16870 		tlen = rsm->r_end - rsm->r_start;
16871 		if (tlen > segsiz)
16872 			tlen = segsiz;
16873 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16874 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16875 			 __func__, __LINE__,
16876 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16877 		sb_offset = rsm->r_start - tp->snd_una;
16878 		cwin = min(tp->snd_wnd, tlen);
16879 		len = cwin;
16880 	}
16881 	if (rack->r_must_retran &&
16882 	    (doing_tlp == 0) &&
16883 	    (rsm == NULL)) {
16884 		/*
16885 		 * Non-Sack and we had a RTO or Sack/non-Sack and a
16886 		 * MTU change, we need to retransmit until we reach
16887 		 * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
16888 		 */
16889 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
16890 			int sendwin, flight;
16891 
16892 			sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16893 			flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16894 			if (flight >= sendwin) {
16895 				so = inp->inp_socket;
16896 				sb = &so->so_snd;
16897 				goto just_return_nolock;
16898 			}
16899 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16900 			if (rsm == NULL) {
16901 				/* TSNH */
16902 				rack->r_must_retran = 0;
16903 				rack->r_ctl.rc_out_at_rto = 0;
16904 				so = inp->inp_socket;
16905 				sb = &so->so_snd;
16906 				goto just_return_nolock;
16907 			}
16908 			if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
16909 				/* It does not have the flag, we are done */
16910 				rack->r_must_retran = 0;
16911 				rack->r_ctl.rc_out_at_rto = 0;
16912 			} else {
16913 				sack_rxmit = 1;
16914 				len = rsm->r_end - rsm->r_start;
16915 				sendalot = 0;
16916 				sb_offset = rsm->r_start - tp->snd_una;
16917 				if (len >= segsiz)
16918 					len = segsiz;
16919 				/*
16920 				 * Delay removing the flag RACK_MUST_RXT so
16921 				 * that the fastpath for retransmit will
16922 				 * work with this rsm.
16923 				 */
16924 
16925 			}
16926 		} else {
16927 			/* We must be done if there is nothing outstanding */
16928 			rack->r_must_retran = 0;
16929 			rack->r_ctl.rc_out_at_rto = 0;
16930 		}
16931 	}
16932 	/*
16933 	 * Enforce a connection sendmap count limit if set
16934 	 * as long as we are not retransmiting.
16935 	 */
16936 	if ((rsm == NULL) &&
16937 	    (rack->do_detection == 0) &&
16938 	    (V_tcp_map_entries_limit > 0) &&
16939 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16940 		counter_u64_add(rack_to_alloc_limited, 1);
16941 		if (!rack->alloc_limit_reported) {
16942 			rack->alloc_limit_reported = 1;
16943 			counter_u64_add(rack_alloc_limited_conns, 1);
16944 		}
16945 		so = inp->inp_socket;
16946 		sb = &so->so_snd;
16947 		goto just_return_nolock;
16948 	}
16949 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16950 		/* we are retransmitting the fin */
16951 		len--;
16952 		if (len) {
16953 			/*
16954 			 * When retransmitting data do *not* include the
16955 			 * FIN. This could happen from a TLP probe.
16956 			 */
16957 			flags &= ~TH_FIN;
16958 		}
16959 	}
16960 	if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16961 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16962 		int ret;
16963 
16964 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
16965 		if (ret == 0)
16966 			return (0);
16967 	}
16968 	so = inp->inp_socket;
16969 	sb = &so->so_snd;
16970 	if (do_a_prefetch == 0) {
16971 		kern_prefetch(sb, &do_a_prefetch);
16972 		do_a_prefetch = 1;
16973 	}
16974 #ifdef NETFLIX_SHARED_CWND
16975 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
16976 	    rack->rack_enable_scwnd) {
16977 		/* We are doing cwnd sharing */
16978 		if (rack->gp_ready &&
16979 		    (rack->rack_attempted_scwnd == 0) &&
16980 		    (rack->r_ctl.rc_scw == NULL) &&
16981 		    tp->t_lib) {
16982 			/* The pcbid is in, lets make an attempt */
16983 			counter_u64_add(rack_try_scwnd, 1);
16984 			rack->rack_attempted_scwnd = 1;
16985 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
16986 								   &rack->r_ctl.rc_scw_index,
16987 								   segsiz);
16988 		}
16989 		if (rack->r_ctl.rc_scw &&
16990 		    (rack->rack_scwnd_is_idle == 1) &&
16991 		    sbavail(&so->so_snd)) {
16992 			/* we are no longer out of data */
16993 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16994 			rack->rack_scwnd_is_idle = 0;
16995 		}
16996 		if (rack->r_ctl.rc_scw) {
16997 			/* First lets update and get the cwnd */
16998 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
16999 								    rack->r_ctl.rc_scw_index,
17000 								    tp->snd_cwnd, tp->snd_wnd, segsiz);
17001 		}
17002 	}
17003 #endif
17004 	/*
17005 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
17006 	 * state flags.
17007 	 */
17008 	if (tp->t_flags & TF_NEEDFIN)
17009 		flags |= TH_FIN;
17010 	if (tp->t_flags & TF_NEEDSYN)
17011 		flags |= TH_SYN;
17012 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17013 		void *end_rsm;
17014 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17015 		if (end_rsm)
17016 			kern_prefetch(end_rsm, &prefetch_rsm);
17017 		prefetch_rsm = 1;
17018 	}
17019 	SOCKBUF_LOCK(sb);
17020 	/*
17021 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
17022 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17023 	 * negative length.  This can also occur when TCP opens up its
17024 	 * congestion window while receiving additional duplicate acks after
17025 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
17026 	 * the fast-retransmit.
17027 	 *
17028 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
17029 	 * set to snd_una, the sb_offset will be 0, and the length may wind
17030 	 * up 0.
17031 	 *
17032 	 * If sack_rxmit is true we are retransmitting from the scoreboard
17033 	 * in which case len is already set.
17034 	 */
17035 	if ((sack_rxmit == 0) &&
17036 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17037 		uint32_t avail;
17038 
17039 		avail = sbavail(sb);
17040 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17041 			sb_offset = tp->snd_nxt - tp->snd_una;
17042 		else
17043 			sb_offset = 0;
17044 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17045 			if (rack->r_ctl.rc_tlp_new_data) {
17046 				/* TLP is forcing out new data */
17047 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17048 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17049 				}
17050 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17051 					if (tp->snd_wnd > sb_offset)
17052 						len = tp->snd_wnd - sb_offset;
17053 					else
17054 						len = 0;
17055 				} else {
17056 					len = rack->r_ctl.rc_tlp_new_data;
17057 				}
17058 				rack->r_ctl.rc_tlp_new_data = 0;
17059 			}  else {
17060 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17061 			}
17062 			if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17063 				/*
17064 				 * For prr=off, we need to send only 1 MSS
17065 				 * at a time. We do this because another sack could
17066 				 * be arriving that causes us to send retransmits and
17067 				 * we don't want to be on a long pace due to a larger send
17068 				 * that keeps us from sending out the retransmit.
17069 				 */
17070 				len = segsiz;
17071 			}
17072 		} else {
17073 			uint32_t outstanding;
17074 			/*
17075 			 * We are inside of a Fast recovery episode, this
17076 			 * is caused by a SACK or 3 dup acks. At this point
17077 			 * we have sent all the retransmissions and we rely
17078 			 * on PRR to dictate what we will send in the form of
17079 			 * new data.
17080 			 */
17081 
17082 			outstanding = tp->snd_max - tp->snd_una;
17083 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17084 				if (tp->snd_wnd > outstanding) {
17085 					len = tp->snd_wnd - outstanding;
17086 					/* Check to see if we have the data */
17087 					if ((sb_offset + len) > avail) {
17088 						/* It does not all fit */
17089 						if (avail > sb_offset)
17090 							len = avail - sb_offset;
17091 						else
17092 							len = 0;
17093 					}
17094 				} else {
17095 					len = 0;
17096 				}
17097 			} else if (avail > sb_offset) {
17098 				len = avail - sb_offset;
17099 			} else {
17100 				len = 0;
17101 			}
17102 			if (len > 0) {
17103 				if (len > rack->r_ctl.rc_prr_sndcnt) {
17104 					len = rack->r_ctl.rc_prr_sndcnt;
17105 				}
17106 				if (len > 0) {
17107 					sub_from_prr = 1;
17108 				}
17109 			}
17110 			if (len > segsiz) {
17111 				/*
17112 				 * We should never send more than a MSS when
17113 				 * retransmitting or sending new data in prr
17114 				 * mode unless the override flag is on. Most
17115 				 * likely the PRR algorithm is not going to
17116 				 * let us send a lot as well :-)
17117 				 */
17118 				if (rack->r_ctl.rc_prr_sendalot == 0) {
17119 					len = segsiz;
17120 				}
17121 			} else if (len < segsiz) {
17122 				/*
17123 				 * Do we send any? The idea here is if the
17124 				 * send empty's the socket buffer we want to
17125 				 * do it. However if not then lets just wait
17126 				 * for our prr_sndcnt to get bigger.
17127 				 */
17128 				long leftinsb;
17129 
17130 				leftinsb = sbavail(sb) - sb_offset;
17131 				if (leftinsb > len) {
17132 					/* This send does not empty the sb */
17133 					len = 0;
17134 				}
17135 			}
17136 		}
17137 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17138 		/*
17139 		 * If you have not established
17140 		 * and are not doing FAST OPEN
17141 		 * no data please.
17142 		 */
17143 		if ((sack_rxmit == 0) &&
17144 		    (!IS_FASTOPEN(tp->t_flags))){
17145 			len = 0;
17146 			sb_offset = 0;
17147 		}
17148 	}
17149 	if (prefetch_so_done == 0) {
17150 		kern_prefetch(so, &prefetch_so_done);
17151 		prefetch_so_done = 1;
17152 	}
17153 	/*
17154 	 * Lop off SYN bit if it has already been sent.  However, if this is
17155 	 * SYN-SENT state and if segment contains data and if we don't know
17156 	 * that foreign host supports TAO, suppress sending segment.
17157 	 */
17158 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17159 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17160 		/*
17161 		 * When sending additional segments following a TFO SYN|ACK,
17162 		 * do not include the SYN bit.
17163 		 */
17164 		if (IS_FASTOPEN(tp->t_flags) &&
17165 		    (tp->t_state == TCPS_SYN_RECEIVED))
17166 			flags &= ~TH_SYN;
17167 	}
17168 	/*
17169 	 * Be careful not to send data and/or FIN on SYN segments. This
17170 	 * measure is needed to prevent interoperability problems with not
17171 	 * fully conformant TCP implementations.
17172 	 */
17173 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17174 		len = 0;
17175 		flags &= ~TH_FIN;
17176 	}
17177 	/*
17178 	 * On TFO sockets, ensure no data is sent in the following cases:
17179 	 *
17180 	 *  - When retransmitting SYN|ACK on a passively-created socket
17181 	 *
17182 	 *  - When retransmitting SYN on an actively created socket
17183 	 *
17184 	 *  - When sending a zero-length cookie (cookie request) on an
17185 	 *    actively created socket
17186 	 *
17187 	 *  - When the socket is in the CLOSED state (RST is being sent)
17188 	 */
17189 	if (IS_FASTOPEN(tp->t_flags) &&
17190 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17191 	     ((tp->t_state == TCPS_SYN_SENT) &&
17192 	      (tp->t_tfo_client_cookie_len == 0)) ||
17193 	     (flags & TH_RST))) {
17194 		sack_rxmit = 0;
17195 		len = 0;
17196 	}
17197 	/* Without fast-open there should never be data sent on a SYN */
17198 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17199 		tp->snd_nxt = tp->iss;
17200 		len = 0;
17201 	}
17202 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17203 		/* We only send 1 MSS if we have a DSACK block */
17204 		add_flag |= RACK_SENT_W_DSACK;
17205 		len = segsiz;
17206 	}
17207 	orig_len = len;
17208 	if (len <= 0) {
17209 		/*
17210 		 * If FIN has been sent but not acked, but we haven't been
17211 		 * called to retransmit, len will be < 0.  Otherwise, window
17212 		 * shrank after we sent into it.  If window shrank to 0,
17213 		 * cancel pending retransmit, pull snd_nxt back to (closed)
17214 		 * window, and set the persist timer if it isn't already
17215 		 * going.  If the window didn't close completely, just wait
17216 		 * for an ACK.
17217 		 *
17218 		 * We also do a general check here to ensure that we will
17219 		 * set the persist timer when we have data to send, but a
17220 		 * 0-byte window. This makes sure the persist timer is set
17221 		 * even if the packet hits one of the "goto send" lines
17222 		 * below.
17223 		 */
17224 		len = 0;
17225 		if ((tp->snd_wnd == 0) &&
17226 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17227 		    (tp->snd_una == tp->snd_max) &&
17228 		    (sb_offset < (int)sbavail(sb))) {
17229 			rack_enter_persist(tp, rack, cts);
17230 		}
17231 	} else if ((rsm == NULL) &&
17232 		   (doing_tlp == 0) &&
17233 		   (len < pace_max_seg)) {
17234 		/*
17235 		 * We are not sending a maximum sized segment for
17236 		 * some reason. Should we not send anything (think
17237 		 * sws or persists)?
17238 		 */
17239 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17240 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17241 		    (len < minseg) &&
17242 		    (len < (int)(sbavail(sb) - sb_offset))) {
17243 			/*
17244 			 * Here the rwnd is less than
17245 			 * the minimum pacing size, this is not a retransmit,
17246 			 * we are established and
17247 			 * the send is not the last in the socket buffer
17248 			 * we send nothing, and we may enter persists
17249 			 * if nothing is outstanding.
17250 			 */
17251 			len = 0;
17252 			if (tp->snd_max == tp->snd_una) {
17253 				/*
17254 				 * Nothing out we can
17255 				 * go into persists.
17256 				 */
17257 				rack_enter_persist(tp, rack, cts);
17258 			}
17259 		     } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17260 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17261 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17262 			   (len < minseg)) {
17263 			/*
17264 			 * Here we are not retransmitting, and
17265 			 * the cwnd is not so small that we could
17266 			 * not send at least a min size (rxt timer
17267 			 * not having gone off), We have 2 segments or
17268 			 * more already in flight, its not the tail end
17269 			 * of the socket buffer  and the cwnd is blocking
17270 			 * us from sending out a minimum pacing segment size.
17271 			 * Lets not send anything.
17272 			 */
17273 			len = 0;
17274 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17275 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17276 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17277 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17278 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
17279 			/*
17280 			 * Here we have a send window but we have
17281 			 * filled it up and we can't send another pacing segment.
17282 			 * We also have in flight more than 2 segments
17283 			 * and we are not completing the sb i.e. we allow
17284 			 * the last bytes of the sb to go out even if
17285 			 * its not a full pacing segment.
17286 			 */
17287 			len = 0;
17288 		} else if ((rack->r_ctl.crte != NULL) &&
17289 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17290 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17291 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17292 			   (len < (int)(sbavail(sb) - sb_offset))) {
17293 			/*
17294 			 * Here we are doing hardware pacing, this is not a TLP,
17295 			 * we are not sending a pace max segment size, there is rwnd
17296 			 * room to send at least N pace_max_seg, the cwnd is greater
17297 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
17298 			 * more segments in flight and its not the tail of the socket buffer.
17299 			 *
17300 			 * We don't want to send instead we need to get more ack's in to
17301 			 * allow us to send a full pacing segment. Normally, if we are pacing
17302 			 * about the right speed, we should have finished our pacing
17303 			 * send as most of the acks have come back if we are at the
17304 			 * right rate. This is a bit fuzzy since return path delay
17305 			 * can delay the acks, which is why we want to make sure we
17306 			 * have cwnd space to have a bit more than a max pace segments in flight.
17307 			 *
17308 			 * If we have not gotten our acks back we are pacing at too high a
17309 			 * rate delaying will not hurt and will bring our GP estimate down by
17310 			 * injecting the delay. If we don't do this we will send
17311 			 * 2 MSS out in response to the acks being clocked in which
17312 			 * defeats the point of hw-pacing (i.e. to help us get
17313 			 * larger TSO's out).
17314 			 */
17315 			len = 0;
17316 
17317 		}
17318 
17319 	}
17320 	/* len will be >= 0 after this point. */
17321 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17322 	rack_sndbuf_autoscale(rack);
17323 	/*
17324 	 * Decide if we can use TCP Segmentation Offloading (if supported by
17325 	 * hardware).
17326 	 *
17327 	 * TSO may only be used if we are in a pure bulk sending state.  The
17328 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17329 	 * options prevent using TSO.  With TSO the TCP header is the same
17330 	 * (except for the sequence number) for all generated packets.  This
17331 	 * makes it impossible to transmit any options which vary per
17332 	 * generated segment or packet.
17333 	 *
17334 	 * IPv4 handling has a clear separation of ip options and ip header
17335 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17336 	 * the right thing below to provide length of just ip options and thus
17337 	 * checking for ipoptlen is enough to decide if ip options are present.
17338 	 */
17339 	ipoptlen = 0;
17340 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17341 	/*
17342 	 * Pre-calculate here as we save another lookup into the darknesses
17343 	 * of IPsec that way and can actually decide if TSO is ok.
17344 	 */
17345 #ifdef INET6
17346 	if (isipv6 && IPSEC_ENABLED(ipv6))
17347 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
17348 #ifdef INET
17349 	else
17350 #endif
17351 #endif				/* INET6 */
17352 #ifdef INET
17353 		if (IPSEC_ENABLED(ipv4))
17354 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
17355 #endif				/* INET */
17356 #endif
17357 
17358 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17359 	ipoptlen += ipsec_optlen;
17360 #endif
17361 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17362 	    (tp->t_port == 0) &&
17363 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
17364 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17365 	    ipoptlen == 0)
17366 		tso = 1;
17367 	{
17368 		uint32_t outstanding;
17369 
17370 		outstanding = tp->snd_max - tp->snd_una;
17371 		if (tp->t_flags & TF_SENTFIN) {
17372 			/*
17373 			 * If we sent a fin, snd_max is 1 higher than
17374 			 * snd_una
17375 			 */
17376 			outstanding--;
17377 		}
17378 		if (sack_rxmit) {
17379 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17380 				flags &= ~TH_FIN;
17381 		} else {
17382 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17383 				   sbused(sb)))
17384 				flags &= ~TH_FIN;
17385 		}
17386 	}
17387 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17388 	    (long)TCP_MAXWIN << tp->rcv_scale);
17389 
17390 	/*
17391 	 * Sender silly window avoidance.   We transmit under the following
17392 	 * conditions when len is non-zero:
17393 	 *
17394 	 * - We have a full segment (or more with TSO) - This is the last
17395 	 * buffer in a write()/send() and we are either idle or running
17396 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
17397 	 * then 1/2 the maximum send window's worth of data (receiver may be
17398 	 * limited the window size) - we need to retransmit
17399 	 */
17400 	if (len) {
17401 		if (len >= segsiz) {
17402 			goto send;
17403 		}
17404 		/*
17405 		 * NOTE! on localhost connections an 'ack' from the remote
17406 		 * end may occur synchronously with the output and cause us
17407 		 * to flush a buffer queued with moretocome.  XXX
17408 		 *
17409 		 */
17410 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
17411 		    (idle || (tp->t_flags & TF_NODELAY)) &&
17412 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17413 		    (tp->t_flags & TF_NOPUSH) == 0) {
17414 			pass = 2;
17415 			goto send;
17416 		}
17417 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
17418 			pass = 22;
17419 			goto send;
17420 		}
17421 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17422 			pass = 4;
17423 			goto send;
17424 		}
17425 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
17426 			pass = 5;
17427 			goto send;
17428 		}
17429 		if (sack_rxmit) {
17430 			pass = 6;
17431 			goto send;
17432 		}
17433 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17434 		    (ctf_outstanding(tp) < (segsiz * 2))) {
17435 			/*
17436 			 * We have less than two MSS outstanding (delayed ack)
17437 			 * and our rwnd will not let us send a full sized
17438 			 * MSS. Lets go ahead and let this small segment
17439 			 * out because we want to try to have at least two
17440 			 * packets inflight to not be caught by delayed ack.
17441 			 */
17442 			pass = 12;
17443 			goto send;
17444 		}
17445 	}
17446 	/*
17447 	 * Sending of standalone window updates.
17448 	 *
17449 	 * Window updates are important when we close our window due to a
17450 	 * full socket buffer and are opening it again after the application
17451 	 * reads data from it.  Once the window has opened again and the
17452 	 * remote end starts to send again the ACK clock takes over and
17453 	 * provides the most current window information.
17454 	 *
17455 	 * We must avoid the silly window syndrome whereas every read from
17456 	 * the receive buffer, no matter how small, causes a window update
17457 	 * to be sent.  We also should avoid sending a flurry of window
17458 	 * updates when the socket buffer had queued a lot of data and the
17459 	 * application is doing small reads.
17460 	 *
17461 	 * Prevent a flurry of pointless window updates by only sending an
17462 	 * update when we can increase the advertized window by more than
17463 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
17464 	 * full or is very small be more aggressive and send an update
17465 	 * whenever we can increase by two mss sized segments. In all other
17466 	 * situations the ACK's to new incoming data will carry further
17467 	 * window increases.
17468 	 *
17469 	 * Don't send an independent window update if a delayed ACK is
17470 	 * pending (it will get piggy-backed on it) or the remote side
17471 	 * already has done a half-close and won't send more data.  Skip
17472 	 * this if the connection is in T/TCP half-open state.
17473 	 */
17474 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17475 	    !(tp->t_flags & TF_DELACK) &&
17476 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
17477 		/*
17478 		 * "adv" is the amount we could increase the window, taking
17479 		 * into account that we are limited by TCP_MAXWIN <<
17480 		 * tp->rcv_scale.
17481 		 */
17482 		int32_t adv;
17483 		int oldwin;
17484 
17485 		adv = recwin;
17486 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17487 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
17488 			if (adv > oldwin)
17489 			    adv -= oldwin;
17490 			else {
17491 				/* We can't increase the window */
17492 				adv = 0;
17493 			}
17494 		} else
17495 			oldwin = 0;
17496 
17497 		/*
17498 		 * If the new window size ends up being the same as or less
17499 		 * than the old size when it is scaled, then don't force
17500 		 * a window update.
17501 		 */
17502 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17503 			goto dontupdate;
17504 
17505 		if (adv >= (int32_t)(2 * segsiz) &&
17506 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17507 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17508 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17509 			pass = 7;
17510 			goto send;
17511 		}
17512 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17513 			pass = 23;
17514 			goto send;
17515 		}
17516 	}
17517 dontupdate:
17518 
17519 	/*
17520 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17521 	 * is also a catch-all for the retransmit timer timeout case.
17522 	 */
17523 	if (tp->t_flags & TF_ACKNOW) {
17524 		pass = 8;
17525 		goto send;
17526 	}
17527 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17528 		pass = 9;
17529 		goto send;
17530 	}
17531 	/*
17532 	 * If our state indicates that FIN should be sent and we have not
17533 	 * yet done so, then we need to send.
17534 	 */
17535 	if ((flags & TH_FIN) &&
17536 	    (tp->snd_nxt == tp->snd_una)) {
17537 		pass = 11;
17538 		goto send;
17539 	}
17540 	/*
17541 	 * No reason to send a segment, just return.
17542 	 */
17543 just_return:
17544 	SOCKBUF_UNLOCK(sb);
17545 just_return_nolock:
17546 	{
17547 		int app_limited = CTF_JR_SENT_DATA;
17548 
17549 		if (tot_len_this_send > 0) {
17550 			/* Make sure snd_nxt is up to max */
17551 			rack->r_ctl.fsb.recwin = recwin;
17552 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17553 			if ((error == 0) &&
17554 			    rack_use_rfo &&
17555 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
17556 			    (ipoptlen == 0) &&
17557 			    (tp->snd_nxt == tp->snd_max) &&
17558 			    (tp->rcv_numsacks == 0) &&
17559 			    rack->r_fsb_inited &&
17560 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
17561 			    (rack->r_must_retran == 0) &&
17562 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
17563 			    (len > 0) && (orig_len > 0) &&
17564 			    (orig_len > len) &&
17565 			    ((orig_len - len) >= segsiz) &&
17566 			    ((optlen == 0) ||
17567 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17568 				/* We can send at least one more MSS using our fsb */
17569 
17570 				rack->r_fast_output = 1;
17571 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17572 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17573 				rack->r_ctl.fsb.tcp_flags = flags;
17574 				rack->r_ctl.fsb.left_to_send = orig_len - len;
17575 				if (hw_tls)
17576 					rack->r_ctl.fsb.hw_tls = 1;
17577 				else
17578 					rack->r_ctl.fsb.hw_tls = 0;
17579 				KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17580 					("rack:%p left_to_send:%u sbavail:%u out:%u",
17581 					rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17582 					 (tp->snd_max - tp->snd_una)));
17583 				if (rack->r_ctl.fsb.left_to_send < segsiz)
17584 					rack->r_fast_output = 0;
17585 				else {
17586 					if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17587 						rack->r_ctl.fsb.rfo_apply_push = 1;
17588 					else
17589 						rack->r_ctl.fsb.rfo_apply_push = 0;
17590 				}
17591 			} else
17592 				rack->r_fast_output = 0;
17593 
17594 
17595 			rack_log_fsb(rack, tp, so, flags,
17596 				     ipoptlen, orig_len, len, 0,
17597 				     1, optlen, __LINE__, 1);
17598 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17599 				tp->snd_nxt = tp->snd_max;
17600 		} else {
17601 			int end_window = 0;
17602 			uint32_t seq = tp->gput_ack;
17603 
17604 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17605 			if (rsm) {
17606 				/*
17607 				 * Mark the last sent that we just-returned (hinting
17608 				 * that delayed ack may play a role in any rtt measurement).
17609 				 */
17610 				rsm->r_just_ret = 1;
17611 			}
17612 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17613 			rack->r_ctl.rc_agg_delayed = 0;
17614 			rack->r_early = 0;
17615 			rack->r_late = 0;
17616 			rack->r_ctl.rc_agg_early = 0;
17617 			if ((ctf_outstanding(tp) +
17618 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17619 				 minseg)) >= tp->snd_wnd) {
17620 				/* We are limited by the rwnd */
17621 				app_limited = CTF_JR_RWND_LIMITED;
17622 				if (IN_FASTRECOVERY(tp->t_flags))
17623 				    rack->r_ctl.rc_prr_sndcnt = 0;
17624 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
17625 				/* We are limited by whats available -- app limited */
17626 				app_limited = CTF_JR_APP_LIMITED;
17627 				if (IN_FASTRECOVERY(tp->t_flags))
17628 				    rack->r_ctl.rc_prr_sndcnt = 0;
17629 			} else if ((idle == 0) &&
17630 				   ((tp->t_flags & TF_NODELAY) == 0) &&
17631 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17632 				   (len < segsiz)) {
17633 				/*
17634 				 * No delay is not on and the
17635 				 * user is sending less than 1MSS. This
17636 				 * brings out SWS avoidance so we
17637 				 * don't send. Another app-limited case.
17638 				 */
17639 				app_limited = CTF_JR_APP_LIMITED;
17640 			} else if (tp->t_flags & TF_NOPUSH) {
17641 				/*
17642 				 * The user has requested no push of
17643 				 * the last segment and we are
17644 				 * at the last segment. Another app
17645 				 * limited case.
17646 				 */
17647 				app_limited = CTF_JR_APP_LIMITED;
17648 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17649 				/* Its the cwnd */
17650 				app_limited = CTF_JR_CWND_LIMITED;
17651 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
17652 				   (rack->rack_no_prr == 0) &&
17653 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17654 				app_limited = CTF_JR_PRR;
17655 			} else {
17656 				/* Now why here are we not sending? */
17657 #ifdef NOW
17658 #ifdef INVARIANTS
17659 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17660 #endif
17661 #endif
17662 				app_limited = CTF_JR_ASSESSING;
17663 			}
17664 			/*
17665 			 * App limited in some fashion, for our pacing GP
17666 			 * measurements we don't want any gap (even cwnd).
17667 			 * Close  down the measurement window.
17668 			 */
17669 			if (rack_cwnd_block_ends_measure &&
17670 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
17671 			     (app_limited == CTF_JR_PRR))) {
17672 				/*
17673 				 * The reason we are not sending is
17674 				 * the cwnd (or prr). We have been configured
17675 				 * to end the measurement window in
17676 				 * this case.
17677 				 */
17678 				end_window = 1;
17679 			} else if (rack_rwnd_block_ends_measure &&
17680 				   (app_limited == CTF_JR_RWND_LIMITED)) {
17681 				/*
17682 				 * We are rwnd limited and have been
17683 				 * configured to end the measurement
17684 				 * window in this case.
17685 				 */
17686 				end_window = 1;
17687 			} else if (app_limited == CTF_JR_APP_LIMITED) {
17688 				/*
17689 				 * A true application limited period, we have
17690 				 * ran out of data.
17691 				 */
17692 				end_window = 1;
17693 			} else if (app_limited == CTF_JR_ASSESSING) {
17694 				/*
17695 				 * In the assessing case we hit the end of
17696 				 * the if/else and had no known reason
17697 				 * This will panic us under invariants..
17698 				 *
17699 				 * If we get this out in logs we need to
17700 				 * investagate which reason we missed.
17701 				 */
17702 				end_window = 1;
17703 			}
17704 			if (end_window) {
17705 				uint8_t log = 0;
17706 
17707 				/* Adjust the Gput measurement */
17708 				if ((tp->t_flags & TF_GPUTINPROG) &&
17709 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
17710 					tp->gput_ack = tp->snd_max;
17711 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17712 						/*
17713 						 * There is not enough to measure.
17714 						 */
17715 						tp->t_flags &= ~TF_GPUTINPROG;
17716 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17717 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
17718 									   tp->gput_seq,
17719 									   0, 0, 18, __LINE__, NULL, 0);
17720 					} else
17721 						log = 1;
17722 				}
17723 				/* Mark the last packet has app limited */
17724 				rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17725 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17726 					if (rack->r_ctl.rc_app_limited_cnt == 0)
17727 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17728 					else {
17729 						/*
17730 						 * Go out to the end app limited and mark
17731 						 * this new one as next and move the end_appl up
17732 						 * to this guy.
17733 						 */
17734 						if (rack->r_ctl.rc_end_appl)
17735 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17736 						rack->r_ctl.rc_end_appl = rsm;
17737 					}
17738 					rsm->r_flags |= RACK_APP_LIMITED;
17739 					rack->r_ctl.rc_app_limited_cnt++;
17740 				}
17741 				if (log)
17742 					rack_log_pacing_delay_calc(rack,
17743 								   rack->r_ctl.rc_app_limited_cnt, seq,
17744 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17745 			}
17746 		}
17747 		/* Check if we need to go into persists or not */
17748 		if ((tp->snd_max == tp->snd_una) &&
17749 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
17750 		    sbavail(sb) &&
17751 		    (sbavail(sb) > tp->snd_wnd) &&
17752 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17753 			/* Yes lets make sure to move to persist before timer-start */
17754 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17755 		}
17756 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17757 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17758 	}
17759 #ifdef NETFLIX_SHARED_CWND
17760 	if ((sbavail(sb) == 0) &&
17761 	    rack->r_ctl.rc_scw) {
17762 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17763 		rack->rack_scwnd_is_idle = 1;
17764 	}
17765 #endif
17766 #ifdef TCP_ACCOUNTING
17767 	if (tot_len_this_send > 0) {
17768 		crtsc = get_cyclecount();
17769 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17770 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
17771 		}
17772 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17773 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17774 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17775 		}
17776 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17777 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17778 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17779 		}
17780 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17781 	} else {
17782 		crtsc = get_cyclecount();
17783 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17784 			tp->tcp_cnt_counters[SND_LIMITED]++;
17785 		}
17786 		counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17787 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17788 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17789 		}
17790 		counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17791 	}
17792 	sched_unpin();
17793 #endif
17794 	return (0);
17795 
17796 send:
17797 	if (rsm || sack_rxmit)
17798 		counter_u64_add(rack_nfto_resend, 1);
17799 	else
17800 		counter_u64_add(rack_non_fto_send, 1);
17801 	if ((flags & TH_FIN) &&
17802 	    sbavail(sb)) {
17803 		/*
17804 		 * We do not transmit a FIN
17805 		 * with data outstanding. We
17806 		 * need to make it so all data
17807 		 * is acked first.
17808 		 */
17809 		flags &= ~TH_FIN;
17810 	}
17811 	/* Enforce stack imposed max seg size if we have one */
17812 	if (rack->r_ctl.rc_pace_max_segs &&
17813 	    (len > rack->r_ctl.rc_pace_max_segs)) {
17814 		mark = 1;
17815 		len = rack->r_ctl.rc_pace_max_segs;
17816 	}
17817 	SOCKBUF_LOCK_ASSERT(sb);
17818 	if (len > 0) {
17819 		if (len >= segsiz)
17820 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17821 		else
17822 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17823 	}
17824 	/*
17825 	 * Before ESTABLISHED, force sending of initial options unless TCP
17826 	 * set not to do any options. NOTE: we assume that the IP/TCP header
17827 	 * plus TCP options always fit in a single mbuf, leaving room for a
17828 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17829 	 * + optlen <= MCLBYTES
17830 	 */
17831 	optlen = 0;
17832 #ifdef INET6
17833 	if (isipv6)
17834 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17835 	else
17836 #endif
17837 		hdrlen = sizeof(struct tcpiphdr);
17838 
17839 	/*
17840 	 * Compute options for segment. We only have to care about SYN and
17841 	 * established connection segments.  Options for SYN-ACK segments
17842 	 * are handled in TCP syncache.
17843 	 */
17844 	to.to_flags = 0;
17845 	if ((tp->t_flags & TF_NOOPT) == 0) {
17846 		/* Maximum segment size. */
17847 		if (flags & TH_SYN) {
17848 			tp->snd_nxt = tp->iss;
17849 			to.to_mss = tcp_mssopt(&inp->inp_inc);
17850 			if (tp->t_port)
17851 				to.to_mss -= V_tcp_udp_tunneling_overhead;
17852 			to.to_flags |= TOF_MSS;
17853 
17854 			/*
17855 			 * On SYN or SYN|ACK transmits on TFO connections,
17856 			 * only include the TFO option if it is not a
17857 			 * retransmit, as the presence of the TFO option may
17858 			 * have caused the original SYN or SYN|ACK to have
17859 			 * been dropped by a middlebox.
17860 			 */
17861 			if (IS_FASTOPEN(tp->t_flags) &&
17862 			    (tp->t_rxtshift == 0)) {
17863 				if (tp->t_state == TCPS_SYN_RECEIVED) {
17864 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17865 					to.to_tfo_cookie =
17866 						(u_int8_t *)&tp->t_tfo_cookie.server;
17867 					to.to_flags |= TOF_FASTOPEN;
17868 					wanted_cookie = 1;
17869 				} else if (tp->t_state == TCPS_SYN_SENT) {
17870 					to.to_tfo_len =
17871 						tp->t_tfo_client_cookie_len;
17872 					to.to_tfo_cookie =
17873 						tp->t_tfo_cookie.client;
17874 					to.to_flags |= TOF_FASTOPEN;
17875 					wanted_cookie = 1;
17876 					/*
17877 					 * If we wind up having more data to
17878 					 * send with the SYN than can fit in
17879 					 * one segment, don't send any more
17880 					 * until the SYN|ACK comes back from
17881 					 * the other end.
17882 					 */
17883 					sendalot = 0;
17884 				}
17885 			}
17886 		}
17887 		/* Window scaling. */
17888 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17889 			to.to_wscale = tp->request_r_scale;
17890 			to.to_flags |= TOF_SCALE;
17891 		}
17892 		/* Timestamps. */
17893 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
17894 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17895 			to.to_tsval = ms_cts + tp->ts_offset;
17896 			to.to_tsecr = tp->ts_recent;
17897 			to.to_flags |= TOF_TS;
17898 		}
17899 		/* Set receive buffer autosizing timestamp. */
17900 		if (tp->rfbuf_ts == 0 &&
17901 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
17902 			tp->rfbuf_ts = tcp_ts_getticks();
17903 		/* Selective ACK's. */
17904 		if (tp->t_flags & TF_SACK_PERMIT) {
17905 			if (flags & TH_SYN)
17906 				to.to_flags |= TOF_SACKPERM;
17907 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17908 				 tp->rcv_numsacks > 0) {
17909 				to.to_flags |= TOF_SACK;
17910 				to.to_nsacks = tp->rcv_numsacks;
17911 				to.to_sacks = (u_char *)tp->sackblks;
17912 			}
17913 		}
17914 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17915 		/* TCP-MD5 (RFC2385). */
17916 		if (tp->t_flags & TF_SIGNATURE)
17917 			to.to_flags |= TOF_SIGNATURE;
17918 #endif				/* TCP_SIGNATURE */
17919 
17920 		/* Processing the options. */
17921 		hdrlen += optlen = tcp_addoptions(&to, opt);
17922 		/*
17923 		 * If we wanted a TFO option to be added, but it was unable
17924 		 * to fit, ensure no data is sent.
17925 		 */
17926 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17927 		    !(to.to_flags & TOF_FASTOPEN))
17928 			len = 0;
17929 	}
17930 	if (tp->t_port) {
17931 		if (V_tcp_udp_tunneling_port == 0) {
17932 			/* The port was removed?? */
17933 			SOCKBUF_UNLOCK(&so->so_snd);
17934 #ifdef TCP_ACCOUNTING
17935 			crtsc = get_cyclecount();
17936 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17937 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17938 			}
17939 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
17940 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17941 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17942 			}
17943 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
17944 			sched_unpin();
17945 #endif
17946 			return (EHOSTUNREACH);
17947 		}
17948 		hdrlen += sizeof(struct udphdr);
17949 	}
17950 #ifdef INET6
17951 	if (isipv6)
17952 		ipoptlen = ip6_optlen(tp->t_inpcb);
17953 	else
17954 #endif
17955 		if (tp->t_inpcb->inp_options)
17956 			ipoptlen = tp->t_inpcb->inp_options->m_len -
17957 				offsetof(struct ipoption, ipopt_list);
17958 		else
17959 			ipoptlen = 0;
17960 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17961 	ipoptlen += ipsec_optlen;
17962 #endif
17963 
17964 	/*
17965 	 * Adjust data length if insertion of options will bump the packet
17966 	 * length beyond the t_maxseg length. Clear the FIN bit because we
17967 	 * cut off the tail of the segment.
17968 	 */
17969 	if (len + optlen + ipoptlen > tp->t_maxseg) {
17970 		if (tso) {
17971 			uint32_t if_hw_tsomax;
17972 			uint32_t moff;
17973 			int32_t max_len;
17974 
17975 			/* extract TSO information */
17976 			if_hw_tsomax = tp->t_tsomax;
17977 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
17978 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
17979 			KASSERT(ipoptlen == 0,
17980 				("%s: TSO can't do IP options", __func__));
17981 
17982 			/*
17983 			 * Check if we should limit by maximum payload
17984 			 * length:
17985 			 */
17986 			if (if_hw_tsomax != 0) {
17987 				/* compute maximum TSO length */
17988 				max_len = (if_hw_tsomax - hdrlen -
17989 					   max_linkhdr);
17990 				if (max_len <= 0) {
17991 					len = 0;
17992 				} else if (len > max_len) {
17993 					sendalot = 1;
17994 					len = max_len;
17995 					mark = 2;
17996 				}
17997 			}
17998 			/*
17999 			 * Prevent the last segment from being fractional
18000 			 * unless the send sockbuf can be emptied:
18001 			 */
18002 			max_len = (tp->t_maxseg - optlen);
18003 			if ((sb_offset + len) < sbavail(sb)) {
18004 				moff = len % (u_int)max_len;
18005 				if (moff != 0) {
18006 					mark = 3;
18007 					len -= moff;
18008 				}
18009 			}
18010 			/*
18011 			 * In case there are too many small fragments don't
18012 			 * use TSO:
18013 			 */
18014 			if (len <= segsiz) {
18015 				mark = 4;
18016 				tso = 0;
18017 			}
18018 			/*
18019 			 * Send the FIN in a separate segment after the bulk
18020 			 * sending is done. We don't trust the TSO
18021 			 * implementations to clear the FIN flag on all but
18022 			 * the last segment.
18023 			 */
18024 			if (tp->t_flags & TF_NEEDFIN) {
18025 				sendalot = 4;
18026 			}
18027 		} else {
18028 			mark = 5;
18029 			if (optlen + ipoptlen >= tp->t_maxseg) {
18030 				/*
18031 				 * Since we don't have enough space to put
18032 				 * the IP header chain and the TCP header in
18033 				 * one packet as required by RFC 7112, don't
18034 				 * send it. Also ensure that at least one
18035 				 * byte of the payload can be put into the
18036 				 * TCP segment.
18037 				 */
18038 				SOCKBUF_UNLOCK(&so->so_snd);
18039 				error = EMSGSIZE;
18040 				sack_rxmit = 0;
18041 				goto out;
18042 			}
18043 			len = tp->t_maxseg - optlen - ipoptlen;
18044 			sendalot = 5;
18045 		}
18046 	} else {
18047 		tso = 0;
18048 		mark = 6;
18049 	}
18050 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18051 		("%s: len > IP_MAXPACKET", __func__));
18052 #ifdef DIAGNOSTIC
18053 #ifdef INET6
18054 	if (max_linkhdr + hdrlen > MCLBYTES)
18055 #else
18056 		if (max_linkhdr + hdrlen > MHLEN)
18057 #endif
18058 			panic("tcphdr too big");
18059 #endif
18060 
18061 	/*
18062 	 * This KASSERT is here to catch edge cases at a well defined place.
18063 	 * Before, those had triggered (random) panic conditions further
18064 	 * down.
18065 	 */
18066 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18067 	if ((len == 0) &&
18068 	    (flags & TH_FIN) &&
18069 	    (sbused(sb))) {
18070 		/*
18071 		 * We have outstanding data, don't send a fin by itself!.
18072 		 */
18073 		goto just_return;
18074 	}
18075 	/*
18076 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
18077 	 * and initialize the header from the template for sends on this
18078 	 * connection.
18079 	 */
18080 	hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
18081 	if (len) {
18082 		uint32_t max_val;
18083 		uint32_t moff;
18084 
18085 		if (rack->r_ctl.rc_pace_max_segs)
18086 			max_val = rack->r_ctl.rc_pace_max_segs;
18087 		else if (rack->rc_user_set_max_segs)
18088 			max_val = rack->rc_user_set_max_segs * segsiz;
18089 		else
18090 			max_val = len;
18091 		/*
18092 		 * We allow a limit on sending with hptsi.
18093 		 */
18094 		if (len > max_val) {
18095 			mark = 7;
18096 			len = max_val;
18097 		}
18098 #ifdef INET6
18099 		if (MHLEN < hdrlen + max_linkhdr)
18100 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18101 		else
18102 #endif
18103 			m = m_gethdr(M_NOWAIT, MT_DATA);
18104 
18105 		if (m == NULL) {
18106 			SOCKBUF_UNLOCK(sb);
18107 			error = ENOBUFS;
18108 			sack_rxmit = 0;
18109 			goto out;
18110 		}
18111 		m->m_data += max_linkhdr;
18112 		m->m_len = hdrlen;
18113 
18114 		/*
18115 		 * Start the m_copy functions from the closest mbuf to the
18116 		 * sb_offset in the socket buffer chain.
18117 		 */
18118 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
18119 		s_mb = mb;
18120 		s_moff = moff;
18121 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18122 			m_copydata(mb, moff, (int)len,
18123 				   mtod(m, caddr_t)+hdrlen);
18124 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18125 				sbsndptr_adv(sb, mb, len);
18126 			m->m_len += len;
18127 		} else {
18128 			struct sockbuf *msb;
18129 
18130 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18131 				msb = NULL;
18132 			else
18133 				msb = sb;
18134 			m->m_next = tcp_m_copym(
18135 				mb, moff, &len,
18136 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18137 				((rsm == NULL) ? hw_tls : 0)
18138 #ifdef NETFLIX_COPY_ARGS
18139 				, &filled_all
18140 #endif
18141 				);
18142 			if (len <= (tp->t_maxseg - optlen)) {
18143 				/*
18144 				 * Must have ran out of mbufs for the copy
18145 				 * shorten it to no longer need tso. Lets
18146 				 * not put on sendalot since we are low on
18147 				 * mbufs.
18148 				 */
18149 				tso = 0;
18150 			}
18151 			if (m->m_next == NULL) {
18152 				SOCKBUF_UNLOCK(sb);
18153 				(void)m_free(m);
18154 				error = ENOBUFS;
18155 				sack_rxmit = 0;
18156 				goto out;
18157 			}
18158 		}
18159 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18160 			if (rsm && (rsm->r_flags & RACK_TLP)) {
18161 				/*
18162 				 * TLP should not count in retran count, but
18163 				 * in its own bin
18164 				 */
18165 				counter_u64_add(rack_tlp_retran, 1);
18166 				counter_u64_add(rack_tlp_retran_bytes, len);
18167 			} else {
18168 				tp->t_sndrexmitpack++;
18169 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18170 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18171 			}
18172 #ifdef STATS
18173 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18174 						 len);
18175 #endif
18176 		} else {
18177 			KMOD_TCPSTAT_INC(tcps_sndpack);
18178 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18179 #ifdef STATS
18180 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18181 						 len);
18182 #endif
18183 		}
18184 		/*
18185 		 * If we're sending everything we've got, set PUSH. (This
18186 		 * will keep happy those implementations which only give
18187 		 * data to the user when a buffer fills or a PUSH comes in.)
18188 		 */
18189 		if (sb_offset + len == sbused(sb) &&
18190 		    sbused(sb) &&
18191 		    !(flags & TH_SYN)) {
18192 			flags |= TH_PUSH;
18193 			add_flag |= RACK_HAD_PUSH;
18194 		}
18195 
18196 		SOCKBUF_UNLOCK(sb);
18197 	} else {
18198 		SOCKBUF_UNLOCK(sb);
18199 		if (tp->t_flags & TF_ACKNOW)
18200 			KMOD_TCPSTAT_INC(tcps_sndacks);
18201 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
18202 			KMOD_TCPSTAT_INC(tcps_sndctrl);
18203 		else
18204 			KMOD_TCPSTAT_INC(tcps_sndwinup);
18205 
18206 		m = m_gethdr(M_NOWAIT, MT_DATA);
18207 		if (m == NULL) {
18208 			error = ENOBUFS;
18209 			sack_rxmit = 0;
18210 			goto out;
18211 		}
18212 #ifdef INET6
18213 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18214 		    MHLEN >= hdrlen) {
18215 			M_ALIGN(m, hdrlen);
18216 		} else
18217 #endif
18218 			m->m_data += max_linkhdr;
18219 		m->m_len = hdrlen;
18220 	}
18221 	SOCKBUF_UNLOCK_ASSERT(sb);
18222 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18223 #ifdef MAC
18224 	mac_inpcb_create_mbuf(inp, m);
18225 #endif
18226 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18227 #ifdef INET6
18228 		if (isipv6)
18229 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18230 		else
18231 #endif				/* INET6 */
18232 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18233 		th = rack->r_ctl.fsb.th;
18234 		udp = rack->r_ctl.fsb.udp;
18235 		if (udp) {
18236 #ifdef INET6
18237 			if (isipv6)
18238 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18239 			else
18240 #endif				/* INET6 */
18241 				ulen = hdrlen + len - sizeof(struct ip);
18242 			udp->uh_ulen = htons(ulen);
18243 		}
18244 	} else {
18245 #ifdef INET6
18246 		if (isipv6) {
18247 			ip6 = mtod(m, struct ip6_hdr *);
18248 			if (tp->t_port) {
18249 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18250 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18251 				udp->uh_dport = tp->t_port;
18252 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18253 				udp->uh_ulen = htons(ulen);
18254 				th = (struct tcphdr *)(udp + 1);
18255 			} else
18256 				th = (struct tcphdr *)(ip6 + 1);
18257 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
18258 		} else
18259 #endif				/* INET6 */
18260 		{
18261 			ip = mtod(m, struct ip *);
18262 #ifdef TCPDEBUG
18263 			ipov = (struct ipovly *)ip;
18264 #endif
18265 			if (tp->t_port) {
18266 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18267 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18268 				udp->uh_dport = tp->t_port;
18269 				ulen = hdrlen + len - sizeof(struct ip);
18270 				udp->uh_ulen = htons(ulen);
18271 				th = (struct tcphdr *)(udp + 1);
18272 			} else
18273 				th = (struct tcphdr *)(ip + 1);
18274 			tcpip_fillheaders(inp, tp->t_port, ip, th);
18275 		}
18276 	}
18277 	/*
18278 	 * Fill in fields, remembering maximum advertised window for use in
18279 	 * delaying messages about window sizes. If resending a FIN, be sure
18280 	 * not to use a new sequence number.
18281 	 */
18282 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18283 	    tp->snd_nxt == tp->snd_max)
18284 		tp->snd_nxt--;
18285 	/*
18286 	 * If we are starting a connection, send ECN setup SYN packet. If we
18287 	 * are on a retransmit, we may resend those bits a number of times
18288 	 * as per RFC 3168.
18289 	 */
18290 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
18291 		flags |= tcp_ecn_output_syn_sent(tp);
18292 	}
18293 	/* Also handle parallel SYN for ECN */
18294 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18295 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
18296 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
18297 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18298 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18299 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18300 #ifdef INET6
18301 		if (isipv6) {
18302 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18303 			ip6->ip6_flow |= htonl(ect << 20);
18304 		}
18305 		else
18306 #endif
18307 		{
18308 			ip->ip_tos &= ~IPTOS_ECN_MASK;
18309 			ip->ip_tos |= ect;
18310 		}
18311 	}
18312 	/*
18313 	 * If we are doing retransmissions, then snd_nxt will not reflect
18314 	 * the first unsent octet.  For ACK only packets, we do not want the
18315 	 * sequence number of the retransmitted packet, we want the sequence
18316 	 * number of the next unsent octet.  So, if there is no data (and no
18317 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
18318 	 * ti_seq.  But if we are in persist state, snd_max might reflect
18319 	 * one byte beyond the right edge of the window, so use snd_nxt in
18320 	 * that case, since we know we aren't doing a retransmission.
18321 	 * (retransmit and persist are mutually exclusive...)
18322 	 */
18323 	if (sack_rxmit == 0) {
18324 		if (len || (flags & (TH_SYN | TH_FIN))) {
18325 			th->th_seq = htonl(tp->snd_nxt);
18326 			rack_seq = tp->snd_nxt;
18327 		} else {
18328 			th->th_seq = htonl(tp->snd_max);
18329 			rack_seq = tp->snd_max;
18330 		}
18331 	} else {
18332 		th->th_seq = htonl(rsm->r_start);
18333 		rack_seq = rsm->r_start;
18334 	}
18335 	th->th_ack = htonl(tp->rcv_nxt);
18336 	tcp_set_flags(th, flags);
18337 	/*
18338 	 * Calculate receive window.  Don't shrink window, but avoid silly
18339 	 * window syndrome.
18340 	 * If a RST segment is sent, advertise a window of zero.
18341 	 */
18342 	if (flags & TH_RST) {
18343 		recwin = 0;
18344 	} else {
18345 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18346 		    recwin < (long)segsiz) {
18347 			recwin = 0;
18348 		}
18349 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18350 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18351 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18352 	}
18353 
18354 	/*
18355 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18356 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18357 	 * handled in syncache.
18358 	 */
18359 	if (flags & TH_SYN)
18360 		th->th_win = htons((u_short)
18361 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18362 	else {
18363 		/* Avoid shrinking window with window scaling. */
18364 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
18365 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18366 	}
18367 	/*
18368 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18369 	 * window.  This may cause the remote transmitter to stall.  This
18370 	 * flag tells soreceive() to disable delayed acknowledgements when
18371 	 * draining the buffer.  This can occur if the receiver is
18372 	 * attempting to read more data than can be buffered prior to
18373 	 * transmitting on the connection.
18374 	 */
18375 	if (th->th_win == 0) {
18376 		tp->t_sndzerowin++;
18377 		tp->t_flags |= TF_RXWIN0SENT;
18378 	} else
18379 		tp->t_flags &= ~TF_RXWIN0SENT;
18380 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
18381 	/* Now are we using fsb?, if so copy the template data to the mbuf */
18382 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18383 		uint8_t *cpto;
18384 
18385 		cpto = mtod(m, uint8_t *);
18386 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18387 		/*
18388 		 * We have just copied in:
18389 		 * IP/IP6
18390 		 * <optional udphdr>
18391 		 * tcphdr (no options)
18392 		 *
18393 		 * We need to grab the correct pointers into the mbuf
18394 		 * for both the tcp header, and possibly the udp header (if tunneling).
18395 		 * We do this by using the offset in the copy buffer and adding it
18396 		 * to the mbuf base pointer (cpto).
18397 		 */
18398 #ifdef INET6
18399 		if (isipv6)
18400 			ip6 = mtod(m, struct ip6_hdr *);
18401 		else
18402 #endif				/* INET6 */
18403 			ip = mtod(m, struct ip *);
18404 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18405 		/* If we have a udp header lets set it into the mbuf as well */
18406 		if (udp)
18407 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18408 	}
18409 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18410 	if (to.to_flags & TOF_SIGNATURE) {
18411 		/*
18412 		 * Calculate MD5 signature and put it into the place
18413 		 * determined before.
18414 		 * NOTE: since TCP options buffer doesn't point into
18415 		 * mbuf's data, calculate offset and use it.
18416 		 */
18417 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18418 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18419 			/*
18420 			 * Do not send segment if the calculation of MD5
18421 			 * digest has failed.
18422 			 */
18423 			goto out;
18424 		}
18425 	}
18426 #endif
18427 	if (optlen) {
18428 		bcopy(opt, th + 1, optlen);
18429 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18430 	}
18431 	/*
18432 	 * Put TCP length in extended header, and then checksum extended
18433 	 * header and data.
18434 	 */
18435 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18436 #ifdef INET6
18437 	if (isipv6) {
18438 		/*
18439 		 * ip6_plen is not need to be filled now, and will be filled
18440 		 * in ip6_output.
18441 		 */
18442 		if (tp->t_port) {
18443 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18444 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18445 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18446 			th->th_sum = htons(0);
18447 			UDPSTAT_INC(udps_opackets);
18448 		} else {
18449 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18450 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18451 			th->th_sum = in6_cksum_pseudo(ip6,
18452 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18453 						      0);
18454 		}
18455 	}
18456 #endif
18457 #if defined(INET6) && defined(INET)
18458 	else
18459 #endif
18460 #ifdef INET
18461 	{
18462 		if (tp->t_port) {
18463 			m->m_pkthdr.csum_flags = CSUM_UDP;
18464 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18465 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18466 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18467 			th->th_sum = htons(0);
18468 			UDPSTAT_INC(udps_opackets);
18469 		} else {
18470 			m->m_pkthdr.csum_flags = CSUM_TCP;
18471 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18472 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18473 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18474 									IPPROTO_TCP + len + optlen));
18475 		}
18476 		/* IP version must be set here for ipv4/ipv6 checking later */
18477 		KASSERT(ip->ip_v == IPVERSION,
18478 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18479 	}
18480 #endif
18481 	/*
18482 	 * Enable TSO and specify the size of the segments. The TCP pseudo
18483 	 * header checksum is always provided. XXX: Fixme: This is currently
18484 	 * not the case for IPv6.
18485 	 */
18486 	if (tso) {
18487 		KASSERT(len > tp->t_maxseg - optlen,
18488 			("%s: len <= tso_segsz", __func__));
18489 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18490 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18491 	}
18492 	KASSERT(len + hdrlen == m_length(m, NULL),
18493 		("%s: mbuf chain different than expected: %d + %u != %u",
18494 		 __func__, len, hdrlen, m_length(m, NULL)));
18495 
18496 #ifdef TCP_HHOOK
18497 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18498 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
18499 #endif
18500 	/* We're getting ready to send; log now. */
18501 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18502 		union tcp_log_stackspecific log;
18503 
18504 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18505 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18506 		if (rack->rack_no_prr)
18507 			log.u_bbr.flex1 = 0;
18508 		else
18509 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18510 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18511 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18512 		log.u_bbr.flex4 = orig_len;
18513 		if (filled_all)
18514 			log.u_bbr.flex5 = 0x80000000;
18515 		else
18516 			log.u_bbr.flex5 = 0;
18517 		/* Save off the early/late values */
18518 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18519 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18520 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18521 		if (rsm || sack_rxmit) {
18522 			if (doing_tlp)
18523 				log.u_bbr.flex8 = 2;
18524 			else
18525 				log.u_bbr.flex8 = 1;
18526 		} else {
18527 			if (doing_tlp)
18528 				log.u_bbr.flex8 = 3;
18529 			else
18530 				log.u_bbr.flex8 = 0;
18531 		}
18532 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18533 		log.u_bbr.flex7 = mark;
18534 		log.u_bbr.flex7 <<= 8;
18535 		log.u_bbr.flex7 |= pass;
18536 		log.u_bbr.pkts_out = tp->t_maxseg;
18537 		log.u_bbr.timeStamp = cts;
18538 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18539 		log.u_bbr.lt_epoch = cwnd_to_use;
18540 		log.u_bbr.delivered = sendalot;
18541 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18542 				     len, &log, false, NULL, NULL, 0, &tv);
18543 	} else
18544 		lgb = NULL;
18545 
18546 	/*
18547 	 * Fill in IP length and desired time to live and send to IP level.
18548 	 * There should be a better way to handle ttl and tos; we could keep
18549 	 * them in the template, but need a way to checksum without them.
18550 	 */
18551 	/*
18552 	 * m->m_pkthdr.len should have been set before cksum calcuration,
18553 	 * because in6_cksum() need it.
18554 	 */
18555 #ifdef INET6
18556 	if (isipv6) {
18557 		/*
18558 		 * we separately set hoplimit for every segment, since the
18559 		 * user might want to change the value via setsockopt. Also,
18560 		 * desired default hop limit might be changed via Neighbor
18561 		 * Discovery.
18562 		 */
18563 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18564 
18565 		/*
18566 		 * Set the packet size here for the benefit of DTrace
18567 		 * probes. ip6_output() will set it properly; it's supposed
18568 		 * to include the option header lengths as well.
18569 		 */
18570 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18571 
18572 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18573 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18574 		else
18575 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18576 
18577 		if (tp->t_state == TCPS_SYN_SENT)
18578 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18579 
18580 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18581 		/* TODO: IPv6 IP6TOS_ECT bit on */
18582 		error = ip6_output(m,
18583 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18584 				   inp->in6p_outputopts,
18585 #else
18586 				   NULL,
18587 #endif
18588 				   &inp->inp_route6,
18589 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18590 				   NULL, NULL, inp);
18591 
18592 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18593 			mtu = inp->inp_route6.ro_nh->nh_mtu;
18594 	}
18595 #endif				/* INET6 */
18596 #if defined(INET) && defined(INET6)
18597 	else
18598 #endif
18599 #ifdef INET
18600 	{
18601 		ip->ip_len = htons(m->m_pkthdr.len);
18602 #ifdef INET6
18603 		if (inp->inp_vflag & INP_IPV6PROTO)
18604 			ip->ip_ttl = in6_selecthlim(inp, NULL);
18605 #endif				/* INET6 */
18606 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18607 		/*
18608 		 * If we do path MTU discovery, then we set DF on every
18609 		 * packet. This might not be the best thing to do according
18610 		 * to RFC3390 Section 2. However the tcp hostcache migitates
18611 		 * the problem so it affects only the first tcp connection
18612 		 * with a host.
18613 		 *
18614 		 * NB: Don't set DF on small MTU/MSS to have a safe
18615 		 * fallback.
18616 		 */
18617 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18618 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18619 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18620 				ip->ip_off |= htons(IP_DF);
18621 			}
18622 		} else {
18623 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18624 		}
18625 
18626 		if (tp->t_state == TCPS_SYN_SENT)
18627 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18628 
18629 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
18630 
18631 		error = ip_output(m,
18632 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18633 				  inp->inp_options,
18634 #else
18635 				  NULL,
18636 #endif
18637 				  &inp->inp_route,
18638 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18639 				  inp);
18640 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18641 			mtu = inp->inp_route.ro_nh->nh_mtu;
18642 	}
18643 #endif				/* INET */
18644 
18645 out:
18646 	if (lgb) {
18647 		lgb->tlb_errno = error;
18648 		lgb = NULL;
18649 	}
18650 	/*
18651 	 * In transmit state, time the transmission and arrange for the
18652 	 * retransmit.  In persist state, just set snd_max.
18653 	 */
18654 	if (error == 0) {
18655 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18656 		if (rsm && doing_tlp) {
18657 			rack->rc_last_sent_tlp_past_cumack = 0;
18658 			rack->rc_last_sent_tlp_seq_valid = 1;
18659 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18660 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18661 		}
18662 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
18663 		if (rsm && (doing_tlp == 0)) {
18664 			/* Set we retransmitted */
18665 			rack->rc_gp_saw_rec = 1;
18666 		} else {
18667 			if (cwnd_to_use > tp->snd_ssthresh) {
18668 				/* Set we sent in CA */
18669 				rack->rc_gp_saw_ca = 1;
18670 			} else {
18671 				/* Set we sent in SS */
18672 				rack->rc_gp_saw_ss = 1;
18673 			}
18674 		}
18675 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18676 		    (tp->t_flags & TF_SACK_PERMIT) &&
18677 		    tp->rcv_numsacks > 0)
18678 			tcp_clean_dsack_blocks(tp);
18679 		tot_len_this_send += len;
18680 		if (len == 0)
18681 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18682 		else if (len == 1) {
18683 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18684 		} else if (len > 1) {
18685 			int idx;
18686 
18687 			idx = (len / segsiz) + 3;
18688 			if (idx >= TCP_MSS_ACCT_ATIMER)
18689 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18690 			else
18691 				counter_u64_add(rack_out_size[idx], 1);
18692 		}
18693 	}
18694 	if ((rack->rack_no_prr == 0) &&
18695 	    sub_from_prr &&
18696 	    (error == 0)) {
18697 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18698 			rack->r_ctl.rc_prr_sndcnt -= len;
18699 		else
18700 			rack->r_ctl.rc_prr_sndcnt = 0;
18701 	}
18702 	sub_from_prr = 0;
18703 	if (doing_tlp) {
18704 		/* Make sure the TLP is added */
18705 		add_flag |= RACK_TLP;
18706 	} else if (rsm) {
18707 		/* If its a resend without TLP then it must not have the flag */
18708 		rsm->r_flags &= ~RACK_TLP;
18709 	}
18710 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18711 			rack_to_usec_ts(&tv),
18712 			rsm, add_flag, s_mb, s_moff, hw_tls);
18713 
18714 
18715 	if ((error == 0) &&
18716 	    (len > 0) &&
18717 	    (tp->snd_una == tp->snd_max))
18718 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
18719 	{
18720 		tcp_seq startseq = tp->snd_nxt;
18721 
18722 		/* Track our lost count */
18723 		if (rsm && (doing_tlp == 0))
18724 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18725 		/*
18726 		 * Advance snd_nxt over sequence space of this segment.
18727 		 */
18728 		if (error)
18729 			/* We don't log or do anything with errors */
18730 			goto nomore;
18731 		if (doing_tlp == 0) {
18732 			if (rsm == NULL) {
18733 				/*
18734 				 * Not a retransmission of some
18735 				 * sort, new data is going out so
18736 				 * clear our TLP count and flag.
18737 				 */
18738 				rack->rc_tlp_in_progress = 0;
18739 				rack->r_ctl.rc_tlp_cnt_out = 0;
18740 			}
18741 		} else {
18742 			/*
18743 			 * We have just sent a TLP, mark that it is true
18744 			 * and make sure our in progress is set so we
18745 			 * continue to check the count.
18746 			 */
18747 			rack->rc_tlp_in_progress = 1;
18748 			rack->r_ctl.rc_tlp_cnt_out++;
18749 		}
18750 		if (flags & (TH_SYN | TH_FIN)) {
18751 			if (flags & TH_SYN)
18752 				tp->snd_nxt++;
18753 			if (flags & TH_FIN) {
18754 				tp->snd_nxt++;
18755 				tp->t_flags |= TF_SENTFIN;
18756 			}
18757 		}
18758 		/* In the ENOBUFS case we do *not* update snd_max */
18759 		if (sack_rxmit)
18760 			goto nomore;
18761 
18762 		tp->snd_nxt += len;
18763 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18764 			if (tp->snd_una == tp->snd_max) {
18765 				/*
18766 				 * Update the time we just added data since
18767 				 * none was outstanding.
18768 				 */
18769 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18770 				tp->t_acktime = ticks;
18771 			}
18772 			tp->snd_max = tp->snd_nxt;
18773 			/*
18774 			 * Time this transmission if not a retransmission and
18775 			 * not currently timing anything.
18776 			 * This is only relevant in case of switching back to
18777 			 * the base stack.
18778 			 */
18779 			if (tp->t_rtttime == 0) {
18780 				tp->t_rtttime = ticks;
18781 				tp->t_rtseq = startseq;
18782 				KMOD_TCPSTAT_INC(tcps_segstimed);
18783 			}
18784 			if (len &&
18785 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
18786 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18787 		}
18788 		/*
18789 		 * If we are doing FO we need to update the mbuf position and subtract
18790 		 * this happens when the peer sends us duplicate information and
18791 		 * we thus want to send a DSACK.
18792 		 *
18793 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18794 		 * turned off? If not then we are going to echo multiple DSACK blocks
18795 		 * out (with the TSO), which we should not be doing.
18796 		 */
18797 		if (rack->r_fast_output && len) {
18798 			if (rack->r_ctl.fsb.left_to_send > len)
18799 				rack->r_ctl.fsb.left_to_send -= len;
18800 			else
18801 				rack->r_ctl.fsb.left_to_send = 0;
18802 			if (rack->r_ctl.fsb.left_to_send < segsiz)
18803 				rack->r_fast_output = 0;
18804 			if (rack->r_fast_output) {
18805 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18806 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18807 			}
18808 		}
18809 	}
18810 nomore:
18811 	if (error) {
18812 		rack->r_ctl.rc_agg_delayed = 0;
18813 		rack->r_early = 0;
18814 		rack->r_late = 0;
18815 		rack->r_ctl.rc_agg_early = 0;
18816 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
18817 		/*
18818 		 * Failures do not advance the seq counter above. For the
18819 		 * case of ENOBUFS we will fall out and retry in 1ms with
18820 		 * the hpts. Everything else will just have to retransmit
18821 		 * with the timer.
18822 		 *
18823 		 * In any case, we do not want to loop around for another
18824 		 * send without a good reason.
18825 		 */
18826 		sendalot = 0;
18827 		switch (error) {
18828 		case EPERM:
18829 			tp->t_softerror = error;
18830 #ifdef TCP_ACCOUNTING
18831 			crtsc = get_cyclecount();
18832 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18833 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18834 			}
18835 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18836 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18837 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18838 			}
18839 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18840 			sched_unpin();
18841 #endif
18842 			return (error);
18843 		case ENOBUFS:
18844 			/*
18845 			 * Pace us right away to retry in a some
18846 			 * time
18847 			 */
18848 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18849 			if (rack->rc_enobuf < 0x7f)
18850 				rack->rc_enobuf++;
18851 			if (slot < (10 * HPTS_USEC_IN_MSEC))
18852 				slot = 10 * HPTS_USEC_IN_MSEC;
18853 			if (rack->r_ctl.crte != NULL) {
18854 				counter_u64_add(rack_saw_enobuf_hw, 1);
18855 				tcp_rl_log_enobuf(rack->r_ctl.crte);
18856 			}
18857 			counter_u64_add(rack_saw_enobuf, 1);
18858 			goto enobufs;
18859 		case EMSGSIZE:
18860 			/*
18861 			 * For some reason the interface we used initially
18862 			 * to send segments changed to another or lowered
18863 			 * its MTU. If TSO was active we either got an
18864 			 * interface without TSO capabilits or TSO was
18865 			 * turned off. If we obtained mtu from ip_output()
18866 			 * then update it and try again.
18867 			 */
18868 			if (tso)
18869 				tp->t_flags &= ~TF_TSO;
18870 			if (mtu != 0) {
18871 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
18872 				goto again;
18873 			}
18874 			slot = 10 * HPTS_USEC_IN_MSEC;
18875 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18876 #ifdef TCP_ACCOUNTING
18877 			crtsc = get_cyclecount();
18878 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18879 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18880 			}
18881 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18882 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18883 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18884 			}
18885 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18886 			sched_unpin();
18887 #endif
18888 			return (error);
18889 		case ENETUNREACH:
18890 			counter_u64_add(rack_saw_enetunreach, 1);
18891 		case EHOSTDOWN:
18892 		case EHOSTUNREACH:
18893 		case ENETDOWN:
18894 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
18895 				tp->t_softerror = error;
18896 			}
18897 			/* FALLTHROUGH */
18898 		default:
18899 			slot = 10 * HPTS_USEC_IN_MSEC;
18900 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18901 #ifdef TCP_ACCOUNTING
18902 			crtsc = get_cyclecount();
18903 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18904 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18905 			}
18906 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18907 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18908 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18909 			}
18910 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18911 			sched_unpin();
18912 #endif
18913 			return (error);
18914 		}
18915 	} else {
18916 		rack->rc_enobuf = 0;
18917 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18918 			rack->r_ctl.retran_during_recovery += len;
18919 	}
18920 	KMOD_TCPSTAT_INC(tcps_sndtotal);
18921 
18922 	/*
18923 	 * Data sent (as far as we can tell). If this advertises a larger
18924 	 * window than any other segment, then remember the size of the
18925 	 * advertised window. Any pending ACK has now been sent.
18926 	 */
18927 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18928 		tp->rcv_adv = tp->rcv_nxt + recwin;
18929 
18930 	tp->last_ack_sent = tp->rcv_nxt;
18931 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18932 enobufs:
18933 	if (sendalot) {
18934 		/* Do we need to turn off sendalot? */
18935 		if (rack->r_ctl.rc_pace_max_segs &&
18936 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18937 			/* We hit our max. */
18938 			sendalot = 0;
18939 		} else if ((rack->rc_user_set_max_segs) &&
18940 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18941 			/* We hit the user defined max */
18942 			sendalot = 0;
18943 		}
18944 	}
18945 	if ((error == 0) && (flags & TH_FIN))
18946 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18947 	if (flags & TH_RST) {
18948 		/*
18949 		 * We don't send again after sending a RST.
18950 		 */
18951 		slot = 0;
18952 		sendalot = 0;
18953 		if (error == 0)
18954 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18955 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18956 		/*
18957 		 * Get our pacing rate, if an error
18958 		 * occurred in sending (ENOBUF) we would
18959 		 * hit the else if with slot preset. Other
18960 		 * errors return.
18961 		 */
18962 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18963 	}
18964 	if (rsm &&
18965 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18966 	    rack->use_rack_rr) {
18967 		/* Its a retransmit and we use the rack cheat? */
18968 		if ((slot == 0) ||
18969 		    (rack->rc_always_pace == 0) ||
18970 		    (rack->r_rr_config == 1)) {
18971 			/*
18972 			 * We have no pacing set or we
18973 			 * are using old-style rack or
18974 			 * we are overriden to use the old 1ms pacing.
18975 			 */
18976 			slot = rack->r_ctl.rc_min_to;
18977 		}
18978 	}
18979 	/* We have sent clear the flag */
18980 	rack->r_ent_rec_ns = 0;
18981 	if (rack->r_must_retran) {
18982 		if (rsm) {
18983 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
18984 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
18985 				/*
18986 				 * We have retransmitted all.
18987 				 */
18988 				rack->r_must_retran = 0;
18989 				rack->r_ctl.rc_out_at_rto = 0;
18990 			}
18991 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18992 			/*
18993 			 * Sending new data will also kill
18994 			 * the loop.
18995 			 */
18996 			rack->r_must_retran = 0;
18997 			rack->r_ctl.rc_out_at_rto = 0;
18998 		}
18999 	}
19000 	rack->r_ctl.fsb.recwin = recwin;
19001 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19002 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19003 		/*
19004 		 * We hit an RTO and now have past snd_max at the RTO
19005 		 * clear all the WAS flags.
19006 		 */
19007 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19008 	}
19009 	if (slot) {
19010 		/* set the rack tcb into the slot N */
19011 		if ((error == 0) &&
19012 		    rack_use_rfo &&
19013 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19014 		    (rsm == NULL) &&
19015 		    (tp->snd_nxt == tp->snd_max) &&
19016 		    (ipoptlen == 0) &&
19017 		    (tp->rcv_numsacks == 0) &&
19018 		    rack->r_fsb_inited &&
19019 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19020 		    (rack->r_must_retran == 0) &&
19021 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19022 		    (len > 0) && (orig_len > 0) &&
19023 		    (orig_len > len) &&
19024 		    ((orig_len - len) >= segsiz) &&
19025 		    ((optlen == 0) ||
19026 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19027 			/* We can send at least one more MSS using our fsb */
19028 
19029 			rack->r_fast_output = 1;
19030 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19031 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19032 			rack->r_ctl.fsb.tcp_flags = flags;
19033 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19034 			if (hw_tls)
19035 				rack->r_ctl.fsb.hw_tls = 1;
19036 			else
19037 				rack->r_ctl.fsb.hw_tls = 0;
19038 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19039 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19040 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19041 				 (tp->snd_max - tp->snd_una)));
19042 			if (rack->r_ctl.fsb.left_to_send < segsiz)
19043 				rack->r_fast_output = 0;
19044 			else {
19045 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19046 					rack->r_ctl.fsb.rfo_apply_push = 1;
19047 				else
19048 					rack->r_ctl.fsb.rfo_apply_push = 0;
19049 			}
19050 		} else
19051 			rack->r_fast_output = 0;
19052 		rack_log_fsb(rack, tp, so, flags,
19053 			     ipoptlen, orig_len, len, error,
19054 			     (rsm == NULL), optlen, __LINE__, 2);
19055 	} else if (sendalot) {
19056 		int ret;
19057 
19058 		sack_rxmit = 0;
19059 		if ((error == 0) &&
19060 		    rack_use_rfo &&
19061 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19062 		    (rsm == NULL) &&
19063 		    (ipoptlen == 0) &&
19064 		    (tp->rcv_numsacks == 0) &&
19065 		    (tp->snd_nxt == tp->snd_max) &&
19066 		    (rack->r_must_retran == 0) &&
19067 		    rack->r_fsb_inited &&
19068 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19069 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19070 		    (len > 0) && (orig_len > 0) &&
19071 		    (orig_len > len) &&
19072 		    ((orig_len - len) >= segsiz) &&
19073 		    ((optlen == 0) ||
19074 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19075 			/* we can use fast_output for more */
19076 
19077 			rack->r_fast_output = 1;
19078 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19079 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19080 			rack->r_ctl.fsb.tcp_flags = flags;
19081 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19082 			if (hw_tls)
19083 				rack->r_ctl.fsb.hw_tls = 1;
19084 			else
19085 				rack->r_ctl.fsb.hw_tls = 0;
19086 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19087 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19088 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19089 				 (tp->snd_max - tp->snd_una)));
19090 			if (rack->r_ctl.fsb.left_to_send < segsiz) {
19091 				rack->r_fast_output = 0;
19092 			}
19093 			if (rack->r_fast_output) {
19094 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19095 					rack->r_ctl.fsb.rfo_apply_push = 1;
19096 				else
19097 					rack->r_ctl.fsb.rfo_apply_push = 0;
19098 				rack_log_fsb(rack, tp, so, flags,
19099 					     ipoptlen, orig_len, len, error,
19100 					     (rsm == NULL), optlen, __LINE__, 3);
19101 				error = 0;
19102 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19103 				if (ret >= 0)
19104 					return (ret);
19105 			        else if (error)
19106 					goto nomore;
19107 
19108 			}
19109 		}
19110 		goto again;
19111 	}
19112 	/* Assure when we leave that snd_nxt will point to top */
19113 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19114 		tp->snd_nxt = tp->snd_max;
19115 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19116 #ifdef TCP_ACCOUNTING
19117 	crtsc = get_cyclecount() - ts_val;
19118 	if (tot_len_this_send) {
19119 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19120 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
19121 		}
19122 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19123 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19124 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19125 		}
19126 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19127 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19128 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19129 		}
19130 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19131 	} else {
19132 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19133 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
19134 		}
19135 		counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19136 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19137 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19138 		}
19139 		counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19140 	}
19141 	sched_unpin();
19142 #endif
19143 	if (error == ENOBUFS)
19144 		error = 0;
19145 	return (error);
19146 }
19147 
19148 static void
19149 rack_update_seg(struct tcp_rack *rack)
19150 {
19151 	uint32_t orig_val;
19152 
19153 	orig_val = rack->r_ctl.rc_pace_max_segs;
19154 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19155 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
19156 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19157 }
19158 
19159 static void
19160 rack_mtu_change(struct tcpcb *tp)
19161 {
19162 	/*
19163 	 * The MSS may have changed
19164 	 */
19165 	struct tcp_rack *rack;
19166 	struct rack_sendmap *rsm;
19167 
19168 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19169 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19170 		/*
19171 		 * The MTU has changed we need to resend everything
19172 		 * since all we have sent is lost. We first fix
19173 		 * up the mtu though.
19174 		 */
19175 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19176 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
19177 		rack_remxt_tmr(tp);
19178 		rack->r_fast_output = 0;
19179 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19180 						rack->r_ctl.rc_sacked);
19181 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19182 		rack->r_must_retran = 1;
19183 		/* Mark all inflight to needing to be rxt'd */
19184 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19185 			rsm->r_flags |= RACK_MUST_RXT;
19186 		}
19187 	}
19188 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19189 	/* We don't use snd_nxt to retransmit */
19190 	tp->snd_nxt = tp->snd_max;
19191 }
19192 
19193 static int
19194 rack_set_profile(struct tcp_rack *rack, int prof)
19195 {
19196 	int err = EINVAL;
19197 	if (prof == 1) {
19198 		/* pace_always=1 */
19199 		if (rack->rc_always_pace == 0) {
19200 			if (tcp_can_enable_pacing() == 0)
19201 				return (EBUSY);
19202 		}
19203 		rack->rc_always_pace = 1;
19204 		if (rack->use_fixed_rate || rack->gp_ready)
19205 			rack_set_cc_pacing(rack);
19206 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19207 		rack->rack_attempt_hdwr_pace = 0;
19208 		/* cmpack=1 */
19209 		if (rack_use_cmp_acks)
19210 			rack->r_use_cmp_ack = 1;
19211 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19212 		    rack->r_use_cmp_ack)
19213 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19214 		/* scwnd=1 */
19215 		rack->rack_enable_scwnd = 1;
19216 		/* dynamic=100 */
19217 		rack->rc_gp_dyn_mul = 1;
19218 		/* gp_inc_ca */
19219 		rack->r_ctl.rack_per_of_gp_ca = 100;
19220 		/* rrr_conf=3 */
19221 		rack->r_rr_config = 3;
19222 		/* npush=2 */
19223 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19224 		/* fillcw=1 */
19225 		rack->rc_pace_to_cwnd = 1;
19226 		rack->rc_pace_fill_if_rttin_range = 0;
19227 		rack->rtt_limit_mul = 0;
19228 		/* noprr=1 */
19229 		rack->rack_no_prr = 1;
19230 		/* lscwnd=1 */
19231 		rack->r_limit_scw = 1;
19232 		/* gp_inc_rec */
19233 		rack->r_ctl.rack_per_of_gp_rec = 90;
19234 		err = 0;
19235 
19236 	} else if (prof == 3) {
19237 		/* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19238 		/* pace_always=1 */
19239 		if (rack->rc_always_pace == 0) {
19240 			if (tcp_can_enable_pacing() == 0)
19241 				return (EBUSY);
19242 		}
19243 		rack->rc_always_pace = 1;
19244 		if (rack->use_fixed_rate || rack->gp_ready)
19245 			rack_set_cc_pacing(rack);
19246 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19247 		rack->rack_attempt_hdwr_pace = 0;
19248 		/* cmpack=1 */
19249 		if (rack_use_cmp_acks)
19250 			rack->r_use_cmp_ack = 1;
19251 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19252 		    rack->r_use_cmp_ack)
19253 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19254 		/* scwnd=1 */
19255 		rack->rack_enable_scwnd = 1;
19256 		/* dynamic=100 */
19257 		rack->rc_gp_dyn_mul = 1;
19258 		/* gp_inc_ca */
19259 		rack->r_ctl.rack_per_of_gp_ca = 100;
19260 		/* rrr_conf=3 */
19261 		rack->r_rr_config = 3;
19262 		/* npush=2 */
19263 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19264 		/* fillcw=2 */
19265 		rack->rc_pace_to_cwnd = 1;
19266 		rack->r_fill_less_agg = 1;
19267 		rack->rc_pace_fill_if_rttin_range = 0;
19268 		rack->rtt_limit_mul = 0;
19269 		/* noprr=1 */
19270 		rack->rack_no_prr = 1;
19271 		/* lscwnd=1 */
19272 		rack->r_limit_scw = 1;
19273 		/* gp_inc_rec */
19274 		rack->r_ctl.rack_per_of_gp_rec = 90;
19275 		err = 0;
19276 
19277 
19278 	} else if (prof == 2) {
19279 		/* cmpack=1 */
19280 		if (rack->rc_always_pace == 0) {
19281 			if (tcp_can_enable_pacing() == 0)
19282 				return (EBUSY);
19283 		}
19284 		rack->rc_always_pace = 1;
19285 		if (rack->use_fixed_rate || rack->gp_ready)
19286 			rack_set_cc_pacing(rack);
19287 		rack->r_use_cmp_ack = 1;
19288 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19289 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19290 		/* pace_always=1 */
19291 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19292 		/* scwnd=1 */
19293 		rack->rack_enable_scwnd = 1;
19294 		/* dynamic=100 */
19295 		rack->rc_gp_dyn_mul = 1;
19296 		rack->r_ctl.rack_per_of_gp_ca = 100;
19297 		/* rrr_conf=3 */
19298 		rack->r_rr_config = 3;
19299 		/* npush=2 */
19300 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19301 		/* fillcw=1 */
19302 		rack->rc_pace_to_cwnd = 1;
19303 		rack->rc_pace_fill_if_rttin_range = 0;
19304 		rack->rtt_limit_mul = 0;
19305 		/* noprr=1 */
19306 		rack->rack_no_prr = 1;
19307 		/* lscwnd=0 */
19308 		rack->r_limit_scw = 0;
19309 		err = 0;
19310 	} else if (prof == 0) {
19311 		/* This changes things back to the default settings */
19312 		err = 0;
19313 		if (rack->rc_always_pace) {
19314 			tcp_decrement_paced_conn();
19315 			rack_undo_cc_pacing(rack);
19316 			rack->rc_always_pace = 0;
19317 		}
19318 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19319 			rack->rc_always_pace = 1;
19320 			if (rack->use_fixed_rate || rack->gp_ready)
19321 				rack_set_cc_pacing(rack);
19322 		} else
19323 			rack->rc_always_pace = 0;
19324 		if (rack_dsack_std_based & 0x1) {
19325 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19326 			rack->rc_rack_tmr_std_based = 1;
19327 		}
19328 		if (rack_dsack_std_based & 0x2) {
19329 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19330 			rack->rc_rack_use_dsack = 1;
19331 		}
19332 		if (rack_use_cmp_acks)
19333 			rack->r_use_cmp_ack = 1;
19334 		else
19335 			rack->r_use_cmp_ack = 0;
19336 		if (rack_disable_prr)
19337 			rack->rack_no_prr = 1;
19338 		else
19339 			rack->rack_no_prr = 0;
19340 		if (rack_gp_no_rec_chg)
19341 			rack->rc_gp_no_rec_chg = 1;
19342 		else
19343 			rack->rc_gp_no_rec_chg = 0;
19344 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19345 			rack->r_mbuf_queue = 1;
19346 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19347 				rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19348 			rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19349 		} else {
19350 			rack->r_mbuf_queue = 0;
19351 			rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19352 		}
19353 		if (rack_enable_shared_cwnd)
19354 			rack->rack_enable_scwnd = 1;
19355 		else
19356 			rack->rack_enable_scwnd = 0;
19357 		if (rack_do_dyn_mul) {
19358 			/* When dynamic adjustment is on CA needs to start at 100% */
19359 			rack->rc_gp_dyn_mul = 1;
19360 			if (rack_do_dyn_mul >= 100)
19361 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19362 		} else {
19363 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19364 			rack->rc_gp_dyn_mul = 0;
19365 		}
19366 		rack->r_rr_config = 0;
19367 		rack->r_ctl.rc_no_push_at_mrtt = 0;
19368 		rack->rc_pace_to_cwnd = 0;
19369 		rack->rc_pace_fill_if_rttin_range = 0;
19370 		rack->rtt_limit_mul = 0;
19371 
19372 		if (rack_enable_hw_pacing)
19373 			rack->rack_hdw_pace_ena = 1;
19374 		else
19375 			rack->rack_hdw_pace_ena = 0;
19376 		if (rack_disable_prr)
19377 			rack->rack_no_prr = 1;
19378 		else
19379 			rack->rack_no_prr = 0;
19380 		if (rack_limits_scwnd)
19381 			rack->r_limit_scw  = 1;
19382 		else
19383 			rack->r_limit_scw  = 0;
19384 		err = 0;
19385 	}
19386 	return (err);
19387 }
19388 
19389 static int
19390 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19391 {
19392 	struct deferred_opt_list *dol;
19393 
19394 	dol = malloc(sizeof(struct deferred_opt_list),
19395 		     M_TCPFSB, M_NOWAIT|M_ZERO);
19396 	if (dol == NULL) {
19397 		/*
19398 		 * No space yikes -- fail out..
19399 		 */
19400 		return (0);
19401 	}
19402 	dol->optname = sopt_name;
19403 	dol->optval = loptval;
19404 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19405 	return (1);
19406 }
19407 
19408 static int
19409 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19410 		    uint32_t optval, uint64_t loptval)
19411 {
19412 	struct epoch_tracker et;
19413 	struct sockopt sopt;
19414 	struct cc_newreno_opts opt;
19415 	uint64_t val;
19416 	int error = 0;
19417 	uint16_t ca, ss;
19418 
19419 	switch (sopt_name) {
19420 
19421 	case TCP_RACK_DSACK_OPT:
19422 		RACK_OPTS_INC(tcp_rack_dsack_opt);
19423 		if (optval & 0x1) {
19424 			rack->rc_rack_tmr_std_based = 1;
19425 		} else {
19426 			rack->rc_rack_tmr_std_based = 0;
19427 		}
19428 		if (optval & 0x2) {
19429 			rack->rc_rack_use_dsack = 1;
19430 		} else {
19431 			rack->rc_rack_use_dsack = 0;
19432 		}
19433 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19434 		break;
19435 	case TCP_RACK_PACING_BETA:
19436 		RACK_OPTS_INC(tcp_rack_beta);
19437 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19438 			/* This only works for newreno. */
19439 			error = EINVAL;
19440 			break;
19441 		}
19442 		if (rack->rc_pacing_cc_set) {
19443 			/*
19444 			 * Set them into the real CC module
19445 			 * whats in the rack pcb is the old values
19446 			 * to be used on restoral/
19447 			 */
19448 			sopt.sopt_dir = SOPT_SET;
19449 			opt.name = CC_NEWRENO_BETA;
19450 			opt.val = optval;
19451 			if (CC_ALGO(tp)->ctl_output != NULL)
19452 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19453 			else {
19454 				error = ENOENT;
19455 				break;
19456 			}
19457 		} else {
19458 			/*
19459 			 * Not pacing yet so set it into our local
19460 			 * rack pcb storage.
19461 			 */
19462 			rack->r_ctl.rc_saved_beta.beta = optval;
19463 		}
19464 		break;
19465 	case TCP_RACK_TIMER_SLOP:
19466 		RACK_OPTS_INC(tcp_rack_timer_slop);
19467 		rack->r_ctl.timer_slop = optval;
19468 		if (rack->rc_tp->t_srtt) {
19469 			/*
19470 			 * If we have an SRTT lets update t_rxtcur
19471 			 * to have the new slop.
19472 			 */
19473 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19474 					   rack_rto_min, rack_rto_max,
19475 					   rack->r_ctl.timer_slop);
19476 		}
19477 		break;
19478 	case TCP_RACK_PACING_BETA_ECN:
19479 		RACK_OPTS_INC(tcp_rack_beta_ecn);
19480 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19481 			/* This only works for newreno. */
19482 			error = EINVAL;
19483 			break;
19484 		}
19485 		if (rack->rc_pacing_cc_set) {
19486 			/*
19487 			 * Set them into the real CC module
19488 			 * whats in the rack pcb is the old values
19489 			 * to be used on restoral/
19490 			 */
19491 			sopt.sopt_dir = SOPT_SET;
19492 			opt.name = CC_NEWRENO_BETA_ECN;
19493 			opt.val = optval;
19494 			if (CC_ALGO(tp)->ctl_output != NULL)
19495 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19496 			else
19497 				error = ENOENT;
19498 		} else {
19499 			/*
19500 			 * Not pacing yet so set it into our local
19501 			 * rack pcb storage.
19502 			 */
19503 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19504 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19505 		}
19506 		break;
19507 	case TCP_DEFER_OPTIONS:
19508 		RACK_OPTS_INC(tcp_defer_opt);
19509 		if (optval) {
19510 			if (rack->gp_ready) {
19511 				/* Too late */
19512 				error = EINVAL;
19513 				break;
19514 			}
19515 			rack->defer_options = 1;
19516 		} else
19517 			rack->defer_options = 0;
19518 		break;
19519 	case TCP_RACK_MEASURE_CNT:
19520 		RACK_OPTS_INC(tcp_rack_measure_cnt);
19521 		if (optval && (optval <= 0xff)) {
19522 			rack->r_ctl.req_measurements = optval;
19523 		} else
19524 			error = EINVAL;
19525 		break;
19526 	case TCP_REC_ABC_VAL:
19527 		RACK_OPTS_INC(tcp_rec_abc_val);
19528 		if (optval > 0)
19529 			rack->r_use_labc_for_rec = 1;
19530 		else
19531 			rack->r_use_labc_for_rec = 0;
19532 		break;
19533 	case TCP_RACK_ABC_VAL:
19534 		RACK_OPTS_INC(tcp_rack_abc_val);
19535 		if ((optval > 0) && (optval < 255))
19536 			rack->rc_labc = optval;
19537 		else
19538 			error = EINVAL;
19539 		break;
19540 	case TCP_HDWR_UP_ONLY:
19541 		RACK_OPTS_INC(tcp_pacing_up_only);
19542 		if (optval)
19543 			rack->r_up_only = 1;
19544 		else
19545 			rack->r_up_only = 0;
19546 		break;
19547 	case TCP_PACING_RATE_CAP:
19548 		RACK_OPTS_INC(tcp_pacing_rate_cap);
19549 		rack->r_ctl.bw_rate_cap = loptval;
19550 		break;
19551 	case TCP_RACK_PROFILE:
19552 		RACK_OPTS_INC(tcp_profile);
19553 		error = rack_set_profile(rack, optval);
19554 		break;
19555 	case TCP_USE_CMP_ACKS:
19556 		RACK_OPTS_INC(tcp_use_cmp_acks);
19557 		if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19558 			/* You can't turn it off once its on! */
19559 			error = EINVAL;
19560 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19561 			rack->r_use_cmp_ack = 1;
19562 			rack->r_mbuf_queue = 1;
19563 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19564 		}
19565 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19566 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19567 		break;
19568 	case TCP_SHARED_CWND_TIME_LIMIT:
19569 		RACK_OPTS_INC(tcp_lscwnd);
19570 		if (optval)
19571 			rack->r_limit_scw = 1;
19572 		else
19573 			rack->r_limit_scw = 0;
19574 		break;
19575  	case TCP_RACK_PACE_TO_FILL:
19576 		RACK_OPTS_INC(tcp_fillcw);
19577 		if (optval == 0)
19578 			rack->rc_pace_to_cwnd = 0;
19579 		else {
19580 			rack->rc_pace_to_cwnd = 1;
19581 			if (optval > 1)
19582 				rack->r_fill_less_agg = 1;
19583 		}
19584 		if ((optval >= rack_gp_rtt_maxmul) &&
19585 		    rack_gp_rtt_maxmul &&
19586 		    (optval < 0xf)) {
19587 			rack->rc_pace_fill_if_rttin_range = 1;
19588 			rack->rtt_limit_mul = optval;
19589 		} else {
19590 			rack->rc_pace_fill_if_rttin_range = 0;
19591 			rack->rtt_limit_mul = 0;
19592 		}
19593 		break;
19594 	case TCP_RACK_NO_PUSH_AT_MAX:
19595 		RACK_OPTS_INC(tcp_npush);
19596 		if (optval == 0)
19597 			rack->r_ctl.rc_no_push_at_mrtt = 0;
19598 		else if (optval < 0xff)
19599 			rack->r_ctl.rc_no_push_at_mrtt = optval;
19600 		else
19601 			error = EINVAL;
19602 		break;
19603 	case TCP_SHARED_CWND_ENABLE:
19604 		RACK_OPTS_INC(tcp_rack_scwnd);
19605 		if (optval == 0)
19606 			rack->rack_enable_scwnd = 0;
19607 		else
19608 			rack->rack_enable_scwnd = 1;
19609 		break;
19610 	case TCP_RACK_MBUF_QUEUE:
19611 		/* Now do we use the LRO mbuf-queue feature */
19612 		RACK_OPTS_INC(tcp_rack_mbufq);
19613 		if (optval || rack->r_use_cmp_ack)
19614 			rack->r_mbuf_queue = 1;
19615 		else
19616 			rack->r_mbuf_queue = 0;
19617 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19618 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19619 		else
19620 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19621 		break;
19622 	case TCP_RACK_NONRXT_CFG_RATE:
19623 		RACK_OPTS_INC(tcp_rack_cfg_rate);
19624 		if (optval == 0)
19625 			rack->rack_rec_nonrxt_use_cr = 0;
19626 		else
19627 			rack->rack_rec_nonrxt_use_cr = 1;
19628 		break;
19629 	case TCP_NO_PRR:
19630 		RACK_OPTS_INC(tcp_rack_noprr);
19631 		if (optval == 0)
19632 			rack->rack_no_prr = 0;
19633 		else if (optval == 1)
19634 			rack->rack_no_prr = 1;
19635 		else if (optval == 2)
19636 			rack->no_prr_addback = 1;
19637 		else
19638 			error = EINVAL;
19639 		break;
19640 	case TCP_TIMELY_DYN_ADJ:
19641 		RACK_OPTS_INC(tcp_timely_dyn);
19642 		if (optval == 0)
19643 			rack->rc_gp_dyn_mul = 0;
19644 		else {
19645 			rack->rc_gp_dyn_mul = 1;
19646 			if (optval >= 100) {
19647 				/*
19648 				 * If the user sets something 100 or more
19649 				 * its the gp_ca value.
19650 				 */
19651 				rack->r_ctl.rack_per_of_gp_ca  = optval;
19652 			}
19653 		}
19654 		break;
19655 	case TCP_RACK_DO_DETECTION:
19656 		RACK_OPTS_INC(tcp_rack_do_detection);
19657 		if (optval == 0)
19658 			rack->do_detection = 0;
19659 		else
19660 			rack->do_detection = 1;
19661 		break;
19662 	case TCP_RACK_TLP_USE:
19663 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19664 			error = EINVAL;
19665 			break;
19666 		}
19667 		RACK_OPTS_INC(tcp_tlp_use);
19668 		rack->rack_tlp_threshold_use = optval;
19669 		break;
19670 	case TCP_RACK_TLP_REDUCE:
19671 		/* RACK TLP cwnd reduction (bool) */
19672 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
19673 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19674 		break;
19675 	/*  Pacing related ones */
19676 	case TCP_RACK_PACE_ALWAYS:
19677 		/*
19678 		 * zero is old rack method, 1 is new
19679 		 * method using a pacing rate.
19680 		 */
19681 		RACK_OPTS_INC(tcp_rack_pace_always);
19682 		if (optval > 0) {
19683 			if (rack->rc_always_pace) {
19684 				error = EALREADY;
19685 				break;
19686 			} else if (tcp_can_enable_pacing()) {
19687 				rack->rc_always_pace = 1;
19688 				if (rack->use_fixed_rate || rack->gp_ready)
19689 					rack_set_cc_pacing(rack);
19690 			}
19691 			else {
19692 				error = ENOSPC;
19693 				break;
19694 			}
19695 		} else {
19696 			if (rack->rc_always_pace) {
19697 				tcp_decrement_paced_conn();
19698 				rack->rc_always_pace = 0;
19699 				rack_undo_cc_pacing(rack);
19700 			}
19701 		}
19702 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19703 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19704 		else
19705 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19706 		/* A rate may be set irate or other, if so set seg size */
19707 		rack_update_seg(rack);
19708 		break;
19709 	case TCP_BBR_RACK_INIT_RATE:
19710 		RACK_OPTS_INC(tcp_initial_rate);
19711 		val = optval;
19712 		/* Change from kbits per second to bytes per second */
19713 		val *= 1000;
19714 		val /= 8;
19715 		rack->r_ctl.init_rate = val;
19716 		if (rack->rc_init_win != rack_default_init_window) {
19717 			uint32_t win, snt;
19718 
19719 			/*
19720 			 * Options don't always get applied
19721 			 * in the order you think. So in order
19722 			 * to assure we update a cwnd we need
19723 			 * to check and see if we are still
19724 			 * where we should raise the cwnd.
19725 			 */
19726 			win = rc_init_window(rack);
19727 			if (SEQ_GT(tp->snd_max, tp->iss))
19728 				snt = tp->snd_max - tp->iss;
19729 			else
19730 				snt = 0;
19731 			if ((snt < win) &&
19732 			    (tp->snd_cwnd < win))
19733 				tp->snd_cwnd = win;
19734 		}
19735 		if (rack->rc_always_pace)
19736 			rack_update_seg(rack);
19737 		break;
19738 	case TCP_BBR_IWINTSO:
19739 		RACK_OPTS_INC(tcp_initial_win);
19740 		if (optval && (optval <= 0xff)) {
19741 			uint32_t win, snt;
19742 
19743 			rack->rc_init_win = optval;
19744 			win = rc_init_window(rack);
19745 			if (SEQ_GT(tp->snd_max, tp->iss))
19746 				snt = tp->snd_max - tp->iss;
19747 			else
19748 				snt = 0;
19749 			if ((snt < win) &&
19750 			    (tp->t_srtt |
19751 #ifdef NETFLIX_PEAKRATE
19752 			     tp->t_maxpeakrate |
19753 #endif
19754 			     rack->r_ctl.init_rate)) {
19755 				/*
19756 				 * We are not past the initial window
19757 				 * and we have some bases for pacing,
19758 				 * so we need to possibly adjust up
19759 				 * the cwnd. Note even if we don't set
19760 				 * the cwnd, its still ok to raise the rc_init_win
19761 				 * which can be used coming out of idle when we
19762 				 * would have a rate.
19763 				 */
19764 				if (tp->snd_cwnd < win)
19765 					tp->snd_cwnd = win;
19766 			}
19767 			if (rack->rc_always_pace)
19768 				rack_update_seg(rack);
19769 		} else
19770 			error = EINVAL;
19771 		break;
19772 	case TCP_RACK_FORCE_MSEG:
19773 		RACK_OPTS_INC(tcp_rack_force_max_seg);
19774 		if (optval)
19775 			rack->rc_force_max_seg = 1;
19776 		else
19777 			rack->rc_force_max_seg = 0;
19778 		break;
19779 	case TCP_RACK_PACE_MAX_SEG:
19780 		/* Max segments size in a pace in bytes */
19781 		RACK_OPTS_INC(tcp_rack_max_seg);
19782 		rack->rc_user_set_max_segs = optval;
19783 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19784 		break;
19785 	case TCP_RACK_PACE_RATE_REC:
19786 		/* Set the fixed pacing rate in Bytes per second ca */
19787 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19788 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19789 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19790 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19791 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19792 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19793 		rack->use_fixed_rate = 1;
19794 		if (rack->rc_always_pace)
19795 			rack_set_cc_pacing(rack);
19796 		rack_log_pacing_delay_calc(rack,
19797 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19798 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19799 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19800 					   __LINE__, NULL,0);
19801 		break;
19802 
19803 	case TCP_RACK_PACE_RATE_SS:
19804 		/* Set the fixed pacing rate in Bytes per second ca */
19805 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19806 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19807 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19808 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19809 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19810 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19811 		rack->use_fixed_rate = 1;
19812 		if (rack->rc_always_pace)
19813 			rack_set_cc_pacing(rack);
19814 		rack_log_pacing_delay_calc(rack,
19815 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19816 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19817 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19818 					   __LINE__, NULL, 0);
19819 		break;
19820 
19821 	case TCP_RACK_PACE_RATE_CA:
19822 		/* Set the fixed pacing rate in Bytes per second ca */
19823 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19824 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19825 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19826 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19827 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19828 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19829 		rack->use_fixed_rate = 1;
19830 		if (rack->rc_always_pace)
19831 			rack_set_cc_pacing(rack);
19832 		rack_log_pacing_delay_calc(rack,
19833 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19834 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19835 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19836 					   __LINE__, NULL, 0);
19837 		break;
19838 	case TCP_RACK_GP_INCREASE_REC:
19839 		RACK_OPTS_INC(tcp_gp_inc_rec);
19840 		rack->r_ctl.rack_per_of_gp_rec = optval;
19841 		rack_log_pacing_delay_calc(rack,
19842 					   rack->r_ctl.rack_per_of_gp_ss,
19843 					   rack->r_ctl.rack_per_of_gp_ca,
19844 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19845 					   __LINE__, NULL, 0);
19846 		break;
19847 	case TCP_RACK_GP_INCREASE_CA:
19848 		RACK_OPTS_INC(tcp_gp_inc_ca);
19849 		ca = optval;
19850 		if (ca < 100) {
19851 			/*
19852 			 * We don't allow any reduction
19853 			 * over the GP b/w.
19854 			 */
19855 			error = EINVAL;
19856 			break;
19857 		}
19858 		rack->r_ctl.rack_per_of_gp_ca = ca;
19859 		rack_log_pacing_delay_calc(rack,
19860 					   rack->r_ctl.rack_per_of_gp_ss,
19861 					   rack->r_ctl.rack_per_of_gp_ca,
19862 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19863 					   __LINE__, NULL, 0);
19864 		break;
19865 	case TCP_RACK_GP_INCREASE_SS:
19866 		RACK_OPTS_INC(tcp_gp_inc_ss);
19867 		ss = optval;
19868 		if (ss < 100) {
19869 			/*
19870 			 * We don't allow any reduction
19871 			 * over the GP b/w.
19872 			 */
19873 			error = EINVAL;
19874 			break;
19875 		}
19876 		rack->r_ctl.rack_per_of_gp_ss = ss;
19877 		rack_log_pacing_delay_calc(rack,
19878 					   rack->r_ctl.rack_per_of_gp_ss,
19879 					   rack->r_ctl.rack_per_of_gp_ca,
19880 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19881 					   __LINE__, NULL, 0);
19882 		break;
19883 	case TCP_RACK_RR_CONF:
19884 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19885 		if (optval && optval <= 3)
19886 			rack->r_rr_config = optval;
19887 		else
19888 			rack->r_rr_config = 0;
19889 		break;
19890 	case TCP_HDWR_RATE_CAP:
19891 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
19892 		if (optval) {
19893 			if (rack->r_rack_hw_rate_caps == 0)
19894 				rack->r_rack_hw_rate_caps = 1;
19895 			else
19896 				error = EALREADY;
19897 		} else {
19898 			rack->r_rack_hw_rate_caps = 0;
19899 		}
19900 		break;
19901 	case TCP_BBR_HDWR_PACE:
19902 		RACK_OPTS_INC(tcp_hdwr_pacing);
19903 		if (optval){
19904 			if (rack->rack_hdrw_pacing == 0) {
19905 				rack->rack_hdw_pace_ena = 1;
19906 				rack->rack_attempt_hdwr_pace = 0;
19907 			} else
19908 				error = EALREADY;
19909 		} else {
19910 			rack->rack_hdw_pace_ena = 0;
19911 #ifdef RATELIMIT
19912 			if (rack->r_ctl.crte != NULL) {
19913 				rack->rack_hdrw_pacing = 0;
19914 				rack->rack_attempt_hdwr_pace = 0;
19915 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19916 				rack->r_ctl.crte = NULL;
19917 			}
19918 #endif
19919 		}
19920 		break;
19921 	/*  End Pacing related ones */
19922 	case TCP_RACK_PRR_SENDALOT:
19923 		/* Allow PRR to send more than one seg */
19924 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
19925 		rack->r_ctl.rc_prr_sendalot = optval;
19926 		break;
19927 	case TCP_RACK_MIN_TO:
19928 		/* Minimum time between rack t-o's in ms */
19929 		RACK_OPTS_INC(tcp_rack_min_to);
19930 		rack->r_ctl.rc_min_to = optval;
19931 		break;
19932 	case TCP_RACK_EARLY_SEG:
19933 		/* If early recovery max segments */
19934 		RACK_OPTS_INC(tcp_rack_early_seg);
19935 		rack->r_ctl.rc_early_recovery_segs = optval;
19936 		break;
19937 	case TCP_RACK_ENABLE_HYSTART:
19938 	{
19939 		if (optval) {
19940 			tp->ccv->flags |= CCF_HYSTART_ALLOWED;
19941 			if (rack_do_hystart > RACK_HYSTART_ON)
19942 				tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND;
19943 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
19944 				tp->ccv->flags |= CCF_HYSTART_CONS_SSTH;
19945 		} else {
19946 			tp->ccv->flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
19947 		}
19948 	}
19949 	break;
19950 	case TCP_RACK_REORD_THRESH:
19951 		/* RACK reorder threshold (shift amount) */
19952 		RACK_OPTS_INC(tcp_rack_reord_thresh);
19953 		if ((optval > 0) && (optval < 31))
19954 			rack->r_ctl.rc_reorder_shift = optval;
19955 		else
19956 			error = EINVAL;
19957 		break;
19958 	case TCP_RACK_REORD_FADE:
19959 		/* Does reordering fade after ms time */
19960 		RACK_OPTS_INC(tcp_rack_reord_fade);
19961 		rack->r_ctl.rc_reorder_fade = optval;
19962 		break;
19963 	case TCP_RACK_TLP_THRESH:
19964 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
19965 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
19966 		if (optval)
19967 			rack->r_ctl.rc_tlp_threshold = optval;
19968 		else
19969 			error = EINVAL;
19970 		break;
19971 	case TCP_BBR_USE_RACK_RR:
19972 		RACK_OPTS_INC(tcp_rack_rr);
19973 		if (optval)
19974 			rack->use_rack_rr = 1;
19975 		else
19976 			rack->use_rack_rr = 0;
19977 		break;
19978 	case TCP_FAST_RSM_HACK:
19979 		RACK_OPTS_INC(tcp_rack_fastrsm_hack);
19980 		if (optval)
19981 			rack->fast_rsm_hack = 1;
19982 		else
19983 			rack->fast_rsm_hack = 0;
19984 		break;
19985 	case TCP_RACK_PKT_DELAY:
19986 		/* RACK added ms i.e. rack-rtt + reord + N */
19987 		RACK_OPTS_INC(tcp_rack_pkt_delay);
19988 		rack->r_ctl.rc_pkt_delay = optval;
19989 		break;
19990 	case TCP_DELACK:
19991 		RACK_OPTS_INC(tcp_rack_delayed_ack);
19992 		if (optval == 0)
19993 			tp->t_delayed_ack = 0;
19994 		else
19995 			tp->t_delayed_ack = 1;
19996 		if (tp->t_flags & TF_DELACK) {
19997 			tp->t_flags &= ~TF_DELACK;
19998 			tp->t_flags |= TF_ACKNOW;
19999 			NET_EPOCH_ENTER(et);
20000 			rack_output(tp);
20001 			NET_EPOCH_EXIT(et);
20002 		}
20003 		break;
20004 
20005 	case TCP_BBR_RACK_RTT_USE:
20006 		RACK_OPTS_INC(tcp_rack_rtt_use);
20007 		if ((optval != USE_RTT_HIGH) &&
20008 		    (optval != USE_RTT_LOW) &&
20009 		    (optval != USE_RTT_AVG))
20010 			error = EINVAL;
20011 		else
20012 			rack->r_ctl.rc_rate_sample_method = optval;
20013 		break;
20014 	case TCP_DATA_AFTER_CLOSE:
20015 		RACK_OPTS_INC(tcp_data_after_close);
20016 		if (optval)
20017 			rack->rc_allow_data_af_clo = 1;
20018 		else
20019 			rack->rc_allow_data_af_clo = 0;
20020 		break;
20021 	default:
20022 		break;
20023 	}
20024 #ifdef NETFLIX_STATS
20025 	tcp_log_socket_option(tp, sopt_name, optval, error);
20026 #endif
20027 	return (error);
20028 }
20029 
20030 
20031 static void
20032 rack_apply_deferred_options(struct tcp_rack *rack)
20033 {
20034 	struct deferred_opt_list *dol, *sdol;
20035 	uint32_t s_optval;
20036 
20037 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20038 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20039 		/* Disadvantage of deferal is you loose the error return */
20040 		s_optval = (uint32_t)dol->optval;
20041 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20042 		free(dol, M_TCPDO);
20043 	}
20044 }
20045 
20046 static void
20047 rack_hw_tls_change(struct tcpcb *tp, int chg)
20048 {
20049 	/*
20050 	 * HW tls state has changed.. fix all
20051 	 * rsm's in flight.
20052 	 */
20053 	struct tcp_rack *rack;
20054 	struct rack_sendmap *rsm;
20055 
20056 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20057 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
20058 		if (chg)
20059 			rsm->r_hw_tls = 1;
20060 		else
20061 			rsm->r_hw_tls = 0;
20062 	}
20063 	if (chg)
20064 		rack->r_ctl.fsb.hw_tls = 1;
20065 	else
20066 		rack->r_ctl.fsb.hw_tls = 0;
20067 }
20068 
20069 static int
20070 rack_pru_options(struct tcpcb *tp, int flags)
20071 {
20072 	if (flags & PRUS_OOB)
20073 		return (EOPNOTSUPP);
20074 	return (0);
20075 }
20076 
20077 static struct tcp_function_block __tcp_rack = {
20078 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
20079 	.tfb_tcp_output = rack_output,
20080 	.tfb_do_queued_segments = ctf_do_queued_segments,
20081 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
20082 	.tfb_tcp_do_segment = rack_do_segment,
20083 	.tfb_tcp_ctloutput = rack_ctloutput,
20084 	.tfb_tcp_fb_init = rack_init,
20085 	.tfb_tcp_fb_fini = rack_fini,
20086 	.tfb_tcp_timer_stop_all = rack_stopall,
20087 	.tfb_tcp_timer_activate = rack_timer_activate,
20088 	.tfb_tcp_timer_active = rack_timer_active,
20089 	.tfb_tcp_timer_stop = rack_timer_stop,
20090 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20091 	.tfb_tcp_handoff_ok = rack_handoff_ok,
20092 	.tfb_tcp_mtu_chg = rack_mtu_change,
20093 	.tfb_pru_options = rack_pru_options,
20094 	.tfb_hwtls_change = rack_hw_tls_change,
20095 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20096 };
20097 
20098 /*
20099  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20100  * socket option arguments.  When it re-acquires the lock after the copy, it
20101  * has to revalidate that the connection is still valid for the socket
20102  * option.
20103  */
20104 static int
20105 rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
20106 {
20107 #ifdef INET6
20108 	struct ip6_hdr *ip6;
20109 #endif
20110 #ifdef INET
20111 	struct ip *ip;
20112 #endif
20113 	struct tcpcb *tp;
20114 	struct tcp_rack *rack;
20115 	uint64_t loptval;
20116 	int32_t error = 0, optval;
20117 
20118 	tp = intotcpcb(inp);
20119 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20120 	if (rack == NULL) {
20121 		INP_WUNLOCK(inp);
20122 		return (EINVAL);
20123 	}
20124 #ifdef INET6
20125 	ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20126 #endif
20127 #ifdef INET
20128 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20129 #endif
20130 
20131 	switch (sopt->sopt_level) {
20132 #ifdef INET6
20133 	case IPPROTO_IPV6:
20134 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
20135 		switch (sopt->sopt_name) {
20136 		case IPV6_USE_MIN_MTU:
20137 			tcp6_use_min_mtu(tp);
20138 			break;
20139 		case IPV6_TCLASS:
20140 			/*
20141 			 * The DSCP codepoint has changed, update the fsb.
20142 			 */
20143 			ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20144 			    (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20145 			break;
20146 		}
20147 		INP_WUNLOCK(inp);
20148 		return (0);
20149 #endif
20150 #ifdef INET
20151 	case IPPROTO_IP:
20152 		switch (sopt->sopt_name) {
20153 		case IP_TOS:
20154 			/*
20155 			 * The DSCP codepoint has changed, update the fsb.
20156 			 */
20157 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
20158 			break;
20159 		case IP_TTL:
20160 			/*
20161 			 * The TTL has changed, update the fsb.
20162 			 */
20163 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20164 			break;
20165 		}
20166 		INP_WUNLOCK(inp);
20167 		return (0);
20168 #endif
20169 	}
20170 
20171 	switch (sopt->sopt_name) {
20172 	case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
20173 	/*  Pacing related ones */
20174 	case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
20175 	case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
20176 	case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
20177 	case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
20178 	case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
20179 	case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
20180 	case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
20181 	case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
20182 	case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
20183 	case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
20184 	case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
20185 	case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
20186 	case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
20187 	case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
20188 	case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
20189 	case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
20190        /* End pacing related */
20191 	case TCP_FAST_RSM_HACK:			/*  URL:frsm_hack */
20192 	case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20193 	case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
20194 	case TCP_RACK_MIN_TO:			/*  URL:min_to */
20195 	case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
20196 	case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
20197 	case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
20198 	case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
20199 	case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
20200 	case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
20201 	case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
20202 	case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
20203 	case TCP_RACK_DO_DETECTION:		/*  URL:detect */
20204 	case TCP_NO_PRR:			/*  URL:noprr */
20205 	case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
20206 	case TCP_DATA_AFTER_CLOSE:		/*  no URL */
20207 	case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
20208 	case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
20209 	case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
20210 	case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
20211 	case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
20212 	case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
20213 	case TCP_RACK_PROFILE:			/*  URL:profile */
20214 	case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
20215 	case TCP_RACK_ABC_VAL:			/*  URL:labc */
20216 	case TCP_REC_ABC_VAL:			/*  URL:reclabc */
20217 	case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
20218 	case TCP_DEFER_OPTIONS:			/*  URL:defer */
20219 	case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
20220 	case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
20221 	case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
20222 	case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
20223 	case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
20224 		break;
20225 	default:
20226 		/* Filter off all unknown options to the base stack */
20227 		return (tcp_default_ctloutput(inp, sopt));
20228 		break;
20229 	}
20230 	INP_WUNLOCK(inp);
20231 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20232 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20233 		/*
20234 		 * We truncate it down to 32 bits for the socket-option trace this
20235 		 * means rates > 34Gbps won't show right, but thats probably ok.
20236 		 */
20237 		optval = (uint32_t)loptval;
20238 	} else {
20239 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20240 		/* Save it in 64 bit form too */
20241 		loptval = optval;
20242 	}
20243 	if (error)
20244 		return (error);
20245 	INP_WLOCK(inp);
20246 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
20247 		INP_WUNLOCK(inp);
20248 		return (ECONNRESET);
20249 	}
20250 	if (tp->t_fb != &__tcp_rack) {
20251 		INP_WUNLOCK(inp);
20252 		return (ENOPROTOOPT);
20253 	}
20254 	if (rack->defer_options && (rack->gp_ready == 0) &&
20255 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20256 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20257 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20258 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20259 		/* Options are beind deferred */
20260 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20261 			INP_WUNLOCK(inp);
20262 			return (0);
20263 		} else {
20264 			/* No memory to defer, fail */
20265 			INP_WUNLOCK(inp);
20266 			return (ENOMEM);
20267 		}
20268 	}
20269 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20270 	INP_WUNLOCK(inp);
20271 	return (error);
20272 }
20273 
20274 static void
20275 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20276 {
20277 
20278 	INP_WLOCK_ASSERT(tp->t_inpcb);
20279 	bzero(ti, sizeof(*ti));
20280 
20281 	ti->tcpi_state = tp->t_state;
20282 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20283 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20284 	if (tp->t_flags & TF_SACK_PERMIT)
20285 		ti->tcpi_options |= TCPI_OPT_SACK;
20286 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20287 		ti->tcpi_options |= TCPI_OPT_WSCALE;
20288 		ti->tcpi_snd_wscale = tp->snd_scale;
20289 		ti->tcpi_rcv_wscale = tp->rcv_scale;
20290 	}
20291 	if (tp->t_flags2 & TF2_ECN_PERMIT)
20292 		ti->tcpi_options |= TCPI_OPT_ECN;
20293 	if (tp->t_flags & TF_FASTOPEN)
20294 		ti->tcpi_options |= TCPI_OPT_TFO;
20295 	/* still kept in ticks is t_rcvtime */
20296 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20297 	/* Since we hold everything in precise useconds this is easy */
20298 	ti->tcpi_rtt = tp->t_srtt;
20299 	ti->tcpi_rttvar = tp->t_rttvar;
20300 	ti->tcpi_rto = tp->t_rxtcur;
20301 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20302 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
20303 	/*
20304 	 * FreeBSD-specific extension fields for tcp_info.
20305 	 */
20306 	ti->tcpi_rcv_space = tp->rcv_wnd;
20307 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
20308 	ti->tcpi_snd_wnd = tp->snd_wnd;
20309 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
20310 	ti->tcpi_snd_nxt = tp->snd_nxt;
20311 	ti->tcpi_snd_mss = tp->t_maxseg;
20312 	ti->tcpi_rcv_mss = tp->t_maxseg;
20313 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20314 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20315 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20316 #ifdef NETFLIX_STATS
20317 	ti->tcpi_total_tlp = tp->t_sndtlppack;
20318 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20319 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20320 #endif
20321 #ifdef TCP_OFFLOAD
20322 	if (tp->t_flags & TF_TOE) {
20323 		ti->tcpi_options |= TCPI_OPT_TOE;
20324 		tcp_offload_tcp_info(tp, ti);
20325 	}
20326 #endif
20327 }
20328 
20329 static int
20330 rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
20331 {
20332 	struct tcpcb *tp;
20333 	struct tcp_rack *rack;
20334 	int32_t error, optval;
20335 	uint64_t val, loptval;
20336 	struct	tcp_info ti;
20337 	/*
20338 	 * Because all our options are either boolean or an int, we can just
20339 	 * pull everything into optval and then unlock and copy. If we ever
20340 	 * add a option that is not a int, then this will have quite an
20341 	 * impact to this routine.
20342 	 */
20343 	error = 0;
20344 	tp = intotcpcb(inp);
20345 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20346 	if (rack == NULL) {
20347 		INP_WUNLOCK(inp);
20348 		return (EINVAL);
20349 	}
20350 	switch (sopt->sopt_name) {
20351 	case TCP_INFO:
20352 		/* First get the info filled */
20353 		rack_fill_info(tp, &ti);
20354 		/* Fix up the rtt related fields if needed */
20355 		INP_WUNLOCK(inp);
20356 		error = sooptcopyout(sopt, &ti, sizeof ti);
20357 		return (error);
20358 	/*
20359 	 * Beta is the congestion control value for NewReno that influences how
20360 	 * much of a backoff happens when loss is detected. It is normally set
20361 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20362 	 * when you exit recovery.
20363 	 */
20364 	case TCP_RACK_PACING_BETA:
20365 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20366 			error = EINVAL;
20367 		else if (rack->rc_pacing_cc_set == 0)
20368 			optval = rack->r_ctl.rc_saved_beta.beta;
20369 		else {
20370 			/*
20371 			 * Reach out into the CC data and report back what
20372 			 * I have previously set. Yeah it looks hackish but
20373 			 * we don't want to report the saved values.
20374 			 */
20375 			if (tp->ccv->cc_data)
20376 				optval = ((struct newreno *)tp->ccv->cc_data)->beta;
20377 			else
20378 				error = EINVAL;
20379 		}
20380 		break;
20381 		/*
20382 		 * Beta_ecn is the congestion control value for NewReno that influences how
20383 		 * much of a backoff happens when a ECN mark is detected. It is normally set
20384 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20385 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
20386 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
20387 		 */
20388 
20389 	case TCP_RACK_PACING_BETA_ECN:
20390 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20391 			error = EINVAL;
20392 		else if (rack->rc_pacing_cc_set == 0)
20393 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20394 		else {
20395 			/*
20396 			 * Reach out into the CC data and report back what
20397 			 * I have previously set. Yeah it looks hackish but
20398 			 * we don't want to report the saved values.
20399 			 */
20400 			if (tp->ccv->cc_data)
20401 				optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
20402 			else
20403 				error = EINVAL;
20404 		}
20405 		break;
20406 	case TCP_RACK_DSACK_OPT:
20407 		optval = 0;
20408 		if (rack->rc_rack_tmr_std_based) {
20409 			optval |= 1;
20410 		}
20411 		if (rack->rc_rack_use_dsack) {
20412 			optval |= 2;
20413 		}
20414 		break;
20415  	case TCP_RACK_ENABLE_HYSTART:
20416 	{
20417 		if (tp->ccv->flags & CCF_HYSTART_ALLOWED) {
20418 			optval = RACK_HYSTART_ON;
20419 			if (tp->ccv->flags & CCF_HYSTART_CAN_SH_CWND)
20420 				optval = RACK_HYSTART_ON_W_SC;
20421 			if (tp->ccv->flags & CCF_HYSTART_CONS_SSTH)
20422 				optval = RACK_HYSTART_ON_W_SC_C;
20423 		} else {
20424 			optval = RACK_HYSTART_OFF;
20425 		}
20426 	}
20427 	break;
20428 	case TCP_FAST_RSM_HACK:
20429 		optval = rack->fast_rsm_hack;
20430 		break;
20431 	case TCP_DEFER_OPTIONS:
20432 		optval = rack->defer_options;
20433 		break;
20434 	case TCP_RACK_MEASURE_CNT:
20435 		optval = rack->r_ctl.req_measurements;
20436 		break;
20437 	case TCP_REC_ABC_VAL:
20438 		optval = rack->r_use_labc_for_rec;
20439 		break;
20440 	case TCP_RACK_ABC_VAL:
20441 		optval = rack->rc_labc;
20442 		break;
20443 	case TCP_HDWR_UP_ONLY:
20444 		optval= rack->r_up_only;
20445 		break;
20446 	case TCP_PACING_RATE_CAP:
20447 		loptval = rack->r_ctl.bw_rate_cap;
20448 		break;
20449 	case TCP_RACK_PROFILE:
20450 		/* You cannot retrieve a profile, its write only */
20451 		error = EINVAL;
20452 		break;
20453 	case TCP_USE_CMP_ACKS:
20454 		optval = rack->r_use_cmp_ack;
20455 		break;
20456 	case TCP_RACK_PACE_TO_FILL:
20457 		optval = rack->rc_pace_to_cwnd;
20458 		if (optval && rack->r_fill_less_agg)
20459 			optval++;
20460 		break;
20461 	case TCP_RACK_NO_PUSH_AT_MAX:
20462 		optval = rack->r_ctl.rc_no_push_at_mrtt;
20463 		break;
20464 	case TCP_SHARED_CWND_ENABLE:
20465 		optval = rack->rack_enable_scwnd;
20466 		break;
20467 	case TCP_RACK_NONRXT_CFG_RATE:
20468 		optval = rack->rack_rec_nonrxt_use_cr;
20469 		break;
20470 	case TCP_NO_PRR:
20471 		if (rack->rack_no_prr  == 1)
20472 			optval = 1;
20473 		else if (rack->no_prr_addback == 1)
20474 			optval = 2;
20475 		else
20476 			optval = 0;
20477 		break;
20478 	case TCP_RACK_DO_DETECTION:
20479 		optval = rack->do_detection;
20480 		break;
20481 	case TCP_RACK_MBUF_QUEUE:
20482 		/* Now do we use the LRO mbuf-queue feature */
20483 		optval = rack->r_mbuf_queue;
20484 		break;
20485 	case TCP_TIMELY_DYN_ADJ:
20486 		optval = rack->rc_gp_dyn_mul;
20487 		break;
20488 	case TCP_BBR_IWINTSO:
20489 		optval = rack->rc_init_win;
20490 		break;
20491 	case TCP_RACK_TLP_REDUCE:
20492 		/* RACK TLP cwnd reduction (bool) */
20493 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20494 		break;
20495 	case TCP_BBR_RACK_INIT_RATE:
20496 		val = rack->r_ctl.init_rate;
20497 		/* convert to kbits per sec */
20498 		val *= 8;
20499 		val /= 1000;
20500 		optval = (uint32_t)val;
20501 		break;
20502 	case TCP_RACK_FORCE_MSEG:
20503 		optval = rack->rc_force_max_seg;
20504 		break;
20505 	case TCP_RACK_PACE_MAX_SEG:
20506 		/* Max segments in a pace */
20507 		optval = rack->rc_user_set_max_segs;
20508 		break;
20509 	case TCP_RACK_PACE_ALWAYS:
20510 		/* Use the always pace method */
20511 		optval = rack->rc_always_pace;
20512 		break;
20513 	case TCP_RACK_PRR_SENDALOT:
20514 		/* Allow PRR to send more than one seg */
20515 		optval = rack->r_ctl.rc_prr_sendalot;
20516 		break;
20517 	case TCP_RACK_MIN_TO:
20518 		/* Minimum time between rack t-o's in ms */
20519 		optval = rack->r_ctl.rc_min_to;
20520 		break;
20521 	case TCP_RACK_EARLY_SEG:
20522 		/* If early recovery max segments */
20523 		optval = rack->r_ctl.rc_early_recovery_segs;
20524 		break;
20525 	case TCP_RACK_REORD_THRESH:
20526 		/* RACK reorder threshold (shift amount) */
20527 		optval = rack->r_ctl.rc_reorder_shift;
20528 		break;
20529 	case TCP_RACK_REORD_FADE:
20530 		/* Does reordering fade after ms time */
20531 		optval = rack->r_ctl.rc_reorder_fade;
20532 		break;
20533 	case TCP_BBR_USE_RACK_RR:
20534 		/* Do we use the rack cheat for rxt */
20535 		optval = rack->use_rack_rr;
20536 		break;
20537 	case TCP_RACK_RR_CONF:
20538 		optval = rack->r_rr_config;
20539 		break;
20540 	case TCP_HDWR_RATE_CAP:
20541 		optval = rack->r_rack_hw_rate_caps;
20542 		break;
20543 	case TCP_BBR_HDWR_PACE:
20544 		optval = rack->rack_hdw_pace_ena;
20545 		break;
20546 	case TCP_RACK_TLP_THRESH:
20547 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20548 		optval = rack->r_ctl.rc_tlp_threshold;
20549 		break;
20550 	case TCP_RACK_PKT_DELAY:
20551 		/* RACK added ms i.e. rack-rtt + reord + N */
20552 		optval = rack->r_ctl.rc_pkt_delay;
20553 		break;
20554 	case TCP_RACK_TLP_USE:
20555 		optval = rack->rack_tlp_threshold_use;
20556 		break;
20557 	case TCP_RACK_PACE_RATE_CA:
20558 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20559 		break;
20560 	case TCP_RACK_PACE_RATE_SS:
20561 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20562 		break;
20563 	case TCP_RACK_PACE_RATE_REC:
20564 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20565 		break;
20566 	case TCP_RACK_GP_INCREASE_SS:
20567 		optval = rack->r_ctl.rack_per_of_gp_ca;
20568 		break;
20569 	case TCP_RACK_GP_INCREASE_CA:
20570 		optval = rack->r_ctl.rack_per_of_gp_ss;
20571 		break;
20572 	case TCP_BBR_RACK_RTT_USE:
20573 		optval = rack->r_ctl.rc_rate_sample_method;
20574 		break;
20575 	case TCP_DELACK:
20576 		optval = tp->t_delayed_ack;
20577 		break;
20578 	case TCP_DATA_AFTER_CLOSE:
20579 		optval = rack->rc_allow_data_af_clo;
20580 		break;
20581 	case TCP_SHARED_CWND_TIME_LIMIT:
20582 		optval = rack->r_limit_scw;
20583 		break;
20584 	case TCP_RACK_TIMER_SLOP:
20585 		optval = rack->r_ctl.timer_slop;
20586 		break;
20587 	default:
20588 		return (tcp_default_ctloutput(inp, sopt));
20589 		break;
20590 	}
20591 	INP_WUNLOCK(inp);
20592 	if (error == 0) {
20593 		if (TCP_PACING_RATE_CAP)
20594 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
20595 		else
20596 			error = sooptcopyout(sopt, &optval, sizeof optval);
20597 	}
20598 	return (error);
20599 }
20600 
20601 static int
20602 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt)
20603 {
20604 	if (sopt->sopt_dir == SOPT_SET) {
20605 		return (rack_set_sockopt(inp, sopt));
20606 	} else if (sopt->sopt_dir == SOPT_GET) {
20607 		return (rack_get_sockopt(inp, sopt));
20608 	} else {
20609 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
20610 	}
20611 }
20612 
20613 static const char *rack_stack_names[] = {
20614 	__XSTRING(STACKNAME),
20615 #ifdef STACKALIAS
20616 	__XSTRING(STACKALIAS),
20617 #endif
20618 };
20619 
20620 static int
20621 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20622 {
20623 	memset(mem, 0, size);
20624 	return (0);
20625 }
20626 
20627 static void
20628 rack_dtor(void *mem, int32_t size, void *arg)
20629 {
20630 
20631 }
20632 
20633 static bool rack_mod_inited = false;
20634 
20635 static int
20636 tcp_addrack(module_t mod, int32_t type, void *data)
20637 {
20638 	int32_t err = 0;
20639 	int num_stacks;
20640 
20641 	switch (type) {
20642 	case MOD_LOAD:
20643 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20644 		    sizeof(struct rack_sendmap),
20645 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20646 
20647 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20648 		    sizeof(struct tcp_rack),
20649 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20650 
20651 		sysctl_ctx_init(&rack_sysctl_ctx);
20652 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20653 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20654 		    OID_AUTO,
20655 #ifdef STACKALIAS
20656 		    __XSTRING(STACKALIAS),
20657 #else
20658 		    __XSTRING(STACKNAME),
20659 #endif
20660 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20661 		    "");
20662 		if (rack_sysctl_root == NULL) {
20663 			printf("Failed to add sysctl node\n");
20664 			err = EFAULT;
20665 			goto free_uma;
20666 		}
20667 		rack_init_sysctls();
20668 		num_stacks = nitems(rack_stack_names);
20669 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20670 		    rack_stack_names, &num_stacks);
20671 		if (err) {
20672 			printf("Failed to register %s stack name for "
20673 			    "%s module\n", rack_stack_names[num_stacks],
20674 			    __XSTRING(MODNAME));
20675 			sysctl_ctx_free(&rack_sysctl_ctx);
20676 free_uma:
20677 			uma_zdestroy(rack_zone);
20678 			uma_zdestroy(rack_pcb_zone);
20679 			rack_counter_destroy();
20680 			printf("Failed to register rack module -- err:%d\n", err);
20681 			return (err);
20682 		}
20683 		tcp_lro_reg_mbufq();
20684 		rack_mod_inited = true;
20685 		break;
20686 	case MOD_QUIESCE:
20687 		err = deregister_tcp_functions(&__tcp_rack, true, false);
20688 		break;
20689 	case MOD_UNLOAD:
20690 		err = deregister_tcp_functions(&__tcp_rack, false, true);
20691 		if (err == EBUSY)
20692 			break;
20693 		if (rack_mod_inited) {
20694 			uma_zdestroy(rack_zone);
20695 			uma_zdestroy(rack_pcb_zone);
20696 			sysctl_ctx_free(&rack_sysctl_ctx);
20697 			rack_counter_destroy();
20698 			rack_mod_inited = false;
20699 		}
20700 		tcp_lro_dereg_mbufq();
20701 		err = 0;
20702 		break;
20703 	default:
20704 		return (EOPNOTSUPP);
20705 	}
20706 	return (err);
20707 }
20708 
20709 static moduledata_t tcp_rack = {
20710 	.name = __XSTRING(MODNAME),
20711 	.evhand = tcp_addrack,
20712 	.priv = 0
20713 };
20714 
20715 MODULE_VERSION(MODNAME, 1);
20716 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20717 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
20718