xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 9e5787d2284e187abb5b654d924394a65772e004)
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 <sys/param.h>
36 #include <sys/arb.h>
37 #include <sys/module.h>
38 #include <sys/kernel.h>
39 #ifdef TCP_HHOOK
40 #include <sys/hhook.h>
41 #endif
42 #include <sys/lock.h>
43 #include <sys/malloc.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/mbuf.h>
47 #include <sys/proc.h>		/* for proc0 declaration */
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/systm.h>
52 #ifdef STATS
53 #include <sys/qmath.h>
54 #include <sys/tree.h>
55 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
56 #else
57 #include <sys/tree.h>
58 #endif
59 #include <sys/refcount.h>
60 #include <sys/queue.h>
61 #include <sys/tim_filter.h>
62 #include <sys/smp.h>
63 #include <sys/kthread.h>
64 #include <sys/kern_prefetch.h>
65 #include <sys/protosw.h>
66 
67 #include <vm/uma.h>
68 
69 #include <net/route.h>
70 #include <net/route/nhop.h>
71 #include <net/vnet.h>
72 
73 #define TCPSTATES		/* for logging */
74 
75 #include <netinet/in.h>
76 #include <netinet/in_kdtrace.h>
77 #include <netinet/in_pcb.h>
78 #include <netinet/ip.h>
79 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
80 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
81 #include <netinet/ip_var.h>
82 #include <netinet/ip6.h>
83 #include <netinet6/in6_pcb.h>
84 #include <netinet6/ip6_var.h>
85 #include <netinet/tcp.h>
86 #define	TCPOUTFLAGS
87 #include <netinet/tcp_fsm.h>
88 #include <netinet/tcp_log_buf.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_hpts.h>
93 #include <netinet/tcp_ratelimit.h>
94 #include <netinet/tcpip.h>
95 #include <netinet/cc/cc.h>
96 #include <netinet/tcp_fastopen.h>
97 #include <netinet/tcp_lro.h>
98 #ifdef NETFLIX_SHARED_CWND
99 #include <netinet/tcp_shared_cwnd.h>
100 #endif
101 #ifdef TCPDEBUG
102 #include <netinet/tcp_debug.h>
103 #endif				/* TCPDEBUG */
104 #ifdef TCP_OFFLOAD
105 #include <netinet/tcp_offload.h>
106 #endif
107 #ifdef INET6
108 #include <netinet6/tcp6_var.h>
109 #endif
110 
111 #include <netipsec/ipsec_support.h>
112 
113 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
114 #include <netipsec/ipsec.h>
115 #include <netipsec/ipsec6.h>
116 #endif				/* IPSEC */
117 
118 #include <netinet/udp.h>
119 #include <netinet/udp_var.h>
120 #include <machine/in_cksum.h>
121 
122 #ifdef MAC
123 #include <security/mac/mac_framework.h>
124 #endif
125 #include "sack_filter.h"
126 #include "tcp_rack.h"
127 #include "rack_bbr_common.h"
128 
129 uma_zone_t rack_zone;
130 uma_zone_t rack_pcb_zone;
131 
132 #ifndef TICKS2SBT
133 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
134 #endif
135 
136 struct sysctl_ctx_list rack_sysctl_ctx;
137 struct sysctl_oid *rack_sysctl_root;
138 
139 #define CUM_ACKED 1
140 #define SACKED 2
141 
142 /*
143  * The RACK module incorporates a number of
144  * TCP ideas that have been put out into the IETF
145  * over the last few years:
146  * - Matt Mathis's Rate Halving which slowly drops
147  *    the congestion window so that the ack clock can
148  *    be maintained during a recovery.
149  * - Yuchung Cheng's RACK TCP (for which its named) that
150  *    will stop us using the number of dup acks and instead
151  *    use time as the gage of when we retransmit.
152  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
153  *    of Dukkipati et.al.
154  * RACK depends on SACK, so if an endpoint arrives that
155  * cannot do SACK the state machine below will shuttle the
156  * connection back to using the "default" TCP stack that is
157  * in FreeBSD.
158  *
159  * To implement RACK the original TCP stack was first decomposed
160  * into a functional state machine with individual states
161  * for each of the possible TCP connection states. The do_segement
162  * functions role in life is to mandate the connection supports SACK
163  * initially and then assure that the RACK state matches the conenction
164  * state before calling the states do_segment function. Each
165  * state is simplified due to the fact that the original do_segment
166  * has been decomposed and we *know* what state we are in (no
167  * switches on the state) and all tests for SACK are gone. This
168  * greatly simplifies what each state does.
169  *
170  * TCP output is also over-written with a new version since it
171  * must maintain the new rack scoreboard.
172  *
173  */
174 static int32_t rack_tlp_thresh = 1;
175 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
176 static int32_t rack_tlp_use_greater = 1;
177 static int32_t rack_reorder_thresh = 2;
178 static int32_t rack_reorder_fade = 60000;	/* 0 - never fade, def 60,000
179 						 * - 60 seconds */
180 /* Attack threshold detections */
181 static uint32_t rack_highest_sack_thresh_seen = 0;
182 static uint32_t rack_highest_move_thresh_seen = 0;
183 
184 static int32_t rack_pkt_delay = 1;
185 static int32_t rack_early_recovery = 1;
186 static int32_t rack_send_a_lot_in_prr = 1;
187 static int32_t rack_min_to = 1;	/* Number of ms minimum timeout */
188 static int32_t rack_verbose_logging = 0;
189 static int32_t rack_ignore_data_after_close = 1;
190 static int32_t rack_enable_shared_cwnd = 0;
191 static int32_t rack_limits_scwnd = 1;
192 static int32_t rack_enable_mqueue_for_nonpaced = 0;
193 static int32_t rack_disable_prr = 0;
194 static int32_t use_rack_rr = 1;
195 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
196 static int32_t rack_persist_min = 250;	/* 250ms */
197 static int32_t rack_persist_max = 2000;	/* 2 Second */
198 static int32_t rack_sack_not_required = 0;	/* set to one to allow non-sack to use rack */
199 static int32_t rack_default_init_window = 0; 	/* Use system default */
200 static int32_t rack_limit_time_with_srtt = 0;
201 static int32_t rack_hw_pace_adjust = 0;
202 /*
203  * Currently regular tcp has a rto_min of 30ms
204  * the backoff goes 12 times so that ends up
205  * being a total of 122.850 seconds before a
206  * connection is killed.
207  */
208 static uint32_t rack_def_data_window = 20;
209 static uint32_t rack_goal_bdp = 2;
210 static uint32_t rack_min_srtts = 1;
211 static uint32_t rack_min_measure_usec = 0;
212 static int32_t rack_tlp_min = 10;
213 static int32_t rack_rto_min = 30;	/* 30ms same as main freebsd */
214 static int32_t rack_rto_max = 4000;	/* 4 seconds */
215 static const int32_t rack_free_cache = 2;
216 static int32_t rack_hptsi_segments = 40;
217 static int32_t rack_rate_sample_method = USE_RTT_LOW;
218 static int32_t rack_pace_every_seg = 0;
219 static int32_t rack_delayed_ack_time = 200;	/* 200ms */
220 static int32_t rack_slot_reduction = 4;
221 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
222 static int32_t rack_cwnd_block_ends_measure = 0;
223 static int32_t rack_rwnd_block_ends_measure = 0;
224 
225 static int32_t rack_lower_cwnd_at_tlp = 0;
226 static int32_t rack_use_proportional_reduce = 0;
227 static int32_t rack_proportional_rate = 10;
228 static int32_t rack_tlp_max_resend = 2;
229 static int32_t rack_limited_retran = 0;
230 static int32_t rack_always_send_oldest = 0;
231 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
232 
233 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
234 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
235 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
236 
237 /* Probertt */
238 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
239 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
240 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
241 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
242 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
243 
244 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
245 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
246 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
247 static uint32_t rack_probertt_use_min_rtt_exit = 0;
248 static uint32_t rack_probe_rtt_sets_cwnd = 0;
249 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
250 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in us */
251 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
252 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction  */
253 static uint32_t rack_min_probertt_hold = 200000;	/* Equal to delayed ack time */
254 static uint32_t rack_probertt_filter_life = 10000000;
255 static uint32_t rack_probertt_lower_within = 10;
256 static uint32_t rack_min_rtt_movement = 250;	/* Must move at least 250 useconds to count as a lowering */
257 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
258 static int32_t rack_probertt_clear_is = 1;
259 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
260 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
261 
262 
263 /* Part of pacing */
264 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
265 
266 /* Timely information */
267 /* Combine these two gives the range of 'no change' to bw */
268 /* ie the up/down provide the upper and lower bound  */
269 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
270 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
271 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
272 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
273 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
274 static int32_t rack_gp_decrease_per = 20;	/* 20% decrease in multipler */
275 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multipler */
276 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
277 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
278 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
279 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
280 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
281 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
282 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
283 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
284 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
285 static int32_t rack_use_max_for_nobackoff = 0;
286 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
287 static int32_t rack_timely_no_stopping = 0;
288 static int32_t rack_down_raise_thresh = 100;
289 static int32_t rack_req_segs = 1;
290 
291 /* Weird delayed ack mode */
292 static int32_t rack_use_imac_dack = 0;
293 /* Rack specific counters */
294 counter_u64_t rack_badfr;
295 counter_u64_t rack_badfr_bytes;
296 counter_u64_t rack_rtm_prr_retran;
297 counter_u64_t rack_rtm_prr_newdata;
298 counter_u64_t rack_timestamp_mismatch;
299 counter_u64_t rack_reorder_seen;
300 counter_u64_t rack_paced_segments;
301 counter_u64_t rack_unpaced_segments;
302 counter_u64_t rack_calc_zero;
303 counter_u64_t rack_calc_nonzero;
304 counter_u64_t rack_saw_enobuf;
305 counter_u64_t rack_saw_enetunreach;
306 counter_u64_t rack_per_timer_hole;
307 
308 /* Tail loss probe counters */
309 counter_u64_t rack_tlp_tot;
310 counter_u64_t rack_tlp_newdata;
311 counter_u64_t rack_tlp_retran;
312 counter_u64_t rack_tlp_retran_bytes;
313 counter_u64_t rack_tlp_retran_fail;
314 counter_u64_t rack_to_tot;
315 counter_u64_t rack_to_arm_rack;
316 counter_u64_t rack_to_arm_tlp;
317 counter_u64_t rack_to_alloc;
318 counter_u64_t rack_to_alloc_hard;
319 counter_u64_t rack_to_alloc_emerg;
320 counter_u64_t rack_to_alloc_limited;
321 counter_u64_t rack_alloc_limited_conns;
322 counter_u64_t rack_split_limited;
323 
324 counter_u64_t rack_sack_proc_all;
325 counter_u64_t rack_sack_proc_short;
326 counter_u64_t rack_sack_proc_restart;
327 counter_u64_t rack_sack_attacks_detected;
328 counter_u64_t rack_sack_attacks_reversed;
329 counter_u64_t rack_sack_used_next_merge;
330 counter_u64_t rack_sack_splits;
331 counter_u64_t rack_sack_used_prev_merge;
332 counter_u64_t rack_sack_skipped_acked;
333 counter_u64_t rack_ack_total;
334 counter_u64_t rack_express_sack;
335 counter_u64_t rack_sack_total;
336 counter_u64_t rack_move_none;
337 counter_u64_t rack_move_some;
338 
339 counter_u64_t rack_used_tlpmethod;
340 counter_u64_t rack_used_tlpmethod2;
341 counter_u64_t rack_enter_tlp_calc;
342 counter_u64_t rack_input_idle_reduces;
343 counter_u64_t rack_collapsed_win;
344 counter_u64_t rack_tlp_does_nada;
345 counter_u64_t rack_try_scwnd;
346 
347 /* Temp CPU counters */
348 counter_u64_t rack_find_high;
349 
350 counter_u64_t rack_progress_drops;
351 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
352 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
353 
354 static void
355 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
356 
357 static int
358 rack_process_ack(struct mbuf *m, struct tcphdr *th,
359     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
360     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
361 static int
362 rack_process_data(struct mbuf *m, struct tcphdr *th,
363     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
364     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
365 static void
366 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
367     struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery);
368 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
369 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
370     uint8_t limit_type);
371 static struct rack_sendmap *
372 rack_check_recovery_mode(struct tcpcb *tp,
373     uint32_t tsused);
374 static void
375 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th,
376     uint32_t type);
377 static void rack_counter_destroy(void);
378 static int
379 rack_ctloutput(struct socket *so, struct sockopt *sopt,
380     struct inpcb *inp, struct tcpcb *tp);
381 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
382 static void
383 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line);
384 static void
385 rack_do_segment(struct mbuf *m, struct tcphdr *th,
386     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
387     uint8_t iptos);
388 static void rack_dtor(void *mem, int32_t size, void *arg);
389 static void
390 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm,
391     uint32_t t, uint32_t cts);
392 static void
393 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
394     uint32_t flex1, uint32_t flex2,
395     uint32_t flex3, uint32_t flex4,
396     uint32_t flex5, uint32_t flex6,
397     uint16_t flex7, uint8_t mod);
398 static void
399 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
400    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm);
401 static struct rack_sendmap *
402 rack_find_high_nonack(struct tcp_rack *rack,
403     struct rack_sendmap *rsm);
404 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
405 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
406 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
407 static int
408 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
409     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
410 static void
411 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
412 			    tcp_seq th_ack, int line);
413 static uint32_t
414 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
415 static int32_t rack_handoff_ok(struct tcpcb *tp);
416 static int32_t rack_init(struct tcpcb *tp);
417 static void rack_init_sysctls(void);
418 static void
419 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
420     struct tcphdr *th);
421 static void
422 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
423     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts,
424     uint8_t pass, struct rack_sendmap *hintrsm, uint32_t us_cts);
425 static void
426 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
427     struct rack_sendmap *rsm);
428 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
429 static int32_t rack_output(struct tcpcb *tp);
430 
431 static uint32_t
432 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
433     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
434     uint32_t cts, int *moved_two);
435 static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th);
436 static void rack_remxt_tmr(struct tcpcb *tp);
437 static int
438 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
439     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
440 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
441 static int32_t rack_stopall(struct tcpcb *tp);
442 static void
443 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
444     uint32_t delta);
445 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
446 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
447 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
448 static uint32_t
449 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
450     struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp);
451 static void
452 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
453     struct rack_sendmap *rsm, uint32_t ts);
454 static int
455 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
456     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
457 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
458 static int
459 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
460     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
461     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
462 static int
463 rack_do_closing(struct mbuf *m, struct tcphdr *th,
464     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
465     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
466 static int
467 rack_do_established(struct mbuf *m, struct tcphdr *th,
468     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
469     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
470 static int
471 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
472     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
473     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
474 static int
475 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
476     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
477     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
478 static int
479 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
480     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
481     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
482 static int
483 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
484     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
485     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
486 static int
487 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
488     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
489     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
490 static int
491 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
492     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
493     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
494 struct rack_sendmap *
495 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
496     uint32_t tsused);
497 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
498     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
499 static void
500      tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th);
501 
502 int32_t rack_clear_counter=0;
503 
504 
505 static int
506 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
507 {
508 	uint32_t stat;
509 	int32_t error;
510 
511 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
512 	if (error || req->newptr == NULL)
513 		return error;
514 
515 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
516 	if (error)
517 		return (error);
518 	if (stat == 1) {
519 #ifdef INVARIANTS
520 		printf("Clearing RACK counters\n");
521 #endif
522 		counter_u64_zero(rack_badfr);
523 		counter_u64_zero(rack_badfr_bytes);
524 		counter_u64_zero(rack_rtm_prr_retran);
525 		counter_u64_zero(rack_rtm_prr_newdata);
526 		counter_u64_zero(rack_timestamp_mismatch);
527 		counter_u64_zero(rack_reorder_seen);
528 		counter_u64_zero(rack_tlp_tot);
529 		counter_u64_zero(rack_tlp_newdata);
530 		counter_u64_zero(rack_tlp_retran);
531 		counter_u64_zero(rack_tlp_retran_bytes);
532 		counter_u64_zero(rack_tlp_retran_fail);
533 		counter_u64_zero(rack_to_tot);
534 		counter_u64_zero(rack_to_arm_rack);
535 		counter_u64_zero(rack_to_arm_tlp);
536 		counter_u64_zero(rack_paced_segments);
537 		counter_u64_zero(rack_calc_zero);
538 		counter_u64_zero(rack_calc_nonzero);
539 		counter_u64_zero(rack_unpaced_segments);
540 		counter_u64_zero(rack_saw_enobuf);
541 		counter_u64_zero(rack_saw_enetunreach);
542 		counter_u64_zero(rack_per_timer_hole);
543 		counter_u64_zero(rack_to_alloc_hard);
544 		counter_u64_zero(rack_to_alloc_emerg);
545 		counter_u64_zero(rack_sack_proc_all);
546 		counter_u64_zero(rack_sack_proc_short);
547 		counter_u64_zero(rack_sack_proc_restart);
548 		counter_u64_zero(rack_to_alloc);
549 		counter_u64_zero(rack_to_alloc_limited);
550 		counter_u64_zero(rack_alloc_limited_conns);
551 		counter_u64_zero(rack_split_limited);
552 		counter_u64_zero(rack_find_high);
553 		counter_u64_zero(rack_sack_attacks_detected);
554 		counter_u64_zero(rack_sack_attacks_reversed);
555 		counter_u64_zero(rack_sack_used_next_merge);
556 		counter_u64_zero(rack_sack_used_prev_merge);
557 		counter_u64_zero(rack_sack_splits);
558 		counter_u64_zero(rack_sack_skipped_acked);
559 		counter_u64_zero(rack_ack_total);
560 		counter_u64_zero(rack_express_sack);
561 		counter_u64_zero(rack_sack_total);
562 		counter_u64_zero(rack_move_none);
563 		counter_u64_zero(rack_move_some);
564 		counter_u64_zero(rack_used_tlpmethod);
565 		counter_u64_zero(rack_used_tlpmethod2);
566 		counter_u64_zero(rack_enter_tlp_calc);
567 		counter_u64_zero(rack_progress_drops);
568 		counter_u64_zero(rack_tlp_does_nada);
569 		counter_u64_zero(rack_try_scwnd);
570 		counter_u64_zero(rack_collapsed_win);
571 
572 	}
573 	rack_clear_counter = 0;
574 	return (0);
575 }
576 
577 
578 
579 static void
580 rack_init_sysctls(void)
581 {
582 	struct sysctl_oid *rack_counters;
583 	struct sysctl_oid *rack_attack;
584 	struct sysctl_oid *rack_pacing;
585 	struct sysctl_oid *rack_timely;
586 	struct sysctl_oid *rack_timers;
587 	struct sysctl_oid *rack_tlp;
588 	struct sysctl_oid *rack_misc;
589 	struct sysctl_oid *rack_measure;
590 	struct sysctl_oid *rack_probertt;
591 
592 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
593 	    SYSCTL_CHILDREN(rack_sysctl_root),
594 	    OID_AUTO,
595 	    "sack_attack",
596 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
597 	    "Rack Sack Attack Counters and Controls");
598 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
599 	    SYSCTL_CHILDREN(rack_sysctl_root),
600 	    OID_AUTO,
601 	    "stats",
602 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
603 	    "Rack Counters");
604 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
605 	    SYSCTL_CHILDREN(rack_sysctl_root),
606 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
607 	    &rack_rate_sample_method , USE_RTT_LOW,
608 	    "What method should we use for rate sampling 0=high, 1=low ");
609 	/* Probe rtt related controls */
610 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
611 	    SYSCTL_CHILDREN(rack_sysctl_root),
612 	    OID_AUTO,
613 	    "probertt",
614 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
615 	    "ProbeRTT related Controls");
616 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
617 	    SYSCTL_CHILDREN(rack_probertt),
618 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
619 	    &rack_atexit_prtt_hbp, 130,
620 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
621 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
622 	    SYSCTL_CHILDREN(rack_probertt),
623 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
624 	    &rack_atexit_prtt, 130,
625 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
626 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
627 	    SYSCTL_CHILDREN(rack_probertt),
628 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
629 	    &rack_per_of_gp_probertt, 60,
630 	    "What percentage of goodput do we pace at in probertt");
631 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
632 	    SYSCTL_CHILDREN(rack_probertt),
633 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
634 	    &rack_per_of_gp_probertt_reduce, 10,
635 	    "What percentage of goodput do we reduce every gp_srtt");
636 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
637 	    SYSCTL_CHILDREN(rack_probertt),
638 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
639 	    &rack_per_of_gp_lowthresh, 40,
640 	    "What percentage of goodput do we allow the multiplier to fall to");
641 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
642 	    SYSCTL_CHILDREN(rack_probertt),
643 	    OID_AUTO, "time_between", CTLFLAG_RW,
644 	    & rack_time_between_probertt, 96000000,
645 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
646 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
647 	    SYSCTL_CHILDREN(rack_probertt),
648 	    OID_AUTO, "safety", CTLFLAG_RW,
649 	    &rack_probe_rtt_safety_val, 2000000,
650 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
651 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
652 	    SYSCTL_CHILDREN(rack_probertt),
653 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
654 	    &rack_probe_rtt_sets_cwnd, 0,
655 	    "Do we set the cwnd too (if always_lower is on)");
656 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
657 	    SYSCTL_CHILDREN(rack_probertt),
658 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
659 	    &rack_max_drain_wait, 2,
660 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
661 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
662 	    SYSCTL_CHILDREN(rack_probertt),
663 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
664 	    &rack_must_drain, 1,
665 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
666 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
667 	    SYSCTL_CHILDREN(rack_probertt),
668 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
669 	    &rack_probertt_use_min_rtt_entry, 1,
670 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
671 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
672 	    SYSCTL_CHILDREN(rack_probertt),
673 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
674 	    &rack_probertt_use_min_rtt_exit, 0,
675 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
676 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
677 	    SYSCTL_CHILDREN(rack_probertt),
678 	    OID_AUTO, "length_div", CTLFLAG_RW,
679 	    &rack_probertt_gpsrtt_cnt_div, 0,
680 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
681 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
682 	    SYSCTL_CHILDREN(rack_probertt),
683 	    OID_AUTO, "length_mul", CTLFLAG_RW,
684 	    &rack_probertt_gpsrtt_cnt_mul, 0,
685 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
686 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
687 	    SYSCTL_CHILDREN(rack_probertt),
688 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
689 	    &rack_min_probertt_hold, 200000,
690 	    "What is the minimum time we hold probertt at target");
691 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
692 	    SYSCTL_CHILDREN(rack_probertt),
693 	    OID_AUTO, "filter_life", CTLFLAG_RW,
694 	    &rack_probertt_filter_life, 10000000,
695 	    "What is the time for the filters life in useconds");
696 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
697 	    SYSCTL_CHILDREN(rack_probertt),
698 	    OID_AUTO, "lower_within", CTLFLAG_RW,
699 	    &rack_probertt_lower_within, 10,
700 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
701 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
702 	    SYSCTL_CHILDREN(rack_probertt),
703 	    OID_AUTO, "must_move", CTLFLAG_RW,
704 	    &rack_min_rtt_movement, 250,
705 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
706 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
707 	    SYSCTL_CHILDREN(rack_probertt),
708 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
709 	    &rack_probertt_clear_is, 1,
710 	    "Do we clear I/S counts on exiting probe-rtt");
711 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
712 	    SYSCTL_CHILDREN(rack_probertt),
713 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
714 	    &rack_max_drain_hbp, 1,
715 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
716 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
717 	    SYSCTL_CHILDREN(rack_probertt),
718 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
719 	    &rack_hbp_thresh, 3,
720 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
721 	/* Pacing related sysctls */
722 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
723 	    SYSCTL_CHILDREN(rack_sysctl_root),
724 	    OID_AUTO,
725 	    "pacing",
726 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
727 	    "Pacing related Controls");
728 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
729 	    SYSCTL_CHILDREN(rack_pacing),
730 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
731 	    &rack_max_per_above, 30,
732 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
733 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
734 	    SYSCTL_CHILDREN(rack_pacing),
735 	    OID_AUTO, "pace_to_one", CTLFLAG_RW,
736 	    &rack_pace_one_seg, 0,
737 	    "Do we allow low b/w pacing of 1MSS instead of two");
738 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
739 	    SYSCTL_CHILDREN(rack_pacing),
740 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
741 	    &rack_limit_time_with_srtt, 0,
742 	    "Do we limit pacing time based on srtt");
743 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
744 	    SYSCTL_CHILDREN(rack_pacing),
745 	    OID_AUTO, "init_win", CTLFLAG_RW,
746 	    &rack_default_init_window, 0,
747 	    "Do we have a rack initial window 0 = system default");
748 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
749 	    SYSCTL_CHILDREN(rack_pacing),
750 	    OID_AUTO, "hw_pacing_adjust", CTLFLAG_RW,
751 	    &rack_hw_pace_adjust, 0,
752 	    "What percentage do we raise the MSS by (11 = 1.1%)");
753 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
754 	    SYSCTL_CHILDREN(rack_pacing),
755 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
756 	    &rack_per_of_gp_ss, 250,
757 	    "If non zero, what percentage of goodput to pace at in slow start");
758 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
759 	    SYSCTL_CHILDREN(rack_pacing),
760 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
761 	    &rack_per_of_gp_ca, 150,
762 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
763 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
764 	    SYSCTL_CHILDREN(rack_pacing),
765 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
766 	    &rack_per_of_gp_rec, 200,
767 	    "If non zero, what percentage of goodput to pace at in recovery");
768 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
769 	    SYSCTL_CHILDREN(rack_pacing),
770 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
771 	    &rack_hptsi_segments, 40,
772 	    "What size is the max for TSO segments in pacing and burst mitigation");
773 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
774 	    SYSCTL_CHILDREN(rack_pacing),
775 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
776 	    &rack_slot_reduction, 4,
777 	    "When doing only burst mitigation what is the reduce divisor");
778 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
779 	    SYSCTL_CHILDREN(rack_sysctl_root),
780 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
781 	    &rack_pace_every_seg, 0,
782 	    "If set we use pacing, if clear we use only the original burst mitigation");
783 
784 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
785 	    SYSCTL_CHILDREN(rack_sysctl_root),
786 	    OID_AUTO,
787 	    "timely",
788 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
789 	    "Rack Timely RTT Controls");
790 	/* Timely based GP dynmics */
791 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
792 	    SYSCTL_CHILDREN(rack_timely),
793 	    OID_AUTO, "upper", CTLFLAG_RW,
794 	    &rack_gp_per_bw_mul_up, 2,
795 	    "Rack timely upper range for equal b/w (in percentage)");
796 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
797 	    SYSCTL_CHILDREN(rack_timely),
798 	    OID_AUTO, "lower", CTLFLAG_RW,
799 	    &rack_gp_per_bw_mul_down, 4,
800 	    "Rack timely lower range for equal b/w (in percentage)");
801 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
802 	    SYSCTL_CHILDREN(rack_timely),
803 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
804 	    &rack_gp_rtt_maxmul, 3,
805 	    "Rack timely multipler of lowest rtt for rtt_max");
806 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
807 	    SYSCTL_CHILDREN(rack_timely),
808 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
809 	    &rack_gp_rtt_mindiv, 4,
810 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
811 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
812 	    SYSCTL_CHILDREN(rack_timely),
813 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
814 	    &rack_gp_rtt_minmul, 1,
815 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
816 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
817 	    SYSCTL_CHILDREN(rack_timely),
818 	    OID_AUTO, "decrease", CTLFLAG_RW,
819 	    &rack_gp_decrease_per, 20,
820 	    "Rack timely decrease percentage of our GP multiplication factor");
821 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
822 	    SYSCTL_CHILDREN(rack_timely),
823 	    OID_AUTO, "increase", CTLFLAG_RW,
824 	    &rack_gp_increase_per, 2,
825 	    "Rack timely increase perentage of our GP multiplication factor");
826 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
827 	    SYSCTL_CHILDREN(rack_timely),
828 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
829 	    &rack_per_lower_bound, 50,
830 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
831 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
832 	    SYSCTL_CHILDREN(rack_timely),
833 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
834 	    &rack_per_upper_bound_ss, 0,
835 	    "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
836 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
837 	    SYSCTL_CHILDREN(rack_timely),
838 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
839 	    &rack_per_upper_bound_ca, 0,
840 	    "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
841 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
842 	    SYSCTL_CHILDREN(rack_timely),
843 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
844 	    &rack_do_dyn_mul, 0,
845 	    "Rack timely do we enable dynmaic timely goodput by default");
846 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
847 	    SYSCTL_CHILDREN(rack_timely),
848 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
849 	    &rack_gp_no_rec_chg, 1,
850 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
851 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
852 	    SYSCTL_CHILDREN(rack_timely),
853 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
854 	    &rack_timely_dec_clear, 6,
855 	    "Rack timely what threshold do we count to before another boost during b/w decent");
856 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
857 	    SYSCTL_CHILDREN(rack_timely),
858 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
859 	    &rack_timely_max_push_rise, 3,
860 	    "Rack timely how many times do we push up with b/w increase");
861 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
862 	    SYSCTL_CHILDREN(rack_timely),
863 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
864 	    &rack_timely_max_push_drop, 3,
865 	    "Rack timely how many times do we push back on b/w decent");
866 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
867 	    SYSCTL_CHILDREN(rack_timely),
868 	    OID_AUTO, "min_segs", CTLFLAG_RW,
869 	    &rack_timely_min_segs, 4,
870 	    "Rack timely when setting the cwnd what is the min num segments");
871 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
872 	    SYSCTL_CHILDREN(rack_timely),
873 	    OID_AUTO, "noback_max", CTLFLAG_RW,
874 	    &rack_use_max_for_nobackoff, 0,
875 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
876 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
877 	    SYSCTL_CHILDREN(rack_timely),
878 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
879 	    &rack_timely_int_timely_only, 0,
880 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
881 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
882 	    SYSCTL_CHILDREN(rack_timely),
883 	    OID_AUTO, "nonstop", CTLFLAG_RW,
884 	    &rack_timely_no_stopping, 0,
885 	    "Rack timely don't stop increase");
886 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
887 	    SYSCTL_CHILDREN(rack_timely),
888 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
889 	    &rack_down_raise_thresh, 100,
890 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
891 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
892 	    SYSCTL_CHILDREN(rack_timely),
893 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
894 	    &rack_req_segs, 1,
895 	    "Bottom dragging if not these many segments outstanding and room");
896 
897 	/* TLP and Rack related parameters */
898 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
899 	    SYSCTL_CHILDREN(rack_sysctl_root),
900 	    OID_AUTO,
901 	    "tlp",
902 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
903 	    "TLP and Rack related Controls");
904 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
905 	    SYSCTL_CHILDREN(rack_tlp),
906 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
907 	    &use_rack_rr, 1,
908 	    "Do we use Rack Rapid Recovery");
909 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
910 	    SYSCTL_CHILDREN(rack_tlp),
911 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
912 	    &rack_non_rxt_use_cr, 0,
913 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
914 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
915 	    SYSCTL_CHILDREN(rack_tlp),
916 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
917 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
918 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
919 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
920 	    SYSCTL_CHILDREN(rack_tlp),
921 	    OID_AUTO, "limit", CTLFLAG_RW,
922 	    &rack_tlp_limit, 2,
923 	    "How many TLP's can be sent without sending new data");
924 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
925 	    SYSCTL_CHILDREN(rack_tlp),
926 	    OID_AUTO, "use_greater", CTLFLAG_RW,
927 	    &rack_tlp_use_greater, 1,
928 	    "Should we use the rack_rtt time if its greater than srtt");
929 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
930 	    SYSCTL_CHILDREN(rack_tlp),
931 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
932 	    &rack_tlp_min, 10,
933 	    "TLP minimum timeout per the specification (10ms)");
934 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
935 	    SYSCTL_CHILDREN(rack_tlp),
936 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
937 	    &rack_always_send_oldest, 0,
938 	    "Should we always send the oldest TLP and RACK-TLP");
939 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
940 	    SYSCTL_CHILDREN(rack_tlp),
941 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
942 	    &rack_limited_retran, 0,
943 	    "How many times can a rack timeout drive out sends");
944 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
945 	    SYSCTL_CHILDREN(rack_tlp),
946 	    OID_AUTO, "tlp_retry", CTLFLAG_RW,
947 	    &rack_tlp_max_resend, 2,
948 	    "How many times does TLP retry a single segment or multiple with no ACK");
949 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
950 	    SYSCTL_CHILDREN(rack_tlp),
951 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
952 	    &rack_lower_cwnd_at_tlp, 0,
953 	    "When a TLP completes a retran should we enter recovery");
954 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
955 	    SYSCTL_CHILDREN(rack_tlp),
956 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
957 	    &rack_reorder_thresh, 2,
958 	    "What factor for rack will be added when seeing reordering (shift right)");
959 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
960 	    SYSCTL_CHILDREN(rack_tlp),
961 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
962 	    &rack_tlp_thresh, 1,
963 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
964 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
965 	    SYSCTL_CHILDREN(rack_tlp),
966 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
967 	    &rack_reorder_fade, 0,
968 	    "Does reorder detection fade, if so how many ms (0 means never)");
969 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
970 	    SYSCTL_CHILDREN(rack_tlp),
971 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
972 	    &rack_pkt_delay, 1,
973 	    "Extra RACK time (in ms) besides reordering thresh");
974 
975 	/* Timer related controls */
976 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
977 	    SYSCTL_CHILDREN(rack_sysctl_root),
978 	    OID_AUTO,
979 	    "timers",
980 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
981 	    "Timer related controls");
982 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
983 	    SYSCTL_CHILDREN(rack_timers),
984 	    OID_AUTO, "persmin", CTLFLAG_RW,
985 	    &rack_persist_min, 250,
986 	    "What is the minimum time in milliseconds between persists");
987 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
988 	    SYSCTL_CHILDREN(rack_timers),
989 	    OID_AUTO, "persmax", CTLFLAG_RW,
990 	    &rack_persist_max, 2000,
991 	    "What is the largest delay in milliseconds between persists");
992 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
993 	    SYSCTL_CHILDREN(rack_timers),
994 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
995 	    &rack_delayed_ack_time, 200,
996 	    "Delayed ack time (200ms)");
997 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
998 	    SYSCTL_CHILDREN(rack_timers),
999 	    OID_AUTO, "minrto", CTLFLAG_RW,
1000 	    &rack_rto_min, 0,
1001 	    "Minimum RTO in ms -- set with caution below 1000 due to TLP");
1002 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1003 	    SYSCTL_CHILDREN(rack_timers),
1004 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1005 	    &rack_rto_max, 0,
1006 	    "Maxiumum RTO in ms -- should be at least as large as min_rto");
1007 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1008 	    SYSCTL_CHILDREN(rack_timers),
1009 	    OID_AUTO, "minto", CTLFLAG_RW,
1010 	    &rack_min_to, 1,
1011 	    "Minimum rack timeout in milliseconds");
1012 	/* Measure controls */
1013 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1014 	    SYSCTL_CHILDREN(rack_sysctl_root),
1015 	    OID_AUTO,
1016 	    "measure",
1017 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1018 	    "Measure related controls");
1019 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1020 	    SYSCTL_CHILDREN(rack_measure),
1021 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1022 	    &rack_wma_divisor, 8,
1023 	    "When doing b/w calculation what is the  divisor for the WMA");
1024 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1025 	    SYSCTL_CHILDREN(rack_measure),
1026 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1027 	    &rack_cwnd_block_ends_measure, 0,
1028 	    "Does a cwnd just-return end the measurement window (app limited)");
1029 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1030 	    SYSCTL_CHILDREN(rack_measure),
1031 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1032 	    &rack_rwnd_block_ends_measure, 0,
1033 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1034 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1035 	    SYSCTL_CHILDREN(rack_measure),
1036 	    OID_AUTO, "min_target", CTLFLAG_RW,
1037 	    &rack_def_data_window, 20,
1038 	    "What is the minimum target window (in mss) for a GP measurements");
1039 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1040 	    SYSCTL_CHILDREN(rack_measure),
1041 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1042 	    &rack_goal_bdp, 2,
1043 	    "What is the goal BDP to measure");
1044 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1045 	    SYSCTL_CHILDREN(rack_measure),
1046 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1047 	    &rack_min_srtts, 1,
1048 	    "What is the goal BDP to measure");
1049 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1050 	    SYSCTL_CHILDREN(rack_measure),
1051 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1052 	    &rack_min_measure_usec, 0,
1053 	    "What is the Minimum time time for a measurement if 0, this is off");
1054 	/* Misc rack controls */
1055 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1056 	    SYSCTL_CHILDREN(rack_sysctl_root),
1057 	    OID_AUTO,
1058 	    "misc",
1059 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1060 	    "Misc related controls");
1061 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1062 	    SYSCTL_CHILDREN(rack_misc),
1063 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1064 	    &rack_enable_shared_cwnd, 0,
1065 	    "Should RACK try to use the shared cwnd on connections where allowed");
1066 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1067 	    SYSCTL_CHILDREN(rack_misc),
1068 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1069 	    &rack_limits_scwnd, 1,
1070 	    "Should RACK place low end time limits on the shared cwnd feature");
1071 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1072 	    SYSCTL_CHILDREN(rack_misc),
1073 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1074 	    &rack_enable_mqueue_for_nonpaced, 0,
1075 	    "Should RACK use mbuf queuing for non-paced connections");
1076 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1077 	    SYSCTL_CHILDREN(rack_misc),
1078 	    OID_AUTO, "iMac_dack", CTLFLAG_RW,
1079 	    &rack_use_imac_dack, 0,
1080 	    "Should RACK try to emulate iMac delayed ack");
1081 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1082 	    SYSCTL_CHILDREN(rack_misc),
1083 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1084 	    &rack_disable_prr, 0,
1085 	    "Should RACK not use prr and only pace (must have pacing on)");
1086 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1087 	    SYSCTL_CHILDREN(rack_misc),
1088 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1089 	    &rack_verbose_logging, 0,
1090 	    "Should RACK black box logging be verbose");
1091 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1092 	    SYSCTL_CHILDREN(rack_misc),
1093 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1094 	    &rack_ignore_data_after_close, 1,
1095 	    "Do we hold off sending a RST until all pending data is ack'd");
1096 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1097 	    SYSCTL_CHILDREN(rack_misc),
1098 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1099 	    &rack_sack_not_required, 0,
1100 	    "Do we allow rack to run on connections not supporting SACK");
1101 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1102 	    SYSCTL_CHILDREN(rack_misc),
1103 	    OID_AUTO, "recovery_loss_prop", CTLFLAG_RW,
1104 	    &rack_use_proportional_reduce, 0,
1105 	    "Should we proportionaly reduce cwnd based on the number of losses ");
1106 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1107 	    SYSCTL_CHILDREN(rack_misc),
1108 	    OID_AUTO, "recovery_prop", CTLFLAG_RW,
1109 	    &rack_proportional_rate, 10,
1110 	    "What percent reduction per loss");
1111 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1112 	    SYSCTL_CHILDREN(rack_misc),
1113 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1114 	    &rack_send_a_lot_in_prr, 1,
1115 	    "Send a lot in prr");
1116 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1117 	    SYSCTL_CHILDREN(rack_misc),
1118 	    OID_AUTO, "earlyrecovery", CTLFLAG_RW,
1119 	    &rack_early_recovery, 1,
1120 	    "Do we do early recovery with rack");
1121 	/* Sack Attacker detection stuff */
1122 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1123 	    SYSCTL_CHILDREN(rack_attack),
1124 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1125 	    &rack_highest_sack_thresh_seen, 0,
1126 	    "Highest sack to ack ratio seen");
1127 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1128 	    SYSCTL_CHILDREN(rack_attack),
1129 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1130 	    &rack_highest_move_thresh_seen, 0,
1131 	    "Highest move to non-move ratio seen");
1132 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1133 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1134 	    SYSCTL_CHILDREN(rack_attack),
1135 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1136 	    &rack_ack_total,
1137 	    "Total number of Ack's");
1138 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1139 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1140 	    SYSCTL_CHILDREN(rack_attack),
1141 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1142 	    &rack_express_sack,
1143 	    "Total expresss number of Sack's");
1144 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1145 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1146 	    SYSCTL_CHILDREN(rack_attack),
1147 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1148 	    &rack_sack_total,
1149 	    "Total number of SACKs");
1150 	rack_move_none = counter_u64_alloc(M_WAITOK);
1151 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1152 	    SYSCTL_CHILDREN(rack_attack),
1153 	    OID_AUTO, "move_none", CTLFLAG_RD,
1154 	    &rack_move_none,
1155 	    "Total number of SACK index reuse of postions under threshold");
1156 	rack_move_some = counter_u64_alloc(M_WAITOK);
1157 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1158 	    SYSCTL_CHILDREN(rack_attack),
1159 	    OID_AUTO, "move_some", CTLFLAG_RD,
1160 	    &rack_move_some,
1161 	    "Total number of SACK index reuse of postions over threshold");
1162 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1163 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1164 	    SYSCTL_CHILDREN(rack_attack),
1165 	    OID_AUTO, "attacks", CTLFLAG_RD,
1166 	    &rack_sack_attacks_detected,
1167 	    "Total number of SACK attackers that had sack disabled");
1168 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1169 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1170 	    SYSCTL_CHILDREN(rack_attack),
1171 	    OID_AUTO, "reversed", CTLFLAG_RD,
1172 	    &rack_sack_attacks_reversed,
1173 	    "Total number of SACK attackers that were later determined false positive");
1174 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1175 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1176 	    SYSCTL_CHILDREN(rack_attack),
1177 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1178 	    &rack_sack_used_next_merge,
1179 	    "Total number of times we used the next merge");
1180 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1181 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1182 	    SYSCTL_CHILDREN(rack_attack),
1183 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1184 	    &rack_sack_used_prev_merge,
1185 	    "Total number of times we used the prev merge");
1186 	/* Counters */
1187 	rack_badfr = counter_u64_alloc(M_WAITOK);
1188 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1189 	    SYSCTL_CHILDREN(rack_counters),
1190 	    OID_AUTO, "badfr", CTLFLAG_RD,
1191 	    &rack_badfr, "Total number of bad FRs");
1192 	rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1193 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1194 	    SYSCTL_CHILDREN(rack_counters),
1195 	    OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1196 	    &rack_badfr_bytes, "Total number of bad FRs");
1197 	rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1198 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1199 	    SYSCTL_CHILDREN(rack_counters),
1200 	    OID_AUTO, "prrsndret", CTLFLAG_RD,
1201 	    &rack_rtm_prr_retran,
1202 	    "Total number of prr based retransmits");
1203 	rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1204 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1205 	    SYSCTL_CHILDREN(rack_counters),
1206 	    OID_AUTO, "prrsndnew", CTLFLAG_RD,
1207 	    &rack_rtm_prr_newdata,
1208 	    "Total number of prr based new transmits");
1209 	rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1210 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1211 	    SYSCTL_CHILDREN(rack_counters),
1212 	    OID_AUTO, "tsnf", CTLFLAG_RD,
1213 	    &rack_timestamp_mismatch,
1214 	    "Total number of timestamps that we could not find the reported ts");
1215 	rack_find_high = counter_u64_alloc(M_WAITOK);
1216 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1217 	    SYSCTL_CHILDREN(rack_counters),
1218 	    OID_AUTO, "findhigh", CTLFLAG_RD,
1219 	    &rack_find_high,
1220 	    "Total number of FIN causing find-high");
1221 	rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1222 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1223 	    SYSCTL_CHILDREN(rack_counters),
1224 	    OID_AUTO, "reordering", CTLFLAG_RD,
1225 	    &rack_reorder_seen,
1226 	    "Total number of times we added delay due to reordering");
1227 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1228 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1229 	    SYSCTL_CHILDREN(rack_counters),
1230 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1231 	    &rack_tlp_tot,
1232 	    "Total number of tail loss probe expirations");
1233 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1234 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1235 	    SYSCTL_CHILDREN(rack_counters),
1236 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1237 	    &rack_tlp_newdata,
1238 	    "Total number of tail loss probe sending new data");
1239 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1240 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1241 	    SYSCTL_CHILDREN(rack_counters),
1242 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1243 	    &rack_tlp_retran,
1244 	    "Total number of tail loss probe sending retransmitted data");
1245 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1246 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1247 	    SYSCTL_CHILDREN(rack_counters),
1248 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1249 	    &rack_tlp_retran_bytes,
1250 	    "Total bytes of tail loss probe sending retransmitted data");
1251 	rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1252 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1253 	    SYSCTL_CHILDREN(rack_counters),
1254 	    OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1255 	    &rack_tlp_retran_fail,
1256 	    "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1257 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1258 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1259 	    SYSCTL_CHILDREN(rack_counters),
1260 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1261 	    &rack_to_tot,
1262 	    "Total number of times the rack to expired");
1263 	rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1264 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1265 	    SYSCTL_CHILDREN(rack_counters),
1266 	    OID_AUTO, "arm_rack", CTLFLAG_RD,
1267 	    &rack_to_arm_rack,
1268 	    "Total number of times the rack timer armed");
1269 	rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1270 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1271 	    SYSCTL_CHILDREN(rack_counters),
1272 	    OID_AUTO, "arm_tlp", CTLFLAG_RD,
1273 	    &rack_to_arm_tlp,
1274 	    "Total number of times the tlp timer armed");
1275 	rack_calc_zero = counter_u64_alloc(M_WAITOK);
1276 	rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1277 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1278 	    SYSCTL_CHILDREN(rack_counters),
1279 	    OID_AUTO, "calc_zero", CTLFLAG_RD,
1280 	    &rack_calc_zero,
1281 	    "Total number of times pacing time worked out to zero");
1282 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1283 	    SYSCTL_CHILDREN(rack_counters),
1284 	    OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1285 	    &rack_calc_nonzero,
1286 	    "Total number of times pacing time worked out to non-zero");
1287 	rack_paced_segments = counter_u64_alloc(M_WAITOK);
1288 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1289 	    SYSCTL_CHILDREN(rack_counters),
1290 	    OID_AUTO, "paced", CTLFLAG_RD,
1291 	    &rack_paced_segments,
1292 	    "Total number of times a segment send caused hptsi");
1293 	rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1294 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1295 	    SYSCTL_CHILDREN(rack_counters),
1296 	    OID_AUTO, "unpaced", CTLFLAG_RD,
1297 	    &rack_unpaced_segments,
1298 	    "Total number of times a segment did not cause hptsi");
1299 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1300 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1301 	    SYSCTL_CHILDREN(rack_counters),
1302 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1303 	    &rack_saw_enobuf,
1304 	    "Total number of times a segment did not cause hptsi");
1305 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1306 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1307 	    SYSCTL_CHILDREN(rack_counters),
1308 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1309 	    &rack_saw_enetunreach,
1310 	    "Total number of times a segment did not cause hptsi");
1311 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1312 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1313 	    SYSCTL_CHILDREN(rack_counters),
1314 	    OID_AUTO, "allocs", CTLFLAG_RD,
1315 	    &rack_to_alloc,
1316 	    "Total allocations of tracking structures");
1317 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1318 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1319 	    SYSCTL_CHILDREN(rack_counters),
1320 	    OID_AUTO, "allochard", CTLFLAG_RD,
1321 	    &rack_to_alloc_hard,
1322 	    "Total allocations done with sleeping the hard way");
1323 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1324 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1325 	    SYSCTL_CHILDREN(rack_counters),
1326 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1327 	    &rack_to_alloc_emerg,
1328 	    "Total allocations done from emergency cache");
1329 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1330 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1331 	    SYSCTL_CHILDREN(rack_counters),
1332 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1333 	    &rack_to_alloc_limited,
1334 	    "Total allocations dropped due to limit");
1335 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1336 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1337 	    SYSCTL_CHILDREN(rack_counters),
1338 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1339 	    &rack_alloc_limited_conns,
1340 	    "Connections with allocations dropped due to limit");
1341 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1342 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1343 	    SYSCTL_CHILDREN(rack_counters),
1344 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1345 	    &rack_split_limited,
1346 	    "Split allocations dropped due to limit");
1347 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1348 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1349 	    SYSCTL_CHILDREN(rack_counters),
1350 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1351 	    &rack_sack_proc_all,
1352 	    "Total times we had to walk whole list for sack processing");
1353 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1354 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1355 	    SYSCTL_CHILDREN(rack_counters),
1356 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1357 	    &rack_sack_proc_restart,
1358 	    "Total times we had to walk whole list due to a restart");
1359 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1360 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1361 	    SYSCTL_CHILDREN(rack_counters),
1362 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1363 	    &rack_sack_proc_short,
1364 	    "Total times we took shortcut for sack processing");
1365 	rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1366 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1367 	    SYSCTL_CHILDREN(rack_counters),
1368 	    OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1369 	    &rack_enter_tlp_calc,
1370 	    "Total times we called calc-tlp");
1371 	rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1372 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1373 	    SYSCTL_CHILDREN(rack_counters),
1374 	    OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1375 	    &rack_used_tlpmethod,
1376 	    "Total number of runt sacks");
1377 	rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1378 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1379 	    SYSCTL_CHILDREN(rack_counters),
1380 	    OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1381 	    &rack_used_tlpmethod2,
1382 	    "Total number of times we hit TLP method 2");
1383 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1384 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1385 	    SYSCTL_CHILDREN(rack_attack),
1386 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1387 	    &rack_sack_skipped_acked,
1388 	    "Total number of times we skipped previously sacked");
1389 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1390 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1391 	    SYSCTL_CHILDREN(rack_attack),
1392 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1393 	    &rack_sack_splits,
1394 	    "Total number of times we did the old fashion tree split");
1395 	rack_progress_drops = counter_u64_alloc(M_WAITOK);
1396 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1397 	    SYSCTL_CHILDREN(rack_counters),
1398 	    OID_AUTO, "prog_drops", CTLFLAG_RD,
1399 	    &rack_progress_drops,
1400 	    "Total number of progress drops");
1401 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1402 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1403 	    SYSCTL_CHILDREN(rack_counters),
1404 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1405 	    &rack_input_idle_reduces,
1406 	    "Total number of idle reductions on input");
1407 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1408 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1409 	    SYSCTL_CHILDREN(rack_counters),
1410 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1411 	    &rack_collapsed_win,
1412 	    "Total number of collapsed windows");
1413 	rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1414 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1415 	    SYSCTL_CHILDREN(rack_counters),
1416 	    OID_AUTO, "tlp_nada", CTLFLAG_RD,
1417 	    &rack_tlp_does_nada,
1418 	    "Total number of nada tlp calls");
1419 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1420 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1421 	    SYSCTL_CHILDREN(rack_counters),
1422 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1423 	    &rack_try_scwnd,
1424 	    "Total number of scwnd attempts");
1425 
1426 	rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1427 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1428 	    SYSCTL_CHILDREN(rack_counters),
1429 	    OID_AUTO, "timer_hole", CTLFLAG_RD,
1430 	    &rack_per_timer_hole,
1431 	    "Total persists start in timer hole");
1432 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1433 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1434 	    OID_AUTO, "outsize", CTLFLAG_RD,
1435 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1436 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1437 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1438 	    OID_AUTO, "opts", CTLFLAG_RD,
1439 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1440 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1441 	    SYSCTL_CHILDREN(rack_sysctl_root),
1442 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1443 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1444 }
1445 
1446 static __inline int
1447 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1448 {
1449 	if (SEQ_GEQ(b->r_start, a->r_start) &&
1450 	    SEQ_LT(b->r_start, a->r_end)) {
1451 		/*
1452 		 * The entry b is within the
1453 		 * block a. i.e.:
1454 		 * a --   |-------------|
1455 		 * b --   |----|
1456 		 * <or>
1457 		 * b --       |------|
1458 		 * <or>
1459 		 * b --       |-----------|
1460 		 */
1461 		return (0);
1462 	} else if (SEQ_GEQ(b->r_start, a->r_end)) {
1463 		/*
1464 		 * b falls as either the next
1465 		 * sequence block after a so a
1466 		 * is said to be smaller than b.
1467 		 * i.e:
1468 		 * a --   |------|
1469 		 * b --          |--------|
1470 		 * or
1471 		 * b --              |-----|
1472 		 */
1473 		return (1);
1474 	}
1475 	/*
1476 	 * Whats left is where a is
1477 	 * larger than b. i.e:
1478 	 * a --         |-------|
1479 	 * b --  |---|
1480 	 * or even possibly
1481 	 * b --   |--------------|
1482 	 */
1483 	return (-1);
1484 }
1485 
1486 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1487 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1488 
1489 static uint32_t
1490 rc_init_window(struct tcp_rack *rack)
1491 {
1492 	uint32_t win;
1493 
1494 	if (rack->rc_init_win == 0) {
1495 		/*
1496 		 * Nothing set by the user, use the system stack
1497 		 * default.
1498 		 */
1499 		return(tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1500 	}
1501 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1502 	return(win);
1503 }
1504 
1505 static uint64_t
1506 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1507 {
1508 	if (IN_RECOVERY(rack->rc_tp->t_flags))
1509 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1510 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1511 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1512 	else
1513 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1514 }
1515 
1516 static uint64_t
1517 rack_get_bw(struct tcp_rack *rack)
1518 {
1519 	if (rack->use_fixed_rate) {
1520 		/* Return the fixed pacing rate */
1521 		return (rack_get_fixed_pacing_bw(rack));
1522 	}
1523 	if (rack->r_ctl.gp_bw == 0) {
1524 		/*
1525 		 * We have yet no b/w measurement,
1526 		 * if we have a user set initial bw
1527 		 * return it. If we don't have that and
1528 		 * we have an srtt, use the tcp IW (10) to
1529 		 * calculate a fictional b/w over the SRTT
1530 		 * which is more or less a guess. Note
1531 		 * we don't use our IW from rack on purpose
1532 		 * so if we have like IW=30, we are not
1533 		 * calculating a "huge" b/w.
1534 		 */
1535 		uint64_t bw, srtt;
1536 		if (rack->r_ctl.init_rate)
1537 			return (rack->r_ctl.init_rate);
1538 
1539 		/* Has the user set a max peak rate? */
1540 #ifdef NETFLIX_PEAKRATE
1541 		if (rack->rc_tp->t_maxpeakrate)
1542 			return (rack->rc_tp->t_maxpeakrate);
1543 #endif
1544 		/* Ok lets come up with the IW guess, if we have a srtt */
1545 		if (rack->rc_tp->t_srtt == 0) {
1546 			/*
1547 			 * Go with old pacing method
1548 			 * i.e. burst mitigation only.
1549 			 */
1550 			return (0);
1551 		}
1552 		/* Ok lets get the initial TCP win (not racks) */
1553 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1554 		srtt = ((uint64_t)TICKS_2_USEC(rack->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
1555 		bw *= (uint64_t)USECS_IN_SECOND;
1556 		bw /= srtt;
1557 		return (bw);
1558 	} else {
1559 		uint64_t bw;
1560 
1561 		if(rack->r_ctl.num_avg >= RACK_REQ_AVG) {
1562 			/* Averaging is done, we can return the value */
1563 			bw = rack->r_ctl.gp_bw;
1564 		} else {
1565 			/* Still doing initial average must calculate */
1566 			bw = rack->r_ctl.gp_bw / rack->r_ctl.num_avg;
1567 		}
1568 #ifdef NETFLIX_PEAKRATE
1569 		if ((rack->rc_tp->t_maxpeakrate) &&
1570 		    (bw > rack->rc_tp->t_maxpeakrate)) {
1571 			/* The user has set a peak rate to pace at
1572 			 * don't allow us to pace faster than that.
1573 			 */
1574 			return (rack->rc_tp->t_maxpeakrate);
1575 		}
1576 #endif
1577 		return (bw);
1578 	}
1579 }
1580 
1581 static uint16_t
1582 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1583 {
1584 	if (rack->use_fixed_rate) {
1585 		return (100);
1586 	} else if (rack->in_probe_rtt && (rsm == NULL))
1587 		return(rack->r_ctl.rack_per_of_gp_probertt);
1588 	else if ((IN_RECOVERY(rack->rc_tp->t_flags) &&
1589 		  rack->r_ctl.rack_per_of_gp_rec)) {
1590 		if (rsm) {
1591 			/* a retransmission always use the recovery rate */
1592 			return(rack->r_ctl.rack_per_of_gp_rec);
1593 		} else if (rack->rack_rec_nonrxt_use_cr) {
1594 			/* Directed to use the configured rate */
1595 			goto configured_rate;
1596 		} else if (rack->rack_no_prr &&
1597 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1598 			/* No PRR, lets just use the b/w estimate only */
1599 			return(100);
1600 		} else {
1601 			/*
1602 			 * Here we may have a non-retransmit but we
1603 			 * have no overrides, so just use the recovery
1604 			 * rate (prr is in effect).
1605 			 */
1606 			return(rack->r_ctl.rack_per_of_gp_rec);
1607 		}
1608 	}
1609 configured_rate:
1610 	/* For the configured rate we look at our cwnd vs the ssthresh */
1611 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1612 		return (rack->r_ctl.rack_per_of_gp_ss);
1613 	else
1614 		return(rack->r_ctl.rack_per_of_gp_ca);
1615 }
1616 
1617 static uint64_t
1618 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm)
1619 {
1620 	/*
1621 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
1622 	 */
1623 	uint64_t bw_est;
1624 	uint64_t gain;
1625 
1626 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
1627 	bw_est = bw * gain;
1628 	bw_est /= (uint64_t)100;
1629 	/* Never fall below the minimum (def 64kbps) */
1630 	if (bw_est < RACK_MIN_BW)
1631 		bw_est = RACK_MIN_BW;
1632 	return (bw_est);
1633 }
1634 
1635 static void
1636 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
1637 {
1638 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1639 		union tcp_log_stackspecific log;
1640 		struct timeval tv;
1641 
1642 		if ((mod != 1) && (rack_verbose_logging == 0)) {
1643 			/*
1644 			 * We get 3 values currently for mod
1645 			 * 1 - We are retransmitting and this tells the reason.
1646 			 * 2 - We are clearing a dup-ack count.
1647 			 * 3 - We are incrementing a dup-ack count.
1648 			 *
1649 			 * The clear/increment are only logged
1650 			 * if you have BBverbose on.
1651 			 */
1652 			return;
1653 		}
1654 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1655 		log.u_bbr.flex1 = tsused;
1656 		log.u_bbr.flex2 = thresh;
1657 		log.u_bbr.flex3 = rsm->r_flags;
1658 		log.u_bbr.flex4 = rsm->r_dupack;
1659 		log.u_bbr.flex5 = rsm->r_start;
1660 		log.u_bbr.flex6 = rsm->r_end;
1661 		log.u_bbr.flex8 = mod;
1662 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1663 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1664 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1665 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1666 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1667 		    &rack->rc_inp->inp_socket->so_rcv,
1668 		    &rack->rc_inp->inp_socket->so_snd,
1669 		    BBR_LOG_SETTINGS_CHG, 0,
1670 		    0, &log, false, &tv);
1671 	}
1672 }
1673 
1674 
1675 
1676 static void
1677 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
1678 {
1679 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1680 		union tcp_log_stackspecific log;
1681 		struct timeval tv;
1682 
1683 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1684 		log.u_bbr.flex1 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT);
1685 		log.u_bbr.flex2 = to * 1000;
1686 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
1687 		log.u_bbr.flex4 = slot;
1688 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
1689 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
1690 		log.u_bbr.flex7 = rack->rc_in_persist;
1691 		log.u_bbr.flex8 = which;
1692 		if (rack->rack_no_prr)
1693 			log.u_bbr.pkts_out = 0;
1694 		else
1695 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
1696 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1697 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1698 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1699 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1700 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1701 		    &rack->rc_inp->inp_socket->so_rcv,
1702 		    &rack->rc_inp->inp_socket->so_snd,
1703 		    BBR_LOG_TIMERSTAR, 0,
1704 		    0, &log, false, &tv);
1705 	}
1706 }
1707 
1708 static void
1709 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
1710 {
1711 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1712 		union tcp_log_stackspecific log;
1713 		struct timeval tv;
1714 
1715 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1716 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1717 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1718 		log.u_bbr.flex8 = to_num;
1719 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
1720 		log.u_bbr.flex2 = rack->rc_rack_rtt;
1721 		if (rsm == NULL)
1722 			log.u_bbr.flex3 = 0;
1723 		else
1724 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
1725 		if (rack->rack_no_prr)
1726 			log.u_bbr.flex5 = 0;
1727 		else
1728 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
1729 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1730 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1731 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1732 		    &rack->rc_inp->inp_socket->so_rcv,
1733 		    &rack->rc_inp->inp_socket->so_snd,
1734 		    BBR_LOG_RTO, 0,
1735 		    0, &log, false, &tv);
1736 	}
1737 }
1738 
1739 static void
1740 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
1741 		 struct rack_sendmap *rsm, int conf)
1742 {
1743 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1744 		union tcp_log_stackspecific log;
1745 		struct timeval tv;
1746 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1747 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1748 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1749 		log.u_bbr.flex1 = t;
1750 		log.u_bbr.flex2 = len;
1751 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt * HPTS_USEC_IN_MSEC;
1752 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest * HPTS_USEC_IN_MSEC;
1753 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest * HPTS_USEC_IN_MSEC;
1754 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_rtt_cnt;
1755 		log.u_bbr.flex7 = conf;
1756 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot * (uint64_t)HPTS_USEC_IN_MSEC;
1757 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
1758 		if (rack->rack_no_prr)
1759 			log.u_bbr.pkts_out = 0;
1760 		else
1761 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
1762 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1763 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtt;
1764 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
1765 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1766 		if (rsm) {
1767 			log.u_bbr.pkt_epoch = rsm->r_start;
1768 			log.u_bbr.lost = rsm->r_end;
1769 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
1770 		} else {
1771 
1772 			/* Its a SYN */
1773 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
1774 			log.u_bbr.lost = 0;
1775 			log.u_bbr.cwnd_gain = 0;
1776 		}
1777 		/* Write out general bits of interest rrs here */
1778 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
1779 		log.u_bbr.use_lt_bw <<= 1;
1780 		log.u_bbr.use_lt_bw |= rack->forced_ack;
1781 		log.u_bbr.use_lt_bw <<= 1;
1782 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
1783 		log.u_bbr.use_lt_bw <<= 1;
1784 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
1785 		log.u_bbr.use_lt_bw <<= 1;
1786 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
1787 		log.u_bbr.use_lt_bw <<= 1;
1788 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
1789 		log.u_bbr.use_lt_bw <<= 1;
1790 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
1791 		log.u_bbr.use_lt_bw <<= 1;
1792 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
1793 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
1794 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
1795 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
1796 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
1797 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
1798 		TCP_LOG_EVENTP(tp, NULL,
1799 		    &rack->rc_inp->inp_socket->so_rcv,
1800 		    &rack->rc_inp->inp_socket->so_snd,
1801 		    BBR_LOG_BBRRTT, 0,
1802 		    0, &log, false, &tv);
1803 	}
1804 }
1805 
1806 static void
1807 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
1808 {
1809 	/*
1810 	 * Log the rtt sample we are
1811 	 * applying to the srtt algorithm in
1812 	 * useconds.
1813 	 */
1814 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1815 		union tcp_log_stackspecific log;
1816 		struct timeval tv;
1817 
1818 		/* Convert our ms to a microsecond */
1819 		memset(&log, 0, sizeof(log));
1820 		log.u_bbr.flex1 = rtt * 1000;
1821 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
1822 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
1823 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
1824 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
1825 		log.u_bbr.flex8 = rack->sack_attack_disable;
1826 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1827 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1828 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1829 		    &rack->rc_inp->inp_socket->so_rcv,
1830 		    &rack->rc_inp->inp_socket->so_snd,
1831 		    TCP_LOG_RTT, 0,
1832 		    0, &log, false, &tv);
1833 	}
1834 }
1835 
1836 
1837 static inline void
1838 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
1839 {
1840 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
1841 		union tcp_log_stackspecific log;
1842 		struct timeval tv;
1843 
1844 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1845 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1846 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1847 		log.u_bbr.flex1 = line;
1848 		log.u_bbr.flex2 = tick;
1849 		log.u_bbr.flex3 = tp->t_maxunacktime;
1850 		log.u_bbr.flex4 = tp->t_acktime;
1851 		log.u_bbr.flex8 = event;
1852 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1853 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1854 		TCP_LOG_EVENTP(tp, NULL,
1855 		    &rack->rc_inp->inp_socket->so_rcv,
1856 		    &rack->rc_inp->inp_socket->so_snd,
1857 		    BBR_LOG_PROGRESS, 0,
1858 		    0, &log, false, &tv);
1859 	}
1860 }
1861 
1862 static void
1863 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
1864 {
1865 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1866 		union tcp_log_stackspecific log;
1867 
1868 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1869 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1870 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1871 		log.u_bbr.flex1 = slot;
1872 		if (rack->rack_no_prr)
1873 			log.u_bbr.flex2 = 0;
1874 		else
1875 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
1876 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
1877 		log.u_bbr.flex8 = rack->rc_in_persist;
1878 		log.u_bbr.timeStamp = cts;
1879 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1880 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1881 		    &rack->rc_inp->inp_socket->so_rcv,
1882 		    &rack->rc_inp->inp_socket->so_snd,
1883 		    BBR_LOG_BBRSND, 0,
1884 		    0, &log, false, tv);
1885 	}
1886 }
1887 
1888 static void
1889 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out)
1890 {
1891 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1892 		union tcp_log_stackspecific log;
1893 		struct timeval tv;
1894 
1895 		memset(&log, 0, sizeof(log));
1896 		log.u_bbr.flex1 = did_out;
1897 		log.u_bbr.flex2 = nxt_pkt;
1898 		log.u_bbr.flex3 = way_out;
1899 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
1900 		if (rack->rack_no_prr)
1901 			log.u_bbr.flex5 = 0;
1902 		else
1903 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
1904 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
1905 		log.u_bbr.flex7 = rack->r_wanted_output;
1906 		log.u_bbr.flex8 = rack->rc_in_persist;
1907 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1908 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1909 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1910 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1911 		    &rack->rc_inp->inp_socket->so_rcv,
1912 		    &rack->rc_inp->inp_socket->so_snd,
1913 		    BBR_LOG_DOSEG_DONE, 0,
1914 		    0, &log, false, &tv);
1915 	}
1916 }
1917 
1918 static void
1919 rack_log_type_hrdwtso(struct tcpcb *tp, struct tcp_rack *rack, int len, int mod, int32_t orig_len, int frm)
1920 {
1921 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1922 		union tcp_log_stackspecific log;
1923 		struct timeval tv;
1924 		uint32_t cts;
1925 
1926 		memset(&log, 0, sizeof(log));
1927 		cts = tcp_get_usecs(&tv);
1928 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
1929 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
1930 		log.u_bbr.flex4 = len;
1931 		log.u_bbr.flex5 = orig_len;
1932 		log.u_bbr.flex6 = rack->r_ctl.rc_sacked;
1933 		log.u_bbr.flex7 = mod;
1934 		log.u_bbr.flex8 = frm;
1935 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1936 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1937 		TCP_LOG_EVENTP(tp, NULL,
1938 		    &tp->t_inpcb->inp_socket->so_rcv,
1939 		    &tp->t_inpcb->inp_socket->so_snd,
1940 		    TCP_HDWR_TLS, 0,
1941 		    0, &log, false, &tv);
1942 	}
1943 }
1944 
1945 static void
1946 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
1947 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
1948 {
1949 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1950 		union tcp_log_stackspecific log;
1951 		struct timeval tv;
1952 
1953 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1954 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1955 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1956 		log.u_bbr.flex1 = slot;
1957 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
1958 		log.u_bbr.flex4 = reason;
1959 		if (rack->rack_no_prr)
1960 			log.u_bbr.flex5 = 0;
1961 		else
1962 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
1963 		log.u_bbr.flex7 = hpts_calling;
1964 		log.u_bbr.flex8 = rack->rc_in_persist;
1965 		log.u_bbr.lt_epoch = cwnd_to_use;
1966 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1967 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
1968 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1969 		    &rack->rc_inp->inp_socket->so_rcv,
1970 		    &rack->rc_inp->inp_socket->so_snd,
1971 		    BBR_LOG_JUSTRET, 0,
1972 		    tlen, &log, false, &tv);
1973 	}
1974 }
1975 
1976 static void
1977 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
1978 		   struct timeval *tv, uint32_t flags_on_entry)
1979 {
1980 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1981 		union tcp_log_stackspecific log;
1982 
1983 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1984 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1985 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1986 		log.u_bbr.flex1 = line;
1987 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
1988 		log.u_bbr.flex3 = flags_on_entry;
1989 		log.u_bbr.flex4 = us_cts;
1990 		if (rack->rack_no_prr)
1991 			log.u_bbr.flex5 = 0;
1992 		else
1993 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
1994 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
1995 		log.u_bbr.flex7 = hpts_removed;
1996 		log.u_bbr.flex8 = 1;
1997 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
1998 		log.u_bbr.timeStamp = us_cts;
1999 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2000 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2001 		    &rack->rc_inp->inp_socket->so_rcv,
2002 		    &rack->rc_inp->inp_socket->so_snd,
2003 		    BBR_LOG_TIMERCANC, 0,
2004 		    0, &log, false, tv);
2005 	}
2006 }
2007 
2008 static void
2009 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2010 			  uint32_t flex1, uint32_t flex2,
2011 			  uint32_t flex3, uint32_t flex4,
2012 			  uint32_t flex5, uint32_t flex6,
2013 			  uint16_t flex7, uint8_t mod)
2014 {
2015 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2016 		union tcp_log_stackspecific log;
2017 		struct timeval tv;
2018 
2019 		if (mod == 1) {
2020 			/* No you can't use 1, its for the real to cancel */
2021 			return;
2022 		}
2023 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2024 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2025 		log.u_bbr.flex1 = flex1;
2026 		log.u_bbr.flex2 = flex2;
2027 		log.u_bbr.flex3 = flex3;
2028 		log.u_bbr.flex4 = flex4;
2029 		log.u_bbr.flex5 = flex5;
2030 		log.u_bbr.flex6 = flex6;
2031 		log.u_bbr.flex7 = flex7;
2032 		log.u_bbr.flex8 =  mod;
2033 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2034 		    &rack->rc_inp->inp_socket->so_rcv,
2035 		    &rack->rc_inp->inp_socket->so_snd,
2036 		    BBR_LOG_TIMERCANC, 0,
2037 		    0, &log, false, &tv);
2038 	}
2039 }
2040 
2041 static void
2042 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2043 {
2044 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2045 		union tcp_log_stackspecific log;
2046 		struct timeval tv;
2047 
2048 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2049 		log.u_bbr.flex1 = timers;
2050 		log.u_bbr.flex2 = ret;
2051 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2052 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2053 		log.u_bbr.flex5 = cts;
2054 		if (rack->rack_no_prr)
2055 			log.u_bbr.flex6 = 0;
2056 		else
2057 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2058 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2059 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2060 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2061 		    &rack->rc_inp->inp_socket->so_rcv,
2062 		    &rack->rc_inp->inp_socket->so_snd,
2063 		    BBR_LOG_TO_PROCESS, 0,
2064 		    0, &log, false, &tv);
2065 	}
2066 }
2067 
2068 static void
2069 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2070 {
2071 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2072 		union tcp_log_stackspecific log;
2073 		struct timeval tv;
2074 
2075 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2076 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2077 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2078 		if (rack->rack_no_prr)
2079 			log.u_bbr.flex3 = 0;
2080 		else
2081 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2082 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2083 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2084 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2085 		log.u_bbr.flex8 = frm;
2086 		log.u_bbr.pkts_out = orig_cwnd;
2087 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2088 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2089 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2090 		    &rack->rc_inp->inp_socket->so_rcv,
2091 		    &rack->rc_inp->inp_socket->so_snd,
2092 		    BBR_LOG_BBRUPD, 0,
2093 		    0, &log, false, &tv);
2094 	}
2095 }
2096 
2097 #ifdef NETFLIX_EXP_DETECTION
2098 static void
2099 rack_log_sad(struct tcp_rack *rack, int event)
2100 {
2101 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2102 		union tcp_log_stackspecific log;
2103 		struct timeval tv;
2104 
2105 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2106 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
2107 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2108 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2109 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2110 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2111 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2112 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2113 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2114 		log.u_bbr.lt_epoch |= rack->do_detection;
2115 		log.u_bbr.applimited = tcp_map_minimum;
2116 		log.u_bbr.flex7 = rack->sack_attack_disable;
2117 		log.u_bbr.flex8 = event;
2118 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2119 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2120 		log.u_bbr.delivered = tcp_sad_decay_val;
2121 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2122 		    &rack->rc_inp->inp_socket->so_rcv,
2123 		    &rack->rc_inp->inp_socket->so_snd,
2124 		    TCP_SAD_DETECTION, 0,
2125 		    0, &log, false, &tv);
2126 	}
2127 }
2128 #endif
2129 
2130 static void
2131 rack_counter_destroy(void)
2132 {
2133 	counter_u64_free(rack_ack_total);
2134 	counter_u64_free(rack_express_sack);
2135 	counter_u64_free(rack_sack_total);
2136 	counter_u64_free(rack_move_none);
2137 	counter_u64_free(rack_move_some);
2138 	counter_u64_free(rack_sack_attacks_detected);
2139 	counter_u64_free(rack_sack_attacks_reversed);
2140 	counter_u64_free(rack_sack_used_next_merge);
2141 	counter_u64_free(rack_sack_used_prev_merge);
2142 	counter_u64_free(rack_badfr);
2143 	counter_u64_free(rack_badfr_bytes);
2144 	counter_u64_free(rack_rtm_prr_retran);
2145 	counter_u64_free(rack_rtm_prr_newdata);
2146 	counter_u64_free(rack_timestamp_mismatch);
2147 	counter_u64_free(rack_find_high);
2148 	counter_u64_free(rack_reorder_seen);
2149 	counter_u64_free(rack_tlp_tot);
2150 	counter_u64_free(rack_tlp_newdata);
2151 	counter_u64_free(rack_tlp_retran);
2152 	counter_u64_free(rack_tlp_retran_bytes);
2153 	counter_u64_free(rack_tlp_retran_fail);
2154 	counter_u64_free(rack_to_tot);
2155 	counter_u64_free(rack_to_arm_rack);
2156 	counter_u64_free(rack_to_arm_tlp);
2157 	counter_u64_free(rack_calc_zero);
2158 	counter_u64_free(rack_calc_nonzero);
2159 	counter_u64_free(rack_paced_segments);
2160 	counter_u64_free(rack_unpaced_segments);
2161 	counter_u64_free(rack_saw_enobuf);
2162 	counter_u64_free(rack_saw_enetunreach);
2163 	counter_u64_free(rack_to_alloc);
2164 	counter_u64_free(rack_to_alloc_hard);
2165 	counter_u64_free(rack_to_alloc_emerg);
2166 	counter_u64_free(rack_to_alloc_limited);
2167 	counter_u64_free(rack_alloc_limited_conns);
2168 	counter_u64_free(rack_split_limited);
2169 	counter_u64_free(rack_sack_proc_all);
2170 	counter_u64_free(rack_sack_proc_restart);
2171 	counter_u64_free(rack_sack_proc_short);
2172 	counter_u64_free(rack_enter_tlp_calc);
2173 	counter_u64_free(rack_used_tlpmethod);
2174 	counter_u64_free(rack_used_tlpmethod2);
2175 	counter_u64_free(rack_sack_skipped_acked);
2176 	counter_u64_free(rack_sack_splits);
2177 	counter_u64_free(rack_progress_drops);
2178 	counter_u64_free(rack_input_idle_reduces);
2179 	counter_u64_free(rack_collapsed_win);
2180 	counter_u64_free(rack_tlp_does_nada);
2181 	counter_u64_free(rack_try_scwnd);
2182 	counter_u64_free(rack_per_timer_hole);
2183 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2184 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2185 }
2186 
2187 static struct rack_sendmap *
2188 rack_alloc(struct tcp_rack *rack)
2189 {
2190 	struct rack_sendmap *rsm;
2191 
2192 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
2193 	if (rsm) {
2194 		rack->r_ctl.rc_num_maps_alloced++;
2195 		counter_u64_add(rack_to_alloc, 1);
2196 		return (rsm);
2197 	}
2198 	if (rack->rc_free_cnt) {
2199 		counter_u64_add(rack_to_alloc_emerg, 1);
2200 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2201 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2202 		rack->rc_free_cnt--;
2203 		return (rsm);
2204 	}
2205 	return (NULL);
2206 }
2207 
2208 static struct rack_sendmap *
2209 rack_alloc_full_limit(struct tcp_rack *rack)
2210 {
2211 	if ((V_tcp_map_entries_limit > 0) &&
2212 	    (rack->do_detection == 0) &&
2213 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2214 		counter_u64_add(rack_to_alloc_limited, 1);
2215 		if (!rack->alloc_limit_reported) {
2216 			rack->alloc_limit_reported = 1;
2217 			counter_u64_add(rack_alloc_limited_conns, 1);
2218 		}
2219 		return (NULL);
2220 	}
2221 	return (rack_alloc(rack));
2222 }
2223 
2224 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2225 static struct rack_sendmap *
2226 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2227 {
2228 	struct rack_sendmap *rsm;
2229 
2230 	if (limit_type) {
2231 		/* currently there is only one limit type */
2232 		if (V_tcp_map_split_limit > 0 &&
2233 		    (rack->do_detection == 0) &&
2234 		    rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2235 			counter_u64_add(rack_split_limited, 1);
2236 			if (!rack->alloc_limit_reported) {
2237 				rack->alloc_limit_reported = 1;
2238 				counter_u64_add(rack_alloc_limited_conns, 1);
2239 			}
2240 			return (NULL);
2241 		}
2242 	}
2243 
2244 	/* allocate and mark in the limit type, if set */
2245 	rsm = rack_alloc(rack);
2246 	if (rsm != NULL && limit_type) {
2247 		rsm->r_limit_type = limit_type;
2248 		rack->r_ctl.rc_num_split_allocs++;
2249 	}
2250 	return (rsm);
2251 }
2252 
2253 static void
2254 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2255 {
2256 	if (rsm->r_flags & RACK_APP_LIMITED) {
2257 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
2258 			rack->r_ctl.rc_app_limited_cnt--;
2259 		}
2260 	}
2261 	if (rsm->r_limit_type) {
2262 		/* currently there is only one limit type */
2263 		rack->r_ctl.rc_num_split_allocs--;
2264 	}
2265 	if (rsm == rack->r_ctl.rc_first_appl) {
2266 		if (rack->r_ctl.rc_app_limited_cnt == 0)
2267 			rack->r_ctl.rc_first_appl = NULL;
2268 		else {
2269 			/* Follow the next one out */
2270 			struct rack_sendmap fe;
2271 
2272 			fe.r_start = rsm->r_nseq_appl;
2273 			rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2274 		}
2275 	}
2276 	if (rsm == rack->r_ctl.rc_resend)
2277 		rack->r_ctl.rc_resend = NULL;
2278 	if (rsm == rack->r_ctl.rc_rsm_at_retran)
2279 		rack->r_ctl.rc_rsm_at_retran = NULL;
2280 	if (rsm == rack->r_ctl.rc_end_appl)
2281 		rack->r_ctl.rc_end_appl = NULL;
2282 	if (rack->r_ctl.rc_tlpsend == rsm)
2283 		rack->r_ctl.rc_tlpsend = NULL;
2284 	if (rack->r_ctl.rc_sacklast == rsm)
2285 		rack->r_ctl.rc_sacklast = NULL;
2286 	if (rack->rc_free_cnt < rack_free_cache) {
2287 		memset(rsm, 0, sizeof(struct rack_sendmap));
2288 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
2289 		rsm->r_limit_type = 0;
2290 		rack->rc_free_cnt++;
2291 		return;
2292 	}
2293 	rack->r_ctl.rc_num_maps_alloced--;
2294 	uma_zfree(rack_zone, rsm);
2295 }
2296 
2297 static uint32_t
2298 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2299 {
2300 	uint64_t srtt, bw, len, tim;
2301 	uint32_t segsiz, def_len, minl;
2302 
2303 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2304 	def_len = rack_def_data_window * segsiz;
2305 	if (rack->rc_gp_filled == 0) {
2306 		/*
2307 		 * We have no measurement (IW is in flight?) so
2308 		 * we can only guess using our data_window sysctl
2309 		 * value (usually 100MSS).
2310 		 */
2311 		return (def_len);
2312 	}
2313 	/*
2314 	 * Now we have a number of factors to consider.
2315 	 *
2316 	 * 1) We have a desired BDP which is usually
2317 	 *    at least 2.
2318 	 * 2) We have a minimum number of rtt's usually 1 SRTT
2319 	 *    but we allow it too to be more.
2320 	 * 3) We want to make sure a measurement last N useconds (if
2321 	 *    we have set rack_min_measure_usec.
2322 	 *
2323 	 * We handle the first concern here by trying to create a data
2324 	 * window of max(rack_def_data_window, DesiredBDP). The
2325 	 * second concern we handle in not letting the measurement
2326 	 * window end normally until at least the required SRTT's
2327 	 * have gone by which is done further below in
2328 	 * rack_enough_for_measurement(). Finally the third concern
2329 	 * we also handle here by calculating how long that time
2330 	 * would take at the current BW and then return the
2331 	 * max of our first calculation and that length. Note
2332 	 * that if rack_min_measure_usec is 0, we don't deal
2333 	 * with concern 3. Also for both Concern 1 and 3 an
2334 	 * application limited period could end the measurement
2335 	 * earlier.
2336 	 *
2337 	 * So lets calculate the BDP with the "known" b/w using
2338 	 * the SRTT has our rtt and then multiply it by the
2339 	 * goal.
2340 	 */
2341 	bw = rack_get_bw(rack);
2342 	srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
2343 	len = bw * srtt;
2344 	len /= (uint64_t)HPTS_USEC_IN_SEC;
2345 	len *= max(1, rack_goal_bdp);
2346         /* Now we need to round up to the nearest MSS */
2347 	len = roundup(len, segsiz);
2348 	if (rack_min_measure_usec) {
2349 		/* Now calculate our min length for this b/w */
2350 		tim = rack_min_measure_usec;
2351 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
2352 		if (minl == 0)
2353 			minl = 1;
2354 		minl = roundup(minl, segsiz);
2355 		if (len < minl)
2356 			len = minl;
2357 	}
2358 	/*
2359 	 * Now if we have a very small window we want
2360 	 * to attempt to get the window that is
2361 	 * as small as possible. This happens on
2362 	 * low b/w connections and we don't want to
2363 	 * span huge numbers of rtt's between measurements.
2364 	 *
2365 	 * We basically include 2 over our "MIN window" so
2366 	 * that the measurement can be shortened (possibly) by
2367 	 * an ack'ed packet.
2368 	 */
2369 	if (len < def_len)
2370 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
2371 	else
2372 		return (max((uint32_t)len, def_len));
2373 
2374 }
2375 
2376 static int
2377 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack)
2378 {
2379 	uint32_t tim, srtts, segsiz;
2380 
2381 	/*
2382 	 * Has enough time passed for the GP measurement to be valid?
2383 	 */
2384 	if ((tp->snd_max == tp->snd_una) ||
2385 	    (th_ack == tp->snd_max)){
2386 		/* All is acked */
2387 		return (1);
2388 	}
2389 	if (SEQ_LT(th_ack, tp->gput_seq)) {
2390 		/* Not enough bytes yet */
2391 		return (0);
2392 	}
2393 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2394 	if (SEQ_LT(th_ack, tp->gput_ack) &&
2395 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
2396 		/* Not enough bytes yet */
2397 		return (0);
2398 	}
2399 	if (rack->r_ctl.rc_first_appl &&
2400 	    (rack->r_ctl.rc_first_appl->r_start == th_ack)) {
2401 		/*
2402 		 * We are up to the app limited point
2403 		 * we have to measure irrespective of the time..
2404 		 */
2405 		return (1);
2406 	}
2407 	/* Now what about time? */
2408 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
2409 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
2410 	if (tim >= srtts) {
2411 		return (1);
2412 	}
2413 	/* Nope not even a full SRTT has passed */
2414 	return (0);
2415 }
2416 
2417 
2418 static void
2419 rack_log_timely(struct tcp_rack *rack,
2420 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
2421 		uint64_t up_bnd, int line, uint8_t method)
2422 {
2423 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2424 		union tcp_log_stackspecific log;
2425 		struct timeval tv;
2426 
2427 		memset(&log, 0, sizeof(log));
2428 		log.u_bbr.flex1 = logged;
2429 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
2430 		log.u_bbr.flex2 <<= 4;
2431 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
2432 		log.u_bbr.flex2 <<= 4;
2433 		log.u_bbr.flex2 |= rack->rc_gp_incr;
2434 		log.u_bbr.flex2 <<= 4;
2435 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
2436 		log.u_bbr.flex3 = rack->rc_gp_incr;
2437 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
2438 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
2439 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
2440 		log.u_bbr.flex7 = rack->rc_gp_bwred;
2441 		log.u_bbr.flex8 = method;
2442 		log.u_bbr.cur_del_rate = cur_bw;
2443 		log.u_bbr.delRate = low_bnd;
2444 		log.u_bbr.bw_inuse = up_bnd;
2445 		log.u_bbr.rttProp = rack_get_bw(rack);
2446 		log.u_bbr.pkt_epoch = line;
2447 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
2448 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2449 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2450 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
2451 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
2452 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
2453 		log.u_bbr.cwnd_gain <<= 1;
2454 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
2455 		log.u_bbr.cwnd_gain <<= 1;
2456 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
2457 		log.u_bbr.cwnd_gain <<= 1;
2458 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
2459 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
2460 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2461 		    &rack->rc_inp->inp_socket->so_rcv,
2462 		    &rack->rc_inp->inp_socket->so_snd,
2463 		    TCP_TIMELY_WORK, 0,
2464 		    0, &log, false, &tv);
2465 	}
2466 }
2467 
2468 static int
2469 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
2470 {
2471 	/*
2472 	 * Before we increase we need to know if
2473 	 * the estimate just made was less than
2474 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
2475 	 *
2476 	 * If we already are pacing at a fast enough
2477 	 * rate to push us faster there is no sense of
2478 	 * increasing.
2479 	 *
2480 	 * We first caculate our actual pacing rate (ss or ca multipler
2481 	 * times our cur_bw).
2482 	 *
2483 	 * Then we take the last measured rate and multipy by our
2484 	 * maximum pacing overage to give us a max allowable rate.
2485 	 *
2486 	 * If our act_rate is smaller than our max_allowable rate
2487 	 * then we should increase. Else we should hold steady.
2488 	 *
2489 	 */
2490 	uint64_t act_rate, max_allow_rate;
2491 
2492 	if (rack_timely_no_stopping)
2493 		return (1);
2494 
2495 	if ((cur_bw == 0) || (last_bw_est == 0)) {
2496 		/*
2497 		 * Initial startup case or
2498 		 * everything is acked case.
2499 		 */
2500 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
2501 				__LINE__, 9);
2502 		return (1);
2503 	}
2504 	if (mult <= 100) {
2505 		/*
2506 		 * We can always pace at or slightly above our rate.
2507 		 */
2508 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
2509 				__LINE__, 9);
2510 		return (1);
2511 	}
2512 	act_rate = cur_bw * (uint64_t)mult;
2513 	act_rate /= 100;
2514 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
2515 	max_allow_rate /= 100;
2516 	if (act_rate < max_allow_rate) {
2517 		/*
2518 		 * Here the rate we are actually pacing at
2519 		 * is smaller than 10% above our last measurement.
2520 		 * This means we are pacing below what we would
2521 		 * like to try to achieve (plus some wiggle room).
2522 		 */
2523 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
2524 				__LINE__, 9);
2525 		return (1);
2526 	} else {
2527 		/*
2528 		 * Here we are already pacing at least rack_max_per_above(10%)
2529 		 * what we are getting back. This indicates most likely
2530 		 * that we are being limited (cwnd/rwnd/app) and can't
2531 		 * get any more b/w. There is no sense of trying to
2532 		 * raise up the pacing rate its not speeding us up
2533 		 * and we already are pacing faster than we are getting.
2534 		 */
2535 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
2536 				__LINE__, 8);
2537 		return (0);
2538 	}
2539 }
2540 
2541 static void
2542 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
2543 {
2544 	/*
2545 	 * When we drag bottom, we want to assure
2546 	 * that no multiplier is below 1.0, if so
2547 	 * we want to restore it to at least that.
2548 	 */
2549 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
2550 		/* This is unlikely we usually do not touch recovery */
2551 		rack->r_ctl.rack_per_of_gp_rec = 100;
2552 	}
2553 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
2554 		rack->r_ctl.rack_per_of_gp_ca = 100;
2555 	}
2556 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
2557 		rack->r_ctl.rack_per_of_gp_ss = 100;
2558 	}
2559 }
2560 
2561 static void
2562 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
2563 {
2564 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
2565 		rack->r_ctl.rack_per_of_gp_ca = 100;
2566 	}
2567 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
2568 		rack->r_ctl.rack_per_of_gp_ss = 100;
2569 	}
2570 }
2571 
2572 static void
2573 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
2574 {
2575 	int32_t  calc, logged, plus;
2576 
2577 	logged = 0;
2578 
2579 	if (override) {
2580 		/*
2581 		 * override is passed when we are
2582 		 * loosing b/w and making one last
2583 		 * gasp at trying to not loose out
2584 		 * to a new-reno flow.
2585 		 */
2586 		goto extra_boost;
2587 	}
2588 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
2589 	if (rack->rc_gp_incr &&
2590 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
2591 		/*
2592 		 * Reset and get 5 strokes more before the boost. Note
2593 		 * that the count is 0 based so we have to add one.
2594 		 */
2595 extra_boost:
2596 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
2597 		rack->rc_gp_timely_inc_cnt = 0;
2598 	} else
2599 		plus = (uint32_t)rack_gp_increase_per;
2600 	/* Must be at least 1% increase for true timely increases */
2601 	if ((plus < 1) &&
2602 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
2603 		plus = 1;
2604 	if (rack->rc_gp_saw_rec &&
2605 	    (rack->rc_gp_no_rec_chg == 0) &&
2606 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
2607 				  rack->r_ctl.rack_per_of_gp_rec)) {
2608 		/* We have been in recovery ding it too */
2609 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
2610 		if (calc > 0xffff)
2611 			calc = 0xffff;
2612 		logged |= 1;
2613 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
2614 		if (rack_per_upper_bound_ss &&
2615 		    (rack->rc_dragged_bottom == 0) &&
2616 		    (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
2617 			rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
2618 	}
2619 	if (rack->rc_gp_saw_ca &&
2620 	    (rack->rc_gp_saw_ss == 0) &&
2621 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
2622 				  rack->r_ctl.rack_per_of_gp_ca)) {
2623 		/* In CA */
2624 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
2625 		if (calc > 0xffff)
2626 			calc = 0xffff;
2627 		logged |= 2;
2628 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
2629 		if (rack_per_upper_bound_ca &&
2630 		    (rack->rc_dragged_bottom == 0) &&
2631 		    (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
2632 			rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
2633 	}
2634 	if (rack->rc_gp_saw_ss &&
2635 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
2636 				  rack->r_ctl.rack_per_of_gp_ss)) {
2637 		/* In SS */
2638 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
2639 		if (calc > 0xffff)
2640 			calc = 0xffff;
2641 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
2642 		if (rack_per_upper_bound_ss &&
2643 		    (rack->rc_dragged_bottom == 0) &&
2644 		    (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
2645 			rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
2646 		logged |= 4;
2647 	}
2648 	if (logged &&
2649 	    (rack->rc_gp_incr == 0)){
2650 		/* Go into increment mode */
2651 		rack->rc_gp_incr = 1;
2652 		rack->rc_gp_timely_inc_cnt = 0;
2653 	}
2654 	if (rack->rc_gp_incr &&
2655 	    logged &&
2656 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
2657 		rack->rc_gp_timely_inc_cnt++;
2658 	}
2659 	rack_log_timely(rack,  logged, plus, 0, 0,
2660 			__LINE__, 1);
2661 }
2662 
2663 static uint32_t
2664 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
2665 {
2666 	/*
2667 	 * norm_grad = rtt_diff / minrtt;
2668 	 * new_per = curper  * (1 - B * norm_grad)
2669 	 *
2670 	 * B = rack_gp_decrease_per (default 10%)
2671 	 * rtt_dif = input var current rtt-diff
2672 	 * curper = input var current percentage
2673 	 * minrtt = from rack filter
2674 	 *
2675 	 */
2676 	uint64_t perf;
2677 
2678 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
2679 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
2680 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
2681 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
2682 		     (uint64_t)1000000)) /
2683 		(uint64_t)1000000);
2684 	if (perf > curper) {
2685 		/* TSNH */
2686 		perf = curper - 1;
2687 	}
2688 	return ((uint32_t)perf);
2689 }
2690 
2691 static uint32_t
2692 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
2693 {
2694 	/*
2695 	 *                                   highrttthresh
2696 	 * result = curper * (1 - (B * ( 1 -  ------          ))
2697 	 *                                     gp_srtt
2698 	 *
2699 	 * B = rack_gp_decrease_per (default 10%)
2700 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
2701 	 */
2702 	uint64_t perf;
2703 	uint32_t highrttthresh;
2704 
2705 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
2706 
2707 	perf =  (((uint64_t)curper * ((uint64_t)1000000 -
2708 				    ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
2709 					((uint64_t)highrttthresh * (uint64_t)1000000) /
2710 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
2711 	return (perf);
2712 }
2713 
2714 
2715 static void
2716 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
2717 {
2718 	uint64_t logvar, logvar2, logvar3;
2719 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
2720 
2721 	if (rack->rc_gp_incr) {
2722 		/* Turn off increment counting  */
2723 		rack->rc_gp_incr = 0;
2724 		rack->rc_gp_timely_inc_cnt = 0;
2725 	}
2726 	ss_red = ca_red = rec_red = 0;
2727 	logged = 0;
2728 	/* Calculate the reduction value */
2729 	if (rtt_diff < 0) {
2730 		rtt_diff *= -1;
2731 	}
2732 	/* Must be at least 1% reduction */
2733 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
2734 		/* We have been in recovery ding it too */
2735 		if (timely_says == 2) {
2736 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
2737 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
2738 			if (alt < new_per)
2739 				val = alt;
2740 			else
2741 				val = new_per;
2742 		} else
2743 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
2744 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
2745 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
2746 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
2747 		} else {
2748 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
2749 			rec_red = 0;
2750 		}
2751 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
2752 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
2753 		logged |= 1;
2754 	}
2755 	if (rack->rc_gp_saw_ss) {
2756 		/* Sent in SS */
2757 		if (timely_says == 2) {
2758 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
2759 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
2760 			if (alt < new_per)
2761 				val = alt;
2762 			else
2763 				val = new_per;
2764 		} else
2765 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
2766 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
2767 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
2768 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
2769 		} else {
2770 			ss_red = new_per;
2771 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
2772 			logvar = new_per;
2773 			logvar <<= 32;
2774 			logvar |= alt;
2775 			logvar2 = (uint32_t)rtt;
2776 			logvar2 <<= 32;
2777 			logvar2 |= (uint32_t)rtt_diff;
2778 			logvar3 = rack_gp_rtt_maxmul;
2779 			logvar3 <<= 32;
2780 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
2781 			rack_log_timely(rack, timely_says,
2782 					logvar2, logvar3,
2783 					logvar, __LINE__, 10);
2784 		}
2785 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
2786 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
2787 		logged |= 4;
2788 	} else 	if (rack->rc_gp_saw_ca) {
2789 		/* Sent in CA */
2790 		if (timely_says == 2) {
2791 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
2792 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
2793 			if (alt < new_per)
2794 				val = alt;
2795 			else
2796 				val = new_per;
2797 		} else
2798 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
2799 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
2800 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
2801 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
2802 		} else {
2803 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
2804 			ca_red = 0;
2805 			logvar = new_per;
2806 			logvar <<= 32;
2807 			logvar |= alt;
2808 			logvar2 = (uint32_t)rtt;
2809 			logvar2 <<= 32;
2810 			logvar2 |= (uint32_t)rtt_diff;
2811 			logvar3 = rack_gp_rtt_maxmul;
2812 			logvar3 <<= 32;
2813 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
2814 			rack_log_timely(rack, timely_says,
2815 					logvar2, logvar3,
2816 					logvar, __LINE__, 10);
2817 		}
2818 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
2819 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
2820 		logged |= 2;
2821 	}
2822 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
2823 		rack->rc_gp_timely_dec_cnt++;
2824 		if (rack_timely_dec_clear &&
2825 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
2826 			rack->rc_gp_timely_dec_cnt = 0;
2827 	}
2828 	logvar = ss_red;
2829 	logvar <<= 32;
2830 	logvar |= ca_red;
2831 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
2832 			__LINE__, 2);
2833 }
2834 
2835 static void
2836 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
2837 		     uint32_t rtt, uint32_t line, uint8_t reas)
2838 {
2839 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2840 		union tcp_log_stackspecific log;
2841 		struct timeval tv;
2842 
2843 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2844 		log.u_bbr.flex1 = line;
2845 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
2846 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
2847 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
2848 		log.u_bbr.flex5 = rtt;
2849 		log.u_bbr.flex6 = rack->rc_highly_buffered;
2850 		log.u_bbr.flex6 <<= 1;
2851 		log.u_bbr.flex6 |= rack->forced_ack;
2852 		log.u_bbr.flex6 <<= 1;
2853 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
2854 		log.u_bbr.flex6 <<= 1;
2855 		log.u_bbr.flex6 |= rack->in_probe_rtt;
2856 		log.u_bbr.flex6 <<= 1;
2857 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
2858 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
2859 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
2860 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
2861 		log.u_bbr.flex8 = reas;
2862 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2863 		log.u_bbr.delRate = rack_get_bw(rack);
2864 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
2865 		log.u_bbr.cur_del_rate <<= 32;
2866 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
2867 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
2868 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
2869 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2870 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
2871 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
2872 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
2873 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
2874 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
2875 		log.u_bbr.rttProp = us_cts;
2876 		log.u_bbr.rttProp <<= 32;
2877 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
2878 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2879 		    &rack->rc_inp->inp_socket->so_rcv,
2880 		    &rack->rc_inp->inp_socket->so_snd,
2881 		    BBR_LOG_RTT_SHRINKS, 0,
2882 		    0, &log, false, &rack->r_ctl.act_rcv_time);
2883 	}
2884 }
2885 
2886 static void
2887 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
2888 {
2889 	uint64_t bwdp;
2890 
2891 	bwdp = rack_get_bw(rack);
2892 	bwdp *= (uint64_t)rtt;
2893 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
2894 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
2895 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
2896 		/*
2897 		 * A window protocol must be able to have 4 packets
2898 		 * outstanding as the floor in order to function
2899 		 * (especially considering delayed ack :D).
2900 		 */
2901 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
2902 	}
2903 }
2904 
2905 static void
2906 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
2907 {
2908 	/**
2909 	 * ProbeRTT is a bit different in rack_pacing than in
2910 	 * BBR. It is like BBR in that it uses the lowering of
2911 	 * the RTT as a signal that we saw something new and
2912 	 * counts from there for how long between. But it is
2913 	 * different in that its quite simple. It does not
2914 	 * play with the cwnd and wait until we get down
2915 	 * to N segments outstanding and hold that for
2916 	 * 200ms. Instead it just sets the pacing reduction
2917 	 * rate to a set percentage (70 by default) and hold
2918 	 * that for a number of recent GP Srtt's.
2919 	 */
2920 	uint32_t segsiz;
2921 
2922 	if (rack->rc_gp_dyn_mul == 0)
2923 		return;
2924 
2925 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
2926 		/* We are idle */
2927 		return;
2928 	}
2929 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
2930 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
2931 		/*
2932 		 * Stop the goodput now, the idea here is
2933 		 * that future measurements with in_probe_rtt
2934 		 * won't register if they are not greater so
2935 		 * we want to get what info (if any) is available
2936 		 * now.
2937 		 */
2938 		rack_do_goodput_measurement(rack->rc_tp, rack,
2939 					    rack->rc_tp->snd_una, __LINE__);
2940 	}
2941 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
2942 	rack->r_ctl.rc_time_probertt_entered = us_cts;
2943 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
2944 		     rack->r_ctl.rc_pace_min_segs);
2945 	rack->in_probe_rtt = 1;
2946 	rack->measure_saw_probe_rtt = 1;
2947 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
2948 	rack->r_ctl.rc_time_probertt_starts = 0;
2949 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
2950 	if (rack_probertt_use_min_rtt_entry)
2951 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
2952 	else
2953 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
2954 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
2955 			     __LINE__, RACK_RTTS_ENTERPROBE);
2956 }
2957 
2958 static void
2959 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
2960 {
2961 	struct rack_sendmap *rsm;
2962 	uint32_t segsiz;
2963 
2964 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
2965 		     rack->r_ctl.rc_pace_min_segs);
2966 	rack->in_probe_rtt = 0;
2967 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
2968 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
2969 		/*
2970 		 * Stop the goodput now, the idea here is
2971 		 * that future measurements with in_probe_rtt
2972 		 * won't register if they are not greater so
2973 		 * we want to get what info (if any) is available
2974 		 * now.
2975 		 */
2976 		rack_do_goodput_measurement(rack->rc_tp, rack,
2977 					    rack->rc_tp->snd_una, __LINE__);
2978 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
2979 		/*
2980 		 * We don't have enough data to make a measurement.
2981 		 * So lets just stop and start here after exiting
2982 		 * probe-rtt. We probably are not interested in
2983 		 * the results anyway.
2984 		 */
2985 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
2986 	}
2987 	/*
2988 	 * Measurements through the current snd_max are going
2989 	 * to be limited by the slower pacing rate.
2990 	 *
2991 	 * We need to mark these as app-limited so we
2992 	 * don't collapse the b/w.
2993 	 */
2994 	rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
2995 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
2996 		if (rack->r_ctl.rc_app_limited_cnt == 0)
2997 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
2998 		else {
2999 			/*
3000 			 * Go out to the end app limited and mark
3001 			 * this new one as next and move the end_appl up
3002 			 * to this guy.
3003 			 */
3004 			if (rack->r_ctl.rc_end_appl)
3005 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3006 			rack->r_ctl.rc_end_appl = rsm;
3007 		}
3008 		rsm->r_flags |= RACK_APP_LIMITED;
3009 		rack->r_ctl.rc_app_limited_cnt++;
3010 	}
3011 	/*
3012 	 * Now, we need to examine our pacing rate multipliers.
3013 	 * If its under 100%, we need to kick it back up to
3014 	 * 100%. We also don't let it be over our "max" above
3015 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3016 	 * Note setting clamp_atexit_prtt to 0 has the effect
3017 	 * of setting CA/SS to 100% always at exit (which is
3018 	 * the default behavior).
3019 	 */
3020 	if (rack_probertt_clear_is) {
3021 		rack->rc_gp_incr = 0;
3022 		rack->rc_gp_bwred = 0;
3023 		rack->rc_gp_timely_inc_cnt = 0;
3024 		rack->rc_gp_timely_dec_cnt = 0;
3025 	}
3026 	/* Do we do any clamping at exit? */
3027 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3028 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3029 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3030 	}
3031 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3032 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3033 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3034 	}
3035 	/*
3036 	 * Lets set rtt_diff to 0, so that we will get a "boost"
3037 	 * after exiting.
3038 	 */
3039 	rack->r_ctl.rc_rtt_diff = 0;
3040 
3041 	/* Clear all flags so we start fresh */
3042 	rack->rc_tp->t_bytes_acked = 0;
3043 	rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3044 	/*
3045 	 * If configured to, set the cwnd and ssthresh to
3046 	 * our targets.
3047 	 */
3048 	if (rack_probe_rtt_sets_cwnd) {
3049 		uint64_t ebdp;
3050 		uint32_t setto;
3051 
3052 		/* Set ssthresh so we get into CA once we hit our target */
3053 		if (rack_probertt_use_min_rtt_exit == 1) {
3054 			/* Set to min rtt */
3055 			rack_set_prtt_target(rack, segsiz,
3056 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3057 		} else if (rack_probertt_use_min_rtt_exit == 2) {
3058 			/* Set to current gp rtt */
3059 			rack_set_prtt_target(rack, segsiz,
3060 					     rack->r_ctl.rc_gp_srtt);
3061 		} else if (rack_probertt_use_min_rtt_exit == 3) {
3062 			/* Set to entry gp rtt */
3063 			rack_set_prtt_target(rack, segsiz,
3064 					     rack->r_ctl.rc_entry_gp_rtt);
3065 		} else  {
3066 			uint64_t sum;
3067 			uint32_t setval;
3068 
3069 			sum = rack->r_ctl.rc_entry_gp_rtt;
3070 			sum *= 10;
3071 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3072 			if (sum >= 20) {
3073 				/*
3074 				 * A highly buffered path needs
3075 				 * cwnd space for timely to work.
3076 				 * Lets set things up as if
3077 				 * we are heading back here again.
3078 				 */
3079 				setval = rack->r_ctl.rc_entry_gp_rtt;
3080 			} else if (sum >= 15) {
3081 				/*
3082 				 * Lets take the smaller of the
3083 				 * two since we are just somewhat
3084 				 * buffered.
3085 				 */
3086 				setval = rack->r_ctl.rc_gp_srtt;
3087 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
3088 					setval = rack->r_ctl.rc_entry_gp_rtt;
3089 			} else {
3090 				/*
3091 				 * Here we are not highly buffered
3092 				 * and should pick the min we can to
3093 				 * keep from causing loss.
3094 				 */
3095 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3096 			}
3097 			rack_set_prtt_target(rack, segsiz,
3098 					     setval);
3099 		}
3100 		if (rack_probe_rtt_sets_cwnd > 1) {
3101 			/* There is a percentage here to boost */
3102 			ebdp = rack->r_ctl.rc_target_probertt_flight;
3103 			ebdp *= rack_probe_rtt_sets_cwnd;
3104 			ebdp /= 100;
3105 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3106 		} else
3107 			setto = rack->r_ctl.rc_target_probertt_flight;
3108 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3109 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3110 			/* Enforce a min */
3111 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3112 		}
3113 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
3114 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3115 	}
3116 	rack_log_rtt_shrinks(rack,  us_cts,
3117 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3118 			     __LINE__, RACK_RTTS_EXITPROBE);
3119 	/* Clear times last so log has all the info */
3120 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3121 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3122 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3123 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
3124 }
3125 
3126 static void
3127 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3128 {
3129 	/* Check in on probe-rtt */
3130 	if (rack->rc_gp_filled == 0) {
3131 		/* We do not do p-rtt unless we have gp measurements */
3132 		return;
3133 	}
3134 	if (rack->in_probe_rtt) {
3135 		uint64_t no_overflow;
3136 		uint32_t endtime, must_stay;
3137 
3138 		if (rack->r_ctl.rc_went_idle_time &&
3139 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3140 			/*
3141 			 * We went idle during prtt, just exit now.
3142 			 */
3143 			rack_exit_probertt(rack, us_cts);
3144 		} else if (rack_probe_rtt_safety_val &&
3145 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3146 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3147 			/*
3148 			 * Probe RTT safety value triggered!
3149 			 */
3150 			rack_log_rtt_shrinks(rack,  us_cts,
3151 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3152 					     __LINE__, RACK_RTTS_SAFETY);
3153 			rack_exit_probertt(rack, us_cts);
3154 		}
3155 		/* Calculate the max we will wait */
3156 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3157 		if (rack->rc_highly_buffered)
3158 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3159 		/* Calculate the min we must wait */
3160 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3161 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3162 		    TSTMP_LT(us_cts, endtime)) {
3163 			uint32_t calc;
3164 			/* Do we lower more? */
3165 no_exit:
3166 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3167 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3168 			else
3169 				calc = 0;
3170 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3171 			if (calc) {
3172 				/* Maybe */
3173 				calc *= rack_per_of_gp_probertt_reduce;
3174 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3175 				/* Limit it too */
3176 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3177 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3178 			}
3179 			/* We must reach target or the time set */
3180 			return;
3181 		}
3182 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
3183 			if ((TSTMP_LT(us_cts, must_stay) &&
3184 			     rack->rc_highly_buffered) ||
3185 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3186 			      rack->r_ctl.rc_target_probertt_flight)) {
3187 				/* We are not past the must_stay time */
3188 				goto no_exit;
3189 			}
3190 			rack_log_rtt_shrinks(rack,  us_cts,
3191 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3192 					     __LINE__, RACK_RTTS_REACHTARGET);
3193 			rack->r_ctl.rc_time_probertt_starts = us_cts;
3194 			if (rack->r_ctl.rc_time_probertt_starts == 0)
3195 				rack->r_ctl.rc_time_probertt_starts = 1;
3196 			/* Restore back to our rate we want to pace at in prtt */
3197 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3198 		}
3199 		/*
3200 		 * Setup our end time, some number of gp_srtts plus 200ms.
3201 		 */
3202 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3203 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3204 		if (rack_probertt_gpsrtt_cnt_div)
3205 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3206 		else
3207 			endtime = 0;
3208 		endtime += rack_min_probertt_hold;
3209 		endtime += rack->r_ctl.rc_time_probertt_starts;
3210 		if (TSTMP_GEQ(us_cts,  endtime)) {
3211 			/* yes, exit probertt  */
3212 			rack_exit_probertt(rack, us_cts);
3213  		}
3214 
3215 	} else 	if((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3216 		/* Go into probertt, its been too long since we went lower  */
3217 		rack_enter_probertt(rack, us_cts);
3218 	}
3219 }
3220 
3221 static void
3222 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3223 		       uint32_t rtt, int32_t rtt_diff)
3224 {
3225 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
3226 	uint32_t losses;
3227 
3228 	if ((rack->rc_gp_dyn_mul == 0) ||
3229 	    (rack->use_fixed_rate) ||
3230 	    (rack->in_probe_rtt) ||
3231 	    (rack->rc_always_pace == 0)) {
3232 		/* No dynamic GP multipler in play */
3233 		return;
3234 	}
3235 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3236 	cur_bw = rack_get_bw(rack);
3237 	/* Calculate our up and down range */
3238 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3239 	up_bnd /= 100;
3240 	up_bnd += rack->r_ctl.last_gp_comp_bw;
3241 
3242 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3243 	subfr /= 100;
3244 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3245 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3246 		/*
3247 		 * This is the case where our RTT is above
3248 		 * the max target and we have been configured
3249 		 * to just do timely no bonus up stuff in that case.
3250 		 *
3251 		 * There are two configurations, set to 1, and we
3252 		 * just do timely if we are over our max. If its
3253 		 * set above 1 then we slam the multipliers down
3254 		 * to 100 and then decrement per timely.
3255 		 */
3256 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3257 				__LINE__, 3);
3258 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3259 			rack_validate_multipliers_at_or_below_100(rack);
3260 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3261 	} else if ((last_bw_est < low_bnd) && !losses) {
3262 		/*
3263 		 * We are decreasing this is a bit complicated this
3264 		 * means we are loosing ground. This could be
3265 		 * because another flow entered and we are competing
3266 		 * for b/w with it. This will push the RTT up which
3267 		 * makes timely unusable unless we want to get shoved
3268 		 * into a corner and just be backed off (the age
3269 		 * old problem with delay based CC).
3270 		 *
3271 		 * On the other hand if it was a route change we
3272 		 * would like to stay somewhat contained and not
3273 		 * blow out the buffers.
3274 		 */
3275 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3276 				__LINE__, 3);
3277 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3278 		if (rack->rc_gp_bwred == 0) {
3279 			/* Go into reduction counting */
3280 			rack->rc_gp_bwred = 1;
3281 			rack->rc_gp_timely_dec_cnt = 0;
3282 		}
3283 		if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3284 		    (timely_says == 0)) {
3285 			/*
3286 			 * Push another time with a faster pacing
3287 			 * to try to gain back (we include override to
3288 			 * get a full raise factor).
3289 			 */
3290 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3291 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3292 			    (timely_says == 0) ||
3293 			    (rack_down_raise_thresh == 0)) {
3294 				/*
3295 				 * Do an override up in b/w if we were
3296 				 * below the threshold or if the threshold
3297 				 * is zero we always do the raise.
3298 				 */
3299 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3300 			} else {
3301 				/* Log it stays the same */
3302 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3303 						__LINE__, 11);
3304 
3305 			}
3306 			rack->rc_gp_timely_dec_cnt++;
3307 			/* We are not incrementing really no-count */
3308 			rack->rc_gp_incr = 0;
3309 			rack->rc_gp_timely_inc_cnt = 0;
3310 		} else {
3311 			/*
3312 			 * Lets just use the RTT
3313 			 * information and give up
3314 			 * pushing.
3315 			 */
3316 			goto use_timely;
3317 		}
3318 	}  else if ((timely_says != 2) &&
3319 		    !losses &&
3320 		    (last_bw_est > up_bnd)) {
3321 		/*
3322 		 * We are increasing b/w lets keep going, updating
3323 		 * our b/w and ignoring any timely input, unless
3324 		 * of course we are at our max raise (if there is one).
3325 		 */
3326 
3327 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3328 				__LINE__, 3);
3329 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3330 		if (rack->rc_gp_saw_ss &&
3331 		    rack_per_upper_bound_ss &&
3332 		     (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3333 			    /*
3334 			     * In cases where we can't go higher
3335 			     * we should just use timely.
3336 			     */
3337 			    goto use_timely;
3338 		}
3339 		if (rack->rc_gp_saw_ca &&
3340 		    rack_per_upper_bound_ca &&
3341 		    (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
3342 			    /*
3343 			     * In cases where we can't go higher
3344 			     * we should just use timely.
3345 			     */
3346 			    goto use_timely;
3347 		}
3348 		rack->rc_gp_bwred = 0;
3349 		rack->rc_gp_timely_dec_cnt = 0;
3350 		/* You get a set number of pushes if timely is trying to reduce  */
3351 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
3352 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3353 		} else {
3354  			/* Log it stays the same */
3355 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
3356 			    __LINE__, 12);
3357 
3358 		}
3359 		return;
3360 	} else {
3361 		/*
3362 		 * We are staying between the lower and upper range bounds
3363 		 * so use timely to decide.
3364 		 */
3365 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3366 				__LINE__, 3);
3367 use_timely:
3368 		if (timely_says) {
3369 			rack->rc_gp_incr = 0;
3370 			rack->rc_gp_timely_inc_cnt = 0;
3371 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
3372 			    !losses &&
3373 			    (last_bw_est < low_bnd)) {
3374 				/* We are loosing ground */
3375 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3376 				rack->rc_gp_timely_dec_cnt++;
3377 				/* We are not incrementing really no-count */
3378 				rack->rc_gp_incr = 0;
3379 				rack->rc_gp_timely_inc_cnt = 0;
3380 			} else
3381 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3382 		} else  {
3383 			rack->rc_gp_bwred = 0;
3384 			rack->rc_gp_timely_dec_cnt = 0;
3385 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3386 		}
3387 	}
3388 }
3389 
3390 static int32_t
3391 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
3392 {
3393 	int32_t timely_says;
3394 	uint64_t log_mult, log_rtt_a_diff;
3395 
3396 	log_rtt_a_diff = rtt;
3397 	log_rtt_a_diff <<= 32;
3398 	log_rtt_a_diff |= (uint32_t)rtt_diff;
3399 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
3400 		    rack_gp_rtt_maxmul)) {
3401 		/* Reduce the b/w multipler */
3402 		timely_says = 2;
3403 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3404 		log_mult <<= 32;
3405 		log_mult |= prev_rtt;
3406 		rack_log_timely(rack,  timely_says, log_mult,
3407 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3408 				log_rtt_a_diff, __LINE__, 4);
3409 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3410 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
3411 			    max(rack_gp_rtt_mindiv , 1)))) {
3412 		/* Increase the b/w multipler */
3413 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
3414 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
3415 			 max(rack_gp_rtt_mindiv , 1));
3416 		log_mult <<= 32;
3417 		log_mult |= prev_rtt;
3418 		timely_says = 0;
3419 		rack_log_timely(rack,  timely_says, log_mult ,
3420 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3421 				log_rtt_a_diff, __LINE__, 5);
3422 	} else {
3423 		/*
3424 		 * Use a gradient to find it the timely gradient
3425 		 * is:
3426 		 * grad = rc_rtt_diff / min_rtt;
3427 		 *
3428 		 * anything below or equal to 0 will be
3429 		 * a increase indication. Anything above
3430 		 * zero is a decrease. Note we take care
3431 		 * of the actual gradient calculation
3432 		 * in the reduction (its not needed for
3433 		 * increase).
3434 		 */
3435 		log_mult = prev_rtt;
3436 		if (rtt_diff <= 0) {
3437 			/*
3438 			 * Rttdiff is less than zero, increase the
3439 			 * b/w multipler (its 0 or negative)
3440 			 */
3441 			timely_says = 0;
3442 			rack_log_timely(rack,  timely_says, log_mult,
3443 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
3444 		} else {
3445 			/* Reduce the b/w multipler */
3446 			timely_says = 1;
3447 			rack_log_timely(rack,  timely_says, log_mult,
3448 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
3449 		}
3450 	}
3451 	return (timely_says);
3452 }
3453 
3454 static void
3455 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
3456 			    tcp_seq th_ack, int line)
3457 {
3458 	uint64_t tim, bytes_ps, ltim, stim, utim;
3459 	uint32_t segsiz, bytes, reqbytes, us_cts;
3460 	int32_t gput, new_rtt_diff, timely_says;
3461 
3462 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
3463 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3464 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
3465 		tim = us_cts - tp->gput_ts;
3466 	else
3467 		tim = 0;
3468 
3469 	if (TSTMP_GT(rack->r_ctl.rc_gp_cumack_ts, rack->r_ctl.rc_gp_output_ts))
3470 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
3471 	else
3472 		stim = 0;
3473 	/*
3474 	 * Use the larger of the send time or ack time. This prevents us
3475 	 * from being influenced by ack artifacts to come up with too
3476 	 * high of measurement. Note that since we are spanning over many more
3477 	 * bytes in most of our measurements hopefully that is less likely to
3478 	 * occur.
3479 	 */
3480 	if (tim > stim)
3481 		utim = max(tim, 1);
3482 	else
3483 		utim = max(stim, 1);
3484 	/* Lets validate utim */
3485 	ltim = max(1, (utim/HPTS_USEC_IN_MSEC));
3486 	gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
3487 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
3488 	if ((tim == 0) && (stim == 0)) {
3489 		/*
3490 		 * Invalid measurement time, maybe
3491 		 * all on one ack/one send?
3492 		 */
3493 		bytes = 0;
3494 		bytes_ps = 0;
3495 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3496 					   0, 0, 0, 10, __LINE__, NULL);
3497 		goto skip_measurement;
3498 	}
3499 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
3500 		/* We never made a us_rtt measurement? */
3501 		bytes = 0;
3502 		bytes_ps = 0;
3503 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3504 					   0, 0, 0, 10, __LINE__, NULL);
3505 		goto skip_measurement;
3506 	}
3507 	/*
3508 	 * Calculate the maximum possible b/w this connection
3509 	 * could have. We base our calculation on the lowest
3510 	 * rtt we have seen during the measurement and the
3511 	 * largest rwnd the client has given us in that time. This
3512 	 * forms a BDP that is the maximum that we could ever
3513 	 * get to the client. Anything larger is not valid.
3514 	 *
3515 	 * I originally had code here that rejected measurements
3516 	 * where the time was less than 1/2 the latest us_rtt.
3517 	 * But after thinking on that I realized its wrong since
3518 	 * say you had a 150Mbps or even 1Gbps link, and you
3519 	 * were a long way away.. example I am in Europe (100ms rtt)
3520 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
3521 	 * bytes my time would be 1.2ms, and yet my rtt would say
3522 	 * the measurement was invalid the time was < 50ms. The
3523 	 * same thing is true for 150Mb (8ms of time).
3524 	 *
3525 	 * A better way I realized is to look at what the maximum
3526 	 * the connection could possibly do. This is gated on
3527 	 * the lowest RTT we have seen and the highest rwnd.
3528 	 * We should in theory never exceed that, if we are
3529 	 * then something on the path is storing up packets
3530 	 * and then feeding them all at once to our endpoint
3531 	 * messing up our measurement.
3532 	 */
3533 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
3534 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
3535 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
3536 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3537 		/* No measurement can be made */
3538 		bytes = 0;
3539 		bytes_ps = 0;
3540 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3541 					   0, 0, 0, 10, __LINE__, NULL);
3542 		goto skip_measurement;
3543 	} else
3544 		bytes = (th_ack - tp->gput_seq);
3545 	bytes_ps = (uint64_t)bytes;
3546 	/*
3547 	 * Don't measure a b/w for pacing unless we have gotten at least
3548 	 * an initial windows worth of data in this measurement interval.
3549 	 *
3550 	 * Small numbers of bytes get badly influenced by delayed ack and
3551 	 * other artifacts. Note we take the initial window or our
3552 	 * defined minimum GP (defaulting to 10 which hopefully is the
3553 	 * IW).
3554 	 */
3555 	if (rack->rc_gp_filled == 0) {
3556 		/*
3557 		 * The initial estimate is special. We
3558 		 * have blasted out an IW worth of packets
3559 		 * without a real valid ack ts results. We
3560 		 * then setup the app_limited_needs_set flag,
3561 		 * this should get the first ack in (probably 2
3562 		 * MSS worth) to be recorded as the timestamp.
3563 		 * We thus allow a smaller number of bytes i.e.
3564 		 * IW - 2MSS.
3565 		 */
3566 		reqbytes -= (2 * segsiz);
3567 		/* Also lets fill previous for our first measurement to be neutral */
3568 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
3569 	}
3570 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
3571 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3572 					   rack->r_ctl.rc_app_limited_cnt,
3573 					   0, 0, 10, __LINE__, NULL);
3574 		goto skip_measurement;
3575 	}
3576 	/*
3577 	 * We now need to calculate the Timely like status so
3578 	 * we can update (possibly) the b/w multipliers.
3579 	 */
3580 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
3581 	if (rack->rc_gp_filled == 0) {
3582 		/* No previous reading */
3583 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
3584 	} else {
3585 		if (rack->measure_saw_probe_rtt == 0) {
3586 			/*
3587 			 * We don't want a probertt to be counted
3588 			 * since it will be negative incorrectly. We
3589 			 * expect to be reducing the RTT when we
3590 			 * pace at a slower rate.
3591 			 */
3592 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
3593 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
3594 		}
3595 	}
3596 	timely_says = rack_make_timely_judgement(rack,
3597 		rack->r_ctl.rc_gp_srtt,
3598 		rack->r_ctl.rc_rtt_diff,
3599 	        rack->r_ctl.rc_prev_gp_srtt
3600 		);
3601 	bytes_ps *= HPTS_USEC_IN_SEC;
3602 	bytes_ps /= utim;
3603 	if (bytes_ps > rack->r_ctl.last_max_bw) {
3604 		/*
3605 		 * Something is on path playing
3606 		 * since this b/w is not possible based
3607 		 * on our BDP (highest rwnd and lowest rtt
3608 		 * we saw in the measurement window).
3609 		 *
3610 		 * Another option here would be to
3611 		 * instead skip the measurement.
3612 		 */
3613 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
3614 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
3615 					   11, __LINE__, NULL);
3616 		bytes_ps = rack->r_ctl.last_max_bw;
3617 	}
3618 	/* We store gp for b/w in bytes per second  */
3619 	if (rack->rc_gp_filled == 0) {
3620 		/* Initial measurment */
3621 		if (bytes_ps) {
3622 			rack->r_ctl.gp_bw = bytes_ps;
3623 			rack->rc_gp_filled = 1;
3624 			rack->r_ctl.num_avg = 1;
3625 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
3626 		} else {
3627 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
3628 						   rack->r_ctl.rc_app_limited_cnt,
3629 						   0, 0, 10, __LINE__, NULL);
3630 		}
3631 		if (rack->rc_inp->inp_in_hpts &&
3632 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
3633 			/*
3634 			 * Ok we can't trust the pacer in this case
3635 			 * where we transition from un-paced to paced.
3636 			 * Or for that matter when the burst mitigation
3637 			 * was making a wild guess and got it wrong.
3638 			 * Stop the pacer and clear up all the aggregate
3639 			 * delays etc.
3640 			 */
3641 			tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
3642 			rack->r_ctl.rc_hpts_flags = 0;
3643 			rack->r_ctl.rc_last_output_to = 0;
3644 		}
3645 	} else if (rack->r_ctl.num_avg < RACK_REQ_AVG) {
3646 		/* Still a small number run an average */
3647 		rack->r_ctl.gp_bw += bytes_ps;
3648 		rack->r_ctl.num_avg++;
3649 		if (rack->r_ctl.num_avg >= RACK_REQ_AVG) {
3650 			/* We have collected enought to move forward */
3651 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_avg;
3652 		}
3653 	} else {
3654 		/*
3655 		 * We want to take 1/wma of the goodput and add in to 7/8th
3656 		 * of the old value weighted by the srtt. So if your measurement
3657 		 * period is say 2 SRTT's long you would get 1/4 as the
3658 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
3659 		 *
3660 		 * But we must be careful not to take too much i.e. if the
3661 		 * srtt is say 20ms and the measurement is taken over
3662 		 * 400ms our weight would be 400/20 i.e. 20. On the
3663 		 * other hand if we get a measurement over 1ms with a
3664 		 * 10ms rtt we only want to take a much smaller portion.
3665 		 */
3666 		uint64_t  resid_bw, subpart, addpart, srtt;
3667 
3668 		srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
3669 		if (srtt == 0) {
3670 			/*
3671 			 * Strange why did t_srtt go back to zero?
3672 			 */
3673 			if (rack->r_ctl.rc_rack_min_rtt)
3674 				srtt = (rack->r_ctl.rc_rack_min_rtt * HPTS_USEC_IN_MSEC);
3675 			else
3676 				srtt = HPTS_USEC_IN_MSEC;
3677 		}
3678 		/*
3679 		 * XXXrrs: Note for reviewers, in playing with
3680 		 * dynamic pacing I discovered this GP calculation
3681 		 * as done originally leads to some undesired results.
3682 		 * Basically you can get longer measurements contributing
3683 		 * too much to the WMA. Thus I changed it if you are doing
3684 		 * dynamic adjustments to only do the aportioned adjustment
3685 		 * if we have a very small (time wise) measurement. Longer
3686 		 * measurements just get there weight (defaulting to 1/8)
3687 		 * add to the WMA. We may want to think about changing
3688 		 * this to always do that for both sides i.e. dynamic
3689 		 * and non-dynamic... but considering lots of folks
3690 		 * were playing with this I did not want to change the
3691 		 * calculation per.se. without your thoughts.. Lawerence?
3692 		 * Peter??
3693 		 */
3694 		if (rack->rc_gp_dyn_mul == 0) {
3695 			subpart = rack->r_ctl.gp_bw * utim;
3696 			subpart /= (srtt * 8);
3697 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
3698 				/*
3699 				 * The b/w update takes no more
3700 				 * away then 1/2 our running total
3701 				 * so factor it in.
3702 				 */
3703 				addpart = bytes_ps * utim;
3704 				addpart /= (srtt * 8);
3705 			} else {
3706 				/*
3707 				 * Don't allow a single measurement
3708 				 * to account for more than 1/2 of the
3709 				 * WMA. This could happen on a retransmission
3710 				 * where utim becomes huge compared to
3711 				 * srtt (multiple retransmissions when using
3712 				 * the sending rate which factors in all the
3713 				 * transmissions from the first one).
3714 				 */
3715 				subpart = rack->r_ctl.gp_bw / 2;
3716 				addpart = bytes_ps / 2;
3717 			}
3718 			resid_bw = rack->r_ctl.gp_bw - subpart;
3719 			rack->r_ctl.gp_bw = resid_bw + addpart;
3720 		} else {
3721 			if ((utim / srtt) <= 1) {
3722 				/*
3723 				 * The b/w update was over a small period
3724 				 * of time. The idea here is to prevent a small
3725 				 * measurement time period from counting
3726 				 * too much. So we scale it based on the
3727 				 * time so it attributes less than 1/rack_wma_divisor
3728 				 * of its measurement.
3729 				 */
3730 				subpart = rack->r_ctl.gp_bw * utim;
3731 				subpart /= (srtt * rack_wma_divisor);
3732 				addpart = bytes_ps * utim;
3733 				addpart /= (srtt * rack_wma_divisor);
3734 			} else {
3735 				/*
3736 				 * The scaled measurement was long
3737 				 * enough so lets just add in the
3738 				 * portion of the measurment i.e. 1/rack_wma_divisor
3739 				 */
3740 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
3741 				addpart = bytes_ps / rack_wma_divisor;
3742 			}
3743 			if ((rack->measure_saw_probe_rtt == 0) ||
3744 		            (bytes_ps > rack->r_ctl.gp_bw)) {
3745 				/*
3746 				 * For probe-rtt we only add it in
3747 				 * if its larger, all others we just
3748 				 * add in.
3749 				 */
3750 				resid_bw = rack->r_ctl.gp_bw - subpart;
3751 				rack->r_ctl.gp_bw = resid_bw + addpart;
3752 			}
3753 		}
3754 	}
3755 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
3756 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
3757 		rack_update_multiplier(rack, timely_says, bytes_ps,
3758 				       rack->r_ctl.rc_gp_srtt,
3759 				       rack->r_ctl.rc_rtt_diff);
3760 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
3761 				   rack_get_bw(rack), 3, line, NULL);
3762 	/* reset the gp srtt and setup the new prev */
3763 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
3764 	/* Record the lost count for the next measurement */
3765 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
3766 	/*
3767 	 * We restart our diffs based on the gpsrtt in the
3768 	 * measurement window.
3769 	 */
3770 	rack->rc_gp_rtt_set = 0;
3771 	rack->rc_gp_saw_rec = 0;
3772 	rack->rc_gp_saw_ca = 0;
3773 	rack->rc_gp_saw_ss = 0;
3774 	rack->rc_dragged_bottom = 0;
3775 skip_measurement:
3776 
3777 #ifdef STATS
3778 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3779 				 gput);
3780 	/*
3781 	 * XXXLAS: This is a temporary hack, and should be
3782 	 * chained off VOI_TCP_GPUT when stats(9) grows an
3783 	 * API to deal with chained VOIs.
3784 	 */
3785 	if (tp->t_stats_gput_prev > 0)
3786 		stats_voi_update_abs_s32(tp->t_stats,
3787 					 VOI_TCP_GPUT_ND,
3788 					 ((gput - tp->t_stats_gput_prev) * 100) /
3789 					 tp->t_stats_gput_prev);
3790 #endif
3791 	tp->t_flags &= ~TF_GPUTINPROG;
3792 	tp->t_stats_gput_prev = gput;
3793 	/*
3794 	 * Now are we app limited now and there is space from where we
3795 	 * were to where we want to go?
3796 	 *
3797 	 * We don't do the other case i.e. non-applimited here since
3798 	 * the next send will trigger us picking up the missing data.
3799 	 */
3800 	if (rack->r_ctl.rc_first_appl &&
3801 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
3802 	    rack->r_ctl.rc_app_limited_cnt &&
3803 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
3804 	    ((rack->r_ctl.rc_first_appl->r_start - th_ack) >
3805 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3806 		/*
3807 		 * Yep there is enough outstanding to make a measurement here.
3808 		 */
3809 		struct rack_sendmap *rsm, fe;
3810 
3811 		tp->t_flags |= TF_GPUTINPROG;
3812 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
3813 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
3814 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
3815 		rack->app_limited_needs_set = 0;
3816 		tp->gput_seq = th_ack;
3817 		if (rack->in_probe_rtt)
3818 			rack->measure_saw_probe_rtt = 1;
3819 		else if ((rack->measure_saw_probe_rtt) &&
3820 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
3821 			rack->measure_saw_probe_rtt = 0;
3822 		if ((rack->r_ctl.rc_first_appl->r_start - th_ack) >= rack_get_measure_window(tp, rack)) {
3823 			/* There is a full window to gain info from */
3824 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
3825 		} else {
3826 			/* We can only measure up to the applimited point */
3827 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_start - th_ack);
3828 		}
3829 		/*
3830 		 * Now we need to find the timestamp of the send at tp->gput_seq
3831 		 * for the send based measurement.
3832 		 */
3833 		fe.r_start = tp->gput_seq;
3834 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3835 		if (rsm) {
3836 			/* Ok send-based limit is set */
3837 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
3838 				/*
3839 				 * Move back to include the earlier part
3840 				 * so our ack time lines up right (this may
3841 				 * make an overlapping measurement but thats
3842 				 * ok).
3843 				 */
3844 				tp->gput_seq = rsm->r_start;
3845 			}
3846 			if (rsm->r_flags & RACK_ACKED)
3847 				tp->gput_ts = rsm->r_ack_arrival;
3848 			else
3849 				rack->app_limited_needs_set = 1;
3850 			rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send;
3851 		} else {
3852 			/*
3853 			 * If we don't find the rsm due to some
3854 			 * send-limit set the current time, which
3855 			 * basically disables the send-limit.
3856 			 */
3857 			rack->r_ctl.rc_gp_output_ts = tcp_get_usecs(NULL);
3858 		}
3859 		rack_log_pacing_delay_calc(rack,
3860 					   tp->gput_seq,
3861 					   tp->gput_ack,
3862 					   (uint64_t)rsm,
3863 					   tp->gput_ts,
3864 					   rack->r_ctl.rc_app_limited_cnt,
3865 					   9,
3866 					   __LINE__, NULL);
3867 	}
3868 }
3869 
3870 /*
3871  * CC wrapper hook functions
3872  */
3873 static void
3874 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs,
3875     uint16_t type, int32_t recovery)
3876 {
3877 	INP_WLOCK_ASSERT(tp->t_inpcb);
3878 	tp->ccv->nsegs = nsegs;
3879 	tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th);
3880 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
3881 		uint32_t max;
3882 
3883 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
3884 		if (tp->ccv->bytes_this_ack > max) {
3885 			tp->ccv->bytes_this_ack = max;
3886 		}
3887 	}
3888 	if (rack->r_ctl.cwnd_to_use <= tp->snd_wnd)
3889 		tp->ccv->flags |= CCF_CWND_LIMITED;
3890 	else
3891 		tp->ccv->flags &= ~CCF_CWND_LIMITED;
3892 #ifdef STATS
3893 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
3894 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
3895 #endif
3896 	if ((tp->t_flags & TF_GPUTINPROG) &&
3897 	    rack_enough_for_measurement(tp, rack, th->th_ack)) {
3898 		/* Measure the Goodput */
3899 		rack_do_goodput_measurement(tp, rack, th->th_ack, __LINE__);
3900 #ifdef NETFLIX_PEAKRATE
3901 		if ((type == CC_ACK) &&
3902 		    (tp->t_maxpeakrate)) {
3903 			/*
3904 			 * We update t_peakrate_thr. This gives us roughly
3905 			 * one update per round trip time. Note
3906 			 * it will only be used if pace_always is off i.e
3907 			 * we don't do this for paced flows.
3908 			 */
3909 			tcp_update_peakrate_thr(tp);
3910 		}
3911 #endif
3912 	}
3913 	if (rack->r_ctl.cwnd_to_use > tp->snd_ssthresh) {
3914 		tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
3915 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
3916 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
3917 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
3918 			tp->ccv->flags |= CCF_ABC_SENTAWND;
3919 		}
3920 	} else {
3921 		tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3922 		tp->t_bytes_acked = 0;
3923 	}
3924 	if (CC_ALGO(tp)->ack_received != NULL) {
3925 		/* XXXLAS: Find a way to live without this */
3926 		tp->ccv->curack = th->th_ack;
3927 		CC_ALGO(tp)->ack_received(tp->ccv, type);
3928 	}
3929 #ifdef STATS
3930 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
3931 #endif
3932 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
3933 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
3934 	}
3935 #ifdef NETFLIX_PEAKRATE
3936 	/* we enforce max peak rate if it is set and we are not pacing */
3937 	if ((rack->rc_always_pace == 0) &&
3938 	    tp->t_peakrate_thr &&
3939 	    (tp->snd_cwnd > tp->t_peakrate_thr)) {
3940 		tp->snd_cwnd = tp->t_peakrate_thr;
3941 	}
3942 #endif
3943 }
3944 
3945 static void
3946 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th)
3947 {
3948 	struct tcp_rack *rack;
3949 
3950 	rack = (struct tcp_rack *)tp->t_fb_ptr;
3951 	INP_WLOCK_ASSERT(tp->t_inpcb);
3952 	/*
3953 	 * If we are doing PRR and have enough
3954 	 * room to send <or> we are pacing and prr
3955 	 * is disabled we will want to see if we
3956 	 * can send data (by setting r_wanted_output to
3957 	 * true).
3958 	 */
3959 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
3960 	    rack->rack_no_prr)
3961 		rack->r_wanted_output = 1;
3962 }
3963 
3964 static void
3965 rack_post_recovery(struct tcpcb *tp, struct tcphdr *th)
3966 {
3967 	struct tcp_rack *rack;
3968 	uint32_t orig_cwnd;
3969 
3970 
3971 	orig_cwnd = tp->snd_cwnd;
3972 	INP_WLOCK_ASSERT(tp->t_inpcb);
3973 	rack = (struct tcp_rack *)tp->t_fb_ptr;
3974 	if (rack->rc_not_backing_off == 0) {
3975 		/* only alert CC if we alerted when we entered */
3976 		if (CC_ALGO(tp)->post_recovery != NULL) {
3977 			tp->ccv->curack = th->th_ack;
3978 			CC_ALGO(tp)->post_recovery(tp->ccv);
3979 		}
3980 		if (tp->snd_cwnd > tp->snd_ssthresh) {
3981 			/* Drop us down to the ssthresh (1/2 cwnd at loss) */
3982 			tp->snd_cwnd = tp->snd_ssthresh;
3983 		}
3984 	}
3985 	if ((rack->rack_no_prr == 0) &&
3986 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
3987 		/* Suck the next prr cnt back into cwnd */
3988 		tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt;
3989 		rack->r_ctl.rc_prr_sndcnt = 0;
3990 		rack_log_to_prr(rack, 1, 0);
3991 	}
3992 	rack_log_to_prr(rack, 14, orig_cwnd);
3993 	tp->snd_recover = tp->snd_una;
3994 	EXIT_RECOVERY(tp->t_flags);
3995 }
3996 
3997 static void
3998 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type)
3999 {
4000 	struct tcp_rack *rack;
4001 
4002 	INP_WLOCK_ASSERT(tp->t_inpcb);
4003 
4004 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4005 	switch (type) {
4006 	case CC_NDUPACK:
4007 		tp->t_flags &= ~TF_WASFRECOVERY;
4008 		tp->t_flags &= ~TF_WASCRECOVERY;
4009 		if (!IN_FASTRECOVERY(tp->t_flags)) {
4010 			rack->r_ctl.rc_prr_delivered = 0;
4011 			rack->r_ctl.rc_prr_out = 0;
4012 			if (rack->rack_no_prr == 0) {
4013 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4014 				rack_log_to_prr(rack, 2, 0);
4015 			}
4016 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4017 			tp->snd_recover = tp->snd_max;
4018 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4019 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4020 		}
4021 		break;
4022 	case CC_ECN:
4023 		if (!IN_CONGRECOVERY(tp->t_flags) ||
4024 		    /*
4025 		     * Allow ECN reaction on ACK to CWR, if
4026 		     * that data segment was also CE marked.
4027 		     */
4028 		    SEQ_GEQ(th->th_ack, tp->snd_recover)) {
4029 			EXIT_CONGRECOVERY(tp->t_flags);
4030 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4031 			tp->snd_recover = tp->snd_max + 1;
4032 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4033 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4034 		}
4035 		break;
4036 	case CC_RTO:
4037 		tp->t_dupacks = 0;
4038 		tp->t_bytes_acked = 0;
4039 		EXIT_RECOVERY(tp->t_flags);
4040 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4041 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4042 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
4043 		if (tp->t_flags2 & TF2_ECN_PERMIT)
4044 			tp->t_flags2 |= TF2_ECN_SND_CWR;
4045 		break;
4046 	case CC_RTO_ERR:
4047 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4048 		/* RTO was unnecessary, so reset everything. */
4049 		tp->snd_cwnd = tp->snd_cwnd_prev;
4050 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
4051 		tp->snd_recover = tp->snd_recover_prev;
4052 		if (tp->t_flags & TF_WASFRECOVERY) {
4053 			ENTER_FASTRECOVERY(tp->t_flags);
4054 			tp->t_flags &= ~TF_WASFRECOVERY;
4055 		}
4056 		if (tp->t_flags & TF_WASCRECOVERY) {
4057 			ENTER_CONGRECOVERY(tp->t_flags);
4058 			tp->t_flags &= ~TF_WASCRECOVERY;
4059 		}
4060 		tp->snd_nxt = tp->snd_max;
4061 		tp->t_badrxtwin = 0;
4062 		break;
4063 	}
4064 	/*
4065 	 * If we are below our max rtt, don't
4066 	 * signal the CC control to change things.
4067 	 * instead set it up so that we are in
4068 	 * recovery but not going to back off.
4069 	 */
4070 
4071 	if (rack->rc_highly_buffered) {
4072 		/*
4073 		 * Do we use the higher rtt for
4074 		 * our threshold to not backoff (like CDG)?
4075 		 */
4076 		uint32_t rtt_mul, rtt_div;
4077 
4078 		if (rack_use_max_for_nobackoff) {
4079 			rtt_mul = (rack_gp_rtt_maxmul - 1);
4080 			rtt_div = 1;
4081 		} else {
4082 			rtt_mul = rack_gp_rtt_minmul;
4083 			rtt_div = max(rack_gp_rtt_mindiv , 1);
4084 		}
4085 		if (rack->r_ctl.rc_gp_srtt <= (rack->r_ctl.rc_lowest_us_rtt +
4086 					       ((rack->r_ctl.rc_lowest_us_rtt * rtt_mul) /
4087 						rtt_div))) {
4088 			/* below our min threshold */
4089 			rack->rc_not_backing_off = 1;
4090 			ENTER_RECOVERY(rack->rc_tp->t_flags);
4091 			rack_log_rtt_shrinks(rack, 0,
4092 					     rtt_mul,
4093 					     rtt_div,
4094 					     RACK_RTTS_NOBACKOFF);
4095 			return;
4096 		}
4097 	}
4098 	rack->rc_not_backing_off = 0;
4099 	if (CC_ALGO(tp)->cong_signal != NULL) {
4100 		if (th != NULL)
4101 			tp->ccv->curack = th->th_ack;
4102 		CC_ALGO(tp)->cong_signal(tp->ccv, type);
4103 	}
4104 }
4105 
4106 
4107 
4108 static inline void
4109 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4110 {
4111 	uint32_t i_cwnd;
4112 
4113 	INP_WLOCK_ASSERT(tp->t_inpcb);
4114 
4115 #ifdef NETFLIX_STATS
4116 	KMOD_TCPSTAT_INC(tcps_idle_restarts);
4117 	if (tp->t_state == TCPS_ESTABLISHED)
4118 		KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4119 #endif
4120 	if (CC_ALGO(tp)->after_idle != NULL)
4121 		CC_ALGO(tp)->after_idle(tp->ccv);
4122 
4123 	if (tp->snd_cwnd == 1)
4124 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
4125 	else
4126 		i_cwnd = rc_init_window(rack);
4127 
4128 	/*
4129 	 * Being idle is no differnt than the initial window. If the cc
4130 	 * clamps it down below the initial window raise it to the initial
4131 	 * window.
4132 	 */
4133 	if (tp->snd_cwnd < i_cwnd) {
4134 		tp->snd_cwnd = i_cwnd;
4135 	}
4136 }
4137 
4138 
4139 /*
4140  * Indicate whether this ack should be delayed.  We can delay the ack if
4141  * following conditions are met:
4142  *	- There is no delayed ack timer in progress.
4143  *	- Our last ack wasn't a 0-sized window. We never want to delay
4144  *	  the ack that opens up a 0-sized window.
4145  *	- LRO wasn't used for this segment. We make sure by checking that the
4146  *	  segment size is not larger than the MSS.
4147  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
4148  *	  connection.
4149  */
4150 #define DELAY_ACK(tp, tlen)			 \
4151 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4152 	((tp->t_flags & TF_DELACK) == 0) && 	 \
4153 	(tlen <= tp->t_maxseg) &&		 \
4154 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4155 
4156 static struct rack_sendmap *
4157 rack_find_lowest_rsm(struct tcp_rack *rack)
4158 {
4159 	struct rack_sendmap *rsm;
4160 
4161 	/*
4162 	 * Walk the time-order transmitted list looking for an rsm that is
4163 	 * not acked. This will be the one that was sent the longest time
4164 	 * ago that is still outstanding.
4165 	 */
4166 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4167 		if (rsm->r_flags & RACK_ACKED) {
4168 			continue;
4169 		}
4170 		goto finish;
4171 	}
4172 finish:
4173 	return (rsm);
4174 }
4175 
4176 static struct rack_sendmap *
4177 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4178 {
4179 	struct rack_sendmap *prsm;
4180 
4181 	/*
4182 	 * Walk the sequence order list backward until we hit and arrive at
4183 	 * the highest seq not acked. In theory when this is called it
4184 	 * should be the last segment (which it was not).
4185 	 */
4186 	counter_u64_add(rack_find_high, 1);
4187 	prsm = rsm;
4188 	RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4189 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4190 			continue;
4191 		}
4192 		return (prsm);
4193 	}
4194 	return (NULL);
4195 }
4196 
4197 
4198 static uint32_t
4199 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4200 {
4201 	int32_t lro;
4202 	uint32_t thresh;
4203 
4204 	/*
4205 	 * lro is the flag we use to determine if we have seen reordering.
4206 	 * If it gets set we have seen reordering. The reorder logic either
4207 	 * works in one of two ways:
4208 	 *
4209 	 * If reorder-fade is configured, then we track the last time we saw
4210 	 * re-ordering occur. If we reach the point where enough time as
4211 	 * passed we no longer consider reordering has occuring.
4212 	 *
4213 	 * Or if reorder-face is 0, then once we see reordering we consider
4214 	 * the connection to alway be subject to reordering and just set lro
4215 	 * to 1.
4216 	 *
4217 	 * In the end if lro is non-zero we add the extra time for
4218 	 * reordering in.
4219 	 */
4220 	if (srtt == 0)
4221 		srtt = 1;
4222 	if (rack->r_ctl.rc_reorder_ts) {
4223 		if (rack->r_ctl.rc_reorder_fade) {
4224 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
4225 				lro = cts - rack->r_ctl.rc_reorder_ts;
4226 				if (lro == 0) {
4227 					/*
4228 					 * No time as passed since the last
4229 					 * reorder, mark it as reordering.
4230 					 */
4231 					lro = 1;
4232 				}
4233 			} else {
4234 				/* Negative time? */
4235 				lro = 0;
4236 			}
4237 			if (lro > rack->r_ctl.rc_reorder_fade) {
4238 				/* Turn off reordering seen too */
4239 				rack->r_ctl.rc_reorder_ts = 0;
4240 				lro = 0;
4241 			}
4242 		} else {
4243 			/* Reodering does not fade */
4244 			lro = 1;
4245 		}
4246 	} else {
4247 		lro = 0;
4248 	}
4249 	thresh = srtt + rack->r_ctl.rc_pkt_delay;
4250 	if (lro) {
4251 		/* It must be set, if not you get 1/4 rtt */
4252 		if (rack->r_ctl.rc_reorder_shift)
4253 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
4254 		else
4255 			thresh += (srtt >> 2);
4256 	} else {
4257 		thresh += 1;
4258 	}
4259 	/* We don't let the rack timeout be above a RTO */
4260 	if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) {
4261 		thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur);
4262 	}
4263 	/* And we don't want it above the RTO max either */
4264 	if (thresh > rack_rto_max) {
4265 		thresh = rack_rto_max;
4266 	}
4267 	return (thresh);
4268 }
4269 
4270 static uint32_t
4271 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
4272 		     struct rack_sendmap *rsm, uint32_t srtt)
4273 {
4274 	struct rack_sendmap *prsm;
4275 	uint32_t thresh, len;
4276 	int segsiz;
4277 
4278 	if (srtt == 0)
4279 		srtt = 1;
4280 	if (rack->r_ctl.rc_tlp_threshold)
4281 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
4282 	else
4283 		thresh = (srtt * 2);
4284 
4285 	/* Get the previous sent packet, if any  */
4286 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4287 	counter_u64_add(rack_enter_tlp_calc, 1);
4288 	len = rsm->r_end - rsm->r_start;
4289 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
4290 		/* Exactly like the ID */
4291 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
4292 			uint32_t alt_thresh;
4293 			/*
4294 			 * Compensate for delayed-ack with the d-ack time.
4295 			 */
4296 			counter_u64_add(rack_used_tlpmethod, 1);
4297 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
4298 			if (alt_thresh > thresh)
4299 				thresh = alt_thresh;
4300 		}
4301 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
4302 		/* 2.1 behavior */
4303 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
4304 		if (prsm && (len <= segsiz)) {
4305 			/*
4306 			 * Two packets outstanding, thresh should be (2*srtt) +
4307 			 * possible inter-packet delay (if any).
4308 			 */
4309 			uint32_t inter_gap = 0;
4310 			int idx, nidx;
4311 
4312 			counter_u64_add(rack_used_tlpmethod, 1);
4313 			idx = rsm->r_rtr_cnt - 1;
4314 			nidx = prsm->r_rtr_cnt - 1;
4315 			if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4316 				/* Yes it was sent later (or at the same time) */
4317 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4318 			}
4319 			thresh += inter_gap;
4320 		} else 	if (len <= segsiz) {
4321 			/*
4322 			 * Possibly compensate for delayed-ack.
4323 			 */
4324 			uint32_t alt_thresh;
4325 
4326 			counter_u64_add(rack_used_tlpmethod2, 1);
4327 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
4328 			if (alt_thresh > thresh)
4329 				thresh = alt_thresh;
4330 		}
4331 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
4332 		/* 2.2 behavior */
4333 		if (len <= segsiz) {
4334 			uint32_t alt_thresh;
4335 			/*
4336 			 * Compensate for delayed-ack with the d-ack time.
4337 			 */
4338 			counter_u64_add(rack_used_tlpmethod, 1);
4339 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
4340 			if (alt_thresh > thresh)
4341 				thresh = alt_thresh;
4342 		}
4343 	}
4344  	/* Not above an RTO */
4345 	if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) {
4346 		thresh = TICKS_2_MSEC(tp->t_rxtcur);
4347 	}
4348 	/* Not above a RTO max */
4349 	if (thresh > rack_rto_max) {
4350 		thresh = rack_rto_max;
4351 	}
4352 	/* Apply user supplied min TLP */
4353 	if (thresh < rack_tlp_min) {
4354 		thresh = rack_tlp_min;
4355 	}
4356 	return (thresh);
4357 }
4358 
4359 static uint32_t
4360 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
4361 {
4362 	/*
4363 	 * We want the rack_rtt which is the
4364 	 * last rtt we measured. However if that
4365 	 * does not exist we fallback to the srtt (which
4366 	 * we probably will never do) and then as a last
4367 	 * resort we use RACK_INITIAL_RTO if no srtt is
4368 	 * yet set.
4369 	 */
4370 	if (rack->rc_rack_rtt)
4371 		return(rack->rc_rack_rtt);
4372 	else if (tp->t_srtt == 0)
4373 		return(RACK_INITIAL_RTO);
4374 	return (TICKS_2_MSEC(tp->t_srtt >> TCP_RTT_SHIFT));
4375 }
4376 
4377 static struct rack_sendmap *
4378 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
4379 {
4380 	/*
4381 	 * Check to see that we don't need to fall into recovery. We will
4382 	 * need to do so if our oldest transmit is past the time we should
4383 	 * have had an ack.
4384 	 */
4385 	struct tcp_rack *rack;
4386 	struct rack_sendmap *rsm;
4387 	int32_t idx;
4388 	uint32_t srtt, thresh;
4389 
4390 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4391 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
4392 		return (NULL);
4393 	}
4394 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
4395 	if (rsm == NULL)
4396 		return (NULL);
4397 
4398 	if (rsm->r_flags & RACK_ACKED) {
4399 		rsm = rack_find_lowest_rsm(rack);
4400 		if (rsm == NULL)
4401 			return (NULL);
4402 	}
4403 	idx = rsm->r_rtr_cnt - 1;
4404 	srtt = rack_grab_rtt(tp, rack);
4405 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
4406 	if (TSTMP_LT(tsused, rsm->r_tim_lastsent[idx])) {
4407 		return (NULL);
4408 	}
4409 	if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) {
4410 		return (NULL);
4411 	}
4412 	/* Ok if we reach here we are over-due and this guy can be sent */
4413 	if (IN_RECOVERY(tp->t_flags) == 0) {
4414 		/*
4415 		 * For the one that enters us into recovery record undo
4416 		 * info.
4417 		 */
4418 		rack->r_ctl.rc_rsm_start = rsm->r_start;
4419 		rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
4420 		rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
4421 	}
4422 	rack_cong_signal(tp, NULL, CC_NDUPACK);
4423 	return (rsm);
4424 }
4425 
4426 static uint32_t
4427 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
4428 {
4429 	int32_t t;
4430 	int32_t tt;
4431 	uint32_t ret_val;
4432 
4433 	t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT));
4434 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
4435 	    rack_persist_min, rack_persist_max);
4436 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
4437 		tp->t_rxtshift++;
4438 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
4439 	ret_val = (uint32_t)tt;
4440 	return (ret_val);
4441 }
4442 
4443 static uint32_t
4444 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
4445 {
4446 	/*
4447 	 * Start the FR timer, we do this based on getting the first one in
4448 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
4449 	 * events we need to stop the running timer (if its running) before
4450 	 * starting the new one.
4451 	 */
4452 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
4453 	uint32_t srtt_cur;
4454 	int32_t idx;
4455 	int32_t is_tlp_timer = 0;
4456 	struct rack_sendmap *rsm;
4457 
4458 	if (rack->t_timers_stopped) {
4459 		/* All timers have been stopped none are to run */
4460 		return (0);
4461 	}
4462 	if (rack->rc_in_persist) {
4463 		/* We can't start any timer in persists */
4464 		return (rack_get_persists_timer_val(tp, rack));
4465 	}
4466 	rack->rc_on_min_to = 0;
4467 	if ((tp->t_state < TCPS_ESTABLISHED) ||
4468 	    ((tp->t_flags & TF_SACK_PERMIT) == 0))
4469 		goto activate_rxt;
4470 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
4471 	if ((rsm == NULL) || sup_rack) {
4472 		/* Nothing on the send map */
4473 activate_rxt:
4474 		time_since_sent = 0;
4475 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
4476 		if (rsm) {
4477 			idx = rsm->r_rtr_cnt - 1;
4478 			if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], rack->r_ctl.rc_tlp_rxt_last_time))
4479 				tstmp_touse = rsm->r_tim_lastsent[idx];
4480 			else
4481 				tstmp_touse = rack->r_ctl.rc_tlp_rxt_last_time;
4482 			if (TSTMP_GT(cts, tstmp_touse))
4483 			    time_since_sent = cts - tstmp_touse;
4484 		}
4485 		if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4486 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
4487 			to = TICKS_2_MSEC(tp->t_rxtcur);
4488 			if (to > time_since_sent)
4489 				to -= time_since_sent;
4490 			else
4491 				to = rack->r_ctl.rc_min_to;
4492 			if (to == 0)
4493 				to = 1;
4494 			return (to);
4495 		}
4496 		return (0);
4497 	}
4498 	if (rsm->r_flags & RACK_ACKED) {
4499 		rsm = rack_find_lowest_rsm(rack);
4500 		if (rsm == NULL) {
4501 			/* No lowest? */
4502 			goto activate_rxt;
4503 		}
4504 	}
4505 	if (rack->sack_attack_disable) {
4506 		/*
4507 		 * We don't want to do
4508 		 * any TLP's if you are an attacker.
4509 		 * Though if you are doing what
4510 		 * is expected you may still have
4511 		 * SACK-PASSED marks.
4512 		 */
4513 		goto activate_rxt;
4514 	}
4515 	/* Convert from ms to usecs */
4516 	if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
4517 		if ((tp->t_flags & TF_SENTFIN) &&
4518 		    ((tp->snd_max - tp->snd_una) == 1) &&
4519 		    (rsm->r_flags & RACK_HAS_FIN)) {
4520 			/*
4521 			 * We don't start a rack timer if all we have is a
4522 			 * FIN outstanding.
4523 			 */
4524 			goto activate_rxt;
4525 		}
4526 		if ((rack->use_rack_rr == 0) &&
4527 		    (IN_RECOVERY(tp->t_flags)) &&
4528 		    (rack->rack_no_prr == 0) &&
4529 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
4530 			/*
4531 			 * We are not cheating, in recovery  and
4532 			 * not enough ack's to yet get our next
4533 			 * retransmission out.
4534 			 *
4535 			 * Note that classified attackers do not
4536 			 * get to use the rack-cheat.
4537 			 */
4538 			goto activate_tlp;
4539 		}
4540 		srtt = rack_grab_rtt(tp, rack);
4541 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
4542 		idx = rsm->r_rtr_cnt - 1;
4543 		exp = rsm->r_tim_lastsent[idx] + thresh;
4544 		if (SEQ_GEQ(exp, cts)) {
4545 			to = exp - cts;
4546 			if (to < rack->r_ctl.rc_min_to) {
4547 				to = rack->r_ctl.rc_min_to;
4548 				if (rack->r_rr_config == 3)
4549 					rack->rc_on_min_to = 1;
4550 			}
4551 		} else {
4552 			to = rack->r_ctl.rc_min_to;
4553 			if (rack->r_rr_config == 3)
4554 				rack->rc_on_min_to = 1;
4555 		}
4556 	} else {
4557 		/* Ok we need to do a TLP not RACK */
4558 activate_tlp:
4559 		if ((rack->rc_tlp_in_progress != 0) &&
4560 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
4561 			/*
4562 			 * The previous send was a TLP and we have sent
4563 			 * N TLP's without sending new data.
4564 			 */
4565 			goto activate_rxt;
4566 		}
4567 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
4568 		if (rsm == NULL) {
4569 			/* We found no rsm to TLP with. */
4570 			goto activate_rxt;
4571 		}
4572 		if (rsm->r_flags & RACK_HAS_FIN) {
4573 			/* If its a FIN we dont do TLP */
4574 			rsm = NULL;
4575 			goto activate_rxt;
4576 		}
4577 		idx = rsm->r_rtr_cnt - 1;
4578 		time_since_sent = 0;
4579 		if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], rack->r_ctl.rc_tlp_rxt_last_time))
4580 			tstmp_touse = rsm->r_tim_lastsent[idx];
4581 		else
4582 			tstmp_touse = rack->r_ctl.rc_tlp_rxt_last_time;
4583 		if (TSTMP_GT(cts, tstmp_touse))
4584 		    time_since_sent = cts - tstmp_touse;
4585 		is_tlp_timer = 1;
4586 		if (tp->t_srtt) {
4587 			srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT);
4588 			srtt = TICKS_2_MSEC(srtt_cur);
4589 		} else
4590 			srtt = RACK_INITIAL_RTO;
4591 		/*
4592 		 * If the SRTT is not keeping up and the
4593 		 * rack RTT has spiked we want to use
4594 		 * the last RTT not the smoothed one.
4595 		 */
4596 		if (rack_tlp_use_greater && (srtt < rack_grab_rtt(tp, rack)))
4597 			srtt = rack_grab_rtt(tp, rack);
4598 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
4599 		if (thresh > time_since_sent)
4600 			to = thresh - time_since_sent;
4601 		else {
4602 			to = rack->r_ctl.rc_min_to;
4603 			rack_log_alt_to_to_cancel(rack,
4604 						  thresh,		/* flex1 */
4605 						  time_since_sent,	/* flex2 */
4606 						  tstmp_touse,		/* flex3 */
4607 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
4608 						  rsm->r_tim_lastsent[idx],
4609 						  srtt,
4610 						  idx, 99);
4611 		}
4612 		if (to > TCPTV_REXMTMAX) {
4613 			/*
4614 			 * If the TLP time works out to larger than the max
4615 			 * RTO lets not do TLP.. just RTO.
4616 			 */
4617 			goto activate_rxt;
4618 		}
4619 	}
4620 	if (is_tlp_timer == 0) {
4621 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
4622 	} else {
4623 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
4624 	}
4625 	if (to == 0)
4626 		to = 1;
4627 	return (to);
4628 }
4629 
4630 static void
4631 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
4632 {
4633 	if (rack->rc_in_persist == 0) {
4634 		if (tp->t_flags & TF_GPUTINPROG) {
4635 			/*
4636 			 * Stop the goodput now, the calling of the
4637 			 * measurement function clears the flag.
4638 			 */
4639 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__);
4640 		}
4641 #ifdef NETFLIX_SHARED_CWND
4642 		if (rack->r_ctl.rc_scw) {
4643 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
4644 			rack->rack_scwnd_is_idle = 1;
4645 		}
4646 #endif
4647 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
4648 		if (rack->r_ctl.rc_went_idle_time == 0)
4649 			rack->r_ctl.rc_went_idle_time = 1;
4650 		rack_timer_cancel(tp, rack, cts, __LINE__);
4651 		tp->t_rxtshift = 0;
4652 		rack->rc_in_persist = 1;
4653 	}
4654 }
4655 
4656 static void
4657 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
4658 {
4659 	if (rack->rc_inp->inp_in_hpts)  {
4660 		tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
4661 		rack->r_ctl.rc_hpts_flags  = 0;
4662 	}
4663 #ifdef NETFLIX_SHARED_CWND
4664 	if (rack->r_ctl.rc_scw) {
4665 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
4666 		rack->rack_scwnd_is_idle = 0;
4667 	}
4668 #endif
4669 	if (rack->rc_gp_dyn_mul &&
4670 	    (rack->use_fixed_rate == 0) &&
4671 	    (rack->rc_always_pace)) {
4672 		/*
4673 		 * Do we count this as if a probe-rtt just
4674 		 * finished?
4675 		 */
4676 		uint32_t time_idle, idle_min;
4677 
4678 		time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
4679 		idle_min = rack_min_probertt_hold;
4680 		if (rack_probertt_gpsrtt_cnt_div) {
4681 			uint64_t extra;
4682 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
4683 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
4684 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
4685 			idle_min += (uint32_t)extra;
4686 		}
4687 		if (time_idle >= idle_min)  {
4688 			/* Yes, we count it as a probe-rtt. */
4689 			uint32_t us_cts;
4690 
4691 			us_cts = tcp_get_usecs(NULL);
4692 			if (rack->in_probe_rtt == 0) {
4693 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4694 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
4695 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
4696 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
4697 			} else {
4698 				rack_exit_probertt(rack, us_cts);
4699 			}
4700 		}
4701 
4702 	}
4703 	rack->rc_in_persist = 0;
4704 	rack->r_ctl.rc_went_idle_time = 0;
4705 	tp->t_rxtshift = 0;
4706  	rack->r_ctl.rc_agg_delayed = 0;
4707 	rack->r_early = 0;
4708 	rack->r_late = 0;
4709 	rack->r_ctl.rc_agg_early = 0;
4710 }
4711 
4712 static void
4713 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
4714 		   struct hpts_diag *diag, struct timeval *tv)
4715 {
4716 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
4717 		union tcp_log_stackspecific log;
4718 
4719 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4720 		log.u_bbr.flex1 = diag->p_nxt_slot;
4721 		log.u_bbr.flex2 = diag->p_cur_slot;
4722 		log.u_bbr.flex3 = diag->slot_req;
4723 		log.u_bbr.flex4 = diag->inp_hptsslot;
4724 		log.u_bbr.flex5 = diag->slot_remaining;
4725 		log.u_bbr.flex6 = diag->need_new_to;
4726 		log.u_bbr.flex7 = diag->p_hpts_active;
4727 		log.u_bbr.flex8 = diag->p_on_min_sleep;
4728 		/* Hijack other fields as needed  */
4729 		log.u_bbr.epoch = diag->have_slept;
4730 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
4731 		log.u_bbr.pkts_out = diag->co_ret;
4732 		log.u_bbr.applimited = diag->hpts_sleep_time;
4733 		log.u_bbr.delivered = diag->p_prev_slot;
4734 		log.u_bbr.inflight = diag->p_runningtick;
4735 		log.u_bbr.bw_inuse = diag->wheel_tick;
4736 		log.u_bbr.rttProp = diag->wheel_cts;
4737 		log.u_bbr.timeStamp = cts;
4738 		log.u_bbr.delRate = diag->maxticks;
4739 		log.u_bbr.cur_del_rate = diag->p_curtick;
4740 		log.u_bbr.cur_del_rate <<= 32;
4741 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
4742 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4743 		    &rack->rc_inp->inp_socket->so_rcv,
4744 		    &rack->rc_inp->inp_socket->so_snd,
4745 		    BBR_LOG_HPTSDIAG, 0,
4746 		    0, &log, false, tv);
4747 	}
4748 
4749 }
4750 
4751 static void
4752 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
4753       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
4754 {
4755 	struct hpts_diag diag;
4756 	struct inpcb *inp;
4757 	struct timeval tv;
4758 	uint32_t delayed_ack = 0;
4759 	uint32_t hpts_timeout;
4760 	uint8_t stopped;
4761 	uint32_t left = 0;
4762 	uint32_t us_cts;
4763 
4764 	inp = tp->t_inpcb;
4765 	if ((tp->t_state == TCPS_CLOSED) ||
4766 	    (tp->t_state == TCPS_LISTEN)) {
4767 		return;
4768 	}
4769 	if (inp->inp_in_hpts) {
4770 		/* Already on the pacer */
4771 		return;
4772 	}
4773 	stopped = rack->rc_tmr_stopped;
4774 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
4775 		left = rack->r_ctl.rc_timer_exp - cts;
4776 	}
4777 	rack->r_ctl.rc_timer_exp = 0;
4778 	rack->r_ctl.rc_hpts_flags = 0;
4779 	us_cts = tcp_get_usecs(&tv);
4780 	/* Now early/late accounting */
4781 	if (rack->r_early) {
4782 		/*
4783 		 * We have a early carry over set,
4784 		 * we can always add more time so we
4785 		 * can always make this compensation.
4786 		 */
4787 		slot += rack->r_ctl.rc_agg_early;
4788 		rack->r_early = 0;
4789 		rack->r_ctl.rc_agg_early = 0;
4790 	}
4791 	if (rack->r_late) {
4792 		/*
4793 		 * This is harder, we can
4794 		 * compensate some but it
4795 		 * really depends on what
4796 		 * the current pacing time is.
4797 		 */
4798 		if (rack->r_ctl.rc_agg_delayed >= slot) {
4799 			/*
4800 			 * We can't compensate for it all.
4801 			 * And we have to have some time
4802 			 * on the clock. We always have a min
4803 			 * 10 slots (10 x 10 i.e. 100 usecs).
4804 			 */
4805 			if (slot <= HPTS_TICKS_PER_USEC) {
4806 				/* We gain delay */
4807 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_USEC - slot);
4808 				slot = HPTS_TICKS_PER_USEC;
4809 			} else {
4810 				/* We take off some */
4811 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_USEC);
4812 				slot = HPTS_TICKS_PER_USEC;
4813 			}
4814 		} else {
4815 
4816 			slot -= rack->r_ctl.rc_agg_delayed;
4817 			rack->r_ctl.rc_agg_delayed = 0;
4818 			/* Make sure we have 100 useconds at minimum */
4819 			if (slot < HPTS_TICKS_PER_USEC) {
4820 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_USEC - slot;
4821 				slot = HPTS_TICKS_PER_USEC;
4822 			}
4823 			if (rack->r_ctl.rc_agg_delayed == 0)
4824 				rack->r_late = 0;
4825 		}
4826 	}
4827 	if (slot) {
4828 		/* We are pacing too */
4829 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
4830 	}
4831 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
4832 #ifdef NETFLIX_EXP_DETECTION
4833 	if (rack->sack_attack_disable &&
4834 	    (slot < tcp_sad_pacing_interval)) {
4835 		/*
4836 		 * We have a potential attacker on
4837 		 * the line. We have possibly some
4838 		 * (or now) pacing time set. We want to
4839 		 * slow down the processing of sacks by some
4840 		 * amount (if it is an attacker). Set the default
4841 		 * slot for attackers in place (unless the orginal
4842 		 * interval is longer). Its stored in
4843 		 * micro-seconds, so lets convert to msecs.
4844 		 */
4845 		slot = tcp_sad_pacing_interval;
4846 	}
4847 #endif
4848 	if (tp->t_flags & TF_DELACK) {
4849 		delayed_ack = TICKS_2_MSEC(tcp_delacktime);
4850 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
4851 	}
4852 	if (delayed_ack && ((hpts_timeout == 0) ||
4853 			    (delayed_ack < hpts_timeout)))
4854 		hpts_timeout = delayed_ack;
4855 	else
4856 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4857 	/*
4858 	 * If no timers are going to run and we will fall off the hptsi
4859 	 * wheel, we resort to a keep-alive timer if its configured.
4860 	 */
4861 	if ((hpts_timeout == 0) &&
4862 	    (slot == 0)) {
4863 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4864 		    (tp->t_state <= TCPS_CLOSING)) {
4865 			/*
4866 			 * Ok we have no timer (persists, rack, tlp, rxt  or
4867 			 * del-ack), we don't have segments being paced. So
4868 			 * all that is left is the keepalive timer.
4869 			 */
4870 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
4871 				/* Get the established keep-alive time */
4872 				hpts_timeout = TP_KEEPIDLE(tp);
4873 			} else {
4874 				/* Get the initial setup keep-alive time */
4875 				hpts_timeout = TP_KEEPINIT(tp);
4876 			}
4877 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
4878 			if (rack->in_probe_rtt) {
4879 				/*
4880 				 * We want to instead not wake up a long time from
4881 				 * now but to wake up about the time we would
4882 				 * exit probe-rtt and initiate a keep-alive ack.
4883 				 * This will get us out of probe-rtt and update
4884 				 * our min-rtt.
4885 				 */
4886 				hpts_timeout = (rack_min_probertt_hold / HPTS_USEC_IN_MSEC);
4887 			}
4888 		}
4889 	}
4890 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
4891 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
4892 		/*
4893 		 * RACK, TLP, persists and RXT timers all are restartable
4894 		 * based on actions input .. i.e we received a packet (ack
4895 		 * or sack) and that changes things (rw, or snd_una etc).
4896 		 * Thus we can restart them with a new value. For
4897 		 * keep-alive, delayed_ack we keep track of what was left
4898 		 * and restart the timer with a smaller value.
4899 		 */
4900 		if (left < hpts_timeout)
4901 			hpts_timeout = left;
4902 	}
4903 	if (hpts_timeout) {
4904 		/*
4905 		 * Hack alert for now we can't time-out over 2,147,483
4906 		 * seconds (a bit more than 596 hours), which is probably ok
4907 		 * :).
4908 		 */
4909 		if (hpts_timeout > 0x7ffffffe)
4910 			hpts_timeout = 0x7ffffffe;
4911 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
4912 	}
4913 	if ((rack->rc_gp_filled == 0) &&
4914 	    (hpts_timeout < slot) &&
4915 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
4916 		/*
4917 		 * We have no good estimate yet for the
4918 		 * old clunky burst mitigation or the
4919 		 * real pacing. And the tlp or rxt is smaller
4920 		 * than the pacing calculation. Lets not
4921 		 * pace that long since we know the calculation
4922 		 * so far is not accurate.
4923 		 */
4924 		slot = hpts_timeout;
4925 	}
4926 	rack->r_ctl.last_pacing_time = slot;
4927 	if (slot) {
4928 		rack->r_ctl.rc_last_output_to = us_cts + slot;
4929 		if (rack->rc_always_pace || rack->r_mbuf_queue) {
4930 			if ((rack->rc_gp_filled == 0) ||
4931 			    rack->pacing_longer_than_rtt) {
4932 				inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
4933 			} else {
4934 				inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
4935 				if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
4936 				    (rack->r_rr_config != 3))
4937 					inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
4938 				else
4939 					inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
4940 			}
4941 		}
4942 		if ((rack->use_rack_rr) &&
4943 		    (rack->r_rr_config < 2) &&
4944 		    ((hpts_timeout) && ((hpts_timeout * HPTS_USEC_IN_MSEC) < slot))) {
4945 			/*
4946 			 * Arrange for the hpts to kick back in after the
4947 			 * t-o if the t-o does not cause a send.
4948 			 */
4949 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout),
4950 						   __LINE__, &diag);
4951 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
4952 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
4953 		} else {
4954 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
4955 						   __LINE__, &diag);
4956 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
4957 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
4958 		}
4959 	} else if (hpts_timeout) {
4960 		if (rack->rc_always_pace || rack->r_mbuf_queue) {
4961 			if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)  {
4962 				/* For a rack timer, don't wake us */
4963 				inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
4964 				if  (rack->r_rr_config != 3)
4965 					inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
4966 				else
4967 					inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
4968 			} else {
4969 				/* All other timers wake us up */
4970 				inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
4971 				inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
4972 			}
4973 		}
4974 		(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout),
4975 					   __LINE__, &diag);
4976 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
4977 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
4978 	} else {
4979 		/* No timer starting */
4980 #ifdef INVARIANTS
4981 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
4982 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
4983 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
4984 		}
4985 #endif
4986 	}
4987 	rack->rc_tmr_stopped = 0;
4988 	if (slot)
4989 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
4990 }
4991 
4992 /*
4993  * RACK Timer, here we simply do logging and house keeping.
4994  * the normal rack_output() function will call the
4995  * appropriate thing to check if we need to do a RACK retransmit.
4996  * We return 1, saying don't proceed with rack_output only
4997  * when all timers have been stopped (destroyed PCB?).
4998  */
4999 static int
5000 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5001 {
5002 	/*
5003 	 * This timer simply provides an internal trigger to send out data.
5004 	 * The check_recovery_mode call will see if there are needed
5005 	 * retransmissions, if so we will enter fast-recovery. The output
5006 	 * call may or may not do the same thing depending on sysctl
5007 	 * settings.
5008 	 */
5009 	struct rack_sendmap *rsm;
5010 	int32_t recovery;
5011 
5012 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5013 		return (1);
5014 	}
5015 	recovery = IN_RECOVERY(tp->t_flags);
5016 	counter_u64_add(rack_to_tot, 1);
5017 	if (rack->r_state && (rack->r_state != tp->t_state))
5018 		rack_set_state(tp, rack);
5019 	rack->rc_on_min_to = 0;
5020 	rsm = rack_check_recovery_mode(tp, cts);
5021 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5022 	if (rsm) {
5023 		uint32_t rtt;
5024 
5025 		rack->r_ctl.rc_resend = rsm;
5026 		if (rack->use_rack_rr) {
5027 			/*
5028 			 * Don't accumulate extra pacing delay
5029 			 * we are allowing the rack timer to
5030 			 * over-ride pacing i.e. rrr takes precedence
5031 			 * if the pacing interval is longer than the rrr
5032 			 * time (in other words we get the min pacing
5033 			 * time versus rrr pacing time).
5034 			 */
5035 			rack->r_timer_override = 1;
5036 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5037 		}
5038 		rtt = rack->rc_rack_rtt;
5039 		if (rtt == 0)
5040 			rtt = 1;
5041 		if (rack->rack_no_prr == 0) {
5042 			if ((recovery == 0) &&
5043 			    (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) {
5044 				/*
5045 				 * The rack-timeout that enter's us into recovery
5046 				 * will force out one MSS and set us up so that we
5047 				 * can do one more send in 2*rtt (transitioning the
5048 				 * rack timeout into a rack-tlp).
5049 				 */
5050 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5051 				rack->r_timer_override = 1;
5052 				rack_log_to_prr(rack, 3, 0);
5053 			} else if ((rack->r_ctl.rc_prr_sndcnt < (rsm->r_end - rsm->r_start)) &&
5054 				   rack->use_rack_rr) {
5055 				/*
5056 				 * When a rack timer goes, if the rack rr is
5057 				 * on, arrange it so we can send a full segment
5058 				 * overriding prr (though we pay a price for this
5059 				 * for future new sends).
5060 				 */
5061 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5062 				rack_log_to_prr(rack, 4, 0);
5063 			}
5064 		}
5065 	}
5066 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5067 	if (rsm == NULL) {
5068 		/* restart a timer and return 1 */
5069 		rack_start_hpts_timer(rack, tp, cts,
5070 				      0, 0, 0);
5071 		return (1);
5072 	}
5073 	return (0);
5074 }
5075 
5076 static __inline void
5077 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5078 	       struct rack_sendmap *rsm, uint32_t start)
5079 {
5080 	int idx;
5081 
5082 	nrsm->r_start = start;
5083 	nrsm->r_end = rsm->r_end;
5084 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
5085 	nrsm->r_flags = rsm->r_flags;
5086 	nrsm->r_dupack = rsm->r_dupack;
5087 	nrsm->usec_orig_send = rsm->usec_orig_send;
5088 	nrsm->r_rtr_bytes = 0;
5089 	rsm->r_end = nrsm->r_start;
5090 	nrsm->r_just_ret = rsm->r_just_ret;
5091 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
5092 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
5093 	}
5094 }
5095 
5096 static struct rack_sendmap *
5097 rack_merge_rsm(struct tcp_rack *rack,
5098 	       struct rack_sendmap *l_rsm,
5099 	       struct rack_sendmap *r_rsm)
5100 {
5101 	/*
5102 	 * We are merging two ack'd RSM's,
5103 	 * the l_rsm is on the left (lower seq
5104 	 * values) and the r_rsm is on the right
5105 	 * (higher seq value). The simplest way
5106 	 * to merge these is to move the right
5107 	 * one into the left. I don't think there
5108 	 * is any reason we need to try to find
5109 	 * the oldest (or last oldest retransmitted).
5110 	 */
5111 	struct rack_sendmap *rm;
5112 
5113 	l_rsm->r_end = r_rsm->r_end;
5114 	if (l_rsm->r_dupack < r_rsm->r_dupack)
5115 		l_rsm->r_dupack = r_rsm->r_dupack;
5116 	if (r_rsm->r_rtr_bytes)
5117 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
5118 	if (r_rsm->r_in_tmap) {
5119 		/* This really should not happen */
5120 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
5121 		r_rsm->r_in_tmap = 0;
5122 	}
5123 
5124 	/* Now the flags */
5125 	if (r_rsm->r_flags & RACK_HAS_FIN)
5126 		l_rsm->r_flags |= RACK_HAS_FIN;
5127 	if (r_rsm->r_flags & RACK_TLP)
5128 		l_rsm->r_flags |= RACK_TLP;
5129 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
5130 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
5131 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
5132 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
5133 		/*
5134 		 * If both are app-limited then let the
5135 		 * free lower the count. If right is app
5136 		 * limited and left is not, transfer.
5137 		 */
5138 		l_rsm->r_flags |= RACK_APP_LIMITED;
5139 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
5140 		if (r_rsm == rack->r_ctl.rc_first_appl)
5141 			rack->r_ctl.rc_first_appl = l_rsm;
5142 	}
5143 	rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
5144 #ifdef INVARIANTS
5145 	if (rm != r_rsm) {
5146 		panic("removing head in rack:%p rsm:%p rm:%p",
5147 		      rack, r_rsm, rm);
5148 	}
5149 #endif
5150 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
5151 		/* Transfer the split limit to the map we free */
5152 		r_rsm->r_limit_type = l_rsm->r_limit_type;
5153 		l_rsm->r_limit_type = 0;
5154 	}
5155 	rack_free(rack, r_rsm);
5156 	return(l_rsm);
5157 }
5158 
5159 /*
5160  * TLP Timer, here we simply setup what segment we want to
5161  * have the TLP expire on, the normal rack_output() will then
5162  * send it out.
5163  *
5164  * We return 1, saying don't proceed with rack_output only
5165  * when all timers have been stopped (destroyed PCB?).
5166  */
5167 static int
5168 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5169 {
5170 	/*
5171 	 * Tail Loss Probe.
5172 	 */
5173 	struct rack_sendmap *rsm = NULL;
5174 	struct rack_sendmap *insret;
5175 	struct socket *so;
5176 	uint32_t amm, old_prr_snd = 0;
5177 	uint32_t out, avail;
5178 	int collapsed_win = 0;
5179 
5180 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5181 		return (1);
5182 	}
5183 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
5184 		/* Its not time yet */
5185 		return (0);
5186 	}
5187 	if (ctf_progress_timeout_check(tp, true)) {
5188 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
5189 		tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
5190 		return (1);
5191 	}
5192 	/*
5193 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
5194 	 * need to figure out how to force a full MSS segment out.
5195 	 */
5196 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
5197 	counter_u64_add(rack_tlp_tot, 1);
5198 	if (rack->r_state && (rack->r_state != tp->t_state))
5199 		rack_set_state(tp, rack);
5200 	so = tp->t_inpcb->inp_socket;
5201 	avail = sbavail(&so->so_snd);
5202 	out = tp->snd_max - tp->snd_una;
5203 	if (out > tp->snd_wnd) {
5204 		/* special case, we need a retransmission */
5205 		collapsed_win = 1;
5206 		goto need_retran;
5207 	}
5208 	/*
5209 	 * Check our send oldest always settings, and if
5210 	 * there is an oldest to send jump to the need_retran.
5211 	 */
5212 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
5213 		goto need_retran;
5214 
5215 	if (avail > out) {
5216 		/* New data is available */
5217 		amm = avail - out;
5218 		if (amm > ctf_fixed_maxseg(tp)) {
5219 			amm = ctf_fixed_maxseg(tp);
5220 			if ((amm + out) > tp->snd_wnd) {
5221 				/* We are rwnd limited */
5222 				goto need_retran;
5223 			}
5224 		} else if (amm < ctf_fixed_maxseg(tp)) {
5225 			/* not enough to fill a MTU */
5226 			goto need_retran;
5227 		}
5228 		if (IN_RECOVERY(tp->t_flags)) {
5229 			/* Unlikely */
5230 			if (rack->rack_no_prr == 0) {
5231 				old_prr_snd = rack->r_ctl.rc_prr_sndcnt;
5232 				if (out + amm <= tp->snd_wnd) {
5233 					rack->r_ctl.rc_prr_sndcnt = amm;
5234 					rack_log_to_prr(rack, 4, 0);
5235 				}
5236 			} else
5237 				goto need_retran;
5238 		} else {
5239 			/* Set the send-new override */
5240 			if (out + amm <= tp->snd_wnd)
5241 				rack->r_ctl.rc_tlp_new_data = amm;
5242 			else
5243 				goto need_retran;
5244 		}
5245 		rack->r_ctl.rc_tlpsend = NULL;
5246 		counter_u64_add(rack_tlp_newdata, 1);
5247 		goto send;
5248 	}
5249 need_retran:
5250 	/*
5251 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
5252 	 * optionally the first un-acked segment.
5253 	 */
5254 	if (collapsed_win == 0) {
5255 		if (rack_always_send_oldest)
5256 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5257 		else {
5258 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
5259 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
5260 				rsm = rack_find_high_nonack(rack, rsm);
5261 			}
5262 		}
5263 		if (rsm == NULL) {
5264 			counter_u64_add(rack_tlp_does_nada, 1);
5265 #ifdef TCP_BLACKBOX
5266 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
5267 #endif
5268 			goto out;
5269 		}
5270 	} else {
5271 		/*
5272 		 * We must find the last segment
5273 		 * that was acceptable by the client.
5274 		 */
5275 		RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
5276 			if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
5277 				/* Found one */
5278 				break;
5279 			}
5280 		}
5281 		if (rsm == NULL) {
5282 			/* None? if so send the first */
5283 			rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
5284 			if (rsm == NULL) {
5285 				counter_u64_add(rack_tlp_does_nada, 1);
5286 #ifdef TCP_BLACKBOX
5287 				tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
5288 #endif
5289 				goto out;
5290 			}
5291 		}
5292 	}
5293 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
5294 		/*
5295 		 * We need to split this the last segment in two.
5296 		 */
5297 		struct rack_sendmap *nrsm;
5298 
5299 
5300 		nrsm = rack_alloc_full_limit(rack);
5301 		if (nrsm == NULL) {
5302 			/*
5303 			 * No memory to split, we will just exit and punt
5304 			 * off to the RXT timer.
5305 			 */
5306 			counter_u64_add(rack_tlp_does_nada, 1);
5307 			goto out;
5308 		}
5309 		rack_clone_rsm(rack, nrsm, rsm,
5310 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
5311 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
5312 #ifdef INVARIANTS
5313 		if (insret != NULL) {
5314 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
5315 			      nrsm, insret, rack, rsm);
5316 		}
5317 #endif
5318 		if (rsm->r_in_tmap) {
5319 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5320 			nrsm->r_in_tmap = 1;
5321 		}
5322 		rsm->r_flags &= (~RACK_HAS_FIN);
5323 		rsm = nrsm;
5324 	}
5325 	rack->r_ctl.rc_tlpsend = rsm;
5326 send:
5327 	rack->r_timer_override = 1;
5328 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
5329 	return (0);
5330 out:
5331 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
5332 	return (0);
5333 }
5334 
5335 /*
5336  * Delayed ack Timer, here we simply need to setup the
5337  * ACK_NOW flag and remove the DELACK flag. From there
5338  * the output routine will send the ack out.
5339  *
5340  * We only return 1, saying don't proceed, if all timers
5341  * are stopped (destroyed PCB?).
5342  */
5343 static int
5344 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5345 {
5346 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5347 		return (1);
5348 	}
5349 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
5350 	tp->t_flags &= ~TF_DELACK;
5351 	tp->t_flags |= TF_ACKNOW;
5352 	KMOD_TCPSTAT_INC(tcps_delack);
5353 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5354 	return (0);
5355 }
5356 
5357 /*
5358  * Persists timer, here we simply send the
5359  * same thing as a keepalive will.
5360  * the one byte send.
5361  *
5362  * We only return 1, saying don't proceed, if all timers
5363  * are stopped (destroyed PCB?).
5364  */
5365 static int
5366 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5367 {
5368 	struct tcptemp *t_template;
5369 	struct inpcb *inp;
5370 	int32_t retval = 1;
5371 
5372 	inp = tp->t_inpcb;
5373 
5374 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5375 		return (1);
5376 	}
5377 	if (rack->rc_in_persist == 0)
5378 		return (0);
5379 	if (ctf_progress_timeout_check(tp, false)) {
5380 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
5381 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
5382 		tcp_set_inp_to_drop(inp, ETIMEDOUT);
5383 		return (1);
5384 	}
5385 	KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
5386 	/*
5387 	 * Persistence timer into zero window. Force a byte to be output, if
5388 	 * possible.
5389 	 */
5390 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
5391 	/*
5392 	 * Hack: if the peer is dead/unreachable, we do not time out if the
5393 	 * window is closed.  After a full backoff, drop the connection if
5394 	 * the idle time (no responses to probes) reaches the maximum
5395 	 * backoff that we would use if retransmitting.
5396 	 */
5397 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
5398 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
5399 	    ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
5400 		KMOD_TCPSTAT_INC(tcps_persistdrop);
5401 		retval = 1;
5402 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
5403 		tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
5404 		goto out;
5405 	}
5406 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
5407 	    tp->snd_una == tp->snd_max)
5408 		rack_exit_persist(tp, rack, cts);
5409 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
5410 	/*
5411 	 * If the user has closed the socket then drop a persisting
5412 	 * connection after a much reduced timeout.
5413 	 */
5414 	if (tp->t_state > TCPS_CLOSE_WAIT &&
5415 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
5416 		retval = 1;
5417 		KMOD_TCPSTAT_INC(tcps_persistdrop);
5418 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
5419 		tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
5420 		goto out;
5421 	}
5422 	t_template = tcpip_maketemplate(rack->rc_inp);
5423 	if (t_template) {
5424 		/* only set it if we were answered */
5425 		if (rack->forced_ack == 0) {
5426 			rack->forced_ack = 1;
5427 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
5428 		}
5429 		tcp_respond(tp, t_template->tt_ipgen,
5430 			    &t_template->tt_t, (struct mbuf *)NULL,
5431 			    tp->rcv_nxt, tp->snd_una - 1, 0);
5432 		/* This sends an ack */
5433 		if (tp->t_flags & TF_DELACK)
5434 			tp->t_flags &= ~TF_DELACK;
5435 		free(t_template, M_TEMP);
5436 	}
5437 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
5438 		tp->t_rxtshift++;
5439 out:
5440 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
5441 	rack_start_hpts_timer(rack, tp, cts,
5442 			      0, 0, 0);
5443 	return (retval);
5444 }
5445 
5446 /*
5447  * If a keepalive goes off, we had no other timers
5448  * happening. We always return 1 here since this
5449  * routine either drops the connection or sends
5450  * out a segment with respond.
5451  */
5452 static int
5453 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5454 {
5455 	struct tcptemp *t_template;
5456 	struct inpcb *inp;
5457 
5458 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5459 		return (1);
5460 	}
5461 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
5462 	inp = tp->t_inpcb;
5463 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
5464 	/*
5465 	 * Keep-alive timer went off; send something or drop connection if
5466 	 * idle for too long.
5467 	 */
5468 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
5469 	if (tp->t_state < TCPS_ESTABLISHED)
5470 		goto dropit;
5471 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5472 	    tp->t_state <= TCPS_CLOSING) {
5473 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
5474 			goto dropit;
5475 		/*
5476 		 * Send a packet designed to force a response if the peer is
5477 		 * up and reachable: either an ACK if the connection is
5478 		 * still alive, or an RST if the peer has closed the
5479 		 * connection due to timeout or reboot. Using sequence
5480 		 * number tp->snd_una-1 causes the transmitted zero-length
5481 		 * segment to lie outside the receive window; by the
5482 		 * protocol spec, this requires the correspondent TCP to
5483 		 * respond.
5484 		 */
5485 		KMOD_TCPSTAT_INC(tcps_keepprobe);
5486 		t_template = tcpip_maketemplate(inp);
5487 		if (t_template) {
5488 			if (rack->forced_ack == 0) {
5489 				rack->forced_ack = 1;
5490 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
5491 			}
5492 			tcp_respond(tp, t_template->tt_ipgen,
5493 			    &t_template->tt_t, (struct mbuf *)NULL,
5494 			    tp->rcv_nxt, tp->snd_una - 1, 0);
5495 			free(t_template, M_TEMP);
5496 		}
5497 	}
5498 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
5499 	return (1);
5500 dropit:
5501 	KMOD_TCPSTAT_INC(tcps_keepdrops);
5502 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
5503 	tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
5504 	return (1);
5505 }
5506 
5507 /*
5508  * Retransmit helper function, clear up all the ack
5509  * flags and take care of important book keeping.
5510  */
5511 static void
5512 rack_remxt_tmr(struct tcpcb *tp)
5513 {
5514 	/*
5515 	 * The retransmit timer went off, all sack'd blocks must be
5516 	 * un-acked.
5517 	 */
5518 	struct rack_sendmap *rsm, *trsm = NULL;
5519 	struct tcp_rack *rack;
5520 	int32_t cnt = 0;
5521 
5522 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5523 	rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__);
5524 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
5525 	if (rack->r_state && (rack->r_state != tp->t_state))
5526 		rack_set_state(tp, rack);
5527 	/*
5528 	 * Ideally we would like to be able to
5529 	 * mark SACK-PASS on anything not acked here.
5530 	 * However, if we do that we would burst out
5531 	 * all that data 1ms apart. This would be unwise,
5532 	 * so for now we will just let the normal rxt timer
5533 	 * and tlp timer take care of it.
5534 	 */
5535 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
5536 		if (rsm->r_flags & RACK_ACKED) {
5537 			cnt++;
5538 			rsm->r_dupack = 0;
5539 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
5540 			if (rsm->r_in_tmap == 0) {
5541 				/* We must re-add it back to the tlist */
5542 				if (trsm == NULL) {
5543 					TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
5544 				} else {
5545 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
5546 				}
5547 				rsm->r_in_tmap = 1;
5548 			}
5549 		}
5550 		trsm = rsm;
5551 		if (rsm->r_flags & RACK_ACKED)
5552 			rsm->r_flags |= RACK_WAS_ACKED;
5553 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
5554 	}
5555 	/* Clear the count (we just un-acked them) */
5556 	rack->r_ctl.rc_sacked = 0;
5557 	rack->r_ctl.rc_agg_delayed = 0;
5558 	rack->r_early = 0;
5559 	rack->r_ctl.rc_agg_early = 0;
5560 	rack->r_late = 0;
5561 	/* Clear the tlp rtx mark */
5562 	rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
5563 	rack->r_ctl.rc_prr_sndcnt = 0;
5564 	rack_log_to_prr(rack, 6, 0);
5565 	rack->r_timer_override = 1;
5566 }
5567 
5568 static void
5569 rack_cc_conn_init(struct tcpcb *tp)
5570 {
5571 	struct tcp_rack *rack;
5572 
5573 
5574 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5575 	cc_conn_init(tp);
5576 	/*
5577 	 * We want a chance to stay in slowstart as
5578 	 * we create a connection. TCP spec says that
5579 	 * initially ssthresh is infinite. For our
5580 	 * purposes that is the snd_wnd.
5581 	 */
5582 	if (tp->snd_ssthresh < tp->snd_wnd) {
5583 		tp->snd_ssthresh = tp->snd_wnd;
5584 	}
5585 	/*
5586 	 * We also want to assure a IW worth of
5587 	 * data can get inflight.
5588 	 */
5589 	if (rc_init_window(rack) < tp->snd_cwnd)
5590 		tp->snd_cwnd = rc_init_window(rack);
5591 }
5592 
5593 /*
5594  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
5595  * we will setup to retransmit the lowest seq number outstanding.
5596  */
5597 static int
5598 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5599 {
5600 	int32_t rexmt;
5601 	struct inpcb *inp;
5602 	int32_t retval = 0;
5603 	bool isipv6;
5604 
5605 	inp = tp->t_inpcb;
5606 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5607 		return (1);
5608 	}
5609 	if (ctf_progress_timeout_check(tp, false)) {
5610 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5611 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
5612 		tcp_set_inp_to_drop(inp, ETIMEDOUT);
5613 		return (1);
5614 	}
5615 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
5616 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
5617 	    (tp->snd_una == tp->snd_max)) {
5618 		/* Nothing outstanding .. nothing to do */
5619 		return (0);
5620 	}
5621 	/*
5622 	 * Retransmission timer went off.  Message has not been acked within
5623 	 * retransmit interval.  Back off to a longer retransmit interval
5624 	 * and retransmit one segment.
5625 	 */
5626 	rack_remxt_tmr(tp);
5627 	if ((rack->r_ctl.rc_resend == NULL) ||
5628 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
5629 		/*
5630 		 * If the rwnd collapsed on
5631 		 * the one we are retransmitting
5632 		 * it does not count against the
5633 		 * rxt count.
5634 		 */
5635 		tp->t_rxtshift++;
5636 	}
5637 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
5638 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
5639 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
5640 		retval = 1;
5641 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5642 		tcp_set_inp_to_drop(rack->rc_inp,
5643 		    (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
5644 		goto out;
5645 	}
5646 	if (tp->t_state == TCPS_SYN_SENT) {
5647 		/*
5648 		 * If the SYN was retransmitted, indicate CWND to be limited
5649 		 * to 1 segment in cc_conn_init().
5650 		 */
5651 		tp->snd_cwnd = 1;
5652 	} else if (tp->t_rxtshift == 1) {
5653 		/*
5654 		 * first retransmit; record ssthresh and cwnd so they can be
5655 		 * recovered if this turns out to be a "bad" retransmit. A
5656 		 * retransmit is considered "bad" if an ACK for this segment
5657 		 * is received within RTT/2 interval; the assumption here is
5658 		 * that the ACK was already in flight.  See "On Estimating
5659 		 * End-to-End Network Path Properties" by Allman and Paxson
5660 		 * for more details.
5661 		 */
5662 		tp->snd_cwnd_prev = tp->snd_cwnd;
5663 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
5664 		tp->snd_recover_prev = tp->snd_recover;
5665 		if (IN_FASTRECOVERY(tp->t_flags))
5666 			tp->t_flags |= TF_WASFRECOVERY;
5667 		else
5668 			tp->t_flags &= ~TF_WASFRECOVERY;
5669 		if (IN_CONGRECOVERY(tp->t_flags))
5670 			tp->t_flags |= TF_WASCRECOVERY;
5671 		else
5672 			tp->t_flags &= ~TF_WASCRECOVERY;
5673 		tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5674 		tp->t_flags |= TF_PREVVALID;
5675 	} else
5676 		tp->t_flags &= ~TF_PREVVALID;
5677 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
5678 	if ((tp->t_state == TCPS_SYN_SENT) ||
5679 	    (tp->t_state == TCPS_SYN_RECEIVED))
5680 		rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift]);
5681 	else
5682 		rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5683 	TCPT_RANGESET(tp->t_rxtcur, rexmt,
5684 	   max(MSEC_2_TICKS(rack_rto_min), rexmt),
5685 	   MSEC_2_TICKS(rack_rto_max));
5686 	/*
5687 	 * We enter the path for PLMTUD if connection is established or, if
5688 	 * connection is FIN_WAIT_1 status, reason for the last is that if
5689 	 * amount of data we send is very small, we could send it in couple
5690 	 * of packets and process straight to FIN. In that case we won't
5691 	 * catch ESTABLISHED state.
5692 	 */
5693 #ifdef INET6
5694 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
5695 #else
5696 	isipv6 = false;
5697 #endif
5698 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
5699 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
5700 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
5701 	    ((tp->t_state == TCPS_ESTABLISHED) ||
5702 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
5703 
5704 		/*
5705 		 * Idea here is that at each stage of mtu probe (usually,
5706 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
5707 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5708 		 * should take care of that.
5709 		 */
5710 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5711 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5712 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5713 		    tp->t_rxtshift % 2 == 0)) {
5714 			/*
5715 			 * Enter Path MTU Black-hole Detection mechanism: -
5716 			 * Disable Path MTU Discovery (IP "DF" bit). -
5717 			 * Reduce MTU to lower value than what we negotiated
5718 			 * with peer.
5719 			 */
5720 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5721 				/* Record that we may have found a black hole. */
5722 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5723 				/* Keep track of previous MSS. */
5724 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5725 			}
5726 
5727 			/*
5728 			 * Reduce the MSS to blackhole value or to the
5729 			 * default in an attempt to retransmit.
5730 			 */
5731 #ifdef INET6
5732 			if (isipv6 &&
5733 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5734 				/* Use the sysctl tuneable blackhole MSS. */
5735 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5736 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5737 			} else if (isipv6) {
5738 				/* Use the default MSS. */
5739 				tp->t_maxseg = V_tcp_v6mssdflt;
5740 				/*
5741 				 * Disable Path MTU Discovery when we switch
5742 				 * to minmss.
5743 				 */
5744 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5745 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5746 			}
5747 #endif
5748 #if defined(INET6) && defined(INET)
5749 			else
5750 #endif
5751 #ifdef INET
5752 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5753 				/* Use the sysctl tuneable blackhole MSS. */
5754 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5755 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5756 			} else {
5757 				/* Use the default MSS. */
5758 				tp->t_maxseg = V_tcp_mssdflt;
5759 				/*
5760 				 * Disable Path MTU Discovery when we switch
5761 				 * to minmss.
5762 				 */
5763 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5764 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5765 			}
5766 #endif
5767 		} else {
5768 			/*
5769 			 * If further retransmissions are still unsuccessful
5770 			 * with a lowered MTU, maybe this isn't a blackhole
5771 			 * and we restore the previous MSS and blackhole
5772 			 * detection flags. The limit '6' is determined by
5773 			 * giving each probe stage (1448, 1188, 524) 2
5774 			 * chances to recover.
5775 			 */
5776 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5777 			    (tp->t_rxtshift >= 6)) {
5778 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5779 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5780 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5781 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5782 			}
5783 		}
5784 	}
5785 	/*
5786 	 * If we backed off this far, our srtt estimate is probably bogus.
5787 	 * Clobber it so we'll take the next rtt measurement as our srtt;
5788 	 * move the current srtt into rttvar to keep the current retransmit
5789 	 * times until then.
5790 	 */
5791 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5792 #ifdef INET6
5793 		if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
5794 			in6_losing(tp->t_inpcb);
5795 		else
5796 #endif
5797 			in_losing(tp->t_inpcb);
5798 		tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5799 		tp->t_srtt = 0;
5800 	}
5801 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
5802 	tp->snd_recover = tp->snd_max;
5803 	tp->t_flags |= TF_ACKNOW;
5804 	tp->t_rtttime = 0;
5805 	rack_cong_signal(tp, NULL, CC_RTO);
5806 out:
5807 	return (retval);
5808 }
5809 
5810 static int
5811 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling)
5812 {
5813 	int32_t ret = 0;
5814 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5815 
5816 	if (timers == 0) {
5817 		return (0);
5818 	}
5819 	if (tp->t_state == TCPS_LISTEN) {
5820 		/* no timers on listen sockets */
5821 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5822 			return (0);
5823 		return (1);
5824 	}
5825 	if ((timers & PACE_TMR_RACK) &&
5826 	    rack->rc_on_min_to) {
5827 		/*
5828 		 * For the rack timer when we
5829 		 * are on a min-timeout (which means rrr_conf = 3)
5830 		 * we don't want to check the timer. It may
5831 		 * be going off for a pace and thats ok we
5832 		 * want to send the retransmit (if its ready).
5833 		 *
5834 		 * If its on a normal rack timer (non-min) then
5835 		 * we will check if its expired.
5836 		 */
5837 		goto skip_time_check;
5838 	}
5839 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
5840 		uint32_t left;
5841 
5842 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5843 			ret = -1;
5844 			rack_log_to_processing(rack, cts, ret, 0);
5845 			return (0);
5846 		}
5847 		if (hpts_calling == 0) {
5848 			/*
5849 			 * A user send or queued mbuf (sack) has called us? We
5850 			 * return 0 and let the pacing guards
5851 			 * deal with it if they should or
5852 			 * should not cause a send.
5853 			 */
5854 			ret = -2;
5855 			rack_log_to_processing(rack, cts, ret, 0);
5856 			return (0);
5857 		}
5858 		/*
5859 		 * Ok our timer went off early and we are not paced false
5860 		 * alarm, go back to sleep.
5861 		 */
5862 		ret = -3;
5863 		left = rack->r_ctl.rc_timer_exp - cts;
5864 		tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
5865 		rack_log_to_processing(rack, cts, ret, left);
5866 		return (1);
5867 	}
5868 skip_time_check:
5869 	rack->rc_tmr_stopped = 0;
5870 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5871 	if (timers & PACE_TMR_DELACK) {
5872 		ret = rack_timeout_delack(tp, rack, cts);
5873 	} else if (timers & PACE_TMR_RACK) {
5874 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
5875 		ret = rack_timeout_rack(tp, rack, cts);
5876 	} else if (timers & PACE_TMR_TLP) {
5877 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
5878 		ret = rack_timeout_tlp(tp, rack, cts);
5879 	} else if (timers & PACE_TMR_RXT) {
5880 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
5881 		ret = rack_timeout_rxt(tp, rack, cts);
5882 	} else if (timers & PACE_TMR_PERSIT) {
5883 		ret = rack_timeout_persist(tp, rack, cts);
5884 	} else if (timers & PACE_TMR_KEEP) {
5885 		ret = rack_timeout_keepalive(tp, rack, cts);
5886 	}
5887 	rack_log_to_processing(rack, cts, ret, timers);
5888 	return (ret);
5889 }
5890 
5891 static void
5892 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
5893 {
5894 	struct timeval tv;
5895 	uint32_t us_cts, flags_on_entry;
5896 	uint8_t hpts_removed = 0;
5897 
5898 
5899 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
5900 	us_cts = tcp_get_usecs(&tv);
5901 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
5902 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
5903 	     ((tp->snd_max - tp->snd_una) == 0))) {
5904 		tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5905 		hpts_removed = 1;
5906 		/* If we were not delayed cancel out the flag. */
5907 		if ((tp->snd_max - tp->snd_una) == 0)
5908 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5909 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
5910 	}
5911 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5912 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5913 		if (rack->rc_inp->inp_in_hpts &&
5914 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
5915 			/*
5916 			 * Canceling timer's when we have no output being
5917 			 * paced. We also must remove ourselves from the
5918 			 * hpts.
5919 			 */
5920 			tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5921 			hpts_removed = 1;
5922 		}
5923 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5924 	}
5925 	if (hpts_removed == 0)
5926 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
5927 }
5928 
5929 static void
5930 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
5931 {
5932 	return;
5933 }
5934 
5935 static int
5936 rack_stopall(struct tcpcb *tp)
5937 {
5938 	struct tcp_rack *rack;
5939 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5940 	rack->t_timers_stopped = 1;
5941 	return (0);
5942 }
5943 
5944 static void
5945 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
5946 {
5947 	return;
5948 }
5949 
5950 static int
5951 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
5952 {
5953 	return (0);
5954 }
5955 
5956 static void
5957 rack_stop_all_timers(struct tcpcb *tp)
5958 {
5959 	struct tcp_rack *rack;
5960 
5961 	/*
5962 	 * Assure no timers are running.
5963 	 */
5964 	if (tcp_timer_active(tp, TT_PERSIST)) {
5965 		/* We enter in persists, set the flag appropriately */
5966 		rack = (struct tcp_rack *)tp->t_fb_ptr;
5967 		rack->rc_in_persist = 1;
5968 	}
5969 	tcp_timer_suspend(tp, TT_PERSIST);
5970 	tcp_timer_suspend(tp, TT_REXMT);
5971 	tcp_timer_suspend(tp, TT_KEEP);
5972 	tcp_timer_suspend(tp, TT_DELACK);
5973 }
5974 
5975 static void
5976 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
5977     struct rack_sendmap *rsm, uint32_t ts)
5978 {
5979 	int32_t idx;
5980 
5981 	rsm->r_rtr_cnt++;
5982 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
5983 	rsm->r_dupack = 0;
5984 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
5985 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
5986 		rsm->r_flags |= RACK_OVERMAX;
5987 	}
5988 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
5989 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5990 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5991 	}
5992 	idx = rsm->r_rtr_cnt - 1;
5993 	rsm->r_tim_lastsent[idx] = ts;
5994 	if (rsm->r_flags & RACK_ACKED) {
5995 		/* Problably MTU discovery messing with us */
5996 		rsm->r_flags &= ~RACK_ACKED;
5997 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5998 	}
5999 	if (rsm->r_in_tmap) {
6000 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6001 		rsm->r_in_tmap = 0;
6002 	}
6003 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6004 	rsm->r_in_tmap = 1;
6005 	if (rsm->r_flags & RACK_SACK_PASSED) {
6006 		/* We have retransmitted due to the SACK pass */
6007 		rsm->r_flags &= ~RACK_SACK_PASSED;
6008 		rsm->r_flags |= RACK_WAS_SACKPASS;
6009 	}
6010 }
6011 
6012 
6013 static uint32_t
6014 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
6015     struct rack_sendmap *rsm, uint32_t ts, int32_t *lenp)
6016 {
6017 	/*
6018 	 * We (re-)transmitted starting at rsm->r_start for some length
6019 	 * (possibly less than r_end.
6020 	 */
6021 	struct rack_sendmap *nrsm, *insret;
6022 	uint32_t c_end;
6023 	int32_t len;
6024 
6025 	len = *lenp;
6026 	c_end = rsm->r_start + len;
6027 	if (SEQ_GEQ(c_end, rsm->r_end)) {
6028 		/*
6029 		 * We retransmitted the whole piece or more than the whole
6030 		 * slopping into the next rsm.
6031 		 */
6032 		rack_update_rsm(tp, rack, rsm, ts);
6033 		if (c_end == rsm->r_end) {
6034 			*lenp = 0;
6035 			return (0);
6036 		} else {
6037 			int32_t act_len;
6038 
6039 			/* Hangs over the end return whats left */
6040 			act_len = rsm->r_end - rsm->r_start;
6041 			*lenp = (len - act_len);
6042 			return (rsm->r_end);
6043 		}
6044 		/* We don't get out of this block. */
6045 	}
6046 	/*
6047 	 * Here we retransmitted less than the whole thing which means we
6048 	 * have to split this into what was transmitted and what was not.
6049 	 */
6050 	nrsm = rack_alloc_full_limit(rack);
6051 	if (nrsm == NULL) {
6052 		/*
6053 		 * We can't get memory, so lets not proceed.
6054 		 */
6055 		*lenp = 0;
6056 		return (0);
6057 	}
6058 	/*
6059 	 * So here we are going to take the original rsm and make it what we
6060 	 * retransmitted. nrsm will be the tail portion we did not
6061 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
6062 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
6063 	 * 1, 6 and the new piece will be 6, 11.
6064 	 */
6065 	rack_clone_rsm(rack, nrsm, rsm, c_end);
6066 	nrsm->r_dupack = 0;
6067 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
6068 	insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6069 #ifdef INVARIANTS
6070 	if (insret != NULL) {
6071 		panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6072 		      nrsm, insret, rack, rsm);
6073 	}
6074 #endif
6075 	if (rsm->r_in_tmap) {
6076 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6077 		nrsm->r_in_tmap = 1;
6078 	}
6079 	rsm->r_flags &= (~RACK_HAS_FIN);
6080 	rack_update_rsm(tp, rack, rsm, ts);
6081 	*lenp = 0;
6082 	return (0);
6083 }
6084 
6085 
6086 static void
6087 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
6088     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts,
6089     uint8_t pass, struct rack_sendmap *hintrsm, uint32_t us_cts)
6090 {
6091 	struct tcp_rack *rack;
6092 	struct rack_sendmap *rsm, *nrsm, *insret, fe;
6093 	register uint32_t snd_max, snd_una;
6094 
6095 	/*
6096 	 * Add to the RACK log of packets in flight or retransmitted. If
6097 	 * there is a TS option we will use the TS echoed, if not we will
6098 	 * grab a TS.
6099 	 *
6100 	 * Retransmissions will increment the count and move the ts to its
6101 	 * proper place. Note that if options do not include TS's then we
6102 	 * won't be able to effectively use the ACK for an RTT on a retran.
6103 	 *
6104 	 * Notes about r_start and r_end. Lets consider a send starting at
6105 	 * sequence 1 for 10 bytes. In such an example the r_start would be
6106 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
6107 	 * This means that r_end is actually the first sequence for the next
6108 	 * slot (11).
6109 	 *
6110 	 */
6111 	/*
6112 	 * If err is set what do we do XXXrrs? should we not add the thing?
6113 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
6114 	 * i.e. proceed with add ** do this for now.
6115 	 */
6116 	INP_WLOCK_ASSERT(tp->t_inpcb);
6117 	if (err)
6118 		/*
6119 		 * We don't log errors -- we could but snd_max does not
6120 		 * advance in this case either.
6121 		 */
6122 		return;
6123 
6124 	if (th_flags & TH_RST) {
6125 		/*
6126 		 * We don't log resets and we return immediately from
6127 		 * sending
6128 		 */
6129 		return;
6130 	}
6131 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6132 	snd_una = tp->snd_una;
6133 	if (SEQ_LEQ((seq_out + len), snd_una)) {
6134 		/* Are sending an old segment to induce an ack (keep-alive)? */
6135 		return;
6136 	}
6137 	if (SEQ_LT(seq_out, snd_una)) {
6138 		/* huh? should we panic? */
6139 		uint32_t end;
6140 
6141 		end = seq_out + len;
6142 		seq_out = snd_una;
6143 		if (SEQ_GEQ(end, seq_out))
6144 			len = end - seq_out;
6145 		else
6146 			len = 0;
6147 	}
6148 	snd_max = tp->snd_max;
6149 	if (th_flags & (TH_SYN | TH_FIN)) {
6150 		/*
6151 		 * The call to rack_log_output is made before bumping
6152 		 * snd_max. This means we can record one extra byte on a SYN
6153 		 * or FIN if seq_out is adding more on and a FIN is present
6154 		 * (and we are not resending).
6155 		 */
6156 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
6157 			len++;
6158 		if (th_flags & TH_FIN)
6159 			len++;
6160 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
6161 			/*
6162 			 * The add/update as not been done for the FIN/SYN
6163 			 * yet.
6164 			 */
6165 			snd_max = tp->snd_nxt;
6166 		}
6167 	}
6168 	if (len == 0) {
6169 		/* We don't log zero window probes */
6170 		return;
6171 	}
6172 	rack->r_ctl.rc_time_last_sent = ts;
6173 	if (IN_RECOVERY(tp->t_flags)) {
6174 		rack->r_ctl.rc_prr_out += len;
6175 	}
6176 	/* First question is it a retransmission or new? */
6177 	if (seq_out == snd_max) {
6178 		/* Its new */
6179 again:
6180 		rsm = rack_alloc(rack);
6181 		if (rsm == NULL) {
6182 			/*
6183 			 * Hmm out of memory and the tcb got destroyed while
6184 			 * we tried to wait.
6185 			 */
6186 			return;
6187 		}
6188 		if (th_flags & TH_FIN) {
6189 			rsm->r_flags = RACK_HAS_FIN;
6190 		} else {
6191 			rsm->r_flags = 0;
6192 		}
6193 		rsm->r_tim_lastsent[0] = ts;
6194 		rsm->r_rtr_cnt = 1;
6195 		rsm->r_rtr_bytes = 0;
6196 		rsm->usec_orig_send = us_cts;
6197 		if (th_flags & TH_SYN) {
6198 			/* The data space is one beyond snd_una */
6199 			rsm->r_flags |= RACK_HAS_SIN;
6200 			rsm->r_start = seq_out + 1;
6201 			rsm->r_end = rsm->r_start + (len - 1);
6202 		} else {
6203 			/* Normal case */
6204 			rsm->r_start = seq_out;
6205 			rsm->r_end = rsm->r_start + len;
6206 		}
6207 		rsm->r_dupack = 0;
6208 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6209 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
6210 #ifdef INVARIANTS
6211 		if (insret != NULL) {
6212 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6213 			      nrsm, insret, rack, rsm);
6214 		}
6215 #endif
6216 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6217 		rsm->r_in_tmap = 1;
6218 		/*
6219 		 * Special case detection, is there just a single
6220 		 * packet outstanding when we are not in recovery?
6221 		 *
6222 		 * If this is true mark it so.
6223 		 */
6224 		if ((IN_RECOVERY(tp->t_flags) == 0) &&
6225 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
6226 			struct rack_sendmap *prsm;
6227 
6228 			prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
6229 			if (prsm)
6230 				prsm->r_one_out_nr = 1;
6231 		}
6232 		return;
6233 	}
6234 	/*
6235 	 * If we reach here its a retransmission and we need to find it.
6236 	 */
6237 	memset(&fe, 0, sizeof(fe));
6238 more:
6239 	if (hintrsm && (hintrsm->r_start == seq_out)) {
6240 		rsm = hintrsm;
6241 		hintrsm = NULL;
6242 	} else {
6243 		/* No hints sorry */
6244 		rsm = NULL;
6245 	}
6246 	if ((rsm) && (rsm->r_start == seq_out)) {
6247 		seq_out = rack_update_entry(tp, rack, rsm, ts, &len);
6248 		if (len == 0) {
6249 			return;
6250 		} else {
6251 			goto more;
6252 		}
6253 	}
6254 	/* Ok it was not the last pointer go through it the hard way. */
6255 refind:
6256 	fe.r_start = seq_out;
6257 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
6258 	if (rsm) {
6259 		if (rsm->r_start == seq_out) {
6260 			seq_out = rack_update_entry(tp, rack, rsm, ts, &len);
6261 			if (len == 0) {
6262 				return;
6263 			} else {
6264 				goto refind;
6265 			}
6266 		}
6267 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6268 			/* Transmitted within this piece */
6269 			/*
6270 			 * Ok we must split off the front and then let the
6271 			 * update do the rest
6272 			 */
6273 			nrsm = rack_alloc_full_limit(rack);
6274 			if (nrsm == NULL) {
6275 				rack_update_rsm(tp, rack, rsm, ts);
6276 				return;
6277 			}
6278 			/*
6279 			 * copy rsm to nrsm and then trim the front of rsm
6280 			 * to not include this part.
6281 			 */
6282 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
6283 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6284 #ifdef INVARIANTS
6285 			if (insret != NULL) {
6286 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6287 				      nrsm, insret, rack, rsm);
6288 			}
6289 #endif
6290 			if (rsm->r_in_tmap) {
6291 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6292 				nrsm->r_in_tmap = 1;
6293 			}
6294 			rsm->r_flags &= (~RACK_HAS_FIN);
6295 			seq_out = rack_update_entry(tp, rack, nrsm, ts, &len);
6296 			if (len == 0) {
6297 				return;
6298 			} else if (len > 0)
6299 				goto refind;
6300 		}
6301 	}
6302 	/*
6303 	 * Hmm not found in map did they retransmit both old and on into the
6304 	 * new?
6305 	 */
6306 	if (seq_out == tp->snd_max) {
6307 		goto again;
6308 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
6309 #ifdef INVARIANTS
6310 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6311 		    seq_out, len, tp->snd_una, tp->snd_max);
6312 		printf("Starting Dump of all rack entries\n");
6313 		RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6314 			printf("rsm:%p start:%u end:%u\n",
6315 			    rsm, rsm->r_start, rsm->r_end);
6316 		}
6317 		printf("Dump complete\n");
6318 		panic("seq_out not found rack:%p tp:%p",
6319 		    rack, tp);
6320 #endif
6321 	} else {
6322 #ifdef INVARIANTS
6323 		/*
6324 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
6325 		 * flag)
6326 		 */
6327 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6328 		    seq_out, len, tp->snd_max, tp);
6329 #endif
6330 	}
6331 }
6332 
6333 /*
6334  * Record one of the RTT updates from an ack into
6335  * our sample structure.
6336  */
6337 
6338 static void
6339 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
6340 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
6341 {
6342 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
6343 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
6344 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
6345 	}
6346 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
6347 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
6348 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
6349 	}
6350 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
6351 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
6352 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
6353 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
6354 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
6355 	}
6356 	if ((confidence == 1) &&
6357 	    ((rsm == NULL) ||
6358 	     (rsm->r_just_ret) ||
6359 	     (rsm->r_one_out_nr &&
6360 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
6361 		/*
6362 		 * If the rsm had a just return
6363 		 * hit it then we can't trust the
6364 		 * rtt measurement for buffer deterimination
6365 		 * Note that a confidence of 2, indicates
6366 		 * SACK'd which overrides the r_just_ret or
6367 		 * the r_one_out_nr. If it was a CUM-ACK and
6368 		 * we had only two outstanding, but get an
6369 		 * ack for only 1. Then that also lowers our
6370 		 * confidence.
6371 		 */
6372 		confidence = 0;
6373 	}
6374 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
6375 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
6376 		if (rack->r_ctl.rack_rs.confidence == 0) {
6377 			/*
6378 			 * We take anything with no current confidence
6379 			 * saved.
6380 			 */
6381 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
6382 			rack->r_ctl.rack_rs.confidence = confidence;
6383 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
6384 		} else if (confidence || rack->r_ctl.rack_rs.confidence) {
6385 			/*
6386 			 * Once we have a confident number,
6387 			 * we can update it with a smaller
6388 			 * value since this confident number
6389 			 * may include the DSACK time until
6390 			 * the next segment (the second one) arrived.
6391 			 */
6392 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
6393 			rack->r_ctl.rack_rs.confidence = confidence;
6394 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
6395 		}
6396 
6397 	}
6398 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
6399 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
6400 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
6401 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
6402 }
6403 
6404 /*
6405  * Collect new round-trip time estimate
6406  * and update averages and current timeout.
6407  */
6408 static void
6409 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
6410 {
6411 	int32_t delta;
6412 	uint32_t o_srtt, o_var;
6413 	int32_t hrtt_up = 0;
6414 	int32_t rtt;
6415 
6416 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
6417 		/* No valid sample */
6418 		return;
6419 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
6420 		/* We are to use the lowest RTT seen in a single ack */
6421 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
6422 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
6423 		/* We are to use the highest RTT seen in a single ack */
6424 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
6425 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
6426 		/* We are to use the average RTT seen in a single ack */
6427 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
6428 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
6429 	} else {
6430 #ifdef INVARIANTS
6431 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
6432 #endif
6433 		return;
6434 	}
6435 	if (rtt == 0)
6436 		rtt = 1;
6437 	if (rack->rc_gp_rtt_set == 0) {
6438 		/*
6439 		 * With no RTT we have to accept
6440 		 * even one we are not confident of.
6441 		 */
6442 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
6443 		rack->rc_gp_rtt_set = 1;
6444 	} else if (rack->r_ctl.rack_rs.confidence) {
6445 		/* update the running gp srtt */
6446 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
6447 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
6448 	}
6449 	if (rack->r_ctl.rack_rs.confidence) {
6450 		/*
6451 		 * record the low and high for highly buffered path computation,
6452 		 * we only do this if we are confident (not a retransmission).
6453 		 */
6454 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
6455 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
6456 			hrtt_up = 1;
6457 		}
6458 		if (rack->rc_highly_buffered == 0) {
6459 			/*
6460 			 * Currently once we declare a path has
6461 			 * highly buffered there is no going
6462 			 * back, which may be a problem...
6463 			 */
6464 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
6465 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
6466 						     rack->r_ctl.rc_highest_us_rtt,
6467 						     rack->r_ctl.rc_lowest_us_rtt,
6468 						     RACK_RTTS_SEEHBP);
6469 				rack->rc_highly_buffered = 1;
6470 			}
6471 		}
6472 	}
6473 	if ((rack->r_ctl.rack_rs.confidence) ||
6474 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
6475 		/*
6476 		 * If we are highly confident of it <or> it was
6477 		 * never retransmitted we accept it as the last us_rtt.
6478 		 */
6479 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
6480 		/* The lowest rtt can be set if its was not retransmited */
6481 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
6482 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
6483 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
6484 				rack->r_ctl.rc_lowest_us_rtt = 1;
6485 		}
6486 	}
6487 	rack_log_rtt_sample(rack, rtt);
6488 	o_srtt = tp->t_srtt;
6489 	o_var = tp->t_rttvar;
6490 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6491 	if (tp->t_srtt != 0) {
6492 		/*
6493 		 * srtt is stored as fixed point with 5 bits after the
6494 		 * binary point (i.e., scaled by 8).  The following magic is
6495 		 * equivalent to the smoothing algorithm in rfc793 with an
6496 		 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6497 		 * Adjust rtt to origin 0.
6498 		 */
6499 		delta = ((rtt - 1) << TCP_DELTA_SHIFT)
6500 		    - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6501 
6502 		tp->t_srtt += delta;
6503 		if (tp->t_srtt <= 0)
6504 			tp->t_srtt = 1;
6505 
6506 		/*
6507 		 * We accumulate a smoothed rtt variance (actually, a
6508 		 * smoothed mean difference), then set the retransmit timer
6509 		 * to smoothed rtt + 4 times the smoothed variance. rttvar
6510 		 * is stored as fixed point with 4 bits after the binary
6511 		 * point (scaled by 16).  The following is equivalent to
6512 		 * rfc793 smoothing with an alpha of .75 (rttvar =
6513 		 * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
6514 		 * wired-in beta.
6515 		 */
6516 		if (delta < 0)
6517 			delta = -delta;
6518 		delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6519 		tp->t_rttvar += delta;
6520 		if (tp->t_rttvar <= 0)
6521 			tp->t_rttvar = 1;
6522 		if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
6523 			tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6524 	} else {
6525 		/*
6526 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
6527 		 * variance to half the rtt (so our first retransmit happens
6528 		 * at 3*rtt).
6529 		 */
6530 		tp->t_srtt = rtt << TCP_RTT_SHIFT;
6531 		tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
6532 		tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6533 	}
6534 	KMOD_TCPSTAT_INC(tcps_rttupdated);
6535 	tp->t_rttupdated++;
6536 #ifdef STATS
6537 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
6538 #endif
6539 	tp->t_rxtshift = 0;
6540 
6541 	/*
6542 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
6543 	 * way we do the smoothing, srtt and rttvar will each average +1/2
6544 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
6545 	 * tick of rounding and 1 extra tick because of +-1/2 tick
6546 	 * uncertainty in the firing of the timer.  The bias will give us
6547 	 * exactly the 1.5 tick we need.  But, because the bias is
6548 	 * statistical, we have to test that we don't drop below the minimum
6549 	 * feasible timer (which is 2 ticks).
6550 	 */
6551 	TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6552 	   max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max));
6553 	tp->t_softerror = 0;
6554 }
6555 
6556 static void
6557 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm,
6558     uint32_t t, uint32_t cts)
6559 {
6560 	/*
6561 	 * For this RSM, we acknowledged the data from a previous
6562 	 * transmission, not the last one we made. This means we did a false
6563 	 * retransmit.
6564 	 */
6565 	struct tcp_rack *rack;
6566 
6567 	if (rsm->r_flags & RACK_HAS_FIN) {
6568 		/*
6569 		 * The sending of the FIN often is multiple sent when we
6570 		 * have everything outstanding ack'd. We ignore this case
6571 		 * since its over now.
6572 		 */
6573 		return;
6574 	}
6575 	if (rsm->r_flags & RACK_TLP) {
6576 		/*
6577 		 * We expect TLP's to have this occur.
6578 		 */
6579 		return;
6580 	}
6581 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6582 	/* should we undo cc changes and exit recovery? */
6583 	if (IN_RECOVERY(tp->t_flags)) {
6584 		if (rack->r_ctl.rc_rsm_start == rsm->r_start) {
6585 			/*
6586 			 * Undo what we ratched down and exit recovery if
6587 			 * possible
6588 			 */
6589 			EXIT_RECOVERY(tp->t_flags);
6590 			tp->snd_recover = tp->snd_una;
6591 			if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd)
6592 				tp->snd_cwnd = rack->r_ctl.rc_cwnd_at;
6593 			if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh)
6594 				tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at;
6595 		}
6596 	}
6597 	if (rsm->r_flags & RACK_WAS_SACKPASS) {
6598 		/*
6599 		 * We retransmitted based on a sack and the earlier
6600 		 * retransmission ack'd it - re-ordering is occuring.
6601 		 */
6602 		counter_u64_add(rack_reorder_seen, 1);
6603 		rack->r_ctl.rc_reorder_ts = cts;
6604 	}
6605 	counter_u64_add(rack_badfr, 1);
6606 	counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start));
6607 }
6608 
6609 static void
6610 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
6611 {
6612 	/*
6613 	 * Apply to filter the inbound us-rtt at us_cts.
6614 	 */
6615 	uint32_t old_rtt;
6616 
6617 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
6618 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
6619 			       us_rtt, us_cts);
6620 	if (rack->r_ctl.last_pacing_time &&
6621 	    rack->rc_gp_dyn_mul &&
6622 	    (rack->r_ctl.last_pacing_time > us_rtt))
6623 		rack->pacing_longer_than_rtt = 1;
6624 	else
6625 		rack->pacing_longer_than_rtt = 0;
6626 	if (old_rtt > us_rtt) {
6627 		/* We just hit a new lower rtt time */
6628 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
6629 				     __LINE__, RACK_RTTS_NEWRTT);
6630 		/*
6631 		 * Only count it if its lower than what we saw within our
6632 		 * calculated range.
6633 		 */
6634 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
6635 			if (rack_probertt_lower_within &&
6636 			    rack->rc_gp_dyn_mul &&
6637 			    (rack->use_fixed_rate == 0) &&
6638 			    (rack->rc_always_pace)) {
6639 				/*
6640 				 * We are seeing a new lower rtt very close
6641 				 * to the time that we would have entered probe-rtt.
6642 				 * This is probably due to the fact that a peer flow
6643 				 * has entered probe-rtt. Lets go in now too.
6644 				 */
6645 				uint32_t val;
6646 
6647 				val = rack_probertt_lower_within * rack_time_between_probertt;
6648 				val /= 100;
6649 				if ((rack->in_probe_rtt == 0)  &&
6650 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
6651 					rack_enter_probertt(rack, us_cts);
6652 				}
6653 			}
6654 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6655 		}
6656 	}
6657 }
6658 
6659 static int
6660 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
6661     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
6662 {
6663 	int32_t i;
6664 	uint32_t t, len_acked;
6665 
6666 	if ((rsm->r_flags & RACK_ACKED) ||
6667 	    (rsm->r_flags & RACK_WAS_ACKED))
6668 		/* Already done */
6669 		return (0);
6670 
6671 	if (ack_type == CUM_ACKED) {
6672 		if (SEQ_GT(th_ack, rsm->r_end))
6673 			len_acked = rsm->r_end - rsm->r_start;
6674 		else
6675 			len_acked = th_ack - rsm->r_start;
6676 	} else
6677 		len_acked = rsm->r_end - rsm->r_start;
6678 	if (rsm->r_rtr_cnt == 1) {
6679 		uint32_t us_rtt;
6680 
6681 		t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
6682 		if ((int)t <= 0)
6683 			t = 1;
6684 		if (!tp->t_rttlow || tp->t_rttlow > t)
6685 			tp->t_rttlow = t;
6686 		if (!rack->r_ctl.rc_rack_min_rtt ||
6687 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
6688 			rack->r_ctl.rc_rack_min_rtt = t;
6689 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
6690 				rack->r_ctl.rc_rack_min_rtt = 1;
6691 			}
6692 		}
6693 		us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - rsm->usec_orig_send;
6694 		if (us_rtt == 0)
6695 			us_rtt = 1;
6696 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
6697 		if (ack_type == SACKED)
6698 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
6699 		else {
6700 			/*
6701 			 * For cum-ack we are only confident if what
6702 			 * is being acked is included in a measurement.
6703 			 * Otherwise it could be an idle period that
6704 			 * includes Delayed-ack time.
6705 			 */
6706 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
6707 					    (rack->app_limited_needs_set ? 0 : 1), rsm, rsm->r_rtr_cnt);
6708 		}
6709 		if ((rsm->r_flags & RACK_TLP) &&
6710 		    (!IN_RECOVERY(tp->t_flags))) {
6711 			/* Segment was a TLP and our retrans matched */
6712 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
6713 				rack->r_ctl.rc_rsm_start = tp->snd_max;
6714 				rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
6715 				rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
6716 				rack_cong_signal(tp, NULL, CC_NDUPACK);
6717 				/*
6718 				 * When we enter recovery we need to assure
6719 				 * we send one packet.
6720 				 */
6721 				if (rack->rack_no_prr == 0) {
6722 					rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
6723 					rack_log_to_prr(rack, 7, 0);
6724 				}
6725 			}
6726 		}
6727 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
6728 			/* New more recent rack_tmit_time */
6729 			rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
6730 			rack->rc_rack_rtt = t;
6731 		}
6732 		return (1);
6733 	}
6734 	/*
6735 	 * We clear the soft/rxtshift since we got an ack.
6736 	 * There is no assurance we will call the commit() function
6737 	 * so we need to clear these to avoid incorrect handling.
6738 	 */
6739 	tp->t_rxtshift = 0;
6740 	tp->t_softerror = 0;
6741 	if ((to->to_flags & TOF_TS) &&
6742 	    (ack_type == CUM_ACKED) &&
6743 	    (to->to_tsecr) &&
6744 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
6745 		/*
6746 		 * Now which timestamp does it match? In this block the ACK
6747 		 * must be coming from a previous transmission.
6748 		 */
6749 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
6750 			if (rsm->r_tim_lastsent[i] == to->to_tsecr) {
6751 				t = cts - rsm->r_tim_lastsent[i];
6752 				if ((int)t <= 0)
6753 					t = 1;
6754 				if ((i + 1) < rsm->r_rtr_cnt) {
6755 					/* Likely */
6756 					rack_earlier_retran(tp, rsm, t, cts);
6757 				}
6758 				if (!tp->t_rttlow || tp->t_rttlow > t)
6759 					tp->t_rttlow = t;
6760 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
6761 					rack->r_ctl.rc_rack_min_rtt = t;
6762 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
6763 						rack->r_ctl.rc_rack_min_rtt = 1;
6764 					}
6765 				}
6766 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
6767 				    rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
6768 					/* New more recent rack_tmit_time */
6769 					rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
6770 					rack->rc_rack_rtt = t;
6771 				}
6772 				tcp_rack_xmit_timer(rack, t + 1, len_acked, (t * HPTS_USEC_IN_MSEC), 0, rsm,
6773 						    rsm->r_rtr_cnt);
6774 				return (1);
6775 			}
6776 		}
6777 		goto ts_not_found;
6778 	} else {
6779 		/*
6780 		 * Ok its a SACK block that we retransmitted. or a windows
6781 		 * machine without timestamps. We can tell nothing from the
6782 		 * time-stamp since its not there or the time the peer last
6783 		 * recieved a segment that moved forward its cum-ack point.
6784 		 */
6785 ts_not_found:
6786 		i = rsm->r_rtr_cnt - 1;
6787 		t = cts - rsm->r_tim_lastsent[i];
6788 		if ((int)t <= 0)
6789 			t = 1;
6790 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
6791 			/*
6792 			 * We retransmitted and the ack came back in less
6793 			 * than the smallest rtt we have observed. We most
6794 			 * likey did an improper retransmit as outlined in
6795 			 * 4.2 Step 3 point 2 in the rack-draft.
6796 			 */
6797 			i = rsm->r_rtr_cnt - 2;
6798 			t = cts - rsm->r_tim_lastsent[i];
6799 			rack_earlier_retran(tp, rsm, t, cts);
6800 		} else if (rack->r_ctl.rc_rack_min_rtt) {
6801 			/*
6802 			 * We retransmitted it and the retransmit did the
6803 			 * job.
6804 			 */
6805 			if (!rack->r_ctl.rc_rack_min_rtt ||
6806 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
6807 				rack->r_ctl.rc_rack_min_rtt = t;
6808 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
6809 					rack->r_ctl.rc_rack_min_rtt = 1;
6810 				}
6811 			}
6812 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) {
6813 				/* New more recent rack_tmit_time */
6814 				rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i];
6815 				rack->rc_rack_rtt = t;
6816 			}
6817 			return (1);
6818 		}
6819 	}
6820 	return (0);
6821 }
6822 
6823 /*
6824  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
6825  */
6826 static void
6827 rack_log_sack_passed(struct tcpcb *tp,
6828     struct tcp_rack *rack, struct rack_sendmap *rsm)
6829 {
6830 	struct rack_sendmap *nrsm;
6831 
6832 	nrsm = rsm;
6833 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
6834 	    rack_head, r_tnext) {
6835 		if (nrsm == rsm) {
6836 			/* Skip orginal segment he is acked */
6837 			continue;
6838 		}
6839 		if (nrsm->r_flags & RACK_ACKED) {
6840 			/*
6841 			 * Skip ack'd segments, though we
6842 			 * should not see these, since tmap
6843 			 * should not have ack'd segments.
6844 			 */
6845 			continue;
6846 		}
6847 		if (nrsm->r_flags & RACK_SACK_PASSED) {
6848 			/*
6849 			 * We found one that is already marked
6850 			 * passed, we have been here before and
6851 			 * so all others below this are marked.
6852 			 */
6853 			break;
6854 		}
6855 		nrsm->r_flags |= RACK_SACK_PASSED;
6856 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
6857 	}
6858 }
6859 
6860 static void
6861 rack_need_set_test(struct tcpcb *tp,
6862 		   struct tcp_rack *rack,
6863 		   struct rack_sendmap *rsm,
6864 		   tcp_seq th_ack,
6865 		   int line,
6866 		   int use_which)
6867 {
6868 
6869 	if ((tp->t_flags & TF_GPUTINPROG) &&
6870 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
6871 		/*
6872 		 * We were app limited, and this ack
6873 		 * butts up or goes beyond the point where we want
6874 		 * to start our next measurement. We need
6875 		 * to record the new gput_ts as here and
6876 		 * possibly update the start sequence.
6877 		 */
6878 		uint32_t seq, ts;
6879 
6880 		if (rsm->r_rtr_cnt > 1) {
6881 			/*
6882 			 * This is a retransmit, can we
6883 			 * really make any assessment at this
6884 			 * point?  We are not really sure of
6885 			 * the timestamp, is it this or the
6886 			 * previous transmission?
6887 			 *
6888 			 * Lets wait for something better that
6889 			 * is not retransmitted.
6890 			 */
6891 			return;
6892 		}
6893 		seq = tp->gput_seq;
6894 		ts = tp->gput_ts;
6895 		rack->app_limited_needs_set = 0;
6896 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
6897 		/* Do we start at a new end? */
6898 		if ((use_which == RACK_USE_BEG) &&
6899 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
6900 			/*
6901 			 * When we get an ACK that just eats
6902 			 * up some of the rsm, we set RACK_USE_BEG
6903 			 * since whats at r_start (i.e. th_ack)
6904 			 * is left unacked and thats where the
6905 			 * measurement not starts.
6906 			 */
6907 			tp->gput_seq = rsm->r_start;
6908 			rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send;
6909 		}
6910 		if ((use_which == RACK_USE_END) &&
6911 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
6912 			    /*
6913 			     * We use the end when the cumack
6914 			     * is moving forward and completely
6915 			     * deleting the rsm passed so basically
6916 			     * r_end holds th_ack.
6917 			     *
6918 			     * For SACK's we also want to use the end
6919 			     * since this piece just got sacked and
6920 			     * we want to target anything after that
6921 			     * in our measurement.
6922 			     */
6923 			    tp->gput_seq = rsm->r_end;
6924 			    rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send;
6925 		}
6926 		if (use_which == RACK_USE_END_OR_THACK) {
6927 			/*
6928 			 * special case for ack moving forward,
6929 			 * not a sack, we need to move all the
6930 			 * way up to where this ack cum-ack moves
6931 			 * to.
6932 			 */
6933 			if (SEQ_GT(th_ack, rsm->r_end))
6934 				tp->gput_seq = th_ack;
6935 			else
6936 				tp->gput_seq = rsm->r_end;
6937 			rack->r_ctl.rc_gp_output_ts = rsm->usec_orig_send;
6938 		}
6939 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
6940 			/*
6941 			 * We moved beyond this guy's range, re-calculate
6942 			 * the new end point.
6943 			 */
6944 			if (rack->rc_gp_filled == 0) {
6945 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
6946 			} else {
6947 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
6948 			}
6949 		}
6950 		/*
6951 		 * We are moving the goal post, we may be able to clear the
6952 		 * measure_saw_probe_rtt flag.
6953 		 */
6954 		if ((rack->in_probe_rtt == 0) &&
6955 		    (rack->measure_saw_probe_rtt) &&
6956 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
6957 			rack->measure_saw_probe_rtt = 0;
6958 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
6959 					   seq, tp->gput_seq, 0, 5, line, NULL);
6960 		if (rack->rc_gp_filled &&
6961 		    ((tp->gput_ack - tp->gput_seq) <
6962 		     max(rc_init_window(rack), (MIN_GP_WIN *
6963 						ctf_fixed_maxseg(tp))))) {
6964 			/*
6965 			 * There is no sense of continuing this measurement
6966 			 * because its too small to gain us anything we
6967 			 * trust. Skip it and that way we can start a new
6968 			 * measurement quicker.
6969 			 */
6970 			rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
6971 						   0, 0, 0, 6, __LINE__, NULL);
6972 			tp->t_flags &= ~TF_GPUTINPROG;
6973 		}
6974 	}
6975 }
6976 
6977 static uint32_t
6978 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
6979 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
6980 {
6981 	uint32_t start, end, changed = 0;
6982 	struct rack_sendmap stack_map;
6983 	struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
6984 	int32_t used_ref = 1;
6985 	int moved = 0;
6986 
6987 	start = sack->start;
6988 	end = sack->end;
6989 	rsm = *prsm;
6990 	memset(&fe, 0, sizeof(fe));
6991 do_rest_ofb:
6992 	if ((rsm == NULL) ||
6993 	    (SEQ_LT(end, rsm->r_start)) ||
6994 	    (SEQ_GEQ(start, rsm->r_end)) ||
6995 	    (SEQ_LT(start, rsm->r_start))) {
6996 		/*
6997 		 * We are not in the right spot,
6998 		 * find the correct spot in the tree.
6999 		 */
7000 		used_ref = 0;
7001 		fe.r_start = start;
7002 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7003 		moved++;
7004 	}
7005 	if (rsm == NULL) {
7006 		/* TSNH */
7007 		goto out;
7008 	}
7009 	/* Ok we have an ACK for some piece of this rsm */
7010 	if (rsm->r_start != start) {
7011 		if ((rsm->r_flags & RACK_ACKED) == 0) {
7012 			/**
7013 			 * Need to split this in two pieces the before and after,
7014 			 * the before remains in the map, the after must be
7015 			 * added. In other words we have:
7016 			 * rsm        |--------------|
7017 			 * sackblk        |------->
7018 			 * rsm will become
7019 			 *     rsm    |---|
7020 			 * and nrsm will be  the sacked piece
7021 			 *     nrsm       |----------|
7022 			 *
7023 			 * But before we start down that path lets
7024 			 * see if the sack spans over on top of
7025 			 * the next guy and it is already sacked.
7026 			 */
7027 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7028 			if (next && (next->r_flags & RACK_ACKED) &&
7029 			    SEQ_GEQ(end, next->r_start)) {
7030 				/**
7031 				 * So the next one is already acked, and
7032 				 * we can thus by hookery use our stack_map
7033 				 * to reflect the piece being sacked and
7034 				 * then adjust the two tree entries moving
7035 				 * the start and ends around. So we start like:
7036 				 *  rsm     |------------|             (not-acked)
7037 				 *  next                 |-----------| (acked)
7038 				 *  sackblk        |-------->
7039 				 *  We want to end like so:
7040 				 *  rsm     |------|                   (not-acked)
7041 				 *  next           |-----------------| (acked)
7042 				 *  nrsm           |-----|
7043 				 * Where nrsm is a temporary stack piece we
7044 				 * use to update all the gizmos.
7045 				 */
7046 				/* Copy up our fudge block */
7047 				nrsm = &stack_map;
7048 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
7049 				/* Now adjust our tree blocks */
7050 				rsm->r_end = start;
7051 				next->r_start = start;
7052 				/* Clear out the dup ack count of the remainder */
7053 				rsm->r_dupack = 0;
7054 				rsm->r_just_ret = 0;
7055 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7056 				/* Now lets make sure our fudge block is right */
7057 				nrsm->r_start = start;
7058 				/* Now lets update all the stats and such */
7059 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
7060 				if (rack->app_limited_needs_set)
7061 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
7062 				changed += (nrsm->r_end - nrsm->r_start);
7063 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
7064 				if (nrsm->r_flags & RACK_SACK_PASSED) {
7065 					counter_u64_add(rack_reorder_seen, 1);
7066 					rack->r_ctl.rc_reorder_ts = cts;
7067 				}
7068 				/*
7069 				 * Now we want to go up from rsm (the
7070 				 * one left un-acked) to the next one
7071 				 * in the tmap. We do this so when
7072 				 * we walk backwards we include marking
7073 				 * sack-passed on rsm (The one passed in
7074 				 * is skipped since it is generally called
7075 				 * on something sacked before removing it
7076 				 * from the tmap).
7077 				 */
7078 				if (rsm->r_in_tmap) {
7079 					nrsm = TAILQ_NEXT(rsm, r_tnext);
7080 					/*
7081 					 * Now that we have the next
7082 					 * one walk backwards from there.
7083 					 */
7084 					if (nrsm && nrsm->r_in_tmap)
7085 						rack_log_sack_passed(tp, rack, nrsm);
7086 				}
7087 				/* Now are we done? */
7088 				if (SEQ_LT(end, next->r_end) ||
7089 				    (end == next->r_end)) {
7090 					/* Done with block */
7091 					goto out;
7092 				}
7093 				counter_u64_add(rack_sack_used_next_merge, 1);
7094 				/* Postion for the next block */
7095 				start = next->r_end;
7096 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
7097 				if (rsm == NULL)
7098 					goto out;
7099 			} else {
7100 				/**
7101 				 * We can't use any hookery here, so we
7102 				 * need to split the map. We enter like
7103 				 * so:
7104 				 *  rsm      |--------|
7105 				 *  sackblk       |----->
7106 				 * We will add the new block nrsm and
7107 				 * that will be the new portion, and then
7108 				 * fall through after reseting rsm. So we
7109 				 * split and look like this:
7110 				 *  rsm      |----|
7111 				 *  sackblk       |----->
7112 				 *  nrsm          |---|
7113 				 * We then fall through reseting
7114 				 * rsm to nrsm, so the next block
7115 				 * picks it up.
7116 				 */
7117 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
7118 				if (nrsm == NULL) {
7119 					/*
7120 					 * failed XXXrrs what can we do but loose the sack
7121 					 * info?
7122 					 */
7123 					goto out;
7124 				}
7125 				counter_u64_add(rack_sack_splits, 1);
7126 				rack_clone_rsm(rack, nrsm, rsm, start);
7127 				rsm->r_just_ret = 0;
7128 				insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7129 #ifdef INVARIANTS
7130 				if (insret != NULL) {
7131 					panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7132 					      nrsm, insret, rack, rsm);
7133 				}
7134 #endif
7135 				if (rsm->r_in_tmap) {
7136 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7137 					nrsm->r_in_tmap = 1;
7138 				}
7139 				rsm->r_flags &= (~RACK_HAS_FIN);
7140 				/* Position us to point to the new nrsm that starts the sack blk */
7141 				rsm = nrsm;
7142 			}
7143 		} else {
7144 			/* Already sacked this piece */
7145 			counter_u64_add(rack_sack_skipped_acked, 1);
7146 			moved++;
7147 			if (end == rsm->r_end) {
7148 				/* Done with block */
7149 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7150 				goto out;
7151 			} else if (SEQ_LT(end, rsm->r_end)) {
7152 				/* A partial sack to a already sacked block */
7153 				moved++;
7154 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7155 				goto out;
7156 			} else {
7157 				/*
7158 				 * The end goes beyond this guy
7159 				 * repostion the start to the
7160 				 * next block.
7161 				 */
7162 				start = rsm->r_end;
7163 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7164 				if (rsm == NULL)
7165 					goto out;
7166 			}
7167 		}
7168 	}
7169 	if (SEQ_GEQ(end, rsm->r_end)) {
7170 		/**
7171 		 * The end of this block is either beyond this guy or right
7172 		 * at this guy. I.e.:
7173 		 *  rsm ---                 |-----|
7174 		 *  end                     |-----|
7175 		 *  <or>
7176 		 *  end                     |---------|
7177 		 */
7178 		if ((rsm->r_flags & RACK_ACKED) == 0) {
7179 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
7180 			changed += (rsm->r_end - rsm->r_start);
7181 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7182 			if (rsm->r_in_tmap) /* should be true */
7183 				rack_log_sack_passed(tp, rack, rsm);
7184 			/* Is Reordering occuring? */
7185 			if (rsm->r_flags & RACK_SACK_PASSED) {
7186 				rsm->r_flags &= ~RACK_SACK_PASSED;
7187 				counter_u64_add(rack_reorder_seen, 1);
7188 				rack->r_ctl.rc_reorder_ts = cts;
7189 			}
7190 			if (rack->app_limited_needs_set)
7191 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
7192 			rsm->r_ack_arrival = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
7193 			rsm->r_flags |= RACK_ACKED;
7194 			rsm->r_flags &= ~RACK_TLP;
7195 			if (rsm->r_in_tmap) {
7196 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7197 				rsm->r_in_tmap = 0;
7198 			}
7199 		} else {
7200 			counter_u64_add(rack_sack_skipped_acked, 1);
7201 			moved++;
7202 		}
7203 		if (end == rsm->r_end) {
7204 			/* This block only - done, setup for next  */
7205 			goto out;
7206 		}
7207 		/*
7208 		 * There is more not coverend by this rsm move on
7209 		 * to the next block in the RB tree.
7210 		 */
7211 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7212 		start = rsm->r_end;
7213 		rsm = nrsm;
7214 		if (rsm == NULL)
7215 			goto out;
7216 		goto do_rest_ofb;
7217 	}
7218 	/**
7219 	 * The end of this sack block is smaller than
7220 	 * our rsm i.e.:
7221 	 *  rsm ---                 |-----|
7222 	 *  end                     |--|
7223 	 */
7224 	if ((rsm->r_flags & RACK_ACKED) == 0) {
7225 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7226 		if (prev && (prev->r_flags & RACK_ACKED)) {
7227 			/**
7228 			 * Goal, we want the right remainder of rsm to shrink
7229 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
7230 			 * We want to expand prev to go all the way
7231 			 * to prev->r_end <- end.
7232 			 * so in the tree we have before:
7233 			 *   prev     |--------|         (acked)
7234 			 *   rsm               |-------| (non-acked)
7235 			 *   sackblk           |-|
7236 			 * We churn it so we end up with
7237 			 *   prev     |----------|       (acked)
7238 			 *   rsm                 |-----| (non-acked)
7239 			 *   nrsm              |-| (temporary)
7240 			 */
7241 			nrsm = &stack_map;
7242 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
7243 			prev->r_end = end;
7244 			rsm->r_start = end;
7245 			/* Now adjust nrsm (stack copy) to be
7246 			 * the one that is the small
7247 			 * piece that was "sacked".
7248 			 */
7249 			nrsm->r_end = end;
7250 			rsm->r_dupack = 0;
7251 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7252 			/*
7253 			 * Now nrsm is our new little piece
7254 			 * that is acked (which was merged
7255 			 * to prev). Update the rtt and changed
7256 			 * based on that. Also check for reordering.
7257 			 */
7258 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
7259 			if (rack->app_limited_needs_set)
7260 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
7261 			changed += (nrsm->r_end - nrsm->r_start);
7262 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
7263 			if (nrsm->r_flags & RACK_SACK_PASSED) {
7264 				counter_u64_add(rack_reorder_seen, 1);
7265 				rack->r_ctl.rc_reorder_ts = cts;
7266 			}
7267 			rsm = prev;
7268 			counter_u64_add(rack_sack_used_prev_merge, 1);
7269 		} else {
7270 			/**
7271 			 * This is the case where our previous
7272 			 * block is not acked either, so we must
7273 			 * split the block in two.
7274 			 */
7275 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
7276 			if (nrsm == NULL) {
7277 				/* failed rrs what can we do but loose the sack info? */
7278 				goto out;
7279 			}
7280 			/**
7281 			 * In this case nrsm becomes
7282 			 * nrsm->r_start = end;
7283 			 * nrsm->r_end = rsm->r_end;
7284 			 * which is un-acked.
7285 			 * <and>
7286 			 * rsm->r_end = nrsm->r_start;
7287 			 * i.e. the remaining un-acked
7288 			 * piece is left on the left
7289 			 * hand side.
7290 			 *
7291 			 * So we start like this
7292 			 * rsm      |----------| (not acked)
7293 			 * sackblk  |---|
7294 			 * build it so we have
7295 			 * rsm      |---|         (acked)
7296 			 * nrsm         |------|  (not acked)
7297 			 */
7298 			counter_u64_add(rack_sack_splits, 1);
7299 			rack_clone_rsm(rack, nrsm, rsm, end);
7300 			rsm->r_flags &= (~RACK_HAS_FIN);
7301 			rsm->r_just_ret = 0;
7302 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7303 #ifdef INVARIANTS
7304 			if (insret != NULL) {
7305 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7306 				      nrsm, insret, rack, rsm);
7307 			}
7308 #endif
7309 			if (rsm->r_in_tmap) {
7310 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7311 				nrsm->r_in_tmap = 1;
7312 			}
7313 			nrsm->r_dupack = 0;
7314 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7315 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
7316 			changed += (rsm->r_end - rsm->r_start);
7317 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7318 			if (rsm->r_in_tmap) /* should be true */
7319 				rack_log_sack_passed(tp, rack, rsm);
7320 			/* Is Reordering occuring? */
7321 			if (rsm->r_flags & RACK_SACK_PASSED) {
7322 				rsm->r_flags &= ~RACK_SACK_PASSED;
7323 				counter_u64_add(rack_reorder_seen, 1);
7324 				rack->r_ctl.rc_reorder_ts = cts;
7325 			}
7326 			if (rack->app_limited_needs_set)
7327 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
7328 			rsm->r_ack_arrival = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
7329 			rsm->r_flags |= RACK_ACKED;
7330 			rsm->r_flags &= ~RACK_TLP;
7331 			if (rsm->r_in_tmap) {
7332 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7333 				rsm->r_in_tmap = 0;
7334 			}
7335 		}
7336 	} else if (start != end){
7337 		/*
7338 		 * The block was already acked.
7339 		 */
7340 		counter_u64_add(rack_sack_skipped_acked, 1);
7341 		moved++;
7342 	}
7343 out:
7344 	if (rsm && (rsm->r_flags & RACK_ACKED)) {
7345 		/*
7346 		 * Now can we merge where we worked
7347 		 * with either the previous or
7348 		 * next block?
7349 		 */
7350 		next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7351 		while (next) {
7352 		    if (next->r_flags & RACK_ACKED) {
7353 			/* yep this and next can be merged */
7354 			rsm = rack_merge_rsm(rack, rsm, next);
7355 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7356 		    } else
7357 			    break;
7358 		}
7359 		/* Now what about the previous? */
7360 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7361 		while (prev) {
7362 		    if (prev->r_flags & RACK_ACKED) {
7363 			/* yep the previous and this can be merged */
7364 			rsm = rack_merge_rsm(rack, prev, rsm);
7365 			prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7366 		    } else
7367 			    break;
7368 		}
7369 	}
7370 	if (used_ref == 0) {
7371 		counter_u64_add(rack_sack_proc_all, 1);
7372 	} else {
7373 		counter_u64_add(rack_sack_proc_short, 1);
7374 	}
7375 	/* Save off the next one for quick reference. */
7376 	if (rsm)
7377 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7378 	else
7379 		nrsm = NULL;
7380 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
7381 	/* Pass back the moved. */
7382 	*moved_two = moved;
7383 	return (changed);
7384 }
7385 
7386 static void inline
7387 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
7388 {
7389 	struct rack_sendmap *tmap;
7390 
7391 	tmap = NULL;
7392 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
7393 		/* Its no longer sacked, mark it so */
7394 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7395 #ifdef INVARIANTS
7396 		if (rsm->r_in_tmap) {
7397 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
7398 			      rack, rsm, rsm->r_flags);
7399 		}
7400 #endif
7401 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
7402 		/* Rebuild it into our tmap */
7403 		if (tmap == NULL) {
7404 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7405 			tmap = rsm;
7406 		} else {
7407 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7408 			tmap = rsm;
7409 		}
7410 		tmap->r_in_tmap = 1;
7411 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7412 	}
7413 	/*
7414 	 * Now lets possibly clear the sack filter so we start
7415 	 * recognizing sacks that cover this area.
7416 	 */
7417 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
7418 
7419 }
7420 
7421 static void
7422 rack_do_decay(struct tcp_rack *rack)
7423 {
7424 	struct timeval res;
7425 
7426 #define	timersub(tvp, uvp, vvp)						\
7427 	do {								\
7428 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
7429 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
7430 		if ((vvp)->tv_usec < 0) {				\
7431 			(vvp)->tv_sec--;				\
7432 			(vvp)->tv_usec += 1000000;			\
7433 		}							\
7434 	} while (0)
7435 
7436 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
7437 #undef timersub
7438 
7439 	rack->r_ctl.input_pkt++;
7440 	if ((rack->rc_in_persist) ||
7441 	    (res.tv_sec >= 1) ||
7442 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
7443 		/*
7444 		 * Check for decay of non-SAD,
7445 		 * we want all SAD detection metrics to
7446 		 * decay 1/4 per second (or more) passed.
7447 		 */
7448 		uint32_t pkt_delta;
7449 
7450 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
7451 		/* Update our saved tracking values */
7452 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
7453 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
7454 		/* Now do we escape without decay? */
7455 #ifdef NETFLIX_EXP_DETECTION
7456 		if (rack->rc_in_persist ||
7457 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
7458 		    (pkt_delta < tcp_sad_low_pps)){
7459 			/*
7460 			 * We don't decay idle connections
7461 			 * or ones that have a low input pps.
7462 			 */
7463 			return;
7464 		}
7465 		/* Decay the counters */
7466 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
7467 							tcp_sad_decay_val);
7468 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
7469 							 tcp_sad_decay_val);
7470 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
7471 							       tcp_sad_decay_val);
7472 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
7473 								tcp_sad_decay_val);
7474 #endif
7475 	}
7476 }
7477 
7478 static void
7479 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th)
7480 {
7481 	uint32_t changed, entered_recovery = 0;
7482 	struct tcp_rack *rack;
7483 	struct rack_sendmap *rsm, *rm;
7484 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7485 	register uint32_t th_ack;
7486 	int32_t i, j, k, num_sack_blks = 0;
7487 	uint32_t cts, acked, ack_point, sack_changed = 0;
7488 	int loop_start = 0, moved_two = 0;
7489 	uint32_t tsused;
7490 
7491 
7492 	INP_WLOCK_ASSERT(tp->t_inpcb);
7493 	if (th->th_flags & TH_RST) {
7494 		/* We don't log resets */
7495 		return;
7496 	}
7497 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7498 	cts = tcp_ts_getticks();
7499 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
7500 	changed = 0;
7501 	th_ack = th->th_ack;
7502 	if (rack->sack_attack_disable == 0)
7503 		rack_do_decay(rack);
7504 	if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
7505 		/*
7506 		 * You only get credit for
7507 		 * MSS and greater (and you get extra
7508 		 * credit for larger cum-ack moves).
7509 		 */
7510 		int ac;
7511 
7512 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
7513 		rack->r_ctl.ack_count += ac;
7514 		counter_u64_add(rack_ack_total, ac);
7515 	}
7516 	if (rack->r_ctl.ack_count > 0xfff00000) {
7517 		/*
7518 		 * reduce the number to keep us under
7519 		 * a uint32_t.
7520 		 */
7521 		rack->r_ctl.ack_count /= 2;
7522 		rack->r_ctl.sack_count /= 2;
7523 	}
7524 	if (SEQ_GT(th_ack, tp->snd_una)) {
7525 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
7526 		tp->t_acktime = ticks;
7527 	}
7528 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
7529 		changed = th_ack - rsm->r_start;
7530 	if (changed) {
7531 		/*
7532 		 * The ACK point is advancing to th_ack, we must drop off
7533 		 * the packets in the rack log and calculate any eligble
7534 		 * RTT's.
7535 		 */
7536 		rack->r_wanted_output = 1;
7537 more:
7538 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
7539 		if (rsm == NULL) {
7540 			if ((th_ack - 1) == tp->iss) {
7541 				/*
7542 				 * For the SYN incoming case we will not
7543 				 * have called tcp_output for the sending of
7544 				 * the SYN, so there will be no map. All
7545 				 * other cases should probably be a panic.
7546 				 */
7547 				goto proc_sack;
7548 			}
7549 			if (tp->t_flags & TF_SENTFIN) {
7550 				/* if we send a FIN we will not hav a map */
7551 				goto proc_sack;
7552 			}
7553 #ifdef INVARIANTS
7554 			panic("No rack map tp:%p for th:%p state:%d rack:%p snd_una:%u snd_max:%u snd_nxt:%u chg:%d\n",
7555 			      tp,
7556 			      th, tp->t_state, rack,
7557 			      tp->snd_una, tp->snd_max, tp->snd_nxt, changed);
7558 #endif
7559 			goto proc_sack;
7560 		}
7561 		if (SEQ_LT(th_ack, rsm->r_start)) {
7562 			/* Huh map is missing this */
7563 #ifdef INVARIANTS
7564 			printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
7565 			       rsm->r_start,
7566 			       th_ack, tp->t_state, rack->r_state);
7567 #endif
7568 			goto proc_sack;
7569 		}
7570 		rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
7571 		/* Now do we consume the whole thing? */
7572 		if (SEQ_GEQ(th_ack, rsm->r_end)) {
7573 			/* Its all consumed. */
7574 			uint32_t left;
7575 			uint8_t newly_acked;
7576 
7577 			rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7578 			rsm->r_rtr_bytes = 0;
7579 			/* Record the time of highest cumack sent */
7580 			rack->r_ctl.rc_gp_cumack_ts = rsm->usec_orig_send;
7581 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7582 #ifdef INVARIANTS
7583 			if (rm != rsm) {
7584 				panic("removing head in rack:%p rsm:%p rm:%p",
7585 				      rack, rsm, rm);
7586 			}
7587 #endif
7588 			if (rsm->r_in_tmap) {
7589 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7590 				rsm->r_in_tmap = 0;
7591 			}
7592 			newly_acked = 1;
7593 			if (rsm->r_flags & RACK_ACKED) {
7594 				/*
7595 				 * It was acked on the scoreboard -- remove
7596 				 * it from total
7597 				 */
7598 				rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7599 				newly_acked = 0;
7600 			} else if (rsm->r_flags & RACK_SACK_PASSED) {
7601 				/*
7602 				 * There are segments ACKED on the
7603 				 * scoreboard further up. We are seeing
7604 				 * reordering.
7605 				 */
7606 				rsm->r_flags &= ~RACK_SACK_PASSED;
7607 				counter_u64_add(rack_reorder_seen, 1);
7608 				rsm->r_ack_arrival = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
7609 				rsm->r_flags |= RACK_ACKED;
7610 				rack->r_ctl.rc_reorder_ts = cts;
7611 			}
7612 			left = th_ack - rsm->r_end;
7613 			if (rack->app_limited_needs_set && newly_acked)
7614 				rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
7615 			/* Free back to zone */
7616 			rack_free(rack, rsm);
7617 			if (left) {
7618 				goto more;
7619 			}
7620 			goto proc_sack;
7621 		}
7622 		if (rsm->r_flags & RACK_ACKED) {
7623 			/*
7624 			 * It was acked on the scoreboard -- remove it from
7625 			 * total for the part being cum-acked.
7626 			 */
7627 			rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7628 		}
7629 		/*
7630 		 * Clear the dup ack count for
7631 		 * the piece that remains.
7632 		 */
7633 		rsm->r_dupack = 0;
7634 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7635 		if (rsm->r_rtr_bytes) {
7636 			/*
7637 			 * It was retransmitted adjust the
7638 			 * sack holes for what was acked.
7639 			 */
7640 			int ack_am;
7641 
7642 			ack_am = (th_ack - rsm->r_start);
7643 			if (ack_am >= rsm->r_rtr_bytes) {
7644 				rack->r_ctl.rc_holes_rxt -= ack_am;
7645 				rsm->r_rtr_bytes -= ack_am;
7646 			}
7647 		}
7648 		/*
7649 		 * Update where the piece starts and record
7650 		 * the time of send of highest cumack sent.
7651 		 */
7652 		rack->r_ctl.rc_gp_cumack_ts = rsm->usec_orig_send;
7653 		rsm->r_start = th_ack;
7654 		if (rack->app_limited_needs_set)
7655 			rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
7656 
7657 	}
7658 proc_sack:
7659 	/* Check for reneging */
7660 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
7661 	if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
7662 		/*
7663 		 * The peer has moved snd_una up to
7664 		 * the edge of this send, i.e. one
7665 		 * that it had previously acked. The only
7666 		 * way that can be true if the peer threw
7667 		 * away data (space issues) that it had
7668 		 * previously sacked (else it would have
7669 		 * given us snd_una up to (rsm->r_end).
7670 		 * We need to undo the acked markings here.
7671 		 *
7672 		 * Note we have to look to make sure th_ack is
7673 		 * our rsm->r_start in case we get an old ack
7674 		 * where th_ack is behind snd_una.
7675 		 */
7676 		rack_peer_reneges(rack, rsm, th->th_ack);
7677 	}
7678 	if ((to->to_flags & TOF_SACK) == 0) {
7679 		/* We are done nothing left */
7680 		goto out;
7681 	}
7682 	/* Sack block processing */
7683 	if (SEQ_GT(th_ack, tp->snd_una))
7684 		ack_point = th_ack;
7685 	else
7686 		ack_point = tp->snd_una;
7687 	for (i = 0; i < to->to_nsacks; i++) {
7688 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
7689 		      &sack, sizeof(sack));
7690 		sack.start = ntohl(sack.start);
7691 		sack.end = ntohl(sack.end);
7692 		if (SEQ_GT(sack.end, sack.start) &&
7693 		    SEQ_GT(sack.start, ack_point) &&
7694 		    SEQ_LT(sack.start, tp->snd_max) &&
7695 		    SEQ_GT(sack.end, ack_point) &&
7696 		    SEQ_LEQ(sack.end, tp->snd_max)) {
7697 			sack_blocks[num_sack_blks] = sack;
7698 			num_sack_blks++;
7699 #ifdef NETFLIX_STATS
7700 		} else if (SEQ_LEQ(sack.start, th_ack) &&
7701 			   SEQ_LEQ(sack.end, th_ack)) {
7702 			/*
7703 			 * Its a D-SACK block.
7704 			 */
7705 			tcp_record_dsack(sack.start, sack.end);
7706 #endif
7707 		}
7708 
7709 	}
7710 	/*
7711 	 * Sort the SACK blocks so we can update the rack scoreboard with
7712 	 * just one pass.
7713 	 */
7714 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
7715 					 num_sack_blks, th->th_ack);
7716 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
7717 	if (num_sack_blks == 0)  {
7718 		/* Nothing to sack (DSACKs?) */
7719 		goto out_with_totals;
7720 	}
7721 	if (num_sack_blks < 2) {
7722 		/* Only one, we don't need to sort */
7723 		goto do_sack_work;
7724 	}
7725 	/* Sort the sacks */
7726 	for (i = 0; i < num_sack_blks; i++) {
7727 		for (j = i + 1; j < num_sack_blks; j++) {
7728 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7729 				sack = sack_blocks[i];
7730 				sack_blocks[i] = sack_blocks[j];
7731 				sack_blocks[j] = sack;
7732 			}
7733 		}
7734 	}
7735 	/*
7736 	 * Now are any of the sack block ends the same (yes some
7737 	 * implementations send these)?
7738 	 */
7739 again:
7740 	if (num_sack_blks == 0)
7741 		goto out_with_totals;
7742 	if (num_sack_blks > 1) {
7743 		for (i = 0; i < num_sack_blks; i++) {
7744 			for (j = i + 1; j < num_sack_blks; j++) {
7745 				if (sack_blocks[i].end == sack_blocks[j].end) {
7746 					/*
7747 					 * Ok these two have the same end we
7748 					 * want the smallest end and then
7749 					 * throw away the larger and start
7750 					 * again.
7751 					 */
7752 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7753 						/*
7754 						 * The second block covers
7755 						 * more area use that
7756 						 */
7757 						sack_blocks[i].start = sack_blocks[j].start;
7758 					}
7759 					/*
7760 					 * Now collapse out the dup-sack and
7761 					 * lower the count
7762 					 */
7763 					for (k = (j + 1); k < num_sack_blks; k++) {
7764 						sack_blocks[j].start = sack_blocks[k].start;
7765 						sack_blocks[j].end = sack_blocks[k].end;
7766 						j++;
7767 					}
7768 					num_sack_blks--;
7769 					goto again;
7770 				}
7771 			}
7772 		}
7773 	}
7774 do_sack_work:
7775 	/*
7776 	 * First lets look to see if
7777 	 * we have retransmitted and
7778 	 * can use the transmit next?
7779 	 */
7780 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7781 	if (rsm &&
7782 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
7783 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
7784 		/*
7785 		 * We probably did the FR and the next
7786 		 * SACK in continues as we would expect.
7787 		 */
7788 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
7789 		if (acked) {
7790 			rack->r_wanted_output = 1;
7791 			changed += acked;
7792 			sack_changed += acked;
7793 		}
7794 		if (num_sack_blks == 1) {
7795 			/*
7796 			 * This is what we would expect from
7797 			 * a normal implementation to happen
7798 			 * after we have retransmitted the FR,
7799 			 * i.e the sack-filter pushes down
7800 			 * to 1 block and the next to be retransmitted
7801 			 * is the sequence in the sack block (has more
7802 			 * are acked). Count this as ACK'd data to boost
7803 			 * up the chances of recovering any false positives.
7804 			 */
7805 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
7806 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
7807 			counter_u64_add(rack_express_sack, 1);
7808 			if (rack->r_ctl.ack_count > 0xfff00000) {
7809 				/*
7810 				 * reduce the number to keep us under
7811 				 * a uint32_t.
7812 				 */
7813 				rack->r_ctl.ack_count /= 2;
7814 				rack->r_ctl.sack_count /= 2;
7815 			}
7816 			goto out_with_totals;
7817 		} else {
7818 			/*
7819 			 * Start the loop through the
7820 			 * rest of blocks, past the first block.
7821 			 */
7822 			moved_two = 0;
7823 			loop_start = 1;
7824 		}
7825 	}
7826 	/* Its a sack of some sort */
7827 	rack->r_ctl.sack_count++;
7828 	if (rack->r_ctl.sack_count > 0xfff00000) {
7829 		/*
7830 		 * reduce the number to keep us under
7831 		 * a uint32_t.
7832 		 */
7833 		rack->r_ctl.ack_count /= 2;
7834 		rack->r_ctl.sack_count /= 2;
7835 	}
7836 	counter_u64_add(rack_sack_total, 1);
7837 	if (rack->sack_attack_disable) {
7838 		/* An attacker disablement is in place */
7839 		if (num_sack_blks > 1) {
7840 			rack->r_ctl.sack_count += (num_sack_blks - 1);
7841 			rack->r_ctl.sack_moved_extra++;
7842 			counter_u64_add(rack_move_some, 1);
7843 			if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
7844 				rack->r_ctl.sack_moved_extra /= 2;
7845 				rack->r_ctl.sack_noextra_move /= 2;
7846 			}
7847 		}
7848 		goto out;
7849 	}
7850 	rsm = rack->r_ctl.rc_sacklast;
7851 	for (i = loop_start; i < num_sack_blks; i++) {
7852 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
7853 		if (acked) {
7854 			rack->r_wanted_output = 1;
7855 			changed += acked;
7856 			sack_changed += acked;
7857 		}
7858 		if (moved_two) {
7859 			/*
7860 			 * If we did not get a SACK for at least a MSS and
7861 			 * had to move at all, or if we moved more than our
7862 			 * threshold, it counts against the "extra" move.
7863 			 */
7864 			rack->r_ctl.sack_moved_extra += moved_two;
7865 			counter_u64_add(rack_move_some, 1);
7866 		} else {
7867 			/*
7868 			 * else we did not have to move
7869 			 * any more than we would expect.
7870 			 */
7871 			rack->r_ctl.sack_noextra_move++;
7872 			counter_u64_add(rack_move_none, 1);
7873 		}
7874 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
7875 			/*
7876 			 * If the SACK was not a full MSS then
7877 			 * we add to sack_count the number of
7878 			 * MSS's (or possibly more than
7879 			 * a MSS if its a TSO send) we had to skip by.
7880 			 */
7881 			rack->r_ctl.sack_count += moved_two;
7882 			counter_u64_add(rack_sack_total, moved_two);
7883 		}
7884 		/*
7885 		 * Now we need to setup for the next
7886 		 * round. First we make sure we won't
7887 		 * exceed the size of our uint32_t on
7888 		 * the various counts, and then clear out
7889 		 * moved_two.
7890 		 */
7891 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
7892 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
7893 			rack->r_ctl.sack_moved_extra /= 2;
7894 			rack->r_ctl.sack_noextra_move /= 2;
7895 		}
7896 		if (rack->r_ctl.sack_count > 0xfff00000) {
7897 			rack->r_ctl.ack_count /= 2;
7898 			rack->r_ctl.sack_count /= 2;
7899 		}
7900 		moved_two = 0;
7901 	}
7902 out_with_totals:
7903 	if (num_sack_blks > 1) {
7904 		/*
7905 		 * You get an extra stroke if
7906 		 * you have more than one sack-blk, this
7907 		 * could be where we are skipping forward
7908 		 * and the sack-filter is still working, or
7909 		 * it could be an attacker constantly
7910 		 * moving us.
7911 		 */
7912 		rack->r_ctl.sack_moved_extra++;
7913 		counter_u64_add(rack_move_some, 1);
7914 	}
7915 out:
7916 #ifdef NETFLIX_EXP_DETECTION
7917 	if ((rack->do_detection || tcp_force_detection) &&
7918 	    tcp_sack_to_ack_thresh &&
7919 	    tcp_sack_to_move_thresh &&
7920 	    ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
7921 		/*
7922 		 * We have thresholds set to find
7923 		 * possible attackers and disable sack.
7924 		 * Check them.
7925 		 */
7926 		uint64_t ackratio, moveratio, movetotal;
7927 
7928 		/* Log detecting */
7929 		rack_log_sad(rack, 1);
7930 		ackratio = (uint64_t)(rack->r_ctl.sack_count);
7931 		ackratio *= (uint64_t)(1000);
7932 		if (rack->r_ctl.ack_count)
7933 			ackratio /= (uint64_t)(rack->r_ctl.ack_count);
7934 		else {
7935 			/* We really should not hit here */
7936 			ackratio = 1000;
7937 		}
7938 		if ((rack->sack_attack_disable  == 0) &&
7939 		    (ackratio > rack_highest_sack_thresh_seen))
7940 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
7941 		movetotal = rack->r_ctl.sack_moved_extra;
7942 		movetotal += rack->r_ctl.sack_noextra_move;
7943 		moveratio = rack->r_ctl.sack_moved_extra;
7944 		moveratio *= (uint64_t)1000;
7945 		if (movetotal)
7946 			moveratio /= movetotal;
7947 		else {
7948 			/* No moves, thats pretty good */
7949 			moveratio = 0;
7950 		}
7951 		if ((rack->sack_attack_disable == 0) &&
7952 		    (moveratio > rack_highest_move_thresh_seen))
7953 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
7954 		if (rack->sack_attack_disable == 0) {
7955 			if ((ackratio > tcp_sack_to_ack_thresh) &&
7956 			    (moveratio > tcp_sack_to_move_thresh)) {
7957 				/* Disable sack processing */
7958 				rack->sack_attack_disable = 1;
7959 				if (rack->r_rep_attack == 0) {
7960 					rack->r_rep_attack = 1;
7961 					counter_u64_add(rack_sack_attacks_detected, 1);
7962 				}
7963 				if (tcp_attack_on_turns_on_logging) {
7964 					/*
7965 					 * Turn on logging, used for debugging
7966 					 * false positives.
7967 					 */
7968 					rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
7969 				}
7970 				/* Clamp the cwnd at flight size */
7971 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
7972 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
7973 				rack_log_sad(rack, 2);
7974 			}
7975 		} else {
7976 			/* We are sack-disabled check for false positives */
7977 			if ((ackratio <= tcp_restoral_thresh) ||
7978 			    (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
7979 				rack->sack_attack_disable  = 0;
7980 				rack_log_sad(rack, 3);
7981 				/* Restart counting */
7982 				rack->r_ctl.sack_count = 0;
7983 				rack->r_ctl.sack_moved_extra = 0;
7984 				rack->r_ctl.sack_noextra_move = 1;
7985 				rack->r_ctl.ack_count = max(1,
7986 				      (BYTES_THIS_ACK(tp, th)/ctf_fixed_maxseg(rack->rc_tp)));
7987 
7988 				if (rack->r_rep_reverse == 0) {
7989 					rack->r_rep_reverse = 1;
7990 					counter_u64_add(rack_sack_attacks_reversed, 1);
7991 				}
7992 				/* Restore the cwnd */
7993 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
7994 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
7995 			}
7996 		}
7997 	}
7998 #endif
7999 	if (changed) {
8000 		/* Something changed cancel the rack timer */
8001 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
8002 	}
8003 	tsused = tcp_ts_getticks();
8004 	rsm = tcp_rack_output(tp, rack, tsused);
8005 	if ((!IN_RECOVERY(tp->t_flags)) &&
8006 	    rsm) {
8007 		/* Enter recovery */
8008 		rack->r_ctl.rc_rsm_start = rsm->r_start;
8009 		rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
8010 		rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
8011 		entered_recovery = 1;
8012 		rack_cong_signal(tp, NULL, CC_NDUPACK);
8013 		/*
8014 		 * When we enter recovery we need to assure we send
8015 		 * one packet.
8016 		 */
8017 		if (rack->rack_no_prr == 0) {
8018 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
8019 			rack_log_to_prr(rack, 8, 0);
8020 		}
8021 		rack->r_timer_override = 1;
8022 		rack->r_early = 0;
8023 		rack->r_ctl.rc_agg_early = 0;
8024 	} else if (IN_RECOVERY(tp->t_flags) &&
8025 		   rsm &&
8026  		   (rack->r_rr_config == 3)) {
8027 		/*
8028 		 * Assure we can output and we get no
8029 		 * remembered pace time except the retransmit.
8030 		 */
8031 		rack->r_timer_override = 1;
8032 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8033 		rack->r_ctl.rc_resend = rsm;
8034 	}
8035 	if (IN_RECOVERY(tp->t_flags) &&
8036 	    (rack->rack_no_prr == 0) &&
8037 	    (entered_recovery == 0)) {
8038 		/* Deal with PRR here (in recovery only) */
8039 		uint32_t pipe, snd_una;
8040 
8041 		rack->r_ctl.rc_prr_delivered += changed;
8042 		/* Compute prr_sndcnt */
8043 		if (SEQ_GT(tp->snd_una, th_ack)) {
8044 			snd_una = tp->snd_una;
8045 		} else {
8046 			snd_una = th_ack;
8047 		}
8048 		pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
8049 		if (pipe > tp->snd_ssthresh) {
8050 			long sndcnt;
8051 
8052 			sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
8053 			if (rack->r_ctl.rc_prr_recovery_fs > 0)
8054 				sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
8055 			else {
8056 				rack->r_ctl.rc_prr_sndcnt = 0;
8057 				rack_log_to_prr(rack, 9, 0);
8058 				sndcnt = 0;
8059 			}
8060 			sndcnt++;
8061 			if (sndcnt > (long)rack->r_ctl.rc_prr_out)
8062 				sndcnt -= rack->r_ctl.rc_prr_out;
8063 			else
8064 				sndcnt = 0;
8065 			rack->r_ctl.rc_prr_sndcnt = sndcnt;
8066 			rack_log_to_prr(rack, 10, 0);
8067 		} else {
8068 			uint32_t limit;
8069 
8070 			if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
8071 				limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
8072 			else
8073 				limit = 0;
8074 			if (changed > limit)
8075 				limit = changed;
8076 			limit += ctf_fixed_maxseg(tp);
8077 			if (tp->snd_ssthresh > pipe) {
8078 				rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
8079 				rack_log_to_prr(rack, 11, 0);
8080 			} else {
8081 				rack->r_ctl.rc_prr_sndcnt = min(0, limit);
8082 				rack_log_to_prr(rack, 12, 0);
8083 			}
8084 		}
8085 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
8086 		     ((rack->rc_inp->inp_in_hpts == 0) &&
8087 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
8088 			/*
8089 			 * If you are pacing output you don't want
8090 			 * to override.
8091 			 */
8092 			rack->r_early = 0;
8093 			rack->r_ctl.rc_agg_early = 0;
8094 			rack->r_timer_override = 1;
8095 		}
8096 	}
8097 }
8098 
8099 static void
8100 rack_strike_dupack(struct tcp_rack *rack)
8101 {
8102 	struct rack_sendmap *rsm;
8103 
8104 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
8105 	if (rsm && (rsm->r_dupack < 0xff)) {
8106 		rsm->r_dupack++;
8107 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
8108 			rack->r_wanted_output = 1;
8109 			rack->r_timer_override = 1;
8110 			rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
8111 		} else {
8112 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
8113 		}
8114 	}
8115 }
8116 
8117 static void
8118 rack_check_bottom_drag(struct tcpcb *tp,
8119 		       struct tcp_rack *rack,
8120 		       struct socket *so, int32_t acked)
8121 {
8122 	uint32_t segsiz, minseg;
8123 
8124 	segsiz = ctf_fixed_maxseg(tp);
8125 	minseg = segsiz;
8126 
8127 	if (tp->snd_max == tp->snd_una) {
8128 		/*
8129 		 * We are doing dynamic pacing and we are way
8130 		 * under. Basically everything got acked while
8131 		 * we were still waiting on the pacer to expire.
8132 		 *
8133 		 * This means we need to boost the b/w in
8134 		 * addition to any earlier boosting of
8135 		 * the multipler.
8136 		 */
8137 		rack->rc_dragged_bottom = 1;
8138 		rack_validate_multipliers_at_or_above100(rack);
8139 		/*
8140 		 * Lets use the segment bytes acked plus
8141 		 * the lowest RTT seen as the basis to
8142 		 * form a b/w estimate. This will be off
8143 		 * due to the fact that the true estimate
8144 		 * should be around 1/2 the time of the RTT
8145 		 * but we can settle for that.
8146 		 */
8147 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
8148 		    acked) {
8149 			uint64_t bw, calc_bw, rtt;
8150 
8151 			rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8152 			bw = acked;
8153 			calc_bw = bw * 1000000;
8154 			calc_bw /= rtt;
8155 			if (rack->r_ctl.last_max_bw &&
8156 			    (rack->r_ctl.last_max_bw < calc_bw)) {
8157 				/*
8158 				 * If we have a last calculated max bw
8159 				 * enforce it.
8160 				 */
8161 				calc_bw = rack->r_ctl.last_max_bw;
8162 			}
8163 			/* now plop it in */
8164 			if (rack->rc_gp_filled == 0) {
8165 				if (calc_bw > ONE_POINT_TWO_MEG) {
8166 					/*
8167 					 * If we have no measurement
8168 					 * don't let us set in more than
8169 					 * 1.2Mbps. If we are still too
8170 					 * low after pacing with this we
8171 					 * will hopefully have a max b/w
8172 					 * available to sanity check things.
8173 					 */
8174 					calc_bw = ONE_POINT_TWO_MEG;
8175 				}
8176 				rack->r_ctl.rc_rtt_diff = 0;
8177 				rack->r_ctl.gp_bw = calc_bw;
8178 				rack->rc_gp_filled = 1;
8179 				rack->r_ctl.num_avg = RACK_REQ_AVG;
8180 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
8181 			} else if (calc_bw > rack->r_ctl.gp_bw) {
8182 				rack->r_ctl.rc_rtt_diff = 0;
8183 				rack->r_ctl.num_avg = RACK_REQ_AVG;
8184 				rack->r_ctl.gp_bw = calc_bw;
8185 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
8186 			} else
8187 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
8188 			/*
8189 			 * For acks over 1mss we do a extra boost to simulate
8190 			 * where we would get 2 acks (we want 110 for the mul).
8191 			 */
8192 			if (acked > segsiz)
8193 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
8194 		} else {
8195 			/*
8196 			 * Huh, this should not be, settle
8197 			 * for just an old increase.
8198 			 */
8199 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
8200 		}
8201 	} else if ((IN_RECOVERY(tp->t_flags) == 0) &&
8202 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
8203 					       minseg)) &&
8204 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
8205 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
8206 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
8207 		    (segsiz * rack_req_segs))) {
8208 		/*
8209 		 * We are doing dynamic GP pacing and
8210 		 * we have everything except 1MSS or less
8211 		 * bytes left out. We are still pacing away.
8212 		 * And there is data that could be sent, This
8213 		 * means we are inserting delayed ack time in
8214 		 * our measurements because we are pacing too slow.
8215 		 */
8216 		rack_validate_multipliers_at_or_above100(rack);
8217 		rack->rc_dragged_bottom = 1;
8218 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
8219 	}
8220 }
8221 
8222 /*
8223  * Return value of 1, we do not need to call rack_process_data().
8224  * return value of 0, rack_process_data can be called.
8225  * For ret_val if its 0 the TCP is locked, if its non-zero
8226  * its unlocked and probably unsafe to touch the TCB.
8227  */
8228 static int
8229 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8230     struct tcpcb *tp, struct tcpopt *to,
8231     uint32_t tiwin, int32_t tlen,
8232     int32_t * ofia, int32_t thflags, int32_t * ret_val)
8233 {
8234 	int32_t ourfinisacked = 0;
8235 	int32_t nsegs, acked_amount;
8236 	int32_t acked;
8237 	struct mbuf *mfree;
8238 	struct tcp_rack *rack;
8239 	int32_t under_pacing = 0;
8240 	int32_t recovery = 0;
8241 
8242 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8243 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
8244 		ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
8245 		rack->r_wanted_output = 1;
8246 		return (1);
8247 	}
8248 	if (rack->rc_gp_filled &&
8249 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
8250 		under_pacing = 1;
8251 	}
8252 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
8253 		if (rack->rc_in_persist)
8254 			tp->t_rxtshift = 0;
8255 		if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
8256 			rack_strike_dupack(rack);
8257 		rack_log_ack(tp, to, th);
8258 	}
8259 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8260 		/*
8261 		 * Old ack, behind (or duplicate to) the last one rcv'd
8262 		 * Note: Should mark reordering is occuring! We should also
8263 		 * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1,
8264 		 * 3-3, 4-4 would be reording. As well as ack 1, 3-3 <no
8265 		 * retran and> ack 3
8266 		 */
8267 		return (0);
8268 	}
8269 	/*
8270 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
8271 	 * something we sent.
8272 	 */
8273 	if (tp->t_flags & TF_NEEDSYN) {
8274 		/*
8275 		 * T/TCP: Connection was half-synchronized, and our SYN has
8276 		 * been ACK'd (so connection is now fully synchronized).  Go
8277 		 * to non-starred state, increment snd_una for ACK of SYN,
8278 		 * and check if we can do window scaling.
8279 		 */
8280 		tp->t_flags &= ~TF_NEEDSYN;
8281 		tp->snd_una++;
8282 		/* Do window scaling? */
8283 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
8284 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
8285 			tp->rcv_scale = tp->request_r_scale;
8286 			/* Send window already scaled. */
8287 		}
8288 	}
8289 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8290 	INP_WLOCK_ASSERT(tp->t_inpcb);
8291 
8292 	acked = BYTES_THIS_ACK(tp, th);
8293 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
8294 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
8295 	/*
8296 	 * If we just performed our first retransmit, and the ACK arrives
8297 	 * within our recovery window, then it was a mistake to do the
8298 	 * retransmit in the first place.  Recover our original cwnd and
8299 	 * ssthresh, and proceed to transmit where we left off.
8300 	 */
8301 	if (tp->t_flags & TF_PREVVALID) {
8302 		tp->t_flags &= ~TF_PREVVALID;
8303 		if (tp->t_rxtshift == 1 &&
8304 		    (int)(ticks - tp->t_badrxtwin) < 0)
8305 			rack_cong_signal(tp, th, CC_RTO_ERR);
8306 	}
8307 	if (acked) {
8308 		/* assure we are not backed off */
8309 		tp->t_rxtshift = 0;
8310 		rack->rc_tlp_in_progress = 0;
8311 		rack->r_ctl.rc_tlp_cnt_out = 0;
8312 		/*
8313 		 * If it is the RXT timer we want to
8314 		 * stop it, so we can restart a TLP.
8315 		 */
8316 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
8317 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
8318 #ifdef NETFLIX_HTTP_LOGGING
8319 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
8320 #endif
8321 	}
8322 	/*
8323 	 * If we have a timestamp reply, update smoothed round trip time. If
8324 	 * no timestamp is present but transmit timer is running and timed
8325 	 * sequence number was acked, update smoothed round trip time. Since
8326 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
8327 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
8328 	 * timer.
8329 	 *
8330 	 * Some boxes send broken timestamp replies during the SYN+ACK
8331 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
8332 	 * and blow up the retransmit timer.
8333 	 */
8334 	/*
8335 	 * If all outstanding data is acked, stop retransmit timer and
8336 	 * remember to restart (more output or persist). If there is more
8337 	 * data to be acked, restart retransmit timer, using current
8338 	 * (possibly backed-off) value.
8339 	 */
8340 	if (acked == 0) {
8341 		if (ofia)
8342 			*ofia = ourfinisacked;
8343 		return (0);
8344 	}
8345 	if (rack->r_ctl.rc_early_recovery) {
8346 		if (IN_RECOVERY(tp->t_flags)) {
8347 			if (SEQ_LT(th->th_ack, tp->snd_recover) &&
8348 			    (SEQ_LT(th->th_ack, tp->snd_max))) {
8349 				tcp_rack_partialack(tp, th);
8350 			} else {
8351 				rack_post_recovery(tp, th);
8352 				recovery = 1;
8353 			}
8354 		}
8355 	}
8356 	/*
8357 	 * Let the congestion control algorithm update congestion control
8358 	 * related information. This typically means increasing the
8359 	 * congestion window.
8360 	 */
8361 	rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery);
8362 	SOCKBUF_LOCK(&so->so_snd);
8363 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
8364 	tp->snd_wnd -= acked_amount;
8365 	mfree = sbcut_locked(&so->so_snd, acked_amount);
8366 	if ((sbused(&so->so_snd) == 0) &&
8367 	    (acked > acked_amount) &&
8368 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
8369 	    (tp->t_flags & TF_SENTFIN)) {
8370 		/*
8371 		 * We must be sure our fin
8372 		 * was sent and acked (we can be
8373 		 * in FIN_WAIT_1 without having
8374 		 * sent the fin).
8375 		 */
8376 		ourfinisacked = 1;
8377 	}
8378 	/* NB: sowwakeup_locked() does an implicit unlock. */
8379 	sowwakeup_locked(so);
8380 	m_freem(mfree);
8381 	if (rack->r_ctl.rc_early_recovery == 0) {
8382 		if (IN_RECOVERY(tp->t_flags)) {
8383 			if (SEQ_LT(th->th_ack, tp->snd_recover) &&
8384 			    (SEQ_LT(th->th_ack, tp->snd_max))) {
8385 				tcp_rack_partialack(tp, th);
8386 			} else {
8387 				rack_post_recovery(tp, th);
8388 			}
8389 		}
8390 	}
8391 	tp->snd_una = th->th_ack;
8392 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
8393 		tp->snd_recover = tp->snd_una;
8394 
8395 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
8396 		tp->snd_nxt = tp->snd_una;
8397 	}
8398 	if (under_pacing &&
8399 	    (rack->use_fixed_rate == 0) &&
8400 	    (rack->in_probe_rtt == 0) &&
8401 	    rack->rc_gp_dyn_mul &&
8402 	    rack->rc_always_pace) {
8403 		/* Check if we are dragging bottom */
8404 		rack_check_bottom_drag(tp, rack, so, acked);
8405 	}
8406 	if (tp->snd_una == tp->snd_max) {
8407 		/* Nothing left outstanding */
8408 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
8409 		if (rack->r_ctl.rc_went_idle_time == 0)
8410 			rack->r_ctl.rc_went_idle_time = 1;
8411 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
8412 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8413 			tp->t_acktime = 0;
8414 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
8415 		/* Set need output so persist might get set */
8416 		rack->r_wanted_output = 1;
8417 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
8418 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8419 		    (sbavail(&so->so_snd) == 0) &&
8420 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
8421 			/*
8422 			 * The socket was gone and the
8423 			 * peer sent data, time to
8424 			 * reset him.
8425 			 */
8426 			*ret_val = 1;
8427 			/* tcp_close will kill the inp pre-log the Reset */
8428 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
8429 			tp = tcp_close(tp);
8430 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
8431 			return (1);
8432 
8433 		}
8434 	}
8435 	if (ofia)
8436 		*ofia = ourfinisacked;
8437 	return (0);
8438 }
8439 
8440 static void
8441 rack_collapsed_window(struct tcp_rack *rack)
8442 {
8443 	/*
8444 	 * Now we must walk the
8445 	 * send map and divide the
8446 	 * ones left stranded. These
8447 	 * guys can't cause us to abort
8448 	 * the connection and are really
8449 	 * "unsent". However if a buggy
8450 	 * client actually did keep some
8451 	 * of the data i.e. collapsed the win
8452 	 * and refused to ack and then opened
8453 	 * the win and acked that data. We would
8454 	 * get into an ack war, the simplier
8455 	 * method then of just pretending we
8456 	 * did not send those segments something
8457 	 * won't work.
8458 	 */
8459 	struct rack_sendmap *rsm, *nrsm, fe, *insret;
8460 	tcp_seq max_seq;
8461 
8462 	max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
8463 	memset(&fe, 0, sizeof(fe));
8464 	fe.r_start = max_seq;
8465 	/* Find the first seq past or at maxseq */
8466 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8467 	if (rsm == NULL) {
8468 		/* Nothing to do strange */
8469 		rack->rc_has_collapsed = 0;
8470 		return;
8471 	}
8472 	/*
8473 	 * Now do we need to split at
8474 	 * the collapse point?
8475 	 */
8476 	if (SEQ_GT(max_seq, rsm->r_start)) {
8477 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8478 		if (nrsm == NULL) {
8479 			/* We can't get a rsm, mark all? */
8480 			nrsm = rsm;
8481 			goto no_split;
8482 		}
8483 		/* Clone it */
8484 		rack_clone_rsm(rack, nrsm, rsm, max_seq);
8485 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8486 #ifdef INVARIANTS
8487 		if (insret != NULL) {
8488 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8489 			      nrsm, insret, rack, rsm);
8490 		}
8491 #endif
8492 		if (rsm->r_in_tmap) {
8493 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8494 			nrsm->r_in_tmap = 1;
8495 		}
8496 		/*
8497 		 * Set in the new RSM as the
8498 		 * collapsed starting point
8499 		 */
8500 		rsm = nrsm;
8501 	}
8502 no_split:
8503 	counter_u64_add(rack_collapsed_win, 1);
8504 	RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
8505 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
8506 		rack->rc_has_collapsed = 1;
8507 	}
8508 }
8509 
8510 static void
8511 rack_un_collapse_window(struct tcp_rack *rack)
8512 {
8513 	struct rack_sendmap *rsm;
8514 
8515 	RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
8516 		if (rsm->r_flags & RACK_RWND_COLLAPSED)
8517 			rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8518 		else
8519 			break;
8520 	}
8521 	rack->rc_has_collapsed = 0;
8522 }
8523 
8524 static void
8525 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
8526 			int32_t tlen, int32_t tfo_syn)
8527 {
8528 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
8529 		if (rack->rc_dack_mode &&
8530 		    (tlen > 500) &&
8531 		    (rack->rc_dack_toggle == 1)) {
8532 			goto no_delayed_ack;
8533 		}
8534 		rack_timer_cancel(tp, rack,
8535 				  rack->r_ctl.rc_rcvtime, __LINE__);
8536 		tp->t_flags |= TF_DELACK;
8537 	} else {
8538 no_delayed_ack:
8539 		rack->r_wanted_output = 1;
8540 		tp->t_flags |= TF_ACKNOW;
8541 		if (rack->rc_dack_mode) {
8542 			if (tp->t_flags & TF_DELACK)
8543 				rack->rc_dack_toggle = 1;
8544 			else
8545 				rack->rc_dack_toggle = 0;
8546 		}
8547 	}
8548 }
8549 /*
8550  * Return value of 1, the TCB is unlocked and most
8551  * likely gone, return value of 0, the TCP is still
8552  * locked.
8553  */
8554 static int
8555 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8556     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8557     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8558 {
8559 	/*
8560 	 * Update window information. Don't look at window if no ACK: TAC's
8561 	 * send garbage on first SYN.
8562 	 */
8563 	int32_t nsegs;
8564 	int32_t tfo_syn;
8565 	struct tcp_rack *rack;
8566 
8567 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8568 	INP_WLOCK_ASSERT(tp->t_inpcb);
8569 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8570 	if ((thflags & TH_ACK) &&
8571 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8572 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8573 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8574 		/* keep track of pure window updates */
8575 		if (tlen == 0 &&
8576 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8577 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
8578 		tp->snd_wnd = tiwin;
8579 		tp->snd_wl1 = th->th_seq;
8580 		tp->snd_wl2 = th->th_ack;
8581 		if (tp->snd_wnd > tp->max_sndwnd)
8582 			tp->max_sndwnd = tp->snd_wnd;
8583 		rack->r_wanted_output = 1;
8584 	} else if (thflags & TH_ACK) {
8585 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8586 			tp->snd_wnd = tiwin;
8587 			tp->snd_wl1 = th->th_seq;
8588 			tp->snd_wl2 = th->th_ack;
8589 		}
8590 	}
8591 	if (tp->snd_wnd < ctf_outstanding(tp))
8592 		/* The peer collapsed the window */
8593 		rack_collapsed_window(rack);
8594 	else if (rack->rc_has_collapsed)
8595 		rack_un_collapse_window(rack);
8596 	/* Was persist timer active and now we have window space? */
8597 	if ((rack->rc_in_persist != 0) &&
8598 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
8599 				rack->r_ctl.rc_pace_min_segs))) {
8600 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
8601 		tp->snd_nxt = tp->snd_max;
8602 		/* Make sure we output to start the timer */
8603 		rack->r_wanted_output = 1;
8604 	}
8605 	/* Do we enter persists? */
8606 	if ((rack->rc_in_persist == 0) &&
8607 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
8608 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
8609 	    (tp->snd_max == tp->snd_una) &&
8610 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8611 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8612 		/*
8613 		 * Here the rwnd is less than
8614 		 * the pacing size, we are established,
8615 		 * nothing is outstanding, and there is
8616 		 * data to send. Enter persists.
8617 		 */
8618 		tp->snd_nxt = tp->snd_una;
8619 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
8620 	}
8621 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8622 		m_freem(m);
8623 		return (0);
8624 	}
8625 	/*
8626 	 * don't process the URG bit, ignore them drag
8627 	 * along the up.
8628 	 */
8629 	tp->rcv_up = tp->rcv_nxt;
8630 	INP_WLOCK_ASSERT(tp->t_inpcb);
8631 
8632 	/*
8633 	 * Process the segment text, merging it into the TCP sequencing
8634 	 * queue, and arranging for acknowledgment of receipt if necessary.
8635 	 * This process logically involves adjusting tp->rcv_wnd as data is
8636 	 * presented to the user (this happens in tcp_usrreq.c, case
8637 	 * PRU_RCVD).  If a FIN has already been received on this connection
8638 	 * then we just ignore the text.
8639 	 */
8640 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8641 		   IS_FASTOPEN(tp->t_flags));
8642 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
8643 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8644 		tcp_seq save_start = th->th_seq;
8645 		tcp_seq save_rnxt  = tp->rcv_nxt;
8646 		int     save_tlen  = tlen;
8647 
8648 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8649 		/*
8650 		 * Insert segment which includes th into TCP reassembly
8651 		 * queue with control block tp.  Set thflags to whether
8652 		 * reassembly now includes a segment with FIN.  This handles
8653 		 * the common case inline (segment is the next to be
8654 		 * received on an established connection, and the queue is
8655 		 * empty), avoiding linkage into and removal from the queue
8656 		 * and repetition of various conversions. Set DELACK for
8657 		 * segments received in order, but ack immediately when
8658 		 * segments are out of order (so fast retransmit can work).
8659 		 */
8660 		if (th->th_seq == tp->rcv_nxt &&
8661 		    SEGQ_EMPTY(tp) &&
8662 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
8663 		    tfo_syn)) {
8664 #ifdef NETFLIX_SB_LIMITS
8665 			u_int mcnt, appended;
8666 
8667 			if (so->so_rcv.sb_shlim) {
8668 				mcnt = m_memcnt(m);
8669 				appended = 0;
8670 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8671 				    CFO_NOSLEEP, NULL) == false) {
8672 					counter_u64_add(tcp_sb_shlim_fails, 1);
8673 					m_freem(m);
8674 					return (0);
8675 				}
8676 			}
8677 #endif
8678 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
8679 			tp->rcv_nxt += tlen;
8680 			if (tlen &&
8681 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8682 			    (tp->t_fbyte_in == 0)) {
8683 				tp->t_fbyte_in = ticks;
8684 				if (tp->t_fbyte_in == 0)
8685 					tp->t_fbyte_in = 1;
8686 				if (tp->t_fbyte_out && tp->t_fbyte_in)
8687 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8688 			}
8689 			thflags = th->th_flags & TH_FIN;
8690 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
8691 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8692 			SOCKBUF_LOCK(&so->so_rcv);
8693 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8694 				m_freem(m);
8695 			} else
8696 #ifdef NETFLIX_SB_LIMITS
8697 				appended =
8698 #endif
8699 					sbappendstream_locked(&so->so_rcv, m, 0);
8700 			/* NB: sorwakeup_locked() does an implicit unlock. */
8701 			sorwakeup_locked(so);
8702 #ifdef NETFLIX_SB_LIMITS
8703 			if (so->so_rcv.sb_shlim && appended != mcnt)
8704 				counter_fo_release(so->so_rcv.sb_shlim,
8705 				    mcnt - appended);
8706 #endif
8707 		} else {
8708 			/*
8709 			 * XXX: Due to the header drop above "th" is
8710 			 * theoretically invalid by now.  Fortunately
8711 			 * m_adj() doesn't actually frees any mbufs when
8712 			 * trimming from the head.
8713 			 */
8714 			tcp_seq temp = save_start;
8715 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
8716 			tp->t_flags |= TF_ACKNOW;
8717 		}
8718                 if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0)) {
8719                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8720                                 /*
8721                                  * DSACK actually handled in the fastpath
8722                                  * above.
8723                                  */
8724 				RACK_OPTS_INC(tcp_sack_path_1);
8725                                 tcp_update_sack_list(tp, save_start,
8726                                     save_start + save_tlen);
8727                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8728                                 if ((tp->rcv_numsacks >= 1) &&
8729                                     (tp->sackblks[0].end == save_start)) {
8730                                         /*
8731                                          * Partial overlap, recorded at todrop
8732                                          * above.
8733                                          */
8734 					RACK_OPTS_INC(tcp_sack_path_2a);
8735                                         tcp_update_sack_list(tp,
8736                                             tp->sackblks[0].start,
8737                                             tp->sackblks[0].end);
8738                                 } else {
8739 					RACK_OPTS_INC(tcp_sack_path_2b);
8740                                         tcp_update_dsack_list(tp, save_start,
8741                                             save_start + save_tlen);
8742                                 }
8743                         } else if (tlen >= save_tlen) {
8744                                 /* Update of sackblks. */
8745 				RACK_OPTS_INC(tcp_sack_path_3);
8746                                 tcp_update_dsack_list(tp, save_start,
8747                                     save_start + save_tlen);
8748                         } else if (tlen > 0) {
8749 				RACK_OPTS_INC(tcp_sack_path_4);
8750                                 tcp_update_dsack_list(tp, save_start,
8751                                     save_start + tlen);
8752                         }
8753                 }
8754 	} else {
8755 		m_freem(m);
8756 		thflags &= ~TH_FIN;
8757 	}
8758 
8759 	/*
8760 	 * If FIN is received ACK the FIN and let the user know that the
8761 	 * connection is closing.
8762 	 */
8763 	if (thflags & TH_FIN) {
8764 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8765 			socantrcvmore(so);
8766 			/*
8767 			 * If connection is half-synchronized (ie NEEDSYN
8768 			 * flag on) then delay ACK, so it may be piggybacked
8769 			 * when SYN is sent. Otherwise, since we received a
8770 			 * FIN then no more input can be expected, send ACK
8771 			 * now.
8772 			 */
8773 			if (tp->t_flags & TF_NEEDSYN) {
8774 				rack_timer_cancel(tp, rack,
8775 				    rack->r_ctl.rc_rcvtime, __LINE__);
8776 				tp->t_flags |= TF_DELACK;
8777 			} else {
8778 				tp->t_flags |= TF_ACKNOW;
8779 			}
8780 			tp->rcv_nxt++;
8781 		}
8782 		switch (tp->t_state) {
8783 
8784 			/*
8785 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
8786 			 * CLOSE_WAIT state.
8787 			 */
8788 		case TCPS_SYN_RECEIVED:
8789 			tp->t_starttime = ticks;
8790 			/* FALLTHROUGH */
8791 		case TCPS_ESTABLISHED:
8792 			rack_timer_cancel(tp, rack,
8793 			    rack->r_ctl.rc_rcvtime, __LINE__);
8794 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
8795 			break;
8796 
8797 			/*
8798 			 * If still in FIN_WAIT_1 STATE FIN has not been
8799 			 * acked so enter the CLOSING state.
8800 			 */
8801 		case TCPS_FIN_WAIT_1:
8802 			rack_timer_cancel(tp, rack,
8803 			    rack->r_ctl.rc_rcvtime, __LINE__);
8804 			tcp_state_change(tp, TCPS_CLOSING);
8805 			break;
8806 
8807 			/*
8808 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
8809 			 * starting the time-wait timer, turning off the
8810 			 * other standard timers.
8811 			 */
8812 		case TCPS_FIN_WAIT_2:
8813 			rack_timer_cancel(tp, rack,
8814 			    rack->r_ctl.rc_rcvtime, __LINE__);
8815 			tcp_twstart(tp);
8816 			return (1);
8817 		}
8818 	}
8819 	/*
8820 	 * Return any desired output.
8821 	 */
8822 	if ((tp->t_flags & TF_ACKNOW) ||
8823 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
8824 		rack->r_wanted_output = 1;
8825 	}
8826 	INP_WLOCK_ASSERT(tp->t_inpcb);
8827 	return (0);
8828 }
8829 
8830 /*
8831  * Here nothing is really faster, its just that we
8832  * have broken out the fast-data path also just like
8833  * the fast-ack.
8834  */
8835 static int
8836 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8837     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8838     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
8839 {
8840 	int32_t nsegs;
8841 	int32_t newsize = 0;	/* automatic sockbuf scaling */
8842 	struct tcp_rack *rack;
8843 #ifdef NETFLIX_SB_LIMITS
8844 	u_int mcnt, appended;
8845 #endif
8846 #ifdef TCPDEBUG
8847 	/*
8848 	 * The size of tcp_saveipgen must be the size of the max ip header,
8849 	 * now IPv6.
8850 	 */
8851 	u_char tcp_saveipgen[IP6_HDR_LEN];
8852 	struct tcphdr tcp_savetcp;
8853 	short ostate = 0;
8854 
8855 #endif
8856 	/*
8857 	 * If last ACK falls within this segment's sequence numbers, record
8858 	 * the timestamp. NOTE that the test is modified according to the
8859 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8860 	 */
8861 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
8862 		return (0);
8863 	}
8864 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
8865 		return (0);
8866 	}
8867 	if (tiwin && tiwin != tp->snd_wnd) {
8868 		return (0);
8869 	}
8870 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8871 		return (0);
8872 	}
8873 	if (__predict_false((to->to_flags & TOF_TS) &&
8874 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8875 		return (0);
8876 	}
8877 	if (__predict_false((th->th_ack != tp->snd_una))) {
8878 		return (0);
8879 	}
8880 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8881 		return (0);
8882 	}
8883 	if ((to->to_flags & TOF_TS) != 0 &&
8884 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8885 		tp->ts_recent_age = tcp_ts_getticks();
8886 		tp->ts_recent = to->to_tsval;
8887 	}
8888 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8889 	/*
8890 	 * This is a pure, in-sequence data packet with nothing on the
8891 	 * reassembly queue and we have enough buffer space to take it.
8892 	 */
8893 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
8894 
8895 #ifdef NETFLIX_SB_LIMITS
8896 	if (so->so_rcv.sb_shlim) {
8897 		mcnt = m_memcnt(m);
8898 		appended = 0;
8899 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8900 		    CFO_NOSLEEP, NULL) == false) {
8901 			counter_u64_add(tcp_sb_shlim_fails, 1);
8902 			m_freem(m);
8903 			return (1);
8904 		}
8905 	}
8906 #endif
8907 	/* Clean receiver SACK report if present */
8908 	if (tp->rcv_numsacks)
8909 		tcp_clean_sackreport(tp);
8910 	KMOD_TCPSTAT_INC(tcps_preddat);
8911 	tp->rcv_nxt += tlen;
8912 	if (tlen &&
8913 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8914 	    (tp->t_fbyte_in == 0)) {
8915 		tp->t_fbyte_in = ticks;
8916 		if (tp->t_fbyte_in == 0)
8917 			tp->t_fbyte_in = 1;
8918 		if (tp->t_fbyte_out && tp->t_fbyte_in)
8919 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8920 	}
8921 	/*
8922 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8923 	 */
8924 	tp->snd_wl1 = th->th_seq;
8925 	/*
8926 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8927 	 */
8928 	tp->rcv_up = tp->rcv_nxt;
8929 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
8930 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8931 #ifdef TCPDEBUG
8932 	if (so->so_options & SO_DEBUG)
8933 		tcp_trace(TA_INPUT, ostate, tp,
8934 		    (void *)tcp_saveipgen, &tcp_savetcp, 0);
8935 #endif
8936 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8937 
8938 	/* Add data to socket buffer. */
8939 	SOCKBUF_LOCK(&so->so_rcv);
8940 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8941 		m_freem(m);
8942 	} else {
8943 		/*
8944 		 * Set new socket buffer size. Give up when limit is
8945 		 * reached.
8946 		 */
8947 		if (newsize)
8948 			if (!sbreserve_locked(&so->so_rcv,
8949 			    newsize, so, NULL))
8950 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8951 		m_adj(m, drop_hdrlen);	/* delayed header drop */
8952 #ifdef NETFLIX_SB_LIMITS
8953 		appended =
8954 #endif
8955 			sbappendstream_locked(&so->so_rcv, m, 0);
8956 		ctf_calc_rwin(so, tp);
8957 	}
8958 	/* NB: sorwakeup_locked() does an implicit unlock. */
8959 	sorwakeup_locked(so);
8960 #ifdef NETFLIX_SB_LIMITS
8961 	if (so->so_rcv.sb_shlim && mcnt != appended)
8962 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8963 #endif
8964 	rack_handle_delayed_ack(tp, rack, tlen, 0);
8965 	if (tp->snd_una == tp->snd_max)
8966 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
8967 	return (1);
8968 }
8969 
8970 /*
8971  * This subfunction is used to try to highly optimize the
8972  * fast path. We again allow window updates that are
8973  * in sequence to remain in the fast-path. We also add
8974  * in the __predict's to attempt to help the compiler.
8975  * Note that if we return a 0, then we can *not* process
8976  * it and the caller should push the packet into the
8977  * slow-path.
8978  */
8979 static int
8980 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8981     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8982     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
8983 {
8984 	int32_t acked;
8985 	int32_t nsegs;
8986 #ifdef TCPDEBUG
8987 	/*
8988 	 * The size of tcp_saveipgen must be the size of the max ip header,
8989 	 * now IPv6.
8990 	 */
8991 	u_char tcp_saveipgen[IP6_HDR_LEN];
8992 	struct tcphdr tcp_savetcp;
8993 	short ostate = 0;
8994 #endif
8995 	int32_t under_pacing = 0;
8996 	struct tcp_rack *rack;
8997 
8998 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8999 		/* Old ack, behind (or duplicate to) the last one rcv'd */
9000 		return (0);
9001 	}
9002 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
9003 		/* Above what we have sent? */
9004 		return (0);
9005 	}
9006 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
9007 		/* We are retransmitting */
9008 		return (0);
9009 	}
9010 	if (__predict_false(tiwin == 0)) {
9011 		/* zero window */
9012 		return (0);
9013 	}
9014 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
9015 		/* We need a SYN or a FIN, unlikely.. */
9016 		return (0);
9017 	}
9018 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
9019 		/* Timestamp is behind .. old ack with seq wrap? */
9020 		return (0);
9021 	}
9022 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
9023 		/* Still recovering */
9024 		return (0);
9025 	}
9026 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9027 	if (rack->r_ctl.rc_sacked) {
9028 		/* We have sack holes on our scoreboard */
9029 		return (0);
9030 	}
9031 	/* Ok if we reach here, we can process a fast-ack */
9032 	if (rack->rc_gp_filled &&
9033 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
9034 		under_pacing = 1;
9035 	}
9036 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
9037 	rack_log_ack(tp, to, th);
9038 	/* Did the window get updated? */
9039 	if (tiwin != tp->snd_wnd) {
9040 		tp->snd_wnd = tiwin;
9041 		tp->snd_wl1 = th->th_seq;
9042 		if (tp->snd_wnd > tp->max_sndwnd)
9043 			tp->max_sndwnd = tp->snd_wnd;
9044 	}
9045 	/* Do we exit persists? */
9046 	if ((rack->rc_in_persist != 0) &&
9047 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
9048 			       rack->r_ctl.rc_pace_min_segs))) {
9049 		rack_exit_persist(tp, rack, cts);
9050 	}
9051 	/* Do we enter persists? */
9052 	if ((rack->rc_in_persist == 0) &&
9053 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
9054 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
9055 	    (tp->snd_max == tp->snd_una) &&
9056 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
9057 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
9058 		/*
9059 		 * Here the rwnd is less than
9060 		 * the pacing size, we are established,
9061 		 * nothing is outstanding, and there is
9062 		 * data to send. Enter persists.
9063 		 */
9064 		tp->snd_nxt = tp->snd_una;
9065 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
9066 	}
9067 	/*
9068 	 * If last ACK falls within this segment's sequence numbers, record
9069 	 * the timestamp. NOTE that the test is modified according to the
9070 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
9071 	 */
9072 	if ((to->to_flags & TOF_TS) != 0 &&
9073 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
9074 		tp->ts_recent_age = tcp_ts_getticks();
9075 		tp->ts_recent = to->to_tsval;
9076 	}
9077 	/*
9078 	 * This is a pure ack for outstanding data.
9079 	 */
9080 	KMOD_TCPSTAT_INC(tcps_predack);
9081 
9082 	/*
9083 	 * "bad retransmit" recovery.
9084 	 */
9085 	if (tp->t_flags & TF_PREVVALID) {
9086 		tp->t_flags &= ~TF_PREVVALID;
9087 		if (tp->t_rxtshift == 1 &&
9088 		    (int)(ticks - tp->t_badrxtwin) < 0)
9089 			rack_cong_signal(tp, th, CC_RTO_ERR);
9090 	}
9091 	/*
9092 	 * Recalculate the transmit timer / rtt.
9093 	 *
9094 	 * Some boxes send broken timestamp replies during the SYN+ACK
9095 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
9096 	 * and blow up the retransmit timer.
9097 	 */
9098 	acked = BYTES_THIS_ACK(tp, th);
9099 
9100 #ifdef TCP_HHOOK
9101 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
9102 	hhook_run_tcp_est_in(tp, th, to);
9103 #endif
9104 
9105 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
9106 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
9107 	sbdrop(&so->so_snd, acked);
9108 	if (acked) {
9109 		/* assure we are not backed off */
9110 		tp->t_rxtshift = 0;
9111 		rack->rc_tlp_in_progress = 0;
9112 		rack->r_ctl.rc_tlp_cnt_out = 0;
9113 		/*
9114 		 * If it is the RXT timer we want to
9115 		 * stop it, so we can restart a TLP.
9116 		 */
9117 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
9118 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9119 #ifdef NETFLIX_HTTP_LOGGING
9120 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
9121 #endif
9122 	}
9123 	/*
9124 	 * Let the congestion control algorithm update congestion control
9125 	 * related information. This typically means increasing the
9126 	 * congestion window.
9127 	 */
9128 	rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0);
9129 
9130 	tp->snd_una = th->th_ack;
9131 	if (tp->snd_wnd < ctf_outstanding(tp)) {
9132 		/* The peer collapsed the window */
9133 		rack_collapsed_window(rack);
9134 	} else if (rack->rc_has_collapsed)
9135 		rack_un_collapse_window(rack);
9136 
9137 	/*
9138 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
9139 	 */
9140 	tp->snd_wl2 = th->th_ack;
9141 	tp->t_dupacks = 0;
9142 	m_freem(m);
9143 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
9144 
9145 	/*
9146 	 * If all outstanding data are acked, stop retransmit timer,
9147 	 * otherwise restart timer using current (possibly backed-off)
9148 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
9149 	 * If data are ready to send, let tcp_output decide between more
9150 	 * output or persist.
9151 	 */
9152 #ifdef TCPDEBUG
9153 	if (so->so_options & SO_DEBUG)
9154 		tcp_trace(TA_INPUT, ostate, tp,
9155 		    (void *)tcp_saveipgen,
9156 		    &tcp_savetcp, 0);
9157 #endif
9158 	if (under_pacing &&
9159 	    (rack->use_fixed_rate == 0) &&
9160 	    (rack->in_probe_rtt == 0) &&
9161 	    rack->rc_gp_dyn_mul &&
9162 	    rack->rc_always_pace) {
9163 		/* Check if we are dragging bottom */
9164 		rack_check_bottom_drag(tp, rack, so, acked);
9165 	}
9166 	if (tp->snd_una == tp->snd_max) {
9167 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
9168 		if (rack->r_ctl.rc_went_idle_time == 0)
9169 			rack->r_ctl.rc_went_idle_time = 1;
9170 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
9171 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
9172 			tp->t_acktime = 0;
9173 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9174 	}
9175 	/* Wake up the socket if we have room to write more */
9176 	sowwakeup(so);
9177 	if (sbavail(&so->so_snd)) {
9178 		rack->r_wanted_output = 1;
9179 	}
9180 	return (1);
9181 }
9182 
9183 /*
9184  * Return value of 1, the TCB is unlocked and most
9185  * likely gone, return value of 0, the TCP is still
9186  * locked.
9187  */
9188 static int
9189 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
9190     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9191     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9192 {
9193 	int32_t ret_val = 0;
9194 	int32_t todrop;
9195 	int32_t ourfinisacked = 0;
9196 	struct tcp_rack *rack;
9197 
9198 	ctf_calc_rwin(so, tp);
9199 	/*
9200 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
9201 	 * SYN, drop the input. if seg contains a RST, then drop the
9202 	 * connection. if seg does not contain SYN, then drop it. Otherwise
9203 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
9204 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
9205 	 * contains an ECE and ECN support is enabled, the stream is ECN
9206 	 * capable. if SYN has been acked change to ESTABLISHED else
9207 	 * SYN_RCVD state arrange for segment to be acked (eventually)
9208 	 * continue processing rest of data/controls.
9209 	 */
9210 	if ((thflags & TH_ACK) &&
9211 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
9212 	    SEQ_GT(th->th_ack, tp->snd_max))) {
9213 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9214 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9215 		return (1);
9216 	}
9217 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
9218 		TCP_PROBE5(connect__refused, NULL, tp,
9219 		    mtod(m, const char *), tp, th);
9220 		tp = tcp_drop(tp, ECONNREFUSED);
9221 		ctf_do_drop(m, tp);
9222 		return (1);
9223 	}
9224 	if (thflags & TH_RST) {
9225 		ctf_do_drop(m, tp);
9226 		return (1);
9227 	}
9228 	if (!(thflags & TH_SYN)) {
9229 		ctf_do_drop(m, tp);
9230 		return (1);
9231 	}
9232 	tp->irs = th->th_seq;
9233 	tcp_rcvseqinit(tp);
9234 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9235 	if (thflags & TH_ACK) {
9236 		int tfo_partial = 0;
9237 
9238 		KMOD_TCPSTAT_INC(tcps_connects);
9239 		soisconnected(so);
9240 #ifdef MAC
9241 		mac_socketpeer_set_from_mbuf(m, so);
9242 #endif
9243 		/* Do window scaling on this connection? */
9244 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9245 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9246 			tp->rcv_scale = tp->request_r_scale;
9247 		}
9248 		tp->rcv_adv += min(tp->rcv_wnd,
9249 		    TCP_MAXWIN << tp->rcv_scale);
9250 		/*
9251 		 * If not all the data that was sent in the TFO SYN
9252 		 * has been acked, resend the remainder right away.
9253 		 */
9254 		if (IS_FASTOPEN(tp->t_flags) &&
9255 		    (tp->snd_una != tp->snd_max)) {
9256 			tp->snd_nxt = th->th_ack;
9257 			tfo_partial = 1;
9258 		}
9259 		/*
9260 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
9261 		 * will be turned on later.
9262 		 */
9263 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
9264 			rack_timer_cancel(tp, rack,
9265 					  rack->r_ctl.rc_rcvtime, __LINE__);
9266 			tp->t_flags |= TF_DELACK;
9267 		} else {
9268 			rack->r_wanted_output = 1;
9269 			tp->t_flags |= TF_ACKNOW;
9270 			rack->rc_dack_toggle = 0;
9271 		}
9272 		if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
9273 		    (V_tcp_do_ecn == 1)) {
9274 			tp->t_flags2 |= TF2_ECN_PERMIT;
9275 			KMOD_TCPSTAT_INC(tcps_ecn_shs);
9276 		}
9277 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
9278 			/*
9279 			 * We advance snd_una for the
9280 			 * fast open case. If th_ack is
9281 			 * acknowledging data beyond
9282 			 * snd_una we can't just call
9283 			 * ack-processing since the
9284 			 * data stream in our send-map
9285 			 * will start at snd_una + 1 (one
9286 			 * beyond the SYN). If its just
9287 			 * equal we don't need to do that
9288 			 * and there is no send_map.
9289 			 */
9290 			tp->snd_una++;
9291 		}
9292 		/*
9293 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
9294 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
9295 		 */
9296 		tp->t_starttime = ticks;
9297 		if (tp->t_flags & TF_NEEDFIN) {
9298 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
9299 			tp->t_flags &= ~TF_NEEDFIN;
9300 			thflags &= ~TH_SYN;
9301 		} else {
9302 			tcp_state_change(tp, TCPS_ESTABLISHED);
9303 			TCP_PROBE5(connect__established, NULL, tp,
9304 			    mtod(m, const char *), tp, th);
9305 			rack_cc_conn_init(tp);
9306 		}
9307 	} else {
9308 		/*
9309 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
9310 		 * open.  If segment contains CC option and there is a
9311 		 * cached CC, apply TAO test. If it succeeds, connection is *
9312 		 * half-synchronized. Otherwise, do 3-way handshake:
9313 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
9314 		 * there was no CC option, clear cached CC value.
9315 		 */
9316 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
9317 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
9318 	}
9319 	INP_WLOCK_ASSERT(tp->t_inpcb);
9320 	/*
9321 	 * Advance th->th_seq to correspond to first data byte. If data,
9322 	 * trim to stay within window, dropping FIN if necessary.
9323 	 */
9324 	th->th_seq++;
9325 	if (tlen > tp->rcv_wnd) {
9326 		todrop = tlen - tp->rcv_wnd;
9327 		m_adj(m, -todrop);
9328 		tlen = tp->rcv_wnd;
9329 		thflags &= ~TH_FIN;
9330 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
9331 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
9332 	}
9333 	tp->snd_wl1 = th->th_seq - 1;
9334 	tp->rcv_up = th->th_seq;
9335 	/*
9336 	 * Client side of transaction: already sent SYN and data. If the
9337 	 * remote host used T/TCP to validate the SYN, our data will be
9338 	 * ACK'd; if so, enter normal data segment processing in the middle
9339 	 * of step 5, ack processing. Otherwise, goto step 6.
9340 	 */
9341 	if (thflags & TH_ACK) {
9342 		/* For syn-sent we need to possibly update the rtt */
9343 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
9344 			uint32_t t;
9345 
9346 			t = tcp_ts_getticks() - to->to_tsecr;
9347 			if (!tp->t_rttlow || tp->t_rttlow > t)
9348 				tp->t_rttlow = t;
9349 			tcp_rack_xmit_timer(rack, t + 1, 1, (t * HPTS_USEC_IN_MSEC), 0, NULL, 2);
9350 			tcp_rack_xmit_timer_commit(rack, tp);
9351 		}
9352 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
9353 			return (ret_val);
9354 		/* We may have changed to FIN_WAIT_1 above */
9355 		if (tp->t_state == TCPS_FIN_WAIT_1) {
9356 			/*
9357 			 * In FIN_WAIT_1 STATE in addition to the processing
9358 			 * for the ESTABLISHED state if our FIN is now
9359 			 * acknowledged then enter FIN_WAIT_2.
9360 			 */
9361 			if (ourfinisacked) {
9362 				/*
9363 				 * If we can't receive any more data, then
9364 				 * closing user can proceed. Starting the
9365 				 * timer is contrary to the specification,
9366 				 * but if we don't get a FIN we'll hang
9367 				 * forever.
9368 				 *
9369 				 * XXXjl: we should release the tp also, and
9370 				 * use a compressed state.
9371 				 */
9372 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9373 					soisdisconnected(so);
9374 					tcp_timer_activate(tp, TT_2MSL,
9375 					    (tcp_fast_finwait2_recycle ?
9376 					    tcp_finwait2_timeout :
9377 					    TP_MAXIDLE(tp)));
9378 				}
9379 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
9380 			}
9381 		}
9382 	}
9383 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9384 	   tiwin, thflags, nxt_pkt));
9385 }
9386 
9387 /*
9388  * Return value of 1, the TCB is unlocked and most
9389  * likely gone, return value of 0, the TCP is still
9390  * locked.
9391  */
9392 static int
9393 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
9394     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9395     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9396 {
9397 	struct tcp_rack *rack;
9398 	int32_t ret_val = 0;
9399 	int32_t ourfinisacked = 0;
9400 
9401 	ctf_calc_rwin(so, tp);
9402 	if ((thflags & TH_ACK) &&
9403 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
9404 	    SEQ_GT(th->th_ack, tp->snd_max))) {
9405 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9406 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9407 		return (1);
9408 	}
9409 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9410 	if (IS_FASTOPEN(tp->t_flags)) {
9411 		/*
9412 		 * When a TFO connection is in SYN_RECEIVED, the
9413 		 * only valid packets are the initial SYN, a
9414 		 * retransmit/copy of the initial SYN (possibly with
9415 		 * a subset of the original data), a valid ACK, a
9416 		 * FIN, or a RST.
9417 		 */
9418 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
9419 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9420 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9421 			return (1);
9422 		} else if (thflags & TH_SYN) {
9423 			/* non-initial SYN is ignored */
9424 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
9425 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
9426 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
9427 				ctf_do_drop(m, NULL);
9428 				return (0);
9429 			}
9430 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
9431 			ctf_do_drop(m, NULL);
9432 			return (0);
9433 		}
9434 	}
9435 	if ((thflags & TH_RST) ||
9436 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9437 		return (ctf_process_rst(m, th, so, tp));
9438 	/*
9439 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9440 	 * it's less than ts_recent, drop it.
9441 	 */
9442 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9443 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9444 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9445 			return (ret_val);
9446 	}
9447 	/*
9448 	 * In the SYN-RECEIVED state, validate that the packet belongs to
9449 	 * this connection before trimming the data to fit the receive
9450 	 * window.  Check the sequence number versus IRS since we know the
9451 	 * sequence numbers haven't wrapped.  This is a partial fix for the
9452 	 * "LAND" DoS attack.
9453 	 */
9454 	if (SEQ_LT(th->th_seq, tp->irs)) {
9455 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9456 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9457 		return (1);
9458 	}
9459 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9460 		return (ret_val);
9461 	}
9462 	/*
9463 	 * If last ACK falls within this segment's sequence numbers, record
9464 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9465 	 * from the latest proposal of the tcplw@cray.com list (Braden
9466 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9467 	 * with our earlier PAWS tests, so this check should be solely
9468 	 * predicated on the sequence space of this segment. 3) That we
9469 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9470 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9471 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9472 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9473 	 * p.869. In such cases, we can still calculate the RTT correctly
9474 	 * when RCV.NXT == Last.ACK.Sent.
9475 	 */
9476 	if ((to->to_flags & TOF_TS) != 0 &&
9477 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9478 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9479 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9480 		tp->ts_recent_age = tcp_ts_getticks();
9481 		tp->ts_recent = to->to_tsval;
9482 	}
9483 	tp->snd_wnd = tiwin;
9484 	/*
9485 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9486 	 * is on (half-synchronized state), then queue data for later
9487 	 * processing; else drop segment and return.
9488 	 */
9489 	if ((thflags & TH_ACK) == 0) {
9490 		if (IS_FASTOPEN(tp->t_flags)) {
9491 			rack_cc_conn_init(tp);
9492 		}
9493 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9494 		    tiwin, thflags, nxt_pkt));
9495 	}
9496 	KMOD_TCPSTAT_INC(tcps_connects);
9497 	soisconnected(so);
9498 	/* Do window scaling? */
9499 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9500 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9501 		tp->rcv_scale = tp->request_r_scale;
9502 	}
9503 	/*
9504 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
9505 	 * FIN-WAIT-1
9506 	 */
9507 	tp->t_starttime = ticks;
9508 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
9509 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9510 		tp->t_tfo_pending = NULL;
9511 	}
9512 	if (tp->t_flags & TF_NEEDFIN) {
9513 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
9514 		tp->t_flags &= ~TF_NEEDFIN;
9515 	} else {
9516 		tcp_state_change(tp, TCPS_ESTABLISHED);
9517 		TCP_PROBE5(accept__established, NULL, tp,
9518 		    mtod(m, const char *), tp, th);
9519 		/*
9520 		 * TFO connections call cc_conn_init() during SYN
9521 		 * processing.  Calling it again here for such connections
9522 		 * is not harmless as it would undo the snd_cwnd reduction
9523 		 * that occurs when a TFO SYN|ACK is retransmitted.
9524 		 */
9525 		if (!IS_FASTOPEN(tp->t_flags))
9526 			rack_cc_conn_init(tp);
9527 	}
9528 	/*
9529 	 * Account for the ACK of our SYN prior to
9530 	 * regular ACK processing below, except for
9531 	 * simultaneous SYN, which is handled later.
9532 	 */
9533 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
9534 		tp->snd_una++;
9535 	/*
9536 	 * If segment contains data or ACK, will call tcp_reass() later; if
9537 	 * not, do so now to pass queued data to user.
9538 	 */
9539 	if (tlen == 0 && (thflags & TH_FIN) == 0)
9540 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9541 		    (struct mbuf *)0);
9542 	tp->snd_wl1 = th->th_seq - 1;
9543 	/* For syn-recv we need to possibly update the rtt */
9544 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
9545 		uint32_t t;
9546 
9547 		t = tcp_ts_getticks() - to->to_tsecr;
9548 		if (!tp->t_rttlow || tp->t_rttlow > t)
9549 			tp->t_rttlow = t;
9550 		tcp_rack_xmit_timer(rack, t + 1, 1, (t * HPTS_USEC_IN_MSEC), 0, NULL, 2);
9551 		tcp_rack_xmit_timer_commit(rack, tp);
9552 	}
9553 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9554 		return (ret_val);
9555 	}
9556 	if (tp->t_state == TCPS_FIN_WAIT_1) {
9557 		/* We could have went to FIN_WAIT_1 (or EST) above */
9558 		/*
9559 		 * In FIN_WAIT_1 STATE in addition to the processing for the
9560 		 * ESTABLISHED state if our FIN is now acknowledged then
9561 		 * enter FIN_WAIT_2.
9562 		 */
9563 		if (ourfinisacked) {
9564 			/*
9565 			 * If we can't receive any more data, then closing
9566 			 * user can proceed. Starting the timer is contrary
9567 			 * to the specification, but if we don't get a FIN
9568 			 * we'll hang forever.
9569 			 *
9570 			 * XXXjl: we should release the tp also, and use a
9571 			 * compressed state.
9572 			 */
9573 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9574 				soisdisconnected(so);
9575 				tcp_timer_activate(tp, TT_2MSL,
9576 				    (tcp_fast_finwait2_recycle ?
9577 				    tcp_finwait2_timeout :
9578 				    TP_MAXIDLE(tp)));
9579 			}
9580 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
9581 		}
9582 	}
9583 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9584 	    tiwin, thflags, nxt_pkt));
9585 }
9586 
9587 /*
9588  * Return value of 1, the TCB is unlocked and most
9589  * likely gone, return value of 0, the TCP is still
9590  * locked.
9591  */
9592 static int
9593 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9594     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9595     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9596 {
9597 	int32_t ret_val = 0;
9598 	struct tcp_rack *rack;
9599 
9600 	/*
9601 	 * Header prediction: check for the two common cases of a
9602 	 * uni-directional data xfer.  If the packet has no control flags,
9603 	 * is in-sequence, the window didn't change and we're not
9604 	 * retransmitting, it's a candidate.  If the length is zero and the
9605 	 * ack moved forward, we're the sender side of the xfer.  Just free
9606 	 * the data acked & wake any higher level process that was blocked
9607 	 * waiting for space.  If the length is non-zero and the ack didn't
9608 	 * move, we're the receiver side.  If we're getting packets in-order
9609 	 * (the reassembly queue is empty), add the data toc The socket
9610 	 * buffer and note that we need a delayed ack. Make sure that the
9611 	 * hidden state-flags are also off. Since we check for
9612 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9613 	 */
9614 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9615 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9616 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
9617 	    __predict_true(SEGQ_EMPTY(tp)) &&
9618 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
9619 		if (tlen == 0) {
9620 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9621 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
9622 				return (0);
9623 			}
9624 		} else {
9625 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9626 			    tiwin, nxt_pkt, iptos)) {
9627 				return (0);
9628 			}
9629 		}
9630 	}
9631 	ctf_calc_rwin(so, tp);
9632 
9633 	if ((thflags & TH_RST) ||
9634 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9635 		return (ctf_process_rst(m, th, so, tp));
9636 
9637 	/*
9638 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9639 	 * synchronized state.
9640 	 */
9641 	if (thflags & TH_SYN) {
9642 		ctf_challenge_ack(m, th, tp, &ret_val);
9643 		return (ret_val);
9644 	}
9645 	/*
9646 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9647 	 * it's less than ts_recent, drop it.
9648 	 */
9649 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9650 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9651 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9652 			return (ret_val);
9653 	}
9654 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9655 		return (ret_val);
9656 	}
9657 	/*
9658 	 * If last ACK falls within this segment's sequence numbers, record
9659 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9660 	 * from the latest proposal of the tcplw@cray.com list (Braden
9661 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9662 	 * with our earlier PAWS tests, so this check should be solely
9663 	 * predicated on the sequence space of this segment. 3) That we
9664 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9665 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9666 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9667 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9668 	 * p.869. In such cases, we can still calculate the RTT correctly
9669 	 * when RCV.NXT == Last.ACK.Sent.
9670 	 */
9671 	if ((to->to_flags & TOF_TS) != 0 &&
9672 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9673 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9674 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9675 		tp->ts_recent_age = tcp_ts_getticks();
9676 		tp->ts_recent = to->to_tsval;
9677 	}
9678 	/*
9679 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9680 	 * is on (half-synchronized state), then queue data for later
9681 	 * processing; else drop segment and return.
9682 	 */
9683 	if ((thflags & TH_ACK) == 0) {
9684 		if (tp->t_flags & TF_NEEDSYN) {
9685 
9686 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9687 			    tiwin, thflags, nxt_pkt));
9688 
9689 		} else if (tp->t_flags & TF_ACKNOW) {
9690 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9691 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output= 1;
9692 			return (ret_val);
9693 		} else {
9694 			ctf_do_drop(m, NULL);
9695 			return (0);
9696 		}
9697 	}
9698 	/*
9699 	 * Ack processing.
9700 	 */
9701 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9702 		return (ret_val);
9703 	}
9704 	if (sbavail(&so->so_snd)) {
9705 		if (ctf_progress_timeout_check(tp, true)) {
9706 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
9707 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
9708 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9709 			return (1);
9710 		}
9711 	}
9712 	/* State changes only happen in rack_process_data() */
9713 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9714 	    tiwin, thflags, nxt_pkt));
9715 }
9716 
9717 /*
9718  * Return value of 1, the TCB is unlocked and most
9719  * likely gone, return value of 0, the TCP is still
9720  * locked.
9721  */
9722 static int
9723 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9724     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9725     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9726 {
9727 	int32_t ret_val = 0;
9728 
9729 	ctf_calc_rwin(so, tp);
9730 	if ((thflags & TH_RST) ||
9731 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9732 		return (ctf_process_rst(m, th, so, tp));
9733 	/*
9734 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9735 	 * synchronized state.
9736 	 */
9737 	if (thflags & TH_SYN) {
9738 		ctf_challenge_ack(m, th, tp, &ret_val);
9739 		return (ret_val);
9740 	}
9741 	/*
9742 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9743 	 * it's less than ts_recent, drop it.
9744 	 */
9745 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9746 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9747 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9748 			return (ret_val);
9749 	}
9750 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9751 		return (ret_val);
9752 	}
9753 	/*
9754 	 * If last ACK falls within this segment's sequence numbers, record
9755 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9756 	 * from the latest proposal of the tcplw@cray.com list (Braden
9757 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9758 	 * with our earlier PAWS tests, so this check should be solely
9759 	 * predicated on the sequence space of this segment. 3) That we
9760 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9761 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9762 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9763 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9764 	 * p.869. In such cases, we can still calculate the RTT correctly
9765 	 * when RCV.NXT == Last.ACK.Sent.
9766 	 */
9767 	if ((to->to_flags & TOF_TS) != 0 &&
9768 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9769 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9770 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9771 		tp->ts_recent_age = tcp_ts_getticks();
9772 		tp->ts_recent = to->to_tsval;
9773 	}
9774 	/*
9775 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9776 	 * is on (half-synchronized state), then queue data for later
9777 	 * processing; else drop segment and return.
9778 	 */
9779 	if ((thflags & TH_ACK) == 0) {
9780 		if (tp->t_flags & TF_NEEDSYN) {
9781 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9782 			    tiwin, thflags, nxt_pkt));
9783 
9784 		} else if (tp->t_flags & TF_ACKNOW) {
9785 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9786 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
9787 			return (ret_val);
9788 		} else {
9789 			ctf_do_drop(m, NULL);
9790 			return (0);
9791 		}
9792 	}
9793 	/*
9794 	 * Ack processing.
9795 	 */
9796 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9797 		return (ret_val);
9798 	}
9799 	if (sbavail(&so->so_snd)) {
9800 		if (ctf_progress_timeout_check(tp, true)) {
9801 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
9802 						tp, tick, PROGRESS_DROP, __LINE__);
9803 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
9804 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9805 			return (1);
9806 		}
9807 	}
9808 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9809 	    tiwin, thflags, nxt_pkt));
9810 }
9811 
9812 static int
9813 rack_check_data_after_close(struct mbuf *m,
9814     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
9815 {
9816 	struct tcp_rack *rack;
9817 
9818 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9819 	if (rack->rc_allow_data_af_clo == 0) {
9820 	close_now:
9821 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9822 		/* tcp_close will kill the inp pre-log the Reset */
9823 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9824 		tp = tcp_close(tp);
9825 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
9826 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
9827 		return (1);
9828 	}
9829 	if (sbavail(&so->so_snd) == 0)
9830 		goto close_now;
9831 	/* Ok we allow data that is ignored and a followup reset */
9832 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9833 	tp->rcv_nxt = th->th_seq + *tlen;
9834 	tp->t_flags2 |= TF2_DROP_AF_DATA;
9835 	rack->r_wanted_output = 1;
9836 	*tlen = 0;
9837 	return (0);
9838 }
9839 
9840 /*
9841  * Return value of 1, the TCB is unlocked and most
9842  * likely gone, return value of 0, the TCP is still
9843  * locked.
9844  */
9845 static int
9846 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9847     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9848     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9849 {
9850 	int32_t ret_val = 0;
9851 	int32_t ourfinisacked = 0;
9852 
9853 	ctf_calc_rwin(so, tp);
9854 
9855 	if ((thflags & TH_RST) ||
9856 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9857 		return (ctf_process_rst(m, th, so, tp));
9858 	/*
9859 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9860 	 * synchronized state.
9861 	 */
9862 	if (thflags & TH_SYN) {
9863 		ctf_challenge_ack(m, th, tp, &ret_val);
9864 		return (ret_val);
9865 	}
9866 	/*
9867 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9868 	 * it's less than ts_recent, drop it.
9869 	 */
9870 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9871 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9872 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9873 			return (ret_val);
9874 	}
9875 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9876 		return (ret_val);
9877 	}
9878 	/*
9879 	 * If new data are received on a connection after the user processes
9880 	 * are gone, then RST the other end.
9881 	 */
9882 	if ((so->so_state & SS_NOFDREF) && tlen) {
9883 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
9884 			return (1);
9885 	}
9886 	/*
9887 	 * If last ACK falls within this segment's sequence numbers, record
9888 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
9889 	 * from the latest proposal of the tcplw@cray.com list (Braden
9890 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
9891 	 * with our earlier PAWS tests, so this check should be solely
9892 	 * predicated on the sequence space of this segment. 3) That we
9893 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9894 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9895 	 * SEG.Len, This modified check allows us to overcome RFC1323's
9896 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9897 	 * p.869. In such cases, we can still calculate the RTT correctly
9898 	 * when RCV.NXT == Last.ACK.Sent.
9899 	 */
9900 	if ((to->to_flags & TOF_TS) != 0 &&
9901 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9902 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9903 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9904 		tp->ts_recent_age = tcp_ts_getticks();
9905 		tp->ts_recent = to->to_tsval;
9906 	}
9907 	/*
9908 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9909 	 * is on (half-synchronized state), then queue data for later
9910 	 * processing; else drop segment and return.
9911 	 */
9912 	if ((thflags & TH_ACK) == 0) {
9913 		if (tp->t_flags & TF_NEEDSYN) {
9914 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9915 			    tiwin, thflags, nxt_pkt));
9916 		} else if (tp->t_flags & TF_ACKNOW) {
9917 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9918 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
9919 			return (ret_val);
9920 		} else {
9921 			ctf_do_drop(m, NULL);
9922 			return (0);
9923 		}
9924 	}
9925 	/*
9926 	 * Ack processing.
9927 	 */
9928 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9929 		return (ret_val);
9930 	}
9931 	if (ourfinisacked) {
9932 		/*
9933 		 * If we can't receive any more data, then closing user can
9934 		 * proceed. Starting the timer is contrary to the
9935 		 * specification, but if we don't get a FIN we'll hang
9936 		 * forever.
9937 		 *
9938 		 * XXXjl: we should release the tp also, and use a
9939 		 * compressed state.
9940 		 */
9941 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9942 			soisdisconnected(so);
9943 			tcp_timer_activate(tp, TT_2MSL,
9944 			    (tcp_fast_finwait2_recycle ?
9945 			    tcp_finwait2_timeout :
9946 			    TP_MAXIDLE(tp)));
9947 		}
9948 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
9949 	}
9950 	if (sbavail(&so->so_snd)) {
9951 		if (ctf_progress_timeout_check(tp, true)) {
9952 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
9953 						tp, tick, PROGRESS_DROP, __LINE__);
9954 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
9955 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9956 			return (1);
9957 		}
9958 	}
9959 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
9960 	    tiwin, thflags, nxt_pkt));
9961 }
9962 
9963 /*
9964  * Return value of 1, the TCB is unlocked and most
9965  * likely gone, return value of 0, the TCP is still
9966  * locked.
9967  */
9968 static int
9969 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9970     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9971     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9972 {
9973 	int32_t ret_val = 0;
9974 	int32_t ourfinisacked = 0;
9975 
9976 	ctf_calc_rwin(so, tp);
9977 
9978 	if ((thflags & TH_RST) ||
9979 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
9980 		return (ctf_process_rst(m, th, so, tp));
9981 	/*
9982 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9983 	 * synchronized state.
9984 	 */
9985 	if (thflags & TH_SYN) {
9986 		ctf_challenge_ack(m, th, tp, &ret_val);
9987 		return (ret_val);
9988 	}
9989 	/*
9990 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9991 	 * it's less than ts_recent, drop it.
9992 	 */
9993 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9994 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9995 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9996 			return (ret_val);
9997 	}
9998 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9999 		return (ret_val);
10000 	}
10001 	/*
10002 	 * If new data are received on a connection after the user processes
10003 	 * are gone, then RST the other end.
10004 	 */
10005 	if ((so->so_state & SS_NOFDREF) && tlen) {
10006 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
10007 			return (1);
10008 	}
10009 	/*
10010 	 * If last ACK falls within this segment's sequence numbers, record
10011 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
10012 	 * from the latest proposal of the tcplw@cray.com list (Braden
10013 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
10014 	 * with our earlier PAWS tests, so this check should be solely
10015 	 * predicated on the sequence space of this segment. 3) That we
10016 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10017 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10018 	 * SEG.Len, This modified check allows us to overcome RFC1323's
10019 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10020 	 * p.869. In such cases, we can still calculate the RTT correctly
10021 	 * when RCV.NXT == Last.ACK.Sent.
10022 	 */
10023 	if ((to->to_flags & TOF_TS) != 0 &&
10024 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
10025 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
10026 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
10027 		tp->ts_recent_age = tcp_ts_getticks();
10028 		tp->ts_recent = to->to_tsval;
10029 	}
10030 	/*
10031 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
10032 	 * is on (half-synchronized state), then queue data for later
10033 	 * processing; else drop segment and return.
10034 	 */
10035 	if ((thflags & TH_ACK) == 0) {
10036 		if (tp->t_flags & TF_NEEDSYN) {
10037 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10038 			    tiwin, thflags, nxt_pkt));
10039 		} else if (tp->t_flags & TF_ACKNOW) {
10040 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
10041 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output= 1;
10042 			return (ret_val);
10043 		} else {
10044 			ctf_do_drop(m, NULL);
10045 			return (0);
10046 		}
10047 	}
10048 	/*
10049 	 * Ack processing.
10050 	 */
10051 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
10052 		return (ret_val);
10053 	}
10054 	if (ourfinisacked) {
10055 		tcp_twstart(tp);
10056 		m_freem(m);
10057 		return (1);
10058 	}
10059 	if (sbavail(&so->so_snd)) {
10060 		if (ctf_progress_timeout_check(tp, true)) {
10061 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
10062 						tp, tick, PROGRESS_DROP, __LINE__);
10063 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
10064 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10065 			return (1);
10066 		}
10067 	}
10068 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10069 	    tiwin, thflags, nxt_pkt));
10070 }
10071 
10072 /*
10073  * Return value of 1, the TCB is unlocked and most
10074  * likely gone, return value of 0, the TCP is still
10075  * locked.
10076  */
10077 static int
10078 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10079     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10080     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10081 {
10082 	int32_t ret_val = 0;
10083 	int32_t ourfinisacked = 0;
10084 
10085 	ctf_calc_rwin(so, tp);
10086 
10087 	if ((thflags & TH_RST) ||
10088 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
10089 		return (ctf_process_rst(m, th, so, tp));
10090 	/*
10091 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
10092 	 * synchronized state.
10093 	 */
10094 	if (thflags & TH_SYN) {
10095 		ctf_challenge_ack(m, th, tp, &ret_val);
10096 		return (ret_val);
10097 	}
10098 	/*
10099 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10100 	 * it's less than ts_recent, drop it.
10101 	 */
10102 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10103 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10104 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10105 			return (ret_val);
10106 	}
10107 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
10108 		return (ret_val);
10109 	}
10110 	/*
10111 	 * If new data are received on a connection after the user processes
10112 	 * are gone, then RST the other end.
10113 	 */
10114 	if ((so->so_state & SS_NOFDREF) && tlen) {
10115 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
10116 			return (1);
10117 	}
10118 	/*
10119 	 * If last ACK falls within this segment's sequence numbers, record
10120 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
10121 	 * from the latest proposal of the tcplw@cray.com list (Braden
10122 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
10123 	 * with our earlier PAWS tests, so this check should be solely
10124 	 * predicated on the sequence space of this segment. 3) That we
10125 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10126 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10127 	 * SEG.Len, This modified check allows us to overcome RFC1323's
10128 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10129 	 * p.869. In such cases, we can still calculate the RTT correctly
10130 	 * when RCV.NXT == Last.ACK.Sent.
10131 	 */
10132 	if ((to->to_flags & TOF_TS) != 0 &&
10133 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
10134 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
10135 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
10136 		tp->ts_recent_age = tcp_ts_getticks();
10137 		tp->ts_recent = to->to_tsval;
10138 	}
10139 	/*
10140 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
10141 	 * is on (half-synchronized state), then queue data for later
10142 	 * processing; else drop segment and return.
10143 	 */
10144 	if ((thflags & TH_ACK) == 0) {
10145 		if (tp->t_flags & TF_NEEDSYN) {
10146 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10147 			    tiwin, thflags, nxt_pkt));
10148 		} else if (tp->t_flags & TF_ACKNOW) {
10149 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
10150 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
10151 			return (ret_val);
10152 		} else {
10153 			ctf_do_drop(m, NULL);
10154 			return (0);
10155 		}
10156 	}
10157 	/*
10158 	 * case TCPS_LAST_ACK: Ack processing.
10159 	 */
10160 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
10161 		return (ret_val);
10162 	}
10163 	if (ourfinisacked) {
10164 		tp = tcp_close(tp);
10165 		ctf_do_drop(m, tp);
10166 		return (1);
10167 	}
10168 	if (sbavail(&so->so_snd)) {
10169 		if (ctf_progress_timeout_check(tp, true)) {
10170 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
10171 						tp, tick, PROGRESS_DROP, __LINE__);
10172 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
10173 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10174 			return (1);
10175 		}
10176 	}
10177 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10178 	    tiwin, thflags, nxt_pkt));
10179 }
10180 
10181 
10182 /*
10183  * Return value of 1, the TCB is unlocked and most
10184  * likely gone, return value of 0, the TCP is still
10185  * locked.
10186  */
10187 static int
10188 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
10189     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10190     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10191 {
10192 	int32_t ret_val = 0;
10193 	int32_t ourfinisacked = 0;
10194 
10195 	ctf_calc_rwin(so, tp);
10196 
10197 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
10198 	if ((thflags & TH_RST) ||
10199 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
10200 		return (ctf_process_rst(m, th, so, tp));
10201 	/*
10202 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
10203 	 * synchronized state.
10204 	 */
10205 	if (thflags & TH_SYN) {
10206 		ctf_challenge_ack(m, th, tp, &ret_val);
10207 		return (ret_val);
10208 	}
10209 	/*
10210 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10211 	 * it's less than ts_recent, drop it.
10212 	 */
10213 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10214 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10215 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10216 			return (ret_val);
10217 	}
10218 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
10219 		return (ret_val);
10220 	}
10221 	/*
10222 	 * If new data are received on a connection after the user processes
10223 	 * are gone, then RST the other end.
10224 	 */
10225 	if ((so->so_state & SS_NOFDREF) &&
10226 	    tlen) {
10227 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
10228 			return (1);
10229 	}
10230 	/*
10231 	 * If last ACK falls within this segment's sequence numbers, record
10232 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
10233 	 * from the latest proposal of the tcplw@cray.com list (Braden
10234 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
10235 	 * with our earlier PAWS tests, so this check should be solely
10236 	 * predicated on the sequence space of this segment. 3) That we
10237 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
10238 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
10239 	 * SEG.Len, This modified check allows us to overcome RFC1323's
10240 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
10241 	 * p.869. In such cases, we can still calculate the RTT correctly
10242 	 * when RCV.NXT == Last.ACK.Sent.
10243 	 */
10244 	if ((to->to_flags & TOF_TS) != 0 &&
10245 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
10246 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
10247 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
10248 		tp->ts_recent_age = tcp_ts_getticks();
10249 		tp->ts_recent = to->to_tsval;
10250 	}
10251 	/*
10252 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
10253 	 * is on (half-synchronized state), then queue data for later
10254 	 * processing; else drop segment and return.
10255 	 */
10256 	if ((thflags & TH_ACK) == 0) {
10257 		if (tp->t_flags & TF_NEEDSYN) {
10258 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10259 			    tiwin, thflags, nxt_pkt));
10260 		} else if (tp->t_flags & TF_ACKNOW) {
10261 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
10262 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
10263 			return (ret_val);
10264 		} else {
10265 			ctf_do_drop(m, NULL);
10266 			return (0);
10267 		}
10268 	}
10269 	/*
10270 	 * Ack processing.
10271 	 */
10272 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
10273 		return (ret_val);
10274 	}
10275 	if (sbavail(&so->so_snd)) {
10276 		if (ctf_progress_timeout_check(tp, true)) {
10277 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
10278 						tp, tick, PROGRESS_DROP, __LINE__);
10279 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
10280 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10281 			return (1);
10282 		}
10283 	}
10284 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10285 	    tiwin, thflags, nxt_pkt));
10286 }
10287 
10288 static void inline
10289 rack_clear_rate_sample(struct tcp_rack *rack)
10290 {
10291 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
10292 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
10293 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
10294 }
10295 
10296 static void
10297 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line)
10298 {
10299 	uint64_t bw_est, rate_wanted;
10300 	int chged = 0;
10301 	uint32_t user_max;
10302 
10303 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
10304 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
10305 		chged = 1;
10306 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
10307 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
10308 		if (user_max != rack->r_ctl.rc_pace_max_segs)
10309 			chged = 1;
10310 	}
10311 	if (rack->rc_force_max_seg) {
10312 		rack->r_ctl.rc_pace_max_segs = user_max;
10313 	} else if (rack->use_fixed_rate) {
10314 		bw_est = rack_get_bw(rack);
10315 		if ((rack->r_ctl.crte == NULL) ||
10316 		    (bw_est != rack->r_ctl.crte->rate))  {
10317 			rack->r_ctl.rc_pace_max_segs = user_max;
10318 		} else {
10319 			/* We are pacing right at the hardware rate */
10320 			uint32_t segsiz;
10321 
10322 			segsiz = min(ctf_fixed_maxseg(tp),
10323 				     rack->r_ctl.rc_pace_min_segs);
10324 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
10325 				                           bw_est, segsiz, 0,
10326 							   rack->r_ctl.crte, NULL);
10327 		}
10328 	} else if (rack->rc_always_pace) {
10329 		if (rack->r_ctl.gp_bw ||
10330 #ifdef NETFLIX_PEAKRATE
10331 		    rack->rc_tp->t_maxpeakrate ||
10332 #endif
10333 		    rack->r_ctl.init_rate) {
10334 			/* We have a rate of some sort set */
10335 			uint32_t  orig;
10336 
10337 			bw_est = rack_get_bw(rack);
10338 			orig = rack->r_ctl.rc_pace_max_segs;
10339 			rate_wanted = rack_get_output_bw(rack, bw_est, NULL);
10340 			if (rate_wanted) {
10341 				/* We have something */
10342 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
10343 										   rate_wanted,
10344 										   ctf_fixed_maxseg(rack->rc_tp));
10345 			} else
10346 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
10347 			if (orig != rack->r_ctl.rc_pace_max_segs)
10348 				chged = 1;
10349 		} else if ((rack->r_ctl.gp_bw == 0) &&
10350 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
10351 			/*
10352 			 * If we have nothing limit us to bursting
10353 			 * out IW sized pieces.
10354 			 */
10355 			chged = 1;
10356 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
10357 		}
10358 	}
10359 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
10360 		chged = 1;
10361 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
10362 	}
10363 	if (chged)
10364 		rack_log_type_hrdwtso(tp, rack, 0, rack->rc_inp->inp_socket->so_snd.sb_flags, line, 2);
10365 }
10366 
10367 static int
10368 rack_init(struct tcpcb *tp)
10369 {
10370 	struct tcp_rack *rack = NULL;
10371 	struct rack_sendmap *insret;
10372 	uint32_t iwin, snt, us_cts;
10373 
10374 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
10375 	if (tp->t_fb_ptr == NULL) {
10376 		/*
10377 		 * We need to allocate memory but cant. The INP and INP_INFO
10378 		 * locks and they are recusive (happens during setup. So a
10379 		 * scheme to drop the locks fails :(
10380 		 *
10381 		 */
10382 		return (ENOMEM);
10383 	}
10384 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
10385 
10386 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10387 	RB_INIT(&rack->r_ctl.rc_mtree);
10388 	TAILQ_INIT(&rack->r_ctl.rc_free);
10389 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
10390 	rack->rc_tp = tp;
10391 	if (tp->t_inpcb) {
10392 		rack->rc_inp = tp->t_inpcb;
10393 	}
10394 	/* Probably not needed but lets be sure */
10395 	rack_clear_rate_sample(rack);
10396 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
10397 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
10398 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
10399 	if (use_rack_rr)
10400 		rack->use_rack_rr = 1;
10401 	if (V_tcp_delack_enabled)
10402 		tp->t_delayed_ack = 1;
10403 	else
10404 		tp->t_delayed_ack = 0;
10405 	if (rack_enable_shared_cwnd)
10406 		rack->rack_enable_scwnd = 1;
10407 	rack->rc_user_set_max_segs = rack_hptsi_segments;
10408 	rack->rc_force_max_seg = 0;
10409 	if (rack_use_imac_dack)
10410 		rack->rc_dack_mode = 1;
10411 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
10412 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
10413 	rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce;
10414 	rack->r_ctl.rc_prop_rate = rack_proportional_rate;
10415 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
10416 	rack->r_ctl.rc_early_recovery = rack_early_recovery;
10417 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
10418 	rack->r_ctl.rc_highest_us_rtt = 0;
10419 	if (rack_disable_prr)
10420 		rack->rack_no_prr = 1;
10421 	if (rack_gp_no_rec_chg)
10422 		rack->rc_gp_no_rec_chg = 1;
10423 	rack->rc_always_pace = rack_pace_every_seg;
10424 	if (rack_enable_mqueue_for_nonpaced)
10425 		rack->r_mbuf_queue = 1;
10426 	else
10427 		rack->r_mbuf_queue = 0;
10428 	if  (rack->r_mbuf_queue || rack->rc_always_pace)
10429 		tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
10430 	else
10431 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10432 	rack_set_pace_segments(tp, rack, __LINE__);
10433 	if (rack_limits_scwnd)
10434 		rack->r_limit_scw  = 1;
10435 	else
10436 		rack->r_limit_scw  = 0;
10437 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
10438 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
10439 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
10440 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
10441 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
10442 	rack->r_ctl.rc_min_to = rack_min_to;
10443 	microuptime(&rack->r_ctl.act_rcv_time);
10444 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
10445 	rack->r_running_late = 0;
10446 	rack->r_running_early = 0;
10447 	rack->rc_init_win = rack_default_init_window;
10448 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
10449 	if (rack_do_dyn_mul) {
10450 		/* When dynamic adjustment is on CA needs to start at 100% */
10451 		rack->rc_gp_dyn_mul = 1;
10452 		if (rack_do_dyn_mul >= 100)
10453 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
10454 	} else
10455 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
10456 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
10457 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
10458 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
10459 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
10460 				rack_probertt_filter_life);
10461 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10462 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
10463 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
10464 	rack->r_ctl.rc_time_probertt_starts = 0;
10465 	/* Do we force on detection? */
10466 #ifdef NETFLIX_EXP_DETECTION
10467 	if (tcp_force_detection)
10468 		rack->do_detection = 1;
10469 	else
10470 #endif
10471 		rack->do_detection = 0;
10472 	if (rack_non_rxt_use_cr)
10473 		rack->rack_rec_nonrxt_use_cr = 1;
10474 	if (tp->snd_una != tp->snd_max) {
10475 		/* Create a send map for the current outstanding data */
10476 		struct rack_sendmap *rsm;
10477 
10478 		rsm = rack_alloc(rack);
10479 		if (rsm == NULL) {
10480 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
10481 			tp->t_fb_ptr = NULL;
10482 			return (ENOMEM);
10483 		}
10484 		rsm->r_flags = RACK_OVERMAX;
10485 		rsm->r_tim_lastsent[0] = rack->r_ctl.rc_tlp_rxt_last_time;
10486 		rsm->r_rtr_cnt = 1;
10487 		rsm->r_rtr_bytes = 0;
10488 		rsm->r_start = tp->snd_una;
10489 		rsm->r_end = tp->snd_max;
10490 		rsm->usec_orig_send = us_cts;
10491 		rsm->r_dupack = 0;
10492 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
10493 #ifdef INVARIANTS
10494 		if (insret != NULL) {
10495 			panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
10496 			      insret, rack, rsm);
10497 		}
10498 #endif
10499 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10500 		rsm->r_in_tmap = 1;
10501 	}
10502 	/* Cancel the GP measurement in progress */
10503 	tp->t_flags &= ~TF_GPUTINPROG;
10504 	if (SEQ_GT(tp->snd_max, tp->iss))
10505 		snt = tp->snd_max - tp->iss;
10506 	else
10507 		snt = 0;
10508 	iwin = rc_init_window(rack);
10509 	if (snt < iwin) {
10510 		/* We are not past the initial window
10511 		 * so we need to make sure cwnd is
10512 		 * correct.
10513 		 */
10514 		if (tp->snd_cwnd < iwin)
10515 			tp->snd_cwnd = iwin;
10516 		/*
10517 		 * If we are within the initial window
10518 		 * we want ssthresh to be unlimited. Setting
10519 		 * it to the rwnd (which the default stack does
10520 		 * and older racks) is not really a good idea
10521 		 * since we want to be in SS and grow both the
10522 		 * cwnd and the rwnd (via dynamic rwnd growth). If
10523 		 * we set it to the rwnd then as the peer grows its
10524 		 * rwnd we will be stuck in CA and never hit SS.
10525 		 *
10526 		 * Its far better to raise it up high (this takes the
10527 		 * risk that there as been a loss already, probably
10528 		 * we should have an indicator in all stacks of loss
10529 		 * but we don't), but considering the normal use this
10530 		 * is a risk worth taking. The consequences of not
10531 		 * hitting SS are far worse than going one more time
10532 		 * into it early on (before we have sent even a IW).
10533 		 * It is highly unlikely that we will have had a loss
10534 		 * before getting the IW out.
10535 		 */
10536 		tp->snd_ssthresh = 0xffffffff;
10537 	}
10538 	rack_stop_all_timers(tp);
10539 	rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0);
10540 	rack_log_rtt_shrinks(rack,  us_cts,  0,
10541 			     __LINE__, RACK_RTTS_INIT);
10542 	return (0);
10543 }
10544 
10545 static int
10546 rack_handoff_ok(struct tcpcb *tp)
10547 {
10548 	if ((tp->t_state == TCPS_CLOSED) ||
10549 	    (tp->t_state == TCPS_LISTEN)) {
10550 		/* Sure no problem though it may not stick */
10551 		return (0);
10552 	}
10553 	if ((tp->t_state == TCPS_SYN_SENT) ||
10554 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
10555 		/*
10556 		 * We really don't know you have to get to ESTAB or beyond
10557 		 * to tell.
10558 		 */
10559 		return (EAGAIN);
10560 	}
10561 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
10562 		return (0);
10563 	}
10564 	/*
10565 	 * If we reach here we don't do SACK on this connection so we can
10566 	 * never do rack.
10567 	 */
10568 	return (EINVAL);
10569 }
10570 
10571 static void
10572 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10573 {
10574 	if (tp->t_fb_ptr) {
10575 		struct tcp_rack *rack;
10576 		struct rack_sendmap *rsm, *nrsm, *rm;
10577 
10578 		rack = (struct tcp_rack *)tp->t_fb_ptr;
10579 #ifdef NETFLIX_SHARED_CWND
10580 		if (rack->r_ctl.rc_scw) {
10581 			uint32_t limit;
10582 
10583 			if (rack->r_limit_scw)
10584 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
10585 			else
10586 				limit = 0;
10587 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
10588 						  rack->r_ctl.rc_scw_index,
10589 						  limit);
10590 			rack->r_ctl.rc_scw = NULL;
10591 		}
10592 #endif
10593 		/* rack does not use force data but other stacks may clear it */
10594 		tp->t_flags &= ~TF_FORCEDATA;
10595 		if (tp->t_inpcb) {
10596 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10597 			tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
10598 			tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
10599 		}
10600 #ifdef TCP_BLACKBOX
10601 		tcp_log_flowend(tp);
10602 #endif
10603 		RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
10604 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
10605 #ifdef INVARIANTS
10606 			if (rm != rsm) {
10607 				panic("At fini, rack:%p rsm:%p rm:%p",
10608 				      rack, rsm, rm);
10609 			}
10610 #endif
10611 			uma_zfree(rack_zone, rsm);
10612 		}
10613 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
10614 		while (rsm) {
10615 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
10616 			uma_zfree(rack_zone, rsm);
10617 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
10618 		}
10619 		rack->rc_free_cnt = 0;
10620 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
10621 		tp->t_fb_ptr = NULL;
10622 	}
10623 	/* Cancel the GP measurement in progress */
10624 	tp->t_flags &= ~TF_GPUTINPROG;
10625 	/* Make sure snd_nxt is correctly set */
10626 	tp->snd_nxt = tp->snd_max;
10627 }
10628 
10629 
10630 static void
10631 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
10632 {
10633 	switch (tp->t_state) {
10634 	case TCPS_SYN_SENT:
10635 		rack->r_state = TCPS_SYN_SENT;
10636 		rack->r_substate = rack_do_syn_sent;
10637 		break;
10638 	case TCPS_SYN_RECEIVED:
10639 		rack->r_state = TCPS_SYN_RECEIVED;
10640 		rack->r_substate = rack_do_syn_recv;
10641 		break;
10642 	case TCPS_ESTABLISHED:
10643 		rack_set_pace_segments(tp, rack, __LINE__);
10644 		rack->r_state = TCPS_ESTABLISHED;
10645 		rack->r_substate = rack_do_established;
10646 		break;
10647 	case TCPS_CLOSE_WAIT:
10648 		rack->r_state = TCPS_CLOSE_WAIT;
10649 		rack->r_substate = rack_do_close_wait;
10650 		break;
10651 	case TCPS_FIN_WAIT_1:
10652 		rack->r_state = TCPS_FIN_WAIT_1;
10653 		rack->r_substate = rack_do_fin_wait_1;
10654 		break;
10655 	case TCPS_CLOSING:
10656 		rack->r_state = TCPS_CLOSING;
10657 		rack->r_substate = rack_do_closing;
10658 		break;
10659 	case TCPS_LAST_ACK:
10660 		rack->r_state = TCPS_LAST_ACK;
10661 		rack->r_substate = rack_do_lastack;
10662 		break;
10663 	case TCPS_FIN_WAIT_2:
10664 		rack->r_state = TCPS_FIN_WAIT_2;
10665 		rack->r_substate = rack_do_fin_wait_2;
10666 		break;
10667 	case TCPS_LISTEN:
10668 	case TCPS_CLOSED:
10669 	case TCPS_TIME_WAIT:
10670 	default:
10671 		break;
10672 	};
10673 }
10674 
10675 
10676 static void
10677 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
10678 {
10679 	/*
10680 	 * We received an ack, and then did not
10681 	 * call send or were bounced out due to the
10682 	 * hpts was running. Now a timer is up as well, is
10683 	 * it the right timer?
10684 	 */
10685 	struct rack_sendmap *rsm;
10686 	int tmr_up;
10687 
10688 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
10689 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
10690 		return;
10691 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
10692 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
10693 	    (tmr_up == PACE_TMR_RXT)) {
10694 		/* Should be an RXT */
10695 		return;
10696 	}
10697 	if (rsm == NULL) {
10698 		/* Nothing outstanding? */
10699 		if (tp->t_flags & TF_DELACK) {
10700 			if (tmr_up == PACE_TMR_DELACK)
10701 				/* We are supposed to have delayed ack up and we do */
10702 				return;
10703 		} else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
10704 			/*
10705 			 * if we hit enobufs then we would expect the possiblity
10706 			 * of nothing outstanding and the RXT up (and the hptsi timer).
10707 			 */
10708 			return;
10709 		} else if (((V_tcp_always_keepalive ||
10710 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
10711 			    (tp->t_state <= TCPS_CLOSING)) &&
10712 			   (tmr_up == PACE_TMR_KEEP) &&
10713 			   (tp->snd_max == tp->snd_una)) {
10714 			/* We should have keep alive up and we do */
10715 			return;
10716 		}
10717 	}
10718 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
10719 		   ((tmr_up == PACE_TMR_TLP) ||
10720 		    (tmr_up == PACE_TMR_RACK) ||
10721 		    (tmr_up == PACE_TMR_RXT))) {
10722 		/*
10723 		 * Either a Rack, TLP or RXT is fine if  we
10724 		 * have outstanding data.
10725 		 */
10726 		return;
10727 	} else if (tmr_up == PACE_TMR_DELACK) {
10728 		/*
10729 		 * If the delayed ack was going to go off
10730 		 * before the rtx/tlp/rack timer were going to
10731 		 * expire, then that would be the timer in control.
10732 		 * Note we don't check the time here trusting the
10733 		 * code is correct.
10734 		 */
10735 		return;
10736 	}
10737 	/*
10738 	 * Ok the timer originally started is not what we want now.
10739 	 * We will force the hpts to be stopped if any, and restart
10740 	 * with the slot set to what was in the saved slot.
10741 	 */
10742 	if (rack->rc_inp->inp_in_hpts) {
10743 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
10744 			uint32_t us_cts;
10745 
10746 			us_cts = tcp_get_usecs(NULL);
10747 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
10748 				rack->r_early = 1;
10749 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
10750 			}
10751 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
10752 		}
10753 		tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
10754 	}
10755 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10756 	rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0);
10757 }
10758 
10759 static int
10760 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
10761     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
10762     int32_t nxt_pkt, struct timeval *tv)
10763 {
10764 	int32_t thflags, retval, did_out = 0;
10765 	int32_t way_out = 0;
10766 	uint32_t cts;
10767 	uint32_t tiwin;
10768 	struct timespec ts;
10769 	struct tcpopt to;
10770 	struct tcp_rack *rack;
10771 	struct rack_sendmap *rsm;
10772 	int32_t prev_state = 0;
10773 	uint32_t us_cts;
10774 	/*
10775 	 * tv passed from common code is from either M_TSTMP_LRO or
10776 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present. The
10777 	 * rack_pacing stack assumes tv always refers to 'now', so we overwrite
10778 	 * tv here to guarantee that.
10779 	 */
10780 	if (m->m_flags & M_TSTMP_LRO)
10781 		tcp_get_usecs(tv);
10782 
10783 	cts = tcp_tv_to_mssectick(tv);
10784 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10785 
10786 	if ((m->m_flags & M_TSTMP) ||
10787 	    (m->m_flags & M_TSTMP_LRO)) {
10788 		mbuf_tstmp2timespec(m, &ts);
10789 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
10790 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
10791 	} else
10792 		rack->r_ctl.act_rcv_time = *tv;
10793 	kern_prefetch(rack, &prev_state);
10794 	prev_state = 0;
10795 	thflags = th->th_flags;
10796 
10797 	NET_EPOCH_ASSERT();
10798 	INP_WLOCK_ASSERT(tp->t_inpcb);
10799 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
10800 	    __func__));
10801 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
10802 	    __func__));
10803 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
10804 		union tcp_log_stackspecific log;
10805 		struct timeval ltv;
10806 #ifdef NETFLIX_HTTP_LOGGING
10807 		struct http_sendfile_track *http_req;
10808 
10809 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
10810 			http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
10811 		} else {
10812 			http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
10813 		}
10814 #endif
10815 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
10816 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
10817 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
10818 		if (rack->rack_no_prr == 0)
10819 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
10820 		else
10821 			log.u_bbr.flex1 = 0;
10822 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
10823 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10824 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
10825 		log.u_bbr.flex3 = m->m_flags;
10826 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
10827 		if (m->m_flags & M_TSTMP) {
10828 			/* Record the hardware timestamp if present */
10829 			mbuf_tstmp2timespec(m, &ts);
10830 			ltv.tv_sec = ts.tv_sec;
10831 			ltv.tv_usec = ts.tv_nsec / 1000;
10832 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
10833 		} else if (m->m_flags & M_TSTMP_LRO) {
10834 			/* Record the LRO the arrival timestamp */
10835 			mbuf_tstmp2timespec(m, &ts);
10836 			ltv.tv_sec = ts.tv_sec;
10837 			ltv.tv_usec = ts.tv_nsec / 1000;
10838 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
10839 		}
10840 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
10841 		/* Log the rcv time */
10842 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
10843 #ifdef NETFLIX_HTTP_LOGGING
10844 		log.u_bbr.applimited = tp->t_http_closed;
10845 		log.u_bbr.applimited <<= 8;
10846 		log.u_bbr.applimited |= tp->t_http_open;
10847 		log.u_bbr.applimited <<= 8;
10848 		log.u_bbr.applimited |= tp->t_http_req;
10849 		if (http_req) {
10850 			/* Copy out any client req info */
10851 			/* seconds */
10852 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
10853 			/* useconds */
10854 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
10855 			log.u_bbr.rttProp = http_req->timestamp;
10856 			log.u_bbr.cur_del_rate = http_req->start;
10857 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
10858 				log.u_bbr.flex8 |= 1;
10859 			} else {
10860 				log.u_bbr.flex8 |= 2;
10861 				log.u_bbr.bw_inuse = http_req->end;
10862 			}
10863 			log.u_bbr.flex6 = http_req->start_seq;
10864 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
10865 				log.u_bbr.flex8 |= 4;
10866 				log.u_bbr.epoch = http_req->end_seq;
10867 			}
10868 		}
10869 #endif
10870 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
10871 		    tlen, &log, true, &ltv);
10872 	}
10873 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
10874 		way_out = 4;
10875 		retval = 0;
10876 		goto done_with_input;
10877 	}
10878 	/*
10879 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
10880 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
10881 	 */
10882 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
10883 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
10884 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10885 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10886 		return(1);
10887 	}
10888 	/*
10889 	 * Segment received on connection. Reset idle time and keep-alive
10890 	 * timer. XXX: This should be done after segment validation to
10891 	 * ignore broken/spoofed segs.
10892 	 */
10893 	if  (tp->t_idle_reduce &&
10894 	     (tp->snd_max == tp->snd_una) &&
10895 	     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
10896 		counter_u64_add(rack_input_idle_reduces, 1);
10897 		rack_cc_after_idle(rack, tp);
10898 	}
10899 	tp->t_rcvtime = ticks;
10900 	/*
10901 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
10902 	 * the scale is zero.
10903 	 */
10904 	tiwin = th->th_win << tp->snd_scale;
10905 #ifdef STATS
10906 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
10907 #endif
10908 	if (tiwin > rack->r_ctl.rc_high_rwnd)
10909 		rack->r_ctl.rc_high_rwnd = tiwin;
10910 	/*
10911 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
10912 	 * this to occur after we've validated the segment.
10913 	 */
10914 	if (tp->t_flags2 & TF2_ECN_PERMIT) {
10915 		if (thflags & TH_CWR) {
10916 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
10917 			tp->t_flags |= TF_ACKNOW;
10918 		}
10919 		switch (iptos & IPTOS_ECN_MASK) {
10920 		case IPTOS_ECN_CE:
10921 			tp->t_flags2 |= TF2_ECN_SND_ECE;
10922 			KMOD_TCPSTAT_INC(tcps_ecn_ce);
10923 			break;
10924 		case IPTOS_ECN_ECT0:
10925 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
10926 			break;
10927 		case IPTOS_ECN_ECT1:
10928 			KMOD_TCPSTAT_INC(tcps_ecn_ect1);
10929 			break;
10930 		}
10931 
10932 		/* Process a packet differently from RFC3168. */
10933 		cc_ecnpkt_handler(tp, th, iptos);
10934 
10935 		/* Congestion experienced. */
10936 		if (thflags & TH_ECE) {
10937 			rack_cong_signal(tp, th, CC_ECN);
10938 		}
10939 	}
10940 	/*
10941 	 * Parse options on any incoming segment.
10942 	 */
10943 	tcp_dooptions(&to, (u_char *)(th + 1),
10944 	    (th->th_off << 2) - sizeof(struct tcphdr),
10945 	    (thflags & TH_SYN) ? TO_SYN : 0);
10946 
10947 	/*
10948 	 * If echoed timestamp is later than the current time, fall back to
10949 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
10950 	 * were used when this connection was established.
10951 	 */
10952 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
10953 		to.to_tsecr -= tp->ts_offset;
10954 		if (TSTMP_GT(to.to_tsecr, cts))
10955 			to.to_tsecr = 0;
10956 	}
10957 
10958 	/*
10959 	 * If its the first time in we need to take care of options and
10960 	 * verify we can do SACK for rack!
10961 	 */
10962 	if (rack->r_state == 0) {
10963 		/* Should be init'd by rack_init() */
10964 		KASSERT(rack->rc_inp != NULL,
10965 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
10966 		if (rack->rc_inp == NULL) {
10967 			rack->rc_inp = tp->t_inpcb;
10968 		}
10969 
10970 		/*
10971 		 * Process options only when we get SYN/ACK back. The SYN
10972 		 * case for incoming connections is handled in tcp_syncache.
10973 		 * According to RFC1323 the window field in a SYN (i.e., a
10974 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
10975 		 * this is traditional behavior, may need to be cleaned up.
10976 		 */
10977 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
10978 			/* Handle parallel SYN for ECN */
10979 			if (!(thflags & TH_ACK) &&
10980 			    ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
10981 			    ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
10982 				tp->t_flags2 |= TF2_ECN_PERMIT;
10983 				tp->t_flags2 |= TF2_ECN_SND_ECE;
10984 				TCPSTAT_INC(tcps_ecn_shs);
10985 			}
10986 			if ((to.to_flags & TOF_SCALE) &&
10987 			    (tp->t_flags & TF_REQ_SCALE)) {
10988 				tp->t_flags |= TF_RCVD_SCALE;
10989 				tp->snd_scale = to.to_wscale;
10990 			} else
10991 				tp->t_flags &= ~TF_REQ_SCALE;
10992 			/*
10993 			 * Initial send window.  It will be updated with the
10994 			 * next incoming segment to the scaled value.
10995 			 */
10996 			tp->snd_wnd = th->th_win;
10997 			if ((to.to_flags & TOF_TS) &&
10998 			    (tp->t_flags & TF_REQ_TSTMP)) {
10999 				tp->t_flags |= TF_RCVD_TSTMP;
11000 				tp->ts_recent = to.to_tsval;
11001 				tp->ts_recent_age = cts;
11002 			} else
11003 				tp->t_flags &= ~TF_REQ_TSTMP;
11004 			if (to.to_flags & TOF_MSS)
11005 				tcp_mss(tp, to.to_mss);
11006 			if ((tp->t_flags & TF_SACK_PERMIT) &&
11007 			    (to.to_flags & TOF_SACKPERM) == 0)
11008 				tp->t_flags &= ~TF_SACK_PERMIT;
11009 			if (IS_FASTOPEN(tp->t_flags)) {
11010 				if (to.to_flags & TOF_FASTOPEN) {
11011 					uint16_t mss;
11012 
11013 					if (to.to_flags & TOF_MSS)
11014 						mss = to.to_mss;
11015 					else
11016 						if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
11017 							mss = TCP6_MSS;
11018 						else
11019 							mss = TCP_MSS;
11020 					tcp_fastopen_update_cache(tp, mss,
11021 					    to.to_tfo_len, to.to_tfo_cookie);
11022 				} else
11023 					tcp_fastopen_disable_path(tp);
11024 			}
11025 		}
11026 		/*
11027 		 * At this point we are at the initial call. Here we decide
11028 		 * if we are doing RACK or not. We do this by seeing if
11029 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
11030 		 * The code now does do dup-ack counting so if you don't
11031 		 * switch back you won't get rack & TLP, but you will still
11032 		 * get this stack.
11033 		 */
11034 
11035 		if ((rack_sack_not_required == 0) &&
11036 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
11037 			tcp_switch_back_to_default(tp);
11038 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
11039 			    tlen, iptos);
11040 			return (1);
11041 		}
11042 		/* Set the flag */
11043 		rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
11044 		tcp_set_hpts(tp->t_inpcb);
11045 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
11046 	}
11047 	if (thflags & TH_FIN)
11048 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
11049 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
11050 	if ((rack->rc_gp_dyn_mul) &&
11051 	    (rack->use_fixed_rate == 0) &&
11052 	    (rack->rc_always_pace)) {
11053 		/* Check in on probertt */
11054 		rack_check_probe_rtt(rack, us_cts);
11055 	}
11056 	if (rack->forced_ack) {
11057 		uint32_t us_rtt;
11058 
11059 		/*
11060 		 * A persist or keep-alive was forced out, update our
11061 		 * min rtt time. Note we do not worry about lost
11062 		 * retransmissions since KEEP-ALIVES and persists
11063 		 * are usually way long on times of sending (though
11064 		 * if we were really paranoid or worried we could
11065 		 * at least use timestamps if available to validate).
11066 		 */
11067 		rack->forced_ack = 0;
11068 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
11069 		if (us_rtt == 0)
11070 			us_rtt = 1;
11071 		rack_log_rtt_upd(tp, rack, us_rtt, 0, NULL, 3);
11072 		rack_apply_updated_usrtt(rack, us_rtt, us_cts);
11073 	}
11074 	/*
11075 	 * This is the one exception case where we set the rack state
11076 	 * always. All other times (timers etc) we must have a rack-state
11077 	 * set (so we assure we have done the checks above for SACK).
11078 	 */
11079 	rack->r_ctl.rc_rcvtime = cts;
11080 	if (rack->r_state != tp->t_state)
11081 		rack_set_state(tp, rack);
11082 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
11083 	    (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
11084 		kern_prefetch(rsm, &prev_state);
11085 	prev_state = rack->r_state;
11086 	rack_clear_rate_sample(rack);
11087 	retval = (*rack->r_substate) (m, th, so,
11088 	    tp, &to, drop_hdrlen,
11089 	    tlen, tiwin, thflags, nxt_pkt, iptos);
11090 #ifdef INVARIANTS
11091 	if ((retval == 0) &&
11092 	    (tp->t_inpcb == NULL)) {
11093 		panic("retval:%d tp:%p t_inpcb:NULL state:%d",
11094 		    retval, tp, prev_state);
11095 	}
11096 #endif
11097 	if (retval == 0) {
11098 		/*
11099 		 * If retval is 1 the tcb is unlocked and most likely the tp
11100 		 * is gone.
11101 		 */
11102 		INP_WLOCK_ASSERT(tp->t_inpcb);
11103 		if ((rack->rc_gp_dyn_mul) &&
11104 		    (rack->rc_always_pace) &&
11105 		    (rack->use_fixed_rate == 0) &&
11106 		    rack->in_probe_rtt &&
11107 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
11108 			/*
11109 			 * If we are going for target, lets recheck before
11110 			 * we output.
11111 			 */
11112 			rack_check_probe_rtt(rack, us_cts);
11113 		}
11114 		if (rack->set_pacing_done_a_iw == 0) {
11115 			/* How much has been acked? */
11116 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
11117 				/* We have enough to set in the pacing segment size */
11118 				rack->set_pacing_done_a_iw = 1;
11119 				rack_set_pace_segments(tp, rack, __LINE__);
11120 			}
11121 		}
11122 		tcp_rack_xmit_timer_commit(rack, tp);
11123 		if (nxt_pkt == 0) {
11124 			if (rack->r_wanted_output != 0) {
11125 do_output_now:
11126 				did_out = 1;
11127 				(void)tp->t_fb->tfb_tcp_output(tp);
11128 			}
11129 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
11130 		}
11131 		if ((nxt_pkt == 0) &&
11132 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11133 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
11134 		     (tp->t_flags & TF_DELACK) ||
11135 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11136 		      (tp->t_state <= TCPS_CLOSING)))) {
11137 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
11138 			if ((tp->snd_max == tp->snd_una) &&
11139 			    ((tp->t_flags & TF_DELACK) == 0) &&
11140 			    (rack->rc_inp->inp_in_hpts) &&
11141 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11142 				/* keep alive not needed if we are hptsi output yet */
11143 				;
11144 			} else {
11145 				int late = 0;
11146 				if (rack->rc_inp->inp_in_hpts) {
11147 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
11148 						us_cts = tcp_get_usecs(NULL);
11149 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
11150 							rack->r_early = 1;
11151 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
11152 						} else
11153 							late = 1;
11154 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11155 					}
11156 					tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
11157 				}
11158 				if (late && (did_out == 0)) {
11159 					/*
11160 					 * We are late in the sending
11161 					 * and we did not call the output
11162 					 * (this probably should not happen).
11163 					 */
11164 					goto do_output_now;
11165 				}
11166 				rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0);
11167 			}
11168 			way_out = 1;
11169 		} else if (nxt_pkt == 0) {
11170 			/* Do we have the correct timer running? */
11171 			rack_timer_audit(tp, rack, &so->so_snd);
11172 			way_out = 2;
11173 		}
11174 	done_with_input:
11175 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out);
11176 		if (did_out)
11177 			rack->r_wanted_output = 0;
11178 #ifdef INVARIANTS
11179 		if (tp->t_inpcb == NULL) {
11180 			panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
11181 			      did_out,
11182 			      retval, tp, prev_state);
11183 		}
11184 #endif
11185 	}
11186 	return (retval);
11187 }
11188 
11189 void
11190 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
11191     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11192 {
11193 	struct timeval tv;
11194 
11195 	/* First lets see if we have old packets */
11196 	if (tp->t_in_pkt) {
11197 		if (ctf_do_queued_segments(so, tp, 1)) {
11198 			m_freem(m);
11199 			return;
11200 		}
11201 	}
11202 	if (m->m_flags & M_TSTMP_LRO) {
11203 		tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
11204 		tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
11205 	} else {
11206 		/* Should not be should we kassert instead? */
11207 		tcp_get_usecs(&tv);
11208 	}
11209 	if(rack_do_segment_nounlock(m, th, so, tp,
11210 				    drop_hdrlen, tlen, iptos, 0, &tv) == 0)
11211 		INP_WUNLOCK(tp->t_inpcb);
11212 }
11213 
11214 struct rack_sendmap *
11215 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
11216 {
11217 	struct rack_sendmap *rsm = NULL;
11218 	int32_t idx;
11219 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
11220 
11221 	/* Return the next guy to be re-transmitted */
11222 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
11223 		return (NULL);
11224 	}
11225 	if (tp->t_flags & TF_SENTFIN) {
11226 		/* retran the end FIN? */
11227 		return (NULL);
11228 	}
11229 	/* ok lets look at this one */
11230 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11231 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
11232 		goto check_it;
11233 	}
11234 	rsm = rack_find_lowest_rsm(rack);
11235 	if (rsm == NULL) {
11236 		return (NULL);
11237 	}
11238 check_it:
11239 	if (rsm->r_flags & RACK_ACKED) {
11240 		return (NULL);
11241 	}
11242 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
11243 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
11244 		/* Its not yet ready */
11245 		return (NULL);
11246 	}
11247 	srtt = rack_grab_rtt(tp, rack);
11248 	idx = rsm->r_rtr_cnt - 1;
11249 	ts_low = rsm->r_tim_lastsent[idx];
11250 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
11251 	if ((tsused == ts_low) ||
11252 	    (TSTMP_LT(tsused, ts_low))) {
11253 		/* No time since sending */
11254 		return (NULL);
11255 	}
11256 	if ((tsused - ts_low) < thresh) {
11257 		/* It has not been long enough yet */
11258 		return (NULL);
11259 	}
11260 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
11261 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
11262 	     (rack->sack_attack_disable == 0))) {
11263 		/*
11264 		 * We have passed the dup-ack threshold <or>
11265 		 * a SACK has indicated this is missing.
11266 		 * Note that if you are a declared attacker
11267 		 * it is only the dup-ack threshold that
11268 		 * will cause retransmits.
11269 		 */
11270 		/* log retransmit reason */
11271 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
11272 		return (rsm);
11273 	}
11274 	return (NULL);
11275 }
11276 
11277 static void
11278 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
11279 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
11280 			   int line, struct rack_sendmap *rsm)
11281 {
11282 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
11283 		union tcp_log_stackspecific log;
11284 		struct timeval tv;
11285 
11286 		memset(&log, 0, sizeof(log));
11287 		log.u_bbr.flex1 = slot;
11288 		log.u_bbr.flex2 = len;
11289 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
11290 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
11291 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
11292 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
11293 		log.u_bbr.use_lt_bw = rack->app_limited_needs_set;
11294 		log.u_bbr.use_lt_bw <<= 1;
11295 		log.u_bbr.use_lt_bw = rack->rc_gp_filled;
11296 		log.u_bbr.use_lt_bw <<= 1;
11297 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
11298 		log.u_bbr.use_lt_bw <<= 1;
11299 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
11300 		log.u_bbr.pkt_epoch = line;
11301 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
11302 		log.u_bbr.bw_inuse = bw_est;
11303 		log.u_bbr.delRate = bw;
11304 		if (rack->r_ctl.gp_bw == 0)
11305 			log.u_bbr.cur_del_rate = 0;
11306 		else
11307 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
11308 		log.u_bbr.rttProp = len_time;
11309 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
11310 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
11311 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
11312 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
11313 			/* We are in slow start */
11314 			log.u_bbr.flex7 = 1;
11315 		} else {
11316 			/* we are on congestion avoidance */
11317 			log.u_bbr.flex7 = 0;
11318 		}
11319 		log.u_bbr.flex8 = method;
11320 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11321 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11322 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
11323 		log.u_bbr.cwnd_gain <<= 1;
11324 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
11325 		log.u_bbr.cwnd_gain <<= 1;
11326 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
11327 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
11328 		    &rack->rc_inp->inp_socket->so_rcv,
11329 		    &rack->rc_inp->inp_socket->so_snd,
11330 		    BBR_LOG_HPTSI_CALC, 0,
11331 		    0, &log, false, &tv);
11332 	}
11333 }
11334 
11335 static uint32_t
11336 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
11337 {
11338 	uint32_t new_tso, user_max;
11339 
11340 	user_max = rack->rc_user_set_max_segs * mss;
11341 	if (rack->rc_force_max_seg) {
11342 		return (user_max);
11343 	}
11344 	if (rack->use_fixed_rate &&
11345 	    ((rack->r_ctl.crte == NULL) ||
11346 	     (bw != rack->r_ctl.crte->rate))) {
11347 		/* Use the user mss since we are not exactly matched */
11348 		return (user_max);
11349 	}
11350 	new_tso = tcp_get_pacing_burst_size(bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
11351 	if (new_tso > user_max)
11352 		new_tso = user_max;
11353 	return(new_tso);
11354 }
11355 
11356 static void
11357 rack_log_hdwr_pacing(struct tcp_rack *rack, const struct ifnet *ifp,
11358 		     uint64_t rate, uint64_t hw_rate, int line,
11359 		     int error)
11360 {
11361 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
11362 		union tcp_log_stackspecific log;
11363 		struct timeval tv;
11364 
11365 		memset(&log, 0, sizeof(log));
11366 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
11367 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
11368 		log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
11369 		log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
11370 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11371 		log.u_bbr.bw_inuse = rate;
11372 		log.u_bbr.flex5 = line;
11373 		log.u_bbr.flex6 = error;
11374 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
11375 		log.u_bbr.flex8 = rack->use_fixed_rate;
11376 		log.u_bbr.flex8 <<= 1;
11377 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
11378 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
11379 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
11380 		    &rack->rc_inp->inp_socket->so_rcv,
11381 		    &rack->rc_inp->inp_socket->so_snd,
11382 		    BBR_LOG_HDWR_PACE, 0,
11383 		    0, &log, false, &tv);
11384 	}
11385 }
11386 
11387 static int32_t
11388 pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz)
11389 {
11390 	uint64_t lentim, fill_bw;
11391 
11392 	/* Lets first see if we are full, if so continue with normal rate */
11393 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
11394 		return (slot);
11395 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
11396 		return (slot);
11397 	if (rack->r_ctl.rc_last_us_rtt == 0)
11398 		return (slot);
11399 	if (rack->rc_pace_fill_if_rttin_range &&
11400 	    (rack->r_ctl.rc_last_us_rtt >=
11401 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
11402 		/* The rtt is huge, N * smallest, lets not fill */
11403 		return (slot);
11404 	}
11405 	/*
11406 	 * first lets calculate the b/w based on the last us-rtt
11407 	 * and the sndwnd.
11408 	 */
11409 	fill_bw = rack->r_ctl.cwnd_to_use;
11410 	/* Take the rwnd if its smaller */
11411 	if (fill_bw > rack->rc_tp->snd_wnd)
11412 		fill_bw = rack->rc_tp->snd_wnd;
11413 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
11414 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
11415 	/* We are below the min b/w */
11416 	if (fill_bw < RACK_MIN_BW)
11417 		return (slot);
11418 	/*
11419 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
11420 	 * in a rtt, what does that time wise equate too?
11421 	 */
11422 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
11423 	lentim /= fill_bw;
11424 	if (lentim < slot) {
11425 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
11426 					   0, lentim, 12, __LINE__, NULL);
11427 		return ((int32_t)lentim);
11428 	} else
11429 		return (slot);
11430 }
11431 
11432 static int32_t
11433 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
11434 {
11435 	struct rack_sendmap *lrsm;
11436 	int32_t slot = 0;
11437 	int err;
11438 
11439 	if (rack->rc_always_pace == 0) {
11440 		/*
11441 		 * We use the most optimistic possible cwnd/srtt for
11442 		 * sending calculations. This will make our
11443 		 * calculation anticipate getting more through
11444 		 * quicker then possible. But thats ok we don't want
11445 		 * the peer to have a gap in data sending.
11446 		 */
11447 		uint32_t srtt, cwnd, tr_perms = 0;
11448 		int32_t reduce = 0;
11449 
11450 	old_method:
11451 		/*
11452 		 * We keep no precise pacing with the old method
11453 		 * instead we use the pacer to mitigate bursts.
11454 		 */
11455 		rack->r_ctl.rc_agg_delayed = 0;
11456 		rack->r_early = 0;
11457 		rack->r_late = 0;
11458 		rack->r_ctl.rc_agg_early = 0;
11459 		if (rack->r_ctl.rc_rack_min_rtt)
11460 			srtt = rack->r_ctl.rc_rack_min_rtt;
11461 		else
11462 			srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT));
11463 		if (rack->r_ctl.rc_rack_largest_cwnd)
11464 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
11465 		else
11466 			cwnd = rack->r_ctl.cwnd_to_use;
11467 		tr_perms = cwnd / srtt;
11468 		if (tr_perms == 0) {
11469 			tr_perms = ctf_fixed_maxseg(tp);
11470 		}
11471 		/*
11472 		 * Calculate how long this will take to drain, if
11473 		 * the calculation comes out to zero, thats ok we
11474 		 * will use send_a_lot to possibly spin around for
11475 		 * more increasing tot_len_this_send to the point
11476 		 * that its going to require a pace, or we hit the
11477 		 * cwnd. Which in that case we are just waiting for
11478 		 * a ACK.
11479 		 */
11480 		slot = len / tr_perms;
11481 		/* Now do we reduce the time so we don't run dry? */
11482 		if (slot && rack_slot_reduction) {
11483 			reduce = (slot / rack_slot_reduction);
11484 			if (reduce < slot) {
11485 				slot -= reduce;
11486 			} else
11487 				slot = 0;
11488 		}
11489 		slot *=  HPTS_USEC_IN_MSEC;
11490 		if (rsm == NULL) {
11491 			/*
11492 			 * We always consider ourselves app limited with old style
11493 			 * that are not retransmits. This could be the initial
11494 			 * measurement, but thats ok its all setup and specially
11495 			 * handled. If another send leaks out, then that too will
11496 			 * be mark app-limited.
11497 			 */
11498 			lrsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
11499 			if (lrsm && ((lrsm->r_flags & RACK_APP_LIMITED) == 0)) {
11500 				rack->r_ctl.rc_first_appl = lrsm;
11501 				lrsm->r_flags |= RACK_APP_LIMITED;
11502 				rack->r_ctl.rc_app_limited_cnt++;
11503 			}
11504 		}
11505 		rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL);
11506 	} else {
11507 		uint64_t bw_est, res, lentim, rate_wanted;
11508 		uint32_t orig_val, srtt, segs, oh;
11509 
11510 		if ((rack->r_rr_config == 1) && rsm) {
11511 			return (rack->r_ctl.rc_min_to * HPTS_USEC_IN_MSEC);
11512 		}
11513 		if (rack->use_fixed_rate) {
11514 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
11515 		} else if ((rack->r_ctl.init_rate == 0) &&
11516 #ifdef NETFLIX_PEAKRATE
11517 			   (rack->rc_tp->t_maxpeakrate == 0) &&
11518 #endif
11519 			   (rack->r_ctl.gp_bw == 0)) {
11520 			/* no way to yet do an estimate */
11521 			bw_est = rate_wanted = 0;
11522 		} else {
11523 			bw_est = rack_get_bw(rack);
11524 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm);
11525 		}
11526 		if ((bw_est == 0) || (rate_wanted == 0)) {
11527 			/*
11528 			 * No way yet to make a b/w estimate or
11529 			 * our raise is set incorrectly.
11530 			 */
11531 			goto old_method;
11532 		}
11533 		/* We need to account for all the overheads */
11534 		segs = (len + segsiz - 1) / segsiz;
11535 		/*
11536 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
11537 		 * and how much data we put in each packet. Yes this
11538 		 * means we may be off if we are larger than 1500 bytes
11539 		 * or smaller. But this just makes us more conservative.
11540 		 */
11541 		if (ETHERNET_SEGMENT_SIZE > segsiz)
11542 			oh = ETHERNET_SEGMENT_SIZE - segsiz;
11543 		else
11544 			oh = 0;
11545 		segs *= oh;
11546 		lentim = (uint64_t)(len + segs)  * (uint64_t)HPTS_USEC_IN_SEC;
11547 		res = lentim / rate_wanted;
11548 		slot = (uint32_t)res;
11549 		orig_val = rack->r_ctl.rc_pace_max_segs;
11550 		rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
11551 		/* Did we change the TSO size, if so log it */
11552 		if (rack->r_ctl.rc_pace_max_segs != orig_val)
11553 			rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL);
11554 		if ((rack->rc_pace_to_cwnd) &&
11555 		    (rack->in_probe_rtt == 0) &&
11556 		    (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
11557 			/*
11558 			 * We want to pace at our rate *or* faster to
11559 			 * fill the cwnd to the max if its not full.
11560 			 */
11561 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz);
11562 		}
11563 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
11564 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
11565 			if ((rack->rack_hdw_pace_ena) &&
11566 			    (rack->rack_hdrw_pacing == 0) &&
11567 			    (rack->rack_attempt_hdwr_pace == 0)) {
11568 				/*
11569 				 * Lets attempt to turn on hardware pacing
11570 				 * if we can.
11571 				 */
11572 				rack->rack_attempt_hdwr_pace = 1;
11573 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
11574 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
11575 								       rate_wanted,
11576 								       RS_PACING_GEQ,
11577 								       &err);
11578 				if (rack->r_ctl.crte) {
11579 					rack->rack_hdrw_pacing = 1;
11580 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(rate_wanted, segsiz,
11581 												 0, rack->r_ctl.crte,
11582 												 NULL);
11583 					rack_log_hdwr_pacing(rack, rack->rc_inp->inp_route.ro_nh->nh_ifp,
11584 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
11585 							     err);
11586 				}
11587 			} else if (rack->rack_hdrw_pacing &&
11588 				   (rack->r_ctl.crte->rate != rate_wanted)) {
11589 				/* Do we need to adjust our rate? */
11590 				const struct tcp_hwrate_limit_table *nrte;
11591 
11592 				nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
11593 							   rack->rc_tp,
11594 							   rack->rc_inp->inp_route.ro_nh->nh_ifp,
11595 							   rate_wanted,
11596 							   RS_PACING_GEQ,
11597 							   &err);
11598 				if (nrte == NULL) {
11599 					/* Lost the rate */
11600 					rack->rack_hdrw_pacing = 0;
11601 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
11602 				} else if (nrte != rack->r_ctl.crte) {
11603 					rack->r_ctl.crte = nrte;
11604 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(rate_wanted,
11605 												 segsiz, 0,
11606 												 rack->r_ctl.crte,
11607 												 NULL);
11608 					rack_log_hdwr_pacing(rack, rack->rc_inp->inp_route.ro_nh->nh_ifp,
11609 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
11610 							     err);
11611 				}
11612 
11613 			}
11614 		}
11615 		if (rack_limit_time_with_srtt &&
11616 		    (rack->use_fixed_rate == 0) &&
11617 #ifdef NETFLIX_PEAKRATE
11618 		    (rack->rc_tp->t_maxpeakrate == 0) &&
11619 #endif
11620 		    (rack->rack_hdrw_pacing == 0)) {
11621 			/*
11622 			 * Sanity check, we do not allow the pacing delay
11623 			 * to be longer than the SRTT of the path. If it is
11624 			 * a slow path, then adding a packet should increase
11625 			 * the RTT and compensate for this i.e. the srtt will
11626 			 * be greater so the allowed pacing time will be greater.
11627 			 *
11628 			 * Note this restriction is not for where a peak rate
11629 			 * is set, we are doing fixed pacing or hardware pacing.
11630 			 */
11631 			if (rack->rc_tp->t_srtt)
11632 				srtt = (TICKS_2_USEC(rack->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
11633 			else
11634 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
11635 			if (srtt < slot) {
11636 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL);
11637 				slot = srtt;
11638 			}
11639 		}
11640 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm);
11641 	}
11642 	if (slot)
11643 		counter_u64_add(rack_calc_nonzero, 1);
11644 	else
11645 		counter_u64_add(rack_calc_zero, 1);
11646 	return (slot);
11647 }
11648 
11649 static void
11650 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
11651     tcp_seq startseq, uint32_t sb_offset)
11652 {
11653 	struct rack_sendmap *my_rsm = NULL;
11654 	struct rack_sendmap fe;
11655 
11656 	if (tp->t_state < TCPS_ESTABLISHED) {
11657 		/*
11658 		 * We don't start any measurements if we are
11659 		 * not at least established.
11660 		 */
11661 		return;
11662 	}
11663 	tp->t_flags |= TF_GPUTINPROG;
11664 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
11665 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
11666 	tp->gput_seq = startseq;
11667 	rack->app_limited_needs_set = 0;
11668 	if (rack->in_probe_rtt)
11669 		rack->measure_saw_probe_rtt = 1;
11670 	else if ((rack->measure_saw_probe_rtt) &&
11671 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
11672 		rack->measure_saw_probe_rtt = 0;
11673 	if (rack->rc_gp_filled)
11674 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
11675 	else {
11676 		/* Special case initial measurement */
11677 		rack->r_ctl.rc_gp_output_ts = tp->gput_ts = tcp_get_usecs(NULL);
11678 	}
11679 	/*
11680 	 * We take a guess out into the future,
11681 	 * if we have no measurement and no
11682 	 * initial rate, we measure the first
11683 	 * initial-windows worth of data to
11684 	 * speed up getting some GP measurement and
11685 	 * thus start pacing.
11686 	 */
11687 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
11688 		rack->app_limited_needs_set = 1;
11689 		tp->gput_ack = startseq + max(rc_init_window(rack),
11690 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
11691 		rack_log_pacing_delay_calc(rack,
11692 					   tp->gput_seq,
11693 					   tp->gput_ack,
11694 					   0,
11695 					   tp->gput_ts,
11696 					   rack->r_ctl.rc_app_limited_cnt,
11697 					   9,
11698 					   __LINE__, NULL);
11699 		return;
11700 	}
11701 	if (sb_offset) {
11702 		/*
11703 		 * We are out somewhere in the sb
11704 		 * can we use the already outstanding data?
11705 		 */
11706 
11707 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
11708 			/*
11709 			 * Yes first one is good and in this case
11710 			 * the tp->gput_ts is correctly set based on
11711 			 * the last ack that arrived (no need to
11712 			 * set things up when an ack comes in).
11713 			 */
11714 			my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
11715 			if ((my_rsm == NULL) ||
11716 			    (my_rsm->r_rtr_cnt != 1)) {
11717 				/* retransmission? */
11718 				goto use_latest;
11719 			}
11720 		} else {
11721 			if (rack->r_ctl.rc_first_appl == NULL) {
11722 				/*
11723 				 * If rc_first_appl is NULL
11724 				 * then the cnt should be 0.
11725 				 * This is probably an error, maybe
11726 				 * a KASSERT would be approprate.
11727 				 */
11728 				goto use_latest;
11729 			}
11730 			/*
11731 			 * If we have a marker pointer to the last one that is
11732 			 * app limited we can use that, but we need to set
11733 			 * things up so that when it gets ack'ed we record
11734 			 * the ack time (if its not already acked).
11735 			 */
11736 			rack->app_limited_needs_set = 1;
11737 			/*
11738 			 * We want to get to the rsm that is either
11739 			 * next with space i.e. over 1 MSS or the one
11740 			 * after that (after the app-limited).
11741 			 */
11742 			my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
11743 					 rack->r_ctl.rc_first_appl);
11744 			if (my_rsm) {
11745 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
11746 					/* Have to use the next one */
11747 					my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
11748 							 my_rsm);
11749 				else {
11750 					/* Use after the first MSS of it is acked */
11751 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
11752 					goto start_set;
11753 				}
11754 			}
11755 			if ((my_rsm == NULL) ||
11756 			    (my_rsm->r_rtr_cnt != 1)) {
11757 				/*
11758 				 * Either its a retransmit or
11759 				 * the last is the app-limited one.
11760 				 */
11761 				goto use_latest;
11762 			}
11763 		}
11764 		tp->gput_seq = my_rsm->r_start;
11765 start_set:
11766 		if (my_rsm->r_flags & RACK_ACKED) {
11767 			/*
11768 			 * This one has been acked use the arrival ack time
11769 			 */
11770 			tp->gput_ts = my_rsm->r_ack_arrival;
11771 			rack->app_limited_needs_set = 0;
11772 		}
11773 		rack->r_ctl.rc_gp_output_ts = my_rsm->usec_orig_send;
11774 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
11775 		rack_log_pacing_delay_calc(rack,
11776 					   tp->gput_seq,
11777 					   tp->gput_ack,
11778 					   (uint64_t)my_rsm,
11779 					   tp->gput_ts,
11780 					   rack->r_ctl.rc_app_limited_cnt,
11781 					   9,
11782 					   __LINE__, NULL);
11783 		return;
11784 	}
11785 
11786 use_latest:
11787 	/*
11788 	 * We don't know how long we may have been
11789 	 * idle or if this is the first-send. Lets
11790 	 * setup the flag so we will trim off
11791 	 * the first ack'd data so we get a true
11792 	 * measurement.
11793 	 */
11794 	rack->app_limited_needs_set = 1;
11795 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
11796 	/* Find this guy so we can pull the send time */
11797 	fe.r_start = startseq;
11798 	my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
11799 	if (my_rsm) {
11800 		rack->r_ctl.rc_gp_output_ts = my_rsm->usec_orig_send;
11801 		if (my_rsm->r_flags & RACK_ACKED) {
11802 			/*
11803 			 * Unlikely since its probably what was
11804 			 * just transmitted (but I am paranoid).
11805 			 */
11806 			tp->gput_ts = my_rsm->r_ack_arrival;
11807 			rack->app_limited_needs_set = 0;
11808 		}
11809 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
11810 			/* This also is unlikely */
11811 			tp->gput_seq = my_rsm->r_start;
11812 		}
11813 	} else {
11814 		/*
11815 		 * TSNH unless we have some send-map limit,
11816 		 * and even at that it should not be hitting
11817 		 * that limit (we should have stopped sending).
11818 		 */
11819 		rack->r_ctl.rc_gp_output_ts = tcp_get_usecs(NULL);
11820 	}
11821 	rack_log_pacing_delay_calc(rack,
11822 				   tp->gput_seq,
11823 				   tp->gput_ack,
11824 				   (uint64_t)my_rsm,
11825 				   tp->gput_ts,
11826 				   rack->r_ctl.rc_app_limited_cnt,
11827 				   9, __LINE__, NULL);
11828 }
11829 
11830 static inline uint32_t
11831 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
11832     uint32_t avail, int32_t sb_offset)
11833 {
11834 	uint32_t len;
11835 	uint32_t sendwin;
11836 
11837 	if (tp->snd_wnd > cwnd_to_use)
11838 		sendwin = cwnd_to_use;
11839 	else
11840 		sendwin = tp->snd_wnd;
11841 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
11842 		/* We never want to go over our peers rcv-window */
11843 		len = 0;
11844 	} else {
11845 		uint32_t flight;
11846 
11847 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
11848 		if (flight >= sendwin) {
11849 			/*
11850 			 * We have in flight what we are allowed by cwnd (if
11851 			 * it was rwnd blocking it would have hit above out
11852 			 * >= tp->snd_wnd).
11853 			 */
11854 			return (0);
11855 		}
11856 		len = sendwin - flight;
11857 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11858 			/* We would send too much (beyond the rwnd) */
11859 			len = tp->snd_wnd - ctf_outstanding(tp);
11860 		}
11861 		if ((len + sb_offset) > avail) {
11862 			/*
11863 			 * We don't have that much in the SB, how much is
11864 			 * there?
11865 			 */
11866 			len = avail - sb_offset;
11867 		}
11868 	}
11869 	return (len);
11870 }
11871 
11872 static int
11873 rack_output(struct tcpcb *tp)
11874 {
11875 	struct socket *so;
11876 	uint32_t recwin;
11877 	uint32_t sb_offset;
11878 	int32_t len, flags, error = 0;
11879 	struct mbuf *m;
11880 	struct mbuf *mb;
11881 	uint32_t if_hw_tsomaxsegcount = 0;
11882 	uint32_t if_hw_tsomaxsegsize;
11883 	int32_t segsiz, minseg;
11884 	long tot_len_this_send = 0;
11885 	struct ip *ip = NULL;
11886 #ifdef TCPDEBUG
11887 	struct ipovly *ipov = NULL;
11888 #endif
11889 	struct udphdr *udp = NULL;
11890 	struct tcp_rack *rack;
11891 	struct tcphdr *th;
11892 	uint8_t pass = 0;
11893 	uint8_t mark = 0;
11894 	uint8_t wanted_cookie = 0;
11895 	u_char opt[TCP_MAXOLEN];
11896 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
11897 	uint32_t rack_seq;
11898 
11899 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
11900 	unsigned ipsec_optlen = 0;
11901 
11902 #endif
11903 	int32_t idle, sendalot;
11904 	int32_t sub_from_prr = 0;
11905 	volatile int32_t sack_rxmit;
11906 	struct rack_sendmap *rsm = NULL;
11907 	int32_t tso, mtu;
11908 	struct tcpopt to;
11909 	int32_t slot = 0;
11910 	int32_t sup_rack = 0;
11911 	uint32_t cts, us_cts, delayed, early;
11912 	uint8_t hpts_calling, new_data_tlp = 0, doing_tlp = 0;
11913 	uint32_t cwnd_to_use;
11914 	int32_t do_a_prefetch;
11915 	int32_t prefetch_rsm = 0;
11916 	int32_t orig_len;
11917 	struct timeval tv;
11918 	int32_t prefetch_so_done = 0;
11919 	struct tcp_log_buffer *lgb = NULL;
11920 	struct inpcb *inp;
11921 	struct sockbuf *sb;
11922 #ifdef INET6
11923 	struct ip6_hdr *ip6 = NULL;
11924 	int32_t isipv6;
11925 #endif
11926 	uint8_t filled_all = 0;
11927 	bool hw_tls = false;
11928 
11929 	/* setup and take the cache hits here */
11930 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11931 	inp = rack->rc_inp;
11932 	so = inp->inp_socket;
11933 	sb = &so->so_snd;
11934 	kern_prefetch(sb, &do_a_prefetch);
11935 	do_a_prefetch = 1;
11936 	hpts_calling = inp->inp_hpts_calls;
11937 	hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
11938 
11939 	NET_EPOCH_ASSERT();
11940 	INP_WLOCK_ASSERT(inp);
11941 #ifdef TCP_OFFLOAD
11942 	if (tp->t_flags & TF_TOE)
11943 		return (tcp_offload_output(tp));
11944 #endif
11945 	/*
11946 	 * For TFO connections in SYN_RECEIVED, only allow the initial
11947 	 * SYN|ACK and those sent by the retransmit timer.
11948 	 */
11949 	if (IS_FASTOPEN(tp->t_flags) &&
11950 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
11951 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
11952 	    (rack->r_ctl.rc_resend == NULL))         /* not a retransmit */
11953 		return (0);
11954 #ifdef INET6
11955 	if (rack->r_state) {
11956 		/* Use the cache line loaded if possible */
11957 		isipv6 = rack->r_is_v6;
11958 	} else {
11959 		isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
11960 	}
11961 #endif
11962 	early = 0;
11963 	us_cts = tcp_get_usecs(&tv);
11964 	cts = tcp_tv_to_mssectick(&tv);
11965 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
11966 	    inp->inp_in_hpts) {
11967 		/*
11968 		 * We are on the hpts for some timer but not hptsi output.
11969 		 * Remove from the hpts unconditionally.
11970 		 */
11971 		rack_timer_cancel(tp, rack, cts, __LINE__);
11972 	}
11973 	/* Are we pacing and late? */
11974 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11975 	    TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) {
11976 		/* We are delayed */
11977 		delayed = us_cts - rack->r_ctl.rc_last_output_to;
11978 	} else {
11979 		delayed = 0;
11980 	}
11981 	/* Do the timers, which may override the pacer  */
11982 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
11983 		if (rack_process_timers(tp, rack, cts, hpts_calling)) {
11984 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
11985 			return (0);
11986 		}
11987 	}
11988 	if ((rack->r_timer_override) ||
11989 	    (delayed) ||
11990 	    (tp->t_state < TCPS_ESTABLISHED)) {
11991 		if (tp->t_inpcb->inp_in_hpts)
11992 			tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
11993 	} else if (tp->t_inpcb->inp_in_hpts) {
11994 		/*
11995 		 * On the hpts you can't pass even if ACKNOW is on, we will
11996 		 * when the hpts fires.
11997 		 */
11998 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
11999 		return (0);
12000 	}
12001 	inp->inp_hpts_calls = 0;
12002 	/* Finish out both pacing early and late accounting */
12003 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
12004 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12005 		early = rack->r_ctl.rc_last_output_to - us_cts;
12006 	} else
12007 		early = 0;
12008 	if (delayed) {
12009 		rack->r_ctl.rc_agg_delayed += delayed;
12010 		rack->r_late = 1;
12011 	} else if (early) {
12012 		rack->r_ctl.rc_agg_early += early;
12013 		rack->r_early = 1;
12014 	}
12015 	/* Now that early/late accounting is done turn off the flag */
12016 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12017 	rack->r_wanted_output = 0;
12018 	rack->r_timer_override = 0;
12019 	/*
12020 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
12021 	 * only allow the initial SYN or SYN|ACK and those sent
12022 	 * by the retransmit timer.
12023 	 */
12024 	if (IS_FASTOPEN(tp->t_flags) &&
12025 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
12026 	     (tp->t_state == TCPS_SYN_SENT)) &&
12027 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
12028 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
12029 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12030 		goto just_return_nolock;
12031 	}
12032 	/*
12033 	 * Determine length of data that should be transmitted, and flags
12034 	 * that will be used. If there is some data or critical controls
12035 	 * (SYN, RST) to send, then transmit; otherwise, investigate
12036 	 * further.
12037 	 */
12038 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
12039 	if (tp->t_idle_reduce) {
12040 		if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
12041 			rack_cc_after_idle(rack, tp);
12042 	}
12043 	tp->t_flags &= ~TF_LASTIDLE;
12044 	if (idle) {
12045 		if (tp->t_flags & TF_MORETOCOME) {
12046 			tp->t_flags |= TF_LASTIDLE;
12047 			idle = 0;
12048 		}
12049 	}
12050 	if ((tp->snd_una == tp->snd_max) &&
12051 	    rack->r_ctl.rc_went_idle_time &&
12052 	    TSTMP_GT(us_cts, rack->r_ctl.rc_went_idle_time)) {
12053 		idle = us_cts - rack->r_ctl.rc_went_idle_time;
12054 		if (idle > rack_min_probertt_hold) {
12055 			/* Count as a probe rtt */
12056 			if (rack->in_probe_rtt == 0) {
12057 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12058 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
12059 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
12060 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
12061 			} else {
12062 				rack_exit_probertt(rack, us_cts);
12063 			}
12064 		}
12065 		idle = 0;
12066 	}
12067 again:
12068 	/*
12069 	 * If we've recently taken a timeout, snd_max will be greater than
12070 	 * snd_nxt.  There may be SACK information that allows us to avoid
12071 	 * resending already delivered data.  Adjust snd_nxt accordingly.
12072 	 */
12073 	sendalot = 0;
12074 	us_cts = tcp_get_usecs(&tv);
12075 	cts = tcp_tv_to_mssectick(&tv);
12076 	tso = 0;
12077 	mtu = 0;
12078 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
12079 	minseg = segsiz;
12080 	sb_offset = tp->snd_max - tp->snd_una;
12081 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12082 #ifdef NETFLIX_SHARED_CWND
12083 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
12084 	    rack->rack_enable_scwnd) {
12085 		/* We are doing cwnd sharing */
12086 		if (rack->rc_gp_filled &&
12087 		    (rack->rack_attempted_scwnd == 0) &&
12088 		    (rack->r_ctl.rc_scw == NULL) &&
12089 		    tp->t_lib) {
12090 			/* The pcbid is in, lets make an attempt */
12091 			counter_u64_add(rack_try_scwnd, 1);
12092 			rack->rack_attempted_scwnd = 1;
12093 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
12094 								   &rack->r_ctl.rc_scw_index,
12095 								   segsiz);
12096 		}
12097 		if (rack->r_ctl.rc_scw &&
12098 		    (rack->rack_scwnd_is_idle == 1) &&
12099 		    (rack->rc_in_persist == 0) &&
12100 		    sbavail(sb)) {
12101 			/* we are no longer out of data */
12102 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
12103 			rack->rack_scwnd_is_idle = 0;
12104 		}
12105 		if (rack->r_ctl.rc_scw) {
12106 			/* First lets update and get the cwnd */
12107 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
12108 								    rack->r_ctl.rc_scw_index,
12109 								    tp->snd_cwnd, tp->snd_wnd, segsiz);
12110 		}
12111 	}
12112 #endif
12113 	flags = tcp_outflags[tp->t_state];
12114 	while (rack->rc_free_cnt < rack_free_cache) {
12115 		rsm = rack_alloc(rack);
12116 		if (rsm == NULL) {
12117 			if (inp->inp_hpts_calls)
12118 				/* Retry in a ms */
12119 				slot = (1 * HPTS_USEC_IN_MSEC);
12120 			goto just_return_nolock;
12121 		}
12122 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
12123 		rack->rc_free_cnt++;
12124 		rsm = NULL;
12125 	}
12126 	if (inp->inp_hpts_calls)
12127 		inp->inp_hpts_calls = 0;
12128 	sack_rxmit = 0;
12129 	len = 0;
12130 	rsm = NULL;
12131 	if (flags & TH_RST) {
12132 		SOCKBUF_LOCK(sb);
12133 		goto send;
12134 	}
12135 	if (rack->r_ctl.rc_resend) {
12136 		/* Retransmit timer */
12137 		rsm = rack->r_ctl.rc_resend;
12138 		rack->r_ctl.rc_resend = NULL;
12139 		rsm->r_flags &= ~RACK_TLP;
12140 		len = rsm->r_end - rsm->r_start;
12141 		sack_rxmit = 1;
12142 		sendalot = 0;
12143 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
12144 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
12145 			 __func__, __LINE__,
12146 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
12147 		sb_offset = rsm->r_start - tp->snd_una;
12148 		if (len >= segsiz)
12149 			len = segsiz;
12150 	} else if ((rack->rc_in_persist == 0) &&
12151 		   ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) {
12152 		/* We have a retransmit that takes precedence */
12153 		rsm->r_flags &= ~RACK_TLP;
12154 		if ((!IN_RECOVERY(tp->t_flags)) &&
12155 		    ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) {
12156 			/* Enter recovery if not induced by a time-out */
12157 			rack->r_ctl.rc_rsm_start = rsm->r_start;
12158 			rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
12159 			rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
12160 			rack_cong_signal(tp, NULL, CC_NDUPACK);
12161 			/*
12162 			 * When we enter recovery we need to assure we send
12163 			 * one packet.
12164 			 */
12165 			if (rack->rack_no_prr == 0) {
12166 				rack->r_ctl.rc_prr_sndcnt = segsiz;
12167 				rack_log_to_prr(rack, 13, 0);
12168 			}
12169 		}
12170 #ifdef INVARIANTS
12171 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12172 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
12173 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
12174 		}
12175 #endif
12176 		len = rsm->r_end - rsm->r_start;
12177 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
12178 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
12179 			 __func__, __LINE__,
12180 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
12181 		sb_offset = rsm->r_start - tp->snd_una;
12182 		/* Can we send it within the PRR boundary? */
12183 		if (rack->rack_no_prr == 0) {
12184 			if ((rack->use_rack_rr == 0) && (len > rack->r_ctl.rc_prr_sndcnt)) {
12185 				/* It does not fit */
12186 				if ((ctf_flight_size(tp, rack->r_ctl.rc_sacked) > len) &&
12187 				    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
12188 					/*
12189 					 * prr is less than a segment, we
12190 					 * have more acks due in besides
12191 					 * what we need to resend. Lets not send
12192 					 * to avoid sending small pieces of
12193 					 * what we need to retransmit.
12194 					 */
12195 					len = 0;
12196 					goto just_return_nolock;
12197 				}
12198 				len = rack->r_ctl.rc_prr_sndcnt;
12199 			}
12200 		}
12201 		sendalot = 0;
12202 		if (len >= segsiz)
12203 			len = segsiz;
12204 		if (len > 0) {
12205 			sub_from_prr = 1;
12206 			sack_rxmit = 1;
12207 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
12208 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12209 			    min(len, segsiz));
12210 			counter_u64_add(rack_rtm_prr_retran, 1);
12211 		}
12212 	} else 	if (rack->r_ctl.rc_tlpsend) {
12213 		/* Tail loss probe */
12214 		long cwin;
12215 		long tlen;
12216 
12217 		doing_tlp = 1;
12218 		/*
12219 		 * Check if we can do a TLP with a RACK'd packet
12220 		 * this can happen if we are not doing the rack
12221 		 * cheat and we skipped to a TLP and it
12222 		 * went off.
12223 		 */
12224 		rsm = rack->r_ctl.rc_tlpsend;
12225 		rsm->r_flags |= RACK_TLP;
12226 		rack->r_ctl.rc_tlpsend = NULL;
12227 		sack_rxmit = 1;
12228 		tlen = rsm->r_end - rsm->r_start;
12229 		if (tlen > segsiz)
12230 			tlen = segsiz;
12231 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
12232 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
12233 			 __func__, __LINE__,
12234 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
12235 		sb_offset = rsm->r_start - tp->snd_una;
12236 		cwin = min(tp->snd_wnd, tlen);
12237 		len = cwin;
12238 	}
12239 	/*
12240 	 * Enforce a connection sendmap count limit if set
12241 	 * as long as we are not retransmiting.
12242 	 */
12243 	if ((rsm == NULL) &&
12244 	    (rack->do_detection == 0) &&
12245 	    (V_tcp_map_entries_limit > 0) &&
12246 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
12247 		counter_u64_add(rack_to_alloc_limited, 1);
12248 		if (!rack->alloc_limit_reported) {
12249 			rack->alloc_limit_reported = 1;
12250 			counter_u64_add(rack_alloc_limited_conns, 1);
12251 		}
12252 		goto just_return_nolock;
12253 	}
12254 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
12255 		/* we are retransmitting the fin */
12256 		len--;
12257 		if (len) {
12258 			/*
12259 			 * When retransmitting data do *not* include the
12260 			 * FIN. This could happen from a TLP probe.
12261 			 */
12262 			flags &= ~TH_FIN;
12263 		}
12264 	}
12265 #ifdef INVARIANTS
12266 	/* For debugging */
12267 	rack->r_ctl.rc_rsm_at_retran = rsm;
12268 #endif
12269 	/*
12270 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
12271 	 * state flags.
12272 	 */
12273 	if (tp->t_flags & TF_NEEDFIN)
12274 		flags |= TH_FIN;
12275 	if (tp->t_flags & TF_NEEDSYN)
12276 		flags |= TH_SYN;
12277 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12278 		void *end_rsm;
12279 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
12280 		if (end_rsm)
12281 			kern_prefetch(end_rsm, &prefetch_rsm);
12282 		prefetch_rsm = 1;
12283 	}
12284 	SOCKBUF_LOCK(sb);
12285 	/*
12286 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
12287 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12288 	 * negative length.  This can also occur when TCP opens up its
12289 	 * congestion window while receiving additional duplicate acks after
12290 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
12291 	 * the fast-retransmit.
12292 	 *
12293 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
12294 	 * set to snd_una, the sb_offset will be 0, and the length may wind
12295 	 * up 0.
12296 	 *
12297 	 * If sack_rxmit is true we are retransmitting from the scoreboard
12298 	 * in which case len is already set.
12299 	 */
12300 	if ((sack_rxmit == 0) && TCPS_HAVEESTABLISHED(tp->t_state)) {
12301 		uint32_t avail;
12302 
12303 		avail = sbavail(sb);
12304 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
12305 			sb_offset = tp->snd_nxt - tp->snd_una;
12306 		else
12307 			sb_offset = 0;
12308 		if ((IN_RECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
12309 			if (rack->r_ctl.rc_tlp_new_data) {
12310 				/* TLP is forcing out new data */
12311 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
12312 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
12313 				}
12314 				if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd)
12315 					len = tp->snd_wnd;
12316 				else
12317 					len = rack->r_ctl.rc_tlp_new_data;
12318 				rack->r_ctl.rc_tlp_new_data = 0;
12319 				new_data_tlp = doing_tlp = 1;
12320 			}  else
12321 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
12322 			if (IN_RECOVERY(tp->t_flags) && (len > segsiz)) {
12323 				/*
12324 				 * For prr=off, we need to send only 1 MSS
12325 				 * at a time. We do this because another sack could
12326 				 * be arriving that causes us to send retransmits and
12327 				 * we don't want to be on a long pace due to a larger send
12328 				 * that keeps us from sending out the retransmit.
12329 				 */
12330 				len = segsiz;
12331 			}
12332 		} else {
12333 			uint32_t outstanding;
12334 
12335 			/*
12336 			 * We are inside of a SACK recovery episode and are
12337 			 * sending new data, having retransmitted all the
12338 			 * data possible so far in the scoreboard.
12339 			 */
12340 			outstanding = tp->snd_max - tp->snd_una;
12341 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
12342 				if (tp->snd_wnd > outstanding) {
12343 					len = tp->snd_wnd - outstanding;
12344 					/* Check to see if we have the data */
12345 					if ((sb_offset + len) > avail) {
12346 						/* It does not all fit */
12347 						if (avail > sb_offset)
12348 							len = avail - sb_offset;
12349 						else
12350 							len = 0;
12351 					}
12352 				} else
12353 					len = 0;
12354 			} else if (avail > sb_offset)
12355 				len = avail - sb_offset;
12356 			else
12357 				len = 0;
12358 			if (len > 0) {
12359 				if (len > rack->r_ctl.rc_prr_sndcnt)
12360 					len = rack->r_ctl.rc_prr_sndcnt;
12361 				if (len > 0) {
12362 					sub_from_prr = 1;
12363 					counter_u64_add(rack_rtm_prr_newdata, 1);
12364 				}
12365 			}
12366 			if (len > segsiz) {
12367 				/*
12368 				 * We should never send more than a MSS when
12369 				 * retransmitting or sending new data in prr
12370 				 * mode unless the override flag is on. Most
12371 				 * likely the PRR algorithm is not going to
12372 				 * let us send a lot as well :-)
12373 				 */
12374 				if (rack->r_ctl.rc_prr_sendalot == 0)
12375 					len = segsiz;
12376 			} else if (len < segsiz) {
12377 				/*
12378 				 * Do we send any? The idea here is if the
12379 				 * send empty's the socket buffer we want to
12380 				 * do it. However if not then lets just wait
12381 				 * for our prr_sndcnt to get bigger.
12382 				 */
12383 				long leftinsb;
12384 
12385 				leftinsb = sbavail(sb) - sb_offset;
12386 				if (leftinsb > len) {
12387 					/* This send does not empty the sb */
12388 					len = 0;
12389 				}
12390 			}
12391 		}
12392 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
12393 		/*
12394 		 * If you have not established
12395 		 * and are not doing FAST OPEN
12396 		 * no data please.
12397 		 */
12398 		if ((sack_rxmit == 0) &&
12399 		    (!IS_FASTOPEN(tp->t_flags))){
12400 			len = 0;
12401 			sb_offset = 0;
12402 		}
12403 	}
12404 	if (prefetch_so_done == 0) {
12405 		kern_prefetch(so, &prefetch_so_done);
12406 		prefetch_so_done = 1;
12407 	}
12408 	/*
12409 	 * Lop off SYN bit if it has already been sent.  However, if this is
12410 	 * SYN-SENT state and if segment contains data and if we don't know
12411 	 * that foreign host supports TAO, suppress sending segment.
12412 	 */
12413 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
12414 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
12415 		/*
12416 		 * When sending additional segments following a TFO SYN|ACK,
12417 		 * do not include the SYN bit.
12418 		 */
12419 		if (IS_FASTOPEN(tp->t_flags) &&
12420 		    (tp->t_state == TCPS_SYN_RECEIVED))
12421 			flags &= ~TH_SYN;
12422 	}
12423 	/*
12424 	 * Be careful not to send data and/or FIN on SYN segments. This
12425 	 * measure is needed to prevent interoperability problems with not
12426 	 * fully conformant TCP implementations.
12427 	 */
12428 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12429 		len = 0;
12430 		flags &= ~TH_FIN;
12431 	}
12432 	/*
12433 	 * On TFO sockets, ensure no data is sent in the following cases:
12434 	 *
12435 	 *  - When retransmitting SYN|ACK on a passively-created socket
12436 	 *
12437 	 *  - When retransmitting SYN on an actively created socket
12438 	 *
12439 	 *  - When sending a zero-length cookie (cookie request) on an
12440 	 *    actively created socket
12441 	 *
12442 	 *  - When the socket is in the CLOSED state (RST is being sent)
12443 	 */
12444 	if (IS_FASTOPEN(tp->t_flags) &&
12445 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12446 	     ((tp->t_state == TCPS_SYN_SENT) &&
12447 	      (tp->t_tfo_client_cookie_len == 0)) ||
12448 	     (flags & TH_RST))) {
12449 		sack_rxmit = 0;
12450 		len = 0;
12451 	}
12452 	/* Without fast-open there should never be data sent on a SYN */
12453 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
12454 		tp->snd_nxt = tp->iss;
12455 		len = 0;
12456 	}
12457 	orig_len = len;
12458 	if (len <= 0) {
12459 		/*
12460 		 * If FIN has been sent but not acked, but we haven't been
12461 		 * called to retransmit, len will be < 0.  Otherwise, window
12462 		 * shrank after we sent into it.  If window shrank to 0,
12463 		 * cancel pending retransmit, pull snd_nxt back to (closed)
12464 		 * window, and set the persist timer if it isn't already
12465 		 * going.  If the window didn't close completely, just wait
12466 		 * for an ACK.
12467 		 *
12468 		 * We also do a general check here to ensure that we will
12469 		 * set the persist timer when we have data to send, but a
12470 		 * 0-byte window. This makes sure the persist timer is set
12471 		 * even if the packet hits one of the "goto send" lines
12472 		 * below.
12473 		 */
12474 		len = 0;
12475 		if ((tp->snd_wnd == 0) &&
12476 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12477 		    (tp->snd_una == tp->snd_max) &&
12478 		    (sb_offset < (int)sbavail(sb))) {
12479 			tp->snd_nxt = tp->snd_una;
12480 			rack_enter_persist(tp, rack, cts);
12481 		}
12482 	} else if ((rsm == NULL) &&
12483 		   ((doing_tlp == 0) || (new_data_tlp == 1)) &&
12484 		   (len < rack->r_ctl.rc_pace_max_segs)) {
12485 		/*
12486 		 * We are not sending a maximum sized segment for
12487 		 * some reason. Should we not send anything (think
12488 		 * sws or persists)?
12489 		 */
12490 		if ((tp->snd_wnd < min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)), minseg)) &&
12491 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12492 		    (len < minseg) &&
12493 		    (len < (int)(sbavail(sb) - sb_offset))) {
12494 			/*
12495 			 * Here the rwnd is less than
12496 			 * the minimum pacing size, this is not a retransmit,
12497 			 * we are established and
12498 			 * the send is not the last in the socket buffer
12499 			 * we send nothing, and we may enter persists
12500 			 * if nothing is outstanding.
12501 			 */
12502 			len = 0;
12503 			if (tp->snd_max == tp->snd_una) {
12504 				/*
12505 				 * Nothing out we can
12506 				 * go into persists.
12507 				 */
12508 				rack_enter_persist(tp, rack, cts);
12509 				tp->snd_nxt = tp->snd_una;
12510 			}
12511 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
12512 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
12513 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12514 			   (len < minseg)) {
12515 			/*
12516 			 * Here we are not retransmitting, and
12517 			 * the cwnd is not so small that we could
12518 			 * not send at least a min size (rxt timer
12519 			 * not having gone off), We have 2 segments or
12520 			 * more already in flight, its not the tail end
12521 			 * of the socket buffer  and the cwnd is blocking
12522 			 * us from sending out a minimum pacing segment size.
12523 			 * Lets not send anything.
12524 			 */
12525 			len = 0;
12526 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12527 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
12528 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
12529 			   (len < (int)(sbavail(sb) - sb_offset)) &&
12530 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
12531 			/*
12532 			 * Here we have a send window but we have
12533 			 * filled it up and we can't send another pacing segment.
12534 			 * We also have in flight more than 2 segments
12535 			 * and we are not completing the sb i.e. we allow
12536 			 * the last bytes of the sb to go out even if
12537 			 * its not a full pacing segment.
12538 			 */
12539 			len = 0;
12540 		}
12541 	}
12542 	/* len will be >= 0 after this point. */
12543 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12544 	tcp_sndbuf_autoscale(tp, so, min(tp->snd_wnd, cwnd_to_use));
12545 	/*
12546 	 * Decide if we can use TCP Segmentation Offloading (if supported by
12547 	 * hardware).
12548 	 *
12549 	 * TSO may only be used if we are in a pure bulk sending state.  The
12550 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12551 	 * options prevent using TSO.  With TSO the TCP header is the same
12552 	 * (except for the sequence number) for all generated packets.  This
12553 	 * makes it impossible to transmit any options which vary per
12554 	 * generated segment or packet.
12555 	 *
12556 	 * IPv4 handling has a clear separation of ip options and ip header
12557 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
12558 	 * the right thing below to provide length of just ip options and thus
12559 	 * checking for ipoptlen is enough to decide if ip options are present.
12560 	 */
12561 
12562 #ifdef INET6
12563 	if (isipv6)
12564 		ipoptlen = ip6_optlen(tp->t_inpcb);
12565 	else
12566 #endif
12567 		if (tp->t_inpcb->inp_options)
12568 			ipoptlen = tp->t_inpcb->inp_options->m_len -
12569 				offsetof(struct ipoption, ipopt_list);
12570 		else
12571 			ipoptlen = 0;
12572 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12573 	/*
12574 	 * Pre-calculate here as we save another lookup into the darknesses
12575 	 * of IPsec that way and can actually decide if TSO is ok.
12576 	 */
12577 #ifdef INET6
12578 	if (isipv6 && IPSEC_ENABLED(ipv6))
12579 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
12580 #ifdef INET
12581 	else
12582 #endif
12583 #endif				/* INET6 */
12584 #ifdef INET
12585 		if (IPSEC_ENABLED(ipv4))
12586 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
12587 #endif				/* INET */
12588 #endif
12589 
12590 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12591 	ipoptlen += ipsec_optlen;
12592 #endif
12593 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
12594 	    (tp->t_port == 0) &&
12595 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
12596 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
12597 	    ipoptlen == 0)
12598 		tso = 1;
12599 	{
12600 		uint32_t outstanding;
12601 
12602 		outstanding = tp->snd_max - tp->snd_una;
12603 		if (tp->t_flags & TF_SENTFIN) {
12604 			/*
12605 			 * If we sent a fin, snd_max is 1 higher than
12606 			 * snd_una
12607 			 */
12608 			outstanding--;
12609 		}
12610 		if (sack_rxmit) {
12611 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
12612 				flags &= ~TH_FIN;
12613 		} else {
12614 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
12615 				   sbused(sb)))
12616 				flags &= ~TH_FIN;
12617 		}
12618 	}
12619 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12620 	    (long)TCP_MAXWIN << tp->rcv_scale);
12621 
12622 	/*
12623 	 * Sender silly window avoidance.   We transmit under the following
12624 	 * conditions when len is non-zero:
12625 	 *
12626 	 * - We have a full segment (or more with TSO) - This is the last
12627 	 * buffer in a write()/send() and we are either idle or running
12628 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
12629 	 * then 1/2 the maximum send window's worth of data (receiver may be
12630 	 * limited the window size) - we need to retransmit
12631 	 */
12632 	if (len) {
12633 		if (len >= segsiz) {
12634 			goto send;
12635 		}
12636 		/*
12637 		 * NOTE! on localhost connections an 'ack' from the remote
12638 		 * end may occur synchronously with the output and cause us
12639 		 * to flush a buffer queued with moretocome.  XXX
12640 		 *
12641 		 */
12642 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
12643 		    (idle || (tp->t_flags & TF_NODELAY)) &&
12644 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
12645 		    (tp->t_flags & TF_NOPUSH) == 0) {
12646 			pass = 2;
12647 			goto send;
12648 		}
12649 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
12650 			pass = 22;
12651 			goto send;
12652 		}
12653 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
12654 			pass = 4;
12655 			goto send;
12656 		}
12657 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
12658 			pass = 5;
12659 			goto send;
12660 		}
12661 		if (sack_rxmit) {
12662 			pass = 6;
12663 			goto send;
12664 		}
12665 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
12666 		    (ctf_outstanding(tp) < (segsiz * 2))) {
12667 			/*
12668 			 * We have less than two MSS outstanding (delayed ack)
12669 			 * and our rwnd will not let us send a full sized
12670 			 * MSS. Lets go ahead and let this small segment
12671 			 * out because we want to try to have at least two
12672 			 * packets inflight to not be caught by delayed ack.
12673 			 */
12674 			pass = 12;
12675 			goto send;
12676 		}
12677 	}
12678 	/*
12679 	 * Sending of standalone window updates.
12680 	 *
12681 	 * Window updates are important when we close our window due to a
12682 	 * full socket buffer and are opening it again after the application
12683 	 * reads data from it.  Once the window has opened again and the
12684 	 * remote end starts to send again the ACK clock takes over and
12685 	 * provides the most current window information.
12686 	 *
12687 	 * We must avoid the silly window syndrome whereas every read from
12688 	 * the receive buffer, no matter how small, causes a window update
12689 	 * to be sent.  We also should avoid sending a flurry of window
12690 	 * updates when the socket buffer had queued a lot of data and the
12691 	 * application is doing small reads.
12692 	 *
12693 	 * Prevent a flurry of pointless window updates by only sending an
12694 	 * update when we can increase the advertized window by more than
12695 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
12696 	 * full or is very small be more aggressive and send an update
12697 	 * whenever we can increase by two mss sized segments. In all other
12698 	 * situations the ACK's to new incoming data will carry further
12699 	 * window increases.
12700 	 *
12701 	 * Don't send an independent window update if a delayed ACK is
12702 	 * pending (it will get piggy-backed on it) or the remote side
12703 	 * already has done a half-close and won't send more data.  Skip
12704 	 * this if the connection is in T/TCP half-open state.
12705 	 */
12706 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
12707 	    !(tp->t_flags & TF_DELACK) &&
12708 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
12709 		/*
12710 		 * "adv" is the amount we could increase the window, taking
12711 		 * into account that we are limited by TCP_MAXWIN <<
12712 		 * tp->rcv_scale.
12713 		 */
12714 		int32_t adv;
12715 		int oldwin;
12716 
12717 		adv = recwin;
12718 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
12719 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
12720 			if (adv > oldwin)
12721 			    adv -= oldwin;
12722 			else {
12723 				/* We can't increase the window */
12724 				adv = 0;
12725 			}
12726 		} else
12727 			oldwin = 0;
12728 
12729 		/*
12730 		 * If the new window size ends up being the same as or less
12731 		 * than the old size when it is scaled, then don't force
12732 		 * a window update.
12733 		 */
12734 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
12735 			goto dontupdate;
12736 
12737 		if (adv >= (int32_t)(2 * segsiz) &&
12738 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
12739 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
12740 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
12741 			pass = 7;
12742 			goto send;
12743 		}
12744 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
12745 			pass = 23;
12746 			goto send;
12747 		}
12748 	}
12749 dontupdate:
12750 
12751 	/*
12752 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
12753 	 * is also a catch-all for the retransmit timer timeout case.
12754 	 */
12755 	if (tp->t_flags & TF_ACKNOW) {
12756 		pass = 8;
12757 		goto send;
12758 	}
12759 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
12760 		pass = 9;
12761 		goto send;
12762 	}
12763 	/*
12764 	 * If our state indicates that FIN should be sent and we have not
12765 	 * yet done so, then we need to send.
12766 	 */
12767 	if ((flags & TH_FIN) &&
12768 	    (tp->snd_nxt == tp->snd_una)) {
12769 		pass = 11;
12770 		goto send;
12771 	}
12772 	/*
12773 	 * No reason to send a segment, just return.
12774 	 */
12775 just_return:
12776 	SOCKBUF_UNLOCK(sb);
12777 just_return_nolock:
12778 	{
12779 		int app_limited = CTF_JR_SENT_DATA;
12780 
12781 		if (tot_len_this_send > 0) {
12782 			/* Make sure snd_nxt is up to max */
12783 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
12784 				tp->snd_nxt = tp->snd_max;
12785 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
12786 		} else {
12787 			int end_window = 0;
12788 			uint32_t seq = tp->gput_ack;
12789 
12790 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
12791 			if (rsm) {
12792 				/*
12793 				 * Mark the last sent that we just-returned (hinting
12794 				 * that delayed ack may play a role in any rtt measurement).
12795 				 */
12796 				rsm->r_just_ret = 1;
12797 			}
12798 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
12799 			rack->r_ctl.rc_agg_delayed = 0;
12800 			rack->r_early = 0;
12801 			rack->r_late = 0;
12802 			rack->r_ctl.rc_agg_early = 0;
12803 			if ((ctf_outstanding(tp) +
12804 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
12805 				 minseg)) >= tp->snd_wnd) {
12806 				/* We are limited by the rwnd */
12807 				app_limited = CTF_JR_RWND_LIMITED;
12808 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
12809 				/* We are limited by whats available -- app limited */
12810 				app_limited = CTF_JR_APP_LIMITED;
12811 			} else if ((idle == 0) &&
12812 				   ((tp->t_flags & TF_NODELAY) == 0) &&
12813 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
12814 				   (len < segsiz)) {
12815 				/*
12816 				 * No delay is not on and the
12817 				 * user is sending less than 1MSS. This
12818 				 * brings out SWS avoidance so we
12819 				 * don't send. Another app-limited case.
12820 				 */
12821 				app_limited = CTF_JR_APP_LIMITED;
12822 			} else if (tp->t_flags & TF_NOPUSH) {
12823 				/*
12824 				 * The user has requested no push of
12825 				 * the last segment and we are
12826 				 * at the last segment. Another app
12827 				 * limited case.
12828 				 */
12829 				app_limited = CTF_JR_APP_LIMITED;
12830 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
12831 				/* Its the cwnd */
12832 				app_limited = CTF_JR_CWND_LIMITED;
12833 			} else if (rack->rc_in_persist == 1) {
12834 				/* We are in persists */
12835 				app_limited = CTF_JR_PERSISTS;
12836 			} else if (IN_RECOVERY(tp->t_flags) &&
12837 				   (rack->rack_no_prr == 0) &&
12838 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
12839 				app_limited = CTF_JR_PRR;
12840 			} else {
12841 				/* Now why here are we not sending? */
12842 #ifdef NOW
12843 #ifdef INVARIANTS
12844 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
12845 #endif
12846 #endif
12847 				app_limited = CTF_JR_ASSESSING;
12848 			}
12849 			/*
12850 			 * App limited in some fashion, for our pacing GP
12851 			 * measurements we don't want any gap (even cwnd).
12852 			 * Close  down the measurement window.
12853 			 */
12854 			if (rack_cwnd_block_ends_measure &&
12855 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
12856 			     (app_limited == CTF_JR_PRR))) {
12857 				/*
12858 				 * The reason we are not sending is
12859 				 * the cwnd (or prr). We have been configured
12860 				 * to end the measurement window in
12861 				 * this case.
12862 				 */
12863 				end_window = 1;
12864 			} else if (app_limited == CTF_JR_PERSISTS) {
12865 				/*
12866 				 * We never end the measurement window
12867 				 * in persists, though in theory we
12868 				 * should be only entering after everything
12869 				 * is acknowledged (so we will probably
12870 				 * never come here).
12871 				 */
12872 				end_window = 0;
12873 			} else if (rack_rwnd_block_ends_measure &&
12874 				   (app_limited == CTF_JR_RWND_LIMITED)) {
12875 				/*
12876 				 * We are rwnd limited and have been
12877 				 * configured to end the measurement
12878 				 * window in this case.
12879 				 */
12880 				end_window = 1;
12881 			} else if (app_limited == CTF_JR_APP_LIMITED) {
12882 				/*
12883 				 * A true application limited period, we have
12884 				 * ran out of data.
12885 				 */
12886 				end_window = 1;
12887 			} else if (app_limited == CTF_JR_ASSESSING) {
12888 				/*
12889 				 * In the assessing case we hit the end of
12890 				 * the if/else and had no known reason
12891 				 * This will panic us under invariants..
12892 				 *
12893 				 * If we get this out in logs we need to
12894 				 * investagate which reason we missed.
12895 				 */
12896 				end_window = 1;
12897 			}
12898 			if (end_window) {
12899 				uint8_t log = 0;
12900 
12901 				if ((tp->t_flags & TF_GPUTINPROG) &&
12902 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
12903 					/* Mark the last packet has app limited */
12904 					tp->gput_ack = tp->snd_max;
12905 					log = 1;
12906 				}
12907 				rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
12908 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
12909 					if (rack->r_ctl.rc_app_limited_cnt == 0)
12910 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
12911 					else {
12912 						/*
12913 						 * Go out to the end app limited and mark
12914 						 * this new one as next and move the end_appl up
12915 						 * to this guy.
12916 						 */
12917 						if (rack->r_ctl.rc_end_appl)
12918 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
12919 						rack->r_ctl.rc_end_appl = rsm;
12920 					}
12921 					rsm->r_flags |= RACK_APP_LIMITED;
12922 					rack->r_ctl.rc_app_limited_cnt++;
12923 				}
12924 				if (log)
12925 					rack_log_pacing_delay_calc(rack,
12926 								   rack->r_ctl.rc_app_limited_cnt, seq,
12927 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL);
12928 			}
12929 		}
12930 		if (slot) {
12931 			/* set the rack tcb into the slot N */
12932 			counter_u64_add(rack_paced_segments, 1);
12933 		} else if (tot_len_this_send) {
12934 			counter_u64_add(rack_unpaced_segments, 1);
12935 		}
12936 		/* Check if we need to go into persists or not */
12937 		if ((rack->rc_in_persist == 0) &&
12938 		    (tp->snd_max == tp->snd_una) &&
12939 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
12940 		    sbavail(sb) &&
12941 		    (sbavail(sb) > tp->snd_wnd) &&
12942 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
12943 			/* Yes lets make sure to move to persist before timer-start */
12944 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12945 		}
12946 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
12947 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
12948 	}
12949 #ifdef NETFLIX_SHARED_CWND
12950 	if ((sbavail(sb) == 0) &&
12951 	    rack->r_ctl.rc_scw) {
12952 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
12953 		rack->rack_scwnd_is_idle = 1;
12954 	}
12955 #endif
12956 	return (0);
12957 
12958 send:
12959 	if ((flags & TH_FIN) &&
12960 	    sbavail(sb)) {
12961 		/*
12962 		 * We do not transmit a FIN
12963 		 * with data outstanding. We
12964 		 * need to make it so all data
12965 		 * is acked first.
12966 		 */
12967 		flags &= ~TH_FIN;
12968 	}
12969 	/* Enforce stack imposed max seg size if we have one */
12970 	if (rack->r_ctl.rc_pace_max_segs &&
12971 	    (len > rack->r_ctl.rc_pace_max_segs)) {
12972 		mark = 1;
12973 		len = rack->r_ctl.rc_pace_max_segs;
12974 	}
12975 	SOCKBUF_LOCK_ASSERT(sb);
12976 	if (len > 0) {
12977 		if (len >= segsiz)
12978 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
12979 		else
12980 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
12981 	}
12982 	/*
12983 	 * Before ESTABLISHED, force sending of initial options unless TCP
12984 	 * set not to do any options. NOTE: we assume that the IP/TCP header
12985 	 * plus TCP options always fit in a single mbuf, leaving room for a
12986 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
12987 	 * + optlen <= MCLBYTES
12988 	 */
12989 	optlen = 0;
12990 #ifdef INET6
12991 	if (isipv6)
12992 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12993 	else
12994 #endif
12995 		hdrlen = sizeof(struct tcpiphdr);
12996 
12997 	/*
12998 	 * Compute options for segment. We only have to care about SYN and
12999 	 * established connection segments.  Options for SYN-ACK segments
13000 	 * are handled in TCP syncache.
13001 	 */
13002 	to.to_flags = 0;
13003 	if ((tp->t_flags & TF_NOOPT) == 0) {
13004 		/* Maximum segment size. */
13005 		if (flags & TH_SYN) {
13006 			tp->snd_nxt = tp->iss;
13007 			to.to_mss = tcp_mssopt(&inp->inp_inc);
13008 #ifdef NETFLIX_TCPOUDP
13009 			if (tp->t_port)
13010 				to.to_mss -= V_tcp_udp_tunneling_overhead;
13011 #endif
13012 			to.to_flags |= TOF_MSS;
13013 
13014 			/*
13015 			 * On SYN or SYN|ACK transmits on TFO connections,
13016 			 * only include the TFO option if it is not a
13017 			 * retransmit, as the presence of the TFO option may
13018 			 * have caused the original SYN or SYN|ACK to have
13019 			 * been dropped by a middlebox.
13020 			 */
13021 			if (IS_FASTOPEN(tp->t_flags) &&
13022 			    (tp->t_rxtshift == 0)) {
13023 				if (tp->t_state == TCPS_SYN_RECEIVED) {
13024 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
13025 					to.to_tfo_cookie =
13026 						(u_int8_t *)&tp->t_tfo_cookie.server;
13027 					to.to_flags |= TOF_FASTOPEN;
13028 					wanted_cookie = 1;
13029 				} else if (tp->t_state == TCPS_SYN_SENT) {
13030 					to.to_tfo_len =
13031 						tp->t_tfo_client_cookie_len;
13032 					to.to_tfo_cookie =
13033 						tp->t_tfo_cookie.client;
13034 					to.to_flags |= TOF_FASTOPEN;
13035 					wanted_cookie = 1;
13036 					/*
13037 					 * If we wind up having more data to
13038 					 * send with the SYN than can fit in
13039 					 * one segment, don't send any more
13040 					 * until the SYN|ACK comes back from
13041 					 * the other end.
13042 					 */
13043 					sendalot = 0;
13044 				}
13045 			}
13046 		}
13047 		/* Window scaling. */
13048 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
13049 			to.to_wscale = tp->request_r_scale;
13050 			to.to_flags |= TOF_SCALE;
13051 		}
13052 		/* Timestamps. */
13053 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
13054 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
13055 			to.to_tsval = cts + tp->ts_offset;
13056 			to.to_tsecr = tp->ts_recent;
13057 			to.to_flags |= TOF_TS;
13058 		}
13059 		/* Set receive buffer autosizing timestamp. */
13060 		if (tp->rfbuf_ts == 0 &&
13061 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
13062 			tp->rfbuf_ts = tcp_ts_getticks();
13063 		/* Selective ACK's. */
13064 		if (flags & TH_SYN)
13065 			to.to_flags |= TOF_SACKPERM;
13066 		else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13067 			 tp->rcv_numsacks > 0) {
13068 			to.to_flags |= TOF_SACK;
13069 			to.to_nsacks = tp->rcv_numsacks;
13070 			to.to_sacks = (u_char *)tp->sackblks;
13071 		}
13072 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13073 		/* TCP-MD5 (RFC2385). */
13074 		if (tp->t_flags & TF_SIGNATURE)
13075 			to.to_flags |= TOF_SIGNATURE;
13076 #endif				/* TCP_SIGNATURE */
13077 
13078 		/* Processing the options. */
13079 		hdrlen += optlen = tcp_addoptions(&to, opt);
13080 		/*
13081 		 * If we wanted a TFO option to be added, but it was unable
13082 		 * to fit, ensure no data is sent.
13083 		 */
13084 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
13085 		    !(to.to_flags & TOF_FASTOPEN))
13086 			len = 0;
13087 	}
13088 #ifdef NETFLIX_TCPOUDP
13089 	if (tp->t_port) {
13090 		if (V_tcp_udp_tunneling_port == 0) {
13091 			/* The port was removed?? */
13092 			SOCKBUF_UNLOCK(&so->so_snd);
13093 			return (EHOSTUNREACH);
13094 		}
13095 		hdrlen += sizeof(struct udphdr);
13096 	}
13097 #endif
13098 #ifdef INET6
13099 	if (isipv6)
13100 		ipoptlen = ip6_optlen(tp->t_inpcb);
13101 	else
13102 #endif
13103 		if (tp->t_inpcb->inp_options)
13104 			ipoptlen = tp->t_inpcb->inp_options->m_len -
13105 				offsetof(struct ipoption, ipopt_list);
13106 		else
13107 			ipoptlen = 0;
13108 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
13109 	ipoptlen += ipsec_optlen;
13110 #endif
13111 
13112 	/*
13113 	 * Adjust data length if insertion of options will bump the packet
13114 	 * length beyond the t_maxseg length. Clear the FIN bit because we
13115 	 * cut off the tail of the segment.
13116 	 */
13117 	if (len + optlen + ipoptlen > tp->t_maxseg) {
13118 		if (tso) {
13119 			uint32_t if_hw_tsomax;
13120 			uint32_t moff;
13121 			int32_t max_len;
13122 
13123 			/* extract TSO information */
13124 			if_hw_tsomax = tp->t_tsomax;
13125 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
13126 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
13127 			KASSERT(ipoptlen == 0,
13128 				("%s: TSO can't do IP options", __func__));
13129 
13130 			/*
13131 			 * Check if we should limit by maximum payload
13132 			 * length:
13133 			 */
13134 			if (if_hw_tsomax != 0) {
13135 				/* compute maximum TSO length */
13136 				max_len = (if_hw_tsomax - hdrlen -
13137 					   max_linkhdr);
13138 				if (max_len <= 0) {
13139 					len = 0;
13140 				} else if (len > max_len) {
13141 					sendalot = 1;
13142 					len = max_len;
13143 					mark = 2;
13144 				}
13145 			}
13146 			/*
13147 			 * Prevent the last segment from being fractional
13148 			 * unless the send sockbuf can be emptied:
13149 			 */
13150 			max_len = (tp->t_maxseg - optlen);
13151 			if ((sb_offset + len) < sbavail(sb)) {
13152 				moff = len % (u_int)max_len;
13153 				if (moff != 0) {
13154 					mark = 3;
13155 					len -= moff;
13156 				}
13157 			}
13158                         /*
13159 			 * In case there are too many small fragments don't
13160 			 * use TSO:
13161 			 */
13162 			if (len <= segsiz) {
13163 				mark = 4;
13164 				tso = 0;
13165 			}
13166 			/*
13167 			 * Send the FIN in a separate segment after the bulk
13168 			 * sending is done. We don't trust the TSO
13169 			 * implementations to clear the FIN flag on all but
13170 			 * the last segment.
13171 			 */
13172 			if (tp->t_flags & TF_NEEDFIN) {
13173 				sendalot = 4;
13174 			}
13175 		} else {
13176 			mark = 5;
13177 			if (optlen + ipoptlen >= tp->t_maxseg) {
13178 				/*
13179 				 * Since we don't have enough space to put
13180 				 * the IP header chain and the TCP header in
13181 				 * one packet as required by RFC 7112, don't
13182 				 * send it. Also ensure that at least one
13183 				 * byte of the payload can be put into the
13184 				 * TCP segment.
13185 				 */
13186 				SOCKBUF_UNLOCK(&so->so_snd);
13187 				error = EMSGSIZE;
13188 				sack_rxmit = 0;
13189 				goto out;
13190 			}
13191 			len = tp->t_maxseg - optlen - ipoptlen;
13192 			sendalot = 5;
13193 		}
13194 	} else {
13195 		tso = 0;
13196 		mark = 6;
13197 	}
13198 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
13199 		("%s: len > IP_MAXPACKET", __func__));
13200 #ifdef DIAGNOSTIC
13201 #ifdef INET6
13202 	if (max_linkhdr + hdrlen > MCLBYTES)
13203 #else
13204 		if (max_linkhdr + hdrlen > MHLEN)
13205 #endif
13206 			panic("tcphdr too big");
13207 #endif
13208 
13209 	/*
13210 	 * This KASSERT is here to catch edge cases at a well defined place.
13211 	 * Before, those had triggered (random) panic conditions further
13212 	 * down.
13213 	 */
13214 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
13215 	if ((len == 0) &&
13216 	    (flags & TH_FIN) &&
13217 	    (sbused(sb))) {
13218 		/*
13219 		 * We have outstanding data, don't send a fin by itself!.
13220 		 */
13221 		goto just_return;
13222 	}
13223 	/*
13224 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
13225 	 * and initialize the header from the template for sends on this
13226 	 * connection.
13227 	 */
13228 	if (len) {
13229 		uint32_t max_val;
13230 		uint32_t moff;
13231 
13232 		if (rack->r_ctl.rc_pace_max_segs)
13233 			max_val = rack->r_ctl.rc_pace_max_segs;
13234 		else if (rack->rc_user_set_max_segs)
13235 			max_val = rack->rc_user_set_max_segs * segsiz;
13236 		else
13237 			max_val = len;
13238 		/*
13239 		 * We allow a limit on sending with hptsi.
13240 		 */
13241 		if (len > max_val) {
13242 			mark = 7;
13243 			len = max_val;
13244 		}
13245 #ifdef INET6
13246 		if (MHLEN < hdrlen + max_linkhdr)
13247 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
13248 		else
13249 #endif
13250 			m = m_gethdr(M_NOWAIT, MT_DATA);
13251 
13252 		if (m == NULL) {
13253 			SOCKBUF_UNLOCK(sb);
13254 			error = ENOBUFS;
13255 			sack_rxmit = 0;
13256 			goto out;
13257 		}
13258 		m->m_data += max_linkhdr;
13259 		m->m_len = hdrlen;
13260 
13261 		/*
13262 		 * Start the m_copy functions from the closest mbuf to the
13263 		 * sb_offset in the socket buffer chain.
13264 		 */
13265 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
13266 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
13267 			m_copydata(mb, moff, (int)len,
13268 				   mtod(m, caddr_t)+hdrlen);
13269 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
13270 				sbsndptr_adv(sb, mb, len);
13271 			m->m_len += len;
13272 		} else {
13273 			struct sockbuf *msb;
13274 
13275 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
13276 				msb = NULL;
13277 			else
13278 				msb = sb;
13279 			m->m_next = tcp_m_copym(
13280 				mb, moff, &len,
13281 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
13282 				((rsm == NULL) ? hw_tls : 0)
13283 #ifdef NETFLIX_COPY_ARGS
13284 				, &filled_all
13285 #endif
13286 				);
13287 			if (len <= (tp->t_maxseg - optlen)) {
13288 				/*
13289 				 * Must have ran out of mbufs for the copy
13290 				 * shorten it to no longer need tso. Lets
13291 				 * not put on sendalot since we are low on
13292 				 * mbufs.
13293 				 */
13294 				tso = 0;
13295 			}
13296 			if (m->m_next == NULL) {
13297 				SOCKBUF_UNLOCK(sb);
13298 				(void)m_free(m);
13299 				error = ENOBUFS;
13300 				sack_rxmit = 0;
13301 				goto out;
13302 			}
13303 		}
13304 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
13305 			if (rsm && (rsm->r_flags & RACK_TLP)) {
13306 				/*
13307 				 * TLP should not count in retran count, but
13308 				 * in its own bin
13309 				 */
13310 				counter_u64_add(rack_tlp_retran, 1);
13311 				counter_u64_add(rack_tlp_retran_bytes, len);
13312 			} else {
13313 				tp->t_sndrexmitpack++;
13314 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
13315 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
13316 			}
13317 #ifdef STATS
13318 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
13319 						 len);
13320 #endif
13321 		} else {
13322 			KMOD_TCPSTAT_INC(tcps_sndpack);
13323 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
13324 #ifdef STATS
13325 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
13326 						 len);
13327 #endif
13328 		}
13329 		/*
13330 		 * If we're sending everything we've got, set PUSH. (This
13331 		 * will keep happy those implementations which only give
13332 		 * data to the user when a buffer fills or a PUSH comes in.)
13333 		 */
13334 		if (sb_offset + len == sbused(sb) &&
13335 		    sbused(sb) &&
13336 		    !(flags & TH_SYN))
13337 			flags |= TH_PUSH;
13338 
13339 		SOCKBUF_UNLOCK(sb);
13340 	} else {
13341 		SOCKBUF_UNLOCK(sb);
13342 		if (tp->t_flags & TF_ACKNOW)
13343 			KMOD_TCPSTAT_INC(tcps_sndacks);
13344 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
13345 			KMOD_TCPSTAT_INC(tcps_sndctrl);
13346 		else
13347 			KMOD_TCPSTAT_INC(tcps_sndwinup);
13348 
13349 		m = m_gethdr(M_NOWAIT, MT_DATA);
13350 		if (m == NULL) {
13351 			error = ENOBUFS;
13352 			sack_rxmit = 0;
13353 			goto out;
13354 		}
13355 #ifdef INET6
13356 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13357 		    MHLEN >= hdrlen) {
13358 			M_ALIGN(m, hdrlen);
13359 		} else
13360 #endif
13361 			m->m_data += max_linkhdr;
13362 		m->m_len = hdrlen;
13363 	}
13364 	SOCKBUF_UNLOCK_ASSERT(sb);
13365 	m->m_pkthdr.rcvif = (struct ifnet *)0;
13366 #ifdef MAC
13367 	mac_inpcb_create_mbuf(inp, m);
13368 #endif
13369 #ifdef INET6
13370 	if (isipv6) {
13371 		ip6 = mtod(m, struct ip6_hdr *);
13372 #ifdef NETFLIX_TCPOUDP
13373 		if (tp->t_port) {
13374 			udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr));
13375 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13376 			udp->uh_dport = tp->t_port;
13377 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
13378 			udp->uh_ulen = htons(ulen);
13379 			th = (struct tcphdr *)(udp + 1);
13380 		} else
13381 #endif
13382 			th = (struct tcphdr *)(ip6 + 1);
13383 		tcpip_fillheaders(inp,
13384 #ifdef NETFLIX_TCPOUDP
13385 				  tp->t_port,
13386 #endif
13387 				  ip6, th);
13388 	} else
13389 #endif				/* INET6 */
13390 	{
13391 		ip = mtod(m, struct ip *);
13392 #ifdef TCPDEBUG
13393 		ipov = (struct ipovly *)ip;
13394 #endif
13395 #ifdef NETFLIX_TCPOUDP
13396 		if (tp->t_port) {
13397 			udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip));
13398 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13399 			udp->uh_dport = tp->t_port;
13400 			ulen = hdrlen + len - sizeof(struct ip);
13401 			udp->uh_ulen = htons(ulen);
13402 			th = (struct tcphdr *)(udp + 1);
13403 		} else
13404 #endif
13405 			th = (struct tcphdr *)(ip + 1);
13406 		tcpip_fillheaders(inp,
13407 #ifdef NETFLIX_TCPOUDP
13408 				  tp->t_port,
13409 #endif
13410 				  ip, th);
13411 	}
13412 	/*
13413 	 * Fill in fields, remembering maximum advertised window for use in
13414 	 * delaying messages about window sizes. If resending a FIN, be sure
13415 	 * not to use a new sequence number.
13416 	 */
13417 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
13418 	    tp->snd_nxt == tp->snd_max)
13419 		tp->snd_nxt--;
13420 	/*
13421 	 * If we are starting a connection, send ECN setup SYN packet. If we
13422 	 * are on a retransmit, we may resend those bits a number of times
13423 	 * as per RFC 3168.
13424 	 */
13425 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
13426 		if (tp->t_rxtshift >= 1) {
13427 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
13428 				flags |= TH_ECE | TH_CWR;
13429 		} else
13430 			flags |= TH_ECE | TH_CWR;
13431 	}
13432 	/* Handle parallel SYN for ECN */
13433 	if ((tp->t_state == TCPS_SYN_RECEIVED) &&
13434 	    (tp->t_flags2 & TF2_ECN_SND_ECE)) {
13435 		flags |= TH_ECE;
13436 		tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13437 	}
13438 	if (tp->t_state == TCPS_ESTABLISHED &&
13439 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
13440 		/*
13441 		 * If the peer has ECN, mark data packets with ECN capable
13442 		 * transmission (ECT). Ignore pure ack packets,
13443 		 * retransmissions.
13444 		 */
13445 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
13446 		    (sack_rxmit == 0)) {
13447 #ifdef INET6
13448 			if (isipv6)
13449 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
13450 			else
13451 #endif
13452 				ip->ip_tos |= IPTOS_ECN_ECT0;
13453 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13454 			/*
13455 			 * Reply with proper ECN notifications.
13456 			 * Only set CWR on new data segments.
13457 			 */
13458 			if (tp->t_flags2 & TF2_ECN_SND_CWR) {
13459 				flags |= TH_CWR;
13460 				tp->t_flags2 &= ~TF2_ECN_SND_CWR;
13461 			}
13462 		}
13463 		if (tp->t_flags2 & TF2_ECN_SND_ECE)
13464 			flags |= TH_ECE;
13465 	}
13466 	/*
13467 	 * If we are doing retransmissions, then snd_nxt will not reflect
13468 	 * the first unsent octet.  For ACK only packets, we do not want the
13469 	 * sequence number of the retransmitted packet, we want the sequence
13470 	 * number of the next unsent octet.  So, if there is no data (and no
13471 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
13472 	 * ti_seq.  But if we are in persist state, snd_max might reflect
13473 	 * one byte beyond the right edge of the window, so use snd_nxt in
13474 	 * that case, since we know we aren't doing a retransmission.
13475 	 * (retransmit and persist are mutually exclusive...)
13476 	 */
13477 	if (sack_rxmit == 0) {
13478 		if (len || (flags & (TH_SYN | TH_FIN)) ||
13479 		    rack->rc_in_persist) {
13480 			th->th_seq = htonl(tp->snd_nxt);
13481 			rack_seq = tp->snd_nxt;
13482 		} else if (flags & TH_RST) {
13483 			/*
13484 			 * For a Reset send the last cum ack in sequence
13485 			 * (this like any other choice may still generate a
13486 			 * challenge ack, if a ack-update packet is in
13487 			 * flight).
13488 			 */
13489 			th->th_seq = htonl(tp->snd_una);
13490 			rack_seq = tp->snd_una;
13491 		} else {
13492 			th->th_seq = htonl(tp->snd_max);
13493 			rack_seq = tp->snd_max;
13494 		}
13495 	} else {
13496 		th->th_seq = htonl(rsm->r_start);
13497 		rack_seq = rsm->r_start;
13498 	}
13499 	th->th_ack = htonl(tp->rcv_nxt);
13500 	if (optlen) {
13501 		bcopy(opt, th + 1, optlen);
13502 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13503 	}
13504 	th->th_flags = flags;
13505 	/*
13506 	 * Calculate receive window.  Don't shrink window, but avoid silly
13507 	 * window syndrome.
13508 	 * If a RST segment is sent, advertise a window of zero.
13509 	 */
13510 	if (flags & TH_RST) {
13511 		recwin = 0;
13512 	} else {
13513 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
13514 		    recwin < (long)segsiz)
13515 			recwin = 0;
13516 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13517 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
13518 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
13519 	}
13520 
13521 	/*
13522 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13523 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
13524 	 * handled in syncache.
13525 	 */
13526 	if (flags & TH_SYN)
13527 		th->th_win = htons((u_short)
13528 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13529 	else {
13530 		/* Avoid shrinking window with window scaling. */
13531 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
13532 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13533 	}
13534 	/*
13535 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13536 	 * window.  This may cause the remote transmitter to stall.  This
13537 	 * flag tells soreceive() to disable delayed acknowledgements when
13538 	 * draining the buffer.  This can occur if the receiver is
13539 	 * attempting to read more data than can be buffered prior to
13540 	 * transmitting on the connection.
13541 	 */
13542 	if (th->th_win == 0) {
13543 		tp->t_sndzerowin++;
13544 		tp->t_flags |= TF_RXWIN0SENT;
13545 	} else
13546 		tp->t_flags &= ~TF_RXWIN0SENT;
13547 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated  */
13548 
13549 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13550 	if (to.to_flags & TOF_SIGNATURE) {
13551 		/*
13552 		 * Calculate MD5 signature and put it into the place
13553 		 * determined before.
13554 		 * NOTE: since TCP options buffer doesn't point into
13555 		 * mbuf's data, calculate offset and use it.
13556 		 */
13557 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13558 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13559 			/*
13560 			 * Do not send segment if the calculation of MD5
13561 			 * digest has failed.
13562 			 */
13563 			goto out;
13564 		}
13565 	}
13566 #endif
13567 
13568 	/*
13569 	 * Put TCP length in extended header, and then checksum extended
13570 	 * header and data.
13571 	 */
13572 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
13573 #ifdef INET6
13574 	if (isipv6) {
13575 		/*
13576 		 * ip6_plen is not need to be filled now, and will be filled
13577 		 * in ip6_output.
13578 		 */
13579 		if (tp->t_port) {
13580 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13581 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13582 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13583 			th->th_sum = htons(0);
13584 			UDPSTAT_INC(udps_opackets);
13585 		} else {
13586 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13587 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13588 			th->th_sum = in6_cksum_pseudo(ip6,
13589 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
13590 						      0);
13591 		}
13592 	}
13593 #endif
13594 #if defined(INET6) && defined(INET)
13595 	else
13596 #endif
13597 #ifdef INET
13598 	{
13599 		if (tp->t_port) {
13600 			m->m_pkthdr.csum_flags = CSUM_UDP;
13601 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13602 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13603 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13604 			th->th_sum = htons(0);
13605 			UDPSTAT_INC(udps_opackets);
13606 		} else {
13607 			m->m_pkthdr.csum_flags = CSUM_TCP;
13608 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13609 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
13610 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13611 									IPPROTO_TCP + len + optlen));
13612 		}
13613 		/* IP version must be set here for ipv4/ipv6 checking later */
13614 		KASSERT(ip->ip_v == IPVERSION,
13615 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
13616 	}
13617 #endif
13618 	/*
13619 	 * Enable TSO and specify the size of the segments. The TCP pseudo
13620 	 * header checksum is always provided. XXX: Fixme: This is currently
13621 	 * not the case for IPv6.
13622 	 */
13623 	if (tso) {
13624 		KASSERT(len > tp->t_maxseg - optlen,
13625 			("%s: len <= tso_segsz", __func__));
13626 		m->m_pkthdr.csum_flags |= CSUM_TSO;
13627 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
13628 	}
13629 	KASSERT(len + hdrlen == m_length(m, NULL),
13630 		("%s: mbuf chain different than expected: %d + %u != %u",
13631 		 __func__, len, hdrlen, m_length(m, NULL)));
13632 
13633 #ifdef TCP_HHOOK
13634 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
13635 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
13636 #endif
13637 #ifdef TCPDEBUG
13638 	/*
13639 	 * Trace.
13640 	 */
13641 	if (so->so_options & SO_DEBUG) {
13642 		u_short save = 0;
13643 
13644 #ifdef INET6
13645 		if (!isipv6)
13646 #endif
13647 		{
13648 			save = ipov->ih_len;
13649 			ipov->ih_len = htons(m->m_pkthdr.len	/* - hdrlen +
13650 								 * (th->th_off << 2) */ );
13651 		}
13652 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
13653 #ifdef INET6
13654 		if (!isipv6)
13655 #endif
13656 			ipov->ih_len = save;
13657 	}
13658 #endif				/* TCPDEBUG */
13659 
13660 	/* We're getting ready to send; log now. */
13661 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13662 		union tcp_log_stackspecific log;
13663 		struct timeval tv;
13664 
13665 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13666 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13667 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13668 		if (rack->rack_no_prr)
13669 			log.u_bbr.flex1 = 0;
13670 		else
13671 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13672 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
13673 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
13674 		log.u_bbr.flex4 = orig_len;
13675 		if (filled_all)
13676 			log.u_bbr.flex5 = 0x80000000;
13677 		else
13678 			log.u_bbr.flex5 = 0;
13679 		/* Save off the early/late values */
13680 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
13681 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
13682 		log.u_bbr.bw_inuse = rack_get_bw(rack);
13683 		if (rsm || sack_rxmit) {
13684 			if (doing_tlp)
13685 				log.u_bbr.flex8 = 2;
13686 			else
13687 				log.u_bbr.flex8 = 1;
13688 		} else {
13689 			log.u_bbr.flex8 = 0;
13690 		}
13691 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
13692 		log.u_bbr.flex7 = mark;
13693 		log.u_bbr.pkts_out = tp->t_maxseg;
13694 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13695 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13696 		log.u_bbr.lt_epoch = cwnd_to_use;
13697 		log.u_bbr.delivered = sendalot;
13698 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13699 				     len, &log, false, NULL, NULL, 0, &tv);
13700 	} else
13701 		lgb = NULL;
13702 
13703 	/*
13704 	 * Fill in IP length and desired time to live and send to IP level.
13705 	 * There should be a better way to handle ttl and tos; we could keep
13706 	 * them in the template, but need a way to checksum without them.
13707 	 */
13708 	/*
13709 	 * m->m_pkthdr.len should have been set before cksum calcuration,
13710 	 * because in6_cksum() need it.
13711 	 */
13712 #ifdef INET6
13713 	if (isipv6) {
13714 		/*
13715 		 * we separately set hoplimit for every segment, since the
13716 		 * user might want to change the value via setsockopt. Also,
13717 		 * desired default hop limit might be changed via Neighbor
13718 		 * Discovery.
13719 		 */
13720 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13721 
13722 		/*
13723 		 * Set the packet size here for the benefit of DTrace
13724 		 * probes. ip6_output() will set it properly; it's supposed
13725 		 * to include the option header lengths as well.
13726 		 */
13727 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13728 
13729 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
13730 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13731 		else
13732 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13733 
13734 		if (tp->t_state == TCPS_SYN_SENT)
13735 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13736 
13737 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13738 		/* TODO: IPv6 IP6TOS_ECT bit on */
13739 		error = ip6_output(m, inp->in6p_outputopts,
13740 				   &inp->inp_route6,
13741 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13742 				   NULL, NULL, inp);
13743 
13744 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
13745 			mtu = inp->inp_route6.ro_nh->nh_mtu;
13746 	}
13747 #endif				/* INET6 */
13748 #if defined(INET) && defined(INET6)
13749 	else
13750 #endif
13751 #ifdef INET
13752 	{
13753 		ip->ip_len = htons(m->m_pkthdr.len);
13754 #ifdef INET6
13755 		if (inp->inp_vflag & INP_IPV6PROTO)
13756 			ip->ip_ttl = in6_selecthlim(inp, NULL);
13757 #endif				/* INET6 */
13758 		/*
13759 		 * If we do path MTU discovery, then we set DF on every
13760 		 * packet. This might not be the best thing to do according
13761 		 * to RFC3390 Section 2. However the tcp hostcache migitates
13762 		 * the problem so it affects only the first tcp connection
13763 		 * with a host.
13764 		 *
13765 		 * NB: Don't set DF on small MTU/MSS to have a safe
13766 		 * fallback.
13767 		 */
13768 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13769 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13770 			if (tp->t_port == 0 || len < V_tcp_minmss) {
13771 				ip->ip_off |= htons(IP_DF);
13772 			}
13773 		} else {
13774 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13775 		}
13776 
13777 		if (tp->t_state == TCPS_SYN_SENT)
13778 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
13779 
13780 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
13781 
13782 		error = ip_output(m, inp->inp_options, &inp->inp_route,
13783 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
13784 				  inp);
13785 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
13786 			mtu = inp->inp_route.ro_nh->nh_mtu;
13787 	}
13788 #endif				/* INET */
13789 
13790 out:
13791 	if (lgb) {
13792 		lgb->tlb_errno = error;
13793 		lgb = NULL;
13794 	}
13795 	/*
13796 	 * In transmit state, time the transmission and arrange for the
13797 	 * retransmit.  In persist state, just set snd_max.
13798 	 */
13799 	if (error == 0) {
13800 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
13801 		if (rsm && (doing_tlp == 0)) {
13802 			/* Set we retransmitted */
13803 			rack->rc_gp_saw_rec = 1;
13804 		} else {
13805 			if (cwnd_to_use > tp->snd_ssthresh) {
13806 				/* Set we sent in CA */
13807 				rack->rc_gp_saw_ca = 1;
13808 			} else {
13809 				/* Set we sent in SS */
13810 				rack->rc_gp_saw_ss = 1;
13811 			}
13812 		}
13813 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13814 		    (tp->t_flags & TF_SACK_PERMIT) &&
13815 		    tp->rcv_numsacks > 0)
13816 			tcp_clean_dsack_blocks(tp);
13817 		tot_len_this_send += len;
13818 		if (len == 0)
13819 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
13820 		else if (len == 1) {
13821 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
13822 		} else if (len > 1) {
13823 			int idx;
13824 
13825 			idx = (len / segsiz) + 3;
13826 			if (idx >= TCP_MSS_ACCT_ATIMER)
13827 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
13828 			else
13829 				counter_u64_add(rack_out_size[idx], 1);
13830 		}
13831 	}
13832 	if (rack->rack_no_prr == 0) {
13833 		if (sub_from_prr && (error == 0)) {
13834 			if (rack->r_ctl.rc_prr_sndcnt >= len)
13835 				rack->r_ctl.rc_prr_sndcnt -= len;
13836 			else
13837 				rack->r_ctl.rc_prr_sndcnt = 0;
13838 		}
13839  	}
13840 	sub_from_prr = 0;
13841 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts,
13842 			pass, rsm, us_cts);
13843 	if ((error == 0) &&
13844 	    (len > 0) &&
13845 	    (tp->snd_una == tp->snd_max))
13846 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
13847 	/* Now are we in persists? */
13848 	if (rack->rc_in_persist == 0) {
13849 		tcp_seq startseq = tp->snd_nxt;
13850 
13851 		/* Track our lost count */
13852 		if (rsm && (doing_tlp == 0))
13853 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
13854 		/*
13855 		 * Advance snd_nxt over sequence space of this segment.
13856 		 */
13857 		if (error)
13858 			/* We don't log or do anything with errors */
13859 			goto nomore;
13860 		if (doing_tlp == 0) {
13861 			if (rsm == NULL) {
13862 				/*
13863 				 * Not a retransmission of some
13864 				 * sort, new data is going out so
13865 				 * clear our TLP count and flag.
13866 				 */
13867 				rack->rc_tlp_in_progress = 0;
13868 				rack->r_ctl.rc_tlp_cnt_out = 0;
13869 			}
13870 		} else {
13871 			/*
13872 			 * We have just sent a TLP, mark that it is true
13873 			 * and make sure our in progress is set so we
13874 			 * continue to check the count.
13875 			 */
13876 			rack->rc_tlp_in_progress = 1;
13877 			rack->r_ctl.rc_tlp_cnt_out++;
13878 		}
13879 		if (flags & (TH_SYN | TH_FIN)) {
13880 			if (flags & TH_SYN)
13881 				tp->snd_nxt++;
13882 			if (flags & TH_FIN) {
13883 				tp->snd_nxt++;
13884 				tp->t_flags |= TF_SENTFIN;
13885 			}
13886 		}
13887 		/* In the ENOBUFS case we do *not* update snd_max */
13888 		if (sack_rxmit)
13889 			goto nomore;
13890 
13891 		tp->snd_nxt += len;
13892 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
13893 			if (tp->snd_una == tp->snd_max) {
13894 				/*
13895 				 * Update the time we just added data since
13896 				 * none was outstanding.
13897 				 */
13898 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
13899 				tp->t_acktime = ticks;
13900 			}
13901 			tp->snd_max = tp->snd_nxt;
13902 			/*
13903 			 * Time this transmission if not a retransmission and
13904 			 * not currently timing anything.
13905 			 * This is only relevant in case of switching back to
13906 			 * the base stack.
13907 			 */
13908 			if (tp->t_rtttime == 0) {
13909 				tp->t_rtttime = ticks;
13910 				tp->t_rtseq = startseq;
13911 				KMOD_TCPSTAT_INC(tcps_segstimed);
13912 			}
13913 			if (len &&
13914 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
13915 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
13916 		}
13917 	} else {
13918 		/*
13919 		 * Persist case, update snd_max but since we are in persist
13920 		 * mode (no window) we do not update snd_nxt.
13921 		 */
13922 		int32_t xlen = len;
13923 
13924 		if (error)
13925 			goto nomore;
13926 
13927 		if (flags & TH_SYN)
13928 			++xlen;
13929 		if (flags & TH_FIN) {
13930 			++xlen;
13931 			tp->t_flags |= TF_SENTFIN;
13932 		}
13933 		/* In the ENOBUFS case we do *not* update snd_max */
13934 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) {
13935 			if (tp->snd_una == tp->snd_max) {
13936 				/*
13937 				 * Update the time we just added data since
13938 				 * none was outstanding.
13939 				 */
13940 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
13941 				tp->t_acktime = ticks;
13942 			}
13943 			tp->snd_max = tp->snd_nxt + len;
13944 		}
13945 	}
13946 nomore:
13947 	if (error) {
13948 		rack->r_ctl.rc_agg_delayed = 0;
13949 		rack->r_early = 0;
13950 		rack->r_late = 0;
13951 		rack->r_ctl.rc_agg_early = 0;
13952 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
13953 		/*
13954 		 * Failures do not advance the seq counter above. For the
13955 		 * case of ENOBUFS we will fall out and retry in 1ms with
13956 		 * the hpts. Everything else will just have to retransmit
13957 		 * with the timer.
13958 		 *
13959 		 * In any case, we do not want to loop around for another
13960 		 * send without a good reason.
13961 		 */
13962 		sendalot = 0;
13963 		switch (error) {
13964 		case EPERM:
13965 			tp->t_softerror = error;
13966 			return (error);
13967 		case ENOBUFS:
13968 			if (slot == 0) {
13969 				/*
13970 				 * Pace us right away to retry in a some
13971 				 * time
13972 				 */
13973 				slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
13974 				if (rack->rc_enobuf < 126)
13975 					rack->rc_enobuf++;
13976 				if (slot > ((rack->rc_rack_rtt / 2) * HPTS_USEC_IN_MSEC)) {
13977 					slot = (rack->rc_rack_rtt / 2) * HPTS_USEC_IN_MSEC;
13978 				}
13979 				if (slot < (10 * HPTS_USEC_IN_MSEC))
13980 					slot = 10 * HPTS_USEC_IN_MSEC;
13981 			}
13982 			counter_u64_add(rack_saw_enobuf, 1);
13983 			error = 0;
13984 			goto enobufs;
13985 		case EMSGSIZE:
13986 			/*
13987 			 * For some reason the interface we used initially
13988 			 * to send segments changed to another or lowered
13989 			 * its MTU. If TSO was active we either got an
13990 			 * interface without TSO capabilits or TSO was
13991 			 * turned off. If we obtained mtu from ip_output()
13992 			 * then update it and try again.
13993 			 */
13994 			if (tso)
13995 				tp->t_flags &= ~TF_TSO;
13996 			if (mtu != 0) {
13997 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
13998 				goto again;
13999 			}
14000 			slot = 10 * HPTS_USEC_IN_MSEC;
14001 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
14002 			return (error);
14003 		case ENETUNREACH:
14004 			counter_u64_add(rack_saw_enetunreach, 1);
14005 		case EHOSTDOWN:
14006 		case EHOSTUNREACH:
14007 		case ENETDOWN:
14008 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
14009 				tp->t_softerror = error;
14010 			}
14011 			/* FALLTHROUGH */
14012 		default:
14013 			slot = 10 * HPTS_USEC_IN_MSEC;
14014 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
14015 			return (error);
14016 		}
14017 	} else {
14018 		rack->rc_enobuf = 0;
14019 	}
14020 	KMOD_TCPSTAT_INC(tcps_sndtotal);
14021 
14022 	/*
14023 	 * Data sent (as far as we can tell). If this advertises a larger
14024 	 * window than any other segment, then remember the size of the
14025 	 * advertised window. Any pending ACK has now been sent.
14026 	 */
14027 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
14028 		tp->rcv_adv = tp->rcv_nxt + recwin;
14029 	tp->last_ack_sent = tp->rcv_nxt;
14030 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14031 enobufs:
14032 	/* Assure when we leave that snd_nxt will point to top */
14033 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
14034 		tp->snd_nxt = tp->snd_max;
14035 	if (sendalot) {
14036 		/* Do we need to turn off sendalot? */
14037 		if (rack->r_ctl.rc_pace_max_segs &&
14038 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
14039 			/* We hit our max. */
14040 			sendalot = 0;
14041 		} else if ((rack->rc_user_set_max_segs) &&
14042 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
14043 			/* We hit the user defined max */
14044 			sendalot = 0;
14045 		}
14046 	}
14047 	if ((error == 0) && (flags & TH_FIN))
14048 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
14049 	if (flags & TH_RST) {
14050 		/*
14051 		 * We don't send again after sending a RST.
14052 		 */
14053 		slot = 0;
14054 		sendalot = 0;
14055 		if (error == 0)
14056 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
14057 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
14058 		/*
14059 		 * Get our pacing rate, if an error
14060 		 * occured in sending (ENOBUF) we would
14061 		 * hit the else if with slot preset. Other
14062 		 * errors return.
14063 		 */
14064 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
14065 	}
14066 	if (rsm &&
14067 	    rack->use_rack_rr) {
14068 		/* Its a retransmit and we use the rack cheat? */
14069 		if ((slot == 0) ||
14070 		    (rack->rc_always_pace == 0) ||
14071 		    (rack->r_rr_config == 1)) {
14072 			/*
14073 			 * We have no pacing set or we
14074 			 * are using old-style rack or
14075 			 * we are overriden to use the old 1ms pacing.
14076 			 */
14077 			slot = rack->r_ctl.rc_min_to * HPTS_USEC_IN_MSEC;
14078 		}
14079 	}
14080 	if (slot) {
14081 		/* set the rack tcb into the slot N */
14082 		counter_u64_add(rack_paced_segments, 1);
14083 	} else if (sendalot) {
14084 		if (len)
14085 			counter_u64_add(rack_unpaced_segments, 1);
14086 		sack_rxmit = 0;
14087 		goto again;
14088 	} else if (len) {
14089 		counter_u64_add(rack_unpaced_segments, 1);
14090 	}
14091 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
14092 	return (error);
14093 }
14094 
14095 static void
14096 rack_update_seg(struct tcp_rack *rack)
14097 {
14098 	uint32_t orig_val;
14099 
14100 	orig_val = rack->r_ctl.rc_pace_max_segs;
14101 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__);
14102 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
14103 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL);
14104 }
14105 
14106 /*
14107  * rack_ctloutput() must drop the inpcb lock before performing copyin on
14108  * socket option arguments.  When it re-acquires the lock after the copy, it
14109  * has to revalidate that the connection is still valid for the socket
14110  * option.
14111  */
14112 static int
14113 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
14114     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
14115 {
14116 	struct epoch_tracker et;
14117 	uint64_t val;
14118 	int32_t error = 0, optval;
14119 	uint16_t ca, ss;
14120 
14121 
14122 	switch (sopt->sopt_name) {
14123 	case TCP_RACK_PROP_RATE:		/*  URL:prop_rate */
14124 	case TCP_RACK_PROP	:		/*  URL:prop */
14125 	case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
14126 	case TCP_RACK_EARLY_RECOV:		/*  URL:early_recov */
14127 	case TCP_RACK_PACE_REDUCE:		/*  Not used */
14128         /*  Pacing related ones */
14129 	case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
14130 	case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
14131 	case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
14132 	case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
14133 	case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
14134 	case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
14135 	case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
14136 	case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
14137 	case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
14138 	case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
14139 	case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
14140 	case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
14141 	case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
14142        /* End pacing related */
14143 	case TCP_DELACK:
14144 	case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
14145 	case TCP_RACK_MIN_TO:			/*  URL:min_to */
14146 	case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
14147 	case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
14148 	case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
14149 	case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
14150 	case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
14151 	case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
14152 	case TCP_RACK_TLP_INC_VAR:		/*  URL:tlp_inc_var */
14153 	case TCP_RACK_IDLE_REDUCE_HIGH:		/*  URL:idle_reduce_high */
14154 	case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
14155 	case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
14156 	case TCP_RACK_DO_DETECTION:		/*  URL:detect */
14157 	case TCP_NO_PRR:			/*  URL:noprr */
14158 	case TCP_TIMELY_DYN_ADJ:		/*  URL:dynamic */
14159 	case TCP_DATA_AFTER_CLOSE:
14160 	case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
14161 	case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
14162 	case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
14163 	case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
14164 	case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
14165 	case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
14166 	case TCP_RACK_PROFILE:			/*  URL:profile */
14167 		break;
14168 	default:
14169 		return (tcp_default_ctloutput(so, sopt, inp, tp));
14170 		break;
14171 	}
14172 	INP_WUNLOCK(inp);
14173 	error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14174 	if (error)
14175 		return (error);
14176 	INP_WLOCK(inp);
14177 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
14178 		INP_WUNLOCK(inp);
14179 		return (ECONNRESET);
14180 	}
14181 	tp = intotcpcb(inp);
14182 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14183 	switch (sopt->sopt_name) {
14184 	case TCP_RACK_PROFILE:
14185 		RACK_OPTS_INC(tcp_profile);
14186 		if (optval == 1) {
14187 			/* pace_always=1 */
14188 			rack->rc_always_pace = 1;
14189 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14190 			/* scwnd=1 */
14191 			rack->rack_enable_scwnd = 1;
14192 			/* dynamic=100 */
14193 			rack->rc_gp_dyn_mul = 1;
14194 			rack->r_ctl.rack_per_of_gp_ca = 100;
14195 			/* rrr_conf=3 */
14196 			rack->r_rr_config = 3;
14197 			/* npush=2 */
14198 			rack->r_ctl.rc_no_push_at_mrtt = 2;
14199 			/* fillcw=1 */
14200 			rack->rc_pace_to_cwnd = 1;
14201 			rack->rc_pace_fill_if_rttin_range = 0;
14202 			rack->rtt_limit_mul = 0;
14203 			/* noprr=1 */
14204 			rack->rack_no_prr = 1;
14205 			/* lscwnd=1 */
14206 			rack->r_limit_scw = 1;
14207 		} else if (optval == 2) {
14208 			/* pace_always=1 */
14209 			rack->rc_always_pace = 1;
14210 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14211 			/* scwnd=1 */
14212 			rack->rack_enable_scwnd = 1;
14213 			/* dynamic=100 */
14214 			rack->rc_gp_dyn_mul = 1;
14215 			rack->r_ctl.rack_per_of_gp_ca = 100;
14216 			/* rrr_conf=3 */
14217 			rack->r_rr_config = 3;
14218 			/* npush=2 */
14219 			rack->r_ctl.rc_no_push_at_mrtt = 2;
14220 			/* fillcw=1 */
14221 			rack->rc_pace_to_cwnd = 1;
14222 			rack->rc_pace_fill_if_rttin_range = 0;
14223 			rack->rtt_limit_mul = 0;
14224 			/* noprr=1 */
14225 			rack->rack_no_prr = 1;
14226 			/* lscwnd=0 */
14227 			rack->r_limit_scw = 0;
14228 		}
14229 		break;
14230 	case TCP_SHARED_CWND_TIME_LIMIT:
14231 		RACK_OPTS_INC(tcp_lscwnd);
14232 		if (optval)
14233 			rack->r_limit_scw = 1;
14234 		else
14235 			rack->r_limit_scw = 0;
14236 		break;
14237  	case TCP_RACK_PACE_TO_FILL:
14238 		RACK_OPTS_INC(tcp_fillcw);
14239 		if (optval == 0)
14240 			rack->rc_pace_to_cwnd = 0;
14241 		else
14242 			rack->rc_pace_to_cwnd = 1;
14243 		if ((optval >= rack_gp_rtt_maxmul) &&
14244 		    rack_gp_rtt_maxmul &&
14245 		    (optval < 0xf)) {
14246 			rack->rc_pace_fill_if_rttin_range = 1;
14247 			rack->rtt_limit_mul = optval;
14248 		} else {
14249 			rack->rc_pace_fill_if_rttin_range = 0;
14250 			rack->rtt_limit_mul = 0;
14251 		}
14252 		break;
14253 	case TCP_RACK_NO_PUSH_AT_MAX:
14254 		RACK_OPTS_INC(tcp_npush);
14255 		if (optval == 0)
14256 			rack->r_ctl.rc_no_push_at_mrtt = 0;
14257 		else if (optval < 0xff)
14258 			rack->r_ctl.rc_no_push_at_mrtt = optval;
14259 		else
14260 			error = EINVAL;
14261 		break;
14262 	case TCP_SHARED_CWND_ENABLE:
14263 		RACK_OPTS_INC(tcp_rack_scwnd);
14264 		if (optval == 0)
14265 			rack->rack_enable_scwnd = 0;
14266 		else
14267 			rack->rack_enable_scwnd = 1;
14268 		break;
14269 	case TCP_RACK_MBUF_QUEUE:
14270 		/* Now do we use the LRO mbuf-queue feature */
14271 		RACK_OPTS_INC(tcp_rack_mbufq);
14272 		if (optval)
14273 			rack->r_mbuf_queue = 1;
14274 		else
14275 			rack->r_mbuf_queue = 0;
14276 		if  (rack->r_mbuf_queue || rack->rc_always_pace)
14277 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14278 		else
14279 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
14280 		break;
14281 	case TCP_RACK_NONRXT_CFG_RATE:
14282 		RACK_OPTS_INC(tcp_rack_cfg_rate);
14283 		if (optval == 0)
14284 			rack->rack_rec_nonrxt_use_cr = 0;
14285 		else
14286 			rack->rack_rec_nonrxt_use_cr = 1;
14287 		break;
14288 	case TCP_NO_PRR:
14289 		RACK_OPTS_INC(tcp_rack_noprr);
14290 		if (optval == 0)
14291 			rack->rack_no_prr = 0;
14292 		else
14293 			rack->rack_no_prr = 1;
14294 		break;
14295 	case TCP_TIMELY_DYN_ADJ:
14296 		RACK_OPTS_INC(tcp_timely_dyn);
14297 		if (optval == 0)
14298 			rack->rc_gp_dyn_mul = 0;
14299 		else {
14300 			rack->rc_gp_dyn_mul = 1;
14301 			if (optval >= 100) {
14302 				/*
14303 				 * If the user sets something 100 or more
14304 				 * its the gp_ca value.
14305 				 */
14306 				rack->r_ctl.rack_per_of_gp_ca  = optval;
14307 			}
14308 		}
14309 		break;
14310 	case TCP_RACK_DO_DETECTION:
14311 		RACK_OPTS_INC(tcp_rack_do_detection);
14312 		if (optval == 0)
14313 			rack->do_detection = 0;
14314 		else
14315 			rack->do_detection = 1;
14316 		break;
14317 	case TCP_RACK_PROP_RATE:
14318 		if ((optval <= 0) || (optval >= 100)) {
14319 			error = EINVAL;
14320 			break;
14321 		}
14322 		RACK_OPTS_INC(tcp_rack_prop_rate);
14323 		rack->r_ctl.rc_prop_rate = optval;
14324 		break;
14325 	case TCP_RACK_TLP_USE:
14326 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
14327 			error = EINVAL;
14328 			break;
14329 		}
14330 		RACK_OPTS_INC(tcp_tlp_use);
14331 		rack->rack_tlp_threshold_use = optval;
14332 		break;
14333 	case TCP_RACK_PROP:
14334 		/* RACK proportional rate reduction (bool) */
14335 		RACK_OPTS_INC(tcp_rack_prop);
14336 		rack->r_ctl.rc_prop_reduce = optval;
14337 		break;
14338 	case TCP_RACK_TLP_REDUCE:
14339 		/* RACK TLP cwnd reduction (bool) */
14340 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
14341 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
14342 		break;
14343 	case TCP_RACK_EARLY_RECOV:
14344 		/* Should recovery happen early (bool) */
14345 		RACK_OPTS_INC(tcp_rack_early_recov);
14346 		rack->r_ctl.rc_early_recovery = optval;
14347 		break;
14348 
14349         /*  Pacing related ones */
14350 	case TCP_RACK_PACE_ALWAYS:
14351 		/*
14352 		 * zero is old rack method, 1 is new
14353 		 * method using a pacing rate.
14354 		 */
14355 		RACK_OPTS_INC(tcp_rack_pace_always);
14356 		if (optval > 0)
14357 			rack->rc_always_pace = 1;
14358 		else
14359 			rack->rc_always_pace = 0;
14360 		if  (rack->r_mbuf_queue || rack->rc_always_pace)
14361 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
14362 		else
14363 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
14364 		/* A rate may be set irate or other, if so set seg size */
14365 		rack_update_seg(rack);
14366 		break;
14367 	case TCP_BBR_RACK_INIT_RATE:
14368 		RACK_OPTS_INC(tcp_initial_rate);
14369 		val = optval;
14370 		/* Change from kbits per second to bytes per second */
14371 		val *= 1000;
14372 		val /= 8;
14373 		rack->r_ctl.init_rate = val;
14374 		if (rack->rc_init_win != rack_default_init_window) {
14375 			uint32_t win, snt;
14376 
14377 			/*
14378 			 * Options don't always get applied
14379 			 * in the order you think. So in order
14380 			 * to assure we update a cwnd we need
14381 			 * to check and see if we are still
14382 			 * where we should raise the cwnd.
14383 			 */
14384 			win = rc_init_window(rack);
14385 			if (SEQ_GT(tp->snd_max, tp->iss))
14386 				snt = tp->snd_max - tp->iss;
14387 			else
14388 				snt = 0;
14389 			if ((snt < win) &&
14390 			    (tp->snd_cwnd < win))
14391 				tp->snd_cwnd = win;
14392 		}
14393 		if (rack->rc_always_pace)
14394 			rack_update_seg(rack);
14395 		break;
14396 	case TCP_BBR_IWINTSO:
14397 		RACK_OPTS_INC(tcp_initial_win);
14398 		if (optval && (optval <= 0xff)) {
14399 			uint32_t win, snt;
14400 
14401 			rack->rc_init_win = optval;
14402 			win = rc_init_window(rack);
14403 			if (SEQ_GT(tp->snd_max, tp->iss))
14404 				snt = tp->snd_max - tp->iss;
14405 			else
14406 				snt = 0;
14407 			if ((snt < win) &&
14408 			    (tp->t_srtt |
14409 #ifdef NETFLIX_PEAKRATE
14410 			     tp->t_maxpeakrate |
14411 #endif
14412 			     rack->r_ctl.init_rate)) {
14413 				/*
14414 				 * We are not past the initial window
14415 				 * and we have some bases for pacing,
14416 				 * so we need to possibly adjust up
14417 				 * the cwnd. Note even if we don't set
14418 				 * the cwnd, its still ok to raise the rc_init_win
14419 				 * which can be used coming out of idle when we
14420 				 * would have a rate.
14421 				 */
14422 				if (tp->snd_cwnd < win)
14423 					tp->snd_cwnd = win;
14424 			}
14425 			if (rack->rc_always_pace)
14426 				rack_update_seg(rack);
14427 		} else
14428 			error = EINVAL;
14429 		break;
14430 	case TCP_RACK_FORCE_MSEG:
14431 		RACK_OPTS_INC(tcp_rack_force_max_seg);
14432 		if (optval)
14433 			rack->rc_force_max_seg = 1;
14434 		else
14435 			rack->rc_force_max_seg = 0;
14436 		break;
14437 	case TCP_RACK_PACE_MAX_SEG:
14438 		/* Max segments size in a pace in bytes */
14439 		RACK_OPTS_INC(tcp_rack_max_seg);
14440 		rack->rc_user_set_max_segs = optval;
14441 		rack_set_pace_segments(tp, rack, __LINE__);
14442 		break;
14443 	case TCP_RACK_PACE_RATE_REC:
14444 		/* Set the fixed pacing rate in Bytes per second ca */
14445 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
14446 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
14447 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
14448 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
14449 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
14450 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
14451 		rack->use_fixed_rate = 1;
14452 		rack_log_pacing_delay_calc(rack,
14453 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
14454 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
14455 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
14456 					   __LINE__, NULL);
14457 		break;
14458 
14459 	case TCP_RACK_PACE_RATE_SS:
14460 		/* Set the fixed pacing rate in Bytes per second ca */
14461 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
14462 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
14463 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
14464 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
14465 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
14466 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
14467 		rack->use_fixed_rate = 1;
14468 		rack_log_pacing_delay_calc(rack,
14469 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
14470 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
14471 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
14472 					   __LINE__, NULL);
14473 		break;
14474 
14475 	case TCP_RACK_PACE_RATE_CA:
14476 		/* Set the fixed pacing rate in Bytes per second ca */
14477 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
14478 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
14479 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
14480 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
14481 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
14482 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
14483 		rack->use_fixed_rate = 1;
14484 		rack_log_pacing_delay_calc(rack,
14485 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
14486 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
14487 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
14488 					   __LINE__, NULL);
14489 		break;
14490 	case TCP_RACK_GP_INCREASE_REC:
14491 		RACK_OPTS_INC(tcp_gp_inc_rec);
14492 		rack->r_ctl.rack_per_of_gp_rec = optval;
14493 		rack_log_pacing_delay_calc(rack,
14494 					   rack->r_ctl.rack_per_of_gp_ss,
14495 					   rack->r_ctl.rack_per_of_gp_ca,
14496 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
14497 					   __LINE__, NULL);
14498 		break;
14499 	case TCP_RACK_GP_INCREASE_CA:
14500 		RACK_OPTS_INC(tcp_gp_inc_ca);
14501 		ca = optval;
14502 		if (ca < 100) {
14503 			/*
14504 			 * We don't allow any reduction
14505 			 * over the GP b/w.
14506 			 */
14507 			error = EINVAL;
14508 			break;
14509 		}
14510 		rack->r_ctl.rack_per_of_gp_ca = ca;
14511 		rack_log_pacing_delay_calc(rack,
14512 					   rack->r_ctl.rack_per_of_gp_ss,
14513 					   rack->r_ctl.rack_per_of_gp_ca,
14514 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
14515 					   __LINE__, NULL);
14516 		break;
14517 	case TCP_RACK_GP_INCREASE_SS:
14518 		RACK_OPTS_INC(tcp_gp_inc_ss);
14519 		ss = optval;
14520 		if (ss < 100) {
14521 			/*
14522 			 * We don't allow any reduction
14523 			 * over the GP b/w.
14524 			 */
14525 			error = EINVAL;
14526 			break;
14527 		}
14528 		rack->r_ctl.rack_per_of_gp_ss = ss;
14529 		rack_log_pacing_delay_calc(rack,
14530 					   rack->r_ctl.rack_per_of_gp_ss,
14531 					   rack->r_ctl.rack_per_of_gp_ca,
14532 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
14533 					   __LINE__, NULL);
14534 		break;
14535 	case TCP_RACK_RR_CONF:
14536 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
14537 		if (optval && optval <= 3)
14538 			rack->r_rr_config = optval;
14539 		else
14540 			rack->r_rr_config = 0;
14541 		break;
14542 	case TCP_BBR_HDWR_PACE:
14543 		RACK_OPTS_INC(tcp_hdwr_pacing);
14544 		if (optval){
14545 			if (rack->rack_hdrw_pacing == 0) {
14546 				rack->rack_hdw_pace_ena = 1;
14547 				rack->rack_attempt_hdwr_pace = 0;
14548 			} else
14549 				error = EALREADY;
14550 		} else {
14551 			rack->rack_hdw_pace_ena = 0;
14552 #ifdef RATELIMIT
14553 			if (rack->rack_hdrw_pacing) {
14554 				rack->rack_hdrw_pacing = 0;
14555 				in_pcbdetach_txrtlmt(rack->rc_inp);
14556 			}
14557 #endif
14558 		}
14559 		break;
14560         /*  End Pacing related ones */
14561 	case TCP_RACK_PRR_SENDALOT:
14562 		/* Allow PRR to send more than one seg */
14563 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
14564 		rack->r_ctl.rc_prr_sendalot = optval;
14565 		break;
14566 	case TCP_RACK_MIN_TO:
14567 		/* Minimum time between rack t-o's in ms */
14568 		RACK_OPTS_INC(tcp_rack_min_to);
14569 		rack->r_ctl.rc_min_to = optval;
14570 		break;
14571 	case TCP_RACK_EARLY_SEG:
14572 		/* If early recovery max segments */
14573 		RACK_OPTS_INC(tcp_rack_early_seg);
14574 		rack->r_ctl.rc_early_recovery_segs = optval;
14575 		break;
14576 	case TCP_RACK_REORD_THRESH:
14577 		/* RACK reorder threshold (shift amount) */
14578 		RACK_OPTS_INC(tcp_rack_reord_thresh);
14579 		if ((optval > 0) && (optval < 31))
14580 			rack->r_ctl.rc_reorder_shift = optval;
14581 		else
14582 			error = EINVAL;
14583 		break;
14584 	case TCP_RACK_REORD_FADE:
14585 		/* Does reordering fade after ms time */
14586 		RACK_OPTS_INC(tcp_rack_reord_fade);
14587 		rack->r_ctl.rc_reorder_fade = optval;
14588 		break;
14589 	case TCP_RACK_TLP_THRESH:
14590 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14591 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
14592 		if (optval)
14593 			rack->r_ctl.rc_tlp_threshold = optval;
14594 		else
14595 			error = EINVAL;
14596 		break;
14597 	case TCP_BBR_USE_RACK_RR:
14598 		RACK_OPTS_INC(tcp_rack_rr);
14599 		if (optval)
14600 			rack->use_rack_rr = 1;
14601 		else
14602 			rack->use_rack_rr = 0;
14603 		break;
14604 	case TCP_RACK_PKT_DELAY:
14605 		/* RACK added ms i.e. rack-rtt + reord + N */
14606 		RACK_OPTS_INC(tcp_rack_pkt_delay);
14607 		rack->r_ctl.rc_pkt_delay = optval;
14608 		break;
14609 	case TCP_RACK_TLP_INC_VAR:
14610 		/* Does TLP include rtt variance in t-o */
14611 		error = EINVAL;
14612 		break;
14613 	case TCP_RACK_IDLE_REDUCE_HIGH:
14614 		error = EINVAL;
14615 		break;
14616 	case TCP_DELACK:
14617 		if (optval == 0)
14618 			tp->t_delayed_ack = 0;
14619 		else
14620 			tp->t_delayed_ack = 1;
14621 		if (tp->t_flags & TF_DELACK) {
14622 			tp->t_flags &= ~TF_DELACK;
14623 			tp->t_flags |= TF_ACKNOW;
14624 			NET_EPOCH_ENTER(et);
14625 			rack_output(tp);
14626 			NET_EPOCH_EXIT(et);
14627 		}
14628 		break;
14629 
14630 	case TCP_BBR_RACK_RTT_USE:
14631 		if ((optval != USE_RTT_HIGH) &&
14632 		    (optval != USE_RTT_LOW) &&
14633 		    (optval != USE_RTT_AVG))
14634 			error = EINVAL;
14635 		else
14636 			rack->r_ctl.rc_rate_sample_method = optval;
14637 		break;
14638 	case TCP_DATA_AFTER_CLOSE:
14639 		if (optval)
14640 			rack->rc_allow_data_af_clo = 1;
14641 		else
14642 			rack->rc_allow_data_af_clo = 0;
14643 		break;
14644 	case TCP_RACK_PACE_REDUCE:
14645 		/* sysctl only now */
14646 		error = EINVAL;
14647 		break;
14648 	default:
14649 		return (tcp_default_ctloutput(so, sopt, inp, tp));
14650 		break;
14651 	}
14652 #ifdef NETFLIX_STATS
14653 	tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14654 #endif
14655 	INP_WUNLOCK(inp);
14656 	return (error);
14657 }
14658 
14659 static int
14660 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
14661     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
14662 {
14663 	int32_t error, optval;
14664 	uint64_t val;
14665 	/*
14666 	 * Because all our options are either boolean or an int, we can just
14667 	 * pull everything into optval and then unlock and copy. If we ever
14668 	 * add a option that is not a int, then this will have quite an
14669 	 * impact to this routine.
14670 	 */
14671 	error = 0;
14672 	switch (sopt->sopt_name) {
14673 	case TCP_RACK_PROFILE:
14674 		/* You cannot retrieve a profile, its write only */
14675 		error = EINVAL;
14676 		break;
14677 	case TCP_RACK_PACE_TO_FILL:
14678 		optval = rack->rc_pace_to_cwnd;
14679 		break;
14680 	case TCP_RACK_NO_PUSH_AT_MAX:
14681 		optval = rack->r_ctl.rc_no_push_at_mrtt;
14682 		break;
14683 	case TCP_SHARED_CWND_ENABLE:
14684 		optval = rack->rack_enable_scwnd;
14685 		break;
14686 	case TCP_RACK_NONRXT_CFG_RATE:
14687 		optval = rack->rack_rec_nonrxt_use_cr;
14688 		break;
14689 	case TCP_NO_PRR:
14690 		optval = rack->rack_no_prr;
14691 		break;
14692 	case TCP_RACK_DO_DETECTION:
14693 		optval = rack->do_detection;
14694 		break;
14695 	case TCP_RACK_MBUF_QUEUE:
14696 		/* Now do we use the LRO mbuf-queue feature */
14697 		optval = rack->r_mbuf_queue;
14698 		break;
14699 	case TCP_TIMELY_DYN_ADJ:
14700 		optval = rack->rc_gp_dyn_mul;
14701 		break;
14702 	case TCP_BBR_IWINTSO:
14703 		optval = rack->rc_init_win;
14704 		break;
14705 	case TCP_RACK_PROP_RATE:
14706 		optval = rack->r_ctl.rc_prop_rate;
14707 		break;
14708 	case TCP_RACK_PROP:
14709 		/* RACK proportional rate reduction (bool) */
14710 		optval = rack->r_ctl.rc_prop_reduce;
14711 		break;
14712 	case TCP_RACK_TLP_REDUCE:
14713 		/* RACK TLP cwnd reduction (bool) */
14714 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
14715 		break;
14716 	case TCP_RACK_EARLY_RECOV:
14717 		/* Should recovery happen early (bool) */
14718 		optval = rack->r_ctl.rc_early_recovery;
14719 		break;
14720 	case TCP_RACK_PACE_REDUCE:
14721 		/* RACK Hptsi reduction factor (divisor) */
14722 		error = EINVAL;
14723 		break;
14724 	case TCP_BBR_RACK_INIT_RATE:
14725 		val = rack->r_ctl.init_rate;
14726 		/* convert to kbits per sec */
14727 		val *= 8;
14728 		val /= 1000;
14729 		optval = (uint32_t)val;
14730 		break;
14731 	case TCP_RACK_FORCE_MSEG:
14732 		optval = rack->rc_force_max_seg;
14733 		break;
14734 	case TCP_RACK_PACE_MAX_SEG:
14735 		/* Max segments in a pace */
14736 		optval = rack->rc_user_set_max_segs;
14737 		break;
14738 	case TCP_RACK_PACE_ALWAYS:
14739 		/* Use the always pace method */
14740 		optval = rack->rc_always_pace;
14741 		break;
14742 	case TCP_RACK_PRR_SENDALOT:
14743 		/* Allow PRR to send more than one seg */
14744 		optval = rack->r_ctl.rc_prr_sendalot;
14745 		break;
14746 	case TCP_RACK_MIN_TO:
14747 		/* Minimum time between rack t-o's in ms */
14748 		optval = rack->r_ctl.rc_min_to;
14749 		break;
14750 	case TCP_RACK_EARLY_SEG:
14751 		/* If early recovery max segments */
14752 		optval = rack->r_ctl.rc_early_recovery_segs;
14753 		break;
14754 	case TCP_RACK_REORD_THRESH:
14755 		/* RACK reorder threshold (shift amount) */
14756 		optval = rack->r_ctl.rc_reorder_shift;
14757 		break;
14758 	case TCP_RACK_REORD_FADE:
14759 		/* Does reordering fade after ms time */
14760 		optval = rack->r_ctl.rc_reorder_fade;
14761 		break;
14762 	case TCP_BBR_USE_RACK_RR:
14763 		/* Do we use the rack cheat for rxt */
14764 		optval = rack->use_rack_rr;
14765 		break;
14766 	case TCP_RACK_RR_CONF:
14767 		optval = rack->r_rr_config;
14768 		break;
14769 	case TCP_BBR_HDWR_PACE:
14770 		optval = rack->rack_hdw_pace_ena;
14771 		break;
14772 	case TCP_RACK_TLP_THRESH:
14773 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
14774 		optval = rack->r_ctl.rc_tlp_threshold;
14775 		break;
14776 	case TCP_RACK_PKT_DELAY:
14777 		/* RACK added ms i.e. rack-rtt + reord + N */
14778 		optval = rack->r_ctl.rc_pkt_delay;
14779 		break;
14780 	case TCP_RACK_TLP_USE:
14781 		optval = rack->rack_tlp_threshold_use;
14782 		break;
14783 	case TCP_RACK_TLP_INC_VAR:
14784 		/* Does TLP include rtt variance in t-o */
14785 		error = EINVAL;
14786 		break;
14787 	case TCP_RACK_IDLE_REDUCE_HIGH:
14788 		error = EINVAL;
14789 		break;
14790 	case TCP_RACK_PACE_RATE_CA:
14791 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
14792 		break;
14793 	case TCP_RACK_PACE_RATE_SS:
14794 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
14795 		break;
14796 	case TCP_RACK_PACE_RATE_REC:
14797 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
14798 		break;
14799 	case TCP_RACK_GP_INCREASE_SS:
14800 		optval = rack->r_ctl.rack_per_of_gp_ca;
14801 		break;
14802 	case TCP_RACK_GP_INCREASE_CA:
14803 		optval = rack->r_ctl.rack_per_of_gp_ss;
14804 		break;
14805 	case TCP_BBR_RACK_RTT_USE:
14806 		optval = rack->r_ctl.rc_rate_sample_method;
14807 		break;
14808 	case TCP_DELACK:
14809 		optval = tp->t_delayed_ack;
14810 		break;
14811 	case TCP_DATA_AFTER_CLOSE:
14812 		optval = rack->rc_allow_data_af_clo;
14813 		break;
14814 	case TCP_SHARED_CWND_TIME_LIMIT:
14815 		optval = rack->r_limit_scw;
14816 		break;
14817 	default:
14818 		return (tcp_default_ctloutput(so, sopt, inp, tp));
14819 		break;
14820 	}
14821 	INP_WUNLOCK(inp);
14822 	if (error == 0) {
14823 		error = sooptcopyout(sopt, &optval, sizeof optval);
14824 	}
14825 	return (error);
14826 }
14827 
14828 static int
14829 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
14830 {
14831 	int32_t error = EINVAL;
14832 	struct tcp_rack *rack;
14833 
14834 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14835 	if (rack == NULL) {
14836 		/* Huh? */
14837 		goto out;
14838 	}
14839 	if (sopt->sopt_dir == SOPT_SET) {
14840 		return (rack_set_sockopt(so, sopt, inp, tp, rack));
14841 	} else if (sopt->sopt_dir == SOPT_GET) {
14842 		return (rack_get_sockopt(so, sopt, inp, tp, rack));
14843 	}
14844 out:
14845 	INP_WUNLOCK(inp);
14846 	return (error);
14847 }
14848 
14849 static int
14850 rack_pru_options(struct tcpcb *tp, int flags)
14851 {
14852 	if (flags & PRUS_OOB)
14853 		return (EOPNOTSUPP);
14854 	return (0);
14855 }
14856 
14857 static struct tcp_function_block __tcp_rack = {
14858 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
14859 	.tfb_tcp_output = rack_output,
14860 	.tfb_do_queued_segments = ctf_do_queued_segments,
14861 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
14862 	.tfb_tcp_do_segment = rack_do_segment,
14863 	.tfb_tcp_ctloutput = rack_ctloutput,
14864 	.tfb_tcp_fb_init = rack_init,
14865 	.tfb_tcp_fb_fini = rack_fini,
14866 	.tfb_tcp_timer_stop_all = rack_stopall,
14867 	.tfb_tcp_timer_activate = rack_timer_activate,
14868 	.tfb_tcp_timer_active = rack_timer_active,
14869 	.tfb_tcp_timer_stop = rack_timer_stop,
14870 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
14871 	.tfb_tcp_handoff_ok = rack_handoff_ok,
14872 	.tfb_pru_options = rack_pru_options,
14873 };
14874 
14875 static const char *rack_stack_names[] = {
14876 	__XSTRING(STACKNAME),
14877 #ifdef STACKALIAS
14878 	__XSTRING(STACKALIAS),
14879 #endif
14880 };
14881 
14882 static int
14883 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
14884 {
14885 	memset(mem, 0, size);
14886 	return (0);
14887 }
14888 
14889 static void
14890 rack_dtor(void *mem, int32_t size, void *arg)
14891 {
14892 
14893 }
14894 
14895 static bool rack_mod_inited = false;
14896 
14897 static int
14898 tcp_addrack(module_t mod, int32_t type, void *data)
14899 {
14900 	int32_t err = 0;
14901 	int num_stacks;
14902 
14903 	switch (type) {
14904 	case MOD_LOAD:
14905 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
14906 		    sizeof(struct rack_sendmap),
14907 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
14908 
14909 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
14910 		    sizeof(struct tcp_rack),
14911 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
14912 
14913 		sysctl_ctx_init(&rack_sysctl_ctx);
14914 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
14915 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
14916 		    OID_AUTO,
14917 #ifdef STACKALIAS
14918 		    __XSTRING(STACKALIAS),
14919 #else
14920 		    __XSTRING(STACKNAME),
14921 #endif
14922 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
14923 		    "");
14924 		if (rack_sysctl_root == NULL) {
14925 			printf("Failed to add sysctl node\n");
14926 			err = EFAULT;
14927 			goto free_uma;
14928 		}
14929 		rack_init_sysctls();
14930 		num_stacks = nitems(rack_stack_names);
14931 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
14932 		    rack_stack_names, &num_stacks);
14933 		if (err) {
14934 			printf("Failed to register %s stack name for "
14935 			    "%s module\n", rack_stack_names[num_stacks],
14936 			    __XSTRING(MODNAME));
14937 			sysctl_ctx_free(&rack_sysctl_ctx);
14938 free_uma:
14939 			uma_zdestroy(rack_zone);
14940 			uma_zdestroy(rack_pcb_zone);
14941 			rack_counter_destroy();
14942 			printf("Failed to register rack module -- err:%d\n", err);
14943 			return (err);
14944 		}
14945 		tcp_lro_reg_mbufq();
14946 		rack_mod_inited = true;
14947 		break;
14948 	case MOD_QUIESCE:
14949 		err = deregister_tcp_functions(&__tcp_rack, true, false);
14950 		break;
14951 	case MOD_UNLOAD:
14952 		err = deregister_tcp_functions(&__tcp_rack, false, true);
14953 		if (err == EBUSY)
14954 			break;
14955 		if (rack_mod_inited) {
14956 			uma_zdestroy(rack_zone);
14957 			uma_zdestroy(rack_pcb_zone);
14958 			sysctl_ctx_free(&rack_sysctl_ctx);
14959 			rack_counter_destroy();
14960 			rack_mod_inited = false;
14961 		}
14962 		tcp_lro_dereg_mbufq();
14963 		err = 0;
14964 		break;
14965 	default:
14966 		return (EOPNOTSUPP);
14967 	}
14968 	return (err);
14969 }
14970 
14971 static moduledata_t tcp_rack = {
14972 	.name = __XSTRING(MODNAME),
14973 	.evhand = tcp_addrack,
14974 	.priv = 0
14975 };
14976 
14977 MODULE_VERSION(MODNAME, 1);
14978 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
14979 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
14980